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.
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:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
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.
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
35 * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
38 #include <linux/init.h>
39 #include <linux/string.h>
40 #include <linux/slab.h>
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
46 #include "mthca_dev.h"
47 #include "mthca_cmd.h"
48 #include "mthca_memfree.h"
49 #include "mthca_wqe.h"
52 MTHCA_MAX_DIRECT_QP_SIZE
= 4 * PAGE_SIZE
,
53 MTHCA_ACK_REQ_FREQ
= 10,
54 MTHCA_FLIGHT_LIMIT
= 9,
55 MTHCA_UD_HEADER_SIZE
= 72, /* largest UD header possible */
56 MTHCA_INLINE_HEADER_SIZE
= 4, /* data segment overhead for inline */
57 MTHCA_INLINE_CHUNK_SIZE
= 16 /* inline data segment chunk */
61 MTHCA_QP_STATE_RST
= 0,
62 MTHCA_QP_STATE_INIT
= 1,
63 MTHCA_QP_STATE_RTR
= 2,
64 MTHCA_QP_STATE_RTS
= 3,
65 MTHCA_QP_STATE_SQE
= 4,
66 MTHCA_QP_STATE_SQD
= 5,
67 MTHCA_QP_STATE_ERR
= 6,
68 MTHCA_QP_STATE_DRAINING
= 7
80 MTHCA_QP_PM_MIGRATED
= 0x3,
81 MTHCA_QP_PM_ARMED
= 0x0,
82 MTHCA_QP_PM_REARM
= 0x1
86 /* qp_context flags */
87 MTHCA_QP_BIT_DE
= 1 << 8,
89 MTHCA_QP_BIT_SRE
= 1 << 15,
90 MTHCA_QP_BIT_SWE
= 1 << 14,
91 MTHCA_QP_BIT_SAE
= 1 << 13,
92 MTHCA_QP_BIT_SIC
= 1 << 4,
93 MTHCA_QP_BIT_SSC
= 1 << 3,
95 MTHCA_QP_BIT_RRE
= 1 << 15,
96 MTHCA_QP_BIT_RWE
= 1 << 14,
97 MTHCA_QP_BIT_RAE
= 1 << 13,
98 MTHCA_QP_BIT_RIC
= 1 << 4,
99 MTHCA_QP_BIT_RSC
= 1 << 3
102 struct mthca_qp_path
{
111 __be32 sl_tclass_flowlabel
;
113 } __attribute__((packed
));
115 struct mthca_qp_context
{
117 __be32 tavor_sched_queue
; /* Reserved on Arbel */
119 u8 rq_size_stride
; /* Reserved on Tavor */
120 u8 sq_size_stride
; /* Reserved on Tavor */
121 u8 rlkey_arbel_sched_queue
; /* Reserved on Tavor */
126 struct mthca_qp_path pri_path
;
127 struct mthca_qp_path alt_path
;
134 __be32 next_send_psn
;
136 __be32 snd_wqe_base_l
; /* Next send WQE on Tavor */
137 __be32 snd_db_index
; /* (debugging only entries) */
138 __be32 last_acked_psn
;
141 __be32 rnr_nextrecvpsn
;
144 __be32 rcv_wqe_base_l
; /* Next recv WQE on Tavor */
145 __be32 rcv_db_index
; /* (debugging only entries) */
149 __be16 rq_wqe_counter
; /* reserved on Tavor */
150 __be16 sq_wqe_counter
; /* reserved on Tavor */
152 } __attribute__((packed
));
154 struct mthca_qp_param
{
155 __be32 opt_param_mask
;
157 struct mthca_qp_context context
;
159 } __attribute__((packed
));
162 MTHCA_QP_OPTPAR_ALT_ADDR_PATH
= 1 << 0,
163 MTHCA_QP_OPTPAR_RRE
= 1 << 1,
164 MTHCA_QP_OPTPAR_RAE
= 1 << 2,
165 MTHCA_QP_OPTPAR_RWE
= 1 << 3,
166 MTHCA_QP_OPTPAR_PKEY_INDEX
= 1 << 4,
167 MTHCA_QP_OPTPAR_Q_KEY
= 1 << 5,
168 MTHCA_QP_OPTPAR_RNR_TIMEOUT
= 1 << 6,
169 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH
= 1 << 7,
170 MTHCA_QP_OPTPAR_SRA_MAX
= 1 << 8,
171 MTHCA_QP_OPTPAR_RRA_MAX
= 1 << 9,
172 MTHCA_QP_OPTPAR_PM_STATE
= 1 << 10,
173 MTHCA_QP_OPTPAR_PORT_NUM
= 1 << 11,
174 MTHCA_QP_OPTPAR_RETRY_COUNT
= 1 << 12,
175 MTHCA_QP_OPTPAR_ALT_RNR_RETRY
= 1 << 13,
176 MTHCA_QP_OPTPAR_ACK_TIMEOUT
= 1 << 14,
177 MTHCA_QP_OPTPAR_RNR_RETRY
= 1 << 15,
178 MTHCA_QP_OPTPAR_SCHED_QUEUE
= 1 << 16
181 static const u8 mthca_opcode
[] = {
182 [IB_WR_SEND
] = MTHCA_OPCODE_SEND
,
183 [IB_WR_SEND_WITH_IMM
] = MTHCA_OPCODE_SEND_IMM
,
184 [IB_WR_RDMA_WRITE
] = MTHCA_OPCODE_RDMA_WRITE
,
185 [IB_WR_RDMA_WRITE_WITH_IMM
] = MTHCA_OPCODE_RDMA_WRITE_IMM
,
186 [IB_WR_RDMA_READ
] = MTHCA_OPCODE_RDMA_READ
,
187 [IB_WR_ATOMIC_CMP_AND_SWP
] = MTHCA_OPCODE_ATOMIC_CS
,
188 [IB_WR_ATOMIC_FETCH_AND_ADD
] = MTHCA_OPCODE_ATOMIC_FA
,
191 static int is_sqp(struct mthca_dev
*dev
, struct mthca_qp
*qp
)
193 return qp
->qpn
>= dev
->qp_table
.sqp_start
&&
194 qp
->qpn
<= dev
->qp_table
.sqp_start
+ 3;
197 static int is_qp0(struct mthca_dev
*dev
, struct mthca_qp
*qp
)
199 return qp
->qpn
>= dev
->qp_table
.sqp_start
&&
200 qp
->qpn
<= dev
->qp_table
.sqp_start
+ 1;
203 static void *get_recv_wqe(struct mthca_qp
*qp
, int n
)
206 return qp
->queue
.direct
.buf
+ (n
<< qp
->rq
.wqe_shift
);
208 return qp
->queue
.page_list
[(n
<< qp
->rq
.wqe_shift
) >> PAGE_SHIFT
].buf
+
209 ((n
<< qp
->rq
.wqe_shift
) & (PAGE_SIZE
- 1));
212 static void *get_send_wqe(struct mthca_qp
*qp
, int n
)
215 return qp
->queue
.direct
.buf
+ qp
->send_wqe_offset
+
216 (n
<< qp
->sq
.wqe_shift
);
218 return qp
->queue
.page_list
[(qp
->send_wqe_offset
+
219 (n
<< qp
->sq
.wqe_shift
)) >>
221 ((qp
->send_wqe_offset
+ (n
<< qp
->sq
.wqe_shift
)) &
225 static void mthca_wq_init(struct mthca_wq
*wq
)
227 spin_lock_init(&wq
->lock
);
229 wq
->last_comp
= wq
->max
- 1;
234 void mthca_qp_event(struct mthca_dev
*dev
, u32 qpn
,
235 enum ib_event_type event_type
)
238 struct ib_event event
;
240 spin_lock(&dev
->qp_table
.lock
);
241 qp
= mthca_array_get(&dev
->qp_table
.qp
, qpn
& (dev
->limits
.num_qps
- 1));
243 atomic_inc(&qp
->refcount
);
244 spin_unlock(&dev
->qp_table
.lock
);
247 mthca_warn(dev
, "Async event for bogus QP %08x\n", qpn
);
251 event
.device
= &dev
->ib_dev
;
252 event
.event
= event_type
;
253 event
.element
.qp
= &qp
->ibqp
;
254 if (qp
->ibqp
.event_handler
)
255 qp
->ibqp
.event_handler(&event
, qp
->ibqp
.qp_context
);
257 if (atomic_dec_and_test(&qp
->refcount
))
261 static int to_mthca_state(enum ib_qp_state ib_state
)
264 case IB_QPS_RESET
: return MTHCA_QP_STATE_RST
;
265 case IB_QPS_INIT
: return MTHCA_QP_STATE_INIT
;
266 case IB_QPS_RTR
: return MTHCA_QP_STATE_RTR
;
267 case IB_QPS_RTS
: return MTHCA_QP_STATE_RTS
;
268 case IB_QPS_SQD
: return MTHCA_QP_STATE_SQD
;
269 case IB_QPS_SQE
: return MTHCA_QP_STATE_SQE
;
270 case IB_QPS_ERR
: return MTHCA_QP_STATE_ERR
;
275 enum { RC
, UC
, UD
, RD
, RDEE
, MLX
, NUM_TRANS
};
277 static int to_mthca_st(int transport
)
280 case RC
: return MTHCA_QP_ST_RC
;
281 case UC
: return MTHCA_QP_ST_UC
;
282 case UD
: return MTHCA_QP_ST_UD
;
283 case RD
: return MTHCA_QP_ST_RD
;
284 case MLX
: return MTHCA_QP_ST_MLX
;
289 static void store_attrs(struct mthca_sqp
*sqp
, struct ib_qp_attr
*attr
,
292 if (attr_mask
& IB_QP_PKEY_INDEX
)
293 sqp
->pkey_index
= attr
->pkey_index
;
294 if (attr_mask
& IB_QP_QKEY
)
295 sqp
->qkey
= attr
->qkey
;
296 if (attr_mask
& IB_QP_SQ_PSN
)
297 sqp
->send_psn
= attr
->sq_psn
;
300 static void init_port(struct mthca_dev
*dev
, int port
)
304 struct mthca_init_ib_param param
;
306 memset(¶m
, 0, sizeof param
);
308 param
.port_width
= dev
->limits
.port_width_cap
;
309 param
.vl_cap
= dev
->limits
.vl_cap
;
310 param
.mtu_cap
= dev
->limits
.mtu_cap
;
311 param
.gid_cap
= dev
->limits
.gid_table_len
;
312 param
.pkey_cap
= dev
->limits
.pkey_table_len
;
314 err
= mthca_INIT_IB(dev
, ¶m
, port
, &status
);
316 mthca_warn(dev
, "INIT_IB failed, return code %d.\n", err
);
318 mthca_warn(dev
, "INIT_IB returned status %02x.\n", status
);
321 static __be32
get_hw_access_flags(struct mthca_qp
*qp
, struct ib_qp_attr
*attr
,
326 u32 hw_access_flags
= 0;
328 if (attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
)
329 dest_rd_atomic
= attr
->max_dest_rd_atomic
;
331 dest_rd_atomic
= qp
->resp_depth
;
333 if (attr_mask
& IB_QP_ACCESS_FLAGS
)
334 access_flags
= attr
->qp_access_flags
;
336 access_flags
= qp
->atomic_rd_en
;
339 access_flags
&= IB_ACCESS_REMOTE_WRITE
;
341 if (access_flags
& IB_ACCESS_REMOTE_READ
)
