xref: /linux/drivers/infiniband/hw/mthca/mthca_qp.c (revision 20d0021394c1b070bf04b22c5bc8fdb437edd4c5)
1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  *
33  * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
34  */
35 
36 #include <linux/init.h>
37 
38 #include <ib_verbs.h>
39 #include <ib_cache.h>
40 #include <ib_pack.h>
41 
42 #include "mthca_dev.h"
43 #include "mthca_cmd.h"
44 #include "mthca_memfree.h"
45 
46 enum {
47 	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
48 	MTHCA_ACK_REQ_FREQ       = 10,
49 	MTHCA_FLIGHT_LIMIT       = 9,
50 	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
51 	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
52 	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
53 };
54 
55 enum {
56 	MTHCA_QP_STATE_RST  = 0,
57 	MTHCA_QP_STATE_INIT = 1,
58 	MTHCA_QP_STATE_RTR  = 2,
59 	MTHCA_QP_STATE_RTS  = 3,
60 	MTHCA_QP_STATE_SQE  = 4,
61 	MTHCA_QP_STATE_SQD  = 5,
62 	MTHCA_QP_STATE_ERR  = 6,
63 	MTHCA_QP_STATE_DRAINING = 7
64 };
65 
66 enum {
67 	MTHCA_QP_ST_RC 	= 0x0,
68 	MTHCA_QP_ST_UC 	= 0x1,
69 	MTHCA_QP_ST_RD 	= 0x2,
70 	MTHCA_QP_ST_UD 	= 0x3,
71 	MTHCA_QP_ST_MLX = 0x7
72 };
73 
74 enum {
75 	MTHCA_QP_PM_MIGRATED = 0x3,
76 	MTHCA_QP_PM_ARMED    = 0x0,
77 	MTHCA_QP_PM_REARM    = 0x1
78 };
79 
80 enum {
81 	/* qp_context flags */
82 	MTHCA_QP_BIT_DE  = 1 <<  8,
83 	/* params1 */
84 	MTHCA_QP_BIT_SRE = 1 << 15,
85 	MTHCA_QP_BIT_SWE = 1 << 14,
86 	MTHCA_QP_BIT_SAE = 1 << 13,
87 	MTHCA_QP_BIT_SIC = 1 <<  4,
88 	MTHCA_QP_BIT_SSC = 1 <<  3,
89 	/* params2 */
90 	MTHCA_QP_BIT_RRE = 1 << 15,
91 	MTHCA_QP_BIT_RWE = 1 << 14,
92 	MTHCA_QP_BIT_RAE = 1 << 13,
93 	MTHCA_QP_BIT_RIC = 1 <<  4,
94 	MTHCA_QP_BIT_RSC = 1 <<  3
95 };
96 
97 struct mthca_qp_path {
98 	u32 port_pkey;
99 	u8  rnr_retry;
100 	u8  g_mylmc;
101 	u16 rlid;
102 	u8  ackto;
103 	u8  mgid_index;
104 	u8  static_rate;
105 	u8  hop_limit;
106 	u32 sl_tclass_flowlabel;
107 	u8  rgid[16];
108 } __attribute__((packed));
109 
110 struct mthca_qp_context {
111 	u32 flags;
112 	u32 tavor_sched_queue;	/* Reserved on Arbel */
113 	u8  mtu_msgmax;
114 	u8  rq_size_stride;	/* Reserved on Tavor */
115 	u8  sq_size_stride;	/* Reserved on Tavor */
116 	u8  rlkey_arbel_sched_queue;	/* Reserved on Tavor */
117 	u32 usr_page;
118 	u32 local_qpn;
119 	u32 remote_qpn;
120 	u32 reserved1[2];
121 	struct mthca_qp_path pri_path;
122 	struct mthca_qp_path alt_path;
123 	u32 rdd;
124 	u32 pd;
125 	u32 wqe_base;
126 	u32 wqe_lkey;
127 	u32 params1;
128 	u32 reserved2;
129 	u32 next_send_psn;
130 	u32 cqn_snd;
131 	u32 snd_wqe_base_l;	/* Next send WQE on Tavor */
132 	u32 snd_db_index;	/* (debugging only entries) */
133 	u32 last_acked_psn;
134 	u32 ssn;
135 	u32 params2;
136 	u32 rnr_nextrecvpsn;
137 	u32 ra_buff_indx;
138 	u32 cqn_rcv;
139 	u32 rcv_wqe_base_l;	/* Next recv WQE on Tavor */
140 	u32 rcv_db_index;	/* (debugging only entries) */
141 	u32 qkey;
142 	u32 srqn;
143 	u32 rmsn;
144 	u16 rq_wqe_counter;	/* reserved on Tavor */
145 	u16 sq_wqe_counter;	/* reserved on Tavor */
146 	u32 reserved3[18];
147 } __attribute__((packed));
148 
149 struct mthca_qp_param {
150 	u32 opt_param_mask;
151 	u32 reserved1;
152 	struct mthca_qp_context context;
153 	u32 reserved2[62];
154 } __attribute__((packed));
155 
156 enum {
157 	MTHCA_QP_OPTPAR_ALT_ADDR_PATH     = 1 << 0,
158 	MTHCA_QP_OPTPAR_RRE               = 1 << 1,
159 	MTHCA_QP_OPTPAR_RAE               = 1 << 2,
160 	MTHCA_QP_OPTPAR_RWE               = 1 << 3,
161 	MTHCA_QP_OPTPAR_PKEY_INDEX        = 1 << 4,
162 	MTHCA_QP_OPTPAR_Q_KEY             = 1 << 5,
163 	MTHCA_QP_OPTPAR_RNR_TIMEOUT       = 1 << 6,
164 	MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
165 	MTHCA_QP_OPTPAR_SRA_MAX           = 1 << 8,
166 	MTHCA_QP_OPTPAR_RRA_MAX           = 1 << 9,
167 	MTHCA_QP_OPTPAR_PM_STATE          = 1 << 10,
168 	MTHCA_QP_OPTPAR_PORT_NUM          = 1 << 11,
169 	MTHCA_QP_OPTPAR_RETRY_COUNT       = 1 << 12,
170 	MTHCA_QP_OPTPAR_ALT_RNR_RETRY     = 1 << 13,
171 	MTHCA_QP_OPTPAR_ACK_TIMEOUT       = 1 << 14,
172 	MTHCA_QP_OPTPAR_RNR_RETRY         = 1 << 15,
173 	MTHCA_QP_OPTPAR_SCHED_QUEUE       = 1 << 16
174 };
175 
176 enum {
177 	MTHCA_NEXT_DBD       = 1 << 7,
178 	MTHCA_NEXT_FENCE     = 1 << 6,
179 	MTHCA_NEXT_CQ_UPDATE = 1 << 3,
180 	MTHCA_NEXT_EVENT_GEN = 1 << 2,
181 	MTHCA_NEXT_SOLICIT   = 1 << 1,
182 
183 	MTHCA_MLX_VL15       = 1 << 17,
184 	MTHCA_MLX_SLR        = 1 << 16
185 };
186 
187 enum {
188 	MTHCA_INVAL_LKEY = 0x100
189 };
190 
191 struct mthca_next_seg {
192 	u32 nda_op;		/* [31:6] next WQE [4:0] next opcode */
193 	u32 ee_nds;		/* [31:8] next EE  [7] DBD [6] F [5:0] next WQE size */
194 	u32 flags;		/* [3] CQ [2] Event [1] Solicit */
195 	u32 imm;		/* immediate data */
196 };
197 
198 struct mthca_tavor_ud_seg {
199 	u32 reserved1;
200 	u32 lkey;
201 	u64 av_addr;
202 	u32 reserved2[4];
203 	u32 dqpn;
204 	u32 qkey;
205 	u32 reserved3[2];
206 };
207 
208 struct mthca_arbel_ud_seg {
209 	u32 av[8];
210 	u32 dqpn;
211 	u32 qkey;
212 	u32 reserved[2];
213 };
214 
215 struct mthca_bind_seg {
216 	u32 flags;		/* [31] Atomic [30] rem write [29] rem read */
217 	u32 reserved;
218 	u32 new_rkey;
219 	u32 lkey;
220 	u64 addr;
221 	u64 length;
222 };
223 
224 struct mthca_raddr_seg {
225 	u64 raddr;
226 	u32 rkey;
227 	u32 reserved;
228 };
229 
230 struct mthca_atomic_seg {
231 	u64 swap_add;
232 	u64 compare;
233 };
234 
235 struct mthca_data_seg {
236 	u32 byte_count;
237 	u32 lkey;
238 	u64 addr;
239 };
240 
241 struct mthca_mlx_seg {
242 	u32 nda_op;
243 	u32 nds;
244 	u32 flags;		/* [17] VL15 [16] SLR [14:12] static rate
245 				   [11:8] SL [3] C [2] E */
246 	u16 rlid;
247 	u16 vcrc;
248 };
249 
250 static const u8 mthca_opcode[] = {
251 	[IB_WR_SEND]                 = MTHCA_OPCODE_SEND,
252 	[IB_WR_SEND_WITH_IMM]        = MTHCA_OPCODE_SEND_IMM,
253 	[IB_WR_RDMA_WRITE]           = MTHCA_OPCODE_RDMA_WRITE,
254 	[IB_WR_RDMA_WRITE_WITH_IMM]  = MTHCA_OPCODE_RDMA_WRITE_IMM,
255 	[IB_WR_RDMA_READ]            = MTHCA_OPCODE_RDMA_READ,
256 	[IB_WR_ATOMIC_CMP_AND_SWP]   = MTHCA_OPCODE_ATOMIC_CS,
257 	[IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
258 };
259 
260 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
261 {
262 	return qp->qpn >= dev->qp_table.sqp_start &&
263 		qp->qpn <= dev->qp_table.sqp_start + 3;
264 }
265 
266 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
267 {
268 	return qp->qpn >= dev->qp_table.sqp_start &&
269 		qp->qpn <= dev->qp_table.sqp_start + 1;
270 }
271 
272 static void *get_recv_wqe(struct mthca_qp *qp, int n)
273 {
274 	if (qp->is_direct)
275 		return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
276 	else
277 		return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
278 			((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
279 }
280 
281 static void *get_send_wqe(struct mthca_qp *qp, int n)
282 {
283 	if (qp->is_direct)
284 		return qp->queue.direct.buf + qp->send_wqe_offset +
285 			(n << qp->sq.wqe_shift);
286 	else
287 		return qp->queue.page_list[(qp->send_wqe_offset +
288 					    (n << qp->sq.wqe_shift)) >>
289 					   PAGE_SHIFT].buf +
290 			((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
291 			 (PAGE_SIZE - 1));
292 }
293 
294 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
295 		    enum ib_event_type event_type)
296 {
297 	struct mthca_qp *qp;
298 	struct ib_event event;
299 
300 	spin_lock(&dev->qp_table.lock);
301 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
302 	if (qp)
303 		atomic_inc(&qp->refcount);
304 	spin_unlock(&dev->qp_table.lock);
305 
306 	if (!qp) {
307 		mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
308 		return;
309 	}
310 
311 	event.device      = &dev->ib_dev;
312 	event.event       = event_type;
313 	event.element.qp  = &qp->ibqp;
314 	if (qp->ibqp.event_handler)
315 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
316 
317 	if (atomic_dec_and_test(&qp->refcount))
