2 * Copyright (C) 2012 Cray, Inc.
4 * Copyright (c) 2013, 2014, Intel Corporation.
6 * Author: Nic Henke <nic@cray.com>
7 * Author: James Shimek <jshimek@cray.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 /* Primary entry points from LNET. There are no guarantees against reentrance. */
34 .lnd_startup = kgnilnd_startup,
35 .lnd_shutdown = kgnilnd_shutdown,
36 .lnd_ctl = kgnilnd_ctl,
37 .lnd_send = kgnilnd_send,
38 .lnd_recv = kgnilnd_recv,
39 .lnd_eager_recv = kgnilnd_eager_recv,
40 .lnd_query = kgnilnd_query,
43 kgn_data_t kgnilnd_data;
46 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
48 struct task_struct *thrd;
50 thrd = kthread_run(fn, arg, "%s_%02d", name, id);
54 atomic_inc(&kgnilnd_data.kgn_nthreads);
58 /* bind scheduler threads to cpus */
60 kgnilnd_start_sd_threads(void)
64 struct task_struct *task;
66 for_each_online_cpu(cpu) {
67 /* don't bind to cpu 0 - all interrupts are processed here */
71 task = kthread_create(kgnilnd_scheduler, (void *)((long)i),
72 "%s_%02d", "kgnilnd_sd", i);
74 kthread_bind(task, cpu);
75 wake_up_process(task);
77 CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i,
81 atomic_inc(&kgnilnd_data.kgn_nthreads);
83 if (++i >= *kgnilnd_tunables.kgn_sched_threads) {
91 /* needs write_lock on kgn_peer_conn_lock */
93 kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
96 struct list_head *ctmp, *cnxt;
100 loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
102 list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
103 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
105 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
111 if (conn->gnc_device != newconn->gnc_device)
114 /* This is a two connection loopback - one talking to the other */
116 newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
117 newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) {
118 CDEBUG(D_NET, "skipping prune of %p, "
119 "loopback and matching stamps"
120 " connstamp "LPU64"("LPU64")"
121 " peerstamp "LPU64"("LPU64")\n",
122 conn, newconn->gnc_my_connstamp,
123 conn->gnc_peer_connstamp,
124 newconn->gnc_peer_connstamp,
125 conn->gnc_my_connstamp);
129 if (conn->gnc_peerstamp != newconn->gnc_peerstamp) {
130 LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp,
131 "conn 0x%p peerstamp "LPU64" >= "
132 "newconn 0x%p peerstamp "LPU64"\n",
133 conn, conn->gnc_peerstamp,
134 newconn, newconn->gnc_peerstamp);
136 CDEBUG(D_NET, "Closing stale conn nid: %s "
137 " peerstamp:"LPX64"("LPX64")\n",
138 libcfs_nid2str(peer->gnp_nid),
139 conn->gnc_peerstamp, newconn->gnc_peerstamp);
142 LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp,
143 "conn 0x%p peer_connstamp "LPU64" >= "
144 "newconn 0x%p peer_connstamp "LPU64"\n",
145 conn, conn->gnc_peer_connstamp,
146 newconn, newconn->gnc_peer_connstamp);
148 CDEBUG(D_NET, "Closing stale conn nid: %s"
149 " connstamp:"LPU64"("LPU64")\n",
150 libcfs_nid2str(peer->gnp_nid),
151 conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
155 kgnilnd_close_conn_locked(conn, -ESTALE);
159 CWARN("Closed %d stale conns to %s\n", count, libcfs_nid2str(peer->gnp_nid));
166 kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
169 struct list_head *tmp;
173 loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
175 list_for_each(tmp, &peer->gnp_conns) {
176 conn = list_entry(tmp, kgn_conn_t, gnc_list);
177 CDEBUG(D_NET, "checking conn 0x%p for peer %s"
178 " lo %d new "LPU64" existing "LPU64
179 " new peer "LPU64" existing peer "LPU64
180 " new dev %p existing dev %p\n",
181 conn, libcfs_nid2str(peer->gnp_nid),
183 newconn->gnc_peerstamp, conn->gnc_peerstamp,
184 newconn->gnc_peer_connstamp, conn->gnc_peer_connstamp,
185 newconn->gnc_device, conn->gnc_device);
187 /* conn is in the process of closing */
188 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
191 /* 'newconn' is from an earlier version of 'peer'!!! */
192 if (newconn->gnc_peerstamp < conn->gnc_peerstamp)
195 /* 'conn' is from an earlier version of 'peer': it will be
196 * removed when we cull stale conns later on... */
197 if (newconn->gnc_peerstamp > conn->gnc_peerstamp)
200 /* Different devices are OK */
201 if (conn->gnc_device != newconn->gnc_device)
204 /* It's me connecting to myself */
206 newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
207 newconn->gnc_peer_connstamp == conn->gnc_my_connstamp)
210 /* 'newconn' is an earlier connection from 'peer'!!! */
211 if (newconn->gnc_peer_connstamp < conn->gnc_peer_connstamp)
214 /* 'conn' is an earlier connection from 'peer': it will be
215 * removed when we cull stale conns later on... */
216 if (newconn->gnc_peer_connstamp > conn->gnc_peer_connstamp)
219 /* 'newconn' has the SAME connection stamp; 'peer' isn't
220 * playing the game... */
228 kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
234 LASSERT (!in_interrupt());
235 atomic_inc(&kgnilnd_data.kgn_nconns);
237 /* divide by 2 to allow for complete reset and immediate reconnect */
238 if (atomic_read(&kgnilnd_data.kgn_nconns) >= GNILND_MAX_CQID/2) {
239 CERROR("Too many conn are live: %d > %d\n",
240 atomic_read(&kgnilnd_data.kgn_nconns), GNILND_MAX_CQID/2);
241 atomic_dec(&kgnilnd_data.kgn_nconns);
245 LIBCFS_ALLOC(conn, sizeof(*conn));
247 atomic_dec(&kgnilnd_data.kgn_nconns);
251 conn->gnc_tx_ref_table =
252 kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *));
253 if (conn->gnc_tx_ref_table == NULL) {
254 CERROR("Can't allocate conn tx_ref_table\n");
255 GOTO(failed, rc = -ENOMEM);
258 mutex_init(&conn->gnc_smsg_mutex);
259 mutex_init(&conn->gnc_rdma_mutex);
260 atomic_set(&conn->gnc_refcount, 1);
261 atomic_set(&conn->gnc_reaper_noop, 0);
262 atomic_set(&conn->gnc_sched_noop, 0);
263 atomic_set(&conn->gnc_tx_in_use, 0);
264 INIT_LIST_HEAD(&conn->gnc_list);
265 INIT_LIST_HEAD(&conn->gnc_hashlist);
266 INIT_LIST_HEAD(&conn->gnc_schedlist);
267 INIT_LIST_HEAD(&conn->gnc_fmaq);
268 INIT_LIST_HEAD(&conn->gnc_mdd_list);
269 spin_lock_init(&conn->gnc_list_lock);
270 spin_lock_init(&conn->gnc_tx_lock);
271 conn->gnc_magic = GNILND_CONN_MAGIC;
273 /* set tx id to nearly the end to make sure we find wrapping
275 conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10;
277 /* if this fails, we have conflicts and MAX_TX is too large */
278 CLASSERT(GNILND_MAX_MSG_ID < GNILND_MSGID_CLOSE);
280 /* get a new unique CQ id for this conn */
281 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
282 conn->gnc_my_connstamp = kgnilnd_data.kgn_connstamp++;
283 conn->gnc_cqid = kgnilnd_get_cqid_locked();
284 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
286 if (conn->gnc_cqid == 0) {
287 CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
288 GOTO(failed, rc = -E2BIG);
291 CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
292 conn->gnc_cqid, conn);
294 /* need to be set before gnc_ephandle to allow kgnilnd_destroy_conn_ep to
296 conn->gnc_device = dev;
298 conn->gnc_timeout = MAX(*kgnilnd_tunables.kgn_timeout,
300 kgnilnd_update_reaper_timeout(conn->gnc_timeout);
302 /* this is the ep_handle for doing SMSG & BTE */
303 mutex_lock(&dev->gnd_cq_mutex);
304 rrc = kgnilnd_ep_create(dev->gnd_handle, dev->gnd_snd_fma_cqh,
305 &conn->gnc_ephandle);
306 mutex_unlock(&dev->gnd_cq_mutex);
307 if (rrc != GNI_RC_SUCCESS)
308 GOTO(failed, rc = -ENETDOWN);
310 CDEBUG(D_NET, "created conn 0x%p ep_hndl 0x%p\n",
311 conn, conn->gnc_ephandle);
313 /* add ref for EP canceling */
314 kgnilnd_conn_addref(conn);
315 atomic_inc(&dev->gnd_neps);
321 atomic_dec(&kgnilnd_data.kgn_nconns);
322 LIBCFS_FREE(conn->gnc_tx_ref_table, GNILND_MAX_MSG_ID * sizeof(void *));
323 LIBCFS_FREE(conn, sizeof(*conn));
327 /* needs to be called with kgn_peer_conn_lock held (read or write) */
329 kgnilnd_find_conn_locked(kgn_peer_t *peer)
331 kgn_conn_t *conn = NULL;
333 /* if we are in reset, this conn is going to die soon */
334 if (unlikely(kgnilnd_data.kgn_in_reset)) {
338 /* just return the first ESTABLISHED connection */
339 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
340 /* kgnilnd_finish_connect doesn't put connections on the
341 * peer list until they are actually established */
342 LASSERTF(conn->gnc_state >= GNILND_CONN_ESTABLISHED,
343 "found conn %p state %s on peer %p (%s)\n",
344 conn, kgnilnd_conn_state2str(conn), peer,
345 libcfs_nid2str(peer->gnp_nid));
346 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
354 /* needs write_lock on kgn_peer_conn_lock held */
356 kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer) {
358 kgn_device_t *dev = peer->gnp_net->gnn_dev;
361 conn = kgnilnd_find_conn_locked(peer);
367 /* if the peer was previously connecting, check if we should
368 * trigger another connection attempt yet. */
369 if (time_before(jiffies, peer->gnp_reconnect_time)) {
373 /* This check prevents us from creating a new connection to a peer while we are
374 * still in the process of closing an existing connection to the peer.