342 hw_access_flags
|= MTHCA_QP_BIT_RRE
;
343 if (access_flags
& IB_ACCESS_REMOTE_ATOMIC
)
344 hw_access_flags
|= MTHCA_QP_BIT_RAE
;
345 if (access_flags
& IB_ACCESS_REMOTE_WRITE
)
346 hw_access_flags
|= MTHCA_QP_BIT_RWE
;
348 return cpu_to_be32(hw_access_flags
);
351 static inline enum ib_qp_state
to_ib_qp_state(int mthca_state
)
353 switch (mthca_state
) {
354 case MTHCA_QP_STATE_RST
: return IB_QPS_RESET
;
355 case MTHCA_QP_STATE_INIT
: return IB_QPS_INIT
;
356 case MTHCA_QP_STATE_RTR
: return IB_QPS_RTR
;
357 case MTHCA_QP_STATE_RTS
: return IB_QPS_RTS
;
358 case MTHCA_QP_STATE_DRAINING
:
359 case MTHCA_QP_STATE_SQD
: return IB_QPS_SQD
;
360 case MTHCA_QP_STATE_SQE
: return IB_QPS_SQE
;
361 case MTHCA_QP_STATE_ERR
: return IB_QPS_ERR
;
366 static inline enum ib_mig_state
to_ib_mig_state(int mthca_mig_state
)
368 switch (mthca_mig_state
) {
369 case 0: return IB_MIG_ARMED
;
370 case 1: return IB_MIG_REARM
;
371 case 3: return IB_MIG_MIGRATED
;
376 static int to_ib_qp_access_flags(int mthca_flags
)
380 if (mthca_flags
& MTHCA_QP_BIT_RRE
)
381 ib_flags
|= IB_ACCESS_REMOTE_READ
;
382 if (mthca_flags
& MTHCA_QP_BIT_RWE
)
383 ib_flags
|= IB_ACCESS_REMOTE_WRITE
;
384 if (mthca_flags
& MTHCA_QP_BIT_RAE
)
385 ib_flags
|= IB_ACCESS_REMOTE_ATOMIC
;
390 static void to_ib_ah_attr(struct mthca_dev
*dev
, struct ib_ah_attr
*ib_ah_attr
,
391 struct mthca_qp_path
*path
)
393 memset(ib_ah_attr
, 0, sizeof *path
);
394 ib_ah_attr
->port_num
= (be32_to_cpu(path
->port_pkey
) >> 24) & 0x3;
395 ib_ah_attr
->dlid
= be16_to_cpu(path
->rlid
);
396 ib_ah_attr
->sl
= be32_to_cpu(path
->sl_tclass_flowlabel
) >> 28;
397 ib_ah_attr
->src_path_bits
= path
->g_mylmc
& 0x7f;
398 ib_ah_attr
->static_rate
= path
->static_rate
& 0x7;
399 ib_ah_attr
->ah_flags
= (path
->g_mylmc
& (1 << 7)) ? IB_AH_GRH
: 0;
400 if (ib_ah_attr
->ah_flags
) {
401 ib_ah_attr
->grh
.sgid_index
= path
->mgid_index
& (dev
->limits
.gid_table_len
- 1);
402 ib_ah_attr
->grh
.hop_limit
= path
->hop_limit
;
403 ib_ah_attr
->grh
.traffic_class
=
404 (be32_to_cpu(path
->sl_tclass_flowlabel
) >> 20) & 0xff;
405 ib_ah_attr
->grh
.flow_label
=
406 be32_to_cpu(path
->sl_tclass_flowlabel
) & 0xfffff;
407 memcpy(ib_ah_attr
->grh
.dgid
.raw
,
408 path
->rgid
, sizeof ib_ah_attr
->grh
.dgid
.raw
);
412 int mthca_query_qp(struct ib_qp
*ibqp
, struct ib_qp_attr
*qp_attr
, int qp_attr_mask
,
413 struct ib_qp_init_attr
*qp_init_attr
)
415 struct mthca_dev
*dev
= to_mdev(ibqp
->device
);
416 struct mthca_qp
*qp
= to_mqp(ibqp
);
418 struct mthca_mailbox
*mailbox
;
419 struct mthca_qp_param
*qp_param
;
420 struct mthca_qp_context
*context
;
424 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
426 return PTR_ERR(mailbox
);
428 err
= mthca_QUERY_QP(dev
, qp
->qpn
, 0, mailbox
, &status
);
432 mthca_warn(dev
, "QUERY_QP returned status %02x\n", status
);
437 qp_param
= mailbox
->buf
;
438 context
= &qp_param
->context
;
439 mthca_state
= be32_to_cpu(context
->flags
) >> 28;
441 qp_attr
->qp_state
= to_ib_qp_state(mthca_state
);
442 qp_attr
->cur_qp_state
= qp_attr
->qp_state
;
443 qp_attr
->path_mtu
= context
->mtu_msgmax
>> 5;
444 qp_attr
->path_mig_state
=
445 to_ib_mig_state((be32_to_cpu(context
->flags
) >> 11) & 0x3);
446 qp_attr
->qkey
= be32_to_cpu(context
->qkey
);
447 qp_attr
->rq_psn
= be32_to_cpu(context
->rnr_nextrecvpsn
) & 0xffffff;
448 qp_attr
->sq_psn
= be32_to_cpu(context
->next_send_psn
) & 0xffffff;
449 qp_attr
->dest_qp_num
= be32_to_cpu(context
->remote_qpn
) & 0xffffff;
450 qp_attr
->qp_access_flags
=
451 to_ib_qp_access_flags(be32_to_cpu(context
->params2
));
452 qp_attr
->cap
.max_send_wr
= qp
->sq
.max
;
453 qp_attr
->cap
.max_recv_wr
= qp
->rq
.max
;
454 qp_attr
->cap
.max_send_sge
= qp
->sq
.max_gs
;
455 qp_attr
->cap
.max_recv_sge
= qp
->rq
.max_gs
;
456 qp_attr
->cap
.max_inline_data
= qp
->max_inline_data
;
458 to_ib_ah_attr(dev
, &qp_attr
->ah_attr
, &context
->pri_path
);
459 to_ib_ah_attr(dev
, &qp_attr
->alt_ah_attr
, &context
->alt_path
);
461 qp_attr
->pkey_index
= be32_to_cpu(context
->pri_path
.port_pkey
) & 0x7f;
462 qp_attr
->alt_pkey_index
= be32_to_cpu(context
->alt_path
.port_pkey
) & 0x7f;
464 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
465 qp_attr
->sq_draining
= mthca_state
== MTHCA_QP_STATE_DRAINING
;
467 qp_attr
->max_rd_atomic
= 1 << ((be32_to_cpu(context
->params1
) >> 21) & 0x7);
469 qp_attr
->max_dest_rd_atomic
=
470 1 << ((be32_to_cpu(context
->params2
) >> 21) & 0x7);
471 qp_attr
->min_rnr_timer
=
472 (be32_to_cpu(context
->rnr_nextrecvpsn
) >> 24) & 0x1f;
473 qp_attr
->port_num
= qp_attr
->ah_attr
.port_num
;
474 qp_attr
->timeout
= context
->pri_path
.ackto
>> 3;
475 qp_attr
->retry_cnt
= (be32_to_cpu(context
->params1
) >> 16) & 0x7;
476 qp_attr
->rnr_retry
= context
->pri_path
.rnr_retry
>> 5;
477 qp_attr
->alt_port_num
= qp_attr
->alt_ah_attr
.port_num
;
478 qp_attr
->alt_timeout
= context
->alt_path
.ackto
>> 3;
479 qp_init_attr
->cap
= qp_attr
->cap
;
482 mthca_free_mailbox(dev
, mailbox
);
486 static int mthca_path_set(struct mthca_dev
*dev
, struct ib_ah_attr
*ah
,
487 struct mthca_qp_path
*path
)
489 path
->g_mylmc
= ah
->src_path_bits
& 0x7f;
490 path
->rlid
= cpu_to_be16(ah
->dlid
);
491 path
->static_rate
= !!ah
->static_rate
;
493 if (ah
->ah_flags
& IB_AH_GRH
) {
494 if (ah
->grh
.sgid_index
>= dev
->limits
.gid_table_len
) {
495 mthca_dbg(dev
, "sgid_index (%u) too large. max is %d\n",
496 ah
->grh
.sgid_index
, dev
->limits
.gid_table_len
-1);
500 path
->g_mylmc
|= 1 << 7;
501 path
->mgid_index
= ah
->grh
.sgid_index
;
502 path
->hop_limit
= ah
->grh
.hop_limit
;
503 path
->sl_tclass_flowlabel
=
504 cpu_to_be32((ah
->sl
<< 28) |
505 (ah
->grh
.traffic_class
<< 20) |
506 (ah
->grh
.flow_label
));
507 memcpy(path
->rgid
, ah
->grh
.dgid
.raw
, 16);
509 path
->sl_tclass_flowlabel
= cpu_to_be32(ah
->sl
<< 28);
514 int mthca_modify_qp(struct ib_qp
*ibqp
, struct ib_qp_attr
*attr
, int attr_mask
)
516 struct mthca_dev
*dev
= to_mdev(ibqp
->device
);
517 struct mthca_qp
*qp
= to_mqp(ibqp
);
518 enum ib_qp_state cur_state
, new_state
;
519 struct mthca_mailbox
*mailbox
;
520 struct mthca_qp_param
*qp_param
;
521 struct mthca_qp_context
*qp_context
;
526 if (attr_mask
& IB_QP_CUR_STATE
) {
527 cur_state
= attr
->cur_qp_state
;
529 spin_lock_irq(&qp
->sq
.lock
);
530 spin_lock(&qp
->rq
.lock
);
531 cur_state
= qp
->state
;
532 spin_unlock(&qp
->rq
.lock
);
533 spin_unlock_irq(&qp
->sq
.lock
);
536 new_state
= attr_mask
& IB_QP_STATE
? attr
->qp_state
: cur_state
;
538 if (!ib_modify_qp_is_ok(cur_state
, new_state
, ibqp
->qp_type
, attr_mask
)) {
539 mthca_dbg(dev
, "Bad QP transition (transport %d) "
540 "%d->%d with attr 0x%08x\n",
541 qp
->transport
, cur_state
, new_state
,
546 if ((attr_mask
& IB_QP_PKEY_INDEX
) &&
547 attr
->pkey_index
>= dev
->limits
.pkey_table_len
) {
548 mthca_dbg(dev
, "P_Key index (%u) too large. max is %d\n",
549 attr
->pkey_index
, dev
->limits
.pkey_table_len
-1);
553 if ((attr_mask
& IB_QP_PORT
) &&
554 (attr
->port_num
== 0 || attr
->port_num
> dev
->limits
.num_ports
)) {
555 mthca_dbg(dev
, "Port number (%u) is invalid\n", attr
->port_num
);
559 if (attr_mask
& IB_QP_MAX_QP_RD_ATOMIC
&&
560 attr
->max_rd_atomic
> dev
->limits
.max_qp_init_rdma
) {
561 mthca_dbg(dev
, "Max rdma_atomic as initiator %u too large (max is %d)\n",
562 attr
->max_rd_atomic
, dev
->limits