318 		wake_up(&qp->wait);
319 }
320 
321 static int to_mthca_state(enum ib_qp_state ib_state)
322 {
323 	switch (ib_state) {
324 	case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
325 	case IB_QPS_INIT:  return MTHCA_QP_STATE_INIT;
326 	case IB_QPS_RTR:   return MTHCA_QP_STATE_RTR;
327 	case IB_QPS_RTS:   return MTHCA_QP_STATE_RTS;
328 	case IB_QPS_SQD:   return MTHCA_QP_STATE_SQD;
329 	case IB_QPS_SQE:   return MTHCA_QP_STATE_SQE;
330 	case IB_QPS_ERR:   return MTHCA_QP_STATE_ERR;
331 	default:                return -1;
332 	}
333 }
334 
335 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
336 
337 static int to_mthca_st(int transport)
338 {
339 	switch (transport) {
340 	case RC:  return MTHCA_QP_ST_RC;
341 	case UC:  return MTHCA_QP_ST_UC;
342 	case UD:  return MTHCA_QP_ST_UD;
343 	case RD:  return MTHCA_QP_ST_RD;
344 	case MLX: return MTHCA_QP_ST_MLX;
345 	default:  return -1;
346 	}
347 }
348 
349 static const struct {
350 	int trans;
351 	u32 req_param[NUM_TRANS];
352 	u32 opt_param[NUM_TRANS];
353 } state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
354 	[IB_QPS_RESET] = {
355 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
356 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
357 		[IB_QPS_INIT]  = {
358 			.trans = MTHCA_TRANS_RST2INIT,
359 			.req_param = {
360 				[UD]  = (IB_QP_PKEY_INDEX |
361 					 IB_QP_PORT       |
362 					 IB_QP_QKEY),
363 				[UC]  = (IB_QP_PKEY_INDEX |
364 					 IB_QP_PORT       |
365 					 IB_QP_ACCESS_FLAGS),
366 				[RC]  = (IB_QP_PKEY_INDEX |
367 					 IB_QP_PORT       |
368 					 IB_QP_ACCESS_FLAGS),
369 				[MLX] = (IB_QP_PKEY_INDEX |
370 					 IB_QP_QKEY),
371 			},
372 			/* bug-for-bug compatibility with VAPI: */
373 			.opt_param = {
374 				[MLX] = IB_QP_PORT
375 			}
376 		},
377 	},
378 	[IB_QPS_INIT]  = {
379 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
380 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
381 		[IB_QPS_INIT]  = {
382 			.trans = MTHCA_TRANS_INIT2INIT,
383 			.opt_param = {
384 				[UD]  = (IB_QP_PKEY_INDEX |
385 					 IB_QP_PORT       |
386 					 IB_QP_QKEY),
387 				[UC]  = (IB_QP_PKEY_INDEX |
388 					 IB_QP_PORT       |
389 					 IB_QP_ACCESS_FLAGS),
390 				[RC]  = (IB_QP_PKEY_INDEX |
391 					 IB_QP_PORT       |
392 					 IB_QP_ACCESS_FLAGS),
393 				[MLX] = (IB_QP_PKEY_INDEX |
394 					 IB_QP_QKEY),
395 			}
396 		},
397 		[IB_QPS_RTR]   = {
398 			.trans = MTHCA_TRANS_INIT2RTR,
399 			.req_param = {
400 				[UC]  = (IB_QP_AV                  |
401 					 IB_QP_PATH_MTU            |
402 					 IB_QP_DEST_QPN            |
403 					 IB_QP_RQ_PSN              |
404 					 IB_QP_MAX_DEST_RD_ATOMIC),
405 				[RC]  = (IB_QP_AV                  |
406 					 IB_QP_PATH_MTU            |
407 					 IB_QP_DEST_QPN            |
408 					 IB_QP_RQ_PSN              |
409 					 IB_QP_MAX_DEST_RD_ATOMIC  |
410 					 IB_QP_MIN_RNR_TIMER),
411 			},
412 			.opt_param = {
413 				[UD]  = (IB_QP_PKEY_INDEX |
414 					 IB_QP_QKEY),
415 				[UC]  = (IB_QP_ALT_PATH     |
416 					 IB_QP_ACCESS_FLAGS |
417 					 IB_QP_PKEY_INDEX),
418 				[RC]  = (IB_QP_ALT_PATH     |
419 					 IB_QP_ACCESS_FLAGS |
420 					 IB_QP_PKEY_INDEX),
421 				[MLX] = (IB_QP_PKEY_INDEX |
422 					 IB_QP_QKEY),
423 			}
424 		}
425 	},
426 	[IB_QPS_RTR]   = {
427 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
428 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
429 		[IB_QPS_RTS]   = {
430 			.trans = MTHCA_TRANS_RTR2RTS,
431 			.req_param = {
432 				[UD]  = IB_QP_SQ_PSN,
433 				[UC]  = (IB_QP_SQ_PSN            |
434 					 IB_QP_MAX_QP_RD_ATOMIC),
435 				[RC]  = (IB_QP_TIMEOUT           |
436 					 IB_QP_RETRY_CNT         |
437 					 IB_QP_RNR_RETRY         |
438 					 IB_QP_SQ_PSN            |
439 					 IB_QP_MAX_QP_RD_ATOMIC),
440 				[MLX] = IB_QP_SQ_PSN,
441 			},
442 			.opt_param = {
443 				[UD]  = (IB_QP_CUR_STATE             |
444 					 IB_QP_QKEY),
445 				[UC]  = (IB_QP_CUR_STATE             |
446 					 IB_QP_ALT_PATH              |
447 					 IB_QP_ACCESS_FLAGS          |
448 					 IB_QP_PKEY_INDEX            |
449 					 IB_QP_PATH_MIG_STATE),
450 				[RC]  = (IB_QP_CUR_STATE             |
451 					 IB_QP_ALT_PATH              |
452 					 IB_QP_ACCESS_FLAGS          |
453 					 IB_QP_PKEY_INDEX            |
454 					 IB_QP_MIN_RNR_TIMER         |
455 					 IB_QP_PATH_MIG_STATE),
456 				[MLX] = (IB_QP_CUR_STATE             |
457 					 IB_QP_QKEY),
458 			}
459 		}
460 	},
461 	[IB_QPS_RTS]   = {
462 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
463 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
464 		[IB_QPS_RTS]   = {
465 			.trans = MTHCA_TRANS_RTS2RTS,
466 			.opt_param = {
467 				[UD]  = (IB_QP_CUR_STATE             |
468 					 IB_QP_QKEY),
469 				[UC]  = (IB_QP_ACCESS_FLAGS          |
470 					 IB_QP_ALT_PATH              |
471 					 IB_QP_PATH_MIG_STATE),
472 				[RC]  = (IB_QP_ACCESS_FLAGS          |
473 					 IB_QP_ALT_PATH              |
474 					 IB_QP_PATH_MIG_STATE        |
475 					 IB_QP_MIN_RNR_TIMER),
476 				[MLX] = (IB_QP_CUR_STATE             |
477 					 IB_QP_QKEY),
478 			}
479 		},
480 		[IB_QPS_SQD]   = {
481 			.trans = MTHCA_TRANS_RTS2SQD,
482 		},
483 	},
484 	[IB_QPS_SQD]   = {
485 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
486 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
487 		[IB_QPS_RTS]   = {
488 			.trans = MTHCA_TRANS_SQD2RTS,
489 			.opt_param = {
490 				[UD]  = (IB_QP_CUR_STATE             |
491 					 IB_QP_QKEY),
492 				[UC]  = (IB_QP_CUR_STATE             |
493 					 IB_QP_ALT_PATH              |
494 					 IB_QP_ACCESS_FLAGS          |
495 					 IB_QP_PATH_MIG_STATE),
496 				[RC]  = (IB_QP_CUR_STATE             |
497 					 IB_QP_ALT_PATH              |
498 					 IB_QP_ACCESS_FLAGS          |
499 					 IB_QP_MIN_RNR_TIMER         |
500 					 IB_QP_PATH_MIG_STATE),
501 				[MLX] = (IB_QP_CUR_STATE             |
502 					 IB_QP_QKEY),
503 			}
504 		},
505 		[IB_QPS_SQD]   = {
506 			.trans = MTHCA_TRANS_SQD2SQD,
507 			.opt_param = {
508 				[UD]  = (IB_QP_PKEY_INDEX            |
509 					 IB_QP_QKEY),
510 				[UC]  = (IB_QP_AV                    |
511 					 IB_QP_MAX_QP_RD_ATOMIC      |
512 					 IB_QP_MAX_DEST_RD_ATOMIC    |
513 					 IB_QP_CUR_STATE             |
514 					 IB_QP_ALT_PATH              |
515 					 IB_QP_ACCESS_FLAGS          |
516 					 IB_QP_PKEY_INDEX            |
517 					 IB_QP_PATH_MIG_STATE),
518 				[RC]  = (IB_QP_AV                    |
519 					 IB_QP_TIMEOUT               |
520 					 IB_QP_RETRY_CNT             |
521 					 IB_QP_RNR_RETRY             |
522 					 IB_QP_MAX_QP_RD_ATOMIC      |
523 					 IB_QP_MAX_DEST_RD_ATOMIC    |
524 					 IB_QP_CUR_STATE             |
525 					 IB_QP_ALT_PATH              |
526 					 IB_QP_ACCESS_FLAGS          |
527 					 IB_QP_PKEY_INDEX            |
528 					 IB_QP_MIN_RNR_TIMER         |
529 					 IB_QP_PATH_MIG_STATE),
530 				[MLX] = (IB_QP_PKEY_INDEX            |
531 					 IB_QP_QKEY),
532 			}
533 		}
534 	},
535 	[IB_QPS_SQE]   = {
536 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
537 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
538 		[IB_QPS_RTS]   = {
539 			.trans = MTHCA_TRANS_SQERR2RTS,
540 			.opt_param = {
541 				[UD]  = (IB_QP_CUR_STATE             |
542 					 IB_QP_QKEY),
543 				[UC]  = (IB_QP_CUR_STATE),
544 				[RC]  = (IB_QP_CUR_STATE             |
545 					 IB_QP_MIN_RNR_TIMER),
546 				[MLX] = (IB_QP_CUR_STATE             |
547 					 IB_QP_QKEY),
548 			}
549 		}
550 	},
551 	[IB_QPS_ERR] = {
552 		[IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
553 		[IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR }
554 	}
555 };
556 
557 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
558 			int attr_mask)
559 {
560 	if (attr_mask & IB_QP_PKEY_INDEX)
561 		sqp->pkey_index = attr->pkey_index;
562 	if (attr_mask & IB_QP_QKEY)
563 		sqp->qkey = attr->qkey;
