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) 2011, 2015, 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.c
38 * Author: Eric Barton <eric@bartonsoftware.com>
41 #include <asm/div64.h>
45 static lnd_t the_o2iblnd
= {
47 .lnd_startup
= kiblnd_startup
,
48 .lnd_shutdown
= kiblnd_shutdown
,
49 .lnd_ctl
= kiblnd_ctl
,
50 .lnd_query
= kiblnd_query
,
51 .lnd_send
= kiblnd_send
,
52 .lnd_recv
= kiblnd_recv
,
55 kib_data_t kiblnd_data
;
57 static __u32
kiblnd_cksum(void *ptr
, int nob
)
63 sum
= ((sum
<< 1) | (sum
>> 31)) + *c
++;
65 /* ensure I don't return 0 (== no checksum) */
66 return !sum
? 1 : sum
;
69 static char *kiblnd_msgtype2str(int type
)
72 case IBLND_MSG_CONNREQ
:
75 case IBLND_MSG_CONNACK
:
81 case IBLND_MSG_IMMEDIATE
:
84 case IBLND_MSG_PUT_REQ
:
87 case IBLND_MSG_PUT_NAK
:
90 case IBLND_MSG_PUT_ACK
:
93 case IBLND_MSG_PUT_DONE
:
96 case IBLND_MSG_GET_REQ
:
99 case IBLND_MSG_GET_DONE
:
107 static int kiblnd_msgtype2size(int type
)
109 const int hdr_size
= offsetof(kib_msg_t
, ibm_u
);
112 case IBLND_MSG_CONNREQ
:
113 case IBLND_MSG_CONNACK
:
114 return hdr_size
+ sizeof(kib_connparams_t
);
119 case IBLND_MSG_IMMEDIATE
:
120 return offsetof(kib_msg_t
, ibm_u
.immediate
.ibim_payload
[0]);
122 case IBLND_MSG_PUT_REQ
:
123 return hdr_size
+ sizeof(kib_putreq_msg_t
);
125 case IBLND_MSG_PUT_ACK
:
126 return hdr_size
+ sizeof(kib_putack_msg_t
);
128 case IBLND_MSG_GET_REQ
:
129 return hdr_size
+ sizeof(kib_get_msg_t
);
131 case IBLND_MSG_PUT_NAK
:
132 case IBLND_MSG_PUT_DONE
:
133 case IBLND_MSG_GET_DONE
:
134 return hdr_size
+ sizeof(kib_completion_msg_t
);
140 static int kiblnd_unpack_rd(kib_msg_t
*msg
, int flip
)
147 LASSERT(msg
->ibm_type
== IBLND_MSG_GET_REQ
||
148 msg
->ibm_type
== IBLND_MSG_PUT_ACK
);
150 rd
= msg
->ibm_type
== IBLND_MSG_GET_REQ
?
151 &msg
->ibm_u
.get
.ibgm_rd
:
152 &msg
->ibm_u
.putack
.ibpam_rd
;
155 __swab32s(&rd
->rd_key
);
156 __swab32s(&rd
->rd_nfrags
);
161 if (n
<= 0 || n
> IBLND_MAX_RDMA_FRAGS
) {
162 CERROR("Bad nfrags: %d, should be 0 < n <= %d\n",
163 n
, IBLND_MAX_RDMA_FRAGS
);
167 nob
= offsetof(kib_msg_t
, ibm_u
) +
168 kiblnd_rd_msg_size(rd
, msg
->ibm_type
, n
);
170 if (msg
->ibm_nob
< nob
) {
171 CERROR("Short %s: %d(%d)\n",
172 kiblnd_msgtype2str(msg
->ibm_type
), msg
->ibm_nob
, nob
);
179 for (i
= 0; i
< n
; i
++) {
180 __swab32s(&rd
->rd_frags
[i
].rf_nob
);
181 __swab64s(&rd
->rd_frags
[i
].rf_addr
);
187 void kiblnd_pack_msg(lnet_ni_t
*ni
, kib_msg_t
*msg
, int version
,
188 int credits
, lnet_nid_t dstnid
, __u64 dststamp
)
190 kib_net_t
*net
= ni
->ni_data
;
193 * CAVEAT EMPTOR! all message fields not set here should have been
194 * initialised previously.
196 msg
->ibm_magic
= IBLND_MSG_MAGIC
;
197 msg
->ibm_version
= version
;
199 msg
->ibm_credits
= credits
;
202 msg
->ibm_srcnid
= ni
->ni_nid
;
203 msg
->ibm_srcstamp
= net
->ibn_incarnation
;
204 msg
->ibm_dstnid
= dstnid
;
205 msg
->ibm_dststamp
= dststamp
;
207 if (*kiblnd_tunables
.kib_cksum
) {
208 /* NB ibm_cksum zero while computing cksum */
209 msg
->ibm_cksum
= kiblnd_cksum(msg
, msg
->ibm_nob
);
213 int kiblnd_unpack_msg(kib_msg_t
*msg
, int nob
)
215 const int hdr_size
= offsetof(kib_msg_t
, ibm_u
);
221 /* 6 bytes are enough to have received magic + version */
223 CERROR("Short message: %d\n", nob
);
227 if (msg
->ibm_magic
== IBLND_MSG_MAGIC
) {
229 } else if (msg
->ibm_magic
== __swab32(IBLND_MSG_MAGIC
)) {
232 CERROR("Bad magic: %08x\n", msg
->ibm_magic
);
236 version
= flip
? __swab16(msg
->ibm_version
) : msg
->ibm_version
;
237 if (version
!= IBLND_MSG_VERSION
&&
238 version
!= IBLND_MSG_VERSION_1
) {
239 CERROR("Bad version: %x\n", version
);
243 if (nob
< hdr_size
) {
244 CERROR("Short message: %d\n", nob
);
248 msg_nob
= flip
? __swab32(msg
->ibm_nob
) : msg
->ibm_nob
;
250 CERROR("Short message: got %d, wanted %d\n", nob
, msg_nob
);
255 * checksum must be computed with ibm_cksum zero and BEFORE anything
258 msg_cksum
= flip
? __swab32(msg
->ibm_cksum
) : msg
->ibm_cksum
;
261 msg_cksum
!= kiblnd_cksum(msg
, msg_nob
)) {
262 CERROR("Bad checksum\n");
266 msg
->ibm_cksum
= msg_cksum
;
269 /* leave magic unflipped as a clue to peer endianness */
270 msg
->ibm_version
= version
;
271 CLASSERT(sizeof(msg
->ibm_type
) == 1);
272 CLASSERT(sizeof(msg
->ibm_credits
) == 1);
273 msg
->ibm_nob
= msg_nob
;
274 __swab64s(&msg
->ibm_srcnid
);
275 __swab64s(&msg
->ibm_srcstamp
);
276 __swab64s(&msg
->ibm_dstnid
);
277 __swab64s(&msg
->ibm_dststamp
);
280 if (msg
->ibm_srcnid
== LNET_NID_ANY
) {
281 CERROR("Bad src nid: %s\n", libcfs_nid2str(msg
->ibm_srcnid
));
285 if (msg_nob
< kiblnd_msgtype2size(msg
->ibm_type
)) {
286 CERROR("Short %s: %d(%d)\n", kiblnd_msgtype2str(msg
->ibm_type
),
287 msg_nob
, kiblnd_msgtype2size(msg
->ibm_type
));
291 switch (msg
->ibm_type
) {
293 CERROR("Unknown message type %x\n", msg
->ibm_type
);
297 case IBLND_MSG_IMMEDIATE
:
298 case IBLND_MSG_PUT_REQ
:
301 case IBLND_MSG_PUT_ACK
:
302 case IBLND_MSG_GET_REQ
:
303 if (kiblnd_unpack_rd(msg
, flip
))
307 case IBLND_MSG_PUT_NAK
:
308 case IBLND_MSG_PUT_DONE
:
309 case IBLND_MSG_GET_DONE
:
311 __swab32s(&msg
->ibm_u
.completion
.ibcm_status
);
314 case IBLND_MSG_CONNREQ
:
315 case IBLND_MSG_CONNACK
:
317 __swab16s(&msg
->ibm_u
.connparams
.ibcp_queue_depth
);
318 __swab16s(&msg
->ibm_u
.connparams
.ibcp_max_frags
);
319 __swab32s(&msg
->ibm_u
.connparams
.ibcp_max_msg_size
);
326 int kiblnd_create_peer(lnet_ni_t
*ni
, kib_peer_t
**peerp
, lnet_nid_t nid
)
329 kib_net_t
*net
= ni
->ni_data
;
330 int cpt
= lnet_cpt_of_nid(nid
);
334 LASSERT(nid
!= LNET_NID_ANY
);
336 LIBCFS_CPT_ALLOC(peer
, lnet_cpt_table(), cpt
, sizeof(*peer
));
338 CERROR("Cannot allocate peer\n");
345 peer
->ibp_last_alive
= 0;
346 atomic_set(&peer
->ibp_refcount
, 1); /* 1 ref for caller */
348 INIT_LIST_HEAD(&peer
->ibp_list
); /* not in the peer table yet */
349 INIT_LIST_HEAD(&peer
->ibp_conns
);
350 INIT_LIST_HEAD(&peer
->ibp_tx_queue
);
352 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
354 /* always called with a ref on ni, which prevents ni being shutdown */
355 LASSERT(!net
->ibn_shutdown
);
357 /* npeers only grows with the global lock held */
358 atomic_inc(&net
->ibn_npeers
);
360 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
366 void kiblnd_destroy_peer(kib_peer_t
*peer
)
368 kib_net_t
*net
= peer
->ibp_ni
->ni_data
;
371 LASSERT(!atomic_read(&peer
->ibp_refcount
));
372 LASSERT(!kiblnd_peer_active(peer
));
373 LASSERT(!peer
->ibp_connecting
);
374 LASSERT(!peer
->ibp_accepting
);
375 LASSERT(list_empty(&peer
->ibp_conns
));
376 LASSERT(list_empty(&peer
->ibp_tx_queue
));
378 LIBCFS_FREE(peer
, sizeof(*peer
));
381 * NB a peer's connections keep a reference on their peer until
382 * they are destroyed, so we can be assured that _all_ state to do
383 * with this peer has been cleaned up when its refcount drops to
386 atomic_dec(&net
->ibn_npeers
);
389 kib_peer_t
*kiblnd_find_peer_locked(lnet_nid_t nid
)
392 * the caller is responsible for accounting the additional reference
395 struct list_head
*peer_list
= kiblnd_nid2peerlist(nid
);
396 struct list_head
*tmp
;
399 list_for_each(tmp
, peer_list
) {
400 peer
= list_entry(tmp
, kib_peer_t
, ibp_list
);
402 LASSERT(peer
->ibp_connecting
> 0 || /* creating conns */
403 peer
->ibp_accepting
> 0 ||
404 !list_empty(&peer
->ibp_conns
)); /* active conn */
406 if (peer
->ibp_nid
!= nid
)
409 CDEBUG(D_NET
, "got peer [%p] -> %s (%d) version: %x\n",
410 peer
, libcfs_nid2str(nid
),
411 atomic_read(&peer
->ibp_refcount
),
418 void kiblnd_unlink_peer_locked(kib_peer_t
*peer
)
420 LASSERT(list_empty(&peer
->ibp_conns
));
422 LASSERT(kiblnd_peer_active(peer
));
423 list_del_init(&peer
->ibp_list
);
