4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd_cb.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
44 kiblnd_tx_done(lnet_ni_t
*ni
, kib_tx_t
*tx
)
46 lnet_msg_t
*lntmsg
[2];
47 kib_net_t
*net
= ni
->ni_data
;
52 LASSERT(!in_interrupt());
53 LASSERT(!tx
->tx_queued
); /* mustn't be queued for sending */
54 LASSERT(tx
->tx_sending
== 0); /* mustn't be awaiting sent callback */
55 LASSERT(!tx
->tx_waiting
); /* mustn't be awaiting peer response */
56 LASSERT(tx
->tx_pool
!= NULL
);
58 kiblnd_unmap_tx(ni
, tx
);
60 /* tx may have up to 2 lnet msgs to finalise */
61 lntmsg
[0] = tx
->tx_lntmsg
[0]; tx
->tx_lntmsg
[0] = NULL
;
62 lntmsg
[1] = tx
->tx_lntmsg
[1]; tx
->tx_lntmsg
[1] = NULL
;
65 if (tx
->tx_conn
!= NULL
) {
66 LASSERT(ni
== tx
->tx_conn
->ibc_peer
->ibp_ni
);
68 kiblnd_conn_decref(tx
->tx_conn
);
75 kiblnd_pool_free_node(&tx
->tx_pool
->tpo_pool
, &tx
->tx_list
);
77 /* delay finalize until my descs have been freed */
78 for (i
= 0; i
< 2; i
++) {
79 if (lntmsg
[i
] == NULL
)
82 lnet_finalize(ni
, lntmsg
[i
], rc
);
87 kiblnd_txlist_done(lnet_ni_t
*ni
, struct list_head
*txlist
, int status
)
91 while (!list_empty(txlist
)) {
92 tx
= list_entry(txlist
->next
, kib_tx_t
, tx_list
);
94 list_del(&tx
->tx_list
);
97 tx
->tx_status
= status
;
98 kiblnd_tx_done(ni
, tx
);
103 kiblnd_get_idle_tx(lnet_ni_t
*ni
, lnet_nid_t target
)
105 kib_net_t
*net
= (kib_net_t
*)ni
->ni_data
;
106 struct list_head
*node
;
108 kib_tx_poolset_t
*tps
;
110 tps
= net
->ibn_tx_ps
[lnet_cpt_of_nid(target
)];
111 node
= kiblnd_pool_alloc_node(&tps
->tps_poolset
);
114 tx
= container_of(node
, kib_tx_t
, tx_list
);
116 LASSERT(tx
->tx_nwrq
== 0);
117 LASSERT(!tx
->tx_queued
);
118 LASSERT(tx
->tx_sending
== 0);
119 LASSERT(!tx
->tx_waiting
);
120 LASSERT(tx
->tx_status
== 0);
121 LASSERT(tx
->tx_conn
== NULL
);
122 LASSERT(tx
->tx_lntmsg
[0] == NULL
);
123 LASSERT(tx
->tx_lntmsg
[1] == NULL
);
124 LASSERT(tx
->tx_u
.pmr
== NULL
);
125 LASSERT(tx
->tx_nfrags
== 0);
131 kiblnd_drop_rx(kib_rx_t
*rx
)
133 kib_conn_t
*conn
= rx
->rx_conn
;
134 struct kib_sched_info
*sched
= conn
->ibc_sched
;
137 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
138 LASSERT(conn
->ibc_nrx
> 0);
140 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
142 kiblnd_conn_decref(conn
);
146 kiblnd_post_rx(kib_rx_t
*rx
, int credit
)
148 kib_conn_t
*conn
= rx
->rx_conn
;
149 kib_net_t
*net
= conn
->ibc_peer
->ibp_ni
->ni_data
;
150 struct ib_recv_wr
*bad_wrq
= NULL
;
154 LASSERT(net
!= NULL
);
155 LASSERT(!in_interrupt());
156 LASSERT(credit
== IBLND_POSTRX_NO_CREDIT
||
157 credit
== IBLND_POSTRX_PEER_CREDIT
||
158 credit
== IBLND_POSTRX_RSRVD_CREDIT
);
160 mr
= kiblnd_find_dma_mr(conn
->ibc_hdev
, rx
->rx_msgaddr
, IBLND_MSG_SIZE
);
163 rx
->rx_sge
.lkey
= mr
->lkey
;
164 rx
->rx_sge
.addr
= rx
->rx_msgaddr
;
165 rx
->rx_sge
.length
= IBLND_MSG_SIZE
;
167 rx
->rx_wrq
.next
= NULL
;
168 rx
->rx_wrq
.sg_list
= &rx
->rx_sge
;
169 rx
->rx_wrq
.num_sge
= 1;
170 rx
->rx_wrq
.wr_id
= kiblnd_ptr2wreqid(rx
, IBLND_WID_RX
);
172 LASSERT(conn
->ibc_state
>= IBLND_CONN_INIT
);
173 LASSERT(rx
->rx_nob
>= 0); /* not posted */
175 if (conn
->ibc_state
> IBLND_CONN_ESTABLISHED
) {
176 kiblnd_drop_rx(rx
); /* No more posts for this rx */
180 rx
->rx_nob
= -1; /* flag posted */
182 rc
= ib_post_recv(conn
->ibc_cmid
->qp
, &rx
->rx_wrq
, &bad_wrq
);
184 CERROR("Can't post rx for %s: %d, bad_wrq: %p\n",
185 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), rc
, bad_wrq
);
189 if (conn
->ibc_state
< IBLND_CONN_ESTABLISHED
) /* Initial post */
193 kiblnd_close_conn(conn
, rc
);
194 kiblnd_drop_rx(rx
); /* No more posts for this rx */
198 if (credit
== IBLND_POSTRX_NO_CREDIT
)
201 spin_lock(&conn
->ibc_lock
);
202 if (credit
== IBLND_POSTRX_PEER_CREDIT
)
203 conn
->ibc_outstanding_credits
++;
205 conn
->ibc_reserved_credits
++;
206 spin_unlock(&conn
->ibc_lock
);
208 kiblnd_check_sends(conn
);
213 kiblnd_find_waiting_tx_locked(kib_conn_t
*conn
, int txtype
, __u64 cookie
)
215 struct list_head
*tmp
;
217 list_for_each(tmp
, &conn
->ibc_active_txs
) {
218 kib_tx_t
*tx
= list_entry(tmp
, kib_tx_t
, tx_list
);
220 LASSERT(!tx
->tx_queued
);
221 LASSERT(tx
->tx_sending
!= 0 || tx
->tx_waiting
);
223 if (tx
->tx_cookie
!= cookie
)
226 if (tx
->tx_waiting
&&
227 tx
->tx_msg
->ibm_type
== txtype
)
230 CWARN("Bad completion: %swaiting, type %x (wanted %x)\n",
231 tx
->tx_waiting
? "" : "NOT ",
232 tx
->tx_msg
->ibm_type
, txtype
);
238 kiblnd_handle_completion(kib_conn_t
*conn
, int txtype
, int status
, __u64 cookie
)
241 lnet_ni_t
*ni
= conn
->ibc_peer
->ibp_ni
;
244 spin_lock(&conn
->ibc_lock
);
246 tx
= kiblnd_find_waiting_tx_locked(conn
, txtype
, cookie
);
248 spin_unlock(&conn
->ibc_lock
);
250 CWARN("Unmatched completion type %x cookie %#llx from %s\n",
251 txtype
, cookie
, libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
252 kiblnd_close_conn(conn
, -EPROTO
);
256 if (tx
->tx_status
== 0) { /* success so far */
257 if (status
< 0) { /* failed? */
258 tx
->tx_status
= status
;
259 } else if (txtype
== IBLND_MSG_GET_REQ
) {
260 lnet_set_reply_msg_len(ni
, tx
->tx_lntmsg
[1], status
);
266 idle
= !tx
->tx_queued
&& (tx
->tx_sending
== 0);
268 list_del(&tx
->tx_list
);
270 spin_unlock(&conn
->ibc_lock
);
273 kiblnd_tx_done(ni
, tx
);
277 kiblnd_send_completion(kib_conn_t
*conn
, int type
, int status
, __u64 cookie
)
279 lnet_ni_t
*ni
= conn
->ibc_peer
->ibp_ni
;
280 kib_tx_t
*tx
= kiblnd_get_idle_tx(ni
, conn
->ibc_peer
->ibp_nid
);
283 CERROR("Can't get tx for completion %x for %s\n",
284 type
, libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
288 tx
->tx_msg
->ibm_u
.completion
.ibcm_status
= status
;
289 tx
->tx_msg
->ibm_u
.completion
.ibcm_cookie
= cookie
;
290 kiblnd_init_tx_msg(ni
, tx
, type
, sizeof(kib_completion_msg_t
));
292 kiblnd_queue_tx(tx
, conn
);
296 kiblnd_handle_rx(kib_rx_t
*rx
)
298 kib_msg_t
*msg
= rx
->rx_msg
;
299 kib_conn_t
*conn
= rx
->rx_conn
;
300 lnet_ni_t
*ni
= conn
->ibc_peer
->ibp_ni
;
301 int credits
= msg
->ibm_credits
;
307 LASSERT(conn
->ibc_state
>= IBLND_CONN_ESTABLISHED
);
309 CDEBUG(D_NET
, "Received %x[%d] from %s\n",
310 msg
->ibm_type
, credits
,
311 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
314 /* Have I received credits that will let me send? */
315 spin_lock(&conn
->ibc_lock
);
317 if (conn
->ibc_credits
+ credits
>
318 IBLND_MSG_QUEUE_SIZE(conn
->ibc_version
)) {
319 rc2
= conn
->ibc_credits
;
320 spin_unlock(&conn
->ibc_lock
);
322 CERROR("Bad credits from %s: %d + %d > %d\n",
323 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
),
325 IBLND_MSG_QUEUE_SIZE(conn
->ibc_version
));
327 kiblnd_close_conn(conn
, -EPROTO
);
328 kiblnd_post_rx(rx
, IBLND_POSTRX_NO_CREDIT
);
332 conn
->ibc_credits
+= credits
;
334 /* This ensures the credit taken by NOOP can be returned */
335 if (msg
->ibm_type
== IBLND_MSG_NOOP
&&
336 !IBLND_OOB_CAPABLE(conn
->ibc_version
)) /* v1 only */
337 conn
->ibc_outstanding_credits
++;
339 spin_unlock(&conn
->ibc_lock
);
340 kiblnd_check_sends(conn
);
343 switch (msg
->ibm_type
) {
345 CERROR("Bad IBLND message type %x from %s\n",
346 msg
->ibm_type
, libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
347 post_credit
= IBLND_POSTRX_NO_CREDIT
;
352 if (IBLND_OOB_CAPABLE(conn
->ibc_version
)) {
353 post_credit
= IBLND_POSTRX_NO_CREDIT
;
357 if (credits
!= 0) /* credit already posted */
358 post_credit
= IBLND_POSTRX_NO_CREDIT
;
359 else /* a keepalive NOOP */
360 post_credit
= IBLND_POSTRX_PEER_CREDIT
;
363 case IBLND_MSG_IMMEDIATE
:
364 post_credit
= IBLND_POSTRX_DONT_POST
;
365 rc
= lnet_parse(ni
, &msg
->ibm_u
.immediate
.ibim_hdr
,
366 msg
->ibm_srcnid
, rx
, 0);
367 if (rc
< 0) /* repost on error */
368 post_credit
= IBLND_POSTRX_PEER_CREDIT
;
371 case IBLND_MSG_PUT_REQ
:
372 post_credit
= IBLND_POSTRX_DONT_POST
;
373 rc
= lnet_parse(ni
, &msg
->ibm_u
.putreq
.ibprm_hdr
,
374 msg
->ibm_srcnid
, rx
, 1);
375 if (rc
< 0) /* repost on error */
376 post_credit
= IBLND_POSTRX_PEER_CREDIT
;
379 case IBLND_MSG_PUT_NAK
:
380 CWARN("PUT_NACK from %s\n",
381 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
382 post_credit
= IBLND_POSTRX_RSRVD_CREDIT
;
383 kiblnd_handle_completion(conn
, IBLND_MSG_PUT_REQ
,
384 msg
->ibm_u
.completion
.ibcm_status
,
385 msg
->ibm_u
.completion
.ibcm_cookie
);
388 case IBLND_MSG_PUT_ACK
:
389 post_credit
= IBLND_POSTRX_RSRVD_CREDIT
;
391 spin_lock(&conn
->ibc_lock
);
392 tx
= kiblnd_find_waiting_tx_locked(conn
, IBLND_MSG_PUT_REQ
,
393 msg
->ibm_u
.putack
.ibpam_src_cookie
);
395 list_del(&tx
->tx_list
);
396 spin_unlock(&conn
->ibc_lock
);
399 CERROR("Unmatched PUT_ACK from %s\n",
400 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
405 LASSERT(tx
->tx_waiting
);
406 /* CAVEAT EMPTOR: I could be racing with tx_complete, but...
407 * (a) I can overwrite tx_msg since my peer has received it!