564 	if (attr_mask & IB_QP_SQ_PSN)
565 		sqp->send_psn = attr->sq_psn;
566 }
567 
568 static void init_port(struct mthca_dev *dev, int port)
569 {
570 	int err;
571 	u8 status;
572 	struct mthca_init_ib_param param;
573 
574 	memset(&param, 0, sizeof param);
575 
576 	param.enable_1x = 1;
577 	param.enable_4x = 1;
578 	param.vl_cap    = dev->limits.vl_cap;
579 	param.mtu_cap   = dev->limits.mtu_cap;
580 	param.gid_cap   = dev->limits.gid_table_len;
581 	param.pkey_cap  = dev->limits.pkey_table_len;
582 
583 	err = mthca_INIT_IB(dev, &param, port, &status);
584 	if (err)
585 		mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
586 	if (status)
587 		mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
588 }
589 
590 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
591 {
592 	struct mthca_dev *dev = to_mdev(ibqp->device);
593 	struct mthca_qp *qp = to_mqp(ibqp);
594 	enum ib_qp_state cur_state, new_state;
595 	struct mthca_mailbox *mailbox;
596 	struct mthca_qp_param *qp_param;
597 	struct mthca_qp_context *qp_context;
598 	u32 req_param, opt_param;
599 	u8 status;
600 	int err;
601 
602 	if (attr_mask & IB_QP_CUR_STATE) {
603 		if (attr->cur_qp_state != IB_QPS_RTR &&
604 		    attr->cur_qp_state != IB_QPS_RTS &&
605 		    attr->cur_qp_state != IB_QPS_SQD &&
606 		    attr->cur_qp_state != IB_QPS_SQE)
607 			return -EINVAL;
608 		else
609 			cur_state = attr->cur_qp_state;
610 	} else {
611 		spin_lock_irq(&qp->sq.lock);
612 		spin_lock(&qp->rq.lock);
613 		cur_state = qp->state;
614 		spin_unlock(&qp->rq.lock);
615 		spin_unlock_irq(&qp->sq.lock);
616 	}
617 
618 	if (attr_mask & IB_QP_STATE) {
619                if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
620 			return -EINVAL;
621 		new_state = attr->qp_state;
622 	} else
623 		new_state = cur_state;
624 
625 	if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
626 		mthca_dbg(dev, "Illegal QP transition "
627 			  "%d->%d\n", cur_state, new_state);
628 		return -EINVAL;
629 	}
630 
631 	req_param = state_table[cur_state][new_state].req_param[qp->transport];
632 	opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
633 
634 	if ((req_param & attr_mask) != req_param) {
635 		mthca_dbg(dev, "QP transition "
636 			  "%d->%d missing req attr 0x%08x\n",
637 			  cur_state, new_state,
638 			  req_param & ~attr_mask);
639 		return -EINVAL;
640 	}
641 
642 	if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
643 		mthca_dbg(dev, "QP transition (transport %d) "
644 			  "%d->%d has extra attr 0x%08x\n",
645 			  qp->transport,
646 			  cur_state, new_state,
647 			  attr_mask & ~(req_param | opt_param |
648 						 IB_QP_STATE));
649 		return -EINVAL;
650 	}
651 
652 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
653 	if (IS_ERR(mailbox))
654 		return PTR_ERR(mailbox);
655 	qp_param = mailbox->buf;
656 	qp_context = &qp_param->context;
657 	memset(qp_param, 0, sizeof *qp_param);
658 
659 	qp_context->flags      = cpu_to_be32((to_mthca_state(new_state) << 28) |
660 					     (to_mthca_st(qp->transport) << 16));
661 	qp_context->flags     |= cpu_to_be32(MTHCA_QP_BIT_DE);
662 	if (!(attr_mask & IB_QP_PATH_MIG_STATE))
663 		qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
664 	else {
665 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
666 		switch (attr->path_mig_state) {
667 		case IB_MIG_MIGRATED:
668 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
669 			break;
670 		case IB_MIG_REARM:
671 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
672 			break;
673 		case IB_MIG_ARMED:
674 			qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
675 			break;
676 		}
677 	}
678 
679 	/* leave tavor_sched_queue as 0 */
680 
681 	if (qp->transport == MLX || qp->transport == UD)
682 		qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
683 	else if (attr_mask & IB_QP_PATH_MTU)
684 		qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
685 
686 	if (mthca_is_memfree(dev)) {
687 		qp_context->rq_size_stride =
688 			((ffs(qp->rq.max) - 1) << 3) | (qp->rq.wqe_shift - 4);
689 		qp_context->sq_size_stride =
690 			((ffs(qp->sq.max) - 1) << 3) | (qp->sq.wqe_shift - 4);
691 	}
692 
693 	/* leave arbel_sched_queue as 0 */
694 
695 	if (qp->ibqp.uobject)
696 		qp_context->usr_page =
697 			cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
698 	else
699 		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
700 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
701 	if (attr_mask & IB_QP_DEST_QPN) {
702 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
703 	}
704 
705 	if (qp->transport == MLX)
706 		qp_context->pri_path.port_pkey |=
707 			cpu_to_be32(to_msqp(qp)->port << 24);
708 	else {
709 		if (attr_mask & IB_QP_PORT) {
710 			qp_context->pri_path.port_pkey |=
711 				cpu_to_be32(attr->port_num << 24);
712 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
713 		}
714 	}
715 
716 	if (attr_mask & IB_QP_PKEY_INDEX) {
717 		qp_context->pri_path.port_pkey |=
718 			cpu_to_be32(attr->pkey_index);
719 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
720 	}
721 
722 	if (attr_mask & IB_QP_RNR_RETRY) {
723 		qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
724 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
725 	}
726 
727 	if (attr_mask & IB_QP_AV) {
728 		qp_context->pri_path.g_mylmc     = attr->ah_attr.src_path_bits & 0x7f;
729 		qp_context->pri_path.rlid        = cpu_to_be16(attr->ah_attr.dlid);
730 		qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
731 		if (attr->ah_attr.ah_flags & IB_AH_GRH) {
732 			qp_context->pri_path.g_mylmc |= 1 << 7;
733 			qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
734 			qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
735 			qp_context->pri_path.sl_tclass_flowlabel =
736 				cpu_to_be32((attr->ah_attr.sl << 28)                |
737 					    (attr->ah_attr.grh.traffic_class << 20) |
738 					    (attr->ah_attr.grh.flow_label));
739 			memcpy(qp_context->pri_path.rgid,
740 			       attr->ah_attr.grh.dgid.raw, 16);
741 		} else {
742 			qp_context->pri_path.sl_tclass_flowlabel =
743 				cpu_to_be32(attr->ah_attr.sl << 28);
744 		}
745 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
746 	}
747 
748 	if (attr_mask & IB_QP_TIMEOUT) {
749 		qp_context->pri_path.ackto = attr->timeout;
750 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
751 	}
752 
753 	/* XXX alt_path */
754 
755 	/* leave rdd as 0 */
756 	qp_context->pd         = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
757 	/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
758 	qp_context->wqe_lkey   = cpu_to_be32(qp->mr.ibmr.lkey);
759 	qp_context->params1    = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
760 					     (MTHCA_FLIGHT_LIMIT << 24) |
761 					     MTHCA_QP_BIT_SRE           |
762 					     MTHCA_QP_BIT_SWE           |
763 					     MTHCA_QP_BIT_SAE);
764 	if (qp->sq_policy == IB_SIGNAL_ALL_WR)
765 		qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
766 	if (attr_mask & IB_QP_RETRY_CNT) {
767 		qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
768 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
769 	}
770 
771 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
772 		qp_context->params1 |= cpu_to_be32(min(attr->max_rd_atomic ?
773 						       ffs(attr->max_rd_atomic) - 1 : 0,
774 						       7) << 21);
775 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
776 	}
777 
778 	if (attr_mask & IB_QP_SQ_PSN)
779 		qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
780 	qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
781 
782 	if (mthca_is_memfree(dev)) {
783 		qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
784 		qp_context->snd_db_index   = cpu_to_be32(qp->sq.db_index);
785 	}
786 
787 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
788 		/*
789 		 * Only enable RDMA/atomics if we have responder
790 		 * resources set to a non-zero value.