.max_qp_init_rdma
);
566 if (attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
&&
567 attr
->max_dest_rd_atomic
> 1 << dev
->qp_table
.rdb_shift
) {
568 mthca_dbg(dev
, "Max rdma_atomic as responder %u too large (max %d)\n",
569 attr
->max_dest_rd_atomic
, 1 << dev
->qp_table
.rdb_shift
);
573 mailbox
= mthca_alloc_mailbox(dev
, GFP_KERNEL
);
575 return PTR_ERR(mailbox
);
576 qp_param
= mailbox
->buf
;
577 qp_context
= &qp_param
->context
;
578 memset(qp_param
, 0, sizeof *qp_param
);
580 qp_context
->flags
= cpu_to_be32((to_mthca_state(new_state
) << 28) |
581 (to_mthca_st(qp
->transport
) << 16));
582 qp_context
->flags
|= cpu_to_be32(MTHCA_QP_BIT_DE
);
583 if (!(attr_mask
& IB_QP_PATH_MIG_STATE
))
584 qp_context
->flags
|= cpu_to_be32(MTHCA_QP_PM_MIGRATED
<< 11);
586 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE
);
587 switch (attr
->path_mig_state
) {
588 case IB_MIG_MIGRATED
:
589 qp_context
->flags
|= cpu_to_be32(MTHCA_QP_PM_MIGRATED
<< 11);
592 qp_context
->flags
|= cpu_to_be32(MTHCA_QP_PM_REARM
<< 11);
595 qp_context
->flags
|= cpu_to_be32(MTHCA_QP_PM_ARMED
<< 11);
600 /* leave tavor_sched_queue as 0 */
602 if (qp
->transport
== MLX
|| qp
->transport
== UD
)
603 qp_context
->mtu_msgmax
= (IB_MTU_2048
<< 5) | 11;
604 else if (attr_mask
& IB_QP_PATH_MTU
) {
605 if (attr
->path_mtu
< IB_MTU_256
|| attr
->path_mtu
> IB_MTU_2048
) {
606 mthca_dbg(dev
, "path MTU (%u) is invalid\n",
610 qp_context
->mtu_msgmax
= (attr
->path_mtu
<< 5) | 31;
613 if (mthca_is_memfree(dev
)) {
615 qp_context
->rq_size_stride
= long_log2(qp
->rq
.max
) << 3;
616 qp_context
->rq_size_stride
|= qp
->rq
.wqe_shift
- 4;
619 qp_context
->sq_size_stride
= long_log2(qp
->sq
.max
) << 3;
620 qp_context
->sq_size_stride
|= qp
->sq
.wqe_shift
- 4;
623 /* leave arbel_sched_queue as 0 */
625 if (qp
->ibqp
.uobject
)
626 qp_context
->usr_page
=
627 cpu_to_be32(to_mucontext(qp
->ibqp
.uobject
->context
)->uar
.index
);
629 qp_context
->usr_page
= cpu_to_be32(dev
->driver_uar
.index
);
630 qp_context
->local_qpn
= cpu_to_be32(qp
->qpn
);
631 if (attr_mask
& IB_QP_DEST_QPN
) {
632 qp_context
->remote_qpn
= cpu_to_be32(attr
->dest_qp_num
);
635 if (qp
->transport
== MLX
)
636 qp_context
->pri_path
.port_pkey
|=
637 cpu_to_be32(to_msqp(qp
)->port
<< 24);
639 if (attr_mask
& IB_QP_PORT
) {
640 qp_context
->pri_path
.port_pkey
|=
641 cpu_to_be32(attr
->port_num
<< 24);
642 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM
);
646 if (attr_mask
& IB_QP_PKEY_INDEX
) {
647 qp_context
->pri_path
.port_pkey
|=
648 cpu_to_be32(attr
->pkey_index
);
649 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX
);
652 if (attr_mask
& IB_QP_RNR_RETRY
) {
653 qp_context
->alt_path
.rnr_retry
= qp_context
->pri_path
.rnr_retry
=
654 attr
->rnr_retry
<< 5;
655 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY
|
656 MTHCA_QP_OPTPAR_ALT_RNR_RETRY
);
659 if (attr_mask
& IB_QP_AV
) {
660 if (mthca_path_set(dev
, &attr
->ah_attr
, &qp_context
->pri_path
))
663 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH
);
666 if (attr_mask
& IB_QP_TIMEOUT
) {
667 qp_context
->pri_path
.ackto
= attr
->timeout
<< 3;
668 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT
);
671 if (attr_mask
& IB_QP_ALT_PATH
) {
672 if (attr
->alt_pkey_index
>= dev
->limits
.pkey_table_len
) {
673 mthca_dbg(dev
, "Alternate P_Key index (%u) too large. max is %d\n",
674 attr
->alt_pkey_index
, dev
->limits
.pkey_table_len
-1);
678 if (attr
->alt_port_num
== 0 || attr
->alt_port_num
> dev
->limits
.num_ports
) {
679 mthca_dbg(dev
, "Alternate port number (%u) is invalid\n",
684 if (mthca_path_set(dev
, &attr
->alt_ah_attr
, &qp_context
->alt_path
))
687 qp_context
->alt_path
.port_pkey
|= cpu_to_be32(attr
->alt_pkey_index
|
688 attr
->alt_port_num
<< 24);
689 qp_context
->alt_path
.ackto
= attr
->alt_timeout
<< 3;
690 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH
);
694 qp_context
->pd
= cpu_to_be32(to_mpd(ibqp
->pd
)->pd_num
);
695 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
696 qp_context
->wqe_lkey
= cpu_to_be32(qp
->mr
.ibmr
.lkey
);
697 qp_context
->params1
= cpu_to_be32((MTHCA_ACK_REQ_FREQ
<< 28) |
698 (MTHCA_FLIGHT_LIMIT
<< 24) |
700 if (qp
->sq_policy
== IB_SIGNAL_ALL_WR
)
701 qp_context
->params1
|= cpu_to_be32(MTHCA_QP_BIT_SSC
);
702 if (attr_mask
& IB_QP_RETRY_CNT
) {
703 qp_context
->params1
|= cpu_to_be32(attr
->retry_cnt
<< 16);
704 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT
);
707 if (attr_mask
& IB_QP_MAX_QP_RD_ATOMIC
) {
708 if (attr
->max_rd_atomic
) {
709 qp_context
->params1
|=
710 cpu_to_be32(MTHCA_QP_BIT_SRE
|
712 qp_context
->params1
|=
713 cpu_to_be32(fls(attr
->max_rd_atomic
- 1) << 21);
715 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX
);
718 if (attr_mask
& IB_QP_SQ_PSN
)
719 qp_context
->next_send_psn
= cpu_to_be32(attr
->sq_psn
);
720 qp_context
->cqn_snd
= cpu_to_be32(to_mcq(ibqp
->send_cq
)->cqn
);
722 if (mthca_is_memfree(dev
)) {
723 qp_context
->snd_wqe_base_l
= cpu_to_be32(qp
->send_wqe_offset
);
724 qp_context
->snd_db_index
= cpu_to_be32(qp
->sq
.db_index
);
727 if (attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
) {
728 if (attr
->max_dest_rd_atomic
)
729 qp_context
->params2
|=
730 cpu_to_be32(fls(attr
->max_dest_rd_atomic
- 1) << 21);
732 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX
);
735 if (attr_mask
& (IB_QP_ACCESS_FLAGS
| IB_QP_MAX_DEST_RD_ATOMIC
)) {
736 qp_context
->params2
|= get_hw_access_flags(qp
, attr
, attr_mask
);
737 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_RWE
|
738 MTHCA_QP_OPTPAR_RRE
|
739 MTHCA_QP_OPTPAR_RAE
);
742 qp_context
->params2
|= cpu_to_be32(MTHCA_QP_BIT_RSC
);
745 qp_context
->params2
|= cpu_to_be32(MTHCA_QP_BIT_RIC
);
747 if (attr_mask
& IB_QP_MIN_RNR_TIMER
) {
748 qp_context
->rnr_nextrecvpsn
|= cpu_to_be32(attr
->min_rnr_timer
<< 24);
749 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT
);
751 if (attr_mask
& IB_QP_RQ_PSN
)
752 qp_context
->rnr_nextrecvpsn
|= cpu_to_be32(attr
->rq_psn
);
754 qp_context
->ra_buff_indx
=
755 cpu_to_be32(dev
->qp_table
.rdb_base
+
756 ((qp
->qpn
& (dev
->limits
.num_qps
- 1)) * MTHCA_RDB_ENTRY_SIZE
<<
757 dev
->qp_table
.rdb_shift
));
759 qp_context
->cqn_rcv
= cpu_to_be32(to_mcq(ibqp
->recv_cq
)->cqn
);
761 if (mthca_is_memfree(dev
))
762 qp_context
->rcv_db_index
= cpu_to_be32(qp
->rq
.db_index
);
764 if (attr_mask
& IB_QP_QKEY
) {
765 qp_context
->qkey
= cpu_to_be32(attr
->qkey
);
766 qp_param
->opt_param_mask
|= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY
);
770 qp_context
->srqn
= cpu_to_be32(1 << 24 |
771 to_msrq(ibqp
->srq
)->srqn
);
773 if (cur_state
== IB_QPS_RTS
&& new_state
== IB_QPS_SQD
&&
774 attr_mask
& IB_QP_EN_SQD_ASYNC_NOTIFY
&&
775 attr
->en_sqd_async_notify
)
778 err
= mthca_MODIFY_QP(dev
, cur_state
, new_state
, qp
->qpn
, 0,
779 mailbox
, sqd_event
, &status
);
783 mthca_warn(dev
, "modify QP %d->%d returned status %02x.\n",
784 cur_state
, new_state
, status
);
789 qp
->state
= new_state
;
790 if (attr_mask
& IB_QP_ACCESS_FLAGS
)
791 qp
->atomic_rd_en
= attr
->qp_access_flags
;
792 if (attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
)
793 qp
->resp_depth
= attr
->max_dest_rd_atomic
;
796 store_attrs(to_msqp(qp
), attr
, attr_mask
);
799 * If we moved QP0 to RTR, bring the IB link up; if we moved
800 * QP0 to RESET or ERROR, bring the link back down.
