3 * This file is provided under a dual BSD/GPLv2 license. When using or
4 * redistributing this file, you may do so under either license.
8 * Copyright(c) 2015 Intel Corporation.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License as
12 * published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful, but
15 * WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * General Public License for more details.
21 * Copyright(c) 2015 Intel Corporation.
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
27 * - Redistributions of source code must retain the above copyright
28 * notice, this list of conditions and the following disclaimer.
29 * - Redistributions in binary form must reproduce the above copyright
30 * notice, this list of conditions and the following disclaimer in
31 * the documentation and/or other materials provided with the
33 * - Neither the name of Intel Corporation nor the names of its
34 * contributors may be used to endorse or promote products derived
35 * from this software without specific prior written permission.
37 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
38 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
39 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
40 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
41 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
42 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
43 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
44 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
45 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
46 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
47 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
51 #include <rdma/ib_mad.h>
52 #include <rdma/ib_user_verbs.h>
54 #include <linux/module.h>
55 #include <linux/utsname.h>
56 #include <linux/rculist.h>
58 #include <linux/random.h>
59 #include <linux/vmalloc.h>
68 unsigned int hfi1_lkey_table_size
= 16;
69 module_param_named(lkey_table_size
, hfi1_lkey_table_size
, uint
,
71 MODULE_PARM_DESC(lkey_table_size
,
72 "LKEY table size in bits (2^n, 1 <= n <= 23)");
74 static unsigned int hfi1_max_pds
= 0xFFFF;
75 module_param_named(max_pds
, hfi1_max_pds
, uint
, S_IRUGO
);
76 MODULE_PARM_DESC(max_pds
,
77 "Maximum number of protection domains to support");
79 static unsigned int hfi1_max_ahs
= 0xFFFF;
80 module_param_named(max_ahs
, hfi1_max_ahs
, uint
, S_IRUGO
);
81 MODULE_PARM_DESC(max_ahs
, "Maximum number of address handles to support");
83 unsigned int hfi1_max_cqes
= 0x2FFFF;
84 module_param_named(max_cqes
, hfi1_max_cqes
, uint
, S_IRUGO
);
85 MODULE_PARM_DESC(max_cqes
,
86 "Maximum number of completion queue entries to support");
88 unsigned int hfi1_max_cqs
= 0x1FFFF;
89 module_param_named(max_cqs
, hfi1_max_cqs
, uint
, S_IRUGO
);
90 MODULE_PARM_DESC(max_cqs
, "Maximum number of completion queues to support");
92 unsigned int hfi1_max_qp_wrs
= 0x3FFF;
93 module_param_named(max_qp_wrs
, hfi1_max_qp_wrs
, uint
, S_IRUGO
);
94 MODULE_PARM_DESC(max_qp_wrs
, "Maximum number of QP WRs to support");
96 unsigned int hfi1_max_qps
= 16384;
97 module_param_named(max_qps
, hfi1_max_qps
, uint
, S_IRUGO
);
98 MODULE_PARM_DESC(max_qps
, "Maximum number of QPs to support");
100 unsigned int hfi1_max_sges
= 0x60;
101 module_param_named(max_sges
, hfi1_max_sges
, uint
, S_IRUGO
);
102 MODULE_PARM_DESC(max_sges
, "Maximum number of SGEs to support");
104 unsigned int hfi1_max_mcast_grps
= 16384;
105 module_param_named(max_mcast_grps
, hfi1_max_mcast_grps
, uint
, S_IRUGO
);
106 MODULE_PARM_DESC(max_mcast_grps
,
107 "Maximum number of multicast groups to support");
109 unsigned int hfi1_max_mcast_qp_attached
= 16;
110 module_param_named(max_mcast_qp_attached
, hfi1_max_mcast_qp_attached
,
112 MODULE_PARM_DESC(max_mcast_qp_attached
,
113 "Maximum number of attached QPs to support");
115 unsigned int hfi1_max_srqs
= 1024;
116 module_param_named(max_srqs
, hfi1_max_srqs
, uint
, S_IRUGO
);
117 MODULE_PARM_DESC(max_srqs
, "Maximum number of SRQs to support");
119 unsigned int hfi1_max_srq_sges
= 128;
120 module_param_named(max_srq_sges
, hfi1_max_srq_sges
, uint
, S_IRUGO
);
121 MODULE_PARM_DESC(max_srq_sges
, "Maximum number of SRQ SGEs to support");
123 unsigned int hfi1_max_srq_wrs
= 0x1FFFF;
124 module_param_named(max_srq_wrs
, hfi1_max_srq_wrs
, uint
, S_IRUGO
);
125 MODULE_PARM_DESC(max_srq_wrs
, "Maximum number of SRQ WRs support");
127 static void verbs_sdma_complete(
128 struct sdma_txreq
*cookie
,
132 /* Length of buffer to create verbs txreq cache name */
133 #define TXREQ_NAME_LEN 24
136 * Note that it is OK to post send work requests in the SQE and ERR
137 * states; hfi1_do_send() will process them and generate error
138 * completions as per IB 1.2 C10-96.
140 const int ib_hfi1_state_ops
[IB_QPS_ERR
+ 1] = {
142 [IB_QPS_INIT
] = HFI1_POST_RECV_OK
,
143 [IB_QPS_RTR
] = HFI1_POST_RECV_OK
| HFI1_PROCESS_RECV_OK
,
144 [IB_QPS_RTS
] = HFI1_POST_RECV_OK
| HFI1_PROCESS_RECV_OK
|
145 HFI1_POST_SEND_OK
| HFI1_PROCESS_SEND_OK
|
146 HFI1_PROCESS_NEXT_SEND_OK
,
147 [IB_QPS_SQD
] = HFI1_POST_RECV_OK
| HFI1_PROCESS_RECV_OK
|
148 HFI1_POST_SEND_OK
| HFI1_PROCESS_SEND_OK
,
149 [IB_QPS_SQE
] = HFI1_POST_RECV_OK
| HFI1_PROCESS_RECV_OK
|
150 HFI1_POST_SEND_OK
| HFI1_FLUSH_SEND
,
151 [IB_QPS_ERR
] = HFI1_POST_RECV_OK
| HFI1_FLUSH_RECV
|
152 HFI1_POST_SEND_OK
| HFI1_FLUSH_SEND
,
155 struct hfi1_ucontext
{
156 struct ib_ucontext ibucontext
;
159 static inline struct hfi1_ucontext
*to_iucontext(struct ib_ucontext
162 return container_of(ibucontext
, struct hfi1_ucontext
, ibucontext
);
165 static inline void _hfi1_schedule_send(struct hfi1_qp
*qp
);
168 * Translate ib_wr_opcode into ib_wc_opcode.
170 const enum ib_wc_opcode ib_hfi1_wc_opcode
[] = {
171 [IB_WR_RDMA_WRITE
] = IB_WC_RDMA_WRITE
,
172 [IB_WR_RDMA_WRITE_WITH_IMM
] = IB_WC_RDMA_WRITE
,
173 [IB_WR_SEND
] = IB_WC_SEND
,
174 [IB_WR_SEND_WITH_IMM
] = IB_WC_SEND
,
175 [IB_WR_RDMA_READ
] = IB_WC_RDMA_READ
,
176 [IB_WR_ATOMIC_CMP_AND_SWP
] = IB_WC_COMP_SWAP
,
177 [IB_WR_ATOMIC_FETCH_AND_ADD
] = IB_WC_FETCH_ADD
181 * Length of header by opcode, 0 --> not supported
183 const u8 hdr_len_by_opcode
[256] = {
185 [IB_OPCODE_RC_SEND_FIRST
] = 12 + 8,
186 [IB_OPCODE_RC_SEND_MIDDLE
] = 12 + 8,
187 [IB_OPCODE_RC_SEND_LAST
] = 12 + 8,
188 [IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE
] = 12 + 8 + 4,
189 [IB_OPCODE_RC_SEND_ONLY
] = 12 + 8,
190 [IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE
] = 12 + 8 + 4,
191 [IB_OPCODE_RC_RDMA_WRITE_FIRST
] = 12 + 8 + 16,
192 [IB_OPCODE_RC_RDMA_WRITE_MIDDLE
] = 12 + 8,
193 [IB_OPCODE_RC_RDMA_WRITE_LAST
] = 12 + 8,
194 [IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE
] = 12 + 8 + 4,
195 [IB_OPCODE_RC_RDMA_WRITE_ONLY
] = 12 + 8 + 16,
196 [IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE
] = 12 + 8 + 20,
197 [IB_OPCODE_RC_RDMA_READ_REQUEST
] = 12 + 8 + 16,
198 [IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST
] = 12 + 8 + 4,
199 [IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE
] = 12 + 8,
200 [IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST
] = 12 + 8 + 4,
201 [IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY
] = 12 + 8 + 4,
202 [IB_OPCODE_RC_ACKNOWLEDGE
] = 12 + 8 + 4,
203 [IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE
] = 12 + 8 + 4,
204 [IB_OPCODE_RC_COMPARE_SWAP
] = 12 + 8 + 28,
205 [IB_OPCODE_RC_FETCH_ADD
] = 12 + 8 + 28,
207 [IB_OPCODE_UC_SEND_FIRST
] = 12 + 8,
208 [IB_OPCODE_UC_SEND_MIDDLE
] = 12 + 8,
209 [IB_OPCODE_UC_SEND_LAST
] = 12 + 8,
