xref: /freebsd/sys/dev/mthca/mthca_qp.c (revision 64038db825d64fb4827fc8ee264ea0fa1a046d82)
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 } __attribute__((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 } __attribute__((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 } __attribute__((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, path->static_rate & 0xf, port_num));
415 	if (path->g_mylmc & (1 << 7)) {
416 		u32 tc_fl = be32_to_cpu(path->sl_tclass_flowlabel);
417 
418 		rdma_ah_set_grh(ah_attr, NULL,
419 				tc_fl & 0xfffff,
420 				path->mgid_index &
421 				(dev->limits.gid_table_len - 1),
422 				path->hop_limit,
423 				(tc_fl >> 20) & 0xff);
424 		rdma_ah_set_dgid_raw(ah_attr, path->rgid);
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 
439 	mutex_lock(&qp->mutex);
440 
441 	if (qp->state == IB_QPS_RESET) {
442 		qp_attr->qp_state = IB_QPS_RESET;
443 		goto done;
444 	}
445 
446 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
447 	if (IS_ERR(mailbox)) {
448 		err = PTR_ERR(mailbox);
449 		goto out;
450 	}
451 
452 	err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
453 	if (err) {
454 		mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
455 		goto out_mailbox;
456 	}
457 
458 	qp_param    = mailbox->buf;
459 	context     = &qp_param->context;
460 	mthca_state = be32_to_cpu(context->flags) >> 28;
461 
462 	qp->state		     = to_ib_qp_state(mthca_state);
463 	qp_attr->qp_state	     = qp->state;
464 	qp_attr->path_mtu 	     = context->mtu_msgmax >> 5;
465 	qp_attr->path_mig_state      =
466 		to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
467 	qp_attr->qkey 		     = be32_to_cpu(context->qkey);
468 	qp_attr->rq_psn 	     = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
469 	qp_attr->sq_psn 	     = be32_to_cpu(context->next_send_psn) & 0xffffff;
470 	qp_attr->dest_qp_num 	     = be32_to_cpu(context->remote_qpn) & 0xffffff;
471 	qp_attr->qp_access_flags     =
472 		to_ib_qp_access_flags(be32_to_cpu(context->params2));
473 
474 	if (qp->transport == RC || qp->transport == UC) {
475 		to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
476 		to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
477 		qp_attr->alt_pkey_index =
478 			be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
479 		qp_attr->alt_port_num 	=
480 			rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
481 	}
482 
483 	qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
484 	qp_attr->port_num   =
485 		(be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
486 
487 	/* qp_attr->en_sqd_async_notify is only applicable in modify qp */
488 	qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
489 
490 	qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
491 
492 	qp_attr->max_dest_rd_atomic =
493 		1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
494 	qp_attr->min_rnr_timer 	    =
495 		(be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
496 	qp_attr->timeout 	    = context->pri_path.ackto >> 3;
497 	qp_attr->retry_cnt 	    = (be32_to_cpu(context->params1) >> 16) & 0x7;
498 	qp_attr->rnr_retry 	    = context->pri_path.rnr_retry >> 5;
499 	qp_attr->alt_timeout 	    = context->alt_path.ackto >> 3;
500 
501 done:
502 	qp_attr->cur_qp_state	     = qp_attr->qp_state;
503 	qp_attr->cap.max_send_wr     = qp->sq.max;
504 	qp_attr->cap.max_recv_wr     = qp->rq.max;
505 	qp_attr->cap.max_send_sge    = qp->sq.max_gs;
506 	qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
507 	qp_attr->cap.max_inline_data = qp->max_inline_data;
508 
509 	qp_init_attr->cap	     = qp_attr->cap;
510 	qp_init_attr->sq_sig_type    = qp->sq_policy;
511 
512 out_mailbox:
513 	mthca_free_mailbox(dev, mailbox);
514 
515 out:
516 	mutex_unlock(&qp->mutex);
517 	return err;
518 }
519 
520 static int mthca_path_set(struct mthca_dev *dev, const struct rdma_ah_attr *ah,
521 			  struct mthca_qp_path *path, u8 port)
522 {
523 	path->g_mylmc     = rdma_ah_get_path_bits(ah) & 0x7f;
524 	path->rlid        = cpu_to_be16(rdma_ah_get_dlid(ah));
525 	path->static_rate = mthca_get_rate(dev, rdma_ah_get_static_rate(ah),
526 					   port);
527 
528 	if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
529 		const struct ib_global_route *grh = rdma_ah_read_grh(ah);
530 
531 		if (grh->sgid_index >= dev->limits.gid_table_len) {
532 			mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
533 				  grh->sgid_index, dev->limits.gid_table_len-1);
534 			return -1;
535 		}
536 
537 		path->g_mylmc   |= 1 << 7;
538 		path->mgid_index = grh->sgid_index;
539 		path->hop_limit  = grh->hop_limit;
540 		path->sl_tclass_flowlabel =
541 			cpu_to_be32((rdma_ah_get_sl(ah) << 28) |
542 				    (grh->traffic_class << 20) |
543 				    (grh->flow_label));
544 		memcpy(path->rgid, grh->dgid.raw, 16);
545 	} else
546 		path->sl_tclass_flowlabel = cpu_to_be32(rdma_ah_get_sl(ah) <<
547 							28);
548 
549 	return 0;
550 }
551 
552 static int __mthca_modify_qp(struct ib_qp *ibqp,
553 			     const struct ib_qp_attr *attr, int attr_mask,
554 			     enum ib_qp_state cur_state,
555 			     enum ib_qp_state new_state,
556 			     struct ib_udata *udata)
557 {
558 	struct mthca_dev *dev = to_mdev(ibqp->device);
559 	struct mthca_qp *qp = to_mqp(ibqp);
560 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
561 		udata, struct mthca_ucontext, ibucontext);
562 	struct mthca_mailbox *mailbox;
563 	struct mthca_qp_param *qp_param;
564 	struct mthca_qp_context *qp_context;
565 	u32 sqd_event = 0;
566 	int err = -EINVAL;
567 
568 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
569 	if (IS_ERR(mailbox)) {
570 		err = PTR_ERR(mailbox);
571 		goto out;
572 	}
573 	qp_param = mailbox->buf;
574 	qp_context = &qp_param->context;
575 	memset(qp_param, 0, sizeof *qp_param);
576 
577 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
578 					     (to_mthca_st(qp->transport) << 16));
579 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
580 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
581 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
582 	else {
583 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
584 		switch (attr->path_mig_state) {
585 		case IB_MIG_MIGRATED:
586 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
587 			break;
588 		case IB_MIG_REARM:
589 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
590 			break;
591 		case IB_MIG_ARMED:
592 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
593 			break;
594 		}
595 	}
596 
597 	/* leave tavor_sched_queue as 0 */
598 
599 	if (qp->transport == MLX || qp->transport == UD)
600 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
601 	else if (attr_mask & IB_QP_PATH_MTU) {
602 		if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
603 			mthca_dbg(dev, "path MTU (%u) is invalid\n",
604 				  attr->path_mtu);
605 			goto out_mailbox;
606 		}
607 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
608 	}
609 
610 	if (mthca_is_memfree(dev)) {
611 		if (qp->rq.max)
612 			qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
613 		qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
614 
615 		if (qp->sq.max)
616 			qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
617 		qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
618 	}
619 
620 	/* leave arbel_sched_queue as 0 */
621 
622 	if (qp->ibqp.uobject)
623 		qp_context->usr_page = cpu_to_be32(context->uar.index);
624 	else
625 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
626 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
627 	if (attr_mask & IB_QP_DEST_QPN) {
628 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
629 	}
630 
631 	if (qp->transport == MLX)
632 		qp_context->pri_path.port_pkey |=
633 			cpu_to_be32(qp->port << 24);
634 	else {
635 		if (attr_mask & IB_QP_PORT) {
636 			qp_context->pri_path.port_pkey |=
637 				cpu_to_be32(attr->port_num << 24);
638 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
639 		}
640 	}
641 
642 	if (attr_mask & IB_QP_PKEY_INDEX) {
643 		qp_context->pri_path.port_pkey |=
644 			cpu_to_be32(attr->pkey_index);
645 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
646 	}
647 
648 	if (attr_mask & IB_QP_RNR_RETRY) {
649 		qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
650 			attr->rnr_retry << 5;
651 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
652 							MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
653 	}
654 
655 	if (attr_mask & IB_QP_AV) {
656 		if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
657 				   attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
658 			goto out_mailbox;
659 
660 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
661 	}
662 
663 	if (ibqp->qp_type == IB_QPT_RC &&
664 	    cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