802 if (is_qp0(dev
, qp
)) {
803 if (cur_state
!= IB_QPS_RTR
&&
804 new_state
== IB_QPS_RTR
)
805 init_port(dev
, to_msqp(qp
)->port
);
807 if (cur_state
!= IB_QPS_RESET
&&
808 cur_state
!= IB_QPS_ERR
&&
809 (new_state
== IB_QPS_RESET
||
810 new_state
== IB_QPS_ERR
))
811 mthca_CLOSE_IB(dev
, to_msqp(qp
)->port
, &status
);
815 * If we moved a kernel QP to RESET, clean up all old CQ
816 * entries and reinitialize the QP.
818 if (new_state
== IB_QPS_RESET
&& !qp
->ibqp
.uobject
) {
819 mthca_cq_clean(dev
, to_mcq(qp
->ibqp
.send_cq
)->cqn
, qp
->qpn
,
820 qp
->ibqp
.srq
? to_msrq(qp
->ibqp
.srq
) : NULL
);
821 if (qp
->ibqp
.send_cq
!= qp
->ibqp
.recv_cq
)
822 mthca_cq_clean(dev
, to_mcq(qp
->ibqp
.recv_cq
)->cqn
, qp
->qpn
,
823 qp
->ibqp
.srq
? to_msrq(qp
->ibqp
.srq
) : NULL
);
825 mthca_wq_init(&qp
->sq
);
826 qp
->sq
.last
= get_send_wqe(qp
, qp
->sq
.max
- 1);
828 mthca_wq_init(&qp
->rq
);
829 qp
->rq
.last
= get_recv_wqe(qp
, qp
->rq
.max
- 1);
831 if (mthca_is_memfree(dev
)) {
838 mthca_free_mailbox(dev
, mailbox
);
842 static int mthca_max_data_size(struct mthca_dev
*dev
, struct mthca_qp
*qp
, int desc_sz
)
845 * Calculate the maximum size of WQE s/g segments, excluding
846 * the next segment and other non-data segments.
848 int max_data_size
= desc_sz
- sizeof (struct mthca_next_seg
);
850 switch (qp
->transport
) {
852 max_data_size
-= 2 * sizeof (struct mthca_data_seg
);
856 if (mthca_is_memfree(dev
))
857 max_data_size
-= sizeof (struct mthca_arbel_ud_seg
);
859 max_data_size
-= sizeof (struct mthca_tavor_ud_seg
);
863 max_data_size
-= sizeof (struct mthca_raddr_seg
);
867 return max_data_size
;
870 static inline int mthca_max_inline_data(struct mthca_pd
*pd
, int max_data_size
)
872 /* We don't support inline data for kernel QPs (yet). */
873 return pd
->ibpd
.uobject
? max_data_size
- MTHCA_INLINE_HEADER_SIZE
: 0;
876 static void mthca_adjust_qp_caps(struct mthca_dev
*dev
,
880 int max_data_size
= mthca_max_data_size(dev
, qp
,
881 min(dev
->limits
.max_desc_sz
,
882 1 << qp
->sq
.wqe_shift
));
884 qp
->max_inline_data
= mthca_max_inline_data(pd
, max_data_size
);
886 qp
->sq
.max_gs
= min_t(int, dev
->limits
.max_sg
,
887 max_data_size
/ sizeof (struct mthca_data_seg
));
888 qp
->rq
.max_gs
= min_t(int, dev
->limits
.max_sg
,
889 (min(dev
->limits
.max_desc_sz
, 1 << qp
->rq
.wqe_shift
) -
890 sizeof (struct mthca_next_seg
)) /
891 sizeof (struct mthca_data_seg
));
895 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
896 * rq.max_gs and sq.max_gs must all be assigned.
897 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
898 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
901 static int mthca_alloc_wqe_buf(struct mthca_dev
*dev
,
908 size
= sizeof (struct mthca_next_seg
) +
909 qp
->rq
.max_gs
* sizeof (struct mthca_data_seg
);
911 if (size
> dev
->limits
.max_desc_sz
)
914 for (qp
->rq
.wqe_shift
= 6; 1 << qp
->rq
.wqe_shift
< size
;
918 size
= qp
->sq
.max_gs
* sizeof (struct mthca_data_seg
);
919 switch (qp
->transport
) {
921 size
+= 2 * sizeof (struct mthca_data_seg
);
925 size
+= mthca_is_memfree(dev
) ?
926 sizeof (struct mthca_arbel_ud_seg
) :
927 sizeof (struct mthca_tavor_ud_seg
);
931 size
+= sizeof (struct mthca_raddr_seg
);
935 size
+= sizeof (struct mthca_raddr_seg
);
937 * An atomic op will require an atomic segment, a
938 * remote address segment and one scatter entry.
940 size
= max_t(int, size
,
941 sizeof (struct mthca_atomic_seg
) +
942 sizeof (struct mthca_raddr_seg
) +
943 sizeof (struct mthca_data_seg
));
950 /* Make sure that we have enough space for a bind request */
951 size
= max_t(int, size
, sizeof (struct mthca_bind_seg
));
953 size
+= sizeof (struct mthca_next_seg
);
955 if (size
> dev
->limits
.max_desc_sz
)
958 for (qp
->sq
.wqe_shift
= 6; 1 << qp
->sq
.wqe_shift
< size
;
962 qp
->send_wqe_offset
= ALIGN(qp
->rq
.max
<< qp
->rq
.wqe_shift
,
963 1 << qp
->sq
.wqe_shift
);
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.
970 if (pd
->ibpd
.uobject
)
973 size
= PAGE_ALIGN(qp
->send_wqe_offset
+
974 (qp
->sq
.max
<< qp
->sq
.wqe_shift
));
976 qp
->wrid
= kmalloc((qp
->rq
.max
+ qp
->sq
.max
) * sizeof (u64
),
981 err
= mthca_buf_alloc(dev
, size
, MTHCA_MAX_DIRECT_QP_SIZE
,
982 &qp
->queue
, &qp
->is_direct
, pd
, 0, &qp
->mr
);
993 static void mthca_free_wqe_buf(struct mthca_dev
*dev
,
996 mthca_buf_free(dev
, PAGE_ALIGN(qp
->send_wqe_offset
+
997 (qp
->sq
.max
<< qp
->sq
.wqe_shift
)),
998 &qp
->queue
, qp
->is_direct
, &qp
->mr
);
1002 static int mthca_map_memfree(struct mthca_dev
*dev
,
1003 struct mthca_qp
*qp
)
1007 if (mthca_is_memfree(dev
)) {
1008 ret
= mthca_table_get(dev
, dev
->qp_table
.qp_table
, qp
->qpn
);
1012 ret
= mthca_table_get(dev
, dev
->qp_table
.eqp_table
, qp
->qpn
);
1016 ret
= mthca_table_get(dev
, dev
->qp_table
.rdb_table
,
1017 qp
->qpn
<< dev
->qp_table
.rdb_shift
);
1026 mthca_table_put(dev
, dev
->qp_table
.eqp_table
, qp
->qpn
);
1029 mthca_table_put(dev
, dev
->qp_table
.qp_table
, qp
->qpn
);
1034 static void mthca_unmap_memfree(struct mthca_dev
*dev
,
1035 struct mthca_qp
*qp
)
1037 mthca_table_put(dev
, dev
->qp_table
.rdb_table
,
1038 qp
->qpn
<< dev
->qp_table
.rdb_shift
);
1039 mthca_table_put(dev
, dev
->qp_table
.eqp_table
, qp
->qpn
);
1040 mthca_table_put(dev
, dev
->qp_table
.qp_table
, qp
->qpn
);
1043 static int mthca_alloc_memfree(struct mthca_dev
*dev
,
1044 struct mthca_qp
*qp
)
1048 if (mthca_is_memfree(dev
)) {
1049 qp
->rq
.db_index
= mthca_alloc_db(dev
, MTHCA_DB_TYPE_RQ
,
1050 qp
->qpn
, &qp
->rq
.db
);
1051 if (qp
->rq
.db_index
< 0)
1054 qp
->sq
.db_index
= mthca_alloc_db(dev
, MTHCA_DB_TYPE_SQ
,
1055 qp
->qpn
, &qp
->sq
.db
);
1056 if (qp
->sq
.db_index
< 0)
1057 mthca_free_db(dev
, MTHCA_DB_TYPE_RQ
, qp
->rq
.db_index
);
1063 static void mthca_free_memfree(struct mthca_dev
*dev
,
1064 struct mthca_qp
*qp
)
1066 if (mthca_is_memfree(dev
)) {
1067 mthca_free_db(dev
, MTHCA_DB_TYPE_SQ
, qp
->sq
.db_index
);
1068 mthca_free_db(dev
, MTHCA_DB_TYPE_RQ
, qp
->rq
.db_index
);
1072 static int mthca_alloc_qp_common(struct mthca_dev
*dev
,
1073 struct mthca_pd
*pd
,
1074 struct mthca_cq
*send_cq
,
1075 struct mthca_cq
*recv_cq
,
1076 enum ib_sig_type send_policy
,
1077 struct mthca_qp
*qp
)
1082 atomic_set(&qp
->refcount
, 1);
1083 init_waitqueue_head(&qp
->wait
);
1084 qp
->state
= IB_QPS_RESET
;
1085 qp
->atomic_rd_en
= 0;
1087 qp
->sq_policy
= send_policy
;
1088 mthca_wq_init(&qp
->sq
);
1089 mthca_wq_init(&qp
->rq
);
1091 ret
= mthca_map_memfree(dev
, qp
);
1095 ret
= mthca_alloc_wqe_buf(dev
, pd
, qp
);
1097 mthca_unmap_memfree(dev
, qp
);
1101 mthca_adjust_qp_caps(dev
, pd
, qp
);
1104 * If this is a userspace QP, we're done now. The doorbells
1105 * will be allocated and buffers will be initialized in
1108 if (pd
->ibpd
.uobject
)
1111 ret
= mthca_alloc_memfree(dev
, qp
);
1113 mthca_free_wqe_buf(dev
, qp
);
1114 mthca_unmap_memfree(dev
, qp
);
1118 if (mthca_is_memfree(dev
)) {
1119 struct mthca_next_seg
*next
;
1120 struct mthca_data_seg
*scatter
;
1121 int size
= (sizeof (struct mthca_next_seg
) +
1122 qp
->rq
.max_gs
* sizeof (struct mthca_data_seg
)) / 16;
1124 for (i
= 0; i
< qp
->rq
.max
; ++i
) {
1125 next
= get_recv_wqe(qp
, i
);
1126 next
->nda_op
= cpu_to_be32(((i
+ 1) & (qp
->rq
.max
- 1)) <<
1128 next
->ee_nds
= cpu_to_be32(size
);
1130 for (scatter
= (void *) (next
+ 1);
1131 (void *) scatter
< (void *) next
+ (1 << qp
->rq
.wqe_shift
);
1133 scatter
->lkey
= cpu_to_be32(MTHCA_INVAL_LKEY
);
1136 for (i
= 0; i
< qp
->sq
.max
; ++i
) {
1137 next
= get_send_wqe(qp
, i
);
1138 next
->nda_op
= cpu_to_be32((((i
+ 1) & (qp
->sq
.max
- 1)) <<
1140 qp
->send_wqe_offset
);
1144 qp
->sq
.last
= get_send_wqe(qp
, qp
->sq
.max
- 1);
1145 qp
->rq
.last
= get_recv_wqe(qp
, qp
->rq
.max
- 1);
1150 static int mthca_set_qp_size(struct mthca_dev
*dev
, struct ib_qp_cap
*cap
,
1151 struct mthca_pd
*pd
, struct mthca_qp
*qp
)
1153 int max_data_size
= mthca_max_data_size(dev
, qp
, dev
->limits
.max_desc_sz
);
1155 /* Sanity check QP size before proceeding */
1156 if (cap
->max_send_wr
> dev
->limits
.max_wqes
||
1157 cap
->max_recv_wr
> dev
->limits
.max_wqes
||
1158 cap
->max_send_sge
> dev
->limits
.max_sg
||
1159 cap
->max_recv_sge
> dev
->limits
.max_sg
||
1160 cap
->max_inline_data
> mthca_max_inline_data(pd
, max_data_size
))
1164 * For MLX transport we need 2 extra S/G entries:
1165 * one for the header and one for the checksum at the end
1167 if (qp
->transport
== MLX
&& cap
->max_recv_sge
+ 2 > dev
->limits
.max_sg
)
1170 if (mthca_is_memfree(dev
)) {
1171 qp
->rq
.max
= cap
->max_recv_wr
?