210 [IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE
] = 12 + 8 + 4,
211 [IB_OPCODE_UC_SEND_ONLY
] = 12 + 8,
212 [IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE
] = 12 + 8 + 4,
213 [IB_OPCODE_UC_RDMA_WRITE_FIRST
] = 12 + 8 + 16,
214 [IB_OPCODE_UC_RDMA_WRITE_MIDDLE
] = 12 + 8,
215 [IB_OPCODE_UC_RDMA_WRITE_LAST
] = 12 + 8,
216 [IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE
] = 12 + 8 + 4,
217 [IB_OPCODE_UC_RDMA_WRITE_ONLY
] = 12 + 8 + 16,
218 [IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE
] = 12 + 8 + 20,
220 [IB_OPCODE_UD_SEND_ONLY
] = 12 + 8 + 8,
221 [IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE
] = 12 + 8 + 12
224 static const opcode_handler opcode_handler_tbl
[256] = {
226 [IB_OPCODE_RC_SEND_FIRST
] = &hfi1_rc_rcv
,
227 [IB_OPCODE_RC_SEND_MIDDLE
] = &hfi1_rc_rcv
,
228 [IB_OPCODE_RC_SEND_LAST
] = &hfi1_rc_rcv
,
229 [IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE
] = &hfi1_rc_rcv
,
230 [IB_OPCODE_RC_SEND_ONLY
] = &hfi1_rc_rcv
,
231 [IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE
] = &hfi1_rc_rcv
,
232 [IB_OPCODE_RC_RDMA_WRITE_FIRST
] = &hfi1_rc_rcv
,
233 [IB_OPCODE_RC_RDMA_WRITE_MIDDLE
] = &hfi1_rc_rcv
,
234 [IB_OPCODE_RC_RDMA_WRITE_LAST
] = &hfi1_rc_rcv
,
235 [IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE
] = &hfi1_rc_rcv
,
236 [IB_OPCODE_RC_RDMA_WRITE_ONLY
] = &hfi1_rc_rcv
,
237 [IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE
] = &hfi1_rc_rcv
,
238 [IB_OPCODE_RC_RDMA_READ_REQUEST
] = &hfi1_rc_rcv
,
239 [IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST
] = &hfi1_rc_rcv
,
240 [IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE
] = &hfi1_rc_rcv
,
241 [IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST
] = &hfi1_rc_rcv
,
242 [IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY
] = &hfi1_rc_rcv
,
243 [IB_OPCODE_RC_ACKNOWLEDGE
] = &hfi1_rc_rcv
,
244 [IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE
] = &hfi1_rc_rcv
,
245 [IB_OPCODE_RC_COMPARE_SWAP
] = &hfi1_rc_rcv
,
246 [IB_OPCODE_RC_FETCH_ADD
] = &hfi1_rc_rcv
,
248 [IB_OPCODE_UC_SEND_FIRST
] = &hfi1_uc_rcv
,
249 [IB_OPCODE_UC_SEND_MIDDLE
] = &hfi1_uc_rcv
,
250 [IB_OPCODE_UC_SEND_LAST
] = &hfi1_uc_rcv
,
251 [IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE
] = &hfi1_uc_rcv
,
252 [IB_OPCODE_UC_SEND_ONLY
] = &hfi1_uc_rcv
,
253 [IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE
] = &hfi1_uc_rcv
,
254 [IB_OPCODE_UC_RDMA_WRITE_FIRST
] = &hfi1_uc_rcv
,
255 [IB_OPCODE_UC_RDMA_WRITE_MIDDLE
] = &hfi1_uc_rcv
,
256 [IB_OPCODE_UC_RDMA_WRITE_LAST
] = &hfi1_uc_rcv
,
257 [IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE
] = &hfi1_uc_rcv
,
258 [IB_OPCODE_UC_RDMA_WRITE_ONLY
] = &hfi1_uc_rcv
,
259 [IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE
] = &hfi1_uc_rcv
,
261 [IB_OPCODE_UD_SEND_ONLY
] = &hfi1_ud_rcv
,
262 [IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE
] = &hfi1_ud_rcv
,
264 [IB_OPCODE_CNP
] = &hfi1_cnp_rcv
270 __be64 ib_hfi1_sys_image_guid
;
273 * hfi1_copy_sge - copy data to SGE memory
275 * @data: the data to copy
276 * @length: the length of the data
279 struct hfi1_sge_state
*ss
,
280 void *data
, u32 length
,
283 struct hfi1_sge
*sge
= &ss
->sge
;
286 u32 len
= sge
->length
;
290 if (len
> sge
->sge_length
)
291 len
= sge
->sge_length
;
292 WARN_ON_ONCE(len
== 0);
293 memcpy(sge
->vaddr
, data
, len
);
296 sge
->sge_length
-= len
;
297 if (sge
->sge_length
== 0) {
299 hfi1_put_mr(sge
->mr
);
301 *sge
= *ss
->sg_list
++;
302 } else if (sge
->length
== 0 && sge
->mr
->lkey
) {
303 if (++sge
->n
>= HFI1_SEGSZ
) {
304 if (++sge
->m
>= sge
->mr
->mapsz
)
309 sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].vaddr
;
311 sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].length
;
319 * hfi1_skip_sge - skip over SGE memory
321 * @length: the number of bytes to skip
323 void hfi1_skip_sge(struct hfi1_sge_state
*ss
, u32 length
, int release
)
325 struct hfi1_sge
*sge
= &ss
->sge
;
328 u32 len
= sge
->length
;
332 if (len
> sge
->sge_length
)
333 len
= sge
->sge_length
;
334 WARN_ON_ONCE(len
== 0);
337 sge
->sge_length
-= len
;
338 if (sge
->sge_length
== 0) {
340 hfi1_put_mr(sge
->mr
);
342 *sge
= *ss
->sg_list
++;
343 } else if (sge
->length
== 0 && sge
->mr
->lkey
) {
344 if (++sge
->n
>= HFI1_SEGSZ
) {
345 if (++sge
->m
>= sge
->mr
->mapsz
)
350 sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].vaddr
;
352 sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].length
;
359 * post_one_send - post one RC, UC, or UD send work request
360 * @qp: the QP to post on
361 * @wr: the work request to send
363 static int post_one_send(struct hfi1_qp
*qp
, struct ib_send_wr
*wr
)
365 struct hfi1_swqe
*wqe
;
370 struct hfi1_lkey_table
*rkt
;
372 struct hfi1_devdata
*dd
= dd_from_ibdev(qp
->ibqp
.device
);
373 struct hfi1_pportdata
*ppd
;
374 struct hfi1_ibport
*ibp
;
376 /* IB spec says that num_sge == 0 is OK. */
377 if (unlikely(wr
->num_sge
> qp
->s_max_sge
))
380 ppd
= &dd
->pport
[qp
->port_num
- 1];
381 ibp
= &ppd
->ibport_data
;
384 * Don't allow RDMA reads or atomic operations on UC or
385 * undefined operations.
386 * Make sure buffer is large enough to hold the result for atomics.
388 if (qp
->ibqp
.qp_type
== IB_QPT_UC
) {
389 if ((unsigned) wr
->opcode
>= IB_WR_RDMA_READ
)
391 } else if (qp
->ibqp
.qp_type
!= IB_QPT_RC
) {
392 /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
393 if (wr
->opcode
!= IB_WR_SEND
&&
394 wr
->opcode
!= IB_WR_SEND_WITH_IMM
)
396 /* Check UD destination address PD */
397 if (qp
->ibqp
.pd
!= ud_wr(wr
)->ah
->pd
)
399 } else if ((unsigned) wr
->opcode
> IB_WR_ATOMIC_FETCH_AND_ADD
)
401 else if (wr
->opcode
>= IB_WR_ATOMIC_CMP_AND_SWP
&&
403 wr
->sg_list
[0].length
< sizeof(u64
) ||
404 wr
->sg_list
[0].addr
& (sizeof(u64
) - 1)))
406 else if (wr
->opcode
>= IB_WR_RDMA_READ
&& !qp
->s_max_rd_atomic
)
409 next
= qp
->s_head
+ 1;
410 if (next
>= qp
->s_size
)
412 if (next
== qp
->s_last
)
415 rkt
= &to_idev(qp
->ibqp
.device
)->lk_table
;
416 pd
= to_ipd(qp
->ibqp
.pd
);
417 wqe
= get_swqe_ptr(qp
, qp
->s_head
);
420 if (qp
->ibqp
.qp_type
!= IB_QPT_UC
&&
421 qp
->ibqp
.qp_type
!= IB_QPT_RC
)
422 memcpy(&wqe
->ud_wr
, ud_wr(wr
), sizeof(wqe
->ud_wr
));
423 else if (wr
->opcode
== IB_WR_RDMA_WRITE_WITH_IMM
||
424 wr
->opcode
== IB_WR_RDMA_WRITE
||
425 wr
->opcode
== IB_WR_RDMA_READ
)
426 memcpy(&wqe
->rdma_wr
, rdma_wr(wr
), sizeof(wqe
->rdma_wr
));
427 else if (wr
->opcode
== IB_WR_ATOMIC_CMP_AND_SWP
||
428 wr
->opcode
== IB_WR_ATOMIC_FETCH_AND_ADD
)
429 memcpy(&wqe
->atomic_wr
, atomic_wr(wr
), sizeof(wqe
->atomic_wr
));
431 memcpy(&wqe
->wr
, wr
, sizeof(wqe
->wr
));
436 acc
= wr
->opcode
>= IB_WR_RDMA_READ
?
437 IB_ACCESS_LOCAL_WRITE
: 0;
438 for (i
= 0; i
< wr
->num_sge
; i
++) {
439 u32 length
= wr
->sg_list
[i
].length
;
444 ok
= hfi1_lkey_ok(rkt
, pd
, &wqe
->sg_list
[j
],
445 &wr
->sg_list
[i
], acc
);
447 goto bail_inval_free
;
448 wqe
->length
+= length
;
453 if (qp
->ibqp
.qp_type
== IB_QPT_UC
||
454 qp
->ibqp
.qp_type
== IB_QPT_RC
) {
455 if (wqe
->length
> 0x80000000U
)
456 goto bail_inval_free
;
458 struct hfi1_ah
*ah
= to_iah(ud_wr(wr
)->ah
);
460 atomic_inc(&ah
->refcount
);
462 wqe
->ssn
= qp
->s_ssn
++;
468 /* release mr holds */
470 struct hfi1_sge
*sge
= &wqe
->sg_list
[--j
];
472 hfi1_put_mr(sge
->mr
);
478 * post_send - post a send on a QP
479 * @ibqp: the QP to post the send on
480 * @wr: the list of work requests to post
481 * @bad_wr: the first bad WR is put here
483 * This may be called from interrupt context.