665 		u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
666 
667 		if (mthca_is_memfree(dev))
668 			qp_context->rlkey_arbel_sched_queue |= sched_queue;
669 		else
670 			qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
671 
672 		qp_param->opt_param_mask |=
673 			cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
674 	}
675 
676 	if (attr_mask & IB_QP_TIMEOUT) {
677 		qp_context->pri_path.ackto = attr->timeout << 3;
678 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
679 	}
680 
681 	if (attr_mask & IB_QP_ALT_PATH) {
682 		if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
683 			mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
684 				  attr->alt_pkey_index, dev->limits.pkey_table_len-1);
685 			goto out_mailbox;
686 		}
687 
688 		if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
689 			mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
690 				attr->alt_port_num);
691 			goto out_mailbox;
692 		}
693 
694 		if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
695 				   rdma_ah_get_port_num(&attr->alt_ah_attr)))
696 			goto out_mailbox;
697 
698 		qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
699 							      attr->alt_port_num << 24);
700 		qp_context->alt_path.ackto = attr->alt_timeout << 3;
701 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
702 	}
703 
704 	/* leave rdd as 0 */
705 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
706 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
707 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
708 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
709 					     (MTHCA_FLIGHT_LIMIT << 24) |
710 					     MTHCA_QP_BIT_SWE);
711 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
712 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
713 	if (attr_mask & IB_QP_RETRY_CNT) {
714 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
715 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
716 	}
717 
718 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
719 		if (attr->max_rd_atomic) {
720 			qp_context->params1 |=
721 				cpu_to_be32(MTHCA_QP_BIT_SRE |
722 					    MTHCA_QP_BIT_SAE);
723 			qp_context->params1 |=
724 				cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
725 		}
726 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
727 	}
728 
729 	if (attr_mask & IB_QP_SQ_PSN)
730 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
731 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
732 
733 	if (mthca_is_memfree(dev)) {
734 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
735 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
736 	}
737 
738 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
739 		if (attr->max_dest_rd_atomic)
740 			qp_context->params2 |=
741 				cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
742 
743 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
744 	}
745 
746 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
747 		qp_context->params2      |= get_hw_access_flags(qp, attr, attr_mask);
748 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
749 							MTHCA_QP_OPTPAR_RRE |
750 							MTHCA_QP_OPTPAR_RAE);
751 	}
752 
753 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
754 
755 	if (ibqp->srq)
756 		qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
757 
758 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
759 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
760 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
761 	}
762 	if (attr_mask & IB_QP_RQ_PSN)
763 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
764 
765 	qp_context->ra_buff_indx =
766 		cpu_to_be32(dev->qp_table.rdb_base +
767 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
768 			     dev->qp_table.rdb_shift));
769 
770 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
771 
772 	if (mthca_is_memfree(dev))
773 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
774 
775 	if (attr_mask & IB_QP_QKEY) {
776 		qp_context->qkey = cpu_to_be32(attr->qkey);
777 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
778 	}
779 
780 	if (ibqp->srq)
781 		qp_context->srqn = cpu_to_be32(1 << 24 |
782 					       to_msrq(ibqp->srq)->srqn);
783 
784 	if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD	&&
785 	    attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY		&&
786 	    attr->en_sqd_async_notify)
787 		sqd_event = 1 << 31;
788 
789 	err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
790 			      mailbox, sqd_event);
791 	if (err) {
792 		mthca_warn(dev, "modify QP %d->%d returned %d.\n",
793 			   cur_state, new_state, err);
794 		goto out_mailbox;
795 	}
796 
797 	qp->state = new_state;
798 	if (attr_mask & IB_QP_ACCESS_FLAGS)
799 		qp->atomic_rd_en = attr->qp_access_flags;
800 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
801 		qp->resp_depth = attr->max_dest_rd_atomic;
802 	if (attr_mask & IB_QP_PORT)
803 		qp->port = attr->port_num;
804 	if (attr_mask & IB_QP_ALT_PATH)
805 		qp->alt_port = attr->alt_port_num;
806 
807 	if (is_sqp(dev, qp))
808 		store_attrs(to_msqp(qp), attr, attr_mask);
809 
810 	/*
811 	 * If we moved QP0 to RTR, bring the IB link up; if we moved
812 	 * QP0 to RESET or ERROR, bring the link back down.
813 	 */
814 	if (is_qp0(dev, qp)) {
815 		if (cur_state != IB_QPS_RTR &&
816 		    new_state == IB_QPS_RTR)
817 			init_port(dev, qp->port);
818 
819 		if (cur_state != IB_QPS_RESET &&
820 		    cur_state != IB_QPS_ERR &&
821 		    (new_state == IB_QPS_RESET ||
822 		     new_state == IB_QPS_ERR))
823 			mthca_CLOSE_IB(dev, qp->port);
824 	}
825 
826 	/*
827 	 * If we moved a kernel QP to RESET, clean up all old CQ
828 	 * entries and reinitialize the QP.
829 	 */
830 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
831 		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
832 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
833 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
834 			mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
835 
836 		mthca_wq_reset(&qp->sq);
837 		qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
838 
839 		mthca_wq_reset(&qp->rq);
840 		qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
841 
842 		if (mthca_is_memfree(dev)) {
843 			*qp->sq.db = 0;
844 			*qp->rq.db = 0;
845 		}
846 	}
847 
848 out_mailbox:
849 	mthca_free_mailbox(dev, mailbox);
850 out:
851 	return err;
852 }
853 
854 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
855 		    struct ib_udata *udata)
856 {
857 	struct mthca_dev *dev = to_mdev(ibqp->device);
858 	struct mthca_qp *qp = to_mqp(ibqp);
859 	enum ib_qp_state cur_state, new_state;
860 	int err = -EINVAL;
861 
862 	mutex_lock(&qp->mutex);
863 	if (attr_mask & IB_QP_CUR_STATE) {
864 		cur_state = attr->cur_qp_state;
865 	} else {
866 		spin_lock_irq(&qp->sq.lock);
867 		spin_lock(&qp->rq.lock);
868 		cur_state = qp->state;
869 		spin_unlock(&qp->rq.lock);
870 		spin_unlock_irq(&qp->sq.lock);
871 	}
872 
873 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
874 
875 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
876 		mthca_dbg(dev, "Bad QP transition (transport %d) "
877 			  "%d->%d with attr 0x%08x\n",
878 			  qp->transport, cur_state, new_state,
879 			  attr_mask);
880 		goto out;
881 	}
882 
883 	if ((attr_mask & IB_QP_PKEY_INDEX) &&
884 	     attr->pkey_index >= dev->limits.pkey_table_len) {
885 		mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
886 			  attr->pkey_index, dev->limits.pkey_table_len-1);
887 		goto out;
888 	}
889 
890 	if ((attr_mask & IB_QP_PORT) &&
891 	    (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
892 		mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
893 		goto out;
894 	}
895 
896 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
897 	    attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
898 		mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
899 			  attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
900 		goto out;
901 	}
902 
903 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
904 	    attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
905 		mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
906 			  attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
907 		goto out;
908 	}
909 
910 	if (cur_state == new_state && cur_state == IB_QPS_RESET) {
911 		err = 0;
912 		goto out;
913 	}
914 
915 	err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state,
916 				udata);
917 
918 out:
919 	mutex_unlock(&qp->mutex);
920 	return err;
921 }
922 
923 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
924 {
925 	/*
926 	 * Calculate the maximum size of WQE s/g segments, excluding
927 	 * the next segment and other non-data segments.