408 * (b) tx_waiting set tells tx_complete() it's not done. */
410 tx
->tx_nwrq
= 0; /* overwrite PUT_REQ */
412 rc2
= kiblnd_init_rdma(conn
, tx
, IBLND_MSG_PUT_DONE
,
413 kiblnd_rd_size(&msg
->ibm_u
.putack
.ibpam_rd
),
414 &msg
->ibm_u
.putack
.ibpam_rd
,
415 msg
->ibm_u
.putack
.ibpam_dst_cookie
);
417 CERROR("Can't setup rdma for PUT to %s: %d\n",
418 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), rc2
);
420 spin_lock(&conn
->ibc_lock
);
421 tx
->tx_waiting
= 0; /* clear waiting and queue atomically */
422 kiblnd_queue_tx_locked(tx
, conn
);
423 spin_unlock(&conn
->ibc_lock
);
426 case IBLND_MSG_PUT_DONE
:
427 post_credit
= IBLND_POSTRX_PEER_CREDIT
;
428 kiblnd_handle_completion(conn
, IBLND_MSG_PUT_ACK
,
429 msg
->ibm_u
.completion
.ibcm_status
,
430 msg
->ibm_u
.completion
.ibcm_cookie
);
433 case IBLND_MSG_GET_REQ
:
434 post_credit
= IBLND_POSTRX_DONT_POST
;
435 rc
= lnet_parse(ni
, &msg
->ibm_u
.get
.ibgm_hdr
,
436 msg
->ibm_srcnid
, rx
, 1);
437 if (rc
< 0) /* repost on error */
438 post_credit
= IBLND_POSTRX_PEER_CREDIT
;
441 case IBLND_MSG_GET_DONE
:
442 post_credit
= IBLND_POSTRX_RSRVD_CREDIT
;
443 kiblnd_handle_completion(conn
, IBLND_MSG_GET_REQ
,
444 msg
->ibm_u
.completion
.ibcm_status
,
445 msg
->ibm_u
.completion
.ibcm_cookie
);
449 if (rc
< 0) /* protocol error */
450 kiblnd_close_conn(conn
, rc
);
452 if (post_credit
!= IBLND_POSTRX_DONT_POST
)
453 kiblnd_post_rx(rx
, post_credit
);
457 kiblnd_rx_complete(kib_rx_t
*rx
, int status
, int nob
)
459 kib_msg_t
*msg
= rx
->rx_msg
;
460 kib_conn_t
*conn
= rx
->rx_conn
;
461 lnet_ni_t
*ni
= conn
->ibc_peer
->ibp_ni
;
462 kib_net_t
*net
= ni
->ni_data
;
466 LASSERT(net
!= NULL
);
467 LASSERT(rx
->rx_nob
< 0); /* was posted */
468 rx
->rx_nob
= 0; /* isn't now */
470 if (conn
->ibc_state
> IBLND_CONN_ESTABLISHED
)
473 if (status
!= IB_WC_SUCCESS
) {
474 CNETERR("Rx from %s failed: %d\n",
475 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), status
);
482 rc
= kiblnd_unpack_msg(msg
, rx
->rx_nob
);
484 CERROR("Error %d unpacking rx from %s\n",
485 rc
, libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
489 if (msg
->ibm_srcnid
!= conn
->ibc_peer
->ibp_nid
||
490 msg
->ibm_dstnid
!= ni
->ni_nid
||
491 msg
->ibm_srcstamp
!= conn
->ibc_incarnation
||
492 msg
->ibm_dststamp
!= net
->ibn_incarnation
) {
493 CERROR("Stale rx from %s\n",
494 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
499 /* set time last known alive */
500 kiblnd_peer_alive(conn
->ibc_peer
);
502 /* racing with connection establishment/teardown! */
504 if (conn
->ibc_state
< IBLND_CONN_ESTABLISHED
) {
505 rwlock_t
*g_lock
= &kiblnd_data
.kib_global_lock
;
508 write_lock_irqsave(g_lock
, flags
);
509 /* must check holding global lock to eliminate race */
510 if (conn
->ibc_state
< IBLND_CONN_ESTABLISHED
) {
511 list_add_tail(&rx
->rx_list
, &conn
->ibc_early_rxs
);
512 write_unlock_irqrestore(g_lock
, flags
);
515 write_unlock_irqrestore(g_lock
, flags
);
517 kiblnd_handle_rx(rx
);
521 CDEBUG(D_NET
, "rx %p conn %p\n", rx
, conn
);
522 kiblnd_close_conn(conn
, err
);
524 kiblnd_drop_rx(rx
); /* Don't re-post rx. */
528 kiblnd_kvaddr_to_page(unsigned long vaddr
)
532 if (is_vmalloc_addr((void *)vaddr
)) {
533 page
= vmalloc_to_page((void *)vaddr
);
534 LASSERT(page
!= NULL
);
537 #ifdef CONFIG_HIGHMEM
538 if (vaddr
>= PKMAP_BASE
&&
539 vaddr
< (PKMAP_BASE
+ LAST_PKMAP
* PAGE_SIZE
)) {
540 /* No highmem pages only used for bulk (kiov) I/O */
541 CERROR("find page for address in highmem\n");
545 page
= virt_to_page(vaddr
);
546 LASSERT(page
!= NULL
);
551 kiblnd_fmr_map_tx(kib_net_t
*net
, kib_tx_t
*tx
, kib_rdma_desc_t
*rd
, int nob
)
554 __u64
*pages
= tx
->tx_pages
;
555 kib_fmr_poolset_t
*fps
;
562 LASSERT(tx
->tx_pool
!= NULL
);
563 LASSERT(tx
->tx_pool
->tpo_pool
.po_owner
!= NULL
);
565 hdev
= tx
->tx_pool
->tpo_hdev
;
567 for (i
= 0, npages
= 0; i
< rd
->rd_nfrags
; i
++) {
568 for (size
= 0; size
< rd
->rd_frags
[i
].rf_nob
;
569 size
+= hdev
->ibh_page_size
) {
570 pages
[npages
++] = (rd
->rd_frags
[i
].rf_addr
&
571 hdev
->ibh_page_mask
) + size
;
575 cpt
= tx
->tx_pool
->tpo_pool
.po_owner
->ps_cpt
;
577 fps
= net
->ibn_fmr_ps
[cpt
];
578 rc
= kiblnd_fmr_pool_map(fps
, pages
, npages
, 0, &tx
->tx_u
.fmr
);
580 CERROR("Can't map %d pages: %d\n", npages
, rc
);
584 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
586 rd
->rd_key
= (rd
!= tx
->tx_rd
) ? tx
->tx_u
.fmr
.fmr_pfmr
->fmr
->rkey
:
587 tx
->tx_u
.fmr
.fmr_pfmr
->fmr
->lkey
;
588 rd
->rd_frags
[0].rf_addr
&= ~hdev
->ibh_page_mask
;
589 rd
->rd_frags
[0].rf_nob
= nob
;
596 kiblnd_pmr_map_tx(kib_net_t
*net
, kib_tx_t
*tx
, kib_rdma_desc_t
*rd
, int nob
)
599 kib_pmr_poolset_t
*pps
;
604 LASSERT(tx
->tx_pool
!= NULL
);
605 LASSERT(tx
->tx_pool
->tpo_pool
.po_owner
!= NULL
);
607 hdev
= tx
->tx_pool
->tpo_hdev
;
609 iova
= rd
->rd_frags
[0].rf_addr
& ~hdev
->ibh_page_mask
;
611 cpt
= tx
->tx_pool
->tpo_pool
.po_owner
->ps_cpt
;
613 pps
= net
->ibn_pmr_ps
[cpt
];
614 rc
= kiblnd_pmr_pool_map(pps
, hdev
, rd
, &iova
, &tx
->tx_u
.pmr
);
616 CERROR("Failed to create MR by phybuf: %d\n", rc
);
620 /* If rd is not tx_rd, it's going to get sent to a peer, who will need
622 rd
->rd_key
= (rd
!= tx
->tx_rd
) ? tx
->tx_u
.pmr
->pmr_mr
->rkey
:
623 tx
->tx_u
.pmr
->pmr_mr
->lkey
;
625 rd
->rd_frags
[0].rf_addr
= iova
;
626 rd
->rd_frags
[0].rf_nob
= nob
;
632 kiblnd_unmap_tx(lnet_ni_t
*ni
, kib_tx_t
*tx
)
634 kib_net_t
*net
= ni
->ni_data
;
636 LASSERT(net
!= NULL
);
638 if (net
->ibn_fmr_ps
!= NULL
&& tx
->tx_u
.fmr
.fmr_pfmr
!= NULL
) {
639 kiblnd_fmr_pool_unmap(&tx
->tx_u
.fmr
, tx
->tx_status
);
640 tx
->tx_u
.fmr
.fmr_pfmr
= NULL
;
642 } else if (net
->ibn_pmr_ps
!= NULL
&& tx
->tx_u
.pmr
!= NULL
) {
643 kiblnd_pmr_pool_unmap(tx
->tx_u
.pmr
);
647 if (tx
->tx_nfrags
!= 0) {
648 kiblnd_dma_unmap_sg(tx
->tx_pool
->tpo_hdev
->ibh_ibdev
,
649 tx
->tx_frags
, tx
->tx_nfrags
, tx
->tx_dmadir
);
655 kiblnd_map_tx(lnet_ni_t
*ni
, kib_tx_t
*tx
,
656 kib_rdma_desc_t
*rd
, int nfrags
)
658 kib_hca_dev_t
*hdev
= tx
->tx_pool
->tpo_hdev
;
659 kib_net_t
*net
= ni
->ni_data
;
660 struct ib_mr
*mr
= NULL
;
664 /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
666 tx
->tx_dmadir
= (rd
!= tx
->tx_rd
) ? DMA_FROM_DEVICE
: DMA_TO_DEVICE
;
667 tx
->tx_nfrags
= nfrags
;
670 kiblnd_dma_map_sg(hdev
->ibh_ibdev
,
671 tx
->tx_frags
, tx
->tx_nfrags
, tx
->tx_dmadir
);
673 for (i
= 0, nob
= 0; i
< rd
->rd_nfrags
; i
++) {
674 rd
->rd_frags
[i
].rf_nob
= kiblnd_sg_dma_len(
675 hdev
->ibh_ibdev
, &tx
->tx_frags
[i
]);
676 rd
->rd_frags
[i
].rf_addr
= kiblnd_sg_dma_address(
677 hdev
->ibh_ibdev
, &tx
->tx_frags
[i
]);
678 nob
+= rd
->rd_frags
[i
].rf_nob
;
681 /* looking for pre-mapping MR */
682 mr
= kiblnd_find_rd_dma_mr(hdev
, rd
);
684 /* found pre-mapping MR */
685 rd
->rd_key
= (rd
!= tx
->tx_rd
) ? mr
->rkey
: mr
->lkey
;
689 if (net
->ibn_fmr_ps
!= NULL
)
690 return kiblnd_fmr_map_tx(net
, tx
, rd
, nob
);
691 else if (net
->ibn_pmr_ps
!= NULL
)
692 return kiblnd_pmr_map_tx(net
, tx
, rd
, nob
);
699 kiblnd_setup_rd_iov(lnet_ni_t
*ni
, kib_tx_t
*tx
, kib_rdma_desc_t
*rd
,
700 unsigned int niov
, struct kvec
*iov
, int offset
, int nob
)
702 kib_net_t
*net
= ni
->ni_data
;
704 struct scatterlist
*sg
;
711 LASSERT(net
!= NULL
);
713 while (offset
>= iov
->iov_len
) {
714 offset
-= iov
->iov_len
;
724 vaddr
= ((unsigned long)iov
->iov_base
) + offset
;
725 page_offset
= vaddr
& (PAGE_SIZE
- 1);
726 page
= kiblnd_kvaddr_to_page(vaddr
);
728 CERROR("Can't find page\n");
732 fragnob
= min((int)(iov
->iov_len
- offset
), nob
);
733 fragnob
= min(fragnob
, (int)PAGE_SIZE
- page_offset
);
735 sg_set_page(sg
, page
, fragnob
, page_offset
);
738 if (offset
+ fragnob
< iov
->iov_len
) {
748 return kiblnd_map_tx(ni
, tx
, rd
, sg
- tx
->tx_frags
);
752 kiblnd_setup_rd_kiov(lnet_ni_t
*ni
, kib_tx_t
*tx
, kib_rdma_desc_t
*rd
,
753 int nkiov
, lnet_kiov_t
*kiov
, int offset
, int nob
)
755 kib_net_t
*net
= ni
->ni_data
;
756 struct scatterlist
*sg
;
759 CDEBUG(D_NET
, "niov %d offset %d nob %d\n", nkiov
, offset
, nob
);
763 LASSERT(net
!= NULL
);
765 while (offset
>= kiov
->kiov_len
) {
766 offset
-= kiov
->kiov_len
;
776 fragnob
= min((int)(kiov
->kiov_len
- offset
), nob
);
778 sg_set_page(sg
, kiov
->kiov_page
, fragnob
,
779 kiov
->kiov_offset
+ offset
);
788 return kiblnd_map_tx(ni
, tx
, rd
, sg
- tx
->tx_frags
);
792 kiblnd_post_tx_locked(kib_conn_t
*conn
, kib_tx_t
*tx
, int credit
)
793 __releases(conn
->ibc_lock
)
794 __acquires(conn
->ibc_lock
)
796 kib_msg_t
*msg
= tx
->tx_msg
;
797 kib_peer_t
*peer
= conn
->ibc_peer
;
798 int ver
= conn
->ibc_version
;
801 struct ib_send_wr
*bad_wrq
;
803 LASSERT(tx
->tx_queued
);
804 /* We rely on this for QP sizing */
805 LASSERT(tx
->tx_nwrq
> 0);
806 LASSERT(tx
->tx_nwrq
<= 1 + IBLND_RDMA_FRAGS(ver
));
808 LASSERT(credit
== 0 || credit
== 1);
809 LASSERT(conn
->ibc_outstanding_credits
>= 0);
810 LASSERT(conn
->ibc_outstanding_credits
<= IBLND_MSG_QUEUE_SIZE(ver
));
811 LASSERT(conn
->ibc_credits
>= 0);
812 LASSERT(conn
->ibc_credits
<= IBLND_MSG_QUEUE_SIZE(ver
));
814 if (conn
->ibc_nsends_posted
== IBLND_CONCURRENT_SENDS(ver
)) {
815 /* tx completions outstanding... */
816 CDEBUG(D_NET
, "%s: posted enough\n",
817 libcfs_nid2str(peer
->ibp_nid
));
821 if (credit
!= 0 && conn
->ibc_credits
== 0) { /* no credits */
822 CDEBUG(D_NET
, "%s: no credits\n",
823 libcfs_nid2str(peer
->ibp_nid
));
827 if (credit
!= 0 && !IBLND_OOB_CAPABLE(ver
) &&
828 conn
->ibc_credits
== 1 && /* last credit reserved */
829 msg
->ibm_type
!= IBLND_MSG_NOOP
) { /* for NOOP */
830 CDEBUG(D_NET
, "%s: not using last credit\n",
831 libcfs_nid2str(peer
->ibp_nid
));
835 /* NB don't drop ibc_lock before bumping tx_sending */
836 list_del(&tx
->tx_list
);
839 if (msg
->ibm_type
== IBLND_MSG_NOOP
&&
840 (!kiblnd_need_noop(conn
) || /* redundant NOOP */
841 (IBLND_OOB_CAPABLE(ver
) && /* posted enough NOOP */
842 conn
->ibc_noops_posted
== IBLND_OOB_MSGS(ver
)))) {
843 /* OK to drop when posted enough NOOPs, since
844 * kiblnd_check_sends will queue NOOP again when
845 * posted NOOPs complete */
846 spin_unlock(&conn
->ibc_lock
);
847 kiblnd_tx_done(peer
->ibp_ni
, tx
);
848 spin_lock(&conn
->ibc_lock
);
849 CDEBUG(D_NET
, "%s(%d): redundant or enough NOOP\n",
850 libcfs_nid2str(peer
->ibp_nid
),
851 conn
->ibc_noops_posted
);
855 kiblnd_pack_msg(peer
->ibp_ni
, msg
, ver
, conn
->ibc_outstanding_credits
,
856 peer
->ibp_nid
, conn
->ibc_incarnation
);
858 conn
->ibc_credits
-= credit
;
859 conn
->ibc_outstanding_credits
= 0;
860 conn
->ibc_nsends_posted
++;
861 if (msg
->ibm_type
== IBLND_MSG_NOOP
)
862 conn
->ibc_noops_posted
++;
864 /* CAVEAT EMPTOR! This tx could be the PUT_DONE of an RDMA
865 * PUT. If so, it was first queued here as a PUT_REQ, sent and
866 * stashed on ibc_active_txs, matched by an incoming PUT_ACK,
867 * and then re-queued here. It's (just) possible that
868 * tx_sending is non-zero if we've not done the tx_complete()
869 * from the first send; hence the ++ rather than = below. */
871 list_add(&tx
->tx_list
, &conn
->ibc_active_txs
);
873 /* I'm still holding ibc_lock! */
874 if (conn
->ibc_state
!= IBLND_CONN_ESTABLISHED
) {
876 } else if (tx
->tx_pool
->tpo_pool
.po_failed
||
877 conn
->ibc_hdev
!= tx
->tx_pool
->tpo_hdev
) {
878 /* close_conn will launch failover */
881 rc
= ib_post_send(conn
->ibc_cmid
->qp
,
882 tx
->tx_wrq
, &bad_wrq
);
885 conn
->ibc_last_send
= jiffies
;
890 /* NB credits are transferred in the actual
891 * message, which can only be the last work item */
892 conn
->ibc_credits
+= credit
;
893 conn
->ibc_outstanding_credits
+= msg
->ibm_credits
;
894 conn
->ibc_nsends_posted
--;
895 if (msg
->ibm_type
== IBLND_MSG_NOOP
)
896 conn
->ibc_noops_posted
--;
902 done