424 /* lose peerlist's ref */
425 kiblnd_peer_decref(peer
);
428 static int kiblnd_get_peer_info(lnet_ni_t
*ni
, int index
,
429 lnet_nid_t
*nidp
, int *count
)
432 struct list_head
*ptmp
;
436 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
438 for (i
= 0; i
< kiblnd_data
.kib_peer_hash_size
; i
++) {
439 list_for_each(ptmp
, &kiblnd_data
.kib_peers
[i
]) {
440 peer
= list_entry(ptmp
, kib_peer_t
, ibp_list
);
441 LASSERT(peer
->ibp_connecting
> 0 ||
442 peer
->ibp_accepting
> 0 ||
443 !list_empty(&peer
->ibp_conns
));
445 if (peer
->ibp_ni
!= ni
)
451 *nidp
= peer
->ibp_nid
;
452 *count
= atomic_read(&peer
->ibp_refcount
);
454 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
,
460 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
464 static void kiblnd_del_peer_locked(kib_peer_t
*peer
)
466 struct list_head
*ctmp
;
467 struct list_head
*cnxt
;
470 if (list_empty(&peer
->ibp_conns
)) {
471 kiblnd_unlink_peer_locked(peer
);
473 list_for_each_safe(ctmp
, cnxt
, &peer
->ibp_conns
) {
474 conn
= list_entry(ctmp
, kib_conn_t
, ibc_list
);
476 kiblnd_close_conn_locked(conn
, 0);
478 /* NB closing peer's last conn unlinked it. */
481 * NB peer now unlinked; might even be freed if the peer table had the
486 static int kiblnd_del_peer(lnet_ni_t
*ni
, lnet_nid_t nid
)
489 struct list_head
*ptmp
;
490 struct list_head
*pnxt
;
498 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
500 if (nid
!= LNET_NID_ANY
) {
501 lo
= kiblnd_nid2peerlist(nid
) - kiblnd_data
.kib_peers
;
502 hi
= kiblnd_nid2peerlist(nid
) - kiblnd_data
.kib_peers
;
505 hi
= kiblnd_data
.kib_peer_hash_size
- 1;
508 for (i
= lo
; i
<= hi
; i
++) {
509 list_for_each_safe(ptmp
, pnxt
, &kiblnd_data
.kib_peers
[i
]) {
510 peer
= list_entry(ptmp
, kib_peer_t
, ibp_list
);
511 LASSERT(peer
->ibp_connecting
> 0 ||
512 peer
->ibp_accepting
> 0 ||
513 !list_empty(&peer
->ibp_conns
));
515 if (peer
->ibp_ni
!= ni
)
518 if (!(nid
== LNET_NID_ANY
|| peer
->ibp_nid
== nid
))
521 if (!list_empty(&peer
->ibp_tx_queue
)) {
522 LASSERT(list_empty(&peer
->ibp_conns
));
524 list_splice_init(&peer
->ibp_tx_queue
,
528 kiblnd_del_peer_locked(peer
);
529 rc
= 0; /* matched something */
533 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
535 kiblnd_txlist_done(ni
, &zombies
, -EIO
);
540 static kib_conn_t
*kiblnd_get_conn_by_idx(lnet_ni_t
*ni
, int index
)
543 struct list_head
*ptmp
;
545 struct list_head
*ctmp
;
549 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
551 for (i
= 0; i
< kiblnd_data
.kib_peer_hash_size
; i
++) {
552 list_for_each(ptmp
, &kiblnd_data
.kib_peers
[i
]) {
553 peer
= list_entry(ptmp
, kib_peer_t
, ibp_list
);
554 LASSERT(peer
->ibp_connecting
> 0 ||
555 peer
->ibp_accepting
> 0 ||
556 !list_empty(&peer
->ibp_conns
));
558 if (peer
->ibp_ni
!= ni
)
561 list_for_each(ctmp
, &peer
->ibp_conns
) {
565 conn
= list_entry(ctmp
, kib_conn_t
,
567 kiblnd_conn_addref(conn
);
568 read_unlock_irqrestore(
569 &kiblnd_data
.kib_global_lock
,
576 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
580 int kiblnd_translate_mtu(int value
)
600 static void kiblnd_setup_mtu_locked(struct rdma_cm_id
*cmid
)
604 /* XXX There is no path record for iWARP, set by netdev->change_mtu? */
605 if (!cmid
->route
.path_rec
)
608 mtu
= kiblnd_translate_mtu(*kiblnd_tunables
.kib_ib_mtu
);
611 cmid
->route
.path_rec
->mtu
= mtu
;
614 static int kiblnd_get_completion_vector(kib_conn_t
*conn
, int cpt
)
620 lnet_nid_t nid
= conn
->ibc_peer
->ibp_nid
;
622 vectors
= conn
->ibc_cmid
->device
->num_comp_vectors
;
626 mask
= cfs_cpt_cpumask(lnet_cpt_table(), cpt
);
630 /* hash NID to CPU id in this partition... */
631 off
= do_div(nid
, cpumask_weight(mask
));
632 for_each_cpu(i
, mask
) {
641 kib_conn_t
*kiblnd_create_conn(kib_peer_t
*peer
, struct rdma_cm_id
*cmid
,
642 int state
, int version
)
646 * If the new conn is created successfully it takes over the caller's
647 * ref on 'peer'. It also "owns" 'cmid' and destroys it when it itself
648 * is destroyed. On failure, the caller's ref on 'peer' remains and
649 * she must dispose of 'cmid'. (Actually I'd block forever if I tried
650 * to destroy 'cmid' here since I'm called from the CM which still has
651 * its ref on 'cmid').
653 rwlock_t
*glock
= &kiblnd_data
.kib_global_lock
;
654 kib_net_t
*net
= peer
->ibp_ni
->ni_data
;
656 struct ib_qp_init_attr
*init_qp_attr
;
657 struct kib_sched_info
*sched
;
658 struct ib_cq_init_attr cq_attr
= {};
667 LASSERT(!in_interrupt());
671 cpt
= lnet_cpt_of_nid(peer
->ibp_nid
);
672 sched
= kiblnd_data
.kib_scheds
[cpt
];
674 LASSERT(sched
->ibs_nthreads
> 0);
676 LIBCFS_CPT_ALLOC(init_qp_attr
, lnet_cpt_table(), cpt
,
677 sizeof(*init_qp_attr
));
679 CERROR("Can't allocate qp_attr for %s\n",
680 libcfs_nid2str(peer
->ibp_nid
));
684 LIBCFS_CPT_ALLOC(conn
, lnet_cpt_table(), cpt
, sizeof(*conn
));
686 CERROR("Can't allocate connection for %s\n",
687 libcfs_nid2str(peer
->ibp_nid
));
691 conn
->ibc_state
= IBLND_CONN_INIT
;
692 conn
->ibc_version
= version
;
693 conn
->ibc_peer
= peer
; /* I take the caller's ref */
694 cmid
->context
= conn
; /* for future CM callbacks */
695 conn
->ibc_cmid
= cmid
;
697 INIT_LIST_HEAD(&conn
->ibc_early_rxs
);
698 INIT_LIST_HEAD(&conn
->ibc_tx_noops
);
699 INIT_LIST_HEAD(&conn
->ibc_tx_queue
);
700 INIT_LIST_HEAD(&conn
->ibc_tx_queue_rsrvd
);
701 INIT_LIST_HEAD(&conn
->ibc_tx_queue_nocred
);
702 INIT_LIST_HEAD(&conn
->ibc_active_txs
);
703 spin_lock_init(&conn
->ibc_lock
);
705 LIBCFS_CPT_ALLOC(conn
->ibc_connvars
, lnet_cpt_table(), cpt
,
706 sizeof(*conn
->ibc_connvars
));
707 if (!conn
->ibc_connvars
) {
708 CERROR("Can't allocate in-progress connection state\n");
712 write_lock_irqsave(glock
, flags
);
713 if (dev
->ibd_failover
) {
714 write_unlock_irqrestore(glock
, flags
);
715 CERROR("%s: failover in progress\n", dev
->ibd_ifname
);
719 if (dev
->ibd_hdev
->ibh_ibdev
!= cmid
->device
) {
720 /* wakeup failover thread and teardown connection */
721 if (kiblnd_dev_can_failover(dev
)) {
722 list_add_tail(&dev
->ibd_fail_list
,
723 &kiblnd_data
.kib_failed_devs
);
724 wake_up(&kiblnd_data
.kib_failover_waitq
);
727 write_unlock_irqrestore(glock
, flags
);
728 CERROR("cmid HCA(%s), kib_dev(%s) need failover\n",
729 cmid
->device
->name
, dev
->ibd_ifname
);
733 kiblnd_hdev_addref_locked(dev
->ibd_hdev
);
734 conn
->ibc_hdev
= dev
->ibd_hdev
;
736 kiblnd_setup_mtu_locked(cmid
);
738 write_unlock_irqrestore(glock
, flags
);
740 LIBCFS_CPT_ALLOC(conn
->ibc_rxs
, lnet_cpt_table(), cpt
,
741 IBLND_RX_MSGS(version
) * sizeof(kib_rx_t
));
742 if (!conn
->ibc_rxs
) {
743 CERROR("Cannot allocate RX buffers\n");
747 rc
= kiblnd_alloc_pages(&conn
->ibc_rx_pages
, cpt
,
748 IBLND_RX_MSG_PAGES(version
));
752 kiblnd_map_rx_descs(conn
);
754 cq_attr
.cqe
= IBLND_CQ_ENTRIES(version
);
755 cq_attr
.comp_vector
= kiblnd_get_completion_vector(conn
, cpt
);
756 cq
= ib_create_cq(cmid
->device
,
757 kiblnd_cq_completion
, kiblnd_cq_event
, conn
,
760 CERROR("Can't create CQ: %ld, cqe: %d\n",
761 PTR_ERR(cq
), IBLND_CQ_ENTRIES(version
));
767 rc
= ib_req_notify_cq(cq
, IB_CQ_NEXT_COMP
);
769 CERROR("Can't request completion notificiation: %d\n", rc
);
773 init_qp_attr
->event_handler
= kiblnd_qp_event
;
774 init_qp_attr
->qp_context
= conn
;
775 init_qp_attr
->cap
.max_send_wr
= IBLND_SEND_WRS(version
);
776 init_qp_attr
->cap
.max_recv_wr
= IBLND_RECV_WRS(version
);
777 init_qp_attr
->cap
.max_send_sge
= 1;
778 init_qp_attr
->cap
.max_recv_sge
= 1;
779 init_qp_attr
->sq_sig_type
= IB_SIGNAL_REQ_WR
;
780 init_qp_attr
->qp_type
= IB_QPT_RC
;
781 init_qp_attr
->send_cq
= cq
;
782 init_qp_attr
->recv_cq
= cq
;
784 conn
->ibc_sched
= sched
;
786 rc
= rdma_create_qp(cmid
, conn
->ibc_hdev
->ibh_pd