928 	 */
929 	int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
930 
931 	switch (qp->transport) {
932 	case MLX:
933 		max_data_size -= 2 * sizeof (struct mthca_data_seg);
934 		break;
935 
936 	case UD:
937 		if (mthca_is_memfree(dev))
938 			max_data_size -= sizeof (struct mthca_arbel_ud_seg);
939 		else
940 			max_data_size -= sizeof (struct mthca_tavor_ud_seg);
941 		break;
942 
943 	default:
944 		max_data_size -= sizeof (struct mthca_raddr_seg);
945 		break;
946 	}
947 
948 	return max_data_size;
949 }
950 
951 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
952 {
953 	/* We don't support inline data for kernel QPs (yet). */
954 	return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
955 }
956 
957 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
958 				 struct mthca_pd *pd,
959 				 struct mthca_qp *qp)
960 {
961 	int max_data_size = mthca_max_data_size(dev, qp,
962 						min(dev->limits.max_desc_sz,
963 						    1 << qp->sq.wqe_shift));
964 
965 	qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
966 
967 	qp->sq.max_gs = min_t(int, dev->limits.max_sg,
968 			      max_data_size / sizeof (struct mthca_data_seg));
969 	qp->rq.max_gs = min_t(int, dev->limits.max_sg,
970 			       (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
971 				sizeof (struct mthca_next_seg)) /
972 			       sizeof (struct mthca_data_seg));
973 }
974 
975 /*
976  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
977  * rq.max_gs and sq.max_gs must all be assigned.
978  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
979  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
980  * queue)
981  */
982 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
983 			       struct mthca_pd *pd,
984 			       struct mthca_qp *qp,
985 			       struct ib_udata *udata)
986 {
987 	int size;
988 	int err = -ENOMEM;
989 
990 	size = sizeof (struct mthca_next_seg) +
991 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
992 
993 	if (size > dev->limits.max_desc_sz)
994 		return -EINVAL;
995 
996 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
997 	     qp->rq.wqe_shift++)
998 		; /* nothing */
999 
1000 	size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
1001 	switch (qp->transport) {
1002 	case MLX:
1003 		size += 2 * sizeof (struct mthca_data_seg);
1004 		break;
1005 
1006 	case UD:
1007 		size += mthca_is_memfree(dev) ?
1008 			sizeof (struct mthca_arbel_ud_seg) :
1009 			sizeof (struct mthca_tavor_ud_seg);
1010 		break;
1011 
1012 	case UC:
1013 		size += sizeof (struct mthca_raddr_seg);
1014 		break;
1015 
1016 	case RC:
1017 		size += sizeof (struct mthca_raddr_seg);
1018 		/*
1019 		 * An atomic op will require an atomic segment, a
1020 		 * remote address segment and one scatter entry.
1021 		 */
1022 		size = max_t(int, size,
1023 			     sizeof (struct mthca_atomic_seg) +
1024 			     sizeof (struct mthca_raddr_seg) +
1025 			     sizeof (struct mthca_data_seg));
1026 		break;
1027 
1028 	default:
1029 		break;
1030 	}
1031 
1032 	/* Make sure that we have enough space for a bind request */
1033 	size = max_t(int, size, sizeof (struct mthca_bind_seg));
1034 
1035 	size += sizeof (struct mthca_next_seg);
1036 
1037 	if (size > dev->limits.max_desc_sz)
1038 		return -EINVAL;
1039 
1040 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1041 	     qp->sq.wqe_shift++)
1042 		; /* nothing */
1043 
1044 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1045 				    1 << qp->sq.wqe_shift);
1046 
1047 	/*
1048 	 * If this is a userspace QP, we don't actually have to
1049 	 * allocate anything.  All we need is to calculate the WQE
1050 	 * sizes and the send_wqe_offset, so we're done now.
1051 	 */
1052 	if (udata)
1053 		return 0;
1054 
1055 	size = PAGE_ALIGN(qp->send_wqe_offset +
1056 			  (qp->sq.max << qp->sq.wqe_shift));
1057 
1058 	qp->wrid = kmalloc_array(qp->rq.max + qp->sq.max, sizeof(u64),
1059 				 GFP_KERNEL);
1060 	if (!qp->wrid)
1061 		goto err_out;
1062 
1063 	err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1064 			      &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1065 	if (err)
1066 		goto err_out;
1067 
1068 	return 0;
1069 
1070 err_out:
1071 	kfree(qp->wrid);
1072 	return err;
1073 }
1074 
1075 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1076 			       struct mthca_qp *qp)
1077 {
1078 	mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1079 				       (qp->sq.max << qp->sq.wqe_shift)),
1080 		       &qp->queue, qp->is_direct, &qp->mr);
1081 	kfree(qp->wrid);
1082 }
1083 
1084 static int mthca_map_memfree(struct mthca_dev *dev,
1085 			     struct mthca_qp *qp)
1086 {
1087 	int ret;
1088 
1089 	if (mthca_is_memfree(dev)) {
1090 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1091 		if (ret)
1092 			return ret;
1093 
1094 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1095 		if (ret)
1096 			goto err_qpc;
1097 
1098 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1099 				      qp->qpn << dev->qp_table.rdb_shift);
1100 		if (ret)
1101 			goto err_eqpc;
1102 	}
1103 
1104 	return 0;
1105 
1106 err_eqpc:
1107 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1108 
1109 err_qpc:
1110 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1111 
1112 	return ret;
1113 }
1114 
1115 static void mthca_unmap_memfree(struct mthca_dev *dev,
1116 				struct mthca_qp *qp)
1117 {
1118 	mthca_table_put(dev, dev->qp_table.rdb_table,
1119 			qp->qpn << dev->qp_table.rdb_shift);
1120 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1121 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1122 }
1123 
1124 static int mthca_alloc_memfree(struct mthca_dev *dev,
1125 			       struct mthca_qp *qp)
1126 {
1127 	if (mthca_is_memfree(dev)) {
1128 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1129 						 qp->qpn, &qp->rq.db);
1130 		if (qp->rq.db_index < 0)
1131 			return -ENOMEM;
1132 
1133 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1134 						 qp->qpn, &qp->sq.db);
1135 		if (qp->sq.db_index < 0) {
1136 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1137 			return -ENOMEM;
1138 		}
1139 	}
1140 
1141 	return 0;
1142 }
1143 
1144 static void mthca_free_memfree(struct mthca_dev *dev,
1145 			       struct mthca_qp *qp)
1146 {
1147 	if (mthca_is_memfree(dev)) {
1148 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1149 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1150 	}
1151 }
1152 
1153 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1154 				 struct mthca_pd *pd,
1155 				 struct mthca_cq *send_cq,
1156 				 struct mthca_cq *recv_cq,
1157 				 enum ib_sig_type send_policy,
1158 				 struct mthca_qp *qp,
1159 				 struct ib_udata *udata)
1160 {
1161 	int ret;
1162 	int i;
1163 	struct mthca_next_seg *next;
1164 
1165 	qp->refcount = 1;
1166 	init_waitqueue_head(&qp->wait);
1167 	mutex_init(&qp->mutex);
1168 	qp->state    	 = IB_QPS_RESET;
1169 	qp->atomic_rd_en = 0;
1170 	qp->resp_depth   = 0;
1171 	qp->sq_policy    = send_policy;
1172 	mthca_wq_reset(&qp->sq);
1173 	mthca_wq_reset(&qp->rq);
1174 
1175 	spin_lock_init(&qp->sq.lock);
1176 	spin_lock_init(&qp->rq.lock);
1177 
1178 	ret = mthca_map_memfree(dev, qp);
1179 	if (ret)
1180 		return ret;
1181 
1182 	ret = mthca_alloc_wqe_buf(dev, pd, qp, udata);
1183 	if (ret) {
1184 		mthca_unmap_memfree(dev, qp);
1185 		return ret;
1186 	}
1187 
1188 	mthca_adjust_qp_caps(dev, pd, qp);
1189 
1190 	/*
1191 	 * If this is a userspace QP, we're done now.  The doorbells
1192 	 * will be allocated and buffers will be initialized in
1193 	 * userspace.