= (tx
->tx_sending
== 0);
904 list_del(&tx
->tx_list
);
906 spin_unlock(&conn
->ibc_lock
);
908 if (conn
->ibc_state
== IBLND_CONN_ESTABLISHED
)
909 CERROR("Error %d posting transmit to %s\n",
910 rc
, libcfs_nid2str(peer
->ibp_nid
));
912 CDEBUG(D_NET
, "Error %d posting transmit to %s\n",
913 rc
, libcfs_nid2str(peer
->ibp_nid
));
915 kiblnd_close_conn(conn
, rc
);
918 kiblnd_tx_done(peer
->ibp_ni
, tx
);
920 spin_lock(&conn
->ibc_lock
);
926 kiblnd_check_sends(kib_conn_t
*conn
)
928 int ver
= conn
->ibc_version
;
929 lnet_ni_t
*ni
= conn
->ibc_peer
->ibp_ni
;
932 /* Don't send anything until after the connection is established */
933 if (conn
->ibc_state
< IBLND_CONN_ESTABLISHED
) {
934 CDEBUG(D_NET
, "%s too soon\n",
935 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
939 spin_lock(&conn
->ibc_lock
);
941 LASSERT(conn
->ibc_nsends_posted
<= IBLND_CONCURRENT_SENDS(ver
));
942 LASSERT(!IBLND_OOB_CAPABLE(ver
) ||
943 conn
->ibc_noops_posted
<= IBLND_OOB_MSGS(ver
));
944 LASSERT(conn
->ibc_reserved_credits
>= 0);
946 while (conn
->ibc_reserved_credits
> 0 &&
947 !list_empty(&conn
->ibc_tx_queue_rsrvd
)) {
948 tx
= list_entry(conn
->ibc_tx_queue_rsrvd
.next
,
950 list_del(&tx
->tx_list
);
951 list_add_tail(&tx
->tx_list
, &conn
->ibc_tx_queue
);
952 conn
->ibc_reserved_credits
--;
955 if (kiblnd_need_noop(conn
)) {
956 spin_unlock(&conn
->ibc_lock
);
958 tx
= kiblnd_get_idle_tx(ni
, conn
->ibc_peer
->ibp_nid
);
960 kiblnd_init_tx_msg(ni
, tx
, IBLND_MSG_NOOP
, 0);
962 spin_lock(&conn
->ibc_lock
);
964 kiblnd_queue_tx_locked(tx
, conn
);
967 kiblnd_conn_addref(conn
); /* 1 ref for me.... (see b21911) */
972 if (!list_empty(&conn
->ibc_tx_queue_nocred
)) {
974 tx
= list_entry(conn
->ibc_tx_queue_nocred
.next
,
976 } else if (!list_empty(&conn
->ibc_tx_noops
)) {
977 LASSERT(!IBLND_OOB_CAPABLE(ver
));
979 tx
= list_entry(conn
->ibc_tx_noops
.next
,
981 } else if (!list_empty(&conn
->ibc_tx_queue
)) {
983 tx
= list_entry(conn
->ibc_tx_queue
.next
,
988 if (kiblnd_post_tx_locked(conn
, tx
, credit
) != 0)
992 spin_unlock(&conn
->ibc_lock
);
994 kiblnd_conn_decref(conn
); /* ...until here */
998 kiblnd_tx_complete(kib_tx_t
*tx
, int status
)
1000 int failed
= (status
!= IB_WC_SUCCESS
);
1001 kib_conn_t
*conn
= tx
->tx_conn
;
1004 LASSERT(tx
->tx_sending
> 0);
1007 if (conn
->ibc_state
== IBLND_CONN_ESTABLISHED
)
1008 CNETERR("Tx -> %s cookie %#llx sending %d waiting %d: failed %d\n",
1009 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
),
1010 tx
->tx_cookie
, tx
->tx_sending
, tx
->tx_waiting
,
1013 kiblnd_close_conn(conn
, -EIO
);
1015 kiblnd_peer_alive(conn
->ibc_peer
);
1018 spin_lock(&conn
->ibc_lock
);
1020 /* I could be racing with rdma completion. Whoever makes 'tx' idle
1021 * gets to free it, which also drops its ref on 'conn'. */
1024 conn
->ibc_nsends_posted
--;
1025 if (tx
->tx_msg
->ibm_type
== IBLND_MSG_NOOP
)
1026 conn
->ibc_noops_posted
--;
1029 tx
->tx_waiting
= 0; /* don't wait for peer */
1030 tx
->tx_status
= -EIO
;
1033 idle
= (tx
->tx_sending
== 0) && /* This is the final callback */
1034 !tx
->tx_waiting
&& /* Not waiting for peer */
1035 !tx
->tx_queued
; /* Not re-queued (PUT_DONE) */
1037 list_del(&tx
->tx_list
);
1039 kiblnd_conn_addref(conn
); /* 1 ref for me.... */
1041 spin_unlock(&conn
->ibc_lock
);
1044 kiblnd_tx_done(conn
->ibc_peer
->ibp_ni
, tx
);
1046 kiblnd_check_sends(conn
);
1048 kiblnd_conn_decref(conn
); /* ...until here */
1052 kiblnd_init_tx_msg(lnet_ni_t
*ni
, kib_tx_t
*tx
, int type
, int body_nob
)
1054 kib_hca_dev_t
*hdev
= tx
->tx_pool
->tpo_hdev
;
1055 struct ib_sge
*sge
= &tx
->tx_sge
[tx
->tx_nwrq
];
1056 struct ib_send_wr
*wrq
= &tx
->tx_wrq
[tx
->tx_nwrq
];
1057 int nob
= offsetof(kib_msg_t
, ibm_u
) + body_nob
;
1060 LASSERT(tx
->tx_nwrq
>= 0);
1061 LASSERT(tx
->tx_nwrq
< IBLND_MAX_RDMA_FRAGS
+ 1);
1062 LASSERT(nob
<= IBLND_MSG_SIZE
);
1064 kiblnd_init_msg(tx
->tx_msg
, type
, body_nob
);
1066 mr
= kiblnd_find_dma_mr(hdev
, tx
->tx_msgaddr
, nob
);
1067 LASSERT(mr
!= NULL
);
1069 sge
->lkey
= mr
->lkey
;
1070 sge
->addr
= tx
->tx_msgaddr
;
1073 memset(wrq
, 0, sizeof(*wrq
));
1076 wrq
->wr_id
= kiblnd_ptr2wreqid(tx
, IBLND_WID_TX
);
1079 wrq
->opcode
= IB_WR_SEND
;
1080 wrq
->send_flags
= IB_SEND_SIGNALED
;
1086 kiblnd_init_rdma(kib_conn_t
*conn
, kib_tx_t
*tx
, int type
,
1087 int resid
, kib_rdma_desc_t
*dstrd
, __u64 dstcookie
)
1089 kib_msg_t
*ibmsg
= tx
->tx_msg
;
1090 kib_rdma_desc_t
*srcrd
= tx
->tx_rd
;
1091 struct ib_sge
*sge
= &tx
->tx_sge
[0];
1092 struct ib_send_wr
*wrq
= &tx
->tx_wrq
[0];
1098 LASSERT(!in_interrupt());
1099 LASSERT(tx
->tx_nwrq
== 0);
1100 LASSERT(type
== IBLND_MSG_GET_DONE
||
1101 type
== IBLND_MSG_PUT_DONE
);
1103 srcidx
= dstidx
= 0;
1106 if (srcidx
>= srcrd
->rd_nfrags
) {
1107 CERROR("Src buffer exhausted: %d frags\n", srcidx
);
1112 if (dstidx
== dstrd
->rd_nfrags
) {
1113 CERROR("Dst buffer exhausted: %d frags\n", dstidx
);
1118 if (tx
->tx_nwrq
== IBLND_RDMA_FRAGS(conn
->ibc_version
)) {
1119 CERROR("RDMA too fragmented for %s (%d): %d/%d src %d/%d dst frags\n",
1120 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
),
1121 IBLND_RDMA_FRAGS(conn
->ibc_version
),
1122 srcidx
, srcrd
->rd_nfrags
,
1123 dstidx
, dstrd
->rd_nfrags
);
1128 wrknob
= min(min(kiblnd_rd_frag_size(srcrd
, srcidx
),
1129 kiblnd_rd_frag_size(dstrd
, dstidx
)),
1132 sge
= &tx
->tx_sge
[tx
->tx_nwrq
];
1133 sge
->addr
= kiblnd_rd_frag_addr(srcrd
, srcidx
);
1134 sge
->lkey
= kiblnd_rd_frag_key(srcrd
, srcidx
);
1135 sge
->length
= wrknob
;
1137 wrq
= &tx
->tx_wrq
[tx
->tx_nwrq
];
1139 wrq
->next
= wrq
+ 1;
1140 wrq
->wr_id
= kiblnd_ptr2wreqid(tx
, IBLND_WID_RDMA
);
1143 wrq
->opcode
= IB_WR_RDMA_WRITE
;
1144 wrq
->send_flags
= 0;
1146 wrq
->wr
.rdma
.remote_addr
= kiblnd_rd_frag_addr(dstrd
, dstidx
);
1147 wrq
->wr
.rdma
.rkey
= kiblnd_rd_frag_key(dstrd
, dstidx
);
1149 srcidx
= kiblnd_rd_consume_frag(srcrd
, srcidx
, wrknob
);
1150 dstidx
= kiblnd_rd_consume_frag(dstrd
, dstidx
, wrknob
);
1159 if (rc
< 0) /* no RDMA if completing with failure */
1162 ibmsg
->ibm_u
.completion
.ibcm_status
= rc
;
1163 ibmsg
->ibm_u
.completion
.ibcm_cookie
= dstcookie
;
1164 kiblnd_init_tx_msg(conn
->ibc_peer
->ibp_ni
, tx
,
1165 type
, sizeof(kib_completion_msg_t
));
1171 kiblnd_queue_tx_locked(kib_tx_t
*tx
, kib_conn_t
*conn
)
1173 struct list_head
*q
;
1175 LASSERT(tx
->tx_nwrq
> 0); /* work items set up */
1176 LASSERT(!tx
->tx_queued
); /* not queued for sending already */
1177 LASSERT(conn
->ibc_state
>= IBLND_CONN_ESTABLISHED
);
1180 tx
->tx_deadline
= jiffies
+ (*kiblnd_tunables
.kib_timeout
* HZ
);
1182 if (tx
->tx_conn
== NULL
) {
1183 kiblnd_conn_addref(conn
);
1185 LASSERT(tx
->tx_msg
->ibm_type
!= IBLND_MSG_PUT_DONE
);
1187 /* PUT_DONE first attached to conn as a PUT_REQ */
1188 LASSERT(tx
->tx_conn
== conn
);
1189 LASSERT(tx
->tx_msg
->ibm_type
== IBLND_MSG_PUT_DONE
);
1192 switch (tx
->tx_msg
->ibm_type
) {
1196 case IBLND_MSG_PUT_REQ
:
1197 case IBLND_MSG_GET_REQ
:
1198 q
= &conn
->ibc_tx_queue_rsrvd
;
1201 case IBLND_MSG_PUT_NAK
:
1202 case IBLND_MSG_PUT_ACK
:
1203 case IBLND_MSG_PUT_DONE
:
1204 case IBLND_MSG_GET_DONE
:
1205 q
= &conn
->ibc_tx_queue_nocred
;
1208 case IBLND_MSG_NOOP
:
1209 if (IBLND_OOB_CAPABLE(conn
->ibc_version
))
1210 q
= &conn
->ibc_tx_queue_nocred
;
1212 q
= &conn
->ibc_tx_noops
;
1215 case IBLND_MSG_IMMEDIATE
:
1216 q
= &conn
->ibc_tx_queue
;
1220 list_add_tail(&tx
->tx_list
, q
);
1224 kiblnd_queue_tx(kib_tx_t
*tx
, kib_conn_t
*conn
)
1226 spin_lock(&conn
->ibc_lock
);
1227 kiblnd_queue_tx_locked(tx
, conn
);
1228 spin_unlock(&conn
->ibc_lock
);
1230 kiblnd_check_sends(conn
);
1233 static int kiblnd_resolve_addr(struct rdma_cm_id
*cmid
,
1234 struct sockaddr_in
*srcaddr
,
1235 struct sockaddr_in
*dstaddr
,
1238 unsigned short port
;
1241 /* allow the port to be reused */
1242 rc
= rdma_set_reuseaddr(cmid
, 1);
1244 CERROR("Unable to set reuse on cmid: %d\n", rc
);
1248 /* look for a free privileged port */
1249 for (port
= PROT_SOCK
-1; port
> 0; port
--) {
1250 srcaddr
->sin_port
= htons(port
);
1251 rc
= rdma_resolve_addr(cmid
,
1252 (struct sockaddr
*)srcaddr
,
1253 (struct sockaddr
*)dstaddr
,
1256 CDEBUG(D_NET
, "bound to port %hu\n", port
);
1258 } else if (rc
== -EADDRINUSE
|| rc
== -EADDRNOTAVAIL
) {
1259 CDEBUG(D_NET
, "bind to port %hu failed: %d\n",
1266 CERROR("Failed to bind to a free privileged port\n");
1271 kiblnd_connect_peer(kib_peer_t
*peer
)
1273 struct rdma_cm_id
*cmid
;
1275 kib_net_t
*net
= peer
->ibp_ni
->ni_data
;
1276 struct sockaddr_in srcaddr
;
1277 struct sockaddr_in dstaddr
;
1280 LASSERT(net
!= NULL
);
1281 LASSERT(peer
->ibp_connecting
> 0);
1283 cmid
= kiblnd_rdma_create_id(kiblnd_cm_callback
, peer
, RDMA_PS_TCP
,
1287 CERROR("Can't create CMID for %s: %ld\n",
1288 libcfs_nid2str(peer
->ibp_nid
), PTR_ERR(cmid
));
1294 memset(&srcaddr
, 0, sizeof(srcaddr
));
1295 srcaddr
.sin_family
= AF_INET
;
1296 srcaddr
.sin_addr
.s_addr
= htonl(dev
->ibd_ifip
);
1298 memset(&dstaddr
, 0, sizeof(dstaddr
));
1299 dstaddr
.sin_family
= AF_INET
;
1300 dstaddr
.sin_port
= htons(*kiblnd_tunables
.kib_service
);
1301 dstaddr
.sin_addr
.s_addr
= htonl(LNET_NIDADDR(peer
->ibp_nid
));
1303 kiblnd_peer_addref(peer
); /* cmid's ref */
1305 if (*kiblnd_tunables
.kib_use_priv_port
) {
1306 rc
= kiblnd_resolve_addr(cmid
, &srcaddr
, &dstaddr
,
1307 *kiblnd_tunables
.kib_timeout
* 1000);
1309 rc
= rdma_resolve_addr(cmid
,
1310 (struct sockaddr
*)&srcaddr
,
1311 (struct sockaddr
*)&dstaddr
,
1312 *kiblnd_tunables
.kib_timeout
* 1000);
1315 /* Can't initiate address resolution: */
1316 CERROR("Can't resolve addr for %s: %d\n",
1317 libcfs_nid2str(peer
->ibp_nid
), rc
);
1321 LASSERT(cmid
->device
!= NULL
);
1322 CDEBUG(D_NET
, "%s: connection bound to %s:%pI4h:%s\n",
1323 libcfs_nid2str(peer
->ibp_nid
), dev
->ibd_ifname
,
1324 &dev
->ibd_ifip
, cmid
->device
->name
);
1329 kiblnd_peer_decref(peer
); /* cmid's ref */
1330 rdma_destroy_id(cmid
);
1332 kiblnd_peer_connect_failed(peer
, 1, rc
);
1336 kiblnd_launch_tx(lnet_ni_t
*ni
, kib_tx_t
*tx
, lnet_nid_t nid
)
1341 rwlock_t
*g_lock
= &kiblnd_data
.kib_global_lock
;
1342 unsigned long flags
;
1345 /* If I get here, I've committed to send, so I complete the tx with
1346 * failure on any problems */
1348 LASSERT(tx
== NULL
|| tx
->tx_conn
== NULL
); /* only set when assigned a conn */
1349 LASSERT(tx
== NULL
|| tx
->tx_nwrq
> 0); /* work items have been set up */
1351 /* First time, just use a read lock since I expect to find my peer
1353 read_lock_irqsave(g_lock
, flags
);
1355 peer
= kiblnd_find_peer_locked(nid
);
1356 if (peer
!= NULL
&& !list_empty(&peer
->ibp_conns
)) {
1357 /* Found a peer with an established connection */
1358 conn
= kiblnd_get_conn_locked(peer
);
1359 kiblnd_conn_addref(conn
); /* 1 ref for me... */
1361 read_unlock_irqrestore(g_lock
, flags
);
1364 kiblnd_queue_tx(tx
, conn
);
1365 kiblnd_conn_decref(conn
); /* ...to here */
1369 read_unlock(g_lock
);
1370 /* Re-try with a write lock */
1373 peer
= kiblnd_find_peer_locked(nid
);
1375 if (list_empty(&peer
->ibp_conns
)) {
1376 /* found a peer, but it's still connecting... */
1377 LASSERT(peer
->ibp_connecting
!= 0 ||
1378 peer
->ibp_accepting
!= 0);
1380 list_add_tail(&tx
->tx_list
,
1381 &peer
->ibp_tx_queue
);
1382 write_unlock_irqrestore(g_lock
, flags
);
1384 conn
= kiblnd_get_conn_locked(peer
);
1385 kiblnd_conn_addref(conn
); /* 1 ref for me... */
1387 write_unlock_irqrestore(g_lock
, flags
);
1390 kiblnd_queue_tx(tx
, conn
);
1391 kiblnd_conn_decref(conn
); /* ...to here */
1396 write_unlock_irqrestore(g_lock
, flags
);
1398 /* Allocate a peer ready to add to the peer table and retry */
1399 rc
= kiblnd_create_peer(ni
, &peer
, nid
);
1401 CERROR("Can't create peer %s\n", libcfs_nid2str(nid