791 		 */
792 		if (qp->resp_depth) {
793 			qp_context->params2 |=
794 				cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
795 					    MTHCA_QP_BIT_RWE : 0);
796 			qp_context->params2 |=
797 				cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
798 					    MTHCA_QP_BIT_RRE : 0);
799 			qp_context->params2 |=
800 				cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
801 					    MTHCA_QP_BIT_RAE : 0);
802 		}
803 
804 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
805 							MTHCA_QP_OPTPAR_RRE |
806 							MTHCA_QP_OPTPAR_RAE);
807 
808 		qp->atomic_rd_en = attr->qp_access_flags;
809 	}
810 
811 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
812 		u8 rra_max;
813 
814 		if (qp->resp_depth && !attr->max_dest_rd_atomic) {
815 			/*
816 			 * Lowering our responder resources to zero.
817 			 * Turn off RDMA/atomics as responder.
818 			 * (RWE/RRE/RAE in params2 already zero)
819 			 */
820 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
821 								MTHCA_QP_OPTPAR_RRE |
822 								MTHCA_QP_OPTPAR_RAE);
823 		}
824 
825 		if (!qp->resp_depth && attr->max_dest_rd_atomic) {
826 			/*
827 			 * Increasing our responder resources from
828 			 * zero.  Turn on RDMA/atomics as appropriate.
829 			 */
830 			qp_context->params2 |=
831 				cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_WRITE ?
832 					    MTHCA_QP_BIT_RWE : 0);
833 			qp_context->params2 |=
834 				cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
835 					    MTHCA_QP_BIT_RRE : 0);
836 			qp_context->params2 |=
837 				cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
838 					    MTHCA_QP_BIT_RAE : 0);
839 
840 			qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
841 								MTHCA_QP_OPTPAR_RRE |
842 								MTHCA_QP_OPTPAR_RAE);
843 		}
844 
845 		for (rra_max = 0;
846 		     1 << rra_max < attr->max_dest_rd_atomic &&
847 			     rra_max < dev->qp_table.rdb_shift;
848 		     ++rra_max)
849 			; /* nothing */
850 
851 		qp_context->params2      |= cpu_to_be32(rra_max << 21);
852 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
853 
854 		qp->resp_depth = attr->max_dest_rd_atomic;
855 	}
856 
857 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
858 
859 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
860 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
861 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
862 	}
863 	if (attr_mask & IB_QP_RQ_PSN)
864 		qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
865 
866 	qp_context->ra_buff_indx =
867 		cpu_to_be32(dev->qp_table.rdb_base +
868 			    ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
869 			     dev->qp_table.rdb_shift));
870 
871 	qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
872 
873 	if (mthca_is_memfree(dev))
874 		qp_context->rcv_db_index   = cpu_to_be32(qp->rq.db_index);
875 
876 	if (attr_mask & IB_QP_QKEY) {
877 		qp_context->qkey = cpu_to_be32(attr->qkey);
878 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
879 	}
880 
881 	err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
882 			      qp->qpn, 0, mailbox, 0, &status);
883 	if (status) {
884 		mthca_warn(dev, "modify QP %d returned status %02x.\n",
885 			   state_table[cur_state][new_state].trans, status);
886 		err = -EINVAL;
887 	}
888 
889 	if (!err)
890 		qp->state = new_state;
891 
892 	mthca_free_mailbox(dev, mailbox);
893 
894 	if (is_sqp(dev, qp))
895 		store_attrs(to_msqp(qp), attr, attr_mask);
896 
897 	/*
898 	 * If we are moving QP0 to RTR, bring the IB link up; if we
899 	 * are moving QP0 to RESET or ERROR, bring the link back down.
900 	 */
901 	if (is_qp0(dev, qp)) {
902 		if (cur_state != IB_QPS_RTR &&
903 		    new_state == IB_QPS_RTR)
904 			init_port(dev, to_msqp(qp)->port);
905 
906 		if (cur_state != IB_QPS_RESET &&
907 		    cur_state != IB_QPS_ERR &&
908 		    (new_state == IB_QPS_RESET ||
909 		     new_state == IB_QPS_ERR))
910 			mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
911 	}
912 
913 	return err;
914 }
915 
916 /*
917  * Allocate and register buffer for WQEs.  qp->rq.max, sq.max,
918  * rq.max_gs and sq.max_gs must all be assigned.
919  * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
920  * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
921  * queue)
922  */
923 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
924 			       struct mthca_pd *pd,
925 			       struct mthca_qp *qp)
926 {
927 	int size;
928 	int i;
929 	int npages, shift;
930 	dma_addr_t t;
931 	u64 *dma_list = NULL;
932 	int err = -ENOMEM;
933 
934 	size = sizeof (struct mthca_next_seg) +
935 		qp->rq.max_gs * sizeof (struct mthca_data_seg);
936 
937 	for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
938 	     qp->rq.wqe_shift++)
939 		; /* nothing */
940 
941 	size = sizeof (struct mthca_next_seg) +
942 		qp->sq.max_gs * sizeof (struct mthca_data_seg);
943 	switch (qp->transport) {
944 	case MLX:
945 		size += 2 * sizeof (struct mthca_data_seg);
946 		break;
947 	case UD:
948 		if (mthca_is_memfree(dev))
949 			size += sizeof (struct mthca_arbel_ud_seg);
950 		else
951 			size += sizeof (struct mthca_tavor_ud_seg);
952 		break;
953 	default:
954 		/* bind seg is as big as atomic + raddr segs */
955 		size += sizeof (struct mthca_bind_seg);
956 	}
957 
958 	for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
959 	     qp->sq.wqe_shift++)
960 		; /* nothing */
961 
962 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
963 				    1 << qp->sq.wqe_shift);
964 
965 	/*
966 	 * If this is a userspace QP, we don't actually have to
967 	 * allocate anything.  All we need is to calculate the WQE
968 	 * sizes and the send_wqe_offset, so we're done now.
969 	 */
970 	if (pd->ibpd.uobject)
971 		return 0;
972 
973 	size = PAGE_ALIGN(qp->send_wqe_offset +
974 			  (qp->sq.max << qp->sq.wqe_shift));
975 
976 	qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
977 			   GFP_KERNEL);
978 	if (!qp->wrid)
979 		goto err_out;
980 
981 	if (size <= MTHCA_MAX_DIRECT_QP_SIZE) {
982 		qp->is_direct = 1;
983 		npages = 1;
984 		shift = get_order(size) + PAGE_SHIFT;
985 
986 		if (0)
987 			mthca_dbg(dev, "Creating direct QP of size %d (shift %d)\n",
988 				  size, shift);
989 
990 		qp->queue.direct.buf = dma_alloc_coherent(&dev->pdev->dev, size,
991 							  &t, GFP_KERNEL);
992 		if (!qp->queue.direct.buf)
993 			goto err_out;
994 
995 		pci_unmap_addr_set(&qp->queue.direct, mapping, t);
996 
997 		memset(qp->queue.direct.buf, 0, size);
998 
999 		while (t & ((1 << shift) - 1)) {
1000 			--shift;
1001 			npages *= 2;
1002 		}
1003 
1004 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
1005 		if (!dma_list)
1006 			goto err_out_free;
1007 
1008 		for (i = 0; i < npages; ++i)
1009 			dma_list[i] = t + i * (1 << shift);
1010 	} else {
1011 		qp->is_direct = 0;
1012 		npages = size / PAGE_SIZE;
1013 		shift = PAGE_SHIFT;
1014 
1015 		if (0)
1016 			mthca_dbg(dev, "Creating indirect QP with %d pages\n", npages);
1017 
1018 		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
1019 		if (!dma_list)
1020 			goto err_out;
1021 
1022 		qp->queue.page_list = kmalloc(npages *
1023 					      sizeof *qp->queue.page_list,
1024 					      GFP_KERNEL);
1025 		if (!qp->queue.page_list)
1026 			goto err_out;
1027 
1028 		for (i = 0; i < npages; ++i) {
1029 			qp->queue.page_list[i].buf =
1030 				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
1031 						   &t, GFP_KERNEL);
1032 			if (!qp->queue.page_list[i].buf)
1033 				goto err_out_free;
1034 
1035 			memset(qp->queue.page_list[i].buf, 0, PAGE_SIZE);
1036 
1037 			pci_unmap_addr_set(&qp->queue.page_list[i], mapping, t);
1038 			dma_list[i] = t;
1039 		}
1040 	}
1041 
1042 	err = mthca_mr_alloc_phys(dev, pd->pd_num, dma_list, shift,
1043 				  npages, 0, size,
1044 				  MTHCA_MPT_FLAG_LOCAL_READ,
1045 				  &qp->mr);
1046 	if (err)
1047 		goto err_out_free;
1048 
1049 	kfree(dma_list);
1050 	return 0;
1051 
1052  err_out_free:
1053 	if (qp->is_direct) {
1054 		dma_free_coherent(&dev->pdev->dev, size, qp->queue.direct.buf,
1055 				  pci_unmap_addr(&qp->queue.direct, mapping));
1056 	} else
1057 		for (i = 0; i < npages; ++i) {
1058 			if (qp->queue.page_list[i].buf)
1059 				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
1060 						  qp->queue.page_list[i].buf,
1061 						  pci_unmap_addr(&qp->queue.page_list[i],
1062 								 mapping));
1063 
1064 		}
1065 
1066  err_out:
1067 	kfree(qp->wrid);
1068 	kfree(dma_list);
1069 	return err;
1070 }
1071 
1072 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1073 			       struct mthca_qp *qp)
1074 {
1075 	int i;
1076 	int size = PAGE_ALIGN(qp->send_wqe_offset +
1077 			      (qp->sq.max << qp->sq.wqe_shift));
1078 
1079 	if (qp->is_direct) {
1080 		dma_free_coherent(&dev->pdev->dev, size, qp->queue.direct.buf,
1081 				  pci_unmap_addr(&qp->queue.direct, mapping));
1082 	} else {
1083 		for (i = 0; i < size / PAGE_SIZE; ++i) {
1084 			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
1085 					  qp->queue.page_list[i].buf,
1086 					  pci_unmap_addr(&qp->queue.page_list[i],
1087 							 mapping));
1088 		}
1089 	}
1090 
1091 	kfree(qp->wrid);
1092 }
1093 
1094 static int mthca_map_memfree(struct mthca_dev *dev,
1095 			     struct mthca_qp *qp)
1096 {
1097 	int ret;
1098 
1099 	if (mthca_is_memfree(dev)) {
1100 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1101 		if (ret)
1102 			return ret;
1103 
1104 		ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1105 		if (ret)
1106 			goto err_qpc;
1107 
1108  		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1109  				      qp->qpn << dev->qp_table.rdb_shift);
1110  		if (ret)
1111  			goto err_eqpc;
1112 
1113 	}
1114 
1115 	return 0;
1116 
1117 err_eqpc:
1118 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1119 
1120 err_qpc:
1121 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1122 
1123 	return ret;
1124 }
1125 
1126 static void mthca_unmap_memfree(struct mthca_dev *dev,
1127 				struct mthca_qp *qp)
1128 {
1129 	mthca_table_put(dev, dev->qp_table.rdb_table,
1130 			qp->qpn << dev->qp_table.rdb_shift);
1131 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1132 	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1133 }
1134 
1135 static int mthca_alloc_memfree(struct mthca_dev *dev,
1136 			       struct mthca_qp *qp)
1137 {
1138 	int ret = 0;
1139 
1140 	if (mthca_is_memfree(dev)) {
1141 		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1142 						 qp->qpn, &qp->rq.db);
1143 		if (qp->rq.db_index < 0)
1144 			return ret;
1145 
1146 		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1147 						 qp->qpn, &qp->sq.db);
1148 		if (qp->sq.db_index < 0)
1149 			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1150 	}
1151 
1152 	return ret;
1153 }
1154 
1155 static void mthca_free_memfree(struct mthca_dev *dev,
1156 			       struct mthca_qp *qp)
1157 {
1158 	if (mthca_is_memfree(dev)) {
1159 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1160 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1161 	}
1162 }
1163 
1164 static void mthca_wq_init(struct mthca_wq* wq)
1165 {
1166 	spin_lock_init(&wq->lock);
1167 	wq->next_ind  = 0;
1168 	wq->last_comp = wq->max - 1;
1169 	wq->head      = 0;
1170 	wq->tail      = 0;
1171 	wq->last      = NULL;
1172 }
1173 
1174 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1175 				 struct mthca_pd *pd,
1176 				 struct mthca_cq *send_cq,
1177 				 struct mthca_cq *recv_cq,
1178 				 enum ib_sig_type send_policy,
1179 				 struct mthca_qp *qp)
1180 {
1181 	int ret;
1182 	int i;
1183 
1184 	atomic_set(&qp->refcount, 1);
1185 	qp->state    	 = IB_QPS_RESET;
1186 	qp->atomic_rd_en = 0;
1187 	qp->resp_depth   = 0;
1188 	qp->sq_policy    = send_policy;
1189 	mthca_wq_init(&qp->sq);
1190 	mthca_wq_init(&qp->rq);
1191 
1192 	ret = mthca_map_memfree(dev, qp);
1193 	if (ret)
1194 		return ret;
1195 
1196 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
1197 	if (ret) {
1198 		mthca_unmap_memfree(dev, qp);
1199 		return ret;
1200 	}
1201 
1202 	/*
1203 	 * If this is a userspace QP, we're done now.  The doorbells
1204 	 * will be allocated and buffers will be initialized in
1205 	 * userspace.