, init_qp_attr
);
788 CERROR("Can't create QP: %d, send_wr: %d, recv_wr: %d\n",
789 rc
, init_qp_attr
->cap
.max_send_wr
,
790 init_qp_attr
->cap
.max_recv_wr
);
794 LIBCFS_FREE(init_qp_attr
, sizeof(*init_qp_attr
));
796 /* 1 ref for caller and each rxmsg */
797 atomic_set(&conn
->ibc_refcount
, 1 + IBLND_RX_MSGS(version
));
798 conn
->ibc_nrx
= IBLND_RX_MSGS(version
);
801 for (i
= 0; i
< IBLND_RX_MSGS(version
); i
++) {
802 rc
= kiblnd_post_rx(&conn
->ibc_rxs
[i
],
803 IBLND_POSTRX_NO_CREDIT
);
805 CERROR("Can't post rxmsg: %d\n", rc
);
807 /* Make posted receives complete */
808 kiblnd_abort_receives(conn
);
811 * correct # of posted buffers
812 * NB locking needed now I'm racing with completion
814 spin_lock_irqsave(&sched
->ibs_lock
, flags
);
815 conn
->ibc_nrx
-= IBLND_RX_MSGS(version
) - i
;
816 spin_unlock_irqrestore(&sched
->ibs_lock
, flags
);
819 * cmid will be destroyed by CM(ofed) after cm_callback
820 * returned, so we can't refer it anymore
821 * (by kiblnd_connd()->kiblnd_destroy_conn)
823 rdma_destroy_qp(conn
->ibc_cmid
);
824 conn
->ibc_cmid
= NULL
;
826 /* Drop my own and unused rxbuffer refcounts */
827 while (i
++ <= IBLND_RX_MSGS(version
))
828 kiblnd_conn_decref(conn
);
834 /* Init successful! */
835 LASSERT(state
== IBLND_CONN_ACTIVE_CONNECT
||
836 state
== IBLND_CONN_PASSIVE_WAIT
);
837 conn
->ibc_state
= state
;
840 atomic_inc(&net
->ibn_nconns
);
844 kiblnd_destroy_conn(conn
);
846 LIBCFS_FREE(init_qp_attr
, sizeof(*init_qp_attr
));
851 void kiblnd_destroy_conn(kib_conn_t
*conn
)
853 struct rdma_cm_id
*cmid
= conn
->ibc_cmid
;
854 kib_peer_t
*peer
= conn
->ibc_peer
;
857 LASSERT(!in_interrupt());
858 LASSERT(!atomic_read(&conn
->ibc_refcount
));
859 LASSERT(list_empty(&conn
->ibc_early_rxs
));
860 LASSERT(list_empty(&conn
->ibc_tx_noops
));
861 LASSERT(list_empty(&conn
->ibc_tx_queue
));
862 LASSERT(list_empty(&conn
->ibc_tx_queue_rsrvd
));
863 LASSERT(list_empty(&conn
->ibc_tx_queue_nocred
));
864 LASSERT(list_empty(&conn
->ibc_active_txs
));
865 LASSERT(!conn
->ibc_noops_posted
);
866 LASSERT(!conn
->ibc_nsends_posted
);
868 switch (conn
->ibc_state
) {
870 /* conn must be completely disengaged from the network */
873 case IBLND_CONN_DISCONNECTED
:
874 /* connvars should have been freed already */
875 LASSERT(!conn
->ibc_connvars
);
878 case IBLND_CONN_INIT
:
882 /* conn->ibc_cmid might be destroyed by CM already */
883 if (cmid
&& cmid
->qp
)
884 rdma_destroy_qp(cmid
);
887 rc
= ib_destroy_cq(conn
->ibc_cq
);
889 CWARN("Error destroying CQ: %d\n", rc
);
892 if (conn
->ibc_rx_pages
)
893 kiblnd_unmap_rx_descs(conn
);
896 LIBCFS_FREE(conn
->ibc_rxs
,
897 IBLND_RX_MSGS(conn
->ibc_version
)
901 if (conn
->ibc_connvars
)
902 LIBCFS_FREE(conn
->ibc_connvars
, sizeof(*conn
->ibc_connvars
));
905 kiblnd_hdev_decref(conn
->ibc_hdev
);
907 /* See CAVEAT EMPTOR above in kiblnd_create_conn */
908 if (conn
->ibc_state
!= IBLND_CONN_INIT
) {
909 kib_net_t
*net
= peer
->ibp_ni
->ni_data
;
911 kiblnd_peer_decref(peer
);
912 rdma_destroy_id(cmid
);
913 atomic_dec(&net
->ibn_nconns
);
916 LIBCFS_FREE(conn
, sizeof(*conn
));
919 int kiblnd_close_peer_conns_locked(kib_peer_t
*peer
, int why
)
922 struct list_head
*ctmp
;
923 struct list_head
*cnxt
;
926 list_for_each_safe(ctmp
, cnxt
, &peer
->ibp_conns
) {
927 conn
= list_entry(ctmp
, kib_conn_t
, ibc_list
);
929 CDEBUG(D_NET
, "Closing conn -> %s, version: %x, reason: %d\n",
930 libcfs_nid2str(peer
->ibp_nid
),
931 conn
->ibc_version
, why
);
933 kiblnd_close_conn_locked(conn
, why
);
940 int kiblnd_close_stale_conns_locked(kib_peer_t
*peer
,
941 int version
, __u64 incarnation
)
944 struct list_head
*ctmp
;
945 struct list_head
*cnxt
;
948 list_for_each_safe(ctmp
, cnxt
, &peer
->ibp_conns
) {
949 conn
= list_entry(ctmp
, kib_conn_t
, ibc_list
);
951 if (conn
->ibc_version
== version
&&
952 conn
->ibc_incarnation
== incarnation
)
956 "Closing stale conn -> %s version: %x, incarnation:%#llx(%x, %#llx)\n",
957 libcfs_nid2str(peer
->ibp_nid
),
958 conn
->ibc_version
, conn
->ibc_incarnation
,
959 version
, incarnation
);
961 kiblnd_close_conn_locked(conn
, -ESTALE
);
968 static int kiblnd_close_matching_conns(lnet_ni_t
*ni
, lnet_nid_t nid
)
971 struct list_head
*ptmp
;
972 struct list_head
*pnxt
;
979 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
981 if (nid
!= LNET_NID_ANY
) {
982 lo
= kiblnd_nid2peerlist(nid
) - kiblnd_data
.kib_peers
;
983 hi
= kiblnd_nid2peerlist(nid
) - kiblnd_data
.kib_peers
;
986 hi
= kiblnd_data
.kib_peer_hash_size
- 1;
989 for (i
= lo
; i
<= hi
; i
++) {
990 list_for_each_safe(ptmp
, pnxt
, &kiblnd_data
.kib_peers
[i
]) {
991 peer
= list_entry(ptmp
, kib_peer_t
, ibp_list
);
992 LASSERT(peer
->ibp_connecting
> 0 ||
993 peer
->ibp_accepting
> 0 ||
994 !list_empty(&peer
->ibp_conns
));
996 if (peer
->ibp_ni
!= ni
)
999 if (!(nid
== LNET_NID_ANY
|| nid
== peer
->ibp_nid
))
1002 count
+= kiblnd_close_peer_conns_locked(peer
, 0);
1006 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1008 /* wildcards always succeed */
1009 if (nid
== LNET_NID_ANY
)
1012 return !count
? -ENOENT
: 0;
1015 int kiblnd_ctl(lnet_ni_t
*ni
, unsigned int cmd
, void *arg
)
1017 struct libcfs_ioctl_data
*data
= arg
;
1021 case IOC_LIBCFS_GET_PEER
: {
1025 rc
= kiblnd_get_peer_info(ni
, data
->ioc_count
,
1027 data
->ioc_nid
= nid
;
1028 data
->ioc_count
= count
;
1032 case IOC_LIBCFS_DEL_PEER
: {
1033 rc
= kiblnd_del_peer(ni
, data
->ioc_nid
);
1036 case IOC_LIBCFS_GET_CONN
: {
1040 conn
= kiblnd_get_conn_by_idx(ni
, data
->ioc_count
);
1046 LASSERT(conn
->ibc_cmid
);
1047 data
->ioc_nid
= conn
->ibc_peer
->ibp_nid
;
1048 if (!conn
->ibc_cmid
->route
.path_rec
)
1049 data
->ioc_u32
[0] = 0; /* iWarp has no path MTU */
1052 ib_mtu_enum_to_int(conn
->ibc_cmid
->route
.path_rec
->mtu
);
1053 kiblnd_conn_decref(conn
);
1056 case IOC_LIBCFS_CLOSE_CONNECTION
: {
1057 rc
= kiblnd_close_matching_conns(ni
, data
->ioc_nid
);
1068 void kiblnd_query(lnet_ni_t
*ni
, lnet_nid_t nid
, unsigned long *when
)
1070 unsigned long last_alive
= 0;
1071 unsigned long now
= cfs_time_current();
1072 rwlock_t
*glock
= &kiblnd_data
.kib_global_lock
;
1074 unsigned long flags
;
1076 read_lock_irqsave(glock
, flags
);
1078 peer
= kiblnd_find_peer_locked(nid
);
1080 LASSERT(peer
->ibp_connecting
> 0 || /* creating conns */
1081 peer
->ibp_accepting
> 0 ||
1082 !list_empty(&peer
->ibp_conns
)); /* active conn */
1083 last_alive
= peer
->ibp_last_alive
;
1086 read_unlock_irqrestore(glock
, flags
);
1092 * peer is not persistent in hash, trigger peer creation
1093 * and connection establishment with a NULL tx
1096 kiblnd_launch_tx(ni
, NULL
, nid
);
1098 CDEBUG(D_NET
, "Peer %s %p, alive %ld secs ago\n",
1099 libcfs_nid2str(nid
), peer
,
1100 last_alive
? cfs_duration_sec(now
- last_alive
) : -1);
1103 void kiblnd_free_pages(kib_pages_t
*p
)
1105 int npages
= p
->ibp_npages
;
1108 for (i
= 0; i
< npages
; i
++) {
1109 if (p
->ibp_pages
[i
])
1110 __free_page(p
->ibp_pages
[i
]);
1113 LIBCFS_FREE(p
, offsetof(kib_pages_t
, ibp_pages
[npages
]));
1116 int kiblnd_alloc_pages(kib_pages_t
**pp
, int cpt
, int npages
)
1121 LIBCFS_CPT_ALLOC(p
, lnet_cpt_table(), cpt
,
1122 offsetof(kib_pages_t
, ibp_pages
[npages
]));
1124 CERROR("Can't allocate descriptor for %d pages\n", npages
);
1128 memset(p
, 0, offsetof(kib_pages_t
, ibp_pages
[npages
]));
1129 p
->ibp_npages
= npages
;
1131 for (i
= 0; i
< npages
; i
++) {
1132 p
->ibp_pages
[i
] = alloc_pages_node(
1133 cfs_cpt_spread_node(lnet_cpt_table(), cpt
),
1135 if (!p
->ibp_pages
[i
]) {
1136 CERROR("Can't allocate page %d of %d\n", i
, npages
);
1137 kiblnd_free_pages(p
);
1146 void kiblnd_unmap_rx_descs(kib_conn_t
*conn
)
1151 LASSERT(conn
->ibc_rxs
);
1152 LASSERT(conn