));
1403 tx
->tx_status
= -EHOSTUNREACH
;
1405 kiblnd_tx_done(ni
, tx
);
1410 write_lock_irqsave(g_lock
, flags
);
1412 peer2
= kiblnd_find_peer_locked(nid
);
1413 if (peer2
!= NULL
) {
1414 if (list_empty(&peer2
->ibp_conns
)) {
1415 /* found a peer, but it's still connecting... */
1416 LASSERT(peer2
->ibp_connecting
!= 0 ||
1417 peer2
->ibp_accepting
!= 0);
1419 list_add_tail(&tx
->tx_list
,
1420 &peer2
->ibp_tx_queue
);
1421 write_unlock_irqrestore(g_lock
, flags
);
1423 conn
= kiblnd_get_conn_locked(peer2
);
1424 kiblnd_conn_addref(conn
); /* 1 ref for me... */
1426 write_unlock_irqrestore(g_lock
, flags
);
1429 kiblnd_queue_tx(tx
, conn
);
1430 kiblnd_conn_decref(conn
); /* ...to here */
1433 kiblnd_peer_decref(peer
);
1437 /* Brand new peer */
1438 LASSERT(peer
->ibp_connecting
== 0);
1439 peer
->ibp_connecting
= 1;
1441 /* always called with a ref on ni, which prevents ni being shutdown */
1442 LASSERT(((kib_net_t
*)ni
->ni_data
)->ibn_shutdown
== 0);
1445 list_add_tail(&tx
->tx_list
, &peer
->ibp_tx_queue
);
1447 kiblnd_peer_addref(peer
);
1448 list_add_tail(&peer
->ibp_list
, kiblnd_nid2peerlist(nid
));
1450 write_unlock_irqrestore(g_lock
, flags
);
1452 kiblnd_connect_peer(peer
);
1453 kiblnd_peer_decref(peer
);
1457 kiblnd_send(lnet_ni_t
*ni
, void *private, lnet_msg_t
*lntmsg
)
1459 lnet_hdr_t
*hdr
= &lntmsg
->msg_hdr
;
1460 int type
= lntmsg
->msg_type
;
1461 lnet_process_id_t target
= lntmsg
->msg_target
;
1462 int target_is_router
= lntmsg
->msg_target_is_router
;
1463 int routing
= lntmsg
->msg_routing
;
1464 unsigned int payload_niov
= lntmsg
->msg_niov
;
1465 struct kvec
*payload_iov
= lntmsg
->msg_iov
;
1466 lnet_kiov_t
*payload_kiov
= lntmsg
->msg_kiov
;
1467 unsigned int payload_offset
= lntmsg
->msg_offset
;
1468 unsigned int payload_nob
= lntmsg
->msg_len
;
1474 /* NB 'private' is different depending on what we're sending.... */
1476 CDEBUG(D_NET
, "sending %d bytes in %d frags to %s\n",
1477 payload_nob
, payload_niov
, libcfs_id2str(target
));
1479 LASSERT(payload_nob
== 0 || payload_niov
> 0);
1480 LASSERT(payload_niov
<= LNET_MAX_IOV
);
1482 /* Thread context */
1483 LASSERT(!in_interrupt());
1484 /* payload is either all vaddrs or all pages */
1485 LASSERT(!(payload_kiov
!= NULL
&& payload_iov
!= NULL
));
1493 LASSERT(payload_nob
== 0);
1497 if (routing
|| target_is_router
)
1498 break; /* send IMMEDIATE */
1500 /* is the REPLY message too small for RDMA? */
1501 nob
= offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
[lntmsg
->msg_md
->md_length
]);
1502 if (nob
<= IBLND_MSG_SIZE
)
1503 break; /* send IMMEDIATE */
1505 tx
= kiblnd_get_idle_tx(ni
, target
.nid
);
1507 CERROR("Can't allocate txd for GET to %s\n",
1508 libcfs_nid2str(target
.nid
));
1514 if ((lntmsg
->msg_md
->md_options
& LNET_MD_KIOV
) == 0)
1515 rc
= kiblnd_setup_rd_iov(ni
, tx
,
1516 &ibmsg
->ibm_u
.get
.ibgm_rd
,
1517 lntmsg
->msg_md
->md_niov
,
1518 lntmsg
->msg_md
->md_iov
.iov
,
1519 0, lntmsg
->msg_md
->md_length
);
1521 rc
= kiblnd_setup_rd_kiov(ni
, tx
,
1522 &ibmsg
->ibm_u
.get
.ibgm_rd
,
1523 lntmsg
->msg_md
->md_niov
,
1524 lntmsg
->msg_md
->md_iov
.kiov
,
1525 0, lntmsg
->msg_md
->md_length
);
1527 CERROR("Can't setup GET sink for %s: %d\n",
1528 libcfs_nid2str(target
.nid
), rc
);
1529 kiblnd_tx_done(ni
, tx
);
1533 nob
= offsetof(kib_get_msg_t
, ibgm_rd
.rd_frags
[tx
->tx_nfrags
]);
1534 ibmsg
->ibm_u
.get
.ibgm_cookie
= tx
->tx_cookie
;
1535 ibmsg
->ibm_u
.get
.ibgm_hdr
= *hdr
;
1537 kiblnd_init_tx_msg(ni
, tx
, IBLND_MSG_GET_REQ
, nob
);
1539 tx
->tx_lntmsg
[1] = lnet_create_reply_msg(ni
, lntmsg
);
1540 if (tx
->tx_lntmsg
[1] == NULL
) {
1541 CERROR("Can't create reply for GET -> %s\n",
1542 libcfs_nid2str(target
.nid
));
1543 kiblnd_tx_done(ni
, tx
);
1547 tx
->tx_lntmsg
[0] = lntmsg
; /* finalise lntmsg[0,1] on completion */
1548 tx
->tx_waiting
= 1; /* waiting for GET_DONE */
1549 kiblnd_launch_tx(ni
, tx
, target
.nid
);
1552 case LNET_MSG_REPLY
:
1554 /* Is the payload small enough not to need RDMA? */
1555 nob
= offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
[payload_nob
]);
1556 if (nob
<= IBLND_MSG_SIZE
)
1557 break; /* send IMMEDIATE */
1559 tx
= kiblnd_get_idle_tx(ni
, target
.nid
);
1561 CERROR("Can't allocate %s txd for %s\n",
1562 type
== LNET_MSG_PUT
? "PUT" : "REPLY",
1563 libcfs_nid2str(target
.nid
));
1567 if (payload_kiov
== NULL
)
1568 rc
= kiblnd_setup_rd_iov(ni
, tx
, tx
->tx_rd
,
1569 payload_niov
, payload_iov
,
1570 payload_offset
, payload_nob
);
1572 rc
= kiblnd_setup_rd_kiov(ni
, tx
, tx
->tx_rd
,
1573 payload_niov
, payload_kiov
,
1574 payload_offset
, payload_nob
);
1576 CERROR("Can't setup PUT src for %s: %d\n",
1577 libcfs_nid2str(target
.nid
), rc
);
1578 kiblnd_tx_done(ni
, tx
);
1583 ibmsg
->ibm_u
.putreq
.ibprm_hdr
= *hdr
;
1584 ibmsg
->ibm_u
.putreq
.ibprm_cookie
= tx
->tx_cookie
;
1585 kiblnd_init_tx_msg(ni
, tx
, IBLND_MSG_PUT_REQ
, sizeof(kib_putreq_msg_t
));
1587 tx
->tx_lntmsg
[0] = lntmsg
; /* finalise lntmsg on completion */
1588 tx
->tx_waiting
= 1; /* waiting for PUT_{ACK,NAK} */
1589 kiblnd_launch_tx(ni
, tx
, target
.nid
);
1593 /* send IMMEDIATE */
1595 LASSERT(offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
[payload_nob
])
1598 tx
= kiblnd_get_idle_tx(ni
, target
.nid
);
1600 CERROR("Can't send %d to %s: tx descs exhausted\n",
1601 type
, libcfs_nid2str(target
.nid
));
1606 ibmsg
->ibm_u
.immediate
.ibim_hdr
= *hdr
;
1608 if (payload_kiov
!= NULL
)
1609 lnet_copy_kiov2flat(IBLND_MSG_SIZE
, ibmsg
,
1610 offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
),
1611 payload_niov
, payload_kiov
,
1612 payload_offset
, payload_nob
);
1614 lnet_copy_iov2flat(IBLND_MSG_SIZE
, ibmsg
,
1615 offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
),
1616 payload_niov
, payload_iov
,
1617 payload_offset
, payload_nob
);
1619 nob
= offsetof(kib_immediate_msg_t
, ibim_payload
[payload_nob
]);
1620 kiblnd_init_tx_msg(ni
, tx
, IBLND_MSG_IMMEDIATE
, nob
);
1622 tx
->tx_lntmsg
[0] = lntmsg
; /* finalise lntmsg on completion */
1623 kiblnd_launch_tx(ni
, tx
, target
.nid
);
1628 kiblnd_reply(lnet_ni_t
*ni
, kib_rx_t
*rx
, lnet_msg_t
*lntmsg
)
1630 lnet_process_id_t target
= lntmsg
->msg_target
;
1631 unsigned int niov
= lntmsg
->msg_niov
;
1632 struct kvec
*iov
= lntmsg
->msg_iov
;
1633 lnet_kiov_t
*kiov
= lntmsg
->msg_kiov
;
1634 unsigned int offset
= lntmsg
->msg_offset
;
1635 unsigned int nob
= lntmsg
->msg_len
;
1639 tx
= kiblnd_get_idle_tx(ni
, rx
->rx_conn
->ibc_peer
->ibp_nid
);
1641 CERROR("Can't get tx for REPLY to %s\n",
1642 libcfs_nid2str(target
.nid
));
1648 else if (kiov
== NULL
)
1649 rc
= kiblnd_setup_rd_iov(ni
, tx
, tx
->tx_rd
,
1650 niov
, iov
, offset
, nob
);
1652 rc
= kiblnd_setup_rd_kiov(ni
, tx
, tx
->tx_rd
,
1653 niov
, kiov
, offset
, nob
);
1656 CERROR("Can't setup GET src for %s: %d\n",
1657 libcfs_nid2str(target
.nid
), rc
);
1661 rc
= kiblnd_init_rdma(rx
->rx_conn
, tx
,
1662 IBLND_MSG_GET_DONE
, nob
,
1663 &rx
->rx_msg
->ibm_u
.get
.ibgm_rd
,
1664 rx
->rx_msg
->ibm_u
.get
.ibgm_cookie
);
1666 CERROR("Can't setup rdma for GET from %s: %d\n",
1667 libcfs_nid2str(target
.nid
), rc
);
1672 /* No RDMA: local completion may happen now! */
1673 lnet_finalize(ni
, lntmsg
, 0);
1675 /* RDMA: lnet_finalize(lntmsg) when it
1677 tx
->tx_lntmsg
[0] = lntmsg
;
1680 kiblnd_queue_tx(tx
, rx
->rx_conn
);
1684 kiblnd_tx_done(ni
, tx
);
1686 lnet_finalize(ni
, lntmsg
, -EIO
);
1690 kiblnd_recv(lnet_ni_t
*ni
, void *private, lnet_msg_t
*lntmsg
, int delayed
,
1691 unsigned int niov
, struct kvec
*iov
, lnet_kiov_t
*kiov
,
1692 unsigned int offset
, unsigned int mlen
, unsigned int rlen
)
1694 kib_rx_t
*rx
= private;
1695 kib_msg_t
*rxmsg
= rx
->rx_msg
;
1696 kib_conn_t
*conn
= rx
->rx_conn
;
1700 int post_credit
= IBLND_POSTRX_PEER_CREDIT
;
1703 LASSERT(mlen
<= rlen
);
1704 LASSERT(!in_interrupt());
1705 /* Either all pages or all vaddrs */
1706 LASSERT(!(kiov
!= NULL
&& iov
!= NULL
));
1708 switch (rxmsg
->ibm_type
) {
1712 case IBLND_MSG_IMMEDIATE
:
1713 nob
= offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
[rlen
]);
1714 if (nob
> rx
->rx_nob
) {
1715 CERROR("Immediate message from %s too big: %d(%d)\n",
1716 libcfs_nid2str(rxmsg
->ibm_u
.immediate
.ibim_hdr
.src_nid
),
1723 lnet_copy_flat2kiov(niov
, kiov
, offset
,
1724 IBLND_MSG_SIZE
, rxmsg
,
1725 offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
),
1728 lnet_copy_flat2iov(niov
, iov
, offset
,
1729 IBLND_MSG_SIZE
, rxmsg
,
1730 offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
),
1732 lnet_finalize(ni
, lntmsg
, 0);
1735 case IBLND_MSG_PUT_REQ
:
1737 lnet_finalize(ni
, lntmsg
, 0);
1738 kiblnd_send_completion(rx
->rx_conn
, IBLND_MSG_PUT_NAK
, 0,
1739 rxmsg
->ibm_u
.putreq
.ibprm_cookie
);
1743 tx
= kiblnd_get_idle_tx(ni
, conn
->ibc_peer
->ibp_nid
);
1745 CERROR("Can't allocate tx for %s\n",
1746 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
1747 /* Not replying will break the connection */
1754 rc
= kiblnd_setup_rd_iov(ni
, tx
,
1755 &txmsg
->ibm_u
.putack
.ibpam_rd
,
1756 niov
, iov
, offset
, mlen
);
1758 rc
= kiblnd_setup_rd_kiov(ni
, tx
,
1759 &txmsg
->ibm_u
.putack
.ibpam_rd
,
1760 niov
, kiov
, offset
, mlen
);
1762 CERROR("Can't setup PUT sink for %s: %d\n",
1763 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), rc
);
1764 kiblnd_tx_done(ni
, tx
);
1765 /* tell peer it's over */
1766 kiblnd_send_completion(rx
->rx_conn
, IBLND_MSG_PUT_NAK
, rc
,
1767 rxmsg
->ibm_u
.putreq
.ibprm_cookie
);
1771 nob
= offsetof(kib_putack_msg_t
, ibpam_rd
.rd_frags
[tx
->tx_nfrags
]);
1772 txmsg
->ibm_u
.putack
.ibpam_src_cookie
= rxmsg
->ibm_u
.putreq
.ibprm_cookie
;
1773 txmsg
->ibm_u
.putack
.ibpam_dst_cookie
= tx
->tx_cookie
;
1775 kiblnd_init_tx_msg(ni
, tx
, IBLND_MSG_PUT_ACK
, nob
);
1777 tx
->tx_lntmsg
[0] = lntmsg
; /* finalise lntmsg on completion */
1778 tx
->tx_waiting
= 1; /* waiting for PUT_DONE */
1779 kiblnd_queue_tx(tx
, conn
);
1781 /* reposted buffer reserved for PUT_DONE */
1782 post_credit
= IBLND_POSTRX_NO_CREDIT
;
1785 case IBLND_MSG_GET_REQ
:
1786 if (lntmsg
!= NULL
) {
1787 /* Optimized GET; RDMA lntmsg's payload */
1788 kiblnd_reply(ni
, rx
, lntmsg
);
1790 /* GET didn't match anything */
1791 kiblnd_send_completion(rx
->rx_conn
, IBLND_MSG_GET_DONE
,
1793 rxmsg
->ibm_u
.get
.ibgm_cookie
);
1798 kiblnd_post_rx(rx
, post_credit
);
1803 kiblnd_thread_start(int (*fn
)(void *arg
), void *arg
, char *name
)
1805 struct task_struct
*task
= kthread_run(fn
, arg
, "%s", name
);
1808 return PTR_ERR(task
);
1810 atomic_inc(&kiblnd_data
.kib_nthreads
);
1815 kiblnd_thread_fini(void)
1817 atomic_dec(&kiblnd_data
.kib_nthreads
);
1821 kiblnd_peer_alive(kib_peer_t
*peer
)
1823 /* This is racy, but everyone's only writing cfs_time_current() */
1824 peer
->ibp_last_alive
= cfs_time_current();
1829 kiblnd_peer_notify(kib_peer_t
*peer
)
1832 unsigned long last_alive
= 0;
1833 unsigned long flags
;
1835 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1837 if (list_empty(&peer
->ibp_conns
) &&
1838 peer
->ibp_accepting
== 0 &&
1839 peer
->ibp_connecting
== 0 &&
1840 peer
->ibp_error
!= 0) {
1841 error
= peer
->ibp_error
;
1842 peer
->ibp_error
= 0;
1844 last_alive
= peer
->ibp_last_alive
;
1847 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1850 lnet_notify(peer
->ibp_ni
,
1851 peer
->ibp_nid
, 0, last_alive
);
1855 kiblnd_close_conn_locked(kib_conn_t
*conn
, int error
)
1857 /* This just does the immediate housekeeping. 'error' is zero for a
1858 * normal shutdown which can happen only after the connection has been
1859 * established. If the connection is established, schedule the
1860 * connection to be finished off by the connd. Otherwise the connd is
1861 * already dealing with it (either to set it up or tear it down).