1206 	 */
1207 	if (pd->ibpd.uobject)
1208 		return 0;
1209 
1210 	ret = mthca_alloc_memfree(dev, qp);
1211 	if (ret) {
1212 		mthca_free_wqe_buf(dev, qp);
1213 		mthca_unmap_memfree(dev, qp);
1214 		return ret;
1215 	}
1216 
1217 	if (mthca_is_memfree(dev)) {
1218 		struct mthca_next_seg *next;
1219 		struct mthca_data_seg *scatter;
1220 		int size = (sizeof (struct mthca_next_seg) +
1221 			    qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1222 
1223 		for (i = 0; i < qp->rq.max; ++i) {
1224 			next = get_recv_wqe(qp, i);
1225 			next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1226 						   qp->rq.wqe_shift);
1227 			next->ee_nds = cpu_to_be32(size);
1228 
1229 			for (scatter = (void *) (next + 1);
1230 			     (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1231 			     ++scatter)
1232 				scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1233 		}
1234 
1235 		for (i = 0; i < qp->sq.max; ++i) {
1236 			next = get_send_wqe(qp, i);
1237 			next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1238 						    qp->sq.wqe_shift) +
1239 						   qp->send_wqe_offset);
1240 		}
1241 	}
1242 
1243 	return 0;
1244 }
1245 
1246 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1247 			     struct mthca_qp *qp)
1248 {
1249 	/* Sanity check QP size before proceeding */
1250 	if (cap->max_send_wr  > 65536 || cap->max_recv_wr  > 65536 ||
1251 	    cap->max_send_sge > 64    || cap->max_recv_sge > 64)
1252 		return -EINVAL;
1253 
1254 	if (mthca_is_memfree(dev)) {
1255 		qp->rq.max = cap->max_recv_wr ?
1256 			roundup_pow_of_two(cap->max_recv_wr) : 0;
1257 		qp->sq.max = cap->max_send_wr ?
1258 			roundup_pow_of_two(cap->max_send_wr) : 0;
1259 	} else {
1260 		qp->rq.max = cap->max_recv_wr;
1261 		qp->sq.max = cap->max_send_wr;
1262 	}
1263 
1264 	qp->rq.max_gs = cap->max_recv_sge;
1265 	qp->sq.max_gs = max_t(int, cap->max_send_sge,
1266 			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1267 				    MTHCA_INLINE_CHUNK_SIZE) /
1268 			      sizeof (struct mthca_data_seg));
1269 
1270 	/*
1271 	 * For MLX transport we need 2 extra S/G entries:
1272 	 * one for the header and one for the checksum at the end
1273 	 */
1274 	if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1275 	    qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1276 		return -EINVAL;
1277 
1278 	return 0;
1279 }
1280 
1281 int mthca_alloc_qp(struct mthca_dev *dev,
1282 		   struct mthca_pd *pd,
1283 		   struct mthca_cq *send_cq,
1284 		   struct mthca_cq *recv_cq,
1285 		   enum ib_qp_type type,
1286 		   enum ib_sig_type send_policy,
1287 		   struct ib_qp_cap *cap,
1288 		   struct mthca_qp *qp)
1289 {
1290 	int err;
1291 
1292 	err = mthca_set_qp_size(dev, cap, qp);
1293 	if (err)
1294 		return err;
1295 
1296 	switch (type) {
1297 	case IB_QPT_RC: qp->transport = RC; break;
1298 	case IB_QPT_UC: qp->transport = UC; break;
1299 	case IB_QPT_UD: qp->transport = UD; break;
1300 	default: return -EINVAL;
1301 	}
1302 
1303 	qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1304 	if (qp->qpn == -1)
1305 		return -ENOMEM;
1306 
1307 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1308 				    send_policy, qp);
1309 	if (err) {
1310 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1311 		return err;
1312 	}
1313 
1314 	spin_lock_irq(&dev->qp_table.lock);
1315 	mthca_array_set(&dev->qp_table.qp,
1316 			qp->qpn & (dev->limits.num_qps - 1), qp);
1317 	spin_unlock_irq(&dev->qp_table.lock);
1318 
1319 	return 0;
1320 }
1321 
1322 int mthca_alloc_sqp(struct mthca_dev *dev,
1323 		    struct mthca_pd *pd,
1324 		    struct mthca_cq *send_cq,
1325 		    struct mthca_cq *recv_cq,
1326 		    enum ib_sig_type send_policy,
1327 		    struct ib_qp_cap *cap,
1328 		    int qpn,
1329 		    int port,
1330 		    struct mthca_sqp *sqp)
1331 {
1332 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1333 	int err;
1334 
1335 	err = mthca_set_qp_size(dev, cap, &sqp->qp);
1336 	if (err)
1337 		return err;
1338 
1339 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1340 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1341 					     &sqp->header_dma, GFP_KERNEL);
1342 	if (!sqp->header_buf)
1343 		return -ENOMEM;
1344 
1345 	spin_lock_irq(&dev->qp_table.lock);
1346 	if (mthca_array_get(&dev->qp_table.qp, mqpn))
1347 		err = -EBUSY;
1348 	else
1349 		mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1350 	spin_unlock_irq(&dev->qp_table.lock);
1351 
1352 	if (err)
1353 		goto err_out;
1354 
1355 	sqp->port = port;
1356 	sqp->qp.qpn       = mqpn;
1357 	sqp->qp.transport = MLX;
1358 
1359 	err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1360 				    send_policy, &sqp->qp);
1361 	if (err)
1362 		goto err_out_free;
1363 
1364 	atomic_inc(&pd->sqp_count);
1365 
1366 	return 0;
1367 
1368  err_out_free:
1369 	/*
1370 	 * Lock CQs here, so that CQ polling code can do QP lookup
1371 	 * without taking a lock.
1372 	 */
1373 	spin_lock_irq(&send_cq->lock);
1374 	if (send_cq != recv_cq)
1375 		spin_lock(&recv_cq->lock);
1376 
1377 	spin_lock(&dev->qp_table.lock);
1378 	mthca_array_clear(&dev->qp_table.qp, mqpn);
1379 	spin_unlock(&dev->qp_table.lock);
1380 
1381 	if (send_cq != recv_cq)
1382 		spin_unlock(&recv_cq->lock);
1383 	spin_unlock_irq(&send_cq->lock);
1384 
1385  err_out:
1386 	dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1387 			  sqp->header_buf, sqp->header_dma);
1388 
1389 	return err;
1390 }
1391 
1392 void mthca_free_qp(struct mthca_dev *dev,
1393 		   struct mthca_qp *qp)
1394 {
1395 	u8 status;
1396 	struct mthca_cq *send_cq;
1397 	struct mthca_cq *recv_cq;
1398 
1399 	send_cq = to_mcq(qp->ibqp.send_cq);
1400 	recv_cq = to_mcq(qp->ibqp.recv_cq);
1401 
1402 	/*
1403 	 * Lock CQs here, so that CQ polling code can do QP lookup
1404 	 * without taking a lock.