1194 	 */
1195 	if (udata)
1196 		return 0;
1197 
1198 	ret = mthca_alloc_memfree(dev, qp);
1199 	if (ret) {
1200 		mthca_free_wqe_buf(dev, qp);
1201 		mthca_unmap_memfree(dev, qp);
1202 		return ret;
1203 	}
1204 
1205 	if (mthca_is_memfree(dev)) {
1206 		struct mthca_data_seg *scatter;
1207 		int size = (sizeof (struct mthca_next_seg) +
1208 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1209 
1210 		for (i = 0; i < qp->rq.max; ++i) {
1211 			next = get_recv_wqe(qp, i);
1212 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1213 						   qp->rq.wqe_shift);
1214 			next->ee_nds = cpu_to_be32(size);
1215 
1216 			for (scatter = (void *) (next + 1);
1217 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1218 			     ++scatter)
1219 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1220 		}
1221 
1222 		for (i = 0; i < qp->sq.max; ++i) {
1223 			next = get_send_wqe(qp, i);
1224 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1225 						    qp->sq.wqe_shift) +
1226 						   qp->send_wqe_offset);
1227 		}
1228 	} else {
1229 		for (i = 0; i < qp->rq.max; ++i) {
1230 			next = get_recv_wqe(qp, i);
1231 			next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1232 					      qp->rq.wqe_shift) | 1);
1233 		}
1234 	}
1235 
1236 	qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1237 	qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1238 
1239 	return 0;
1240 }
1241 
1242 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1243 			     struct mthca_pd *pd, struct mthca_qp *qp)
1244 {
1245 	int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1246 
1247 	/* Sanity check QP size before proceeding */
1248 	if (cap->max_send_wr  	 > dev->limits.max_wqes ||
1249 	    cap->max_recv_wr  	 > dev->limits.max_wqes ||
1250 	    cap->max_send_sge 	 > dev->limits.max_sg   ||
1251 	    cap->max_recv_sge 	 > dev->limits.max_sg   ||
1252 	    cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1253 		return -EINVAL;
1254 
1255 	/*
1256 	 * For MLX transport we need 2 extra send gather entries:
1257 	 * one for the header and one for the checksum at the end
1258 	 */
1259 	if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1260 		return -EINVAL;
1261 
1262 	if (mthca_is_memfree(dev)) {
1263 		qp->rq.max = cap->max_recv_wr ?
1264 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1265 		qp->sq.max = cap->max_send_wr ?
1266 			roundup_pow_of_two(cap->max_send_wr) : 0;
1267 	} else {
1268 		qp->rq.max = cap->max_recv_wr;
1269 		qp->sq.max = cap->max_send_wr;
1270 	}
1271 
1272 	qp->rq.max_gs = cap->max_recv_sge;
1273 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1274 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1275 				    MTHCA_INLINE_CHUNK_SIZE) /
1276 			      sizeof (struct mthca_data_seg));
1277 
1278 	return 0;
1279 }
1280 
1281 int mthca_alloc_qp(struct mthca_dev *dev,
1282 		   struct mthca_pd *pd,
1283 		   struct mthca_cq *send_cq,
1284 		   struct mthca_cq *recv_cq,
1285 		   enum ib_qp_type type,
1286 		   enum ib_sig_type send_policy,
1287 		   struct ib_qp_cap *cap,
1288 		   struct mthca_qp *qp,
1289 		   struct ib_udata *udata)
1290 {
1291 	int err;
1292 
1293 	switch (type) {
1294 	case IB_QPT_RC: qp->transport = RC; break;
1295 	case IB_QPT_UC: qp->transport = UC; break;
1296 	case IB_QPT_UD: qp->transport = UD; break;
1297 	default: return -EINVAL;
1298 	}
1299 
1300 	err = mthca_set_qp_size(dev, cap, pd, qp);
1301 	if (err)
1302 		return err;
1303 
1304 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1305 	if (qp->qpn == -1)
1306 		return -ENOMEM;
1307 
1308 	/* initialize port to zero for error-catching. */
1309 	qp->port = 0;
1310 
1311 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1312 				    send_policy, qp, udata);
1313 	if (err) {
1314 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1315 		return err;
1316 	}
1317 
1318 	spin_lock_irq(&dev->qp_table.lock);
1319 	mthca_array_set(&dev->qp_table.qp,
1320 			qp->qpn & (dev->limits.num_qps - 1), qp);
1321 	spin_unlock_irq(&dev->qp_table.lock);
1322 
1323 	return 0;
1324 }
1325 
1326 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1327 	__acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1328 {
1329 	if (send_cq == recv_cq) {
1330 		spin_lock_irq(&send_cq->lock);
1331 		__acquire(&recv_cq->lock);
1332 	} else if (send_cq->cqn < recv_cq->cqn) {
1333 		spin_lock_irq(&send_cq->lock);
1334 		spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1335 	} else {
1336 		spin_lock_irq(&recv_cq->lock);
1337 		spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1338 	}
1339 }
1340 
1341 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1342 	__releases(&send_cq->lock) __releases(&recv_cq->lock)
1343 {
1344 	if (send_cq == recv_cq) {
1345 		__release(&recv_cq->lock);
1346 		spin_unlock_irq(&send_cq->lock);
1347 	} else if (send_cq->cqn < recv_cq->cqn) {
1348 		spin_unlock(&recv_cq->lock);
1349 		spin_unlock_irq(&send_cq->lock);
1350 	} else {
1351 		spin_unlock(&send_cq->lock);
1352 		spin_unlock_irq(&recv_cq->lock);
1353 	}
1354 }
1355 
1356 int mthca_alloc_sqp(struct mthca_dev *dev,
1357 		    struct mthca_pd *pd,
1358 		    struct mthca_cq *send_cq,
1359 		    struct mthca_cq *recv_cq,
1360 		    enum ib_sig_type send_policy,
1361 		    struct ib_qp_cap *cap,
1362 		    int qpn,
1363 		    int port,
1364 		    struct mthca_sqp *sqp,
1365 		    struct ib_udata *udata)
1366 {
1367 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1368 	int err;
1369 
1370 	sqp->qp.transport = MLX;
1371 	err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1372 	if (err)
1373 		return err;
1374 
1375 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1376 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1377 					     &sqp->header_dma, GFP_KERNEL);
1378 	if (!sqp->header_buf)
1379 		return -ENOMEM;
1380 
1381 	spin_lock_irq(&dev->qp_table.lock);
1382 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1383 		err = -EBUSY;
1384 	else
1385 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1386 	spin_unlock_irq(&dev->qp_table.lock);
1387 
1388 	if (err)
1389 		goto err_out;
1390 
1391 	sqp->qp.port      = port;
1392 	sqp->qp.qpn       = mqpn;
1393 	sqp->qp.transport = MLX;
1394 
1395 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1396 				    send_policy, &sqp->qp, udata);
1397 	if (err)
1398 		goto err_out_free;
1399 
1400 	atomic_inc(&pd->sqp_count);
1401 
1402 	return 0;
1403 
1404  err_out_free:
1405 	/*
1406 	 * Lock CQs here, so that CQ polling code can do QP lookup
1407 	 * without taking a lock.