1862 * Caller holds kib_global_lock exclusively in irq context */
1863 kib_peer_t
*peer
= conn
->ibc_peer
;
1865 unsigned long flags
;
1867 LASSERT(error
!= 0 || conn
->ibc_state
>= IBLND_CONN_ESTABLISHED
);
1869 if (error
!= 0 && conn
->ibc_comms_error
== 0)
1870 conn
->ibc_comms_error
= error
;
1872 if (conn
->ibc_state
!= IBLND_CONN_ESTABLISHED
)
1873 return; /* already being handled */
1876 list_empty(&conn
->ibc_tx_noops
) &&
1877 list_empty(&conn
->ibc_tx_queue
) &&
1878 list_empty(&conn
->ibc_tx_queue_rsrvd
) &&
1879 list_empty(&conn
->ibc_tx_queue_nocred
) &&
1880 list_empty(&conn
->ibc_active_txs
)) {
1881 CDEBUG(D_NET
, "closing conn to %s\n",
1882 libcfs_nid2str(peer
->ibp_nid
));
1884 CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
1885 libcfs_nid2str(peer
->ibp_nid
), error
,
1886 list_empty(&conn
->ibc_tx_queue
) ? "" : "(sending)",
1887 list_empty(&conn
->ibc_tx_noops
) ? "" : "(sending_noops)",
1888 list_empty(&conn
->ibc_tx_queue_rsrvd
) ? "" : "(sending_rsrvd)",
1889 list_empty(&conn
->ibc_tx_queue_nocred
) ? "" : "(sending_nocred)",
1890 list_empty(&conn
->ibc_active_txs
) ? "" : "(waiting)");
1893 dev
= ((kib_net_t
*)peer
->ibp_ni
->ni_data
)->ibn_dev
;
1894 list_del(&conn
->ibc_list
);
1895 /* connd (see below) takes over ibc_list's ref */
1897 if (list_empty(&peer
->ibp_conns
) && /* no more conns */
1898 kiblnd_peer_active(peer
)) { /* still in peer table */
1899 kiblnd_unlink_peer_locked(peer
);
1901 /* set/clear error on last conn */
1902 peer
->ibp_error
= conn
->ibc_comms_error
;
1905 kiblnd_set_conn_state(conn
, IBLND_CONN_CLOSING
);
1908 kiblnd_dev_can_failover(dev
)) {
1909 list_add_tail(&dev
->ibd_fail_list
,
1910 &kiblnd_data
.kib_failed_devs
);
1911 wake_up(&kiblnd_data
.kib_failover_waitq
);
1914 spin_lock_irqsave(&kiblnd_data
.kib_connd_lock
, flags
);
1916 list_add_tail(&conn
->ibc_list
, &kiblnd_data
.kib_connd_conns
);
1917 wake_up(&kiblnd_data
.kib_connd_waitq
);
1919 spin_unlock_irqrestore(&kiblnd_data
.kib_connd_lock
, flags
);
1923 kiblnd_close_conn(kib_conn_t
*conn
, int error
)
1925 unsigned long flags
;
1927 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1929 kiblnd_close_conn_locked(conn
, error
);
1931 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1935 kiblnd_handle_early_rxs(kib_conn_t
*conn
)
1937 unsigned long flags
;
1941 LASSERT(!in_interrupt());
1942 LASSERT(conn
->ibc_state
>= IBLND_CONN_ESTABLISHED
);
1944 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1945 list_for_each_entry_safe(rx
, tmp
, &conn
->ibc_early_rxs
, rx_list
) {
1946 list_del(&rx
->rx_list
);
1947 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1949 kiblnd_handle_rx(rx
);
1951 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1953 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1957 kiblnd_abort_txs(kib_conn_t
*conn
, struct list_head
*txs
)
1960 struct list_head
*tmp
;
1961 struct list_head
*nxt
;
1964 spin_lock(&conn
->ibc_lock
);
1966 list_for_each_safe(tmp
, nxt
, txs
) {
1967 tx
= list_entry(tmp
, kib_tx_t
, tx_list
);
1969 if (txs
== &conn
->ibc_active_txs
) {
1970 LASSERT(!tx
->tx_queued
);
1971 LASSERT(tx
->tx_waiting
||
1972 tx
->tx_sending
!= 0);
1974 LASSERT(tx
->tx_queued
);
1977 tx
->tx_status
= -ECONNABORTED
;
1980 if (tx
->tx_sending
== 0) {
1982 list_del(&tx
->tx_list
);
1983 list_add(&tx
->tx_list
, &zombies
);
1987 spin_unlock(&conn
->ibc_lock
);
1989 kiblnd_txlist_done(conn
->ibc_peer
->ibp_ni
, &zombies
, -ECONNABORTED
);
1993 kiblnd_finalise_conn(kib_conn_t
*conn
)
1995 LASSERT(!in_interrupt());
1996 LASSERT(conn
->ibc_state
> IBLND_CONN_INIT
);
1998 kiblnd_set_conn_state(conn
, IBLND_CONN_DISCONNECTED
);
2000 /* abort_receives moves QP state to IB_QPS_ERR. This is only required
2001 * for connections that didn't get as far as being connected, because
2002 * rdma_disconnect() does this for free. */
2003 kiblnd_abort_receives(conn
);
2005 /* Complete all tx descs not waiting for sends to complete.
2006 * NB we should be safe from RDMA now that the QP has changed state */
2008 kiblnd_abort_txs(conn
, &conn
->ibc_tx_noops
);
2009 kiblnd_abort_txs(conn
, &conn
->ibc_tx_queue
);
2010 kiblnd_abort_txs(conn
, &conn
->ibc_tx_queue_rsrvd
);
2011 kiblnd_abort_txs(conn
, &conn
->ibc_tx_queue_nocred
);
2012 kiblnd_abort_txs(conn
, &conn
->ibc_active_txs
);
2014 kiblnd_handle_early_rxs(conn
);
2018 kiblnd_peer_connect_failed(kib_peer_t
*peer
, int active
, int error
)
2021 unsigned long flags
;
2023 LASSERT(error
!= 0);
2024 LASSERT(!in_interrupt());
2026 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2029 LASSERT(peer
->ibp_connecting
> 0);
2030 peer
->ibp_connecting
--;
2032 LASSERT(peer
->ibp_accepting
> 0);
2033 peer
->ibp_accepting
--;
2036 if (peer
->ibp_connecting
!= 0 ||
2037 peer
->ibp_accepting
!= 0) {
2038 /* another connection attempt under way... */
2039 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
,
2044 if (list_empty(&peer
->ibp_conns
)) {
2045 /* Take peer's blocked transmits to complete with error */
2046 list_add(&zombies
, &peer
->ibp_tx_queue
);
2047 list_del_init(&peer
->ibp_tx_queue
);
2049 if (kiblnd_peer_active(peer
))
2050 kiblnd_unlink_peer_locked(peer
);
2052 peer
->ibp_error
= error
;
2054 /* Can't have blocked transmits if there are connections */
2055 LASSERT(list_empty(&peer
->ibp_tx_queue
));
2058 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2060 kiblnd_peer_notify(peer
);
2062 if (list_empty(&zombies
))
2065 CNETERR("Deleting messages for %s: connection failed\n",
2066 libcfs_nid2str(peer
->ibp_nid
));
2068 kiblnd_txlist_done(peer
->ibp_ni
, &zombies
, -EHOSTUNREACH
);
2072 kiblnd_connreq_done(kib_conn_t
*conn
, int status
)
2074 kib_peer_t
*peer
= conn
->ibc_peer
;
2077 struct list_head txs
;
2078 unsigned long flags
;
2081 active
= (conn
->ibc_state
== IBLND_CONN_ACTIVE_CONNECT
);
2083 CDEBUG(D_NET
, "%s: active(%d), version(%x), status(%d)\n",
2084 libcfs_nid2str(peer
->ibp_nid
), active
,
2085 conn
->ibc_version
, status
);
2087 LASSERT(!in_interrupt());
2088 LASSERT((conn
->ibc_state
== IBLND_CONN_ACTIVE_CONNECT
&&
2089 peer
->ibp_connecting
> 0) ||
2090 (conn
->ibc_state
== IBLND_CONN_PASSIVE_WAIT
&&
2091 peer
->ibp_accepting
> 0));
2093 LIBCFS_FREE(conn
->ibc_connvars
, sizeof(*conn
->ibc_connvars
));
2094 conn
->ibc_connvars
= NULL
;
2097 /* failed to establish connection */
2098 kiblnd_peer_connect_failed(peer
, active
, status
);
2099 kiblnd_finalise_conn(conn
);
2103 /* connection established */
2104 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2106 conn
->ibc_last_send
= jiffies
;
2107 kiblnd_set_conn_state(conn
, IBLND_CONN_ESTABLISHED
);
2108 kiblnd_peer_alive(peer
);
2110 /* Add conn to peer's list and nuke any dangling conns from a different
2111 * peer instance... */
2112 kiblnd_conn_addref(conn
); /* +1 ref for ibc_list */
2113 list_add(&conn
->ibc_list
, &peer
->ibp_conns
);
2115 peer
->ibp_connecting
--;
2117 peer
->ibp_accepting
--;
2119 if (peer
->ibp_version
== 0) {
2120 peer
->ibp_version
= conn
->ibc_version
;
2121 peer
->ibp_incarnation
= conn
->ibc_incarnation
;
2124 if (peer
->ibp_version
!= conn
->ibc_version
||
2125 peer
->ibp_incarnation
!= conn
->ibc_incarnation
) {
2126 kiblnd_close_stale_conns_locked(peer
, conn
->ibc_version
,
2127 conn
->ibc_incarnation
);
2128 peer
->ibp_version
= conn
->ibc_version
;
2129 peer
->ibp_incarnation
= conn
->ibc_incarnation
;
2132 /* grab pending txs while I have the lock */
2133 list_add(&txs
, &peer
->ibp_tx_queue
);
2134 list_del_init(&peer
->ibp_tx_queue
);
2136 if (!kiblnd_peer_active(peer
) || /* peer has been deleted */
2137 conn
->ibc_comms_error
!= 0) { /* error has happened already */
2138 lnet_ni_t
*ni
= peer
->ibp_ni
;
2140 /* start to shut down connection */
2141 kiblnd_close_conn_locked(conn
, -ECONNABORTED
);
2142 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2144 kiblnd_txlist_done(ni
, &txs
, -ECONNABORTED
);
2149 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2151 /* Schedule blocked txs */
2152 spin_lock(&conn
->ibc_lock
);
2153 list_for_each_entry_safe(tx
, tmp
, &txs
, tx_list
) {
2154 list_del(&tx
->tx_list
);
2156 kiblnd_queue_tx_locked(tx
, conn
);
2158 spin_unlock(&conn
->ibc_lock
);
2160 kiblnd_check_sends(conn
);
2162 /* schedule blocked rxs */
2163 kiblnd_handle_early_rxs(conn
);
2167 kiblnd_reject(struct rdma_cm_id
*cmid
, kib_rej_t
*rej
)
2171 rc
= rdma_reject(cmid
, rej
, sizeof(*rej
));
2174 CWARN("Error %d sending reject\n", rc
);
2178 kiblnd_passive_connect(struct rdma_cm_id
*cmid
, void *priv
, int priv_nob
)
2180 rwlock_t
*g_lock
= &kiblnd_data
.kib_global_lock
;
2181 kib_msg_t
*reqmsg
= priv
;
2187 lnet_ni_t
*ni
= NULL
;
2188 kib_net_t
*net
= NULL
;
2190 struct rdma_conn_param cp
;
2192 int version
= IBLND_MSG_VERSION
;
2193 unsigned long flags
;
2195 struct sockaddr_in
*peer_addr
;
2196 LASSERT(!in_interrupt());
2198 /* cmid inherits 'context' from the corresponding listener id */
2199 ibdev
= (kib_dev_t
*)cmid
->context
;
2200 LASSERT(ibdev
!= NULL
);
2202 memset(&rej
, 0, sizeof(rej
));
2203 rej
.ibr_magic
= IBLND_MSG_MAGIC
;
2204 rej
.ibr_why
= IBLND_REJECT_FATAL
;
2205 rej
.ibr_cp
.ibcp_max_msg_size
= IBLND_MSG_SIZE
;
2207 peer_addr
= (struct sockaddr_in
*)&(cmid
->route
.addr
.dst_addr
);
2208 if (*kiblnd_tunables
.kib_require_priv_port
&&
2209 ntohs(peer_addr
->sin_port
) >= PROT_SOCK
) {
2210 __u32 ip
= ntohl(peer_addr
->sin_addr
.s_addr
);
2211 CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
2212 &ip
, ntohs(peer_addr
->sin_port
));
2216 if (priv_nob
< offsetof(kib_msg_t
, ibm_type
)) {
2217 CERROR("Short connection request\n");
2221 /* Future protocol version compatibility support! If the
2222 * o2iblnd-specific protocol changes, or when LNET unifies
2223 * protocols over all LNDs, the initial connection will