->ibc_hdev
);
1154 for (i
= 0; i
< IBLND_RX_MSGS(conn
->ibc_version
); i
++) {
1155 rx
= &conn
->ibc_rxs
[i
];
1157 LASSERT(rx
->rx_nob
>= 0); /* not posted */
1159 kiblnd_dma_unmap_single(conn
->ibc_hdev
->ibh_ibdev
,
1160 KIBLND_UNMAP_ADDR(rx
, rx_msgunmap
,
1162 IBLND_MSG_SIZE
, DMA_FROM_DEVICE
);
1165 kiblnd_free_pages(conn
->ibc_rx_pages
);
1167 conn
->ibc_rx_pages
= NULL
;
1170 void kiblnd_map_rx_descs(kib_conn_t
*conn
)
1178 for (pg_off
= ipg
= i
= 0; i
< IBLND_RX_MSGS(conn
->ibc_version
); i
++) {
1179 pg
= conn
->ibc_rx_pages
->ibp_pages
[ipg
];
1180 rx
= &conn
->ibc_rxs
[i
];
1183 rx
->rx_msg
= (kib_msg_t
*)(((char *)page_address(pg
)) + pg_off
);
1185 rx
->rx_msgaddr
= kiblnd_dma_map_single(conn
->ibc_hdev
->ibh_ibdev
,
1189 LASSERT(!kiblnd_dma_mapping_error(conn
->ibc_hdev
->ibh_ibdev
,
1191 KIBLND_UNMAP_ADDR_SET(rx
, rx_msgunmap
, rx
->rx_msgaddr
);
1193 CDEBUG(D_NET
, "rx %d: %p %#llx(%#llx)\n",
1194 i
, rx
->rx_msg
, rx
->rx_msgaddr
,
1195 (__u64
)(page_to_phys(pg
) + pg_off
));
1197 pg_off
+= IBLND_MSG_SIZE
;
1198 LASSERT(pg_off
<= PAGE_SIZE
);
1200 if (pg_off
== PAGE_SIZE
) {
1203 LASSERT(ipg
<= IBLND_RX_MSG_PAGES(conn
->ibc_version
));
1208 static void kiblnd_unmap_tx_pool(kib_tx_pool_t
*tpo
)
1210 kib_hca_dev_t
*hdev
= tpo
->tpo_hdev
;
1214 LASSERT(!tpo
->tpo_pool
.po_allocated
);
1219 for (i
= 0; i
< tpo
->tpo_pool
.po_size
; i
++) {
1220 tx
= &tpo
->tpo_tx_descs
[i
];
1221 kiblnd_dma_unmap_single(hdev
->ibh_ibdev
,
1222 KIBLND_UNMAP_ADDR(tx
, tx_msgunmap
,
1224 IBLND_MSG_SIZE
, DMA_TO_DEVICE
);
1227 kiblnd_hdev_decref(hdev
);
1228 tpo
->tpo_hdev
= NULL
;
1231 static kib_hca_dev_t
*kiblnd_current_hdev(kib_dev_t
*dev
)
1233 kib_hca_dev_t
*hdev
;
1234 unsigned long flags
;
1237 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1238 while (dev
->ibd_failover
) {
1239 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1241 CDEBUG(D_NET
, "%s: Wait for failover\n",
1243 schedule_timeout(cfs_time_seconds(1) / 100);
1245 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1248 kiblnd_hdev_addref_locked(dev
->ibd_hdev
);
1249 hdev
= dev
->ibd_hdev
;
1251 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1256 static void kiblnd_map_tx_pool(kib_tx_pool_t
*tpo
)
1258 kib_pages_t
*txpgs
= tpo
->tpo_tx_pages
;
1259 kib_pool_t
*pool
= &tpo
->tpo_pool
;
1260 kib_net_t
*net
= pool
->po_owner
->ps_net
;
1272 /* pre-mapped messages are not bigger than 1 page */
1273 CLASSERT(IBLND_MSG_SIZE
<= PAGE_SIZE
);
1275 /* No fancy arithmetic when we do the buffer calculations */
1276 CLASSERT(!(PAGE_SIZE
% IBLND_MSG_SIZE
));
1278 tpo
->tpo_hdev
= kiblnd_current_hdev(dev
);
1280 for (ipage
= page_offset
= i
= 0; i
< pool
->po_size
; i
++) {
1281 page
= txpgs
->ibp_pages
[ipage
];
1282 tx
= &tpo
->tpo_tx_descs
[i
];
1284 tx
->tx_msg
= (kib_msg_t
*)(((char *)page_address(page
)) +
1287 tx
->tx_msgaddr
= kiblnd_dma_map_single(
1288 tpo
->tpo_hdev
->ibh_ibdev
, tx
->tx_msg
,
1289 IBLND_MSG_SIZE
, DMA_TO_DEVICE
);
1290 LASSERT(!kiblnd_dma_mapping_error(tpo
->tpo_hdev
->ibh_ibdev
,
1292 KIBLND_UNMAP_ADDR_SET(tx
, tx_msgunmap
, tx
->tx_msgaddr
);
1294 list_add(&tx
->tx_list
, &pool
->po_free_list
);
1296 page_offset
+= IBLND_MSG_SIZE
;
1297 LASSERT(page_offset
<= PAGE_SIZE
);
1299 if (page_offset
== PAGE_SIZE
) {
1302 LASSERT(ipage
<= txpgs
->ibp_npages
);
1307 struct ib_mr
*kiblnd_find_dma_mr(kib_hca_dev_t
*hdev
, __u64 addr
, __u64 size
)
1311 LASSERT(hdev
->ibh_mrs
[0]);
1313 if (hdev
->ibh_nmrs
== 1)
1314 return hdev
->ibh_mrs
[0];
1316 index
= addr
>> hdev
->ibh_mr_shift
;
1318 if (index
< hdev
->ibh_nmrs
&&
1319 index
== ((addr
+ size
- 1) >> hdev
->ibh_mr_shift
))
1320 return hdev
->ibh_mrs
[index
];
1325 struct ib_mr
*kiblnd_find_rd_dma_mr(kib_hca_dev_t
*hdev
, kib_rdma_desc_t
*rd
)
1327 struct ib_mr
*prev_mr
;
1331 LASSERT(hdev
->ibh_mrs
[0]);
1333 if (*kiblnd_tunables
.kib_map_on_demand
> 0 &&
1334 *kiblnd_tunables
.kib_map_on_demand
<= rd
->rd_nfrags
)
1337 if (hdev
->ibh_nmrs
== 1)
1338 return hdev
->ibh_mrs
[0];
1340 for (i
= 0, mr
= prev_mr
= NULL
;
1341 i
< rd
->rd_nfrags
; i
++) {
1342 mr
= kiblnd_find_dma_mr(hdev
,
1343 rd
->rd_frags
[i
].rf_addr
,
1344 rd
->rd_frags
[i
].rf_nob
);
1348 if (!mr
|| prev_mr
!= mr
) {
1349 /* Can't covered by one single MR */
1358 static void kiblnd_destroy_fmr_pool(kib_fmr_pool_t
*pool
)
1360 LASSERT(!pool
->fpo_map_count
);
1362 if (pool
->fpo_fmr_pool
)
1363 ib_destroy_fmr_pool(pool
->fpo_fmr_pool
);
1366 kiblnd_hdev_decref(pool
->fpo_hdev
);
1368 LIBCFS_FREE(pool
, sizeof(*pool
));
1371 static void kiblnd_destroy_fmr_pool_list(struct list_head
*head
)
1373 kib_fmr_pool_t
*pool
;
1375 while (!list_empty(head
)) {
1376 pool
= list_entry(head
->next
, kib_fmr_pool_t
, fpo_list
);
1377 list_del(&pool
->fpo_list
);
1378 kiblnd_destroy_fmr_pool(pool
);
1382 static int kiblnd_fmr_pool_size(int ncpts
)
1384 int size
= *kiblnd_tunables
.kib_fmr_pool_size
/ ncpts
;
1386 return max(IBLND_FMR_POOL
, size
);
1389 static int kiblnd_fmr_flush_trigger(int ncpts
)
1391 int size
= *kiblnd_tunables
.kib_fmr_flush_trigger
/ ncpts
;
1393 return max(IBLND_FMR_POOL_FLUSH
, size
);
1396 static int kiblnd_create_fmr_pool(kib_fmr_poolset_t
*fps
,
1397 kib_fmr_pool_t
**pp_fpo
)
1399 /* FMR pool for RDMA */
1400 kib_dev_t
*dev
= fps
->fps_net
->ibn_dev
;
1401 kib_fmr_pool_t
*fpo
;
1402 struct ib_fmr_pool_param param
= {
1403 .max_pages_per_fmr
= LNET_MAX_PAYLOAD
/ PAGE_SIZE
,
1404 .page_shift
= PAGE_SHIFT
,
1405 .access
= (IB_ACCESS_LOCAL_WRITE
|
1406 IB_ACCESS_REMOTE_WRITE
),
1407 .pool_size
= fps
->fps_pool_size
,
1408 .dirty_watermark
= fps
->fps_flush_trigger
,
1409 .flush_function
= NULL
,
1411 .cache
= !!*kiblnd_tunables
.kib_fmr_cache
};
1414 LIBCFS_CPT_ALLOC(fpo
, lnet_cpt_table(), fps
->fps_cpt
, sizeof(*fpo
));
1418 fpo
->fpo_hdev
= kiblnd_current_hdev(dev
);
1420 fpo
->fpo_fmr_pool
= ib_create_fmr_pool(fpo
->fpo_hdev
->ibh_pd
, ¶m
);
1421 if (IS_ERR(fpo
->fpo_fmr_pool
)) {
1422 rc
= PTR_ERR(fpo
->fpo_fmr_pool
);
1423 CERROR("Failed to create FMR pool: %d\n", rc
);
1425 kiblnd_hdev_decref(fpo
->fpo_hdev
);
1426 LIBCFS_FREE(fpo
, sizeof(*fpo
));
1430 fpo
->fpo_deadline
= cfs_time_shift(IBLND_POOL_DEADLINE
);
1431 fpo
->fpo_owner
= fps
;
1437 static void kiblnd_fail_fmr_poolset(kib_fmr_poolset_t
*fps
,
1438 struct list_head
*zombies
)
1440 if (!fps
->fps_net
) /* intialized? */
1443 spin_lock(&fps
->fps_lock
);
1445 while (!list_empty(&fps
->fps_pool_list
)) {
1446 kib_fmr_pool_t
*fpo
= list_entry(fps
->fps_pool_list
.next
,
1447 kib_fmr_pool_t
, fpo_list
);
1448 fpo
->fpo_failed
= 1;
1449 list_del(&fpo
->fpo_list
);
1450 if (!fpo
->fpo_map_count
)
1451 list_add(&fpo
->fpo_list
, zombies
);
1453 list_add(&fpo
->fpo_list
, &fps
->fps_failed_pool_list
);
1456 spin_unlock(&fps
->fps_lock
);
1459 static void kiblnd_fini_fmr_poolset(kib_fmr_poolset_t
*fps
)
1461 if (fps
->fps_net
) { /* initialized? */
1462 kiblnd_destroy_fmr_pool_list(&fps
->fps_failed_pool_list
);
1463 kiblnd_destroy_fmr_pool_list(&fps
->fps_pool_list
);
1467 static int kiblnd_init_fmr_poolset(kib_fmr_poolset_t
*fps
, int cpt
,
1468 kib_net_t
*net
, int pool_size
,
1471 kib_fmr_pool_t
*fpo
;
1474 memset(fps
, 0, sizeof(*fps
));
1478 fps
->fps_pool_size
= pool_size
;
1479 fps
->fps_flush_trigger
= flush_trigger
;
1480 spin_lock_init(&fps
->fps_lock
);
1481 INIT_LIST_HEAD(&fps
->fps_pool_list
);
1482 INIT_LIST_HEAD(&fps
->fps_failed_pool_list
);
1484 rc
= kiblnd_create_fmr_pool(fps
, &fpo
);
1486 list_add_tail(&fpo
->fpo_list
, &fps
->fps_pool_list
);
1491 static int kiblnd_fmr_pool_is_idle(kib_fmr_pool_t
*fpo
, unsigned long now
)
1493 if (fpo
->fpo_map_count
) /* still in use */
1495 if (fpo
->fpo_failed
)
1497 return cfs_time_aftereq(now
, fpo
->fpo_deadline