485 static int post_send(struct ib_qp
*ibqp
, struct ib_send_wr
*wr
,
486 struct ib_send_wr
**bad_wr
)
488 struct hfi1_qp
*qp
= to_iqp(ibqp
);
494 spin_lock_irqsave(&qp
->s_lock
, flags
);
496 /* Check that state is OK to post send. */
497 if (unlikely(!(ib_hfi1_state_ops
[qp
->state
] & HFI1_POST_SEND_OK
))) {
498 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
502 /* sq empty and not list -> call send */
503 call_send
= qp
->s_head
== qp
->s_last
&& !wr
->next
;
505 for (; wr
; wr
= wr
->next
) {
506 err
= post_one_send(qp
, wr
);
514 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
515 if (nreq
&& !call_send
)
516 _hfi1_schedule_send(qp
);
517 if (nreq
&& call_send
)
518 hfi1_do_send(&qp
->s_iowait
.iowork
);
523 * post_receive - post a receive on a QP
524 * @ibqp: the QP to post the receive on
525 * @wr: the WR to post
526 * @bad_wr: the first bad WR is put here
528 * This may be called from interrupt context.
530 static int post_receive(struct ib_qp
*ibqp
, struct ib_recv_wr
*wr
,
531 struct ib_recv_wr
**bad_wr
)
533 struct hfi1_qp
*qp
= to_iqp(ibqp
);
534 struct hfi1_rwq
*wq
= qp
->r_rq
.wq
;
538 /* Check that state is OK to post receive. */
539 if (!(ib_hfi1_state_ops
[qp
->state
] & HFI1_POST_RECV_OK
) || !wq
) {
545 for (; wr
; wr
= wr
->next
) {
546 struct hfi1_rwqe
*wqe
;
550 if ((unsigned) wr
->num_sge
> qp
->r_rq
.max_sge
) {
556 spin_lock_irqsave(&qp
->r_rq
.lock
, flags
);
558 if (next
>= qp
->r_rq
.size
)
560 if (next
== wq
->tail
) {
561 spin_unlock_irqrestore(&qp
->r_rq
.lock
, flags
);
567 wqe
= get_rwqe_ptr(&qp
->r_rq
, wq
->head
);
568 wqe
->wr_id
= wr
->wr_id
;
569 wqe
->num_sge
= wr
->num_sge
;
570 for (i
= 0; i
< wr
->num_sge
; i
++)
571 wqe
->sg_list
[i
] = wr
->sg_list
[i
];
572 /* Make sure queue entry is written before the head index. */
575 spin_unlock_irqrestore(&qp
->r_rq
.lock
, flags
);
584 * Make sure the QP is ready and able to accept the given opcode.
586 static inline int qp_ok(int opcode
, struct hfi1_packet
*packet
)
588 struct hfi1_ibport
*ibp
;
590 if (!(ib_hfi1_state_ops
[packet
->qp
->state
] & HFI1_PROCESS_RECV_OK
))
592 if (((opcode
& OPCODE_QP_MASK
) == packet
->qp
->allowed_ops
) ||
593 (opcode
== IB_OPCODE_CNP
))
596 ibp
= &packet
->rcd
->ppd
->ibport_data
;
603 * hfi1_ib_rcv - process an incoming packet
604 * @packet: data packet information
606 * This is called to process an incoming packet at interrupt level.
608 * Tlen is the length of the header + data + CRC in bytes.
610 void hfi1_ib_rcv(struct hfi1_packet
*packet
)
612 struct hfi1_ctxtdata
*rcd
= packet
->rcd
;
613 struct hfi1_ib_header
*hdr
= packet
->hdr
;
614 u32 tlen
= packet
->tlen
;
615 struct hfi1_pportdata
*ppd
= rcd
->ppd
;
616 struct hfi1_ibport
*ibp
= &ppd
->ibport_data
;
624 lnh
= be16_to_cpu(hdr
->lrh
[0]) & 3;
625 if (lnh
== HFI1_LRH_BTH
)
626 packet
->ohdr
= &hdr
->u
.oth
;
627 else if (lnh
== HFI1_LRH_GRH
) {
630 packet
->ohdr
= &hdr
->u
.l
.oth
;
631 if (hdr
->u
.l
.grh
.next_hdr
!= IB_GRH_NEXT_HDR
)
633 vtf
= be32_to_cpu(hdr
->u
.l
.grh
.version_tclass_flow
);
634 if ((vtf
>> IB_GRH_VERSION_SHIFT
) != IB_GRH_VERSION
)
636 packet
->rcv_flags
|= HFI1_HAS_GRH
;
640 trace_input_ibhdr(rcd
->dd
, hdr
);
642 opcode
= (be32_to_cpu(packet
->ohdr
->bth
[0]) >> 24);
643 inc_opstats(tlen
, &rcd
->opstats
->stats
[opcode
]);
645 /* Get the destination QP number. */
646 qp_num
= be32_to_cpu(packet
->ohdr
->bth
[1]) & HFI1_QPN_MASK
;
647 lid
= be16_to_cpu(hdr
->lrh
[1]);
648 if (unlikely((lid
>= HFI1_MULTICAST_LID_BASE
) &&
649 (lid
!= HFI1_PERMISSIVE_LID
))) {
650 struct hfi1_mcast
*mcast
;
651 struct hfi1_mcast_qp
*p
;
653 if (lnh
!= HFI1_LRH_GRH
)
655 mcast
= hfi1_mcast_find(ibp
, &hdr
->u
.l
.grh
.dgid
);
658 list_for_each_entry_rcu(p
, &mcast
->qp_list
, list
) {
660 spin_lock_irqsave(&packet
->qp
->r_lock
, flags
);
661 if (likely((qp_ok(opcode
, packet
))))
662 opcode_handler_tbl
[opcode
](packet
);
663 spin_unlock_irqrestore(&packet
->qp
->r_lock
, flags
);
666 * Notify hfi1_multicast_detach() if it is waiting for us
669 if (atomic_dec_return(&mcast
->refcount
) <= 1)
670 wake_up(&mcast
->wait
);
673 packet
->qp
= hfi1_lookup_qpn(ibp
, qp_num
);
678 spin_lock_irqsave(&packet
->qp
->r_lock
, flags
);
679 if (likely((qp_ok(opcode
, packet
))))
680 opcode_handler_tbl
[opcode
](packet
);
681 spin_unlock_irqrestore(&packet
->qp
->r_lock
, flags
);
691 * This is called from a timer to check for QPs
692 * which need kernel memory in order to send a packet.
694 static void mem_timer(unsigned long data
)
696 struct hfi1_ibdev
*dev
= (struct hfi1_ibdev
*)data
;
697 struct list_head
*list
= &dev
->memwait
;
698 struct hfi1_qp
*qp
= NULL
;
702 write_seqlock_irqsave(&dev
->iowait_lock
, flags
);
703 if (!list_empty(list
)) {
704 wait
= list_first_entry(list
, struct iowait
, list
);
705 qp
= container_of(wait
, struct hfi1_qp
, s_iowait
);
706 list_del_init(&qp
->s_iowait
.list
);
707 /* refcount held until actual wake up */
708 if (!list_empty(list
))
709 mod_timer(&dev
->mem_timer
, jiffies
+ 1);
711 write_sequnlock_irqrestore(&dev
->iowait_lock
, flags
);
714 hfi1_qp_wakeup(qp
, HFI1_S_WAIT_KMEM
);
717 void update_sge(struct hfi1_sge_state
*ss
, u32 length
)
719 struct hfi1_sge
*sge
= &ss
->sge
;
721 sge
->vaddr
+= length
;
722 sge
->length
-= length
;
723 sge
->sge_length
-= length
;
724 if (sge
->sge_length
== 0) {
726 *sge
= *ss
->sg_list
++;
727 } else if (sge
->length
== 0 && sge
->mr
->lkey
) {
728 if (++sge
->n
>= HFI1_SEGSZ
) {
729 if (++sge
->m
>= sge
->mr
->mapsz
)
733 sge
->vaddr
= sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].vaddr
;
734 sge
->length
= sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].length
;
738 static noinline
struct verbs_txreq
*__get_txreq(struct hfi1_ibdev
*dev
,
741 struct verbs_txreq
*tx
;
744 tx
= kmem_cache_alloc(dev
->verbs_txreq_cache
, GFP_ATOMIC
);
746 spin_lock_irqsave(&qp
->s_lock
, flags
);
747 write_seqlock(&dev
->iowait_lock
);
748 if (ib_hfi1_state_ops
[qp
->state
] & HFI1_PROCESS_RECV_OK
&&
749 list_empty(&qp
->s_iowait
.list
)) {
751 qp
->s_flags
|= HFI1_S_WAIT_TX
;
752 list_add_tail(&qp
->s_iowait
.list
, &dev
->txwait
);
753 trace_hfi1_qpsleep(qp
, HFI1_S_WAIT_TX
);
754 atomic_inc(&qp
->refcount
);
756 qp
->s_flags
&= ~HFI1_S_BUSY
;
757 write_sequnlock(&dev
->iowait_lock
);
758 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
759 tx
= ERR_PTR(-EBUSY
);
764 static inline struct verbs_txreq
*get_txreq(struct hfi1_ibdev
*dev
,
767 struct verbs_txreq
*tx
;
769 tx
= kmem_cache_alloc(dev
->verbs_txreq_cache
, GFP_ATOMIC
);
771 /* call slow path to get the lock */
772 tx
= __get_txreq(dev
, qp
);
780 void hfi1_put_txreq(struct verbs_txreq
*tx
)
782 struct hfi1_ibdev
*dev
;
788 dev
= to_idev(qp
->ibqp
.device
);
794 sdma_txclean(dd_from_dev(dev
), &tx
->txreq
);
796 /* Free verbs_txreq and return to slab cache */
797 kmem_cache_free(dev
->verbs_txreq_cache
, tx
);
800 seq
= read_seqbegin(&dev
->iowait_lock
);
801 if (!list_empty(&dev
->txwait
)) {
804 write_seqlock_irqsave(&dev
->iowait_lock
, flags
);
805 /* Wake up first QP wanting a free struct */
806 wait
= list_first_entry(&dev
->txwait
, struct iowait
,
808 qp
= container_of(wait
, struct hfi1_qp
, s_iowait
);
809 list_del_init(&qp
->s_iowait
.list
);
810 /* refcount held until actual wake up */
811 write_sequnlock_irqrestore(&dev
->iowait_lock
, flags
);
812 hfi1_qp_wakeup(qp
, HFI1_S_WAIT_TX
);
815 } while (read_seqretry(&dev
->iowait_lock
, seq
));
819 * This is called with progress side lock held.