2224 * negotiate a protocol version. I trap this here to avoid
2225 * console errors; the reject tells the peer which protocol I
2227 if (reqmsg
->ibm_magic
== LNET_PROTO_MAGIC
||
2228 reqmsg
->ibm_magic
== __swab32(LNET_PROTO_MAGIC
))
2230 if (reqmsg
->ibm_magic
== IBLND_MSG_MAGIC
&&
2231 reqmsg
->ibm_version
!= IBLND_MSG_VERSION
&&
2232 reqmsg
->ibm_version
!= IBLND_MSG_VERSION_1
)
2234 if (reqmsg
->ibm_magic
== __swab32(IBLND_MSG_MAGIC
) &&
2235 reqmsg
->ibm_version
!= __swab16(IBLND_MSG_VERSION
) &&
2236 reqmsg
->ibm_version
!= __swab16(IBLND_MSG_VERSION_1
))
2239 rc
= kiblnd_unpack_msg(reqmsg
, priv_nob
);
2241 CERROR("Can't parse connection request: %d\n", rc
);
2245 nid
= reqmsg
->ibm_srcnid
;
2246 ni
= lnet_net2ni(LNET_NIDNET(reqmsg
->ibm_dstnid
));
2249 net
= (kib_net_t
*)ni
->ni_data
;
2250 rej
.ibr_incarnation
= net
->ibn_incarnation
;
2253 if (ni
== NULL
|| /* no matching net */
2254 ni
->ni_nid
!= reqmsg
->ibm_dstnid
|| /* right NET, wrong NID! */
2255 net
->ibn_dev
!= ibdev
) { /* wrong device */
2256 CERROR("Can't accept %s on %s (%s:%d:%pI4h): bad dst nid %s\n",
2257 libcfs_nid2str(nid
),
2258 ni
== NULL
? "NA" : libcfs_nid2str(ni
->ni_nid
),
2259 ibdev
->ibd_ifname
, ibdev
->ibd_nnets
,
2261 libcfs_nid2str(reqmsg
->ibm_dstnid
));
2266 /* check time stamp as soon as possible */
2267 if (reqmsg
->ibm_dststamp
!= 0 &&
2268 reqmsg
->ibm_dststamp
!= net
->ibn_incarnation
) {
2269 CWARN("Stale connection request\n");
2270 rej
.ibr_why
= IBLND_REJECT_CONN_STALE
;
2274 /* I can accept peer's version */
2275 version
= reqmsg
->ibm_version
;
2277 if (reqmsg
->ibm_type
!= IBLND_MSG_CONNREQ
) {
2278 CERROR("Unexpected connreq msg type: %x from %s\n",
2279 reqmsg
->ibm_type
, libcfs_nid2str(nid
));
2283 if (reqmsg
->ibm_u
.connparams
.ibcp_queue_depth
!=
2284 IBLND_MSG_QUEUE_SIZE(version
)) {
2285 CERROR("Can't accept %s: incompatible queue depth %d (%d wanted)\n",
2286 libcfs_nid2str(nid
), reqmsg
->ibm_u
.connparams
.ibcp_queue_depth
,
2287 IBLND_MSG_QUEUE_SIZE(version
));
2289 if (version
== IBLND_MSG_VERSION
)
2290 rej
.ibr_why
= IBLND_REJECT_MSG_QUEUE_SIZE
;
2295 if (reqmsg
->ibm_u
.connparams
.ibcp_max_frags
!=
2296 IBLND_RDMA_FRAGS(version
)) {
2297 CERROR("Can't accept %s(version %x): incompatible max_frags %d (%d wanted)\n",
2298 libcfs_nid2str(nid
), version
,
2299 reqmsg
->ibm_u
.connparams
.ibcp_max_frags
,
2300 IBLND_RDMA_FRAGS(version
));
2302 if (version
== IBLND_MSG_VERSION
)
2303 rej
.ibr_why
= IBLND_REJECT_RDMA_FRAGS
;
2309 if (reqmsg
->ibm_u
.connparams
.ibcp_max_msg_size
> IBLND_MSG_SIZE
) {
2310 CERROR("Can't accept %s: message size %d too big (%d max)\n",
2311 libcfs_nid2str(nid
),
2312 reqmsg
->ibm_u
.connparams
.ibcp_max_msg_size
,
2317 /* assume 'nid' is a new peer; create */
2318 rc
= kiblnd_create_peer(ni
, &peer
, nid
);
2320 CERROR("Can't create peer for %s\n", libcfs_nid2str(nid
));
2321 rej
.ibr_why
= IBLND_REJECT_NO_RESOURCES
;
2325 write_lock_irqsave(g_lock
, flags
);
2327 peer2
= kiblnd_find_peer_locked(nid
);
2328 if (peer2
!= NULL
) {
2329 if (peer2
->ibp_version
== 0) {
2330 peer2
->ibp_version
= version
;
2331 peer2
->ibp_incarnation
= reqmsg
->ibm_srcstamp
;
2334 /* not the guy I've talked with */
2335 if (peer2
->ibp_incarnation
!= reqmsg
->ibm_srcstamp
||
2336 peer2
->ibp_version
!= version
) {
2337 kiblnd_close_peer_conns_locked(peer2
, -ESTALE
);
2338 write_unlock_irqrestore(g_lock
, flags
);
2340 CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
2341 libcfs_nid2str(nid
), peer2
->ibp_version
, version
);
2343 kiblnd_peer_decref(peer
);
2344 rej
.ibr_why
= IBLND_REJECT_CONN_STALE
;
2348 /* tie-break connection race in favour of the higher NID */
2349 if (peer2
->ibp_connecting
!= 0 &&
2351 write_unlock_irqrestore(g_lock
, flags
);
2353 CWARN("Conn race %s\n", libcfs_nid2str(peer2
->ibp_nid
));
2355 kiblnd_peer_decref(peer
);
2356 rej
.ibr_why
= IBLND_REJECT_CONN_RACE
;
2360 peer2
->ibp_accepting
++;
2361 kiblnd_peer_addref(peer2
);
2363 write_unlock_irqrestore(g_lock
, flags
);
2364 kiblnd_peer_decref(peer
);
2367 /* Brand new peer */
2368 LASSERT(peer
->ibp_accepting
== 0);
2369 LASSERT(peer
->ibp_version
== 0 &&
2370 peer
->ibp_incarnation
== 0);
2372 peer
->ibp_accepting
= 1;
2373 peer
->ibp_version
= version
;
2374 peer
->ibp_incarnation
= reqmsg
->ibm_srcstamp
;
2376 /* I have a ref on ni that prevents it being shutdown */
2377 LASSERT(net
->ibn_shutdown
== 0);
2379 kiblnd_peer_addref(peer
);
2380 list_add_tail(&peer
->ibp_list
, kiblnd_nid2peerlist(nid
));
2382 write_unlock_irqrestore(g_lock
, flags
);
2385 conn
= kiblnd_create_conn(peer
, cmid
, IBLND_CONN_PASSIVE_WAIT
, version
);
2387 kiblnd_peer_connect_failed(peer
, 0, -ENOMEM
);
2388 kiblnd_peer_decref(peer
);
2389 rej
.ibr_why
= IBLND_REJECT_NO_RESOURCES
;
2393 /* conn now "owns" cmid, so I return success from here on to ensure the
2394 * CM callback doesn't destroy cmid. */
2396 conn
->ibc_incarnation
= reqmsg
->ibm_srcstamp
;
2397 conn
->ibc_credits
= IBLND_MSG_QUEUE_SIZE(version
);
2398 conn
->ibc_reserved_credits
= IBLND_MSG_QUEUE_SIZE(version
);
2399 LASSERT(conn
->ibc_credits
+ conn
->ibc_reserved_credits
+ IBLND_OOB_MSGS(version
)
2400 <= IBLND_RX_MSGS(version
));
2402 ackmsg
= &conn
->ibc_connvars
->cv_msg
;
2403 memset(ackmsg
, 0, sizeof(*ackmsg
));
2405 kiblnd_init_msg(ackmsg
, IBLND_MSG_CONNACK
,
2406 sizeof(ackmsg
->ibm_u
.connparams
));
2407 ackmsg
->ibm_u
.connparams
.ibcp_queue_depth
= IBLND_MSG_QUEUE_SIZE(version
);
2408 ackmsg
->ibm_u
.connparams
.ibcp_max_msg_size
= IBLND_MSG_SIZE
;
2409 ackmsg
->ibm_u
.connparams
.ibcp_max_frags
= IBLND_RDMA_FRAGS(version
);
2411 kiblnd_pack_msg(ni
, ackmsg
, version
, 0, nid
, reqmsg
->ibm_srcstamp
);
2413 memset(&cp
, 0, sizeof(cp
));
2414 cp
.private_data
= ackmsg
;
2415 cp
.private_data_len
= ackmsg
->ibm_nob
;
2416 cp
.responder_resources
= 0; /* No atomic ops or RDMA reads */
2417 cp
.initiator_depth
= 0;
2418 cp
.flow_control
= 1;
2419 cp
.retry_count
= *kiblnd_tunables
.kib_retry_count
;
2420 cp
.rnr_retry_count
= *kiblnd_tunables
.kib_rnr_retry_count
;
2422 CDEBUG(D_NET
, "Accept %s\n", libcfs_nid2str(nid
));
2424 rc
= rdma_accept(cmid
, &cp
);
2426 CERROR("Can't accept %s: %d\n", libcfs_nid2str(nid
), rc
);
2427 rej
.ibr_version
= version
;
2428 rej
.ibr_why
= IBLND_REJECT_FATAL
;
2430 kiblnd_reject(cmid
, &rej
);
2431 kiblnd_connreq_done(conn
, rc
);
2432 kiblnd_conn_decref(conn
);
2442 rej
.ibr_version
= version
;
2443 rej
.ibr_cp
.ibcp_queue_depth
= IBLND_MSG_QUEUE_SIZE(version
);
2444 rej
.ibr_cp
.ibcp_max_frags
= IBLND_RDMA_FRAGS(version
);
2445 kiblnd_reject(cmid
, &rej
);
2447 return -ECONNREFUSED
;
2451 kiblnd_reconnect(kib_conn_t
*conn
, int version
,
2452 __u64 incarnation
, int why
, kib_connparams_t
*cp
)
2454 kib_peer_t
*peer
= conn
->ibc_peer
;
2457 unsigned long flags
;
2459 LASSERT(conn
->ibc_state
== IBLND_CONN_ACTIVE_CONNECT
);
2460 LASSERT(peer
->ibp_connecting
> 0); /* 'conn' at least */
2462 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2464 /* retry connection if it's still needed and no other connection
2465 * attempts (active or passive) are in progress
2466 * NB: reconnect is still needed even when ibp_tx_queue is
2467 * empty if ibp_version != version because reconnect may be
2468 * initiated by kiblnd_query() */
2469 if ((!list_empty(&peer
->ibp_tx_queue
) ||
2470 peer
->ibp_version
!= version
) &&
2471 peer
->ibp_connecting
== 1 &&
2472 peer
->ibp_accepting
== 0) {
2474 peer
->ibp_connecting
++;
2476 peer
->ibp_version
= version
;
2477 peer
->ibp_incarnation
= incarnation
;
2480 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2490 case IBLND_REJECT_CONN_STALE
:
2494 case IBLND_REJECT_CONN_RACE
:
2495 reason
= "conn race";
2498 case IBLND_REJECT_CONN_UNCOMPAT
:
2499 reason
= "version negotiation";
2503 CNETERR("%s: retrying (%s), %x, %x, queue_dep: %d, max_frag: %d, msg_size: %d\n",
2504 libcfs_nid2str(peer
->ibp_nid
),
2505 reason
, IBLND_MSG_VERSION
, version
,
2506 cp
!= NULL
? cp
->ibcp_queue_depth
: IBLND_MSG_QUEUE_SIZE(version
),
2507 cp
!= NULL
? cp
->ibcp_max_frags
: IBLND_RDMA_FRAGS(version
),
2508 cp
!= NULL
? cp
->ibcp_max_msg_size
: IBLND_MSG_SIZE
);
2510 kiblnd_connect_peer(peer
);
2514 kiblnd_rejected(kib_conn_t
*conn
, int reason
, void *priv
, int priv_nob
)
2516 kib_peer_t
*peer
= conn
->ibc_peer
;
2518 LASSERT(!in_interrupt());
2519 LASSERT(conn
->ibc_state
== IBLND_CONN_ACTIVE_CONNECT
);
2522 case IB_CM_REJ_STALE_CONN
:
2523 kiblnd_reconnect(conn
, IBLND_MSG_VERSION
, 0,
2524 IBLND_REJECT_CONN_STALE
, NULL
);
2527 case IB_CM_REJ_INVALID_SERVICE_ID
:
2528 CNETERR("%s rejected: no listener at %d\n",
2529 libcfs_nid2str(peer
->ibp_nid
),
2530 *kiblnd_tunables
.kib_service
);
2533 case IB_CM_REJ_CONSUMER_DEFINED
:
2534 if (priv_nob
>= offsetof(kib_rej_t
, ibr_padding
)) {
2535 kib_rej_t
*rej
= priv
;
2536 kib_connparams_t
*cp
= NULL
;
2538 __u64 incarnation
= -1;
2540 /* NB. default incarnation is -1 because:
2541 * a) V1 will ignore dst incarnation in connreq.
2542 * b) V2 will provide incarnation while rejecting me,
2543 * -1 will be overwrote.
2545 * if I try to connect to a V1 peer with V2 protocol,
2546 * it rejected me then upgrade to V2, I have no idea
2547 * about the upgrading and try to reconnect with V1,
2548 * in this case upgraded V2 can find out I'm trying to
2549 * talk to the old guy and reject me(incarnation is -1).
2552 if (rej
->ibr_magic
== __swab32(IBLND_MSG_MAGIC
) ||
2553 rej
->ibr_magic
== __swab32(LNET_PROTO_MAGIC
)) {
2554 __swab32s(&rej
->ibr_magic
);
2555 __swab16s(&rej
->ibr_version
);
2559 if (priv_nob
>= sizeof(kib_rej_t
) &&
2560 rej
->ibr_version
> IBLND_MSG_VERSION_1
) {
2561 /* priv_nob is always 148 in current version
2562 * of OFED, so we still need to check version.