1408 	 */
1409 	mthca_lock_cqs(send_cq, recv_cq);
1410 
1411 	spin_lock(&dev->qp_table.lock);
1412 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1413 	spin_unlock(&dev->qp_table.lock);
1414 
1415 	mthca_unlock_cqs(send_cq, recv_cq);
1416 
1417  err_out:
1418 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1419 			  sqp->header_buf, sqp->header_dma);
1420 
1421 	return err;
1422 }
1423 
1424 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1425 {
1426 	int c;
1427 
1428 	spin_lock_irq(&dev->qp_table.lock);
1429 	c = qp->refcount;
1430 	spin_unlock_irq(&dev->qp_table.lock);
1431 
1432 	return c;
1433 }
1434 
1435 void mthca_free_qp(struct mthca_dev *dev,
1436 		   struct mthca_qp *qp)
1437 {
1438 	struct mthca_cq *send_cq;
1439 	struct mthca_cq *recv_cq;
1440 
1441 	send_cq = to_mcq(qp->ibqp.send_cq);
1442 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1443 
1444 	/*
1445 	 * Lock CQs here, so that CQ polling code can do QP lookup
1446 	 * without taking a lock.
1447 	 */
1448 	mthca_lock_cqs(send_cq, recv_cq);
1449 
1450 	spin_lock(&dev->qp_table.lock);
1451 	mthca_array_clear(&dev->qp_table.qp,
1452 			  qp->qpn & (dev->limits.num_qps - 1));
1453 	--qp->refcount;
1454 	spin_unlock(&dev->qp_table.lock);
1455 
1456 	mthca_unlock_cqs(send_cq, recv_cq);
1457 
1458 	wait_event(qp->wait, !get_qp_refcount(dev, qp));
1459 
1460 	if (qp->state != IB_QPS_RESET)
1461 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1462 				NULL, 0);
1463 
1464 	/*
1465 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1466 	 * will be cleaned up in userspace, so all we have to do is
1467 	 * unref the mem-free tables and free the QPN in our table.
1468 	 */
1469 	if (!qp->ibqp.uobject) {
1470 		mthca_cq_clean(dev, recv_cq, qp->qpn,
1471 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1472 		if (send_cq != recv_cq)
1473 			mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1474 
1475 		mthca_free_memfree(dev, qp);
1476 		mthca_free_wqe_buf(dev, qp);
1477 	}
1478 
1479 	mthca_unmap_memfree(dev, qp);
1480 
1481 	if (is_sqp(dev, qp)) {
1482 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1483 		dma_free_coherent(&dev->pdev->dev,
1484 				  to_msqp(qp)->header_buf_size,
1485 				  to_msqp(qp)->header_buf,
1486 				  to_msqp(qp)->header_dma);
1487 	} else
1488 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1489 }
1490 
1491 /* Create UD header for an MLX send and build a data segment for it */
1492 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1493 			    int ind, const struct ib_ud_wr *wr,
1494 			    struct mthca_mlx_seg *mlx,
1495 			    struct mthca_data_seg *data)
1496 {
1497 	int header_size;
1498 	int err;
1499 	u16 pkey;
1500 
1501 	ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
1502 			  mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0,
1503 			  &sqp->ud_header);
1504 
1505 	err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header);
1506 	if (err)
1507 		return err;
1508 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1509 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1510 				  (sqp->ud_header.lrh.destination_lid ==
1511 				   IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1512 				  (sqp->ud_header.lrh.service_level << 8));
1513 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1514 	mlx->vcrc = 0;
1515 
1516 	switch (wr->wr.opcode) {
1517 	case IB_WR_SEND:
1518 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1519 		sqp->ud_header.immediate_present = 0;
1520 		break;
1521 	case IB_WR_SEND_WITH_IMM:
1522 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1523 		sqp->ud_header.immediate_present = 1;
1524 		sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
1525 		break;
1526 	default:
1527 		return -EINVAL;
1528 	}
1529 
1530 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1531 	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1532 		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1533 	sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
1534 	if (!sqp->qp.ibqp.qp_num)
1535 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1536 				   sqp->pkey_index, &pkey);
1537 	else
1538 		ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1539 				   wr->pkey_index, &pkey);
1540 	sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1541 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
1542 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1543 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
1544 					       sqp->qkey : wr->remote_qkey);
1545 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1546 
1547 	header_size = ib_ud_header_pack(&sqp->ud_header,
1548 					sqp->header_buf +
1549 					ind * MTHCA_UD_HEADER_SIZE);
1550 
1551 	data->byte_count = cpu_to_be32(header_size);
1552 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1553 	data->addr       = cpu_to_be64(sqp->header_dma +
1554 				       ind * MTHCA_UD_HEADER_SIZE);
1555 
1556 	return 0;
1557 }
1558 
1559 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1560 				    struct ib_cq *ib_cq)
1561 {
1562 	unsigned cur;
1563 	struct mthca_cq *cq;
1564 
1565 	cur = wq->head - wq->tail;
1566 	if (likely(cur + nreq < wq->max))
1567 		return 0;
1568 
1569 	cq = to_mcq(ib_cq);
1570 	spin_lock(&cq->lock);
1571 	cur = wq->head - wq->tail;
1572 	spin_unlock(&cq->lock);
1573 
1574 	return cur + nreq >= wq->max;
1575 }
1576 
1577 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1578 					  u64 remote_addr, u32 rkey)
1579 {
1580 	rseg->raddr    = cpu_to_be64(remote_addr);
1581 	rseg->rkey     = cpu_to_be32(rkey);
1582 	rseg->reserved = 0;
1583 }
1584 
1585 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1586 					   const struct ib_atomic_wr *wr)
1587 {
1588 	if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1589 		aseg->swap_add = cpu_to_be64(wr->swap);
1590 		aseg->compare  = cpu_to_be64(wr->compare_add);
1591 	} else {
1592 		aseg->swap_add = cpu_to_be64(wr->compare_add);
1593 		aseg->compare  = 0;
1594 	}
1595 
1596 }
1597 
1598 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1599 			     const struct ib_ud_wr *wr)
1600 {
1601 	useg->lkey    = cpu_to_be32(to_mah(wr->ah)->key);
1602 	useg->av_addr =	cpu_to_be64(to_mah(wr->ah)->avdma);
1603 	useg->dqpn    =	cpu_to_be32(wr->remote_qpn);
1604 	useg->qkey    =	cpu_to_be32(wr->remote_qkey);
1605 
1606 }
1607 
1608 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1609 			     const struct ib_ud_wr *wr)
1610 {
1611 	memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE);
1612 	useg->dqpn = cpu_to_be32(wr->remote_qpn);
1613 	useg->qkey = cpu_to_be32(wr->remote_qkey);
1614 }
1615 
1616 int mthca_tavor_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1617 			  const struct ib_send_wr **bad_wr)
1618 {
1619 	struct mthca_dev *dev = to_mdev(ibqp->device);
1620 	struct mthca_qp *qp = to_mqp(ibqp);
1621 	void *wqe;
1622 	void *prev_wqe;
1623 	unsigned long flags;
1624 	int err = 0;
1625 	int nreq;
1626 	int i;
1627 	int size;
1628 	/*
1629 	 * f0 and size0 are only used if nreq != 0, and they will
1630 	 * always be initialized the first time through the main loop
1631 	 * before nreq is incremented.  So nreq cannot become non-zero
1632 	 * without initializing f0 and size0, and they are in fact
1633 	 * never used uninitialized.