822 static void verbs_sdma_complete(
823 struct sdma_txreq
*cookie
,
827 struct verbs_txreq
*tx
=
828 container_of(cookie
, struct verbs_txreq
, txreq
);
829 struct hfi1_qp
*qp
= tx
->qp
;
831 spin_lock(&qp
->s_lock
);
833 hfi1_send_complete(qp
, tx
->wqe
, IB_WC_SUCCESS
);
834 else if (qp
->ibqp
.qp_type
== IB_QPT_RC
) {
835 struct hfi1_ib_header
*hdr
;
838 hfi1_rc_send_complete(qp
, hdr
);
842 * This happens when the send engine notes
843 * a QP in the error state and cannot
844 * do the flush work until that QP's
845 * sdma work has finished.
847 if (qp
->s_flags
& HFI1_S_WAIT_DMA
) {
848 qp
->s_flags
&= ~HFI1_S_WAIT_DMA
;
849 hfi1_schedule_send(qp
);
852 spin_unlock(&qp
->s_lock
);
857 static int wait_kmem(struct hfi1_ibdev
*dev
, struct hfi1_qp
*qp
)
862 spin_lock_irqsave(&qp
->s_lock
, flags
);
863 if (ib_hfi1_state_ops
[qp
->state
] & HFI1_PROCESS_RECV_OK
) {
864 write_seqlock(&dev
->iowait_lock
);
865 if (list_empty(&qp
->s_iowait
.list
)) {
866 if (list_empty(&dev
->memwait
))
867 mod_timer(&dev
->mem_timer
, jiffies
+ 1);
868 qp
->s_flags
|= HFI1_S_WAIT_KMEM
;
869 list_add_tail(&qp
->s_iowait
.list
, &dev
->memwait
);
870 trace_hfi1_qpsleep(qp
, HFI1_S_WAIT_KMEM
);
871 atomic_inc(&qp
->refcount
);
873 write_sequnlock(&dev
->iowait_lock
);
874 qp
->s_flags
&= ~HFI1_S_BUSY
;
877 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
883 * This routine calls txadds for each sg entry.
885 * Add failures will revert the sge cursor
887 static int build_verbs_ulp_payload(
888 struct sdma_engine
*sde
,
889 struct hfi1_sge_state
*ss
,
891 struct verbs_txreq
*tx
)
893 struct hfi1_sge
*sg_list
= ss
->sg_list
;
894 struct hfi1_sge sge
= ss
->sge
;
895 u8 num_sge
= ss
->num_sge
;
900 len
= ss
->sge
.length
;
903 if (len
> ss
->sge
.sge_length
)
904 len
= ss
->sge
.sge_length
;
905 WARN_ON_ONCE(len
== 0);
906 ret
= sdma_txadd_kvaddr(
920 ss
->num_sge
= num_sge
;
921 ss
->sg_list
= sg_list
;
926 * Build the number of DMA descriptors needed to send length bytes of data.
928 * NOTE: DMA mapping is held in the tx until completed in the ring or
929 * the tx desc is freed without having been submitted to the ring
931 * This routine insures the following all the helper routine
935 static int build_verbs_tx_desc(
936 struct sdma_engine
*sde
,
937 struct hfi1_sge_state
*ss
,
939 struct verbs_txreq
*tx
,
940 struct ahg_ib_header
*ahdr
,
944 struct hfi1_pio_header
*phdr
;
945 u16 hdrbytes
= tx
->hdr_dwords
<< 2;
948 if (!ahdr
->ahgcount
) {
949 ret
= sdma_txinit_ahg(
957 verbs_sdma_complete
);
960 phdr
->pbc
= cpu_to_le64(pbc
);
961 memcpy(&phdr
->hdr
, &ahdr
->ibh
, hdrbytes
- sizeof(phdr
->pbc
));
963 ret
= sdma_txadd_kvaddr(
967 tx
->hdr_dwords
<< 2);
971 struct hfi1_other_headers
*sohdr
= &ahdr
->ibh
.u
.oth
;
972 struct hfi1_other_headers
*dohdr
= &phdr
->hdr
.u
.oth
;
974 /* needed in rc_send_complete() */
975 phdr
->hdr
.lrh
[0] = ahdr
->ibh
.lrh
[0];
976 if ((be16_to_cpu(phdr
->hdr
.lrh
[0]) & 3) == HFI1_LRH_GRH
) {
977 sohdr
= &ahdr
->ibh
.u
.l
.oth
;
978 dohdr
= &phdr
->hdr
.u
.l
.oth
;
981 dohdr
->bth
[0] = sohdr
->bth
[0];
983 dohdr
->bth
[2] = sohdr
->bth
[2];
984 ret
= sdma_txinit_ahg(
992 verbs_sdma_complete
);
997 /* add the ulp payload - if any. ss can be NULL for acks */
999 ret
= build_verbs_ulp_payload(sde
, ss
, length
, tx
);
1004 int hfi1_verbs_send_dma(struct hfi1_qp
*qp
, struct hfi1_pkt_state
*ps
,
1007 struct ahg_ib_header
*ahdr
= qp
->s_hdr
;
1008 u32 hdrwords
= qp
->s_hdrwords
;
1009 struct hfi1_sge_state
*ss
= qp
->s_cur_sge
;
1010 u32 len
= qp
->s_cur_size
;
1011 u32 plen
= hdrwords
+ ((len
+ 3) >> 2) + 2; /* includes pbc */
1012 struct hfi1_ibdev
*dev
= ps
->dev
;
1013 struct hfi1_pportdata
*ppd
= ps
->ppd
;
1014 struct verbs_txreq
*tx
;
1015 struct sdma_txreq
*stx
;
1020 if (!list_empty(&qp
->s_iowait
.tx_head
)) {
1021 stx
= list_first_entry(
1022 &qp
->s_iowait
.tx_head
,
1025 list_del_init(&stx
->list
);
1026 tx
= container_of(stx
, struct verbs_txreq
, txreq
);
1027 ret
= sdma_send_txreq(tx
->sde
, &qp
->s_iowait
, stx
);
1028 if (unlikely(ret
== -ECOMM
))
1033 tx
= get_txreq(dev
, qp
);
1037 tx
->sde
= qp
->s_sde
;
1039 if (likely(pbc
== 0)) {
1040 u32 vl
= sc_to_vlt(dd_from_ibdev(qp
->ibqp
.device
), sc5
);
1042 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1043 pbc_flags
|= (!!(sc5
& 0x10)) << PBC_DC_INFO_SHIFT
;
1045 pbc
= create_pbc(ppd
, pbc_flags
, qp
->srate_mbps
, vl
, plen
);
1047 tx
->wqe
= qp
->s_wqe
;
1048 tx
->mr
= qp
->s_rdma_mr
;
1050 qp
->s_rdma_mr
= NULL
;
1051 tx
->hdr_dwords
= hdrwords
+ 2;
1052 ret
= build_verbs_tx_desc(tx
->sde
, ss
, len
, tx
, ahdr
, pbc
);
1055 trace_output_ibhdr(dd_from_ibdev(qp
->ibqp
.device
), &ahdr
->ibh
);
1056 ret
= sdma_send_txreq(tx
->sde
, &qp
->s_iowait
, &tx
->txreq
);
1057 if (unlikely(ret
== -ECOMM
))
1062 /* The current one got "sent" */
1065 /* kmalloc or mapping fail */
1067 return wait_kmem(dev
, qp
);
1073 * If we are now in the error state, return zero to flush the
1074 * send work request.
1076 static int no_bufs_available(struct hfi1_qp
*qp
, struct send_context
*sc
)
1078 struct hfi1_devdata
*dd
= sc
->dd
;
1079 struct hfi1_ibdev
*dev
= &dd
->verbs_dev
;
1080 unsigned long flags
;
1084 * Note that as soon as want_buffer() is called and
1085 * possibly before it returns, sc_piobufavail()
1086 * could be called. Therefore, put QP on the I/O wait list before
1087 * enabling the PIO avail interrupt.
1089 spin_lock_irqsave(&qp
->s_lock
, flags
);
1090 if (ib_hfi1_state_ops
[qp
->state
] & HFI1_PROCESS_RECV_OK
) {
1091 write_seqlock(&dev
->iowait_lock
);
1092 if (list_empty(&qp
->s_iowait
.list
)) {
1093 struct hfi1_ibdev
*dev
= &dd
->verbs_dev
;
1097 qp
->s_flags
|= HFI1_S_WAIT_PIO
;
1098 was_empty
= list_empty(&sc
->piowait
);
1099 list_add_tail(&qp
->s_iowait
.list
, &sc
->piowait
);
1100 trace_hfi1_qpsleep(qp
, HFI1_S_WAIT_PIO
);
1101 atomic_inc(&qp
->refcount
);
1102 /* counting: only call wantpiobuf_intr if first user */
1104 hfi1_sc_wantpiobuf_intr(sc
, 1);
1106 write_sequnlock(&dev
->iowait_lock
);
1107 qp
->s_flags
&= ~HFI1_S_BUSY
;
1110 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
1114 struct send_context
*qp_to_send_context(struct hfi1_qp
*qp
, u8 sc5
)
1116 struct hfi1_devdata
*dd
= dd_from_ibdev(qp
->ibqp
.device
);
1117 struct hfi1_pportdata
*ppd
= dd
->pport
+ (qp
->port_num
- 1);
1120 vl
= sc_to_vlt(dd
, sc5
);
1121 if (vl
>= ppd
->vls_supported
&& vl
!= 15)
1123 return dd
->vld
[vl
].sc
;
1126 int hfi1_verbs_send_pio(struct hfi1_qp
*qp
, struct hfi1_pkt_state
*ps
,
1129 struct ahg_ib_header
*ahdr
= qp
->s_hdr
;
1130 u32 hdrwords
= qp
->s_hdrwords
;
1131 struct hfi1_sge_state
*ss
= qp
->s_cur_sge
;
1132 u32 len
= qp
->s_cur_size
;
1133 u32 dwords
= (len
+ 3) >> 2;
1134 u32 plen
= hdrwords
+ dwords
+ 2; /* includes pbc */
1135 struct hfi1_pportdata
*ppd
= ps
->ppd
;
1136 u32
*hdr
= (u32
*)&ahdr
->ibh
;
1139 unsigned long flags
= 0;
1140 struct send_context
*sc
;
1141 struct pio_buf
*pbuf
;
1142 int wc_status
= IB_WC_SUCCESS
;
1144 /* vl15 special case taken care of in ud.c */
1146 sc
= qp_to_send_context(qp
, sc5
);
1150 if (likely(pbc
== 0)) {
1151 u32 vl
= sc_to_vlt(dd_from_ibdev(qp
->ibqp
.device
), sc5
);
1152 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1153 pbc_flags
|= (!!(sc5
& 0x10)) << PBC_DC_INFO_SHIFT
;
1154 pbc
= create_pbc(ppd
, pbc_flags
, qp
->srate_mbps
, vl
, plen
);
1156 pbuf
= sc_buffer_alloc(sc
, plen
, NULL
, NULL
);
1157 if (unlikely(pbuf
== NULL
)) {
1158 if (ppd
->host_link_state
!= HLS_UP_ACTIVE
) {
1160 * If we have filled the PIO buffers to capacity and are
1161 * not in an active state this request is not going to
1162 * go out to so just complete it with an error or else a
1163 * ULP or the core may be stuck waiting.