376 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
377 if (conn->gnc_ephandle != NULL) {
378 CDEBUG(D_NET, "Not connecting non-null ephandle found peer 0x%p->%s\n", peer,
379 libcfs_nid2str(peer->gnp_nid));
384 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
385 /* if we are not connecting, fire up a new connection */
386 /* or if we are anything but IDLE DONT start a new connection */
390 CDEBUG(D_NET, "starting connect to %s\n",
391 libcfs_nid2str(peer->gnp_nid));
392 peer->gnp_connecting = GNILND_PEER_CONNECT;
393 kgnilnd_peer_addref(peer); /* extra ref for connd */
395 spin_lock(&dev->gnd_connd_lock);
396 list_add_tail(&peer->gnp_connd_list, &dev->gnd_connd_peers);
397 spin_unlock(&dev->gnd_connd_lock);
399 kgnilnd_schedule_dgram(dev);
400 CDEBUG(D_NETTRACE, "scheduling new connect\n");
405 /* Caller is responsible for deciding if/when to call this */
407 kgnilnd_destroy_conn_ep(kgn_conn_t *conn)
410 gni_ep_handle_t tmp_ep;
412 /* only if we actually initialized it,
413 * then set NULL to tell kgnilnd_destroy_conn to leave it alone */
415 tmp_ep = xchg(&conn->gnc_ephandle, NULL);
416 if (tmp_ep != NULL) {
417 /* we never re-use the EP, so unbind is not needed */
418 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
419 rrc = kgnilnd_ep_destroy(tmp_ep);
421 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
423 /* if this fails, it could hork up kgni smsg retransmit and others
424 * since we could free the SMSG mbox memory, etc. */
425 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d conn 0x%p ep 0x%p\n",
426 rrc, conn, conn->gnc_ephandle);
428 atomic_dec(&conn->gnc_device->gnd_neps);
430 /* clear out count added in kgnilnd_close_conn_locked
431 * conn will have a peer once it hits finish_connect, where it
432 * is the first spot we'll mark it ESTABLISHED as well */
433 if (conn->gnc_peer) {
434 kgnilnd_admin_decref(conn->gnc_peer->gnp_dirty_eps);
437 /* drop ref for EP */
438 kgnilnd_conn_decref(conn);
443 kgnilnd_destroy_conn(kgn_conn_t *conn)
445 LASSERTF(!in_interrupt() &&
446 !conn->gnc_scheduled &&
447 !conn->gnc_in_purgatory &&
448 conn->gnc_ephandle == NULL &&
449 list_empty(&conn->gnc_list) &&
450 list_empty(&conn->gnc_hashlist) &&
451 list_empty(&conn->gnc_schedlist) &&
452 list_empty(&conn->gnc_mdd_list) &&
453 conn->gnc_magic == GNILND_CONN_MAGIC,
454 "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d\n",
455 conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
457 !!in_interrupt(), conn->gnc_scheduled,
458 conn->gnc_in_purgatory,
461 list_empty(&conn->gnc_list),
462 list_empty(&conn->gnc_hashlist),
463 list_empty(&conn->gnc_schedlist),
464 list_empty(&conn->gnc_mdd_list));
466 /* Tripping these is especially bad, as it means we have items on the
467 * lists that didn't keep their refcount on the connection - or
468 * somebody evil released their own */
469 LASSERTF(list_empty(&conn->gnc_fmaq) &&
470 atomic_read(&conn->gnc_nlive_fma) == 0 &&
471 atomic_read(&conn->gnc_nlive_rdma) == 0,
472 "conn 0x%p fmaq %d@0x%p nfma %d nrdma %d\n",
473 conn, kgnilnd_count_list(&conn->gnc_fmaq), &conn->gnc_fmaq,
474 atomic_read(&conn->gnc_nlive_fma), atomic_read(&conn->gnc_nlive_rdma));
476 CDEBUG(D_NET, "destroying conn %p ephandle %p error %d\n",
477 conn, conn->gnc_ephandle, conn->gnc_error);
479 /* We are freeing this memory remove the magic value from the connection */
482 /* if there is an FMA blk left here, we'll tear it down */
483 if (conn->gnc_fma_blk) {
484 if (conn->gnc_peer) {
485 kgn_mbox_info_t *mbox;
486 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
487 mbox->mbx_prev_nid = conn->gnc_peer->gnp_nid;
489 kgnilnd_release_mbox(conn, 0);
492 if (conn->gnc_peer != NULL)
493 kgnilnd_peer_decref(conn->gnc_peer);
495 if (conn->gnc_tx_ref_table != NULL) {
496 LIBCFS_FREE(conn->gnc_tx_ref_table,
497 GNILND_MAX_MSG_ID * sizeof(void *));
500 LIBCFS_FREE(conn, sizeof(*conn));
501 atomic_dec(&kgnilnd_data.kgn_nconns);
504 /* peer_alive and peer_notify done in the style of the o2iblnd */
506 kgnilnd_peer_alive(kgn_peer_t *peer)
508 set_mb(peer->gnp_last_alive, jiffies);
512 kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive)
523 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DONT_NOTIFY))
526 /* Tell LNet we are giving ups on this peer - but only
527 * if it isn't already reconnected or trying to reconnect */
528 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
530 /* use kgnilnd_find_conn_locked to avoid any conns in the process of being nuked
532 * don't tell LNet if we are in reset - we assume that everyone will be able to
533 * reconnect just fine
535 conn = kgnilnd_find_conn_locked(peer);
537 CDEBUG(D_NETTRACE, "peer 0x%p->%s ting %d conn 0x%p, rst %d error %d\n",
538 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn,
539 kgnilnd_data.kgn_in_reset, error);
541 if (((peer->gnp_connecting == GNILND_PEER_IDLE) &&
543 (!kgnilnd_data.kgn_in_reset) &&
544 (!kgnilnd_conn_clean_errno(error))) || alive) {
548 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
551 /* short circuit if we dont need to notify Lnet */
555 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
558 /* dont do this if this fails since LNET is in shutdown or something else
561 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
562 list_for_each_entry(net , &kgnilnd_data.kgn_nets[i], gnn_list) {