1634 	 */
1635 	int uninitialized_var(size0);
1636 	u32 uninitialized_var(f0);
1637 	int ind;
1638 	u8 op0 = 0;
1639 
1640 	spin_lock_irqsave(&qp->sq.lock, flags);
1641 
1642 	/* XXX check that state is OK to post send */
1643 
1644 	ind = qp->sq.next_ind;
1645 
1646 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1647 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1648 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1649 					" %d max, %d nreq)\n", qp->qpn,
1650 					qp->sq.head, qp->sq.tail,
1651 					qp->sq.max, nreq);
1652 			err = -ENOMEM;
1653 			*bad_wr = wr;
1654 			goto out;
1655 		}
1656 
1657 		wqe = get_send_wqe(qp, ind);
1658 		prev_wqe = qp->sq.last;
1659 		qp->sq.last = wqe;
1660 
1661 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1662 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1663 		((struct mthca_next_seg *) wqe)->flags =
1664 			((wr->send_flags & IB_SEND_SIGNALED) ?
1665 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1666 			((wr->send_flags & IB_SEND_SOLICITED) ?
1667 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1668 			cpu_to_be32(1);
1669 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1670 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1671 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1672 
1673 		wqe += sizeof (struct mthca_next_seg);
1674 		size = sizeof (struct mthca_next_seg) / 16;
1675 
1676 		switch (qp->transport) {
1677 		case RC:
1678 			switch (wr->opcode) {
1679 			case IB_WR_ATOMIC_CMP_AND_SWP:
1680 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1681 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
1682 					      atomic_wr(wr)->rkey);
1683 				wqe += sizeof (struct mthca_raddr_seg);
1684 
1685 				set_atomic_seg(wqe, atomic_wr(wr));
1686 				wqe += sizeof (struct mthca_atomic_seg);
1687 				size += (sizeof (struct mthca_raddr_seg) +
1688 					 sizeof (struct mthca_atomic_seg)) / 16;
1689 				break;
1690 
1691 			case IB_WR_RDMA_WRITE:
1692 			case IB_WR_RDMA_WRITE_WITH_IMM:
1693 			case IB_WR_RDMA_READ:
1694 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1695 					      rdma_wr(wr)->rkey);
1696 				wqe  += sizeof (struct mthca_raddr_seg);
1697 				size += sizeof (struct mthca_raddr_seg) / 16;
1698 				break;
1699 
1700 			default:
1701 				/* No extra segments required for sends */
1702 				break;
1703 			}
1704 
1705 			break;
1706 
1707 		case UC:
1708 			switch (wr->opcode) {
1709 			case IB_WR_RDMA_WRITE:
1710 			case IB_WR_RDMA_WRITE_WITH_IMM:
1711 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1712 					      rdma_wr(wr)->rkey);
1713 				wqe  += sizeof (struct mthca_raddr_seg);
1714 				size += sizeof (struct mthca_raddr_seg) / 16;
1715 				break;
1716 
1717 			default:
1718 				/* No extra segments required for sends */
1719 				break;
1720 			}
1721 
1722 			break;
1723 
1724 		case UD:
1725 			set_tavor_ud_seg(wqe, ud_wr(wr));
1726 			wqe  += sizeof (struct mthca_tavor_ud_seg);
1727 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1728 			break;
1729 
1730 		case MLX:
1731 			err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
1732 					       wqe - sizeof (struct mthca_next_seg),
1733 					       wqe);
1734 			if (err) {
1735 				*bad_wr = wr;
1736 				goto out;
1737 			}
1738 			wqe += sizeof (struct mthca_data_seg);
1739 			size += sizeof (struct mthca_data_seg) / 16;
1740 			break;
1741 		}
1742 
1743 		if (wr->num_sge > qp->sq.max_gs) {
1744 			mthca_err(dev, "too many gathers\n");
1745 			err = -EINVAL;
1746 			*bad_wr = wr;
1747 			goto out;
1748 		}
1749 
1750 		for (i = 0; i < wr->num_sge; ++i) {
1751 			mthca_set_data_seg(wqe, wr->sg_list + i);
1752 			wqe  += sizeof (struct mthca_data_seg);
1753 			size += sizeof (struct mthca_data_seg) / 16;
1754 		}
1755 
1756 		/* Add one more inline data segment for ICRC */
1757 		if (qp->transport == MLX) {
1758 			((struct mthca_data_seg *) wqe)->byte_count =
1759 				cpu_to_be32((1 << 31) | 4);
1760 			((u32 *) wqe)[1] = 0;
1761 			wqe += sizeof (struct mthca_data_seg);
1762 			size += sizeof (struct mthca_data_seg) / 16;
1763 		}
1764 
1765 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1766 
1767 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1768 			mthca_err(dev, "opcode invalid\n");
1769 			err = -EINVAL;
1770 			*bad_wr = wr;
1771 			goto out;
1772 		}
1773 
1774 		((struct mthca_next_seg *) prev_wqe)->nda_op =
1775 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
1776 				     qp->send_wqe_offset) |
1777 				    mthca_opcode[wr->opcode]);
1778 		wmb();
1779 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1780 			cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1781 				    ((wr->send_flags & IB_SEND_FENCE) ?
1782 				    MTHCA_NEXT_FENCE : 0));
1783 
1784 		if (!nreq) {
1785 			size0 = size;
1786 			op0   = mthca_opcode[wr->opcode];
1787 			f0    = wr->send_flags & IB_SEND_FENCE ?
1788 				MTHCA_SEND_DOORBELL_FENCE : 0;
1789 		}
1790 
1791 		++ind;
1792 		if (unlikely(ind >= qp->sq.max))
1793 			ind -= qp->sq.max;
1794 	}
1795 
1796 out:
1797 	if (likely(nreq)) {
1798 		wmb();
1799 
1800 		mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1801 			       qp->send_wqe_offset) | f0 | op0,
1802 			      (qp->qpn << 8) | size0,
1803 			      dev->kar + MTHCA_SEND_DOORBELL,
1804 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1805 		/*
1806 		 * Make sure doorbells don't leak out of SQ spinlock
1807 		 * and reach the HCA out of order:
1808 		 */
1809 		mmiowb();
1810 	}
1811 
1812 	qp->sq.next_ind = ind;
1813 	qp->sq.head    += nreq;
1814 
1815 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1816 	return err;
1817 }
1818 
1819 int mthca_tavor_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1820 			     const struct ib_recv_wr **bad_wr)
1821 {
1822 	struct mthca_dev *dev = to_mdev(ibqp->device);
1823 	struct mthca_qp *qp = to_mqp(ibqp);
1824 	unsigned long flags;
1825 	int err = 0;
1826 	int nreq;
1827 	int i;
1828 	int size;
1829 	/*
1830 	 * size0 is only used if nreq != 0, and it will always be
1831 	 * initialized the first time through the main loop before
1832 	 * nreq is incremented.  So nreq cannot become non-zero
1833 	 * without initializing size0, and it is in fact never used
1834 	 * uninitialized.