);
1500 void kiblnd_fmr_pool_unmap(kib_fmr_t
*fmr
, int status
)
1503 kib_fmr_pool_t
*fpo
= fmr
->fmr_pool
;
1504 kib_fmr_poolset_t
*fps
= fpo
->fpo_owner
;
1505 unsigned long now
= cfs_time_current();
1506 kib_fmr_pool_t
*tmp
;
1509 rc
= ib_fmr_pool_unmap(fmr
->fmr_pfmr
);
1513 rc
= ib_flush_fmr_pool(fpo
->fpo_fmr_pool
);
1517 fmr
->fmr_pool
= NULL
;
1518 fmr
->fmr_pfmr
= NULL
;
1520 spin_lock(&fps
->fps_lock
);
1521 fpo
->fpo_map_count
--; /* decref the pool */
1523 list_for_each_entry_safe(fpo
, tmp
, &fps
->fps_pool_list
, fpo_list
) {
1524 /* the first pool is persistent */
1525 if (fps
->fps_pool_list
.next
== &fpo
->fpo_list
)
1528 if (kiblnd_fmr_pool_is_idle(fpo
, now
)) {
1529 list_move(&fpo
->fpo_list
, &zombies
);
1533 spin_unlock(&fps
->fps_lock
);
1535 if (!list_empty(&zombies
))
1536 kiblnd_destroy_fmr_pool_list(&zombies
);
1539 int kiblnd_fmr_pool_map(kib_fmr_poolset_t
*fps
, __u64
*pages
, int npages
,
1540 __u64 iov
, kib_fmr_t
*fmr
)
1542 struct ib_pool_fmr
*pfmr
;
1543 kib_fmr_pool_t
*fpo
;
1548 spin_lock(&fps
->fps_lock
);
1549 version
= fps
->fps_version
;
1550 list_for_each_entry(fpo
, &fps
->fps_pool_list
, fpo_list
) {
1551 fpo
->fpo_deadline
= cfs_time_shift(IBLND_POOL_DEADLINE
);
1552 fpo
->fpo_map_count
++;
1553 spin_unlock(&fps
->fps_lock
);
1555 pfmr
= ib_fmr_pool_map_phys(fpo
->fpo_fmr_pool
,
1556 pages
, npages
, iov
);
1557 if (likely(!IS_ERR(pfmr
))) {
1558 fmr
->fmr_pool
= fpo
;
1559 fmr
->fmr_pfmr
= pfmr
;
1563 spin_lock(&fps
->fps_lock
);
1564 fpo
->fpo_map_count
--;
1565 if (PTR_ERR(pfmr
) != -EAGAIN
) {
1566 spin_unlock(&fps
->fps_lock
);
1567 return PTR_ERR(pfmr
);
1570 /* EAGAIN and ... */
1571 if (version
!= fps
->fps_version
) {
1572 spin_unlock(&fps
->fps_lock
);
1577 if (fps
->fps_increasing
) {
1578 spin_unlock(&fps
->fps_lock
);
1579 CDEBUG(D_NET
, "Another thread is allocating new FMR pool, waiting for her to complete\n");
1584 if (time_before(cfs_time_current(), fps
->fps_next_retry
)) {
1585 /* someone failed recently */
1586 spin_unlock(&fps
->fps_lock
);
1590 fps
->fps_increasing
= 1;
1591 spin_unlock(&fps
->fps_lock
);
1593 CDEBUG(D_NET
, "Allocate new FMR pool\n");
1594 rc
= kiblnd_create_fmr_pool(fps
, &fpo
);
1595 spin_lock(&fps
->fps_lock
);
1596 fps
->fps_increasing
= 0;
1599 list_add_tail(&fpo
->fpo_list
, &fps
->fps_pool_list
);
1601 fps
->fps_next_retry
= cfs_time_shift(IBLND_POOL_RETRY
);
1603 spin_unlock(&fps
->fps_lock
);
1608 static void kiblnd_fini_pool(kib_pool_t
*pool
)
1610 LASSERT(list_empty(&pool
->po_free_list
));
1611 LASSERT(!pool
->po_allocated
);
1613 CDEBUG(D_NET
, "Finalize %s pool\n", pool
->po_owner
->ps_name
);
1616 static void kiblnd_init_pool(kib_poolset_t
*ps
, kib_pool_t
*pool
, int size
)
1618 CDEBUG(D_NET
, "Initialize %s pool\n", ps
->ps_name
);
1620 memset(pool
, 0, sizeof(*pool
));
1621 INIT_LIST_HEAD(&pool
->po_free_list
);
1622 pool
->po_deadline
= cfs_time_shift(IBLND_POOL_DEADLINE
);
1623 pool
->po_owner
= ps
;
1624 pool
->po_size
= size
;
1627 static void kiblnd_destroy_pool_list(struct list_head
*head
)
1631 while (!list_empty(head
)) {
1632 pool
= list_entry(head
->next
, kib_pool_t
, po_list
);
1633 list_del(&pool
->po_list
);
1635 LASSERT(pool
->po_owner
);
1636 pool
->po_owner
->ps_pool_destroy(pool
);
1640 static void kiblnd_fail_poolset(kib_poolset_t
*ps
, struct list_head
*zombies
)
1642 if (!ps
->ps_net
) /* intialized? */
1645 spin_lock(&ps
->ps_lock
);
1646 while (!list_empty(&ps
->ps_pool_list
)) {
1647 kib_pool_t
*po
= list_entry(ps
->ps_pool_list
.next
,
1648 kib_pool_t
, po_list
);
1650 list_del(&po
->po_list
);
1651 if (!po
->po_allocated
)
1652 list_add(&po
->po_list
, zombies
);
1654 list_add(&po
->po_list
, &ps
->ps_failed_pool_list
);
1656 spin_unlock(&ps
->ps_lock
);
1659 static void kiblnd_fini_poolset(kib_poolset_t
*ps
)
1661 if (ps
->ps_net
) { /* initialized? */
1662 kiblnd_destroy_pool_list(&ps
->ps_failed_pool_list
);
1663 kiblnd_destroy_pool_list(&ps
->ps_pool_list
);
1667 static int kiblnd_init_poolset(kib_poolset_t
*ps
, int cpt
,
1668 kib_net_t
*net
, char *name
, int size
,
1669 kib_ps_pool_create_t po_create
,
1670 kib_ps_pool_destroy_t po_destroy
,
1671 kib_ps_node_init_t nd_init
,
1672 kib_ps_node_fini_t nd_fini
)
1677 memset(ps
, 0, sizeof(*ps
));
1681 ps
->ps_pool_create
= po_create
;
1682 ps
->ps_pool_destroy
= po_destroy
;
1683 ps
->ps_node_init
= nd_init
;
1684 ps
->ps_node_fini
= nd_fini
;
1685 ps
->ps_pool_size
= size
;
1686 if (strlcpy(ps
->ps_name
, name
, sizeof(ps
->ps_name
))
1687 >= sizeof(ps
->ps_name
))
1689 spin_lock_init(&ps
->ps_lock
);
1690 INIT_LIST_HEAD(&ps
->ps_pool_list
);
1691 INIT_LIST_HEAD(&ps
->ps_failed_pool_list
);
1693 rc
= ps
->ps_pool_create(ps
, size
, &pool
);
1695 list_add(&pool
->po_list
, &ps
->ps_pool_list
);
1697 CERROR("Failed to create the first pool for %s\n", ps
->ps_name
);
1702 static int kiblnd_pool_is_idle(kib_pool_t
*pool
, unsigned long now
)
1704 if (pool
->po_allocated
) /* still in use */
1706 if (pool
->po_failed
)
1708 return cfs_time_aftereq(now
, pool
->po_deadline
);
1711 void kiblnd_pool_free_node(kib_pool_t
*pool
, struct list_head
*node
)
1714 kib_poolset_t
*ps
= pool
->po_owner
;
1716 unsigned long now
= cfs_time_current();
1718 spin_lock(&ps
->ps_lock
);
1720 if (ps
->ps_node_fini
)
1721 ps
->ps_node_fini(pool
, node
);
1723 LASSERT(pool
->po_allocated
> 0);
1724 list_add(node
, &pool
->po_free_list
);
1725 pool
->po_allocated
--;
1727 list_for_each_entry_safe(pool
, tmp
, &ps
->ps_pool_list
, po_list
) {
1728 /* the first pool is persistent */
1729 if (ps
->ps_pool_list
.next
== &pool
->po_list
)
1732 if (kiblnd_pool_is_idle(pool
, now
))
1733 list_move(&pool
->po_list
, &zombies
);
1735 spin_unlock(&ps
->ps_lock
);
1737 if (!list_empty(&zombies
))
1738 kiblnd_destroy_pool_list(&zombies
);
1741 struct list_head
*kiblnd_pool_alloc_node(kib_poolset_t
*ps
)
1743 struct list_head
*node
;
1748 spin_lock(&ps
->ps_lock
);
1749 list_for_each_entry(pool
, &ps
->ps_pool_list
, po_list
) {
1750 if (list_empty(&pool
->po_free_list
))
1753 pool
->po_allocated
++;
1754 pool
->po_deadline
= cfs_time_shift(IBLND_POOL_DEADLINE
);
1755 node
= pool
->po_free_list
.next
;
1758 if (ps
->ps_node_init
) {
1759 /* still hold the lock */
1760 ps
->ps_node_init(pool
, node
);
1762 spin_unlock(&ps
->ps_lock
);
1766 /* no available tx pool and ... */
1767 if (ps
->ps_increasing
) {
1768 /* another thread is allocating a new pool */
1769 spin_unlock(&ps
->ps_lock
);
1770 CDEBUG(D_NET
, "Another thread is allocating new %s pool, waiting for her to complete\n",
1776 if (time_before(cfs_time_current(), ps
->ps_next_retry
)) {
1777 /* someone failed recently */
1778 spin_unlock(&ps
->ps_lock
);
1782 ps
->ps_increasing
= 1;
1783 spin_unlock(&ps
->ps_lock
);
1785 CDEBUG(D_NET
, "%s pool exhausted, allocate new pool\n", ps
->ps_name
);
1787 rc
= ps
->ps_pool_create(ps
, ps
->ps_pool_size
, &pool
);
1789 spin_lock(&ps
->ps_lock
);
1790 ps
->ps_increasing
= 0;
1792 list_add_tail(&pool
->po_list
, &ps
->ps_pool_list
);
1794 ps
->ps_next_retry
= cfs_time_shift(IBLND_POOL_RETRY
);
1795 CERROR("Can't allocate new %s pool because out of memory\n",
1798 spin_unlock(&ps
->ps_lock
);
1803 static void kiblnd_destroy_tx_pool(kib_pool_t
*pool
)
1805 kib_tx_pool_t
*tpo
= container_of(pool
, kib_tx_pool_t
, tpo_pool
);
1808 LASSERT(!pool
->po_allocated
);
1810 if (tpo
->tpo_tx_pages
) {
1811 kiblnd_unmap_tx_pool(tpo
);
1812 kiblnd_free_pages(tpo
->tpo_tx_pages
);
1815 if (!tpo
->tpo_tx_descs
)
1818 for (i
= 0; i
< pool
->po_size
; i
++) {
1819 kib_tx_t
*tx
= &tpo
->tpo_tx_descs
[i
];
1821 list_del(&tx
->tx_list
);
1823 LIBCFS_FREE(tx
->tx_pages
,
1825 sizeof(*tx
->tx_pages
));
1827 LIBCFS_FREE(tx
->tx_frags