1167 "alloc failed. state not active, completing");
1168 wc_status
= IB_WC_GENERAL_ERR
;
1172 * This is a normal occurrence. The PIO buffs are full
1173 * up but we are still happily sending, well we could be
1174 * so lets continue to queue the request.
1176 hfi1_cdbg(PIO
, "alloc failed. state active, queuing");
1177 return no_bufs_available(qp
, sc
);
1182 pio_copy(ppd
->dd
, pbuf
, pbc
, hdr
, hdrwords
);
1185 seg_pio_copy_start(pbuf
, pbc
, hdr
, hdrwords
*4);
1187 void *addr
= ss
->sge
.vaddr
;
1188 u32 slen
= ss
->sge
.length
;
1192 update_sge(ss
, slen
);
1193 seg_pio_copy_mid(pbuf
, addr
, slen
);
1196 seg_pio_copy_end(pbuf
);
1200 trace_output_ibhdr(dd_from_ibdev(qp
->ibqp
.device
), &ahdr
->ibh
);
1202 if (qp
->s_rdma_mr
) {
1203 hfi1_put_mr(qp
->s_rdma_mr
);
1204 qp
->s_rdma_mr
= NULL
;
1209 spin_lock_irqsave(&qp
->s_lock
, flags
);
1210 hfi1_send_complete(qp
, qp
->s_wqe
, wc_status
);
1211 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
1212 } else if (qp
->ibqp
.qp_type
== IB_QPT_RC
) {
1213 spin_lock_irqsave(&qp
->s_lock
, flags
);
1214 hfi1_rc_send_complete(qp
, &ahdr
->ibh
);
1215 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
1221 * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1222 * being an entry from the ingress partition key table), return 0
1223 * otherwise. Use the matching criteria for egress partition keys
1224 * specified in the OPAv1 spec., section 9.1l.7.
1226 static inline int egress_pkey_matches_entry(u16 pkey
, u16 ent
)
1228 u16 mkey
= pkey
& PKEY_LOW_15_MASK
;
1229 u16 ment
= ent
& PKEY_LOW_15_MASK
;
1233 * If pkey[15] is set (full partition member),
1234 * is bit 15 in the corresponding table element
1235 * clear (limited member)?
1237 if (pkey
& PKEY_MEMBER_MASK
)
1238 return !!(ent
& PKEY_MEMBER_MASK
);
1245 * egress_pkey_check - return 0 if hdr's pkey matches according to the
1246 * criteria in the OPAv1 spec., section 9.11.7.
1248 static inline int egress_pkey_check(struct hfi1_pportdata
*ppd
,
1249 struct hfi1_ib_header
*hdr
,
1252 struct hfi1_other_headers
*ohdr
;
1253 struct hfi1_devdata
*dd
;
1256 u8 lnh
, sc5
= qp
->s_sc
;
1258 if (!(ppd
->part_enforce
& HFI1_PART_ENFORCE_OUT
))
1261 /* locate the pkey within the headers */
1262 lnh
= be16_to_cpu(hdr
->lrh
[0]) & 3;
1263 if (lnh
== HFI1_LRH_GRH
)
1264 ohdr
= &hdr
->u
.l
.oth
;
1268 pkey
= (u16
)be32_to_cpu(ohdr
->bth
[0]);
1270 /* If SC15, pkey[0:14] must be 0x7fff */
1271 if ((sc5
== 0xf) && ((pkey
& PKEY_LOW_15_MASK
) != PKEY_LOW_15_MASK
))
1275 /* Is the pkey = 0x0, or 0x8000? */
1276 if ((pkey
& PKEY_LOW_15_MASK
) == 0)
1279 /* The most likely matching pkey has index qp->s_pkey_index */
1280 if (unlikely(!egress_pkey_matches_entry(pkey
,
1281 ppd
->pkeys
[qp
->s_pkey_index
]))) {
1282 /* no match - try the entire table */
1283 for (; i
< MAX_PKEY_VALUES
; i
++) {
1284 if (egress_pkey_matches_entry(pkey
, ppd
->pkeys
[i
]))
1289 if (i
< MAX_PKEY_VALUES
)
1292 incr_cntr64(&ppd
->port_xmit_constraint_errors
);
1294 if (!(dd
->err_info_xmit_constraint
.status
& OPA_EI_STATUS_SMASK
)) {
1295 u16 slid
= be16_to_cpu(hdr
->lrh
[3]);
1297 dd
->err_info_xmit_constraint
.status
|= OPA_EI_STATUS_SMASK
;
1298 dd
->err_info_xmit_constraint
.slid
= slid
;
1299 dd
->err_info_xmit_constraint
.pkey
= pkey
;
1305 * hfi1_verbs_send - send a packet
1306 * @qp: the QP to send on
1307 * @ps: the state of the packet to send
1309 * Return zero if packet is sent or queued OK.
1310 * Return non-zero and clear qp->s_flags HFI1_S_BUSY otherwise.
1312 int hfi1_verbs_send(struct hfi1_qp
*qp
, struct hfi1_pkt_state
*ps
)
1314 struct hfi1_devdata
*dd
= dd_from_ibdev(qp
->ibqp
.device
);
1315 struct ahg_ib_header
*ahdr
= qp
->s_hdr
;
1318 unsigned long flags
= 0;
1321 * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1322 * can defer SDMA restart until link goes ACTIVE without
1323 * worrying about just how we got there.
1325 if ((qp
->ibqp
.qp_type
== IB_QPT_SMI
) ||
1326 !(dd
->flags
& HFI1_HAS_SEND_DMA
))
1329 ret
= egress_pkey_check(dd
->pport
, &ahdr
->ibh
, qp
);
1330 if (unlikely(ret
)) {
1332 * The value we are returning here does not get propagated to
1333 * the verbs caller. Thus we need to complete the request with
1334 * error otherwise the caller could be sitting waiting on the
1335 * completion event. Only do this for PIO. SDMA has its own
1336 * mechanism for handling the errors. So for SDMA we can just
1340 hfi1_cdbg(PIO
, "%s() Failed. Completing with err",
1342 spin_lock_irqsave(&qp
->s_lock
, flags
);
1343 hfi1_send_complete(qp
, qp
->s_wqe
, IB_WC_GENERAL_ERR
);
1344 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
1350 ret
= dd
->process_pio_send(qp
, ps
, 0);
1352 #ifdef CONFIG_SDMA_VERBOSITY
1353 dd_dev_err(dd
, "CONFIG SDMA %s:%d %s()\n",
1354 slashstrip(__FILE__
), __LINE__
, __func__
);
1355 dd_dev_err(dd
, "SDMA hdrwords = %u, len = %u\n", qp
->s_hdrwords
,
1358 ret
= dd
->process_dma_send(qp
, ps
, 0);
1364 static int query_device(struct ib_device
*ibdev
,
1365 struct ib_device_attr
*props
,
1366 struct ib_udata
*uhw
)
1368 struct hfi1_devdata
*dd
= dd_from_ibdev(ibdev
);
1369 struct hfi1_ibdev
*dev
= to_idev(ibdev
);
1371 if (uhw
->inlen
|| uhw
->outlen
)
1373 memset(props
, 0, sizeof(*props
));
1375 props
->device_cap_flags
= IB_DEVICE_BAD_PKEY_CNTR
|
1376 IB_DEVICE_BAD_QKEY_CNTR
| IB_DEVICE_SHUTDOWN_PORT
|
1377 IB_DEVICE_SYS_IMAGE_GUID
| IB_DEVICE_RC_RNR_NAK_GEN
|
1378 IB_DEVICE_PORT_ACTIVE_EVENT
| IB_DEVICE_SRQ_RESIZE
;
1380 props
->page_size_cap
= PAGE_SIZE
;
1382 dd
->oui1
<< 16 | dd
->oui2
<< 8 | dd
->oui3
;
1383 props
->vendor_part_id
= dd
->pcidev
->device
;
1384 props
->hw_ver
= dd
->minrev
;
1385 props
->sys_image_guid
= ib_hfi1_sys_image_guid
;
1386 props
->max_mr_size
= ~0ULL;
1387 props
->max_qp
= hfi1_max_qps
;
1388 props
->max_qp_wr
= hfi1_max_qp_wrs
;
1389 props
->max_sge
= hfi1_max_sges
;
1390 props
->max_sge_rd
= hfi1_max_sges
;
1391 props
->max_cq
= hfi1_max_cqs
;
1392 props
->max_ah
= hfi1_max_ahs
;
1393 props
->max_cqe
= hfi1_max_cqes
;
1394 props
->max_mr
= dev
->lk_table
.max
;
1395 props
->max_fmr
= dev
->lk_table
.max
;
1396 props
->max_map_per_fmr
= 32767;
1397 props
->max_pd
= hfi1_max_pds
;
1398 props
->max_qp_rd_atom
= HFI1_MAX_RDMA_ATOMIC
;
1399 props
->max_qp_init_rd_atom
= 255;
1400 /* props->max_res_rd_atom */
1401 props
->max_srq
= hfi1_max_srqs
;
1402 props
->max_srq_wr
= hfi1_max_srq_wrs
;
1403 props
->max_srq_sge
= hfi1_max_srq_sges
;
1404 /* props->local_ca_ack_delay */
1405 props
->atomic_cap
= IB_ATOMIC_GLOB
;
1406 props
->max_pkeys
= hfi1_get_npkeys(dd
);
1407 props
->max_mcast_grp
= hfi1_max_mcast_grps
;
1408 props
->max_mcast_qp_attach
= hfi1_max_mcast_qp_attached
;
1409 props
->max_total_mcast_qp_attach
= props
->max_mcast_qp_attach
*
1410 props
->max_mcast_grp
;
1415 static inline u16
opa_speed_to_ib(u16 in
)
1419 if (in
& OPA_LINK_SPEED_25G
)
1420 out
|= IB_SPEED_EDR
;
1421 if (in
& OPA_LINK_SPEED_12_5G
)
1422 out
|= IB_SPEED_FDR
;
1428 * Convert a single OPA link width (no multiple flags) to an IB value.