563 /* if gnn_shutdown set for any net shutdown is in progress just return */
564 if (net->gnn_shutdown) {
565 up_read(&kgnilnd_data.kgn_net_rw_sem);
573 /* shutdown in progress most likely */
574 up_read(&kgnilnd_data.kgn_net_rw_sem);
578 LIBCFS_ALLOC(nets, nnets * sizeof(*nets));
581 up_read(&kgnilnd_data.kgn_net_rw_sem);
582 CERROR("Failed to allocate nets[%d]\n", nnets);
587 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
588 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
590 kgnilnd_net_addref(net);
594 up_read(&kgnilnd_data.kgn_net_rw_sem);
596 for (i = 0; i < nnets; i++) {
601 peer_nid = kgnilnd_lnd2lnetnid(net->gnn_ni->ni_nid,
604 CDEBUG(D_NET, "peer 0x%p->%s last_alive %lu (%lus ago)\n",
605 peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive,
606 cfs_duration_sec(jiffies - peer->gnp_last_alive));
608 lnet_notify(net->gnn_ni, peer_nid, alive,
609 peer->gnp_last_alive);
611 kgnilnd_net_decref(net);
614 LIBCFS_FREE(nets, nnets * sizeof(*nets));
618 /* need write_lock on kgn_peer_conn_lock */
620 kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
622 kgn_peer_t *peer = conn->gnc_peer;
625 LASSERT(!in_interrupt());
627 /* store error for tx completion */
628 conn->gnc_error = error;
629 peer->gnp_last_errno = error;
631 /* use real error from peer if possible */
632 if (error == -ECONNRESET) {
633 error = conn->gnc_peer_error;
636 /* if we NETERROR, make sure it is rate limited */
637 if (!kgnilnd_conn_clean_errno(error) &&
638 peer->gnp_down == GNILND_RCA_NODE_UP) {
639 CNETERR("closing conn to %s: error %d\n",
640 libcfs_nid2str(peer->gnp_nid), error);
642 CDEBUG(D_NET, "closing conn to %s: error %d\n",
643 libcfs_nid2str(peer->gnp_nid), error);
646 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
647 "conn %p to %s with bogus state %s\n", conn,
648 libcfs_nid2str(conn->gnc_peer->gnp_nid),
649 kgnilnd_conn_state2str(conn));
650 LASSERT(!list_empty(&conn->gnc_hashlist));
651 LASSERT(!list_empty(&conn->gnc_list));
654 /* mark peer count here so any place the EP gets destroyed will
655 * open up the peer count so that a new ESTABLISHED conn is then free
656 * to send new messages -- sending before the previous EPs are destroyed
657 * could end up with messages on the network for the old conn _after_
658 * the new conn and break the mbox safety protocol */
659 kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
661 /* Remove from conn hash table: no new callbacks */
662 list_del_init(&conn->gnc_hashlist);
663 kgnilnd_data.kgn_conn_version++;
664 kgnilnd_conn_decref(conn);
666 /* if we are in reset, go right to CLOSED as there is no scheduler
667 * thread to move from CLOSING to CLOSED */
668 if (unlikely(kgnilnd_data.kgn_in_reset)) {
669 conn->gnc_state = GNILND_CONN_CLOSED;
671 conn->gnc_state = GNILND_CONN_CLOSING;
674 /* leave on peer->gnp_conns to make sure we don't let the reaper
675 * or others try to unlink this peer until the conn is fully
676 * processed for closing */
678 if (kgnilnd_check_purgatory_conn(conn)) {
679 kgnilnd_add_purgatory_locked(conn, conn->gnc_peer);
682 /* Reset RX timeout to ensure we wait for an incoming CLOSE
683 * for the full timeout. If we get a CLOSE we know the
684 * peer has stopped all RDMA. Otherwise if we wait for
685 * the full timeout we can also be sure all RDMA has stopped. */
686 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
689 /* schedule sending CLOSE - if we are in quiesce, this adds to
690 * gnd_ready_conns and allows us to find it in quiesce processing */
691 kgnilnd_schedule_conn(conn);
697 kgnilnd_close_conn(kgn_conn_t *conn, int error)
699 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
700 /* need to check the state here - this call is racy and we don't
701 * know the state until after the lock is grabbed */
702 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
703 kgnilnd_close_conn_locked(conn, error);
705 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
709 kgnilnd_complete_closed_conn(kgn_conn_t *conn)
719 /* Dump log on cksum error - wait until complete phase to let
720 * RX of error happen */
721 if (*kgnilnd_tunables.kgn_checksum_dump &&
722 (conn != NULL && conn->gnc_peer_error == -ENOKEY)) {
723 libcfs_debug_dumplog();
726 /* _CLOSED set in kgnilnd_process_fmaq once we decide to
727 * send the CLOSE or not */
728 LASSERTF(conn->gnc_state == GNILND_CONN_CLOSED,
729 "conn 0x%p->%s with bad state %s\n",
730 conn, conn->gnc_peer ?
731 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
733 kgnilnd_conn_state2str(conn));
735 LASSERT(list_empty(&conn->gnc_hashlist));
737 /* we've sent the close, start nuking */
738 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
739 kgnilnd_schedule_conn(conn);
741 if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
742 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
743 "done, Attempting to recover conn 0x%p "
744 "scheduled %d function: %s line: %d\n", conn,
745 conn->gnc_scheduled, conn->gnc_sched_caller,
746 conn->gnc_sched_line);
750 /* we don't use lists to track things that we can get out of the
753 /* need to hold locks for tx_list_state, sampling it is too racy:
754 * - the lock actually protects tx != NULL, but we can't take the proper
755 * lock until we check tx_list_state, which would be too late and
756 * we could have the TX change under us.