1405 	 */
1406 	spin_lock_irq(&send_cq->lock);
1407 	if (send_cq != recv_cq)
1408 		spin_lock(&recv_cq->lock);
1409 
1410 	spin_lock(&dev->qp_table.lock);
1411 	mthca_array_clear(&dev->qp_table.qp,
1412 			  qp->qpn & (dev->limits.num_qps - 1));
1413 	spin_unlock(&dev->qp_table.lock);
1414 
1415 	if (send_cq != recv_cq)
1416 		spin_unlock(&recv_cq->lock);
1417 	spin_unlock_irq(&send_cq->lock);
1418 
1419 	atomic_dec(&qp->refcount);
1420 	wait_event(qp->wait, !atomic_read(&qp->refcount));
1421 
1422 	if (qp->state != IB_QPS_RESET)
1423 		mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1424 
1425 	/*
1426 	 * If this is a userspace QP, the buffers, MR, CQs and so on
1427 	 * will be cleaned up in userspace, so all we have to do is
1428 	 * unref the mem-free tables and free the QPN in our table.
1429 	 */
1430 	if (!qp->ibqp.uobject) {
1431 		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
1432 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1433 			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
1434 
1435 		mthca_free_mr(dev, &qp->mr);
1436 		mthca_free_memfree(dev, qp);
1437 		mthca_free_wqe_buf(dev, qp);
1438 	}
1439 
1440 	mthca_unmap_memfree(dev, qp);
1441 
1442 	if (is_sqp(dev, qp)) {
1443 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1444 		dma_free_coherent(&dev->pdev->dev,
1445 				  to_msqp(qp)->header_buf_size,
1446 				  to_msqp(qp)->header_buf,
1447 				  to_msqp(qp)->header_dma);
1448 	} else
1449 		mthca_free(&dev->qp_table.alloc, qp->qpn);
1450 }
1451 
1452 /* Create UD header for an MLX send and build a data segment for it */
1453 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1454 			    int ind, struct ib_send_wr *wr,
1455 			    struct mthca_mlx_seg *mlx,
1456 			    struct mthca_data_seg *data)
1457 {
1458 	int header_size;
1459 	int err;
1460 
1461 	ib_ud_header_init(256, /* assume a MAD */
1462 			  sqp->ud_header.grh_present,
1463 			  &sqp->ud_header);
1464 
1465 	err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1466 	if (err)
1467 		return err;
1468 	mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1469 	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1470 				  (sqp->ud_header.lrh.destination_lid == 0xffff ?
1471 				   MTHCA_MLX_SLR : 0) |
1472 				  (sqp->ud_header.lrh.service_level << 8));
1473 	mlx->rlid = sqp->ud_header.lrh.destination_lid;
1474 	mlx->vcrc = 0;
1475 
1476 	switch (wr->opcode) {
1477 	case IB_WR_SEND:
1478 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1479 		sqp->ud_header.immediate_present = 0;
1480 		break;
1481 	case IB_WR_SEND_WITH_IMM:
1482 		sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1483 		sqp->ud_header.immediate_present = 1;
1484 		sqp->ud_header.immediate_data = wr->imm_data;
1485 		break;
1486 	default:
1487 		return -EINVAL;
1488 	}
1489 
1490 	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
1491 	if (sqp->ud_header.lrh.destination_lid == 0xffff)
1492 		sqp->ud_header.lrh.source_lid = 0xffff;
1493 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1494 	if (!sqp->qp.ibqp.qp_num)
1495 		ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1496 				   sqp->pkey_index,
1497 				   &sqp->ud_header.bth.pkey);
1498 	else
1499 		ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1500 				   wr->wr.ud.pkey_index,
1501 				   &sqp->ud_header.bth.pkey);
1502 	cpu_to_be16s(&sqp->ud_header.bth.pkey);
1503 	sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1504 	sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1505 	sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1506 					       sqp->qkey : wr->wr.ud.remote_qkey);
1507 	sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1508 
1509 	header_size = ib_ud_header_pack(&sqp->ud_header,
1510 					sqp->header_buf +
1511 					ind * MTHCA_UD_HEADER_SIZE);
1512 
1513 	data->byte_count = cpu_to_be32(header_size);
1514 	data->lkey       = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1515 	data->addr       = cpu_to_be64(sqp->header_dma +
1516 				       ind * MTHCA_UD_HEADER_SIZE);
1517 
1518 	return 0;
1519 }
1520 
1521 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1522 				    struct ib_cq *ib_cq)
1523 {
1524 	unsigned cur;
1525 	struct mthca_cq *cq;
1526 
1527 	cur = wq->head - wq->tail;
1528 	if (likely(cur + nreq < wq->max))
1529 		return 0;
1530 
1531 	cq = to_mcq(ib_cq);
1532 	spin_lock(&cq->lock);
1533 	cur = wq->head - wq->tail;
1534 	spin_unlock(&cq->lock);
1535 
1536 	return cur + nreq >= wq->max;
1537 }
1538 
1539 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1540 			  struct ib_send_wr **bad_wr)
1541 {
1542 	struct mthca_dev *dev = to_mdev(ibqp->device);
1543 	struct mthca_qp *qp = to_mqp(ibqp);
1544 	void *wqe;
1545 	void *prev_wqe;
1546 	unsigned long flags;
1547 	int err = 0;
1548 	int nreq;
1549 	int i;
1550 	int size;
1551 	int size0 = 0;
1552 	u32 f0 = 0;
1553 	int ind;
1554 	u8 op0 = 0;
1555 
1556 	spin_lock_irqsave(&qp->sq.lock, flags);
1557 
1558 	/* XXX check that state is OK to post send */
1559 
1560 	ind = qp->sq.next_ind;
1561 
1562 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1563 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1564 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1565 					" %d max, %d nreq)\n", qp->qpn,
1566 					qp->sq.head, qp->sq.tail,
1567 					qp->sq.max, nreq);
1568 			err = -ENOMEM;
1569 			*bad_wr = wr;
1570 			goto out;
1571 		}
1572 
1573 		wqe = get_send_wqe(qp, ind);
1574 		prev_wqe = qp->sq.last;
1575 		qp->sq.last = wqe;
1576 
1577 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1578 		((struct mthca_next_seg *) wqe)->ee_nds = 0;
1579 		((struct mthca_next_seg *) wqe)->flags =
1580 			((wr->send_flags & IB_SEND_SIGNALED) ?
1581 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1582 			((wr->send_flags & IB_SEND_SOLICITED) ?
1583 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1584 			cpu_to_be32(1);
1585 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1586 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1587 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1588 
1589 		wqe += sizeof (struct mthca_next_seg);
1590 		size = sizeof (struct mthca_next_seg) / 16;
1591 
1592 		switch (qp->transport) {
1593 		case RC:
1594 			switch (wr->opcode) {
1595 			case IB_WR_ATOMIC_CMP_AND_SWP:
1596 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1597 				((struct mthca_raddr_seg *) wqe)->raddr =
1598 					cpu_to_be64(wr->wr.atomic.remote_addr);
1599 				((struct mthca_raddr_seg *) wqe)->rkey =
1600 					cpu_to_be32(wr->wr.atomic.rkey);
1601 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1602 
1603 				wqe += sizeof (struct mthca_raddr_seg);
1604 
1605 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1606 					((struct mthca_atomic_seg *) wqe)->swap_add =
1607 						cpu_to_be64(wr->wr.atomic.swap);
1608 					((struct mthca_atomic_seg *) wqe)->compare =
1609 						cpu_to_be64(wr->wr.atomic.compare_add);
1610 				} else {
1611 					((struct mthca_atomic_seg *) wqe)->swap_add =
1612 						cpu_to_be64(wr->wr.atomic.compare_add);
1613 					((struct mthca_atomic_seg *) wqe)->compare = 0;
1614 				}
1615 
1616 				wqe += sizeof (struct mthca_atomic_seg);
1617 				size += sizeof (struct mthca_raddr_seg) / 16 +
1618 					sizeof (struct mthca_atomic_seg);
1619 				break;
1620 
1621 			case IB_WR_RDMA_WRITE:
1622 			case IB_WR_RDMA_WRITE_WITH_IMM:
1623 			case IB_WR_RDMA_READ:
1624 				((struct mthca_raddr_seg *) wqe)->raddr =
1625 					cpu_to_be64(wr->wr.rdma.remote_addr);
1626 				((struct mthca_raddr_seg *) wqe)->rkey =
1627 					cpu_to_be32(wr->wr.rdma.rkey);
1628 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1629 				wqe += sizeof (struct mthca_raddr_seg);
1630 				size += sizeof (struct mthca_raddr_seg) / 16;
1631 				break;
1632 
1633 			default:
1634 				/* No extra segments required for sends */
1635 				break;
1636 			}
1637 
1638 			break;
1639 