1429 * A zero OPA link width means link down, which means the IB width value
1432 static inline u16
opa_width_to_ib(u16 in
)
1435 case OPA_LINK_WIDTH_1X
:
1436 /* map 2x and 3x to 1x as they don't exist in IB */
1437 case OPA_LINK_WIDTH_2X
:
1438 case OPA_LINK_WIDTH_3X
:
1440 default: /* link down or unknown, return our largest width */
1441 case OPA_LINK_WIDTH_4X
:
1446 static int query_port(struct ib_device
*ibdev
, u8 port
,
1447 struct ib_port_attr
*props
)
1449 struct hfi1_devdata
*dd
= dd_from_ibdev(ibdev
);
1450 struct hfi1_ibport
*ibp
= to_iport(ibdev
, port
);
1451 struct hfi1_pportdata
*ppd
= ppd_from_ibp(ibp
);
1454 memset(props
, 0, sizeof(*props
));
1455 props
->lid
= lid
? lid
: 0;
1456 props
->lmc
= ppd
->lmc
;
1457 props
->sm_lid
= ibp
->sm_lid
;
1458 props
->sm_sl
= ibp
->sm_sl
;
1459 /* OPA logical states match IB logical states */
1460 props
->state
= driver_lstate(ppd
);
1461 props
->phys_state
= hfi1_ibphys_portstate(ppd
);
1462 props
->port_cap_flags
= ibp
->port_cap_flags
;
1463 props
->gid_tbl_len
= HFI1_GUIDS_PER_PORT
;
1464 props
->max_msg_sz
= 0x80000000;
1465 props
->pkey_tbl_len
= hfi1_get_npkeys(dd
);
1466 props
->bad_pkey_cntr
= ibp
->pkey_violations
;
1467 props
->qkey_viol_cntr
= ibp
->qkey_violations
;
1468 props
->active_width
= (u8
)opa_width_to_ib(ppd
->link_width_active
);
1469 /* see rate_show() in ib core/sysfs.c */
1470 props
->active_speed
= (u8
)opa_speed_to_ib(ppd
->link_speed_active
);
1471 props
->max_vl_num
= ppd
->vls_supported
;
1472 props
->init_type_reply
= 0;
1474 /* Once we are a "first class" citizen and have added the OPA MTUs to
1475 * the core we can advertise the larger MTU enum to the ULPs, for now
1476 * advertise only 4K.
1478 * Those applications which are either OPA aware or pass the MTU enum
1479 * from the Path Records to us will get the new 8k MTU. Those that
1480 * attempt to process the MTU enum may fail in various ways.
1482 props
->max_mtu
= mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu
) ?
1483 4096 : hfi1_max_mtu
), IB_MTU_4096
);
1484 props
->active_mtu
= !valid_ib_mtu(ppd
->ibmtu
) ? props
->max_mtu
:
1485 mtu_to_enum(ppd
->ibmtu
, IB_MTU_2048
);
1486 props
->subnet_timeout
= ibp
->subnet_timeout
;
1491 static int port_immutable(struct ib_device
*ibdev
, u8 port_num
,
1492 struct ib_port_immutable
*immutable
)
1494 struct ib_port_attr attr
;
1497 err
= query_port(ibdev
, port_num
, &attr
);
1501 memset(immutable
, 0, sizeof(*immutable
));
1503 immutable
->pkey_tbl_len
= attr
.pkey_tbl_len
;
1504 immutable
->gid_tbl_len
= attr
.gid_tbl_len
;
1505 immutable
->core_cap_flags
= RDMA_CORE_PORT_INTEL_OPA
;
1506 immutable
->max_mad_size
= OPA_MGMT_MAD_SIZE
;
1511 static int modify_device(struct ib_device
*device
,
1512 int device_modify_mask
,
1513 struct ib_device_modify
*device_modify
)
1515 struct hfi1_devdata
*dd
= dd_from_ibdev(device
);
1519 if (device_modify_mask
& ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID
|
1520 IB_DEVICE_MODIFY_NODE_DESC
)) {
1525 if (device_modify_mask
& IB_DEVICE_MODIFY_NODE_DESC
) {
1526 memcpy(device
->node_desc
, device_modify
->node_desc
, 64);
1527 for (i
= 0; i
< dd
->num_pports
; i
++) {
1528 struct hfi1_ibport
*ibp
= &dd
->pport
[i
].ibport_data
;
1530 hfi1_node_desc_chg(ibp
);
1534 if (device_modify_mask
& IB_DEVICE_MODIFY_SYS_IMAGE_GUID
) {
1535 ib_hfi1_sys_image_guid
=
1536 cpu_to_be64(device_modify
->sys_image_guid
);
1537 for (i
= 0; i
< dd
->num_pports
; i
++) {
1538 struct hfi1_ibport
*ibp
= &dd
->pport
[i
].ibport_data
;
1540 hfi1_sys_guid_chg(ibp
);
1550 static int modify_port(struct ib_device
*ibdev
, u8 port
,
1551 int port_modify_mask
, struct ib_port_modify
*props
)
1553 struct hfi1_ibport
*ibp
= to_iport(ibdev
, port
);
1554 struct hfi1_pportdata
*ppd
= ppd_from_ibp(ibp
);
1557 ibp
->port_cap_flags
|= props
->set_port_cap_mask
;
1558 ibp
->port_cap_flags
&= ~props
->clr_port_cap_mask
;
1559 if (props
->set_port_cap_mask
|| props
->clr_port_cap_mask
)
1560 hfi1_cap_mask_chg(ibp
);
1561 if (port_modify_mask
& IB_PORT_SHUTDOWN
) {
1562 set_link_down_reason(ppd
, OPA_LINKDOWN_REASON_UNKNOWN
, 0,
1563 OPA_LINKDOWN_REASON_UNKNOWN
);
1564 ret
= set_link_state(ppd
, HLS_DN_DOWNDEF
);
1566 if (port_modify_mask
& IB_PORT_RESET_QKEY_CNTR
)
1567 ibp
->qkey_violations
= 0;
1571 static int query_gid(struct ib_device
*ibdev
, u8 port
,
1572 int index
, union ib_gid
*gid
)
1574 struct hfi1_devdata
*dd
= dd_from_ibdev(ibdev
);
1577 if (!port
|| port
> dd
->num_pports
)
1580 struct hfi1_ibport
*ibp
= to_iport(ibdev
, port
);
1581 struct hfi1_pportdata
*ppd
= ppd_from_ibp(ibp
);
1583 gid
->global
.subnet_prefix
= ibp
->gid_prefix
;
1585 gid
->global
.interface_id
= cpu_to_be64(ppd
->guid
);
1586 else if (index
< HFI1_GUIDS_PER_PORT
)
1587 gid
->global
.interface_id
= ibp
->guids
[index
- 1];
1595 static struct ib_pd
*alloc_pd(struct ib_device
*ibdev
,
1596 struct ib_ucontext
*context
,
1597 struct ib_udata
*udata
)
1599 struct hfi1_ibdev
*dev
= to_idev(ibdev
);
1604 * This is actually totally arbitrary. Some correctness tests
1605 * assume there's a maximum number of PDs that can be allocated.
1606 * We don't actually have this limit, but we fail the test if
1607 * we allow allocations of more than we report for this value.
1610 pd
= kmalloc(sizeof(*pd
), GFP_KERNEL
);
1612 ret
= ERR_PTR(-ENOMEM
);
1616 spin_lock(&dev
->n_pds_lock
);
1617 if (dev
->n_pds_allocated
== hfi1_max_pds
) {
1618 spin_unlock(&dev
->n_pds_lock
);
1620 ret
= ERR_PTR(-ENOMEM
);
1624 dev
->n_pds_allocated
++;
1625 spin_unlock(&dev
->n_pds_lock
);
1627 /* ib_alloc_pd() will initialize pd->ibpd. */
1628 pd
->user
= udata
!= NULL
;
1636 static int dealloc_pd(struct ib_pd
*ibpd
)
1638 struct hfi1_pd
*pd
= to_ipd(ibpd
);
1639 struct hfi1_ibdev
*dev
= to_idev(ibpd
->device
);
1641 spin_lock(&dev
->n_pds_lock
);
1642 dev
->n_pds_allocated
--;
1643 spin_unlock(&dev
->n_pds_lock
);
1651 * convert ah port,sl to sc
1653 u8
ah_to_sc(struct ib_device
*ibdev
, struct ib_ah_attr
*ah
)
1655 struct hfi1_ibport
*ibp
= to_iport(ibdev
, ah
->port_num
);
1657 return ibp
->sl_to_sc
[ah
->sl
];
1660 int hfi1_check_ah(struct ib_device
*ibdev
, struct ib_ah_attr
*ah_attr
)
1662 struct hfi1_ibport
*ibp
;
1663 struct hfi1_pportdata
*ppd
;
1664 struct hfi1_devdata
*dd
;
1667 /* A multicast address requires a GRH (see ch. 8.4.1). */
1668 if (ah_attr
->dlid
>= HFI1_MULTICAST_LID_BASE
&&
1669 ah_attr
->dlid
!= HFI1_PERMISSIVE_LID
&&
1670 !(ah_attr
->ah_flags
& IB_AH_GRH
))
1672 if ((ah_attr
->ah_flags
& IB_AH_GRH
) &&
1673 ah_attr
->grh
.sgid_index
>= HFI1_GUIDS_PER_PORT
)
1675 if (ah_attr
->dlid
== 0)
1677 if (ah_attr
->port_num
< 1 ||
1678 ah_attr
->port_num
> ibdev
->phys_port_cnt
)
1680 if (ah_attr
->static_rate
!= IB_RATE_PORT_CURRENT
&&
1681 ib_rate_to_mbps(ah_attr
->static_rate
) < 0)
1683 if (ah_attr
->sl
>= OPA_MAX_SLS
)
1685 /* test the mapping for validity */
1686 ibp
= to_iport(ibdev
, ah_attr
->port_num
);
1687 ppd
= ppd_from_ibp(ibp
);
1688 sc5
= ibp
->sl_to_sc
[ah_attr
->sl
];
1689 dd
= dd_from_ppd(ppd
);
1690 if (sc_to_vlt(dd
, sc5
) > num_vls
&& sc_to_vlt(dd
, sc5
) != 0xf)
1698 * create_ah - create an address handle
1699 * @pd: the protection domain
1700 * @ah_attr: the attributes of the AH
1702 * This may be called from interrupt context.