757 * gnd_rdmaq_lock and gnd_lock and not used together, so taking both
759 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
760 spin_lock(&conn->gnc_device->gnd_lock);
762 for (nrdma = 0; nrdma < GNILND_MAX_MSG_ID; nrdma++) {
763 tx = conn->gnc_tx_ref_table[nrdma];
766 /* only print the first error and if not CLOSE, we often don't see
767 * CQ events for that by the time we get here... and really don't care */
768 if (nlive || tx->tx_msg.gnm_type == GNILND_MSG_CLOSE)
769 tx->tx_state |= GNILND_TX_QUIET_ERROR;
771 GNIDBG_TX(D_NET, tx, "cleaning up on close, nlive %d", nlive);
773 /* don't worry about gnc_lock here as nobody else should be
774 * touching this conn */
775 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
776 list_add_tail(&tx->tx_list, &sinners);
779 spin_unlock(&conn->gnc_device->gnd_lock);
780 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
782 /* nobody should have marked this as needing scheduling after
783 * we called close - so only ref should be us handling it */
784 if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
785 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
786 "done, Attempting to recover conn 0x%p "
787 "scheduled %d function %s line: %d\n", conn,
788 conn->gnc_scheduled, conn->gnc_sched_caller,
789 conn->gnc_sched_line);
791 /* now reset a few to actual counters... */
792 nrdma = atomic_read(&conn->gnc_nlive_rdma);
793 nq_rdma = atomic_read(&conn->gnc_nq_rdma);
795 if (!list_empty(&sinners)) {
796 list_for_each_entry_safe(tx, txn, &sinners, tx_list) {
797 /* clear tx_list to make tx_add_list_locked happy */
798 list_del_init(&tx->tx_list);
799 /* The error codes determine if we hold onto the MDD */
800 kgnilnd_tx_done(tx, conn->gnc_error);
804 logmsg = (nlive + nrdma + nq_rdma);
807 if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_UP) {
808 CNETERR("Closed conn 0x%p->%s (errno %d, peer errno %d): "
809 "canceled %d TX, %d/%d RDMA\n",
810 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
811 conn->gnc_error, conn->gnc_peer_error,
812 nlive, nq_rdma, nrdma);
814 CDEBUG(D_NET, "Closed conn 0x%p->%s (errno %d,"
815 " peer errno %d): canceled %d TX, %d/%d RDMA\n",
816 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
817 conn->gnc_error, conn->gnc_peer_error,
818 nlive, nq_rdma, nrdma);
822 kgnilnd_destroy_conn_ep(conn);
824 /* Bug 765042 - race this with completing a new conn to same peer - we need
825 * finish_connect to detach purgatory before we can do it ourselves here */
826 CFS_RACE(CFS_FAIL_GNI_FINISH_PURG);
828 /* now it is safe to remove from peer list - anyone looking at
829 * gnp_conns now is free to unlink if not on purgatory */
830 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
832 conn->gnc_state = GNILND_CONN_DONE;
834 /* Decrement counter if we are marked by del_conn_or_peers for closing
836 if (conn->gnc_needs_closing)
837 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_conns);
839 /* Remove from peer's list of valid connections if its not in purgatory */
840 if (!conn->gnc_in_purgatory) {
841 list_del_init(&conn->gnc_list);
842 /* Lose peers reference on the conn */
843 kgnilnd_conn_decref(conn);
846 /* NB - only unlinking if we set pending in del_peer_locked from admin or
848 if (kgnilnd_peer_active(conn->gnc_peer) &&
849 conn->gnc_peer->gnp_pending_unlink &&
850 kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
851 kgnilnd_unlink_peer_locked(conn->gnc_peer);
854 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
856 /* I'm telling Mommy! - use peer_error if they initiated close */
857 kgnilnd_peer_notify(conn->gnc_peer,
858 conn->gnc_error == -ECONNRESET ?
859 conn->gnc_peer_error : conn->gnc_error, 0);
865 kgnilnd_set_conn_params(kgn_dgram_t *dgram)
867 kgn_conn_t *conn = dgram->gndg_conn;
868 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
869 kgn_gniparams_t *rem_param = &connreq->gncr_gnparams;
872 gni_smsg_attr_t *remote = &connreq->gncr_gnparams.gnpr_smsg_attr;
874 /* set timeout vals in conn early so we can use them for the NAK */
876 /* use max of the requested and our timeout, peer will do the same */
877 conn->gnc_timeout = MAX(conn->gnc_timeout, connreq->gncr_timeout);
879 /* only ep_bind really mucks around with the CQ */
880 /* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check
881 * is necessary as you can only bind an ep once and we must make sure we dont bind when already bound.
883 if (connreq->gncr_dstnid != LNET_NID_ANY && dgram->gndg_conn_out.gncr_dstnid != connreq->gncr_srcnid) {
884 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
885 rrc = kgnilnd_ep_bind(conn->gnc_ephandle,
886 connreq->gncr_gnparams.gnpr_host_id,
888 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
889 if (rrc != GNI_RC_SUCCESS) {
895 rrc = kgnilnd_ep_set_eventdata(conn->gnc_ephandle, conn->gnc_cqid,
896 connreq->gncr_gnparams.gnpr_cqid);
897 if (rrc != GNI_RC_SUCCESS) {
902 /* Initialize SMSG */
903 rrc = kgnilnd_smsg_init(conn->gnc_ephandle, &conn->gnpr_smsg_attr,
904 &connreq->gncr_gnparams.gnpr_smsg_attr);
905 if (unlikely(rrc == GNI_RC_INVALID_PARAM)) {
906 gni_smsg_attr_t *local = &conn->gnpr_smsg_attr;
907 /* help folks figure out if there is a tunable off, etc. */
908 LCONSOLE_ERROR("SMSG attribute mismatch. Data from local/remote:"
909 " type %d/%d msg_maxsize %u/%u"
910 " mbox_maxcredit %u/%u. Please check kgni"
911 " logs for further data\n",
912 local->msg_type, remote->msg_type,
913 local->msg_maxsize, remote->msg_maxsize,
914 local->mbox_maxcredit, remote->mbox_maxcredit);
916 if (rrc != GNI_RC_SUCCESS) {
921 /* log this for help in debuggin SMSG buffer re-use */
922 CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired"
923 " local cqid %u SMSG %p->%u hndl "LPX64"."LPX64
924 " remote cqid %u SMSG %p->%u hndl "LPX64"."LPX64"\n",
925 conn, libcfs_nid2str(connreq->gncr_srcnid),
926 libcfs_nid2str(connreq->gncr_dstnid),
927 &conn->gnpr_smsg_attr,
929 conn->gnpr_smsg_attr.msg_buffer,
930 conn->gnpr_smsg_attr.mbox_offset,
931 conn->gnpr_smsg_attr.mem_hndl.qword1,
932 conn->gnpr_smsg_attr.mem_hndl.qword2,
933 rem_param->gnpr_cqid,
934 rem_param->gnpr_smsg_attr.msg_buffer,
935 rem_param->gnpr_smsg_attr.mbox_offset,
936 rem_param->gnpr_smsg_attr.mem_hndl.qword1,
937 rem_param->gnpr_smsg_attr.mem_hndl.qword2);
939 conn->gnc_peerstamp = connreq->gncr_peerstamp;
940 conn->gnc_peer_connstamp = connreq->gncr_connstamp;
941 conn->remote_mbox_addr = (void *)((char *)remote->msg_buffer + remote->mbox_offset);
943 /* We update the reaper timeout once we have a valid conn and timeout */
944 kgnilnd_update_reaper_timeout(GNILND_TO2KA(conn->gnc_timeout));
949 rrc = kgnilnd_ep_unbind(conn->gnc_ephandle);
950 /* not sure I can just let this fly */
951 LASSERTF(rrc == GNI_RC_SUCCESS,
952 "bad rc from gni_ep_unbind trying to cleanup: %d\n", rrc);
955 LASSERTF(rc != 0, "SOFTWARE BUG: rc == 0\n");
956 CERROR("Error setting connection params from %s: %d\n",
957 libcfs_nid2str(connreq->gncr_srcnid), rc);
961 /* needs down_read on kgn_net_rw_sem held from before this call until
962 * after the write_lock on kgn_peer_conn_lock - this ensures we stay sane
963 * with kgnilnd_shutdown - it'll get the sem and set shutdown, then get the
964 * kgn_peer_conn_lock to start del_peer'ing. If we hold the sem until after
965 * kgn_peer_conn_lock is held, we guarantee that nobody calls
966 * kgnilnd_add_peer_locked without checking gnn_shutdown */
968 kgnilnd_create_peer_safe(kgn_peer_t **peerp,
976 LASSERT(nid != LNET_NID_ANY);
978 /* We dont pass the net around in the dgram anymore so here is where we find it
979 * this will work unless its in shutdown or the nid has a net that is invalid.
980 * Either way error code needs to be returned in that case.
982 * If the net passed in is not NULL then we can use it, this alleviates looking it
983 * when the calling function has access to the data.