2563 * (define of IB_CM_REJ_PRIVATE_DATA_SIZE) */
2567 __swab64s(&rej
->ibr_incarnation
);
2568 __swab16s(&cp
->ibcp_queue_depth
);
2569 __swab16s(&cp
->ibcp_max_frags
);
2570 __swab32s(&cp
->ibcp_max_msg_size
);
2573 incarnation
= rej
->ibr_incarnation
;
2576 if (rej
->ibr_magic
!= IBLND_MSG_MAGIC
&&
2577 rej
->ibr_magic
!= LNET_PROTO_MAGIC
) {
2578 CERROR("%s rejected: consumer defined fatal error\n",
2579 libcfs_nid2str(peer
->ibp_nid
));
2583 if (rej
->ibr_version
!= IBLND_MSG_VERSION
&&
2584 rej
->ibr_version
!= IBLND_MSG_VERSION_1
) {
2585 CERROR("%s rejected: o2iblnd version %x error\n",
2586 libcfs_nid2str(peer
->ibp_nid
),
2591 if (rej
->ibr_why
== IBLND_REJECT_FATAL
&&
2592 rej
->ibr_version
== IBLND_MSG_VERSION_1
) {
2593 CDEBUG(D_NET
, "rejected by old version peer %s: %x\n",
2594 libcfs_nid2str(peer
->ibp_nid
), rej
->ibr_version
);
2596 if (conn
->ibc_version
!= IBLND_MSG_VERSION_1
)
2597 rej
->ibr_why
= IBLND_REJECT_CONN_UNCOMPAT
;
2600 switch (rej
->ibr_why
) {
2601 case IBLND_REJECT_CONN_RACE
:
2602 case IBLND_REJECT_CONN_STALE
:
2603 case IBLND_REJECT_CONN_UNCOMPAT
:
2604 kiblnd_reconnect(conn
, rej
->ibr_version
,
2605 incarnation
, rej
->ibr_why
, cp
);
2608 case IBLND_REJECT_MSG_QUEUE_SIZE
:
2609 CERROR("%s rejected: incompatible message queue depth %d, %d\n",
2610 libcfs_nid2str(peer
->ibp_nid
),
2611 cp
!= NULL
? cp
->ibcp_queue_depth
:
2612 IBLND_MSG_QUEUE_SIZE(rej
->ibr_version
),
2613 IBLND_MSG_QUEUE_SIZE(conn
->ibc_version
));
2616 case IBLND_REJECT_RDMA_FRAGS
:
2617 CERROR("%s rejected: incompatible # of RDMA fragments %d, %d\n",
2618 libcfs_nid2str(peer
->ibp_nid
),
2619 cp
!= NULL
? cp
->ibcp_max_frags
:
2620 IBLND_RDMA_FRAGS(rej
->ibr_version
),
2621 IBLND_RDMA_FRAGS(conn
->ibc_version
));
2624 case IBLND_REJECT_NO_RESOURCES
:
2625 CERROR("%s rejected: o2iblnd no resources\n",
2626 libcfs_nid2str(peer
->ibp_nid
));
2629 case IBLND_REJECT_FATAL
:
2630 CERROR("%s rejected: o2iblnd fatal error\n",
2631 libcfs_nid2str(peer
->ibp_nid
));
2635 CERROR("%s rejected: o2iblnd reason %d\n",
2636 libcfs_nid2str(peer
->ibp_nid
),
2644 CNETERR("%s rejected: reason %d, size %d\n",
2645 libcfs_nid2str(peer
->ibp_nid
), reason
, priv_nob
);
2649 kiblnd_connreq_done(conn
, -ECONNREFUSED
);
2653 kiblnd_check_connreply(kib_conn_t
*conn
, void *priv
, int priv_nob
)
2655 kib_peer_t
*peer
= conn
->ibc_peer
;
2656 lnet_ni_t
*ni
= peer
->ibp_ni
;
2657 kib_net_t
*net
= ni
->ni_data
;
2658 kib_msg_t
*msg
= priv
;
2659 int ver
= conn
->ibc_version
;
2660 int rc
= kiblnd_unpack_msg(msg
, priv_nob
);
2661 unsigned long flags
;
2663 LASSERT(net
!= NULL
);
2666 CERROR("Can't unpack connack from %s: %d\n",
2667 libcfs_nid2str(peer
->ibp_nid
), rc
);
2671 if (msg
->ibm_type
!= IBLND_MSG_CONNACK
) {
2672 CERROR("Unexpected message %d from %s\n",
2673 msg
->ibm_type
, libcfs_nid2str(peer
->ibp_nid
));
2678 if (ver
!= msg
->ibm_version
) {
2679 CERROR("%s replied version %x is different with requested version %x\n",
2680 libcfs_nid2str(peer
->ibp_nid
), msg
->ibm_version
, ver
);
2685 if (msg
->ibm_u
.connparams
.ibcp_queue_depth
!=
2686 IBLND_MSG_QUEUE_SIZE(ver
)) {
2687 CERROR("%s has incompatible queue depth %d(%d wanted)\n",
2688 libcfs_nid2str(peer
->ibp_nid
),
2689 msg
->ibm_u
.connparams
.ibcp_queue_depth
,
2690 IBLND_MSG_QUEUE_SIZE(ver
));
2695 if (msg
->ibm_u
.connparams
.ibcp_max_frags
!=
2696 IBLND_RDMA_FRAGS(ver
)) {
2697 CERROR("%s has incompatible max_frags %d (%d wanted)\n",
2698 libcfs_nid2str(peer
->ibp_nid
),
2699 msg
->ibm_u
.connparams
.ibcp_max_frags
,
2700 IBLND_RDMA_FRAGS(ver
));
2705 if (msg
->ibm_u
.connparams
.ibcp_max_msg_size
> IBLND_MSG_SIZE
) {
2706 CERROR("%s max message size %d too big (%d max)\n",
2707 libcfs_nid2str(peer
->ibp_nid
),
2708 msg
->ibm_u
.connparams
.ibcp_max_msg_size
,
2714 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2715 if (msg
->ibm_dstnid
== ni
->ni_nid
&&
2716 msg
->ibm_dststamp
== net
->ibn_incarnation
)
2720 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2723 CERROR("Bad connection reply from %s, rc = %d, version: %x max_frags: %d\n",
2724 libcfs_nid2str(peer
->ibp_nid
), rc
,
2725 msg
->ibm_version
, msg
->ibm_u
.connparams
.ibcp_max_frags
);
2729 conn
->ibc_incarnation
= msg
->ibm_srcstamp
;
2731 conn
->ibc_reserved_credits
= IBLND_MSG_QUEUE_SIZE(ver
);
2732 LASSERT(conn
->ibc_credits
+ conn
->ibc_reserved_credits
+ IBLND_OOB_MSGS(ver
)
2733 <= IBLND_RX_MSGS(ver
));
2735 kiblnd_connreq_done(conn
, 0);
2739 /* NB My QP has already established itself, so I handle anything going
2740 * wrong here by setting ibc_comms_error.
2741 * kiblnd_connreq_done(0) moves the conn state to ESTABLISHED, but then
2742 * immediately tears it down. */
2745 conn
->ibc_comms_error
= rc
;
2746 kiblnd_connreq_done(conn
, 0);
2750 kiblnd_active_connect(struct rdma_cm_id
*cmid
)
2752 kib_peer_t
*peer
= (kib_peer_t
*)cmid
->context
;
2755 struct rdma_conn_param cp
;
2758 unsigned long flags
;
2761 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2763 incarnation
= peer
->ibp_incarnation
;
2764 version
= (peer
->ibp_version
== 0) ? IBLND_MSG_VERSION
:
2767 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2769 conn
= kiblnd_create_conn(peer
, cmid
, IBLND_CONN_ACTIVE_CONNECT
, version
);
2771 kiblnd_peer_connect_failed(peer
, 1, -ENOMEM
);
2772 kiblnd_peer_decref(peer
); /* lose cmid's ref */
2776 /* conn "owns" cmid now, so I return success from here on to ensure the
2777 * CM callback doesn't destroy cmid. conn also takes over cmid's ref
2780 msg
= &conn
->ibc_connvars
->cv_msg
;
2782 memset(msg
, 0, sizeof(*msg
));
2783 kiblnd_init_msg(msg
, IBLND_MSG_CONNREQ
, sizeof(msg
->ibm_u
.connparams
));
2784 msg
->ibm_u
.connparams
.ibcp_queue_depth
= IBLND_MSG_QUEUE_SIZE(version
);
2785 msg
->ibm_u
.connparams
.ibcp_max_frags
= IBLND_RDMA_FRAGS(version
);
2786 msg
->ibm_u
.connparams
.ibcp_max_msg_size
= IBLND_MSG_SIZE
;
2788 kiblnd_pack_msg(peer
->ibp_ni
, msg
, version
,
2789 0, peer
->ibp_nid
, incarnation
);
2791 memset(&cp
, 0, sizeof(cp
));
2792 cp
.private_data
= msg
;
2793 cp
.private_data_len
= msg
->ibm_nob
;
2794 cp
.responder_resources
= 0; /* No atomic ops or RDMA reads */
2795 cp
.initiator_depth
= 0;
2796 cp
.flow_control
= 1;
2797 cp
.retry_count
= *kiblnd_tunables
.kib_retry_count
;
2798 cp
.rnr_retry_count
= *kiblnd_tunables
.kib_rnr_retry_count
;
2800 LASSERT(cmid
->context
== (void *)conn
);
2801 LASSERT(conn
->ibc_cmid
== cmid
);
2803 rc
= rdma_connect(cmid
, &cp
);
2805 CERROR("Can't connect to %s: %d\n",
2806 libcfs_nid2str(peer
->ibp_nid
), rc
);
2807 kiblnd_connreq_done(conn
, rc
);
2808 kiblnd_conn_decref(conn
);
2815 kiblnd_cm_callback(struct rdma_cm_id
*cmid
, struct rdma_cm_event
*event
)
2821 switch (event
->event
) {
2823 CERROR("Unexpected event: %d, status: %d\n",
2824 event
->event
, event
->status
);
2827 case RDMA_CM_EVENT_CONNECT_REQUEST
:
2828 /* destroy cmid on failure */
2829 rc
= kiblnd_passive_connect(cmid
,
2830 (void *)KIBLND_CONN_PARAM(event
),
2831 KIBLND_CONN_PARAM_LEN(event
));
2832 CDEBUG(D_NET
, "connreq: %d\n", rc
);
2835 case RDMA_CM_EVENT_ADDR_ERROR
:
2836 peer
= (kib_peer_t
*)cmid
->context
;
2837 CNETERR("%s: ADDR ERROR %d\n",
2838 libcfs_nid2str(peer
->ibp_nid
), event
->status
);
2839 kiblnd_peer_connect_failed(peer
, 1, -EHOSTUNREACH
);
2840 kiblnd_peer_decref(peer
);
2841 return -EHOSTUNREACH
; /* rc != 0 destroys cmid */
2843 case RDMA_CM_EVENT_ADDR_RESOLVED
:
2844 peer
= (kib_peer_t
*)cmid
->context
;
2846 CDEBUG(D_NET
, "%s Addr resolved: %d\n",
2847 libcfs_nid2str(peer
->ibp_nid
), event
->status
);
2849 if (event
->status
!= 0) {
2850 CNETERR("Can't resolve address for %s: %d\n",
2851 libcfs_nid2str(peer
->ibp_nid
), event
->status
);
2854 rc
= rdma_resolve_route(
2855 cmid
, *kiblnd_tunables
.kib_timeout
* 1000);
2858 /* Can't initiate route resolution */
2859 CERROR("Can't resolve route for %s: %d\n",
2860 libcfs_nid2str(peer
->ibp_nid
), rc
);
2862 kiblnd_peer_connect_failed(peer
, 1, rc
);
2863 kiblnd_peer_decref(peer
);
2864 return rc
; /* rc != 0 destroys cmid */
2866 case RDMA_CM_EVENT_ROUTE_ERROR
:
2867 peer
= (kib_peer_t
*)cmid
->context
;
2868 CNETERR("%s: ROUTE ERROR %d\n",
2869 libcfs_nid2str(peer
->ibp_nid
), event
->status
);
2870 kiblnd_peer_connect_failed(peer
, 1, -EHOSTUNREACH
);
2871 kiblnd_peer_decref(peer
);
2872 return -EHOSTUNREACH
; /* rc != 0 destroys cmid */
2874 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
2875 peer
= (kib_peer_t
*)cmid
->context
;
2876 CDEBUG(D_NET
, "%s Route resolved: %d\n",
2877 libcfs_nid2str(peer
->ibp_nid
), event
->status
);
2879 if (event
->status
== 0)
2880 return kiblnd_active_connect(cmid
);
2882 CNETERR("Can't resolve route for %s: %d\n",
2883 libcfs_nid2str(peer
->ibp_nid
), event
->status
);
2884 kiblnd_peer_connect_failed(peer
, 1, event
->status
);
2885 kiblnd_peer_decref(peer
);
2886 return event
->status
; /* rc != 0 destroys cmid */
2888 case RDMA_CM_EVENT_UNREACHABLE
:
2889 conn
= (kib_conn_t
*)cmid
->context
;
2890 LASSERT(conn
->ibc_state
== IBLND_CONN_ACTIVE_CONNECT
||
2891 conn
->ibc_state
== IBLND_CONN_PASSIVE_WAIT
);
2892 CNETERR("%s: UNREACHABLE %d\n",
2893 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), event
->status
);
2894 kiblnd_connreq_done(conn
, -ENETDOWN
);
2895 kiblnd_conn_decref(conn
);
2898 case RDMA_CM_EVENT_CONNECT_ERROR
:
2899 conn
= (kib_conn_t
*)cmid
->context
;
2900 LASSERT(conn
->ibc_state
== IBLND_CONN_ACTIVE_CONNECT
||
2901 conn
->ibc_state
== IBLND_CONN_PASSIVE_WAIT
);
2902 CNETERR("%s: CONNECT ERROR %d\n",
2903 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), event
->status
);
2904 kiblnd_connreq_done(conn
, -ENOTCONN
);
2905 kiblnd_conn_decref(conn
);
2908 case RDMA_CM_EVENT_REJECTED
:
2909 conn
= (kib_conn_t
*)cmid
->context
;
2910 switch (conn
->ibc_state
) {
2914 case IBLND_CONN_PASSIVE_WAIT
:
2915 CERROR("%s: REJECTED %d\n",
2916 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
),
2918 kiblnd_connreq_done(conn
, -ECONNRESET
);
2921 case IBLND_CONN_ACTIVE_CONNECT
:
2922 kiblnd_rejected(conn
, event
->status
,
2923 (void *)KIBLND_CONN_PARAM(event
),
2924 KIBLND_CONN_PARAM_LEN(event
));
2927 kiblnd_conn_decref(conn
);
2930 case RDMA_CM_EVENT_ESTABLISHED
:
2931 conn
= (kib_conn_t
*)cmid
->context
;
2932 switch (conn
->ibc_state
) {
2936 case IBLND_CONN_PASSIVE_WAIT
:
2937 CDEBUG(D_NET
, "ESTABLISHED (passive): %s\n",
2938 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
2939 kiblnd_connreq_done(conn
, 0);
2942 case IBLND_CONN_ACTIVE_CONNECT
:
2943 CDEBUG(D_NET
, "ESTABLISHED(active): %s\n",
2944 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
2945 kiblnd_check_connreply(conn
,
2946 (void *)KIBLND_CONN_PARAM(event
),
2947 KIBLND_CONN_PARAM_LEN(event
));
2950 /* net keeps its ref on conn! */
2953 case RDMA_CM_EVENT_TIMEWAIT_EXIT
:
2954 CDEBUG(D_NET
, "Ignore TIMEWAIT_EXIT event\n");
2956 case RDMA_CM_EVENT_DISCONNECTED
:
2957 conn
= (kib_conn_t
*)cmid
->context
;
2958 if (conn
->ibc_state
< IBLND_CONN_ESTABLISHED
) {
2959 CERROR("%s DISCONNECTED\n",
2960 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
2961 kiblnd_connreq_done(conn
, -ECONNRESET
);
2963 kiblnd_close_conn(conn
, 0);
2965 kiblnd_conn_decref(conn
);
2966 cmid
->context
= NULL
;
2969 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
2970 LCONSOLE_ERROR_MSG(0x131,
2971 "Received notification of device removal\n"
2972 "Please shutdown LNET to allow this to proceed\n");
2973 /* Can't remove network from underneath LNET for now, so I have
2977 case RDMA_CM_EVENT_ADDR_CHANGE
:
2978 LCONSOLE_INFO("Physical link changed (eg hca/port)\n");
2984 kiblnd_check_txs_locked(kib_conn_t
*conn
, struct list_head
*txs
)
2987 struct list_head
*ttmp
;
2989 list_for_each(ttmp
, txs
) {
2990 tx
= list_entry(ttmp
, kib_tx_t
, tx_list
);
2992 if (txs
!= &conn
->ibc_active_txs
) {
2993 LASSERT(tx
->tx_queued
);
2995 LASSERT(!tx
->tx_queued
);
2996 LASSERT(tx
->tx_waiting
|| tx
->tx_sending
!= 0);
2999 if (cfs_time_aftereq(jiffies
, tx
->tx_deadline
)) {
3000 CERROR("Timed out tx: %s, %lu seconds\n",
3001 kiblnd_queue2str(conn
, txs
),
3002 cfs_duration_sec(jiffies
- tx
->tx_deadline
));
3011 kiblnd_conn_timed_out_locked(kib_conn_t
*conn
)
3013 return kiblnd_check_txs_locked(conn