1704 static struct ib_ah
*create_ah(struct ib_pd
*pd
,
1705 struct ib_ah_attr
*ah_attr
)
1709 struct hfi1_ibdev
*dev
= to_idev(pd
->device
);
1710 unsigned long flags
;
1712 if (hfi1_check_ah(pd
->device
, ah_attr
)) {
1713 ret
= ERR_PTR(-EINVAL
);
1717 ah
= kmalloc(sizeof(*ah
), GFP_ATOMIC
);
1719 ret
= ERR_PTR(-ENOMEM
);
1723 spin_lock_irqsave(&dev
->n_ahs_lock
, flags
);
1724 if (dev
->n_ahs_allocated
== hfi1_max_ahs
) {
1725 spin_unlock_irqrestore(&dev
->n_ahs_lock
, flags
);
1727 ret
= ERR_PTR(-ENOMEM
);
1731 dev
->n_ahs_allocated
++;
1732 spin_unlock_irqrestore(&dev
->n_ahs_lock
, flags
);
1734 /* ib_create_ah() will initialize ah->ibah. */
1735 ah
->attr
= *ah_attr
;
1736 atomic_set(&ah
->refcount
, 0);
1744 struct ib_ah
*hfi1_create_qp0_ah(struct hfi1_ibport
*ibp
, u16 dlid
)
1746 struct ib_ah_attr attr
;
1747 struct ib_ah
*ah
= ERR_PTR(-EINVAL
);
1748 struct hfi1_qp
*qp0
;
1750 memset(&attr
, 0, sizeof(attr
));
1752 attr
.port_num
= ppd_from_ibp(ibp
)->port
;
1754 qp0
= rcu_dereference(ibp
->qp
[0]);
1756 ah
= ib_create_ah(qp0
->ibqp
.pd
, &attr
);
1762 * destroy_ah - destroy an address handle
1763 * @ibah: the AH to destroy
1765 * This may be called from interrupt context.
1767 static int destroy_ah(struct ib_ah
*ibah
)
1769 struct hfi1_ibdev
*dev
= to_idev(ibah
->device
);
1770 struct hfi1_ah
*ah
= to_iah(ibah
);
1771 unsigned long flags
;
1773 if (atomic_read(&ah
->refcount
) != 0)
1776 spin_lock_irqsave(&dev
->n_ahs_lock
, flags
);
1777 dev
->n_ahs_allocated
--;
1778 spin_unlock_irqrestore(&dev
->n_ahs_lock
, flags
);
1785 static int modify_ah(struct ib_ah
*ibah
, struct ib_ah_attr
*ah_attr
)
1787 struct hfi1_ah
*ah
= to_iah(ibah
);
1789 if (hfi1_check_ah(ibah
->device
, ah_attr
))
1792 ah
->attr
= *ah_attr
;
1797 static int query_ah(struct ib_ah
*ibah
, struct ib_ah_attr
*ah_attr
)
1799 struct hfi1_ah
*ah
= to_iah(ibah
);
1801 *ah_attr
= ah
->attr
;
1807 * hfi1_get_npkeys - return the size of the PKEY table for context 0
1808 * @dd: the hfi1_ib device
1810 unsigned hfi1_get_npkeys(struct hfi1_devdata
*dd
)
1812 return ARRAY_SIZE(dd
->pport
[0].pkeys
);
1815 static int query_pkey(struct ib_device
*ibdev
, u8 port
, u16 index
,
1818 struct hfi1_devdata
*dd
= dd_from_ibdev(ibdev
);
1821 if (index
>= hfi1_get_npkeys(dd
)) {
1826 *pkey
= hfi1_get_pkey(to_iport(ibdev
, port
), index
);
1834 * alloc_ucontext - allocate a ucontest
1835 * @ibdev: the infiniband device
1836 * @udata: not used by the driver
1839 static struct ib_ucontext
*alloc_ucontext(struct ib_device
*ibdev
,
1840 struct ib_udata
*udata
)
1842 struct hfi1_ucontext
*context
;
1843 struct ib_ucontext
*ret
;
1845 context
= kmalloc(sizeof(*context
), GFP_KERNEL
);
1847 ret
= ERR_PTR(-ENOMEM
);
1851 ret
= &context
->ibucontext
;
1857 static int dealloc_ucontext(struct ib_ucontext
*context
)
1859 kfree(to_iucontext(context
));
1863 static void init_ibport(struct hfi1_pportdata
*ppd
)
1865 struct hfi1_ibport
*ibp
= &ppd
->ibport_data
;
1866 size_t sz
= ARRAY_SIZE(ibp
->sl_to_sc
);
1869 for (i
= 0; i
< sz
; i
++) {
1870 ibp
->sl_to_sc
[i
] = i
;
1871 ibp
->sc_to_sl
[i
] = i
;
1874 spin_lock_init(&ibp
->lock
);
1875 /* Set the prefix to the default value (see ch. 4.1.1) */
1876 ibp
->gid_prefix
= IB_DEFAULT_GID_PREFIX
;
1878 /* Below should only set bits defined in OPA PortInfo.CapabilityMask */
1879 ibp
->port_cap_flags
= IB_PORT_AUTO_MIGR_SUP
|
1880 IB_PORT_CAP_MASK_NOTICE_SUP
;
1881 ibp
->pma_counter_select
[0] = IB_PMA_PORT_XMIT_DATA
;
1882 ibp
->pma_counter_select
[1] = IB_PMA_PORT_RCV_DATA
;
1883 ibp
->pma_counter_select
[2] = IB_PMA_PORT_XMIT_PKTS
;
1884 ibp
->pma_counter_select
[3] = IB_PMA_PORT_RCV_PKTS
;
1885 ibp
->pma_counter_select
[4] = IB_PMA_PORT_XMIT_WAIT
;
1887 RCU_INIT_POINTER(ibp
->qp
[0], NULL
);
1888 RCU_INIT_POINTER(ibp
->qp
[1], NULL
);
1891 static void verbs_txreq_kmem_cache_ctor(void *obj
)
1893 struct verbs_txreq
*tx
= obj
;
1895 memset(tx
, 0, sizeof(*tx
));
1899 * hfi1_register_ib_device - register our device with the infiniband core
1900 * @dd: the device data structure
1901 * Return 0 if successful, errno if unsuccessful.
1903 int hfi1_register_ib_device(struct hfi1_devdata
*dd
)
1905 struct hfi1_ibdev
*dev
= &dd
->verbs_dev
;
1906 struct ib_device
*ibdev
= &dev
->ibdev
;
1907 struct hfi1_pportdata
*ppd
= dd
->pport
;
1908 unsigned i
, lk_tab_size
;
1910 size_t lcpysz
= IB_DEVICE_NAME_MAX
;
1912 char buf
[TXREQ_NAME_LEN
];
1914 ret
= hfi1_qp_init(dev
);
1919 for (i
= 0; i
< dd
->num_pports
; i
++)
1920 init_ibport(ppd
+ i
);
1922 /* Only need to initialize non-zero fields. */
1923 spin_lock_init(&dev
->n_pds_lock
);
1924 spin_lock_init(&dev
->n_ahs_lock
);
1925 spin_lock_init(&dev
->n_cqs_lock
);
1926 spin_lock_init(&dev
->n_qps_lock
);
1927 spin_lock_init(&dev
->n_srqs_lock
);
1928 spin_lock_init(&dev
->n_mcast_grps_lock
);
1929 init_timer(&dev
->mem_timer
);
1930 dev
->mem_timer
.function
= mem_timer
;
1931 dev
->mem_timer
.data
= (unsigned long) dev
;
1934 * The top hfi1_lkey_table_size bits are used to index the
1935 * table. The lower 8 bits can be owned by the user (copied from
1936 * the LKEY). The remaining bits act as a generation number or tag.