986 rc = kgnilnd_find_net(nid, &net);
990 /* find net adds a reference on the net if we are not using
991 * it we must do it manually so the net references are
992 * correct when tearing down the net
994 kgnilnd_net_addref(net);
997 LIBCFS_ALLOC(peer, sizeof(*peer));
999 kgnilnd_net_decref(net);
1002 peer->gnp_nid = nid;
1003 peer->gnp_down = node_state;
1005 /* translate from nid to nic addr & store */
1006 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
1008 kgnilnd_net_decref(net);
1009 LIBCFS_FREE(peer, sizeof(*peer));
1012 CDEBUG(D_NET, "peer 0x%p->%s -> NIC 0x%x\n", peer,
1013 libcfs_nid2str(nid), peer->gnp_host_id);
1015 atomic_set(&peer->gnp_refcount, 1); /* 1 ref for caller */
1016 atomic_set(&peer->gnp_dirty_eps, 0);
1018 INIT_LIST_HEAD(&peer->gnp_list);
1019 INIT_LIST_HEAD(&peer->gnp_connd_list);
1020 INIT_LIST_HEAD(&peer->gnp_conns);
1021 INIT_LIST_HEAD(&peer->gnp_tx_queue);
1023 /* the first reconnect should happen immediately, so we leave
1024 * gnp_reconnect_interval set to 0 */
1026 LASSERTF(net != NULL, "peer 0x%p->%s with NULL net\n",
1027 peer, libcfs_nid2str(nid));
1029 /* must have kgn_net_rw_sem held for this... */
1030 if (net->gnn_shutdown) {
1031 /* shutdown has started already */
1032 kgnilnd_net_decref(net);
1033 LIBCFS_FREE(peer, sizeof(*peer));
1037 peer->gnp_net = net;
1039 atomic_inc(&kgnilnd_data.kgn_npeers);
1046 kgnilnd_destroy_peer(kgn_peer_t *peer)
1048 CDEBUG(D_NET, "peer %s %p deleted\n",
1049 libcfs_nid2str(peer->gnp_nid), peer);
1050 LASSERTF(atomic_read(&peer->gnp_refcount) == 0,
1051 "peer 0x%p->%s refs %d\n",
1052 peer, libcfs_nid2str(peer->gnp_nid),
1053 atomic_read(&peer->gnp_refcount));
1054 LASSERTF(atomic_read(&peer->gnp_dirty_eps) == 0,
1055 "peer 0x%p->%s dirty eps %d\n",
1056 peer, libcfs_nid2str(peer->gnp_nid),
1057 atomic_read(&peer->gnp_dirty_eps));
1058 LASSERTF(peer->gnp_net != NULL, "peer %p (%s) with NULL net\n",
1059 peer, libcfs_nid2str(peer->gnp_nid));
1060 LASSERTF(!kgnilnd_peer_active(peer),
1062 peer, libcfs_nid2str(peer->gnp_nid));
1063 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE || peer->gnp_connecting == GNILND_PEER_KILL,
1064 "peer 0x%p->%s, connecting %d\n",
1065 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1066 LASSERTF(list_empty(&peer->gnp_conns),
1068 peer, libcfs_nid2str(peer->gnp_nid));
1069 LASSERTF(list_empty(&peer->gnp_tx_queue),
1071 peer, libcfs_nid2str(peer->gnp_nid));
1072 LASSERTF(list_empty(&peer->gnp_connd_list),
1074 peer, libcfs_nid2str(peer->gnp_nid));
1076 /* NB a peer's connections keep a reference on their peer until
1077 * they are destroyed, so we can be assured that _all_ state to do
1078 * with this peer has been cleaned up when its refcount drops to
1081 atomic_dec(&kgnilnd_data.kgn_npeers);
1082 kgnilnd_net_decref(peer->gnp_net);
1084 LIBCFS_FREE(peer, sizeof(*peer));
1087 /* the conn might not have made it all the way through to a connected
1088 * state - but we need to purgatory any conn that a remote peer might
1089 * have seen through a posted dgram as well */
1091 kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer)
1093 kgn_mbox_info_t *mbox = NULL;
1096 /* NB - the caller should own conn by removing him from the
1097 * scheduler thread when finishing the close */
1099 LASSERTF(peer != NULL, "conn %p with NULL peer\n", conn);
1101 /* If this is still true, need to add the calls to unlink back in and
1102 * figure out how to close the hole on loopback conns */
1103 LASSERTF(kgnilnd_peer_active(peer), "can't use inactive peer %s (%p)"
1104 " we'll never recover the resources\n",
1105 libcfs_nid2str(peer->gnp_nid), peer);
1107 CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
1110 LASSERTF(conn->gnc_in_purgatory == 0,
1111 "Conn already in purgatory\n");
1112 conn->gnc_in_purgatory = 1;
1114 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1115 mbox->mbx_prev_purg_nid = peer->gnp_nid;
1116 mbox->mbx_add_purgatory = jiffies;
1117 kgnilnd_release_mbox(conn, 1);
1119 LASSERTF(list_empty(&conn->gnc_mdd_list),
1120 "conn 0x%p->%s with active purgatory hold MDD %d\n",
1121 conn, libcfs_nid2str(peer->gnp_nid),
1122 kgnilnd_count_list(&conn->gnc_mdd_list));
1127 /* Instead of detaching everything from purgatory here we just mark the conn as needing
1128 * detach, when the reaper checks the conn the next time it will detach it.
1129 * Calling function requires write_lock held on kgn_peer_conn_lock
1132 kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer) {
1135 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
1136 if (conn->gnc_in_purgatory && !conn->gnc_needs_detach) {
1137 conn->gnc_needs_detach = 1;
1138 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_detach);
1143 /* Calling function needs a write_lock held on kgn_peer_conn_lock */
1145 kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list)
1147 kgn_mbox_info_t *mbox = NULL;
1149 /* if needed, add the conn purgatory data to the list passed in */
1150 if (conn->gnc_in_purgatory) {
1151 CDEBUG(D_NET, "peer %p->%s purg_conn %p@%s mdd_list #tx %d\n",
1152 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1153 conn, kgnilnd_conn_state2str(conn),
1154 kgnilnd_count_list(&conn->gnc_mdd_list));
1156 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1157 mbox->mbx_detach_of_purgatory = jiffies;
1159 /* conn->gnc_list is the entry point on peer->gnp_conns, so detaching it
1160 * here removes it from the list of 'valid' peer connections.
1161 * We put the current conn onto a list of conns to call kgnilnd_release_purgatory_locked()
1162 * and as such the caller of kgnilnd_detach_purgatory_locked() now owns that conn, since its not
1163 * on the peer's conn_list anymore.
1166 list_del_init(&conn->gnc_list);
1168 /* NB - only unlinking if we set pending in del_peer_locked from admin or
1170 if (kgnilnd_peer_active(conn->gnc_peer) &&
1171 conn->gnc_peer->gnp_pending_unlink &&
1172 kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
1173 kgnilnd_unlink_peer_locked(conn->gnc_peer);
1175 /* The reaper will not call detach unless the conn is fully through kgnilnd_complete_closed_conn.
1176 * If the conn is not in a DONE state somehow we are attempting to detach even though
1177 * the conn has not been fully cleaned up. If we detach while the conn is still closing
1178 * we will end up with an orphaned connection that has valid ep_handle, that is not on a
1182 LASSERTF(conn->gnc_state == GNILND_CONN_DONE, "Conn in invalid state %p@%s \n",
1183 conn, kgnilnd_conn_state2str(conn));
1185 /* move from peer to the delayed release list */
1186 list_add_tail(&conn->gnc_list, conn_list);
1191 kgnilnd_release_purgatory_list(struct list_head *conn_list)
1194 kgn_conn_t *conn, *connN;
1195 kgn_mdd_purgatory_t *gmp, *gmpN;
1197 list_for_each_entry_safe(conn, connN, conn_list, gnc_list) {
1198 dev = conn->gnc_device;
1200 kgnilnd_release_mbox(conn, -1);
1201 conn->gnc_in_purgatory = 0;
1203 list_del_init(&conn->gnc_list);
1205 /* gnc_needs_detach is set in kgnilnd_del_conn_or_peer. It is used to keep track
1206 * of conns that have been marked for detach by kgnilnd_del_conn_or_peer.
1207 * The function uses kgn_npending_detach to verify the conn has
1208 * actually been detached.