,
1828 IBLND_MAX_RDMA_FRAGS
*
1829 sizeof(*tx
->tx_frags
));
1831 LIBCFS_FREE(tx
->tx_wrq
,
1832 (1 + IBLND_MAX_RDMA_FRAGS
) *
1833 sizeof(*tx
->tx_wrq
));
1835 LIBCFS_FREE(tx
->tx_sge
,
1836 (1 + IBLND_MAX_RDMA_FRAGS
) *
1837 sizeof(*tx
->tx_sge
));
1839 LIBCFS_FREE(tx
->tx_rd
,
1840 offsetof(kib_rdma_desc_t
,
1841 rd_frags
[IBLND_MAX_RDMA_FRAGS
]));
1844 LIBCFS_FREE(tpo
->tpo_tx_descs
,
1845 pool
->po_size
* sizeof(kib_tx_t
));
1847 kiblnd_fini_pool(pool
);
1848 LIBCFS_FREE(tpo
, sizeof(*tpo
));
1851 static int kiblnd_tx_pool_size(int ncpts
)
1853 int ntx
= *kiblnd_tunables
.kib_ntx
/ ncpts
;
1855 return max(IBLND_TX_POOL
, ntx
);
1858 static int kiblnd_create_tx_pool(kib_poolset_t
*ps
, int size
,
1866 LIBCFS_CPT_ALLOC(tpo
, lnet_cpt_table(), ps
->ps_cpt
, sizeof(*tpo
));
1868 CERROR("Failed to allocate TX pool\n");
1872 pool
= &tpo
->tpo_pool
;
1873 kiblnd_init_pool(ps
, pool
, size
);
1874 tpo
->tpo_tx_descs
= NULL
;
1875 tpo
->tpo_tx_pages
= NULL
;
1877 npg
= (size
* IBLND_MSG_SIZE
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
1878 if (kiblnd_alloc_pages(&tpo
->tpo_tx_pages
, ps
->ps_cpt
, npg
)) {
1879 CERROR("Can't allocate tx pages: %d\n", npg
);
1880 LIBCFS_FREE(tpo
, sizeof(*tpo
));
1884 LIBCFS_CPT_ALLOC(tpo
->tpo_tx_descs
, lnet_cpt_table(), ps
->ps_cpt
,
1885 size
* sizeof(kib_tx_t
));
1886 if (!tpo
->tpo_tx_descs
) {
1887 CERROR("Can't allocate %d tx descriptors\n", size
);
1888 ps
->ps_pool_destroy(pool
);
1892 memset(tpo
->tpo_tx_descs
, 0, size
* sizeof(kib_tx_t
));
1894 for (i
= 0; i
< size
; i
++) {
1895 kib_tx_t
*tx
= &tpo
->tpo_tx_descs
[i
];
1898 if (ps
->ps_net
->ibn_fmr_ps
) {
1899 LIBCFS_CPT_ALLOC(tx
->tx_pages
,
1900 lnet_cpt_table(), ps
->ps_cpt
,
1901 LNET_MAX_IOV
* sizeof(*tx
->tx_pages
));
1906 LIBCFS_CPT_ALLOC(tx
->tx_frags
, lnet_cpt_table(), ps
->ps_cpt
,
1907 IBLND_MAX_RDMA_FRAGS
* sizeof(*tx
->tx_frags
));
1911 sg_init_table(tx
->tx_frags
, IBLND_MAX_RDMA_FRAGS
);
1913 LIBCFS_CPT_ALLOC(tx
->tx_wrq
, lnet_cpt_table(), ps
->ps_cpt
,
1914 (1 + IBLND_MAX_RDMA_FRAGS
) *
1915 sizeof(*tx
->tx_wrq
));
1919 LIBCFS_CPT_ALLOC(tx
->tx_sge
, lnet_cpt_table(), ps
->ps_cpt
,
1920 (1 + IBLND_MAX_RDMA_FRAGS
) *
1921 sizeof(*tx
->tx_sge
));
1925 LIBCFS_CPT_ALLOC(tx
->tx_rd
, lnet_cpt_table(), ps
->ps_cpt
,
1926 offsetof(kib_rdma_desc_t
,
1927 rd_frags
[IBLND_MAX_RDMA_FRAGS
]));
1933 kiblnd_map_tx_pool(tpo
);
1938 ps
->ps_pool_destroy(pool
);
1942 static void kiblnd_tx_init(kib_pool_t
*pool
, struct list_head
*node
)
1944 kib_tx_poolset_t
*tps
= container_of(pool
->po_owner
, kib_tx_poolset_t
,
1946 kib_tx_t
*tx
= list_entry(node
, kib_tx_t
, tx_list
);
1948 tx
->tx_cookie
= tps
->tps_next_tx_cookie
++;
1951 static void kiblnd_net_fini_pools(kib_net_t
*net
)
1955 cfs_cpt_for_each(i
, lnet_cpt_table()) {
1956 kib_tx_poolset_t
*tps
;
1957 kib_fmr_poolset_t
*fps
;
1959 if (net
->ibn_tx_ps
) {
1960 tps
= net
->ibn_tx_ps
[i
];
1961 kiblnd_fini_poolset(&tps
->tps_poolset
);
1964 if (net
->ibn_fmr_ps
) {
1965 fps
= net
->ibn_fmr_ps
[i
];
1966 kiblnd_fini_fmr_poolset(fps
);
1970 if (net
->ibn_tx_ps
) {
1971 cfs_percpt_free(net
->ibn_tx_ps
);
1972 net
->ibn_tx_ps
= NULL
;
1975 if (net
->ibn_fmr_ps
) {
1976 cfs_percpt_free(net
->ibn_fmr_ps
);
1977 net
->ibn_fmr_ps
= NULL
;
1981 static int kiblnd_net_init_pools(kib_net_t
*net
, __u32
*cpts
, int ncpts
)
1983 unsigned long flags
;
1988 read_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
1989 if (!*kiblnd_tunables
.kib_map_on_demand
&&
1990 net
->ibn_dev
->ibd_hdev
->ibh_nmrs
== 1) {
1991 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1992 goto create_tx_pool
;
1995 read_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
1997 if (*kiblnd_tunables
.kib_fmr_pool_size
<
1998 *kiblnd_tunables
.kib_ntx
/ 4) {
1999 CERROR("Can't set fmr pool size (%d) < ntx / 4(%d)\n",
2000 *kiblnd_tunables
.kib_fmr_pool_size
,
2001 *kiblnd_tunables
.kib_ntx
/ 4);
2007 * TX pool must be created later than FMR, see LU-2268
2010 LASSERT(!net
->ibn_tx_ps
);
2013 * premapping can fail if ibd_nmr > 1, so we always create
2014 * FMR pool and map-on-demand if premapping failed
2017 net
->ibn_fmr_ps
= cfs_percpt_alloc(lnet_cpt_table(),
2018 sizeof(kib_fmr_poolset_t
));
2019 if (!net
->ibn_fmr_ps
) {
2020 CERROR("Failed to allocate FMR pool array\n");
2025 for (i
= 0; i
< ncpts
; i
++) {
2026 cpt
= !cpts
? i
: cpts
[i
];
2027 rc
= kiblnd_init_fmr_poolset(net
->ibn_fmr_ps
[cpt
], cpt
, net
,
2028 kiblnd_fmr_pool_size(ncpts
),
2029 kiblnd_fmr_flush_trigger(ncpts
));
2030 if (rc
== -ENOSYS
&& !i
) /* no FMR */
2033 if (rc
) { /* a real error */
2034 CERROR("Can't initialize FMR pool for CPT %d: %d\n",
2041 LASSERT(i
== ncpts
);
2042 goto create_tx_pool
;
2045 cfs_percpt_free(net
->ibn_fmr_ps
);
2046 net
->ibn_fmr_ps
= NULL
;
2048 CWARN("Device does not support FMR\n");
2052 net
->ibn_tx_ps
= cfs_percpt_alloc(lnet_cpt_table(),
2053 sizeof(kib_tx_poolset_t
));
2054 if (!net
->ibn_tx_ps
) {
2055 CERROR("Failed to allocate tx pool array\n");
2060 for (i
= 0; i
< ncpts
; i
++) {
2061 cpt
= !cpts
? i
: cpts
[i
];
2062 rc
= kiblnd_init_poolset(&net
->ibn_tx_ps
[cpt
]->tps_poolset
,
2064 kiblnd_tx_pool_size(ncpts
),
2065 kiblnd_create_tx_pool
,
2066 kiblnd_destroy_tx_pool
,
2067 kiblnd_tx_init
, NULL
);
2069 CERROR("Can't initialize TX pool for CPT %d: %d\n",
2077 kiblnd_net_fini_pools(net
);
2082 static int kiblnd_hdev_get_attr(kib_hca_dev_t
*hdev
)
2085 * It's safe to assume a HCA can handle a page size
2086 * matching that of the native system
2088 hdev
->ibh_page_shift
= PAGE_SHIFT
;
2089 hdev
->ibh_page_size
= 1 << PAGE_SHIFT
;
2090 hdev
->ibh_page_mask
= ~((__u64
)hdev
->ibh_page_size
- 1);
2092 hdev
->ibh_mr_size
= hdev
->ibh_ibdev
->attrs
.max_mr_size
;
2093 if (hdev
->ibh_mr_size
== ~0ULL) {
2094 hdev
->ibh_mr_shift
= 64;
2098 for (hdev
->ibh_mr_shift
= 0;
2099 hdev
->ibh_mr_shift
< 64; hdev
->ibh_mr_shift
++) {
2100 if (hdev
->ibh_mr_size
== (1ULL << hdev
->ibh_mr_shift
) ||
2101 hdev
->ibh_mr_size
== (1ULL << hdev
->ibh_mr_shift
) - 1)
2105 CERROR("Invalid mr size: %#llx\n", hdev
->ibh_mr_size
);
2109 static void kiblnd_hdev_cleanup_mrs(kib_hca_dev_t
*hdev
)
2113 if (!hdev
->ibh_nmrs
|| !hdev
->ibh_mrs
)
2116 for (i
= 0; i
< hdev
->ibh_nmrs
; i
++) {
2117 if (!hdev
->ibh_mrs
[i
])
2120 ib_dereg_mr(hdev
->ibh_mrs
[i
]);
2123 LIBCFS_FREE(hdev
->ibh_mrs
, sizeof(*hdev
->ibh_mrs
) * hdev
->ibh_nmrs
);
2124 hdev
->ibh_mrs
= NULL
;
2128 void kiblnd_hdev_destroy(kib_hca_dev_t
*hdev
)
2130 kiblnd_hdev_cleanup_mrs(hdev
);
2133 ib_dealloc_pd(hdev
->ibh_pd
);
2136 rdma_destroy_id(hdev
->ibh_cmid
);
2138 LIBCFS_FREE(hdev
, sizeof(*hdev
));
2141 static int kiblnd_hdev_setup_mrs(kib_hca_dev_t
*hdev
)
2145 int acflags
= IB_ACCESS_LOCAL_WRITE
| IB_ACCESS_REMOTE_WRITE
;
2147 rc
= kiblnd_hdev_get_attr(hdev
);
2151 LIBCFS_ALLOC(hdev
->ibh_mrs
, 1 * sizeof(*hdev
->ibh_mrs
));
2152 if (!hdev
->ibh_mrs
) {
2153 CERROR("Failed to allocate MRs table\n");
2157 hdev
->ibh_mrs
[0] = NULL
;
2160 mr
= ib_get_dma_mr(hdev
->ibh_pd
, acflags
);
2162 CERROR("Failed ib_get_dma_mr : %ld\n", PTR_ERR(mr
));
2163 kiblnd_hdev_cleanup_mrs(hdev
);
2167 hdev
->ibh_mrs
[0] = mr
;
2173 static int kiblnd_dummy_callback(struct rdma_cm_id
*cmid
,
2174 struct rdma_cm_event
*event
)
2179 static int kiblnd_dev_need_failover(kib_dev_t
*dev
)
2181 struct rdma_cm_id
*cmid
;
2182 struct sockaddr_in srcaddr
;
2183 struct sockaddr_in dstaddr
;
2186 if (!dev
->ibd_hdev
|| /* initializing */
2187 !dev
->ibd_hdev
->ibh_cmid
|| /* listener is dead */
2188 *kiblnd_tunables
.kib_dev_failover
> 1) /* debugging */
2192 * XXX: it's UGLY, but I don't have better way to find
2193 * ib-bonding HCA failover because:
2195 * a. no reliable CM event for HCA failover...