1640 		case UC:
1641 			switch (wr->opcode) {
1642 			case IB_WR_RDMA_WRITE:
1643 			case IB_WR_RDMA_WRITE_WITH_IMM:
1644 				((struct mthca_raddr_seg *) wqe)->raddr =
1645 					cpu_to_be64(wr->wr.rdma.remote_addr);
1646 				((struct mthca_raddr_seg *) wqe)->rkey =
1647 					cpu_to_be32(wr->wr.rdma.rkey);
1648 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1649 				wqe += sizeof (struct mthca_raddr_seg);
1650 				size += sizeof (struct mthca_raddr_seg) / 16;
1651 				break;
1652 
1653 			default:
1654 				/* No extra segments required for sends */
1655 				break;
1656 			}
1657 
1658 			break;
1659 
1660 		case UD:
1661 			((struct mthca_tavor_ud_seg *) wqe)->lkey =
1662 				cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1663 			((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1664 				cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1665 			((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1666 				cpu_to_be32(wr->wr.ud.remote_qpn);
1667 			((struct mthca_tavor_ud_seg *) wqe)->qkey =
1668 				cpu_to_be32(wr->wr.ud.remote_qkey);
1669 
1670 			wqe += sizeof (struct mthca_tavor_ud_seg);
1671 			size += sizeof (struct mthca_tavor_ud_seg) / 16;
1672 			break;
1673 
1674 		case MLX:
1675 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1676 					       wqe - sizeof (struct mthca_next_seg),
1677 					       wqe);
1678 			if (err) {
1679 				*bad_wr = wr;
1680 				goto out;
1681 			}
1682 			wqe += sizeof (struct mthca_data_seg);
1683 			size += sizeof (struct mthca_data_seg) / 16;
1684 			break;
1685 		}
1686 
1687 		if (wr->num_sge > qp->sq.max_gs) {
1688 			mthca_err(dev, "too many gathers\n");
1689 			err = -EINVAL;
1690 			*bad_wr = wr;
1691 			goto out;
1692 		}
1693 
1694 		for (i = 0; i < wr->num_sge; ++i) {
1695 			((struct mthca_data_seg *) wqe)->byte_count =
1696 				cpu_to_be32(wr->sg_list[i].length);
1697 			((struct mthca_data_seg *) wqe)->lkey =
1698 				cpu_to_be32(wr->sg_list[i].lkey);
1699 			((struct mthca_data_seg *) wqe)->addr =
1700 				cpu_to_be64(wr->sg_list[i].addr);
1701 			wqe += sizeof (struct mthca_data_seg);
1702 			size += sizeof (struct mthca_data_seg) / 16;
1703 		}
1704 
1705 		/* Add one more inline data segment for ICRC */
1706 		if (qp->transport == MLX) {
1707 			((struct mthca_data_seg *) wqe)->byte_count =
1708 				cpu_to_be32((1 << 31) | 4);
1709 			((u32 *) wqe)[1] = 0;
1710 			wqe += sizeof (struct mthca_data_seg);
1711 			size += sizeof (struct mthca_data_seg) / 16;
1712 		}
1713 
1714 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
1715 
1716 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1717 			mthca_err(dev, "opcode invalid\n");
1718 			err = -EINVAL;
1719 			*bad_wr = wr;
1720 			goto out;
1721 		}
1722 
1723 		if (prev_wqe) {
1724 			((struct mthca_next_seg *) prev_wqe)->nda_op =
1725 				cpu_to_be32(((ind << qp->sq.wqe_shift) +
1726 					     qp->send_wqe_offset) |
1727 					    mthca_opcode[wr->opcode]);
1728 			wmb();
1729 			((struct mthca_next_seg *) prev_wqe)->ee_nds =
1730 				cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1731 		}
1732 
1733 		if (!size0) {
1734 			size0 = size;
1735 			op0   = mthca_opcode[wr->opcode];
1736 		}
1737 
1738 		++ind;
1739 		if (unlikely(ind >= qp->sq.max))
1740 			ind -= qp->sq.max;
1741 	}
1742 
1743 out:
1744 	if (likely(nreq)) {
1745 		u32 doorbell[2];
1746 
1747 		doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1748 					   qp->send_wqe_offset) | f0 | op0);
1749 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1750 
1751 		wmb();
1752 
1753 		mthca_write64(doorbell,
1754 			      dev->kar + MTHCA_SEND_DOORBELL,
1755 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1756 	}
1757 
1758 	qp->sq.next_ind = ind;
1759 	qp->sq.head    += nreq;
1760 
1761 	spin_unlock_irqrestore(&qp->sq.lock, flags);
1762 	return err;
1763 }
1764 
1765 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1766 			     struct ib_recv_wr **bad_wr)
1767 {
1768 	struct mthca_dev *dev = to_mdev(ibqp->device);
1769 	struct mthca_qp *qp = to_mqp(ibqp);
1770 	unsigned long flags;
1771 	int err = 0;
1772 	int nreq;
1773 	int i;
1774 	int size;
1775 	int size0 = 0;
1776 	int ind;
1777 	void *wqe;
1778 	void *prev_wqe;
1779 
1780 	spin_lock_irqsave(&qp->rq.lock, flags);
1781 
1782 	/* XXX check that state is OK to post receive */
1783 
1784 	ind = qp->rq.next_ind;
1785 
1786 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1787 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1788 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1789 					" %d max, %d nreq)\n", qp->qpn,
1790 					qp->rq.head, qp->rq.tail,
1791 					qp->rq.max, nreq);
1792 			err = -ENOMEM;
1793 			*bad_wr = wr;
1794 			goto out;
1795 		}
1796 
1797 		wqe = get_recv_wqe(qp, ind);
1798 		prev_wqe = qp->rq.last;
1799 		qp->rq.last = wqe;
1800 
1801 		((struct mthca_next_seg *) wqe)->nda_op = 0;
1802 		((struct mthca_next_seg *) wqe)->ee_nds =
1803 			cpu_to_be32(MTHCA_NEXT_DBD);
1804 		((struct mthca_next_seg *) wqe)->flags = 0;
1805 
1806 		wqe += sizeof (struct mthca_next_seg);
1807 		size = sizeof (struct mthca_next_seg) / 16;
1808 
1809 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1810 			err = -EINVAL;
1811 			*bad_wr = wr;
1812 			goto out;
1813 		}
1814 
1815 		for (i = 0; i < wr->num_sge; ++i) {
1816 			((struct mthca_data_seg *) wqe)->byte_count =
1817 				cpu_to_be32(wr->sg_list[i].length);
1818 			((struct mthca_data_seg *) wqe)->lkey =
1819 				cpu_to_be32(wr->sg_list[i].lkey);
1820 			((struct mthca_data_seg *) wqe)->addr =
1821 				cpu_to_be64(wr->sg_list[i].addr);
1822 			wqe += sizeof (struct mthca_data_seg);
1823 			size += sizeof (struct mthca_data_seg) / 16;
1824 		}
1825 
1826 		qp->wrid[ind] = wr->wr_id;
1827 
1828 		if (likely(prev_wqe)) {
1829 			((struct mthca_next_seg *) prev_wqe)->nda_op =
1830 				cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1831 			wmb();
1832 			((struct mthca_next_seg *) prev_wqe)->ee_nds =
1833 				cpu_to_be32(MTHCA_NEXT_DBD | size);
1834 		}
1835 
1836 		if (!size0)
1837 			size0 = size;
1838 
1839 		++ind;
1840 		if (unlikely(ind >= qp->rq.max))
1841 			ind -= qp->rq.max;
1842 	}
1843 
1844 out:
1845 	if (likely(nreq)) {
1846 		u32 doorbell[2];
1847 
1848 		doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1849 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1850 
1851 		wmb();
1852 
1853 		mthca_write64(doorbell,
1854 			      dev->kar + MTHCA_RECEIVE_DOORBELL,
1855 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1856 	}
1857 
1858 	qp->rq.next_ind = ind;
1859 	qp->rq.head    += nreq;
1860 
1861 	spin_unlock_irqrestore(&qp->rq.lock, flags);
1862 	return err;
1863 }
1864 
1865 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1866 			  struct ib_send_wr **bad_wr)
1867 {
1868 	struct mthca_dev *dev = to_mdev(ibqp->device);
1869 	struct mthca_qp *qp = to_mqp(ibqp);
1870 	void *wqe;
1871 	void *prev_wqe;
1872 	unsigned long flags;
1873 	int err = 0;
1874 	int nreq;
1875 	int i;
1876 	int size;
1877 	int size0 = 0;
1878 	u32 f0 = 0;
1879 	int ind;
1880 	u8 op0 = 0;
1881 
1882 	spin_lock_irqsave(&qp->sq.lock, flags);
1883 
1884 	/* XXX check that state is OK to post send */
1885 
1886 	ind = qp->sq.head & (qp->sq.max - 1);
1887 
1888 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1889 		if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1890 			mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1891 					" %d max, %d nreq)\n", qp->qpn,
1892 					qp->sq.head, qp->sq.tail,
1893 					qp->sq.max, nreq);
1894 			err = -ENOMEM;
1895 			*bad_wr = wr;
1896 			goto out;
1897 		}
1898 
1899 		wqe = get_send_wqe(qp, ind);
1900 		prev_wqe = qp->sq.last;
1901 		qp->sq.last = wqe;
1902 
1903 		((struct mthca_next_seg *) wqe)->flags =
1904 			((wr->send_flags & IB_SEND_SIGNALED) ?
1905 			 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1906 			((wr->send_flags & IB_SEND_SOLICITED) ?