1211 if (conn->gnc_needs_detach)
1212 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_detach);
1214 /* if this guy is really dead (we are doing release from reaper),
1215 * make sure we tell LNet - if this is from other context,
1216 * the checks in the function will prevent an errant
1218 kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0);
1220 list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
1223 "dev %p releasing held mdd "LPX64"."LPX64"\n",
1224 conn->gnc_device, gmp->gmp_map_key.qword1,
1225 gmp->gmp_map_key.qword2);
1227 atomic_dec(&dev->gnd_n_mdd_held);
1228 kgnilnd_mem_mdd_release(conn->gnc_device->gnd_handle,
1230 /* ignoring the return code - if kgni/ghal can't find it
1231 * it must be released already */
1233 list_del_init(&gmp->gmp_list);
1234 LIBCFS_FREE(gmp, sizeof(*gmp));
1236 /* lose conn ref for purgatory */
1237 kgnilnd_conn_decref(conn);
1241 /* needs write_lock on kgnilnd_data.kgn_peer_conn_lock held */
1243 kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer)
1247 current_to = peer->gnp_reconnect_interval;
1249 /* we'll try to reconnect fast the first time, then back-off */
1250 if (current_to == 0) {
1251 peer->gnp_reconnect_time = jiffies - 1;
1252 current_to = *kgnilnd_tunables.kgn_min_reconnect_interval;
1254 peer->gnp_reconnect_time = jiffies + cfs_time_seconds(current_to);
1255 /* add 50% of min timeout & retry */
1256 current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2;
1259 current_to = MIN(current_to,
1260 *kgnilnd_tunables.kgn_max_reconnect_interval);
1262 peer->gnp_reconnect_interval = current_to;
1263 CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n",
1264 libcfs_nid2str(peer->gnp_nid), peer->gnp_reconnect_time,
1265 peer->gnp_reconnect_interval);
1268 /* needs kgnilnd_data.kgn_peer_conn_lock held */
1270 kgnilnd_find_peer_locked(lnet_nid_t nid)
1272 struct list_head *peer_list = kgnilnd_nid2peerlist(nid);
1275 /* Chopping nid down to only NIDADDR using LNET_NIDADDR so we only
1276 * have a single peer per device instead of a peer per nid/net combo.
1279 list_for_each_entry(peer, peer_list, gnp_list) {
1280 if (LNET_NIDADDR(nid) != LNET_NIDADDR(peer->gnp_nid))
1283 CDEBUG(D_NET, "got peer [%p] -> %s c %d (%d)\n",
1284 peer, libcfs_nid2str(nid),
1285 peer->gnp_connecting,
1286 atomic_read(&peer->gnp_refcount));
1292 /* need write_lock on kgn_peer_conn_lock */
1294 kgnilnd_unlink_peer_locked(kgn_peer_t *peer)
1296 LASSERTF(list_empty(&peer->gnp_conns),
1298 peer, libcfs_nid2str(peer->gnp_nid));
1299 LASSERTF(list_empty(&peer->gnp_tx_queue),
1301 peer, libcfs_nid2str(peer->gnp_nid));
1302 LASSERTF(kgnilnd_peer_active(peer),
1304 peer, libcfs_nid2str(peer->gnp_nid));
1305 CDEBUG(D_NET, "unlinking peer 0x%p->%s\n",
1306 peer, libcfs_nid2str(peer->gnp_nid));
1308 list_del_init(&peer->gnp_list);
1309 kgnilnd_data.kgn_peer_version++;
1310 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1311 /* lose peerlist's ref */
1312 kgnilnd_peer_decref(peer);
1316 kgnilnd_get_peer_info(int index,
1317 kgn_peer_t **found_peer,
1318 lnet_nid_t *id, __u32 *nic_addr,
1319 int *refcount, int *connecting)
1321 struct list_head *ptmp;
1326 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1328 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1330 list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
1331 peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1336 CDEBUG(D_NET, "found peer %p (%s) at index %d\n",
1337 peer, libcfs_nid2str(peer->gnp_nid), index);
1340 *id = peer->gnp_nid;
1341 *nic_addr = peer->gnp_host_id;
1342 *refcount = atomic_read(&peer->gnp_refcount);
1343 *connecting = peer->gnp_connecting;
1350 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1352 CDEBUG(D_NET, "no gni peer at index %d\n", index);
1356 /* requires write_lock on kgn_peer_conn_lock held */
1358 kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp)
1360 kgn_peer_t *peer, *peer2;
1362 LASSERTF(new_stub_peer != NULL, "bad stub peer for nid %s\n",
1363 libcfs_nid2str(nid));
1365 peer2 = kgnilnd_find_peer_locked(nid);
1366 if (peer2 != NULL) {
1367 /* A peer was created during the lock transition, so drop
1368 * the new one we created */
1369 kgnilnd_peer_decref(new_stub_peer);
1372 peer = new_stub_peer;
1373 /* peer table takes existing ref on peer */
1375 LASSERTF(!kgnilnd_peer_active(peer),
1376 "peer 0x%p->%s already in peer table\n",
1377 peer, libcfs_nid2str(peer->gnp_nid));
1378 list_add_tail(&peer->gnp_list,
1379 kgnilnd_nid2peerlist(nid));
1380 kgnilnd_data.kgn_peer_version++;
1383 LASSERTF(peer->gnp_net != NULL, "peer 0x%p->%s with NULL net\n",
1384 peer, libcfs_nid2str(peer->gnp_nid));
1389 kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp)
1396 if (nid == LNET_NID_ANY)
1399 node_state = kgnilnd_get_node_state(LNET_NIDADDR(nid));
1401 /* NB - this will not block during normal operations -
1402 * the only writer of this is in the startup/shutdown path. */
1403 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1408 rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
1410 up_read(&kgnilnd_data.kgn_net_rw_sem);
1414 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1415 up_read(&kgnilnd_data.kgn_net_rw_sem);
1417 kgnilnd_add_peer_locked(nid, peer, peerp);
1419 CDEBUG(D_NET, "peer 0x%p->%s connecting %d\n",
1420 peerp, libcfs_nid2str((*peerp)->gnp_nid),
1421 (*peerp)->gnp_connecting);
1423 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1427 /* needs write_lock on kgn_peer_conn_lock */
1429 kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies)
1433 /* we do care about state of gnp_connecting - we could be between
1434 * reconnect attempts, so try to find the dgram and cancel the TX
1435 * anyways. If we are in the process of posting DONT do anything;
1436 * once it fails or succeeds we can nuke the connect attempt.
1437 * We have no idea where in kgnilnd_post_dgram we are so we cant
1438 * attempt to cancel until the function is done.