1835 	 */
1836 	int uninitialized_var(size0);
1837 	int ind;
1838 	void *wqe;
1839 	void *prev_wqe;
1840 
1841 	spin_lock_irqsave(&qp->rq.lock, flags);
1842 
1843 	/* XXX check that state is OK to post receive */
1844 
1845 	ind = qp->rq.next_ind;
1846 
1847 	for (nreq = 0; wr; wr = wr->next) {
1848 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1849 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1850 					" %d max, %d nreq)\n", qp->qpn,
1851 					qp->rq.head, qp->rq.tail,
1852 					qp->rq.max, nreq);
1853 			err = -ENOMEM;
1854 			*bad_wr = wr;
1855 			goto out;
1856 		}
1857 
1858 		wqe = get_recv_wqe(qp, ind);
1859 		prev_wqe = qp->rq.last;
1860 		qp->rq.last = wqe;
1861 
1862 		((struct mthca_next_seg *) wqe)->ee_nds =
1863 			cpu_to_be32(MTHCA_NEXT_DBD);
1864 		((struct mthca_next_seg *) wqe)->flags = 0;
1865 
1866 		wqe += sizeof (struct mthca_next_seg);
1867 		size = sizeof (struct mthca_next_seg) / 16;
1868 
1869 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1870 			err = -EINVAL;
1871 			*bad_wr = wr;
1872 			goto out;
1873 		}
1874 
1875 		for (i = 0; i < wr->num_sge; ++i) {
1876 			mthca_set_data_seg(wqe, wr->sg_list + i);
1877 			wqe  += sizeof (struct mthca_data_seg);
1878 			size += sizeof (struct mthca_data_seg) / 16;
1879 		}
1880 
1881 		qp->wrid[ind] = wr->wr_id;
1882 
1883 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
1884 			cpu_to_be32(MTHCA_NEXT_DBD | size);
1885 
1886 		if (!nreq)
1887 			size0 = size;
1888 
1889 		++ind;
1890 		if (unlikely(ind >= qp->rq.max))
1891 			ind -= qp->rq.max;
1892 
1893 		++nreq;
1894 		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1895 			nreq = 0;
1896 
1897 			wmb();
1898 
1899 			mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1900 				      qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1901 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1902 
1903 			qp->rq.next_ind = ind;
1904 			qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1905 		}
1906 	}
1907 
1908 out:
1909 	if (likely(nreq)) {
1910 		wmb();
1911 
1912 		mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1913 			      qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1914 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1915 	}
1916 
1917 	qp->rq.next_ind = ind;
1918 	qp->rq.head    += nreq;
1919 
1920 	/*
1921 	 * Make sure doorbells don't leak out of RQ spinlock and reach
1922 	 * the HCA out of order:
1923 	 */
1924 	mmiowb();
1925 
1926 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1927 	return err;
1928 }
1929 
1930 int mthca_arbel_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1931 			  const struct ib_send_wr **bad_wr)
1932 {
1933 	struct mthca_dev *dev = to_mdev(ibqp->device);
1934 	struct mthca_qp *qp = to_mqp(ibqp);
1935 	u32 dbhi;
1936 	void *wqe;
1937 	void *prev_wqe;
1938 	unsigned long flags;
1939 	int err = 0;
1940 	int nreq;
1941 	int i;
1942 	int size;
1943 	/*
1944 	 * f0 and size0 are only used if nreq != 0, and they will
1945 	 * always be initialized the first time through the main loop
1946 	 * before nreq is incremented.  So nreq cannot become non-zero
1947 	 * without initializing f0 and size0, and they are in fact
1948 	 * never used uninitialized.
1949 	 */
1950 	int uninitialized_var(size0);
1951 	u32 uninitialized_var(f0);
1952 	int ind;
1953 	u8 op0 = 0;
1954 
1955 	spin_lock_irqsave(&qp->sq.lock, flags);
1956 
1957 	/* XXX check that state is OK to post send */
1958 
1959 	ind = qp->sq.head & (qp->sq.max - 1);
1960 
1961 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1962 		if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1963 			nreq = 0;
1964 
1965 			dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1966 				((qp->sq.head & 0xffff) << 8) | f0 | op0;
1967 
1968 			qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1969 
1970 			/*
1971 			 * Make sure that descriptors are written before
1972 			 * doorbell record.
1973 			 */
1974 			wmb();
1975 			*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1976 
1977 			/*
1978 			 * Make sure doorbell record is written before we
1979 			 * write MMIO send doorbell.
1980 			 */
1981 			wmb();
1982 
1983 			mthca_write64(dbhi, (qp->qpn << 8) | size0,
1984 				      dev->kar + MTHCA_SEND_DOORBELL,
1985 				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1986 		}
1987 
1988 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1989 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1990 					" %d max, %d nreq)\n", qp->qpn,
1991 					qp->sq.head, qp->sq.tail,
1992 					qp->sq.max, nreq);
1993 			err = -ENOMEM;
1994 			*bad_wr = wr;
1995 			goto out;
1996 		}
1997 
1998 		wqe = get_send_wqe(qp, ind);
1999 		prev_wqe = qp->sq.last;
2000 		qp->sq.last = wqe;
2001 
2002 		((struct mthca_next_seg *) wqe)->flags =
2003 			((wr->send_flags & IB_SEND_SIGNALED) ?
2004 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
2005 			((wr->send_flags & IB_SEND_SOLICITED) ?
2006 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
2007 			((wr->send_flags & IB_SEND_IP_CSUM) ?
2008 			 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2009 			cpu_to_be32(1);
2010 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2011 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2012 			((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2013 
2014 		wqe += sizeof (struct mthca_next_seg);
2015 		size = sizeof (struct mthca_next_seg) / 16;
2016 
2017 		switch (qp->transport) {
2018 		case RC:
2019 			switch (wr->opcode) {
2020 			case IB_WR_ATOMIC_CMP_AND_SWP:
2021 			case IB_WR_ATOMIC_FETCH_AND_ADD:
2022 				set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2023 					      atomic_wr(wr)->rkey);
2024 				wqe += sizeof (struct mthca_raddr_seg);
2025 
2026 				set_atomic_seg(wqe, atomic_wr(wr));
2027 				wqe  += sizeof (struct mthca_atomic_seg);
2028 				size += (sizeof (struct mthca_raddr_seg) +
2029 					 sizeof (struct mthca_atomic_seg)) / 16;
2030 				break;
2031 
2032 			case IB_WR_RDMA_READ:
2033 			case IB_WR_RDMA_WRITE:
2034 			case IB_WR_RDMA_WRITE_WITH_IMM:
2035 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2036 					      rdma_wr(wr)->rkey);
2037 				wqe  += sizeof (struct mthca_raddr_seg);
2038 				size += sizeof (struct mthca_raddr_seg) / 16;
2039 				break;
2040 
2041 			default:
2042 				/* No extra segments required for sends */
2043 				break;
2044 			}
2045 
2046 			break;
2047 
2048 		case UC:
2049 			switch (wr->opcode) {
2050 			case IB_WR_RDMA_WRITE:
2051 			case IB_WR_RDMA_WRITE_WITH_IMM:
2052 				set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2053 					      rdma_wr(wr)->rkey);
2054 				wqe  += sizeof (struct mthca_raddr_seg);
2055 				size += sizeof (struct mthca_raddr_seg) / 16;
2056 				break;
2057 
2058 			default:
2059 				/* No extra segments required for sends */
2060 				break;
2061 			}
2062 
2063 			break;
2064 
2065 		case UD:
2066 			set_arbel_ud_seg(wqe, ud_wr(wr));
2067 			wqe  += sizeof (struct mthca_arbel_ud_seg);
2068 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
2069 			break;
2070 
2071 		case MLX:
2072 			err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
2073 					       wqe - sizeof (struct mthca_next_seg),
2074 					       wqe);
2075 			if (err) {
2076 				*bad_wr = wr;
2077 				goto out;
2078 			}
2079 			wqe += sizeof (struct mthca_data_seg);
2080 			size += sizeof (struct mthca_data_seg) / 16;
2081 			break;
2082 		}
2083 
2084 		if (wr->num_sge > qp->sq.max_gs) {
2085 			mthca_err(dev, "too many gathers\n");
2086 			err = -EINVAL;
2087 			*bad_wr = wr;
2088 			goto out;
2089 		}
2090 
2091 		for (i = 0; i < wr->num_sge; ++i) {
2092 			mthca_set_data_seg(wqe, wr->sg_list + i);
2093 			wqe  += sizeof (struct mthca_data_seg);
2094 			size += sizeof (struct mthca_data_seg) / 16;
2095 		}
2096 
2097 		/* Add one more inline data segment for ICRC */
2098 		if (qp->transport == MLX) {
2099 			((struct mthca_data_seg *) wqe)->byte_count =
2100 				cpu_to_be32((1 << 31) | 4);
2101 			((u32 *) wqe)[1] = 0;
2102 			wqe += sizeof (struct mthca_data_seg);
2103 			size += sizeof (struct mthca_data_seg) / 16;
2104 		}
2105 
2106 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2107 
2108 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2109 			mthca_err(dev, "opcode invalid\n");
2110 			err = -EINVAL;
2111 			*bad_wr = wr;
2112 			goto out;
2113 		}
2114 
2115 		((struct mthca_next_seg *) prev_wqe)->nda_op =
2116 			cpu_to_be32(((ind << qp->sq.wqe_shift) +
2117 				     qp->send_wqe_offset) |
2118 				    mthca_opcode[wr->opcode]);
2119 		wmb();
2120 		((struct mthca_next_seg *) prev_wqe)->ee_nds =
2121 			cpu_to_be32(MTHCA_NEXT_DBD | size |
2122 				    ((wr->send_flags & IB_SEND_FENCE) ?