2196 * b. no OFED API to get ib_device for current net_device...
2198 * We have only two choices at this point:
2200 * a. rdma_bind_addr(), it will conflict with listener cmid
2201 * b. rdma_resolve_addr() to zero addr
2203 cmid
= kiblnd_rdma_create_id(kiblnd_dummy_callback
, dev
, RDMA_PS_TCP
,
2207 CERROR("Failed to create cmid for failover: %d\n", rc
);
2211 memset(&srcaddr
, 0, sizeof(srcaddr
));
2212 srcaddr
.sin_family
= AF_INET
;
2213 srcaddr
.sin_addr
.s_addr
= (__force u32
)htonl(dev
->ibd_ifip
);
2215 memset(&dstaddr
, 0, sizeof(dstaddr
));
2216 dstaddr
.sin_family
= AF_INET
;
2217 rc
= rdma_resolve_addr(cmid
, (struct sockaddr
*)&srcaddr
,
2218 (struct sockaddr
*)&dstaddr
, 1);
2219 if (rc
|| !cmid
->device
) {
2220 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2221 dev
->ibd_ifname
, &dev
->ibd_ifip
,
2223 rdma_destroy_id(cmid
);
2227 rc
= dev
->ibd_hdev
->ibh_ibdev
!= cmid
->device
; /* true for failover */
2228 rdma_destroy_id(cmid
);
2233 int kiblnd_dev_failover(kib_dev_t
*dev
)
2235 LIST_HEAD(zombie_tpo
);
2236 LIST_HEAD(zombie_ppo
);
2237 LIST_HEAD(zombie_fpo
);
2238 struct rdma_cm_id
*cmid
= NULL
;
2239 kib_hca_dev_t
*hdev
= NULL
;
2242 struct sockaddr_in addr
;
2243 unsigned long flags
;
2247 LASSERT(*kiblnd_tunables
.kib_dev_failover
> 1 ||
2248 dev
->ibd_can_failover
|| !dev
->ibd_hdev
);
2250 rc
= kiblnd_dev_need_failover(dev
);
2254 if (dev
->ibd_hdev
&&
2255 dev
->ibd_hdev
->ibh_cmid
) {
2257 * XXX it's not good to close old listener at here,
2258 * because we can fail to create new listener.
2259 * But we have to close it now, otherwise rdma_bind_addr
2260 * will return EADDRINUSE... How crap!
2262 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2264 cmid
= dev
->ibd_hdev
->ibh_cmid
;
2266 * make next schedule of kiblnd_dev_need_failover()
2269 dev
->ibd_hdev
->ibh_cmid
= NULL
;
2270 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2272 rdma_destroy_id(cmid
);
2275 cmid
= kiblnd_rdma_create_id(kiblnd_cm_callback
, dev
, RDMA_PS_TCP
,
2279 CERROR("Failed to create cmid for failover: %d\n", rc
);
2283 memset(&addr
, 0, sizeof(addr
));
2284 addr
.sin_family
= AF_INET
;
2285 addr
.sin_addr
.s_addr
= (__force u32
)htonl(dev
->ibd_ifip
);
2286 addr
.sin_port
= htons(*kiblnd_tunables
.kib_service
);
2288 /* Bind to failover device or port */
2289 rc
= rdma_bind_addr(cmid
, (struct sockaddr
*)&addr
);
2290 if (rc
|| !cmid
->device
) {
2291 CERROR("Failed to bind %s:%pI4h to device(%p): %d\n",
2292 dev
->ibd_ifname
, &dev
->ibd_ifip
,
2294 rdma_destroy_id(cmid
);
2298 LIBCFS_ALLOC(hdev
, sizeof(*hdev
));
2300 CERROR("Failed to allocate kib_hca_dev\n");
2301 rdma_destroy_id(cmid
);
2306 atomic_set(&hdev
->ibh_ref
, 1);
2307 hdev
->ibh_dev
= dev
;
2308 hdev
->ibh_cmid
= cmid
;
2309 hdev
->ibh_ibdev
= cmid
->device
;
2311 pd
= ib_alloc_pd(cmid
->device
);
2314 CERROR("Can't allocate PD: %d\n", rc
);
2320 rc
= rdma_listen(cmid
, 0);
2322 CERROR("Can't start new listener: %d\n", rc
);
2326 rc
= kiblnd_hdev_setup_mrs(hdev
);
2328 CERROR("Can't setup device: %d\n", rc
);
2332 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2334 swap(dev
->ibd_hdev
, hdev
); /* take over the refcount */
2336 list_for_each_entry(net
, &dev
->ibd_nets
, ibn_list
) {
2337 cfs_cpt_for_each(i
, lnet_cpt_table()) {
2338 kiblnd_fail_poolset(&net
->ibn_tx_ps
[i
]->tps_poolset
,
2341 if (net
->ibn_fmr_ps
)
2342 kiblnd_fail_fmr_poolset(net
->ibn_fmr_ps
[i
],
2347 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2349 if (!list_empty(&zombie_tpo
))
2350 kiblnd_destroy_pool_list(&zombie_tpo
);
2351 if (!list_empty(&zombie_ppo
))
2352 kiblnd_destroy_pool_list(&zombie_ppo
);
2353 if (!list_empty(&zombie_fpo
))
2354 kiblnd_destroy_fmr_pool_list(&zombie_fpo
);
2356 kiblnd_hdev_decref(hdev
);
2359 dev
->ibd_failed_failover
++;
2361 dev
->ibd_failed_failover
= 0;
2366 void kiblnd_destroy_dev(kib_dev_t
*dev
)
2368 LASSERT(!dev
->ibd_nnets
);
2369 LASSERT(list_empty(&dev
->ibd_nets
));
2371 list_del(&dev
->ibd_fail_list
);
2372 list_del(&dev
->ibd_list
);
2375 kiblnd_hdev_decref(dev
->ibd_hdev
);
2377 LIBCFS_FREE(dev
, sizeof(*dev
));
2380 static kib_dev_t
*kiblnd_create_dev(char *ifname
)
2382 struct net_device
*netdev
;
2389 rc
= lnet_ipif_query(ifname
, &up
, &ip
, &netmask
);
2391 CERROR("Can't query IPoIB interface %s: %d\n",
2397 CERROR("Can't query IPoIB interface %s: it's down\n", ifname
);
2401 LIBCFS_ALLOC(dev
, sizeof(*dev
));
2405 netdev
= dev_get_by_name(&init_net
, ifname
);
2407 dev
->ibd_can_failover
= 0;
2409 dev
->ibd_can_failover
= !!(netdev
->flags
& IFF_MASTER
);
2413 INIT_LIST_HEAD(&dev
->ibd_nets
);
2414 INIT_LIST_HEAD(&dev
->ibd_list
); /* not yet in kib_devs */
2415 INIT_LIST_HEAD(&dev
->ibd_fail_list
);
2417 strcpy(&dev
->ibd_ifname
[0], ifname
);
2419 /* initialize the device */
2420 rc
= kiblnd_dev_failover(dev
);
2422 CERROR("Can't initialize device: %d\n", rc
);
2423 LIBCFS_FREE(dev
, sizeof(*dev
));
2427 list_add_tail(&dev
->ibd_list
, &kiblnd_data
.kib_devs
);
2431 static void kiblnd_base_shutdown(void)
2433 struct kib_sched_info
*sched
;
2436 LASSERT(list_empty(&kiblnd_data
.kib_devs
));
2438 switch (kiblnd_data
.kib_init
) {
2442 case IBLND_INIT_ALL
:
2443 case IBLND_INIT_DATA
:
2444 LASSERT(kiblnd_data
.kib_peers
);
2445 for (i
= 0; i
< kiblnd_data
.kib_peer_hash_size
; i
++)
2446 LASSERT(list_empty(&kiblnd_data
.kib_peers
[i
]));
2447 LASSERT(list_empty(&kiblnd_data
.kib_connd_zombies
));
2448 LASSERT(list_empty(&kiblnd_data
.kib_connd_conns
));
2450 /* flag threads to terminate; wake and wait for them to die */
2451 kiblnd_data
.kib_shutdown
= 1;
2454 * NB: we really want to stop scheduler threads net by net
2455 * instead of the whole module, this should be improved
2456 * with dynamic configuration LNet
2458 cfs_percpt_for_each(sched
, i
, kiblnd_data
.kib_scheds
)
2459 wake_up_all(&sched
->ibs_waitq
);
2461 wake_up_all(&kiblnd_data
.kib_connd_waitq
);
2462 wake_up_all(&kiblnd_data
.kib_failover_waitq
);
2465 while (atomic_read(&kiblnd_data
.kib_nthreads
)) {
2468 CDEBUG(((i
& (-i
)) == i
) ? D_WARNING
: D_NET
,
2469 "Waiting for %d threads to terminate\n",
2470 atomic_read(&kiblnd_data
.kib_nthreads
));
2471 set_current_state(TASK_UNINTERRUPTIBLE
);
2472 schedule_timeout(cfs_time_seconds(1));
2477 case IBLND_INIT_NOTHING
:
2481 if (kiblnd_data
.kib_peers
) {
2482 LIBCFS_FREE(kiblnd_data
.kib_peers
,
2483 sizeof(struct list_head
) *
2484 kiblnd_data
.kib_peer_hash_size
);
2487 if (kiblnd_data
.kib_scheds
)
2488 cfs_percpt_free(kiblnd_data
.kib_scheds
);
2490 kiblnd_data
.kib_init
= IBLND_INIT_NOTHING
;
2491 module_put(THIS_MODULE
);
2494 void kiblnd_shutdown(lnet_ni_t
*ni
)
2496 kib_net_t
*net
= ni
->ni_data
;
2497 rwlock_t
*g_lock
= &kiblnd_data
.kib_global_lock
;
2499 unsigned long flags
;
2501 LASSERT(kiblnd_data
.kib_init
== IBLND_INIT_ALL
);
2506 write_lock_irqsave(g_lock
, flags
);
2507 net
->ibn_shutdown
= 1;
2508 write_unlock_irqrestore(g_lock
, flags
);
2510 switch (net
->ibn_init
) {
2514 case IBLND_INIT_ALL
:
2515 /* nuke all existing peers within this net */
2516 kiblnd_del_peer(ni
, LNET_NID_ANY
);
2518 /* Wait for all peer state to clean up */
2520 while (atomic_read(&net
->ibn_npeers
)) {
2522 CDEBUG(((i
& (-i
)) == i
) ? D_WARNING
: D_NET
, /* 2**n? */
2523 "%s: waiting for %d peers to disconnect\n",
2524 libcfs_nid2str(ni
->ni_nid
),
2525 atomic_read(&net
->ibn_npeers
));