1441 /* make sure peer isn't in process of connecting or waiting for connect*/
1442 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1443 if (!(list_empty(&peer->gnp_connd_list))) {
1444 list_del_init(&peer->gnp_connd_list);
1445 /* remove connd ref */
1446 kgnilnd_peer_decref(peer);
1448 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1450 if (peer->gnp_connecting == GNILND_PEER_POSTING || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1451 peer->gnp_connecting = GNILND_PEER_NEEDS_DEATH;
1452 /* We are in process of posting right now the xchg set it up for us to
1453 * cancel the connect so we are finished for now */
1455 /* no need for exchange we have the peer lock and its ready for us to nuke */
1456 LASSERTF(peer->gnp_connecting != GNILND_PEER_POSTING,
1457 "Peer in invalid state 0x%p->%s, connecting %d\n",
1458 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1459 peer->gnp_connecting = GNILND_PEER_IDLE;
1460 set_mb(peer->gnp_last_dgram_errno, -ETIMEDOUT);
1461 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
1465 /* The least we can do is nuke the tx's no matter what.... */
1466 list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1467 kgnilnd_tx_del_state_locked(tx, peer, NULL,
1469 list_add_tail(&tx->tx_list, zombies);
1473 /* needs write_lock on kgn_peer_conn_lock */
1475 kgnilnd_del_peer_locked(kgn_peer_t *peer, int error)
1477 /* this peer could be passive and only held for purgatory,
1478 * take a ref to ensure it doesn't disappear in this function */
1479 kgnilnd_peer_addref(peer);
1481 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1483 /* if purgatory release cleared it out, don't try again */
1484 if (kgnilnd_peer_active(peer)) {
1485 /* always do this to allow kgnilnd_start_connect and
1486 * kgnilnd_finish_connect to catch this before they
1487 * wrap up their operations */
1488 if (kgnilnd_can_unlink_peer_locked(peer)) {
1489 /* already released purgatory, so only active
1491 kgnilnd_unlink_peer_locked(peer);
1493 kgnilnd_close_peer_conns_locked(peer, error);
1494 /* peer unlinks itself when last conn is closed */
1498 /* we are done, release back to the wild */
1499 kgnilnd_peer_decref(peer);
1503 kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
1507 LIST_HEAD (zombies);
1508 struct list_head *ptmp, *pnxt;
1515 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1517 if (nid != LNET_NID_ANY)
1518 lo = hi = kgnilnd_nid2peerlist(nid) - kgnilnd_data.kgn_peers;
1521 hi = *kgnilnd_tunables.kgn_peer_hash_size - 1;
1522 /* wildcards always succeed */
1526 for (i = lo; i <= hi; i++) {
1527 list_for_each_safe(ptmp, pnxt, &kgnilnd_data.kgn_peers[i]) {
1528 peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1530 LASSERTF(peer->gnp_net != NULL,
1531 "peer %p (%s) with NULL net\n",
1532 peer, libcfs_nid2str(peer->gnp_nid));
1534 if (net != NULL && peer->gnp_net != net)
1537 if (!(nid == LNET_NID_ANY || LNET_NIDADDR(peer->gnp_nid) == LNET_NIDADDR(nid)))
1540 /* In both cases, we want to stop any in-flight
1541 * connect attempts */
1542 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1545 case GNILND_DEL_CONN:
1546 kgnilnd_close_peer_conns_locked(peer, error);
1548 case GNILND_DEL_PEER:
1549 peer->gnp_pending_unlink = 1;
1550 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_unlink);
1551 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1552 kgnilnd_del_peer_locked(peer, error);
1554 case GNILND_CLEAR_PURGATORY:
1555 /* Mark everything ready for detach reaper will cleanup
1556 * once we release the kgn_peer_conn_lock
1558 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1559 peer->gnp_last_errno = -EISCONN;
1560 /* clear reconnect so he can reconnect soon */
1561 peer->gnp_reconnect_time = 0;
1562 peer->gnp_reconnect_interval = 0;
1565 CERROR("bad command %d\n", command);
1568 /* we matched something */
1573 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1576 kgnilnd_txlist_done(&zombies, error);
1578 /* This function does not return until the commands it initiated have completed,
1579 * since they have to work there way through the other threads. In the case of shutdown
1580 * threads are not woken up until after this call is initiated so we cannot wait, we just
1581 * need to return. The same applies for stack reset we shouldnt wait as the reset thread
1585 CFS_RACE(CFS_FAIL_GNI_RACE_RESET);
1587 if (error == -ENOTRECOVERABLE || error == -ESHUTDOWN) {
1592 while (atomic_read(&kgnilnd_data.kgn_npending_conns) ||
1593 atomic_read(&kgnilnd_data.kgn_npending_detach) ||
1594 atomic_read(&kgnilnd_data.kgn_npending_unlink)) {
1596 set_current_state(TASK_UNINTERRUPTIBLE);
1597 schedule_timeout(cfs_time_seconds(1));
1600 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n",
1601 atomic_read(&kgnilnd_data.kgn_npending_unlink),
1602 atomic_read(&kgnilnd_data.kgn_npending_conns),
1603 atomic_read(&kgnilnd_data.kgn_npending_detach));
1610 kgnilnd_get_conn_by_idx(int index)
1613 struct list_head *ptmp;
1615 struct list_head *ctmp;
1619 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1620 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1621 list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
1623 peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1625 list_for_each(ctmp, &peer->gnp_conns) {
1626 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1628 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1634 CDEBUG(D_NET, "++conn[%p] -> %s (%d)\n", conn,
1635 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1636 atomic_read(&conn->gnc_refcount));
1637 kgnilnd_conn_addref(conn);
1638 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1642 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1649 kgnilnd_get_conn_info(kgn_peer_t *peer,
1650 int *device_id, __u64 *peerstamp,
1651 int *tx_seq, int *rx_seq,
1652 int *fmaq_len, int *nfma, int *nrdma)
1657 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1659 conn = kgnilnd_find_conn_locked(peer);
1665 *device_id = conn->gnc_device->gnd_host_id;
1666 *peerstamp = conn->gnc_peerstamp;
1667 *tx_seq = atomic_read(&conn->gnc_tx_seq);
1668 *rx_seq = atomic_read(&conn->gnc_rx_seq);
1669 *fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq);
1670 *nfma = atomic_read(&conn->gnc_nlive_fma);
1671 *nrdma = atomic_read(&conn->gnc_nlive_rdma);
1673 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1677 /* needs write_lock on kgn_peer_conn_lock */
1679 kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why)
1682 struct list_head *ctmp, *cnxt;
1685 list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
1686 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1688 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1692 /* we mark gnc_needs closing and increment kgn_npending_conns so that
1693 * kgnilnd_del_conn_or_peer can wait on the other threads closing
1694 * and cleaning up the connection.
1696 if (!conn->gnc_needs_closing) {
1697 conn->gnc_needs_closing = 1;
1698 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_conns);
1700 kgnilnd_close_conn_locked(conn, why);
1706 kgnilnd_report_node_state(lnet_nid_t nid, int down)
1709 kgn_peer_t *peer, *new_peer;
1712 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1713 peer = kgnilnd_find_peer_locked(nid);
1720 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1722 /* Don't add a peer for node up events */
1723 if (down == GNILND_RCA_NODE_UP) {
1727 /* find any valid net - we don't care which one... */
1728 down_read(&kgnilnd_data.kgn_net_rw_sem);
1729 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
1730 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i],
1740 up_read(&kgnilnd_data.kgn_net_rw_sem);
1743 CNETERR("Could not find a net for nid %lld\n", nid);
1747 /* The nid passed in does not yet contain the net portion.
1748 * Let's build it up now
1750 nid = LNET_MKNID(LNET_NIDNET(net->gnn_ni->ni_nid), nid);
1751 rc = kgnilnd_add_peer(net, nid, &new_peer);
1754 CNETERR("Could not add peer for nid %lld, rc %d\n",
1759 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1760 peer = kgnilnd_find_peer_locked(nid);
1763 CNETERR("Could not find peer for nid %lld\n", nid);
1764 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1769 peer->gnp_down = down;
1771 if (down == GNILND_RCA_NODE_DOWN) {
1774 peer->gnp_down_event_time = jiffies;
1775 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1776 conn = kgnilnd_find_conn_locked(peer);
1779 kgnilnd_close_conn_locked(conn, -ENETRESET);
1782 peer->gnp_up_event_time = jiffies;
1785 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1787 if (down == GNILND_RCA_NODE_DOWN) {
1788 /* using ENETRESET so we don't get messages from
1791 kgnilnd_txlist_done(&zombies, -ENETRESET);
1792 kgnilnd_peer_notify(peer, -ECONNRESET, 0);
1793 LCONSOLE_INFO("Received down event for nid %d\n",
1801 kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
1803 struct libcfs_ioctl_data *data = arg;
1804 kgn_net_t *net = ni->ni_data;
1807 LASSERT(ni == net->gnn_ni);
1810 case IOC_LIBCFS_GET_PEER: {
1812 kgn_peer_t *peer = NULL;
1814 __u64 peerstamp = 0;
1815 int peer_refcount = 0, peer_connecting = 0;
1817 int tx_seq = 0, rx_seq = 0;
1818 int fmaq_len = 0, nfma = 0, nrdma = 0;
1820 rc = kgnilnd_get_peer_info(data->ioc_count, &peer,
1821 &nid, &nic_addr, &peer_refcount,
1827 /* LNET_MKNID is used to mask from lnet the multiplexing/demultiplexing of connections and peers
1828 * LNET assumes a conn and peer per net, the LNET_MKNID/LNET_NIDADDR allows us to let Lnet see what it
1829 * wants to see instead of the underlying network that is being used to send the data
1831 data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(nid));
1832 data->ioc_flags = peer_connecting;
1833 data->ioc_count = peer_refcount;
1835 rc = kgnilnd_get_conn_info(peer, &device_id, &peerstamp,
1836 &tx_seq, &rx_seq, &fmaq_len,
1839 /* This is allowable - a persistent peer could not
1840 * have a connection */
1842 /* flag to indicate we are not connected -
1843 * need to print as such */
1844 data->ioc_flags |= (1<<16);
1848 data->ioc_net = device_id;
1849 data->ioc_u64[0] = peerstamp;
1850 data->ioc_u32[0] = fmaq_len;
1851 data->ioc_u32[1] = nfma;
1852 data->ioc_u32[2] = tx_seq;
1853 data->ioc_u32[3] = rx_seq;
1854 data->ioc_u32[4] = nrdma;
1858 case IOC_LIBCFS_ADD_PEER: {
1859 /* just dummy value to allow using common interface */
1861 rc = kgnilnd_add_peer(net, data->ioc_nid, &peer);
1864 case IOC_LIBCFS_DEL_PEER: {
1865 /* NULL is passed in so it affects all peers in existence without regard to network
1866 * as the peer may not exist on the network LNET believes it to be on.