1172 roundup_pow_of_two(cap
->max_recv_wr
) : 0;
1173 qp
->sq
.max
= cap
->max_send_wr
?
1174 roundup_pow_of_two(cap
->max_send_wr
) : 0;
1176 qp
->rq
.max
= cap
->max_recv_wr
;
1177 qp
->sq
.max
= cap
->max_send_wr
;
1180 qp
->rq
.max_gs
= cap
->max_recv_sge
;
1181 qp
->sq
.max_gs
= max_t(int, cap
->max_send_sge
,
1182 ALIGN(cap
->max_inline_data
+ MTHCA_INLINE_HEADER_SIZE
,
1183 MTHCA_INLINE_CHUNK_SIZE
) /
1184 sizeof (struct mthca_data_seg
));
1189 int mthca_alloc_qp(struct mthca_dev
*dev
,
1190 struct mthca_pd
*pd
,
1191 struct mthca_cq
*send_cq
,
1192 struct mthca_cq
*recv_cq
,
1193 enum ib_qp_type type
,
1194 enum ib_sig_type send_policy
,
1195 struct ib_qp_cap
*cap
,
1196 struct mthca_qp
*qp
)
1201 case IB_QPT_RC
: qp
->transport
= RC
; break;
1202 case IB_QPT_UC
: qp
->transport
= UC
; break;
1203 case IB_QPT_UD
: qp
->transport
= UD
; break;
1204 default: return -EINVAL
;
1207 err
= mthca_set_qp_size(dev
, cap
, pd
, qp
);
1211 qp
->qpn
= mthca_alloc(&dev
->qp_table
.alloc
);
1215 err
= mthca_alloc_qp_common(dev
, pd
, send_cq
, recv_cq
,
1218 mthca_free(&dev
->qp_table
.alloc
, qp
->qpn
);
1222 spin_lock_irq(&dev
->qp_table
.lock
);
1223 mthca_array_set(&dev
->qp_table
.qp
,
1224 qp
->qpn
& (dev
->limits
.num_qps
- 1), qp
);
1225 spin_unlock_irq(&dev
->qp_table
.lock
);
1230 int mthca_alloc_sqp(struct mthca_dev
*dev
,
1231 struct mthca_pd
*pd
,
1232 struct mthca_cq
*send_cq
,
1233 struct mthca_cq
*recv_cq
,
1234 enum ib_sig_type send_policy
,
1235 struct ib_qp_cap
*cap
,
1238 struct mthca_sqp
*sqp
)
1240 u32 mqpn
= qpn
* 2 + dev
->qp_table
.sqp_start
+ port
- 1;
1243 sqp
->qp
.transport
= MLX
;
1244 err
= mthca_set_qp_size(dev
, cap
, pd
, &sqp
->qp
);
1248 sqp
->header_buf_size
= sqp
->qp
.sq
.max
* MTHCA_UD_HEADER_SIZE
;
1249 sqp
->header_buf
= dma_alloc_coherent(&dev
->pdev
->dev
, sqp
->header_buf_size
,
1250 &sqp
->header_dma
, GFP_KERNEL
);
1251 if (!sqp
->header_buf
)
1254 spin_lock_irq(&dev
->qp_table
.lock
);
1255 if (mthca_array_get(&dev
->qp_table
.qp
, mqpn
))
1258 mthca_array_set(&dev
->qp_table
.qp
, mqpn
, sqp
);
1259 spin_unlock_irq(&dev
->qp_table
.lock
);
1266 sqp
->qp
.transport
= MLX
;
1268 err
= mthca_alloc_qp_common(dev
, pd
, send_cq
, recv_cq
,
1269 send_policy
, &sqp
->qp
);
1273 atomic_inc(&pd
->sqp_count
);
1279 * Lock CQs here, so that CQ polling code can do QP lookup
1280 * without taking a lock.
1282 spin_lock_irq(&send_cq
->lock
);
1283 if (send_cq
!= recv_cq
)
1284 spin_lock(&recv_cq
->lock
);
1286 spin_lock(&dev
->qp_table
.lock
);
1287 mthca_array_clear(&dev
->qp_table
.qp
, mqpn
);
1288 spin_unlock(&dev
->qp_table
.lock
);
1290 if (send_cq
!= recv_cq
)
1291 spin_unlock(&recv_cq
->lock
);
1292 spin_unlock_irq(&send_cq
->lock
);
1295 dma_free_coherent(&dev
->pdev
->dev
, sqp
->header_buf_size
,
1296 sqp
->header_buf
, sqp
->header_dma
);
1301 void mthca_free_qp(struct mthca_dev
*dev
,
1302 struct mthca_qp
*qp
)
1305 struct mthca_cq
*send_cq
;
1306 struct mthca_cq
*recv_cq
;
1308 send_cq
= to_mcq(qp
->ibqp
.send_cq
);
1309 recv_cq
= to_mcq(qp
->ibqp
.recv_cq
);
1312 * Lock CQs here, so that CQ polling code can do QP lookup
1313 * without taking a lock.
1315 spin_lock_irq(&send_cq
->lock
);
1316 if (send_cq
!= recv_cq
)
1317 spin_lock(&recv_cq
->lock
);
1319 spin_lock(&dev
->qp_table
.lock
);
1320 mthca_array_clear(&dev
->qp_table
.qp
,
1321 qp
->qpn
& (dev
->limits
.num_qps
- 1));
1322 spin_unlock(&dev
->qp_table
.lock
);
1324 if (send_cq
!= recv_cq
)
1325 spin_unlock(&recv_cq
->lock
);
1326 spin_unlock_irq(&send_cq
->lock
);
1328 atomic_dec(&qp
->refcount
);
1329 wait_event(qp
->wait
, !atomic_read(&qp
->refcount
));
1331 if (qp
->state
!= IB_QPS_RESET
)
1332 mthca_MODIFY_QP(dev
, qp
->state
, IB_QPS_RESET
, qp
->qpn
, 0,
1336 * If this is a userspace QP, the buffers, MR, CQs and so on
1337 * will be cleaned up in userspace, so all we have to do is
1338 * unref the mem-free tables and free the QPN in our table.