2123 				     MTHCA_NEXT_FENCE : 0));
2124 
2125 		if (!nreq) {
2126 			size0 = size;
2127 			op0   = mthca_opcode[wr->opcode];
2128 			f0    = wr->send_flags & IB_SEND_FENCE ?
2129 				MTHCA_SEND_DOORBELL_FENCE : 0;
2130 		}
2131 
2132 		++ind;
2133 		if (unlikely(ind >= qp->sq.max))
2134 			ind -= qp->sq.max;
2135 	}
2136 
2137 out:
2138 	if (likely(nreq)) {
2139 		dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2140 
2141 		qp->sq.head += nreq;
2142 
2143 		/*
2144 		 * Make sure that descriptors are written before
2145 		 * doorbell record.
2146 		 */
2147 		wmb();
2148 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2149 
2150 		/*
2151 		 * Make sure doorbell record is written before we
2152 		 * write MMIO send doorbell.
2153 		 */
2154 		wmb();
2155 
2156 		mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2157 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2158 	}
2159 
2160 	/*
2161 	 * Make sure doorbells don't leak out of SQ spinlock and reach
2162 	 * the HCA out of order:
2163 	 */
2164 	mmiowb();
2165 
2166 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2167 	return err;
2168 }
2169 
2170 int mthca_arbel_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2171 			     const struct ib_recv_wr **bad_wr)
2172 {
2173 	struct mthca_dev *dev = to_mdev(ibqp->device);
2174 	struct mthca_qp *qp = to_mqp(ibqp);
2175 	unsigned long flags;
2176 	int err = 0;
2177 	int nreq;
2178 	int ind;
2179 	int i;
2180 	void *wqe;
2181 
2182 	spin_lock_irqsave(&qp->rq.lock, flags);
2183 
2184 	/* XXX check that state is OK to post receive */
2185 
2186 	ind = qp->rq.head & (qp->rq.max - 1);
2187 
2188 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2189 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2190 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2191 					" %d max, %d nreq)\n", qp->qpn,
2192 					qp->rq.head, qp->rq.tail,
2193 					qp->rq.max, nreq);
2194 			err = -ENOMEM;
2195 			*bad_wr = wr;
2196 			goto out;
2197 		}
2198 
2199 		wqe = get_recv_wqe(qp, ind);
2200 
2201 		((struct mthca_next_seg *) wqe)->flags = 0;
2202 
2203 		wqe += sizeof (struct mthca_next_seg);
2204 
2205 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2206 			err = -EINVAL;
2207 			*bad_wr = wr;
2208 			goto out;
2209 		}
2210 
2211 		for (i = 0; i < wr->num_sge; ++i) {
2212 			mthca_set_data_seg(wqe, wr->sg_list + i);
2213 			wqe += sizeof (struct mthca_data_seg);
2214 		}
2215 
2216 		if (i < qp->rq.max_gs)
2217 			mthca_set_data_seg_inval(wqe);
2218 
2219 		qp->wrid[ind] = wr->wr_id;
2220 
2221 		++ind;
2222 		if (unlikely(ind >= qp->rq.max))
2223 			ind -= qp->rq.max;
2224 	}
2225 out:
2226 	if (likely(nreq)) {
2227 		qp->rq.head += nreq;
2228 
2229 		/*
2230 		 * Make sure that descriptors are written before
2231 		 * doorbell record.
2232 		 */
2233 		wmb();
2234 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2235 	}
2236 
2237 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2238 	return err;
2239 }
2240 
2241 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2242 			int index, int *dbd, __be32 *new_wqe)
2243 {
2244 	struct mthca_next_seg *next;
2245 
2246 	/*
2247 	 * For SRQs, all receive WQEs generate a CQE, so we're always
2248 	 * at the end of the doorbell chain.
2249 	 */
2250 	if (qp->ibqp.srq && !is_send) {
2251 		*new_wqe = 0;
2252 		return;
2253 	}
2254 
2255 	if (is_send)
2256 		next = get_send_wqe(qp, index);
2257 	else
2258 		next = get_recv_wqe(qp, index);
2259 
2260 	*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2261 	if (next->ee_nds & cpu_to_be32(0x3f))
2262 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2263 			(next->ee_nds & cpu_to_be32(0x3f));
2264 	else
2265 		*new_wqe = 0;
2266 }
2267 
2268 int mthca_init_qp_table(struct mthca_dev *dev)
2269 {
2270 	int err;
2271 	int i;
2272 
2273 	spin_lock_init(&dev->qp_table.lock);
2274 
2275 	/*
2276 	 * We reserve 2 extra QPs per port for the special QPs.  The
2277 	 * special QP for port 1 has to be even, so round up.
2278 	 */
2279 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2280 	err = mthca_alloc_init(&dev->qp_table.alloc,
2281 			       dev->limits.num_qps,
2282 			       (1 << 24) - 1,
2283 			       dev->qp_table.sqp_start +
2284 			       MTHCA_MAX_PORTS * 2);
2285 	if (err)
2286 		return err;
2287 
2288 	err = mthca_array_init(&dev->qp_table.qp,
2289 			       dev->limits.num_qps);
2290 	if (err) {
2291 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2292 		return err;
2293 	}
2294 
2295 	for (i = 0; i < 2; ++i) {
2296 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2297 				    dev->qp_table.sqp_start + i * 2);
2298 		if (err) {
2299 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2300 				   "%d, aborting.\n", err);
2301 			goto err_out;
2302 		}
2303 	}
2304 	return 0;
2305 
2306  err_out:
2307 	for (i = 0; i < 2; ++i)
2308 		mthca_CONF_SPECIAL_QP(dev, i, 0);
2309 
2310 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2311 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2312 
2313 	return err;
2314 }
2315 
2316 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2317 {
2318 	int i;
2319 
2320 	for (i = 0; i < 2; ++i)
2321 		mthca_CONF_SPECIAL_QP(dev, i, 0);
2322 
2323 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2324 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2325 }
2326