1907 			 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0)   |
1908 			cpu_to_be32(1);
1909 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1910 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1911 			((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1912 
1913 		wqe += sizeof (struct mthca_next_seg);
1914 		size = sizeof (struct mthca_next_seg) / 16;
1915 
1916 		switch (qp->transport) {
1917 		case RC:
1918 			switch (wr->opcode) {
1919 			case IB_WR_ATOMIC_CMP_AND_SWP:
1920 			case IB_WR_ATOMIC_FETCH_AND_ADD:
1921 				((struct mthca_raddr_seg *) wqe)->raddr =
1922 					cpu_to_be64(wr->wr.atomic.remote_addr);
1923 				((struct mthca_raddr_seg *) wqe)->rkey =
1924 					cpu_to_be32(wr->wr.atomic.rkey);
1925 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1926 
1927 				wqe += sizeof (struct mthca_raddr_seg);
1928 
1929 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1930 					((struct mthca_atomic_seg *) wqe)->swap_add =
1931 						cpu_to_be64(wr->wr.atomic.swap);
1932 					((struct mthca_atomic_seg *) wqe)->compare =
1933 						cpu_to_be64(wr->wr.atomic.compare_add);
1934 				} else {
1935 					((struct mthca_atomic_seg *) wqe)->swap_add =
1936 						cpu_to_be64(wr->wr.atomic.compare_add);
1937 					((struct mthca_atomic_seg *) wqe)->compare = 0;
1938 				}
1939 
1940 				wqe += sizeof (struct mthca_atomic_seg);
1941 				size += sizeof (struct mthca_raddr_seg) / 16 +
1942 					sizeof (struct mthca_atomic_seg);
1943 				break;
1944 
1945 			case IB_WR_RDMA_READ:
1946 			case IB_WR_RDMA_WRITE:
1947 			case IB_WR_RDMA_WRITE_WITH_IMM:
1948 				((struct mthca_raddr_seg *) wqe)->raddr =
1949 					cpu_to_be64(wr->wr.rdma.remote_addr);
1950 				((struct mthca_raddr_seg *) wqe)->rkey =
1951 					cpu_to_be32(wr->wr.rdma.rkey);
1952 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1953 				wqe += sizeof (struct mthca_raddr_seg);
1954 				size += sizeof (struct mthca_raddr_seg) / 16;
1955 				break;
1956 
1957 			default:
1958 				/* No extra segments required for sends */
1959 				break;
1960 			}
1961 
1962 			break;
1963 
1964 		case UC:
1965 			switch (wr->opcode) {
1966 			case IB_WR_RDMA_WRITE:
1967 			case IB_WR_RDMA_WRITE_WITH_IMM:
1968 				((struct mthca_raddr_seg *) wqe)->raddr =
1969 					cpu_to_be64(wr->wr.rdma.remote_addr);
1970 				((struct mthca_raddr_seg *) wqe)->rkey =
1971 					cpu_to_be32(wr->wr.rdma.rkey);
1972 				((struct mthca_raddr_seg *) wqe)->reserved = 0;
1973 				wqe += sizeof (struct mthca_raddr_seg);
1974 				size += sizeof (struct mthca_raddr_seg) / 16;
1975 				break;
1976 
1977 			default:
1978 				/* No extra segments required for sends */
1979 				break;
1980 			}
1981 
1982 			break;
1983 
1984 		case UD:
1985 			memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1986 			       to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1987 			((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1988 				cpu_to_be32(wr->wr.ud.remote_qpn);
1989 			((struct mthca_arbel_ud_seg *) wqe)->qkey =
1990 				cpu_to_be32(wr->wr.ud.remote_qkey);
1991 
1992 			wqe += sizeof (struct mthca_arbel_ud_seg);
1993 			size += sizeof (struct mthca_arbel_ud_seg) / 16;
1994 			break;
1995 
1996 		case MLX:
1997 			err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1998 					       wqe - sizeof (struct mthca_next_seg),
1999 					       wqe);
2000 			if (err) {
2001 				*bad_wr = wr;
2002 				goto out;
2003 			}
2004 			wqe += sizeof (struct mthca_data_seg);
2005 			size += sizeof (struct mthca_data_seg) / 16;
2006 			break;
2007 		}
2008 
2009 		if (wr->num_sge > qp->sq.max_gs) {
2010 			mthca_err(dev, "too many gathers\n");
2011 			err = -EINVAL;
2012 			*bad_wr = wr;
2013 			goto out;
2014 		}
2015 
2016 		for (i = 0; i < wr->num_sge; ++i) {
2017 			((struct mthca_data_seg *) wqe)->byte_count =
2018 				cpu_to_be32(wr->sg_list[i].length);
2019 			((struct mthca_data_seg *) wqe)->lkey =
2020 				cpu_to_be32(wr->sg_list[i].lkey);
2021 			((struct mthca_data_seg *) wqe)->addr =
2022 				cpu_to_be64(wr->sg_list[i].addr);
2023 			wqe += sizeof (struct mthca_data_seg);
2024 			size += sizeof (struct mthca_data_seg) / 16;
2025 		}
2026 
2027 		/* Add one more inline data segment for ICRC */
2028 		if (qp->transport == MLX) {
2029 			((struct mthca_data_seg *) wqe)->byte_count =
2030 				cpu_to_be32((1 << 31) | 4);
2031 			((u32 *) wqe)[1] = 0;
2032 			wqe += sizeof (struct mthca_data_seg);
2033 			size += sizeof (struct mthca_data_seg) / 16;
2034 		}
2035 
2036 		qp->wrid[ind + qp->rq.max] = wr->wr_id;
2037 
2038 		if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2039 			mthca_err(dev, "opcode invalid\n");
2040 			err = -EINVAL;
2041 			*bad_wr = wr;
2042 			goto out;
2043 		}
2044 
2045 		if (likely(prev_wqe)) {
2046 			((struct mthca_next_seg *) prev_wqe)->nda_op =
2047 				cpu_to_be32(((ind << qp->sq.wqe_shift) +
2048 					     qp->send_wqe_offset) |
2049 					    mthca_opcode[wr->opcode]);
2050 			wmb();
2051 			((struct mthca_next_seg *) prev_wqe)->ee_nds =
2052 				cpu_to_be32(MTHCA_NEXT_DBD | size);
2053 		}
2054 
2055 		if (!size0) {
2056 			size0 = size;
2057 			op0   = mthca_opcode[wr->opcode];
2058 		}
2059 
2060 		++ind;
2061 		if (unlikely(ind >= qp->sq.max))
2062 			ind -= qp->sq.max;
2063 	}
2064 
2065 out:
2066 	if (likely(nreq)) {
2067 		u32 doorbell[2];
2068 
2069 		doorbell[0] = cpu_to_be32((nreq << 24)                  |
2070 					  ((qp->sq.head & 0xffff) << 8) |
2071 					  f0 | op0);
2072 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2073 
2074 		qp->sq.head += nreq;
2075 
2076 		/*
2077 		 * Make sure that descriptors are written before
2078 		 * doorbell record.
2079 		 */
2080 		wmb();
2081 		*qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2082 
2083 		/*
2084 		 * Make sure doorbell record is written before we
2085 		 * write MMIO send doorbell.
2086 		 */
2087 		wmb();
2088 		mthca_write64(doorbell,
2089 			      dev->kar + MTHCA_SEND_DOORBELL,
2090 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2091 	}
2092 
2093 	spin_unlock_irqrestore(&qp->sq.lock, flags);
2094 	return err;
2095 }
2096 
2097 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2098 			     struct ib_recv_wr **bad_wr)
2099 {
2100 	struct mthca_dev *dev = to_mdev(ibqp->device);
2101 	struct mthca_qp *qp = to_mqp(ibqp);
2102 	unsigned long flags;
2103 	int err = 0;
2104 	int nreq;
2105 	int ind;
2106 	int i;
2107 	void *wqe;
2108 
2109  	spin_lock_irqsave(&qp->rq.lock, flags);
2110 
2111 	/* XXX check that state is OK to post receive */
2112 
2113 	ind = qp->rq.head & (qp->rq.max - 1);
2114 
2115 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
2116 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2117 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2118 					" %d max, %d nreq)\n", qp->qpn,
2119 					qp->rq.head, qp->rq.tail,
2120 					qp->rq.max, nreq);
2121 			err = -ENOMEM;
2122 			*bad_wr = wr;
2123 			goto out;
2124 		}
2125 
2126 		wqe = get_recv_wqe(qp, ind);
2127 
2128 		((struct mthca_next_seg *) wqe)->flags = 0;
2129 
2130 		wqe += sizeof (struct mthca_next_seg);
2131 
2132 		if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2133 			err = -EINVAL;
2134 			*bad_wr = wr;
2135 			goto out;
2136 		}
2137 
2138 		for (i = 0; i < wr->num_sge; ++i) {
2139 			((struct mthca_data_seg *) wqe)->byte_count =
2140 				cpu_to_be32(wr->sg_list[i].length);
2141 			((struct mthca_data_seg *) wqe)->lkey =
2142 				cpu_to_be32(wr->sg_list[i].lkey);
2143 			((struct mthca_data_seg *) wqe)->addr =
2144 				cpu_to_be64(wr->sg_list[i].addr);
2145 			wqe += sizeof (struct mthca_data_seg);
2146 		}
2147 
2148 		if (i < qp->rq.max_gs) {
2149 			((struct mthca_data_seg *) wqe)->byte_count = 0;
2150 			((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2151 			((struct mthca_data_seg *) wqe)->addr = 0;
2152 		}
2153 
2154 		qp->wrid[ind] = wr->wr_id;
2155 
2156 		++ind;
2157 		if (unlikely(ind >= qp->rq.max))
2158 			ind -= qp->rq.max;
2159 	}
2160 out:
2161 	if (likely(nreq)) {
2162 		qp->rq.head += nreq;
2163 
2164 		/*
2165 		 * Make sure that descriptors are written before
2166 		 * doorbell record.
2167 		 */
2168 		wmb();
2169 		*qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2170 	}
2171 
2172 	spin_unlock_irqrestore(&qp->rq.lock, flags);
2173 	return err;
2174 }
2175 
2176 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2177 		       int index, int *dbd, u32 *new_wqe)
2178 {
2179 	struct mthca_next_seg *next;
2180 
2181 	if (is_send)
2182 		next = get_send_wqe(qp, index);
2183 	else
2184 		next = get_recv_wqe(qp, index);
2185 
2186 	if (mthca_is_memfree(dev))
2187 		*dbd = 1;
2188 	else
2189 		*dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2190 	if (next->ee_nds & cpu_to_be32(0x3f))
2191 		*new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2192 			(next->ee_nds & cpu_to_be32(0x3f));
2193 	else
2194 		*new_wqe = 0;
2195 
2196 	return 0;
2197 }
2198 
2199 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2200 {
2201 	int err;
2202 	u8 status;
2203 	int i;
2204 
2205 	spin_lock_init(&dev->qp_table.lock);
2206 
2207 	/*
2208 	 * We reserve 2 extra QPs per port for the special QPs.  The
2209 	 * special QP for port 1 has to be even, so round up.
2210 	 */
2211 	dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2212 	err = mthca_alloc_init(&dev->qp_table.alloc,
2213 			       dev->limits.num_qps,
2214 			       (1 << 24) - 1,
2215 			       dev->qp_table.sqp_start +
2216 			       MTHCA_MAX_PORTS * 2);
2217 	if (err)
2218 		return err;
2219 
2220 	err = mthca_array_init(&dev->qp_table.qp,
2221 			       dev->limits.num_qps);
2222 	if (err) {
2223 		mthca_alloc_cleanup(&dev->qp_table.alloc);
2224 		return err;
2225 	}
2226 
2227 	for (i = 0; i < 2; ++i) {
2228 		err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2229 					    dev->qp_table.sqp_start + i * 2,
2230 					    &status);
2231 		if (err)
2232 			goto err_out;
2233 		if (status) {
2234 			mthca_warn(dev, "CONF_SPECIAL_QP returned "
2235 				   "status %02x, aborting.\n",
2236 				   status);
2237 			err = -EINVAL;
2238 			goto err_out;
2239 		}
2240 	}
2241 	return 0;
2242 
2243  err_out:
2244 	for (i = 0; i < 2; ++i)
2245 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2246 
2247 	mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2248 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2249 
2250 	return err;
2251 }
2252 
2253 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2254 {
2255 	int i;
2256 	u8 status;
2257 
2258 	for (i = 0; i < 2; ++i)
2259 		mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2260 
2261 	mthca_alloc_cleanup(&dev->qp_table.alloc);
2262 }
2263