1340 if (!qp
->ibqp
.uobject
) {
1341 mthca_cq_clean(dev
, to_mcq(qp
->ibqp
.send_cq
)->cqn
, qp
->qpn
,
1342 qp
->ibqp
.srq
? to_msrq(qp
->ibqp
.srq
) : NULL
);
1343 if (qp
->ibqp
.send_cq
!= qp
->ibqp
.recv_cq
)
1344 mthca_cq_clean(dev
, to_mcq(qp
->ibqp
.recv_cq
)->cqn
, qp
->qpn
,
1345 qp
->ibqp
.srq
? to_msrq(qp
->ibqp
.srq
) : NULL
);
1347 mthca_free_memfree(dev
, qp
);
1348 mthca_free_wqe_buf(dev
, qp
);
1351 mthca_unmap_memfree(dev
, qp
);
1353 if (is_sqp(dev
, qp
)) {
1354 atomic_dec(&(to_mpd(qp
->ibqp
.pd
)->sqp_count
));
1355 dma_free_coherent(&dev
->pdev
->dev
,
1356 to_msqp(qp
)->header_buf_size
,
1357 to_msqp(qp
)->header_buf
,
1358 to_msqp(qp
)->header_dma
);
1360 mthca_free(&dev
->qp_table
.alloc
, qp
->qpn
);
1363 /* Create UD header for an MLX send and build a data segment for it */
1364 static int build_mlx_header(struct mthca_dev
*dev
, struct mthca_sqp
*sqp
,
1365 int ind
, struct ib_send_wr
*wr
,
1366 struct mthca_mlx_seg
*mlx
,
1367 struct mthca_data_seg
*data
)
1373 ib_ud_header_init(256, /* assume a MAD */
1374 mthca_ah_grh_present(to_mah(wr
->wr
.ud
.ah
)),
1377 err
= mthca_read_ah(dev
, to_mah(wr
->wr
.ud
.ah
), &sqp
->ud_header
);
1380 mlx
->flags
&= ~cpu_to_be32(MTHCA_NEXT_SOLICIT
| 1);
1381 mlx
->flags
|= cpu_to_be32((!sqp
->qp
.ibqp
.qp_num
? MTHCA_MLX_VL15
: 0) |
1382 (sqp
->ud_header
.lrh
.destination_lid
==
1383 IB_LID_PERMISSIVE
? MTHCA_MLX_SLR
: 0) |
1384 (sqp
->ud_header
.lrh
.service_level
<< 8));
1385 mlx
->rlid
= sqp
->ud_header
.lrh
.destination_lid
;
1388 switch (wr
->opcode
) {
1390 sqp
->ud_header
.bth
.opcode
= IB_OPCODE_UD_SEND_ONLY
;
1391 sqp
->ud_header
.immediate_present
= 0;
1393 case IB_WR_SEND_WITH_IMM
:
1394 sqp
->ud_header
.bth
.opcode
= IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE
;
1395 sqp
->ud_header
.immediate_present
= 1;
1396 sqp
->ud_header
.immediate_data
= wr
->imm_data
;
1402 sqp
->ud_header
.lrh
.virtual_lane
= !sqp
->qp
.ibqp
.qp_num
? 15 : 0;
1403 if (sqp
->ud_header
.lrh
.destination_lid
== IB_LID_PERMISSIVE
)
1404 sqp
->ud_header
.lrh
.source_lid
= IB_LID_PERMISSIVE
;
1405 sqp
->ud_header
.bth
.solicited_event
= !!(wr
->send_flags
& IB_SEND_SOLICITED
);
1406 if (!sqp
->qp
.ibqp
.qp_num
)
1407 ib_get_cached_pkey(&dev
->ib_dev
, sqp
->port
,
1408 sqp
->pkey_index
, &pkey
);
1410 ib_get_cached_pkey(&dev
->ib_dev
, sqp
->port
,
1411 wr
->wr
.ud
.pkey_index
, &pkey
);
1412 sqp
->ud_header
.bth
.pkey
= cpu_to_be16(pkey
);
1413 sqp
->ud_header
.bth
.destination_qpn
= cpu_to_be32(wr
->wr
.ud
.remote_qpn
);
1414 sqp
->ud_header
.bth
.psn
= cpu_to_be32((sqp
->send_psn
++) & ((1 << 24) - 1));
1415 sqp
->ud_header
.deth
.qkey
= cpu_to_be32(wr
->wr
.ud
.remote_qkey
& 0x80000000 ?
1416 sqp
->qkey
: wr
->wr
.ud
.remote_qkey
);
1417 sqp
->ud_header
.deth
.source_qpn
= cpu_to_be32(sqp
->qp
.ibqp
.qp_num
);
1419 header_size
= ib_ud_header_pack(&sqp
->ud_header
,
1421 ind
* MTHCA_UD_HEADER_SIZE
);
1423 data
->byte_count
= cpu_to_be32(header_size
);
1424 data
->lkey
= cpu_to_be32(to_mpd(sqp
->qp
.ibqp
.pd
)->ntmr
.ibmr
.lkey
);
1425 data
->addr
= cpu_to_be64(sqp
->header_dma
+
1426 ind
* MTHCA_UD_HEADER_SIZE
);
1431 static inline int mthca_wq_overflow(struct mthca_wq
*wq
, int nreq
,
1432 struct ib_cq
*ib_cq
)
1435 struct mthca_cq
*cq
;
1437 cur
= wq
->head
- wq
->tail
;
1438 if (likely(cur
+ nreq
< wq
->max
))
1442 spin_lock(&cq
->lock
);
1443 cur
= wq
->head
- wq
->tail
;
1444 spin_unlock(&cq
->lock
);
1446 return cur
+ nreq
>= wq
->max
;
1449 int mthca_tavor_post_send(struct ib_qp
*ibqp
, struct ib_send_wr
*wr
,
1450 struct ib_send_wr
**bad_wr
)
1452 struct mthca_dev
*dev
= to_mdev(ibqp
->device
);
1453 struct mthca_qp
*qp
= to_mqp(ibqp
);
1456 unsigned long flags
;
1466 spin_lock_irqsave(&qp
->sq
.lock
, flags
);
1468 /* XXX check that state is OK to post send */
1470 ind
= qp
->sq
.next_ind
;
1472 for (nreq
= 0; wr
; ++nreq
, wr
= wr
->next
) {
1473 if (mthca_wq_overflow(&qp
->sq
, nreq
, qp
->ibqp
.send_cq
)) {
1474 mthca_err(dev
, "SQ %06x full (%u head, %u tail,"
1475 " %d max, %d nreq)\n", qp
->qpn
,
1476 qp
->sq
.head
, qp
->sq
.tail
,
1483 wqe
= get_send_wqe(qp
, ind
);
1484 prev_wqe
= qp
->sq
.last
;
1487 ((struct mthca_next_seg
*) wqe
)->nda_op
= 0;
1488 ((struct mthca_next_seg
*) wqe
)->ee_nds
= 0;
1489 ((struct mthca_next_seg
*) wqe
)->flags
=
1490 ((wr
->send_flags
& IB_SEND_SIGNALED
) ?
1491 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE
) : 0) |
1492 ((wr
->send_flags
& IB_SEND_SOLICITED
) ?
1493 cpu_to_be32(MTHCA_NEXT_SOLICIT
) : 0) |
1495 if (wr
->opcode
== IB_WR_SEND_WITH_IMM
||
1496 wr
->opcode
== IB_WR_RDMA_WRITE_WITH_IMM
)
1497 ((struct mthca_next_seg
*) wqe
)->imm
= wr
->imm_data
;
1499 wqe
+= sizeof (struct mthca_next_seg
);
1500 size
= sizeof (struct mthca_next_seg
) / 16;
1502 switch (qp
->transport
) {
1504 switch (wr
->opcode
) {
1505 case IB_WR_ATOMIC_CMP_AND_SWP
:
1506 case IB_WR_ATOMIC_FETCH_AND_ADD
:
1507 ((struct mthca_raddr_seg
*) wqe
)->raddr
=
1508 cpu_to_be64(wr
->wr
.atomic
.remote_addr
);
1509 ((struct mthca_raddr_seg
*) wqe
)->rkey
=
1510 cpu_to_be32(wr
->wr
.atomic
.rkey
);
1511 ((struct mthca_raddr_seg
*) wqe
)->reserved
= 0;
1513 wqe
+= sizeof (struct mthca_raddr_seg
);
1515 if (wr
->opcode
== IB_WR_ATOMIC_CMP_AND_SWP
) {
1516 ((struct mthca_atomic_seg
*) wqe
)->swap_add
=
1517 cpu_to_be64(wr
->wr
.atomic
.swap
);
1518 ((struct mthca_atomic_seg
*) wqe
)->compare
=
1519 cpu_to_be64(wr
->wr
.atomic
.compare_add
);
1521 ((struct mthca_atomic_seg
*) wqe
)->swap_add
=
1522 cpu_to_be64(wr
->wr
.atomic
.compare_add
);
1523 ((struct mthca_atomic_seg
*) wqe
)->compare
= 0;
1526 wqe
+= sizeof (struct mthca_atomic_seg
);
1527 size
+= (sizeof (struct mthca_raddr_seg
) +
1528 sizeof (struct mthca_atomic_seg
)) / 16;
1531 case IB_WR_RDMA_WRITE
:
1532 case IB_WR_RDMA_WRITE_WITH_IMM
:
1533 case IB_WR_RDMA_READ
:
1534 ((struct mthca_raddr_seg
*) wqe
)->raddr
=
1535 cpu_to_be64(wr
->wr
.rdma
.remote_addr
);
1536 ((struct mthca_raddr_seg
*) wqe
)->rkey
=
1537 cpu_to_be32(wr
->wr
.rdma
.rkey
);
1538 ((struct mthca_raddr_seg
*) wqe
)->reserved
= 0;
1539 wqe
+= sizeof (struct mthca_raddr_seg
);
1540 size
+= sizeof (struct mthca_raddr_seg
) / 16;
1544 /* No extra segments required for sends */
1551 switch (wr
->opcode
) {
1552 case IB_WR_RDMA_WRITE
:
1553 case IB_WR_RDMA_WRITE_WITH_IMM
:
1554 ((struct mthca_raddr_seg
*) wqe
)->raddr
=
1555 cpu_to_be64(wr
->wr
.rdma
.remote_addr
);
1556 ((struct mthca_raddr_seg
*) wqe
)->rkey
=
1557 cpu_to_be32(wr
->wr
.rdma
.rkey
);
1558 ((struct mthca_raddr_seg
*) wqe
)->reserved
= 0;
1559 wqe
+= sizeof (struct mthca_raddr_seg
);
1560 size
+= sizeof (struct mthca_raddr_seg
) / 16;
1564 /* No extra segments required for sends */
1571 ((struct mthca_tavor_ud_seg
*) wqe
)->lkey
=
1572 cpu_to_be32(to_mah(wr
->wr
.ud
.ah
)->key
);
1573 ((struct mthca_tavor_ud_seg
*) wqe
)->av_addr
=
1574 cpu_to_be64(to_mah(wr
->wr
.ud
.ah
)->avdma
);
1575 ((struct mthca_tavor_ud_seg
*) wqe
)->dqpn
=
1576 cpu_to_be32(wr
->wr
.ud
.remote_qpn
);
1577 ((struct mthca_tavor_ud_seg
*) wqe
)->qkey
=
1578 cpu_to_be32(wr
->wr
.ud
.remote_qkey
);
1580 wqe
+= sizeof (struct mthca_tavor_ud_seg
);
1581 size
+= sizeof (struct mthca_tavor_ud_seg
) / 16;
1585 err
= build_mlx_header(dev
, to_msqp(qp
), ind
, wr
,
1586 wqe
- sizeof (struct mthca_next_seg
),
1592 wqe
+= sizeof (struct mthca_data_seg
);
1593 size
+= sizeof (struct mthca_data_seg
) / 16;
1597 if (wr
->num_sge
> qp
->sq
.max_gs
) {
1598 mthca_err(dev
, "too many gathers\n");
1604 for (i
= 0; i
< wr
->num_sge
; ++i
) {
1605 ((struct mthca_data_seg
*) wqe
)->byte_count
=
1606 cpu_to_be32(wr
->sg_list
[i
].length
);
1607 ((struct mthca_data_seg
*) wqe
)->lkey
=
1608 cpu_to_be32(wr
->sg_list
[i
].lkey
);
1609 ((struct mthca_data_seg
*) wqe
)->addr
=
1610 cpu_to_be64(wr
->sg_list
[i
].addr
);
1611 wqe
+= sizeof (struct mthca_data_seg
);
1612 size
+= sizeof (struct mthca_data_seg
) / 16;
1615 /* Add one more inline data segment for ICRC */
1616 if (qp
->transport
== MLX
) {
1617 ((struct mthca_data_seg
*) wqe
)->byte_count
=
1618 cpu_to_be32((1 << 31) | 4);
1619 ((u32
*) wqe
)[1] = 0;
1620 wqe
+= sizeof (struct mthca_data_seg
);
1621 size
+= sizeof (struct mthca_data_seg
) / 16;
1624 qp
->wrid
[ind
+ qp
->rq
.max
] = wr
->wr_id
;
1626 if (wr
->opcode
>= ARRAY_SIZE(mthca_opcode
)) {
1627 mthca_err(dev
, "opcode invalid\n");
1633 ((struct mthca_next_seg
*) prev_wqe
)->nda_op
=
1634 cpu_to_be32(((ind
<< qp
->sq
.wqe_shift
) +
1635 qp
->send_wqe_offset
) |
1636 mthca_opcode
[wr
->opcode
]);
1638 ((struct mthca_next_seg
*) prev_wqe
)->ee_nds
=
1639 cpu_to_be32((size0
? 0 : MTHCA_NEXT_DBD
) | size
|
1640 ((wr
->send_flags
& IB_SEND_FENCE
) ?