2526 set_current_state(TASK_UNINTERRUPTIBLE
);
2527 schedule_timeout(cfs_time_seconds(1));
2530 kiblnd_net_fini_pools(net
);
2532 write_lock_irqsave(g_lock
, flags
);
2533 LASSERT(net
->ibn_dev
->ibd_nnets
> 0);
2534 net
->ibn_dev
->ibd_nnets
--;
2535 list_del(&net
->ibn_list
);
2536 write_unlock_irqrestore(g_lock
, flags
);
2540 case IBLND_INIT_NOTHING
:
2541 LASSERT(!atomic_read(&net
->ibn_nconns
));
2543 if (net
->ibn_dev
&& !net
->ibn_dev
->ibd_nnets
)
2544 kiblnd_destroy_dev(net
->ibn_dev
);
2549 net
->ibn_init
= IBLND_INIT_NOTHING
;
2552 LIBCFS_FREE(net
, sizeof(*net
));
2555 if (list_empty(&kiblnd_data
.kib_devs
))
2556 kiblnd_base_shutdown();
2559 static int kiblnd_base_startup(void)
2561 struct kib_sched_info
*sched
;
2565 LASSERT(kiblnd_data
.kib_init
== IBLND_INIT_NOTHING
);
2567 try_module_get(THIS_MODULE
);
2568 /* zero pointers, flags etc */
2569 memset(&kiblnd_data
, 0, sizeof(kiblnd_data
));
2571 rwlock_init(&kiblnd_data
.kib_global_lock
);
2573 INIT_LIST_HEAD(&kiblnd_data
.kib_devs
);
2574 INIT_LIST_HEAD(&kiblnd_data
.kib_failed_devs
);
2576 kiblnd_data
.kib_peer_hash_size
= IBLND_PEER_HASH_SIZE
;
2577 LIBCFS_ALLOC(kiblnd_data
.kib_peers
,
2578 sizeof(struct list_head
) * kiblnd_data
.kib_peer_hash_size
);
2579 if (!kiblnd_data
.kib_peers
)
2581 for (i
= 0; i
< kiblnd_data
.kib_peer_hash_size
; i
++)
2582 INIT_LIST_HEAD(&kiblnd_data
.kib_peers
[i
]);
2584 spin_lock_init(&kiblnd_data
.kib_connd_lock
);
2585 INIT_LIST_HEAD(&kiblnd_data
.kib_connd_conns
);
2586 INIT_LIST_HEAD(&kiblnd_data
.kib_connd_zombies
);
2587 init_waitqueue_head(&kiblnd_data
.kib_connd_waitq
);
2588 init_waitqueue_head(&kiblnd_data
.kib_failover_waitq
);
2590 kiblnd_data
.kib_scheds
= cfs_percpt_alloc(lnet_cpt_table(),
2592 if (!kiblnd_data
.kib_scheds
)
2595 cfs_percpt_for_each(sched
, i
, kiblnd_data
.kib_scheds
) {
2598 spin_lock_init(&sched
->ibs_lock
);
2599 INIT_LIST_HEAD(&sched
->ibs_conns
);
2600 init_waitqueue_head(&sched
->ibs_waitq
);
2602 nthrs
= cfs_cpt_weight(lnet_cpt_table(), i
);
2603 if (*kiblnd_tunables
.kib_nscheds
> 0) {
2604 nthrs
= min(nthrs
, *kiblnd_tunables
.kib_nscheds
);
2607 * max to half of CPUs, another half is reserved for
2608 * upper layer modules
2610 nthrs
= min(max(IBLND_N_SCHED
, nthrs
>> 1), nthrs
);
2613 sched
->ibs_nthreads_max
= nthrs
;
2617 kiblnd_data
.kib_error_qpa
.qp_state
= IB_QPS_ERR
;
2619 /* lists/ptrs/locks initialised */
2620 kiblnd_data
.kib_init
= IBLND_INIT_DATA
;
2621 /*****************************************************/
2623 rc
= kiblnd_thread_start(kiblnd_connd
, NULL
, "kiblnd_connd");
2625 CERROR("Can't spawn o2iblnd connd: %d\n", rc
);
2629 if (*kiblnd_tunables
.kib_dev_failover
)
2630 rc
= kiblnd_thread_start(kiblnd_failover_thread
, NULL
,
2634 CERROR("Can't spawn o2iblnd failover thread: %d\n", rc
);
2638 /* flag everything initialised */
2639 kiblnd_data
.kib_init
= IBLND_INIT_ALL
;
2640 /*****************************************************/
2645 kiblnd_base_shutdown();
2649 static int kiblnd_start_schedulers(struct kib_sched_info
*sched
)
2655 if (!sched
->ibs_nthreads
) {
2656 if (*kiblnd_tunables
.kib_nscheds
> 0) {
2657 nthrs
= sched
->ibs_nthreads_max
;
2659 nthrs
= cfs_cpt_weight(lnet_cpt_table(),
2661 nthrs
= min(max(IBLND_N_SCHED
, nthrs
>> 1), nthrs
);
2662 nthrs
= min(IBLND_N_SCHED_HIGH
, nthrs
);
2665 LASSERT(sched
->ibs_nthreads
<= sched
->ibs_nthreads_max
);
2666 /* increase one thread if there is new interface */
2667 nthrs
= sched
->ibs_nthreads
< sched
->ibs_nthreads_max
;
2670 for (i
= 0; i
< nthrs
; i
++) {
2674 id
= KIB_THREAD_ID(sched
->ibs_cpt
, sched
->ibs_nthreads
+ i
);
2675 snprintf(name
, sizeof(name
), "kiblnd_sd_%02ld_%02ld",
2676 KIB_THREAD_CPT(id
), KIB_THREAD_TID(id
));
2677 rc
= kiblnd_thread_start(kiblnd_scheduler
, (void *)id
, name
);
2681 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2682 sched
->ibs_cpt
, sched
->ibs_nthreads
+ i
, rc
);
2686 sched
->ibs_nthreads
+= i
;
2690 static int kiblnd_dev_start_threads(kib_dev_t
*dev
, int newdev
, __u32
*cpts
,
2697 for (i
= 0; i
< ncpts
; i
++) {
2698 struct kib_sched_info
*sched
;
2700 cpt
= !cpts
? i
: cpts
[i
];
2701 sched
= kiblnd_data
.kib_scheds
[cpt
];
2703 if (!newdev
&& sched
->ibs_nthreads
> 0)
2706 rc
= kiblnd_start_schedulers(kiblnd_data
.kib_scheds
[cpt
]);
2708 CERROR("Failed to start scheduler threads for %s\n",
2716 static kib_dev_t
*kiblnd_dev_search(char *ifname
)
2718 kib_dev_t
*alias
= NULL
;
2723 colon
= strchr(ifname
, ':');
2724 list_for_each_entry(dev
, &kiblnd_data
.kib_devs
, ibd_list
) {
2725 if (!strcmp(&dev
->ibd_ifname
[0], ifname
))
2731 colon2
= strchr(dev
->ibd_ifname
, ':');
2737 if (!strcmp(&dev
->ibd_ifname
[0], ifname
))
2748 int kiblnd_startup(lnet_ni_t
*ni
)
2751 kib_dev_t
*ibdev
= NULL
;
2753 struct timespec64 tv
;
2754 unsigned long flags
;
2758 LASSERT(ni
->ni_lnd
== &the_o2iblnd
);
2760 if (kiblnd_data
.kib_init
== IBLND_INIT_NOTHING
) {
2761 rc
= kiblnd_base_startup();
2766 LIBCFS_ALLOC(net
, sizeof(*net
));
2771 ktime_get_real_ts64(&tv
);
2772 net
->ibn_incarnation
= tv
.tv_sec
* USEC_PER_SEC
+
2773 tv
.tv_nsec
/ NSEC_PER_USEC
;
2775 ni
->ni_peertimeout
= *kiblnd_tunables
.kib_peertimeout
;
2776 ni
->ni_maxtxcredits
= *kiblnd_tunables
.kib_credits
;
2777 ni
->ni_peertxcredits
= *kiblnd_tunables
.kib_peertxcredits
;
2778 ni
->ni_peerrtrcredits
= *kiblnd_tunables
.kib_peerrtrcredits
;
2780 if (ni
->ni_interfaces
[0]) {
2781 /* Use the IPoIB interface specified in 'networks=' */
2783 CLASSERT(LNET_MAX_INTERFACES
> 1);
2784 if (ni
->ni_interfaces
[1]) {
2785 CERROR("Multiple interfaces not supported\n");
2789 ifname
= ni
->ni_interfaces
[0];
2791 ifname
= *kiblnd_tunables
.kib_default_ipif
;
2794 if (strlen(ifname
) >= sizeof(ibdev
->ibd_ifname
)) {
2795 CERROR("IPoIB interface name too long: %s\n", ifname
);
2799 ibdev
= kiblnd_dev_search(ifname
);
2802 /* hmm...create kib_dev even for alias */
2803 if (!ibdev
|| strcmp(&ibdev
->ibd_ifname
[0], ifname
))
2804 ibdev
= kiblnd_create_dev(ifname
);
2809 net
->ibn_dev
= ibdev
;
2810 ni
->ni_nid
= LNET_MKNID(LNET_NIDNET(ni
->ni_nid
), ibdev
->ibd_ifip
);
2812 rc
= kiblnd_dev_start_threads(ibdev
, newdev
,
2813 ni
->ni_cpts
, ni
->ni_ncpts
);
2817 rc
= kiblnd_net_init_pools(net
, ni
->ni_cpts
, ni
->ni_ncpts
);
2819 CERROR("Failed to initialize NI pools: %d\n", rc
);
2823 write_lock_irqsave(&kiblnd_data
.kib_global_lock
, flags
);
2825 list_add_tail(&net
->ibn_list
, &ibdev
->ibd_nets
);
2826 write_unlock_irqrestore(&kiblnd_data
.kib_global_lock
, flags
);
2828 net
->ibn_init
= IBLND_INIT_ALL
;
2833 if (!net
->ibn_dev
&& ibdev
)
2834 kiblnd_destroy_dev(ibdev
);
2837 kiblnd_shutdown(ni
);
2839 CDEBUG(D_NET
, "kiblnd_startup failed\n");
2843 static void __exit
kiblnd_module_fini(void)
2845 lnet_unregister_lnd(&the_o2iblnd
);
2848 static int __init
kiblnd_module_init(void)
2852 CLASSERT(sizeof(kib_msg_t
) <= IBLND_MSG_SIZE
);
2853 CLASSERT(offsetof(kib_msg_t
,
2854 ibm_u
.get
.ibgm_rd
.rd_frags
[IBLND_MAX_RDMA_FRAGS
])
2856 CLASSERT(offsetof(kib_msg_t
,
2857 ibm_u
.putack
.ibpam_rd
.rd_frags
[IBLND_MAX_RDMA_FRAGS
])
2860 rc
= kiblnd_tunables_init();
2864 lnet_register_lnd(&the_o2iblnd
);
2869 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2870 MODULE_DESCRIPTION("Kernel OpenIB gen2 LND v2.00");
2871 MODULE_VERSION("2.7.0");
2872 MODULE_LICENSE("GPL");
2874 module_init(kiblnd_module_init
);
2875 module_exit(kiblnd_module_fini
);