1938 spin_lock_init(&dev
->lk_table
.lock
);
1939 dev
->lk_table
.max
= 1 << hfi1_lkey_table_size
;
1940 /* ensure generation is at least 4 bits (keys.c) */
1941 if (hfi1_lkey_table_size
> MAX_LKEY_TABLE_BITS
) {
1942 dd_dev_warn(dd
, "lkey bits %u too large, reduced to %u\n",
1943 hfi1_lkey_table_size
, MAX_LKEY_TABLE_BITS
);
1944 hfi1_lkey_table_size
= MAX_LKEY_TABLE_BITS
;
1946 lk_tab_size
= dev
->lk_table
.max
* sizeof(*dev
->lk_table
.table
);
1947 dev
->lk_table
.table
= (struct hfi1_mregion __rcu
**)
1948 vmalloc(lk_tab_size
);
1949 if (dev
->lk_table
.table
== NULL
) {
1953 RCU_INIT_POINTER(dev
->dma_mr
, NULL
);
1954 for (i
= 0; i
< dev
->lk_table
.max
; i
++)
1955 RCU_INIT_POINTER(dev
->lk_table
.table
[i
], NULL
);
1956 INIT_LIST_HEAD(&dev
->pending_mmaps
);
1957 spin_lock_init(&dev
->pending_lock
);
1958 seqlock_init(&dev
->iowait_lock
);
1959 dev
->mmap_offset
= PAGE_SIZE
;
1960 spin_lock_init(&dev
->mmap_offset_lock
);
1961 INIT_LIST_HEAD(&dev
->txwait
);
1962 INIT_LIST_HEAD(&dev
->memwait
);
1964 descq_cnt
= sdma_get_descq_cnt();
1966 snprintf(buf
, sizeof(buf
), "hfi1_%u_vtxreq_cache", dd
->unit
);
1967 /* SLAB_HWCACHE_ALIGN for AHG */
1968 dev
->verbs_txreq_cache
= kmem_cache_create(buf
,
1969 sizeof(struct verbs_txreq
),
1970 0, SLAB_HWCACHE_ALIGN
,
1971 verbs_txreq_kmem_cache_ctor
);
1972 if (!dev
->verbs_txreq_cache
) {
1974 goto err_verbs_txreq
;
1978 * The system image GUID is supposed to be the same for all
1979 * HFIs in a single system but since there can be other
1980 * device types in the system, we can't be sure this is unique.
1982 if (!ib_hfi1_sys_image_guid
)
1983 ib_hfi1_sys_image_guid
= cpu_to_be64(ppd
->guid
);
1984 lcpysz
= strlcpy(ibdev
->name
, class_name(), lcpysz
);
1985 strlcpy(ibdev
->name
+ lcpysz
, "_%d", IB_DEVICE_NAME_MAX
- lcpysz
);
1986 ibdev
->owner
= THIS_MODULE
;
1987 ibdev
->node_guid
= cpu_to_be64(ppd
->guid
);
1988 ibdev
->uverbs_abi_ver
= HFI1_UVERBS_ABI_VERSION
;
1989 ibdev
->uverbs_cmd_mask
=
1990 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT
) |
1991 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE
) |
1992 (1ull << IB_USER_VERBS_CMD_QUERY_PORT
) |
1993 (1ull << IB_USER_VERBS_CMD_ALLOC_PD
) |
1994 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD
) |
1995 (1ull << IB_USER_VERBS_CMD_CREATE_AH
) |
1996 (1ull << IB_USER_VERBS_CMD_MODIFY_AH
) |
1997 (1ull << IB_USER_VERBS_CMD_QUERY_AH
) |
1998 (1ull << IB_USER_VERBS_CMD_DESTROY_AH
) |
1999 (1ull << IB_USER_VERBS_CMD_REG_MR
) |
2000 (1ull << IB_USER_VERBS_CMD_DEREG_MR
) |
2001 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL
) |
2002 (1ull << IB_USER_VERBS_CMD_CREATE_CQ
) |
2003 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ
) |
2004 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ
) |
2005 (1ull << IB_USER_VERBS_CMD_POLL_CQ
) |
2006 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ
) |
2007 (1ull << IB_USER_VERBS_CMD_CREATE_QP
) |
2008 (1ull << IB_USER_VERBS_CMD_QUERY_QP
) |
2009 (1ull << IB_USER_VERBS_CMD_MODIFY_QP
) |
2010 (1ull << IB_USER_VERBS_CMD_DESTROY_QP
) |
2011 (1ull << IB_USER_VERBS_CMD_POST_SEND
) |
2012 (1ull << IB_USER_VERBS_CMD_POST_RECV
) |
2013 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST
) |
2014 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST
) |
2015 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ
) |
2016 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ
) |
2017 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ
) |
2018 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ
) |
2019 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV
);
2020 ibdev
->node_type
= RDMA_NODE_IB_CA
;
2021 ibdev
->phys_port_cnt
= dd
->num_pports
;
2022 ibdev
->num_comp_vectors
= 1;
2023 ibdev
->dma_device
= &dd
->pcidev
->dev
;
2024 ibdev
->query_device
= query_device
;
2025 ibdev
->modify_device
= modify_device
;
2026 ibdev
->query_port
= query_port
;
2027 ibdev
->modify_port
= modify_port
;
2028 ibdev
->query_pkey
= query_pkey
;
2029 ibdev
->query_gid
= query_gid
;
2030 ibdev
->alloc_ucontext
= alloc_ucontext
;
2031 ibdev
->dealloc_ucontext
= dealloc_ucontext
;
2032 ibdev
->alloc_pd
= alloc_pd
;
2033 ibdev
->dealloc_pd
= dealloc_pd
;
2034 ibdev
->create_ah
= create_ah
;
2035 ibdev
->destroy_ah
= destroy_ah
;
2036 ibdev
->modify_ah
= modify_ah
;
2037 ibdev
->query_ah
= query_ah
;
2038 ibdev
->create_srq
= hfi1_create_srq
;
2039 ibdev
->modify_srq
= hfi1_modify_srq
;
2040 ibdev
->query_srq
= hfi1_query_srq
;
2041 ibdev
->destroy_srq
= hfi1_destroy_srq
;
2042 ibdev
->create_qp
= hfi1_create_qp
;
2043 ibdev
->modify_qp
= hfi1_modify_qp
;
2044 ibdev
->query_qp
= hfi1_query_qp
;
2045 ibdev
->destroy_qp
= hfi1_destroy_qp
;
2046 ibdev
->post_send
= post_send
;
2047 ibdev
->post_recv
= post_receive
;
2048 ibdev
->post_srq_recv
= hfi1_post_srq_receive
;
2049 ibdev
->create_cq
= hfi1_create_cq
;
2050 ibdev
->destroy_cq
= hfi1_destroy_cq
;
2051 ibdev
->resize_cq
= hfi1_resize_cq
;
2052 ibdev
->poll_cq
= hfi1_poll_cq
;
2053 ibdev
->req_notify_cq
= hfi1_req_notify_cq
;
2054 ibdev
->get_dma_mr
= hfi1_get_dma_mr
;
2055 ibdev
->reg_user_mr
= hfi1_reg_user_mr
;
2056 ibdev
->dereg_mr
= hfi1_dereg_mr
;
2057 ibdev
->alloc_mr
= hfi1_alloc_mr
;
2058 ibdev
->alloc_fmr
= hfi1_alloc_fmr
;
2059 ibdev
->map_phys_fmr
= hfi1_map_phys_fmr
;
2060 ibdev
->unmap_fmr
= hfi1_unmap_fmr
;
2061 ibdev
->dealloc_fmr
= hfi1_dealloc_fmr
;
2062 ibdev
->attach_mcast
= hfi1_multicast_attach
;
2063 ibdev
->detach_mcast
= hfi1_multicast_detach
;
2064 ibdev
->process_mad
= hfi1_process_mad
;
2065 ibdev
->mmap
= hfi1_mmap
;
2066 ibdev
->dma_ops
= &hfi1_dma_mapping_ops
;
2067 ibdev
->get_port_immutable
= port_immutable
;
2069 strncpy(ibdev
->node_desc
, init_utsname()->nodename
,
2070 sizeof(ibdev
->node_desc
));
2072 ret
= ib_register_device(ibdev
, hfi1_create_port_files
);
2076 ret
= hfi1_create_agents(dev
);
2080 ret
= hfi1_verbs_register_sysfs(dd
);
2087 hfi1_free_agents(dev
);
2089 ib_unregister_device(ibdev
);
2092 kmem_cache_destroy(dev
->verbs_txreq_cache
);
2093 vfree(dev
->lk_table
.table
);
2097 dd_dev_err(dd
, "cannot register verbs: %d!\n", -ret
);
2102 void hfi1_unregister_ib_device(struct hfi1_devdata
*dd
)
2104 struct hfi1_ibdev
*dev
= &dd
->verbs_dev
;
2105 struct ib_device
*ibdev
= &dev
->ibdev
;
2107 hfi1_verbs_unregister_sysfs(dd
);
2109 hfi1_free_agents(dev
);
2111 ib_unregister_device(ibdev
);
2113 if (!list_empty(&dev
->txwait
))
2114 dd_dev_err(dd
, "txwait list not empty!\n");
2115 if (!list_empty(&dev
->memwait
))
2116 dd_dev_err(dd
, "memwait list not empty!\n");
2118 dd_dev_err(dd
, "DMA MR not NULL!\n");
2121 del_timer_sync(&dev
->mem_timer
);
2122 kmem_cache_destroy(dev
->verbs_txreq_cache
);
2123 vfree(dev
->lk_table
.table
);
2126 void hfi1_cnp_rcv(struct hfi1_packet
*packet
)
2128 struct hfi1_ibport
*ibp
= &packet
->rcd
->ppd
->ibport_data
;
2129 struct hfi1_pportdata
*ppd
= ppd_from_ibp(ibp
);
2130 struct hfi1_ib_header
*hdr
= packet
->hdr
;
2131 struct hfi1_qp
*qp
= packet
->qp
;
2134 u8 sl
, sc5
, sc4_bit
, svc_type
;
2135 bool sc4_set
= has_sc4_bit(packet
);
2137 switch (packet
->qp
->ibqp
.qp_type
) {
2139 rlid
= qp
->remote_ah_attr
.dlid
;
2140 rqpn
= qp
->remote_qpn
;
2141 svc_type
= IB_CC_SVCTYPE_UC
;
2144 rlid
= qp
->remote_ah_attr
.dlid
;
2145 rqpn
= qp
->remote_qpn
;
2146 svc_type
= IB_CC_SVCTYPE_RC
;
2151 svc_type
= IB_CC_SVCTYPE_UD
;
2158 sc4_bit
= sc4_set
<< 4;
2159 sc5
= (be16_to_cpu(hdr
->lrh
[0]) >> 12) & 0xf;
2161 sl
= ibp
->sc_to_sl
[sc5
];
2162 lqpn
= qp
->ibqp
.qp_num
;
2164 process_becn(ppd
, sl
, rlid
, lqpn
, rqpn
, svc_type
);