, &conn
->ibc_tx_queue
) ||
3014 kiblnd_check_txs_locked(conn
, &conn
->ibc_tx_noops
) ||
3015 kiblnd_check_txs_locked(conn
, &conn
->ibc_tx_queue_rsrvd
) ||
3016 kiblnd_check_txs_locked(conn
, &conn
->ibc_tx_queue_nocred
) ||
3017 kiblnd_check_txs_locked(conn
, &conn
->ibc_active_txs
);
3021 kiblnd_check_conns(int idx
)
3024 LIST_HEAD(checksends
);
3025 struct list_head
*peers
= &kiblnd_data
.kib_peers
[idx
];
3026 struct list_head
*ptmp
;
3030 struct list_head
*ctmp
;
3031 unsigned long flags
;
3033 /* NB. We expect to have a look at all the peers and not find any
3034 * RDMAs to time out, so we just use a shared lock while we
3036 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
3038 list_for_each(ptmp
, peers
) {
3039 peer
= list_entry(ptmp
, kib_peer_t
, ibp_list
);
3041 list_for_each(ctmp
, &peer
->ibp_conns
) {
3045 conn
= list_entry(ctmp
, kib_conn_t
, ibc_list
);
3047 LASSERT(conn
->ibc_state
== IBLND_CONN_ESTABLISHED
);
3049 spin_lock(&conn
->ibc_lock
);
3051 sendnoop
= kiblnd_need_noop(conn
);
3052 timedout
= kiblnd_conn_timed_out_locked(conn
);
3053 if (!sendnoop
&& !timedout
) {
3054 spin_unlock(&conn
->ibc_lock
);
3059 CERROR("Timed out RDMA with %s (%lu): c: %u, oc: %u, rc: %u\n",
3060 libcfs_nid2str(peer
->ibp_nid
),
3061 cfs_duration_sec(cfs_time_current() -
3062 peer
->ibp_last_alive
),
3064 conn
->ibc_outstanding_credits
,
3065 conn
->ibc_reserved_credits
);
3066 list_add(&conn
->ibc_connd_list
, &closes
);
3068 list_add(&conn
->ibc_connd_list
,
3071 /* +ref for 'closes' or 'checksends' */
3072 kiblnd_conn_addref(conn
);
3074 spin_unlock(&conn
->ibc_lock
);
3078 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
3080 /* Handle timeout by closing the whole
3081 * connection. We can only be sure RDMA activity
3082 * has ceased once the QP has been modified. */
3083 list_for_each_entry_safe(conn
, tmp
, &closes
, ibc_connd_list
) {
3084 list_del(&conn
->ibc_connd_list
);
3085 kiblnd_close_conn(conn
, -ETIMEDOUT
);
3086 kiblnd_conn_decref(conn
);
3089 /* In case we have enough credits to return via a
3090 * NOOP, but there were no non-blocking tx descs
3091 * free to do it last time... */
3092 while (!list_empty(&checksends
)) {
3093 conn
= list_entry(checksends
.next
,
3094 kib_conn_t
, ibc_connd_list
);
3095 list_del(&conn
->ibc_connd_list
);
3096 kiblnd_check_sends(conn
);
3097 kiblnd_conn_decref(conn
);
3102 kiblnd_disconnect_conn(kib_conn_t
*conn
)
3104 LASSERT(!in_interrupt());
3105 LASSERT(current
== kiblnd_data
.kib_connd
);
3106 LASSERT(conn
->ibc_state
== IBLND_CONN_CLOSING
);
3108 rdma_disconnect(conn
->ibc_cmid
);
3109 kiblnd_finalise_conn(conn
);
3111 kiblnd_peer_notify(conn
->ibc_peer
);
3115 kiblnd_connd(void *arg
)
3118 unsigned long flags
;
3124 unsigned long deadline
= jiffies
;
3126 cfs_block_allsigs();
3128 init_waitqueue_entry(&wait
, current
);
3129 kiblnd_data
.kib_connd
= current
;
3131 spin_lock_irqsave(&kiblnd_data
.kib_connd_lock
, flags
);
3133 while (!kiblnd_data
.kib_shutdown
) {
3137 if (!list_empty(&kiblnd_data
.kib_connd_zombies
)) {
3138 conn
= list_entry(kiblnd_data
. \
3139 kib_connd_zombies
.next
,
3140 kib_conn_t
, ibc_list
);
3141 list_del(&conn
->ibc_list
);
3143 spin_unlock_irqrestore(&kiblnd_data
.kib_connd_lock
,
3147 kiblnd_destroy_conn(conn
);
3149 spin_lock_irqsave(&kiblnd_data
.kib_connd_lock
, flags
);
3152 if (!list_empty(&kiblnd_data
.kib_connd_conns
)) {
3153 conn
= list_entry(kiblnd_data
.kib_connd_conns
.next
,
3154 kib_conn_t
, ibc_list
);
3155 list_del(&conn
->ibc_list
);
3157 spin_unlock_irqrestore(&kiblnd_data
.kib_connd_lock
,
3161 kiblnd_disconnect_conn(conn
);
3162 kiblnd_conn_decref(conn
);
3164 spin_lock_irqsave(&kiblnd_data
.kib_connd_lock
, flags
);
3167 /* careful with the jiffy wrap... */
3168 timeout
= (int)(deadline
- jiffies
);
3172 int chunk
= kiblnd_data
.kib_peer_hash_size
;
3174 spin_unlock_irqrestore(&kiblnd_data
.kib_connd_lock
, flags
);
3177 /* Time to check for RDMA timeouts on a few more
3178 * peers: I do checks every 'p' seconds on a
3179 * proportion of the peer table and I need to check
3180 * every connection 'n' times within a timeout
3181 * interval, to ensure I detect a timeout on any
3182 * connection within (n+1)/n times the timeout
3185 if (*kiblnd_tunables
.kib_timeout
> n
* p
)
3186 chunk
= (chunk
* n
* p
) /
3187 *kiblnd_tunables
.kib_timeout
;
3191 for (i
= 0; i
< chunk
; i
++) {
3192 kiblnd_check_conns(peer_index
);
3193 peer_index
= (peer_index
+ 1) %
3194 kiblnd_data
.kib_peer_hash_size
;
3198 spin_lock_irqsave(&kiblnd_data
.kib_connd_lock
, flags
);
3204 /* Nothing to do for 'timeout' */
3205 set_current_state(TASK_INTERRUPTIBLE
);
3206 add_wait_queue(&kiblnd_data
.kib_connd_waitq
, &wait
);
3207 spin_unlock_irqrestore(&kiblnd_data
.kib_connd_lock
, flags
);
3209 schedule_timeout(timeout
);
3211 remove_wait_queue(&kiblnd_data
.kib_connd_waitq
, &wait
);
3212 spin_lock_irqsave(&kiblnd_data
.kib_connd_lock
, flags
);
3215 spin_unlock_irqrestore(&kiblnd_data
.kib_connd_lock
, flags
);
3217 kiblnd_thread_fini();
3222 kiblnd_qp_event(struct ib_event
*event
, void *arg
)
3224 kib_conn_t
*conn
= arg
;
3226 switch (event
->event
) {
3227 case IB_EVENT_COMM_EST
:
3228 CDEBUG(D_NET
, "%s established\n",
3229 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
));
3233 CERROR("%s: Async QP event type %d\n",
3234 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), event
->event
);
3240 kiblnd_complete(struct ib_wc
*wc
)
3242 switch (kiblnd_wreqid2type(wc
->wr_id
)) {
3246 case IBLND_WID_RDMA
:
3247 /* We only get RDMA completion notification if it fails. All
3248 * subsequent work items, including the final SEND will fail
3249 * too. However we can't print out any more info about the
3250 * failing RDMA because 'tx' might be back on the idle list or
3251 * even reused already if we didn't manage to post all our work
3253 CNETERR("RDMA (tx: %p) failed: %d\n",
3254 kiblnd_wreqid2ptr(wc
->wr_id
), wc
->status
);
3258 kiblnd_tx_complete(kiblnd_wreqid2ptr(wc
->wr_id
), wc
->status
);
3262 kiblnd_rx_complete(kiblnd_wreqid2ptr(wc
->wr_id
), wc
->status
,
3269 kiblnd_cq_completion(struct ib_cq
*cq
, void *arg
)
3271 /* NB I'm not allowed to schedule this conn once its refcount has
3272 * reached 0. Since fundamentally I'm racing with scheduler threads
3273 * consuming my CQ I could be called after all completions have
3274 * occurred. But in this case, ibc_nrx == 0 && ibc_nsends_posted == 0
3275 * and this CQ is about to be destroyed so I NOOP. */
3276 kib_conn_t
*conn
= (kib_conn_t
*)arg
;
3277 struct kib_sched_info
*sched
= conn
->ibc_sched
;
3278 unsigned long flags
;
3280 LASSERT(cq
== conn
->ibc_cq
);
3282 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3284 conn
->ibc_ready
= 1;
3286 if (!conn
->ibc_scheduled
&&
3287 (conn
->ibc_nrx
> 0 ||
3288 conn
->ibc_nsends_posted
> 0)) {
3289 kiblnd_conn_addref(conn
); /* +1 ref for sched_conns */
3290 conn
->ibc_scheduled
= 1;
3291 list_add_tail(&conn
->ibc_sched_list
, &sched
->ibs_conns
);
3293 if (waitqueue_active(&sched
->ibs_waitq
))
3294 wake_up(&sched
->ibs_waitq
);
3297 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
3301 kiblnd_cq_event(struct ib_event
*event
, void *arg
)
3303 kib_conn_t
*conn
= arg
;
3305 CERROR("%s: async CQ event type %d\n",
3306 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), event
->event
);
3310 kiblnd_scheduler(void *arg
)
3312 long id
= (long)arg
;
3313 struct kib_sched_info
*sched
;
3316 unsigned long flags
;
3322 cfs_block_allsigs();
3324 init_waitqueue_entry(&wait
, current
);
3326 sched
= kiblnd_data
.kib_scheds
[KIB_THREAD_CPT(id
)];
3328 rc
= cfs_cpt_bind(lnet_cpt_table(), sched
->ibs_cpt
);
3330 CWARN("Failed to bind on CPT %d, please verify whether all CPUs are healthy and reload modules if necessary, otherwise your system might under risk of low performance\n",
3334 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3336 while (!kiblnd_data
.kib_shutdown
) {
3337 if (busy_loops
++ >= IBLND_RESCHED
) {
3338 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
3343 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3348 if (!list_empty(&sched
->ibs_conns
)) {
3349 conn
= list_entry(sched
->ibs_conns
.next
,
3350 kib_conn_t
, ibc_sched_list
);
3351 /* take over kib_sched_conns' ref on conn... */
3352 LASSERT(conn
->ibc_scheduled
);
3353 list_del(&conn
->ibc_sched_list
);
3354 conn
->ibc_ready
= 0;
3356 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
3358 rc
= ib_poll_cq(conn
->ibc_cq
, 1, &wc
);
3360 rc
= ib_req_notify_cq(conn
->ibc_cq
,
3363 CWARN("%s: ib_req_notify_cq failed: %d, closing connection\n",
3364 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
), rc
);
3365 kiblnd_close_conn(conn
, -EIO
);
3366 kiblnd_conn_decref(conn
);
3367 spin_lock_irqsave(&sched
->ibs_lock
,
3372 rc
= ib_poll_cq(conn
->ibc_cq
, 1, &wc
);
3376 CWARN("%s: ib_poll_cq failed: %d, closing connection\n",
3377 libcfs_nid2str(conn
->ibc_peer
->ibp_nid
),
3379 kiblnd_close_conn(conn
, -EIO
);
3380 kiblnd_conn_decref(conn
);
3381 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3385 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3387 if (rc
!= 0 || conn
->ibc_ready
) {
3388 /* There may be another completion waiting; get
3389 * another scheduler to check while I handle
3391 /* +1 ref for sched_conns */
3392 kiblnd_conn_addref(conn
);
3393 list_add_tail(&conn
->ibc_sched_list
,
3395 if (waitqueue_active(&sched
->ibs_waitq
))
3396 wake_up(&sched
->ibs_waitq
);
3398 conn
->ibc_scheduled
= 0;
3402 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
3403 kiblnd_complete(&wc
);
3405 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3408 kiblnd_conn_decref(conn
); /* ...drop my ref from above */
3415 set_current_state(TASK_INTERRUPTIBLE
);
3416 add_wait_queue_exclusive(&sched
->ibs_waitq
, &wait
);
3417 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
3422 remove_wait_queue(&sched
->ibs_waitq
, &wait
);
3423 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
3426 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
3428 kiblnd_thread_fini();
3433 kiblnd_failover_thread(void *arg
)
3435 rwlock_t
*glock
= &kiblnd_data
.kib_global_lock
;
3438 unsigned long flags
;
3441 LASSERT(*kiblnd_tunables
.kib_dev_failover
!= 0);
3443 cfs_block_allsigs();
3445 init_waitqueue_entry(&wait
, current
);
3446 write_lock_irqsave(glock
, flags
);
3448 while (!kiblnd_data
.kib_shutdown
) {
3449 int do_failover
= 0;
3452 list_for_each_entry(dev
, &kiblnd_data
.kib_failed_devs
,
3454 if (time_before(cfs_time_current(),
3455 dev
->ibd_next_failover
))
3462 list_del_init(&dev
->ibd_fail_list
);
3463 dev
->ibd_failover
= 1;
3464 write_unlock_irqrestore(glock
, flags
);
3466 rc
= kiblnd_dev_failover(dev
);
3468 write_lock_irqsave(glock
, flags
);
3470 LASSERT(dev
->ibd_failover
);
3471 dev
->ibd_failover
= 0;
3472 if (rc
>= 0) { /* Device is OK or failover succeed */
3473 dev
->ibd_next_failover
= cfs_time_shift(3);
3477 /* failed to failover, retry later */
3478 dev
->ibd_next_failover
=
3479 cfs_time_shift(min(dev
->ibd_failed_failover
, 10));
3480 if (kiblnd_dev_can_failover(dev
)) {
3481 list_add_tail(&dev
->ibd_fail_list
,
3482 &kiblnd_data
.kib_failed_devs
);
3488 /* long sleep if no more pending failover */
3489 long_sleep
= list_empty(&kiblnd_data
.kib_failed_devs
);
3491 set_current_state(TASK_INTERRUPTIBLE
);
3492 add_wait_queue(&kiblnd_data
.kib_failover_waitq
, &wait
);
3493 write_unlock_irqrestore(glock
, flags
);
3495 rc
= schedule_timeout(long_sleep
? cfs_time_seconds(10) :
3496 cfs_time_seconds(1));
3497 remove_wait_queue(&kiblnd_data
.kib_failover_waitq
, &wait
);
3498 write_lock_irqsave(glock
, flags
);
3500 if (!long_sleep
|| rc
!= 0)
3503 /* have a long sleep, routine check all active devices,
3504 * we need checking like this because if there is not active
3505 * connection on the dev and no SEND from local, we may listen
3506 * on wrong HCA for ever while there is a bonding failover */
3507 list_for_each_entry(dev
, &kiblnd_data
.kib_devs
, ibd_list
) {
3508 if (kiblnd_dev_can_failover(dev
)) {
3509 list_add_tail(&dev
->ibd_fail_list
,
3510 &kiblnd_data
.kib_failed_devs
);
3515 write_unlock_irqrestore(glock
, flags
);
3517 kiblnd_thread_fini();