1641 MTHCA_NEXT_FENCE
: 0));
1645 op0
= mthca_opcode
[wr
->opcode
];
1649 if (unlikely(ind
>= qp
->sq
.max
))
1657 doorbell
[0] = cpu_to_be32(((qp
->sq
.next_ind
<< qp
->sq
.wqe_shift
) +
1658 qp
->send_wqe_offset
) | f0
| op0
);
1659 doorbell
[1] = cpu_to_be32((qp
->qpn
<< 8) | size0
);
1663 mthca_write64(doorbell
,
1664 dev
->kar
+ MTHCA_SEND_DOORBELL
,
1665 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
1668 qp
->sq
.next_ind
= ind
;
1669 qp
->sq
.head
+= nreq
;
1671 spin_unlock_irqrestore(&qp
->sq
.lock
, flags
);
1675 int mthca_tavor_post_receive(struct ib_qp
*ibqp
, struct ib_recv_wr
*wr
,
1676 struct ib_recv_wr
**bad_wr
)
1678 struct mthca_dev
*dev
= to_mdev(ibqp
->device
);
1679 struct mthca_qp
*qp
= to_mqp(ibqp
);
1681 unsigned long flags
;
1691 spin_lock_irqsave(&qp
->rq
.lock
, flags
);
1693 /* XXX check that state is OK to post receive */
1695 ind
= qp
->rq
.next_ind
;
1697 for (nreq
= 0; wr
; ++nreq
, wr
= wr
->next
) {
1698 if (unlikely(nreq
== MTHCA_TAVOR_MAX_WQES_PER_RECV_DB
)) {
1701 doorbell
[0] = cpu_to_be32((qp
->rq
.next_ind
<< qp
->rq
.wqe_shift
) | size0
);
1702 doorbell
[1] = cpu_to_be32(qp
->qpn
<< 8);
1706 mthca_write64(doorbell
,
1707 dev
->kar
+ MTHCA_RECEIVE_DOORBELL
,
1708 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
1710 qp
->rq
.head
+= MTHCA_TAVOR_MAX_WQES_PER_RECV_DB
;
1714 if (mthca_wq_overflow(&qp
->rq
, nreq
, qp
->ibqp
.recv_cq
)) {
1715 mthca_err(dev
, "RQ %06x full (%u head, %u tail,"
1716 " %d max, %d nreq)\n", qp
->qpn
,
1717 qp
->rq
.head
, qp
->rq
.tail
,
1724 wqe
= get_recv_wqe(qp
, ind
);
1725 prev_wqe
= qp
->rq
.last
;
1728 ((struct mthca_next_seg
*) wqe
)->nda_op
= 0;
1729 ((struct mthca_next_seg
*) wqe
)->ee_nds
=
1730 cpu_to_be32(MTHCA_NEXT_DBD
);
1731 ((struct mthca_next_seg
*) wqe
)->flags
= 0;
1733 wqe
+= sizeof (struct mthca_next_seg
);
1734 size
= sizeof (struct mthca_next_seg
) / 16;
1736 if (unlikely(wr
->num_sge
> qp
->rq
.max_gs
)) {
1742 for (i
= 0; i
< wr
->num_sge
; ++i
) {
1743 ((struct mthca_data_seg
*) wqe
)->byte_count
=
1744 cpu_to_be32(wr
->sg_list
[i
].length
);
1745 ((struct mthca_data_seg
*) wqe
)->lkey
=
1746 cpu_to_be32(wr
->sg_list
[i
].lkey
);
1747 ((struct mthca_data_seg
*) wqe
)->addr
=
1748 cpu_to_be64(wr
->sg_list
[i
].addr
);
1749 wqe
+= sizeof (struct mthca_data_seg
);
1750 size
+= sizeof (struct mthca_data_seg
) / 16;
1753 qp
->wrid
[ind
] = wr
->wr_id
;
1755 ((struct mthca_next_seg
*) prev_wqe
)->nda_op
=
1756 cpu_to_be32((ind
<< qp
->rq
.wqe_shift
) | 1);
1758 ((struct mthca_next_seg
*) prev_wqe
)->ee_nds
=
1759 cpu_to_be32(MTHCA_NEXT_DBD
| size
);
1765 if (unlikely(ind
>= qp
->rq
.max
))
1771 doorbell
[0] = cpu_to_be32((qp
->rq
.next_ind
<< qp
->rq
.wqe_shift
) | size0
);
1772 doorbell
[1] = cpu_to_be32((qp
->qpn
<< 8) | nreq
);
1776 mthca_write64(doorbell
,
1777 dev
->kar
+ MTHCA_RECEIVE_DOORBELL
,
1778 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
1781 qp
->rq
.next_ind
= ind
;
1782 qp
->rq
.head
+= nreq
;
1784 spin_unlock_irqrestore(&qp
->rq
.lock
, flags
);
1788 int mthca_arbel_post_send(struct ib_qp
*ibqp
, struct ib_send_wr
*wr
,
1789 struct ib_send_wr
**bad_wr
)
1791 struct mthca_dev
*dev
= to_mdev(ibqp
->device
);
1792 struct mthca_qp
*qp
= to_mqp(ibqp
);
1796 unsigned long flags
;
1806 spin_lock_irqsave(&qp
->sq
.lock
, flags
);
1808 /* XXX check that state is OK to post send */
1810 ind
= qp
->sq
.head
& (qp
->sq
.max
- 1);
1812 for (nreq
= 0; wr
; ++nreq
, wr
= wr
->next
) {
1813 if (unlikely(nreq
== MTHCA_ARBEL_MAX_WQES_PER_SEND_DB
)) {
1816 doorbell
[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB
<< 24) |
1817 ((qp
->sq
.head
& 0xffff) << 8) |
1819 doorbell
[1] = cpu_to_be32((qp
->qpn
<< 8) | size0
);
1821 qp
->sq
.head
+= MTHCA_ARBEL_MAX_WQES_PER_SEND_DB
;
1825 * Make sure that descriptors are written before
1829 *qp
->sq
.db
= cpu_to_be32(qp
->sq
.head
& 0xffff);
1832 * Make sure doorbell record is written before we
1833 * write MMIO send doorbell.
1836 mthca_write64(doorbell
,
1837 dev
->kar
+ MTHCA_SEND_DOORBELL
,
1838 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
1841 if (mthca_wq_overflow(&qp
->sq
, nreq
, qp
->ibqp
.send_cq
)) {
1842 mthca_err(dev
, "SQ %06x full (%u head, %u tail,"
1843 " %d max, %d nreq)\n", qp
->qpn
,
1844 qp
->sq
.head
, qp
->sq
.tail
,
1851 wqe
= get_send_wqe(qp
, ind
);
1852 prev_wqe
= qp
->sq
.last
;
1855 ((struct mthca_next_seg
*) wqe
)->flags
=
1856 ((wr
->send_flags
& IB_SEND_SIGNALED
) ?
1857 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE
) : 0) |
1858 ((wr
->send_flags
& IB_SEND_SOLICITED
) ?
1859 cpu_to_be32(MTHCA_NEXT_SOLICIT
) : 0) |
1861 if (wr
->opcode
== IB_WR_SEND_WITH_IMM
||
1862 wr
->opcode
== IB_WR_RDMA_WRITE_WITH_IMM
)
1863 ((struct mthca_next_seg
*) wqe
)->imm
= wr
->imm_data
;
1865 wqe
+= sizeof (struct mthca_next_seg
);
1866 size
= sizeof (struct mthca_next_seg
) / 16;
1868 switch (qp
->transport
) {
1870 switch (wr
->opcode
) {
1871 case IB_WR_ATOMIC_CMP_AND_SWP
:
1872 case IB_WR_ATOMIC_FETCH_AND_ADD
:
1873 ((struct mthca_raddr_seg
*) wqe
)->raddr
=
1874 cpu_to_be64(wr
->wr
.atomic
.remote_addr
);
1875 ((struct mthca_raddr_seg
*) wqe
)->rkey
=
1876 cpu_to_be32(wr
->wr
.atomic
.rkey
);
1877 ((struct mthca_raddr_seg
*) wqe
)->reserved
= 0;
1879 wqe
+= sizeof (struct mthca_raddr_seg
);
1881 if (wr
->opcode
== IB_WR_ATOMIC_CMP_AND_SWP
) {
1882 ((struct mthca_atomic_seg
*) wqe
)->swap_add
=
1883 cpu_to_be64(wr
->wr
.atomic
.swap
);
1884 ((struct mthca_atomic_seg
*) wqe
)->compare
=
1885 cpu_to_be64(wr
->wr
.atomic
.compare_add
);
1887 ((struct mthca_atomic_seg
*) wqe
)->swap_add
=
1888 cpu_to_be64(wr
->wr
.atomic
.compare_add
);
1889 ((struct mthca_atomic_seg
*) wqe
)->compare
= 0;
1892 wqe
+= sizeof (struct mthca_atomic_seg
);
1893 size
+= (sizeof (struct mthca_raddr_seg
) +
1894 sizeof (struct mthca_atomic_seg
)) / 16;
1897 case IB_WR_RDMA_READ
:
1898 case IB_WR_RDMA_WRITE
:
1899 case IB_WR_RDMA_WRITE_WITH_IMM
:
1900 ((struct mthca_raddr_seg
*) wqe
)->raddr
=
1901 cpu_to_be64(wr
->wr
.rdma
.remote_addr
);
1902 ((struct mthca_raddr_seg
*) wqe
)->rkey
=
1903 cpu_to_be32(wr
->wr
.rdma
.rkey
);
1904 ((struct mthca_raddr_seg
*) wqe
)->reserved
= 0;
1905 wqe
+= sizeof (struct mthca_raddr_seg
);
1906 size
+= sizeof (struct mthca_raddr_seg
) / 16;
1910 /* No extra segments required for sends */
1917 switch (wr
->opcode
) {
1918 case IB_WR_RDMA_WRITE
:
1919 case IB_WR_RDMA_WRITE_WITH_IMM
:
1920 ((struct mthca_raddr_seg
*) wqe
)->raddr
=
1921 cpu_to_be64(wr
->wr
.rdma
.remote_addr
);
1922 ((struct mthca_raddr_seg
*) wqe
)->rkey
=
1923 cpu_to_be32(wr
->wr
.rdma
.rkey
);
1924 ((struct mthca_raddr_seg
*) wqe
)->reserved
= 0;
1925 wqe
+= sizeof (struct mthca_raddr_seg
);
1926 size
+= sizeof (struct mthca_raddr_seg
) / 16;
1930 /* No extra segments required for sends */
1937 memcpy(((struct mthca_arbel_ud_seg
*) wqe
)->av
,
1938 to_mah(wr
->wr
.ud
.ah
)->av
, MTHCA_AV_SIZE
);
1939 ((struct mthca_arbel_ud_seg
*) wqe
)->dqpn
=
1940 cpu_to_be32(wr
->wr
.ud
.remote_qpn
);
1941 ((struct mthca_arbel_ud_seg
*) wqe
)->qkey
=
1942 cpu_to_be32(wr
->wr
.ud
.remote_qkey
);
1944 wqe
+= sizeof (struct mthca_arbel_ud_seg
);
1945 size
+= sizeof (struct mthca_arbel_ud_seg
) / 16;
1949 err
= build_mlx_header(dev
, to_msqp(qp
), ind
, wr
,
1950 wqe
- sizeof (struct mthca_next_seg
),
1956 wqe
+= sizeof (struct mthca_data_seg
);
1957 size
+= sizeof (struct mthca_data_seg
) / 16;
1961 if (wr
->num_sge
> qp
->sq
.max_gs
) {
1962 mthca_err(dev
, "too many gathers\n");
1968 for (i
= 0; i
< wr
->num_sge
; ++i
) {
1969 ((struct mthca_data_seg
*) wqe
)->byte_count
=
1970 cpu_to_be32(wr
->sg_list
[i
].length
);
1971 ((struct mthca_data_seg
*) wqe
)->lkey
=
1972 cpu_to_be32(wr
->sg_list
[i
].lkey
);
1973 ((struct mthca_data_seg
*) wqe
)->addr
=
1974 cpu_to_be64(wr
->sg_list
[i
].addr
);
1975 wqe
+= sizeof (struct mthca_data_seg
);
1976 size
+= sizeof (struct mthca_data_seg
) / 16;
1979 /* Add one more inline data segment for ICRC */
1980 if (qp
->transport
== MLX
) {
1981 ((struct mthca_data_seg
*) wqe
)->byte_count
=
1982 cpu_to_be32((1 << 31) | 4);
1983 ((u32
*) wqe
)[1] = 0;
1984 wqe
+= sizeof (struct mthca_data_seg
);
1985 size
+= sizeof (struct mthca_data_seg
) / 16;
1988 qp
->wrid
[ind
+ qp
->rq
.max
] = wr
->wr_id
;
1990 if (wr
->opcode
>= ARRAY_SIZE(mthca_opcode
)) {
1991 mthca_err(dev
, "opcode invalid\n");
1997 ((struct mthca_next_seg
*) prev_wqe
)->nda_op
=
1998 cpu_to_be32(((ind
<< qp
->sq
.wqe_shift
) +
1999 qp
->send_wqe_offset
) |
2000 mthca_opcode
[wr
->opcode
]);
2002 ((struct mthca_next_seg
*) prev_wqe
)->ee_nds
=
2003 cpu_to_be32(MTHCA_NEXT_DBD
| size
|
2004 ((wr
->send_flags
& IB_SEND_FENCE
) ?
2005 MTHCA_NEXT_FENCE
: 0));
2009 op0
= mthca_opcode
[wr
->opcode
];
2013 if (unlikely(ind
>= qp
->sq
.max
))
2019 doorbell
[0] = cpu_to_be32((nreq
<< 24) |
2020 ((qp
->sq
.head
& 0xffff) << 8) |
2022 doorbell
[1] = cpu_to_be32((qp
->qpn
<< 8) | size0
);
2024 qp
->sq
.head
+= nreq
;
2027 * Make sure that descriptors are written before
2031 *qp
->sq
.db
= cpu_to_be32(qp
->sq
.head
& 0xffff);
2034 * Make sure doorbell record is written before we
2035 * write MMIO send doorbell.
2038 mthca_write64(doorbell
,
2039 dev
->kar
+ MTHCA_SEND_DOORBELL
,
2040 MTHCA_GET_DOORBELL_LOCK(&dev
->doorbell_lock
));
2043 spin_unlock_irqrestore(&qp
->sq
.lock
, flags
);
2047 int mthca_arbel_post_receive(struct ib_qp
*ibqp
, struct ib_recv_wr
*wr
,
2048 struct ib_recv_wr
**bad_wr
)
2050 struct mthca_dev
*dev
= to_mdev(ibqp
->device
);
2051 struct mthca_qp
*qp
= to_mqp(ibqp
);
2052 unsigned long flags
;
2059 spin_lock_irqsave(&qp
->rq
.lock
, flags
);
2061 /* XXX check that state is OK to post receive */
2063 ind
= qp
->rq
.head
& (qp
->rq
.max
- 1);
2065 for (nreq
= 0; wr
; ++nreq
, wr
= wr
->next
) {
2066 if (mthca_wq_overflow(&qp
->rq
, nreq
, qp
->ibqp
.recv_cq
)) {
2067 mthca_err(dev
, "RQ %06x full (%u head, %u tail,"
2068 " %d max, %d nreq)\n", qp
->qpn
,
2069 qp
->rq
.head
, qp
->rq
.tail
,
2076 wqe
= get_recv_wqe(qp
, ind
);
2078 ((struct mthca_next_seg
*) wqe
)->flags
= 0;
2080 wqe
+= sizeof (struct mthca_next_seg
);
2082 if (unlikely(wr
->num_sge
> qp
->rq
.max_gs
)) {
2088 for (i
= 0; i
< wr
->num_sge
; ++i
) {
2089 ((struct mthca_data_seg
*) wqe
)->byte_count
=
2090 cpu_to_be32(wr
->sg_list
[i
].length
);
2091 ((struct mthca_data_seg
*) wqe
)->lkey
=
2092 cpu_to_be32(wr
->sg_list
[i
].lkey
);
2093 ((struct mthca_data_seg
*) wqe
)->addr
=
2094 cpu_to_be64(wr
->sg_list
[i
].addr
);
2095 wqe
+= sizeof (struct mthca_data_seg
);
2098 if (i
< qp
->rq
.max_gs
) {
2099 ((struct mthca_data_seg
*) wqe
)->byte_count
= 0;
2100 ((struct mthca_data_seg
*) wqe
)->lkey
= cpu_to_be32(MTHCA_INVAL_LKEY
);
2101 ((struct mthca_data_seg
*) wqe
)->addr
= 0;
2104 qp
->wrid
[ind
] = wr
->wr_id
;
2107 if (unlikely(ind
>= qp
->rq
.max
))
2112 qp
->rq
.head
+= nreq
;
2115 * Make sure that descriptors are written before
2119 *qp
->rq
.db
= cpu_to_be32(qp
->rq
.head
& 0xffff);
2122 spin_unlock_irqrestore(&qp
->rq
.lock
, flags
);
2126 void mthca_free_err_wqe(struct mthca_dev
*dev
, struct mthca_qp
*qp
, int is_send
,
2127 int index
, int *dbd
, __be32
*new_wqe
)
2129 struct mthca_next_seg
*next
;
2132 * For SRQs, all WQEs generate a CQE, so we're always at the
2133 * end of the doorbell chain.
2141 next
= get_send_wqe(qp
, index
);
2143 next
= get_recv_wqe(qp
, index
);
2145 *dbd
= !!(next
->ee_nds
& cpu_to_be32(MTHCA_NEXT_DBD
));
2146 if (next
->ee_nds
& cpu_to_be32(0x3f))
2147 *new_wqe
= (next
->nda_op
& cpu_to_be32(~0x3f)) |
2148 (next
->ee_nds
& cpu_to_be32(0x3f));
2153 int __devinit
mthca_init_qp_table(struct mthca_dev
*dev
)
2159 spin_lock_init(&dev
->qp_table
.lock
);
2162 * We reserve 2 extra QPs per port for the special QPs. The
2163 * special QP for port 1 has to be even, so round up.
2165 dev
->qp_table
.sqp_start
= (dev
->limits
.reserved_qps
+ 1) & ~1UL;
2166 err
= mthca_alloc_init(&dev
->qp_table
.alloc
,
2167 dev
->limits
.num_qps
,
2169 dev
->qp_table
.sqp_start
+
2170 MTHCA_MAX_PORTS
* 2);
2174 err
= mthca_array_init(&dev
->qp_table
.qp
,
2175 dev
->limits
.num_qps
);
2177 mthca_alloc_cleanup(&dev
->qp_table
.alloc
);
2181 for (i
= 0; i
< 2; ++i
) {
2182 err
= mthca_CONF_SPECIAL_QP(dev
, i
? IB_QPT_GSI
: IB_QPT_SMI
,
2183 dev
->qp_table
.sqp_start
+ i
* 2,
2188 mthca_warn(dev
, "CONF_SPECIAL_QP returned "
2189 "status %02x, aborting.\n",
2198 for (i
= 0; i
< 2; ++i
)
2199 mthca_CONF_SPECIAL_QP(dev
, i
, 0, &status
);
2201 mthca_array_cleanup(&dev
->qp_table
.qp
, dev
->limits
.num_qps
);
2202 mthca_alloc_cleanup(&dev
->qp_table
.alloc
);
2207 void mthca_cleanup_qp_table(struct mthca_dev
*dev
)
2212 for (i
= 0; i
< 2; ++i
)
2213 mthca_CONF_SPECIAL_QP(dev
, i
, 0, &status
);
2215 mthca_array_cleanup(&dev
->qp_table
.qp
, dev
->limits
.num_qps
);
2216 mthca_alloc_cleanup(&dev
->qp_table
.alloc
);