1868 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1869 GNILND_DEL_PEER, -EUCLEAN);
1872 case IOC_LIBCFS_GET_CONN: {
1873 kgn_conn_t *conn = kgnilnd_get_conn_by_idx(data->ioc_count);
1879 /* LNET_MKNID is used to build the correct address based on what LNET wants to see instead of
1880 * the generic connection that is used to send the data
1882 data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(conn->gnc_peer->gnp_nid));
1883 data->ioc_u32[0] = conn->gnc_device->gnd_id;
1884 kgnilnd_conn_decref(conn);
1888 case IOC_LIBCFS_CLOSE_CONNECTION: {
1889 /* use error = -ENETRESET to indicate it was lctl disconnect */
1890 /* NULL is passed in so it affects all the nets as the connection is virtual
1891 * and may not exist on the network LNET believes it to be on.
1893 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1894 GNILND_DEL_CONN, -ENETRESET);
1897 case IOC_LIBCFS_PUSH_CONNECTION: {
1898 /* we use this to flush purgatory */
1899 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1900 GNILND_CLEAR_PURGATORY, -EUCLEAN);
1903 case IOC_LIBCFS_REGISTER_MYNID: {
1904 /* Ignore if this is a noop */
1905 if (data->ioc_nid == ni->ni_nid) {
1908 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1909 libcfs_nid2str(data->ioc_nid),
1910 libcfs_nid2str(ni->ni_nid));
1921 kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1923 kgn_net_t *net = ni->ni_data;
1925 kgn_peer_t *peer = NULL;
1926 kgn_conn_t *conn = NULL;
1927 lnet_process_id_t id = {
1929 .pid = LNET_PID_LUSTRE,
1933 /* I expect to find him, so only take a read lock */
1934 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1935 peer = kgnilnd_find_peer_locked(nid);
1937 /* LIE if in a quiesce - we will update the timeouts after,
1938 * but we don't want sends failing during it */
1939 if (kgnilnd_data.kgn_quiesce_trigger) {
1941 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1945 /* Update to best guess, might refine on later checks */
1946 *when = peer->gnp_last_alive;
1948 /* we have a peer, how about a conn? */
1949 conn = kgnilnd_find_conn_locked(peer);
1952 /* if there is no conn, check peer last errno to see if clean disconnect
1953 * - if it was, we lie to LNet because we believe a TX would complete
1955 if (kgnilnd_conn_clean_errno(peer->gnp_last_errno)) {
1958 /* we still want to fire a TX and new conn in this case */
1960 /* gnp_last_alive is valid, run for the hills */
1961 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1965 /* if we get here, either we have no peer or no conn for him, so fire off
1966 * new TX to trigger conn setup */
1967 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1969 /* if we couldn't find him, we'll fire up a TX and get connected -
1970 * if we don't do this, after ni_peer_timeout, LNet will declare him dead.
1971 * So really we treat kgnilnd_query as a bit of a 'connect now' type
1972 * event because it'll only do this when it wants to send
1974 * Use a real TX for this to get the proper gnp_tx_queue behavior, etc
1975 * normally we'd use kgnilnd_send_ctlmsg for this, but we don't really
1976 * care that this goes out quickly since we already know we need a new conn
1978 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
1981 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, ni->ni_nid);
1983 kgnilnd_launch_tx(tx, net, &id);
1986 CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lu\n", peer,
1987 libcfs_nid2str(nid), *when);
1992 kgnilnd_dev_init(kgn_device_t *dev)
1996 unsigned int cq_size;
1999 /* size of these CQs should be able to accommodate the outgoing
2000 * RDMA and SMSG transactions. Since we really don't know what we
2001 * really need here, we'll take credits * 2 * 3 to allow a bunch.
2002 * We need to dig into this more with the performance work. */
2003 cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
2005 rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
2006 *kgnilnd_tunables.kgn_pkey, 0,
2008 if (rrc != GNI_RC_SUCCESS) {
2009 CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
2010 GOTO(failed, rc = -ENODEV);
2013 rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
2014 &dev->gnd_host_id, &dev->gnd_handle);
2015 if (rrc != GNI_RC_SUCCESS) {
2016 CERROR("Can't attach CDM to device %d (%d)\n",
2018 GOTO(failed, rc = -ENODEV);
2021 /* a bit gross, but not much we can do - Aries Sim doesn't have
2022 * hardcoded NIC/NID that we can use */
2023 rc = kgnilnd_setup_nic_translation(dev->gnd_host_id);
2025 GOTO(failed, rc = -ENODEV);
2027 /* only dev 0 gets the errors - no need to reset the stack twice
2028 * - this works because we have a single PTAG, if we had more
2029 * then we'd need to have multiple handlers */
2030 if (dev->gnd_id == 0) {
2031 rrc = kgnilnd_subscribe_errors(dev->gnd_handle,
2032 GNI_ERRMASK_CRITICAL |
2033 GNI_ERRMASK_UNKNOWN_TRANSACTION,
2034 0, NULL, kgnilnd_critical_error,
2035 &dev->gnd_err_handle);
2036 if (rrc != GNI_RC_SUCCESS) {
2037 CERROR("Can't subscribe for errors on device %d: rc %d\n",
2039 GOTO(failed, rc = -ENODEV);
2042 rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
2043 kgnilnd_quiesce_end_callback);
2044 if (rc != GNI_RC_SUCCESS) {
2045 CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
2047 GOTO(failed, rc = -ENODEV);
2051 rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock);
2053 CERROR("sock_create returned %d\n", rrc);
2057 rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
2059 /* log messages during startup */
2060 if (kgnilnd_data.kgn_init < GNILND_INIT_ALL) {
2061 CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
2062 dev->gnd_host_id, rc);
2064 GOTO(failed, rc = -ESRCH);
2066 CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);
2068 rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits,
2069 0, kgnilnd_device_callback,
2070 dev->gnd_id, &dev->gnd_snd_rdma_cqh);
2071 if (rrc != GNI_RC_SUCCESS) {
2072 CERROR("Can't create rdma send cq size %u for device "
2073 "%d (%d)\n", cq_size, dev->gnd_id, rrc);
2074 GOTO(failed, rc = -EINVAL);
2077 rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
2078 0, kgnilnd_device_callback, dev->gnd_id,
2079 &dev->gnd_snd_fma_cqh);
2080 if (rrc != GNI_RC_SUCCESS) {
2081 CERROR("Can't create fma send cq size %u for device %d (%d)\n",
2082 cq_size, dev->gnd_id, rrc);
2083 GOTO(failed, rc = -EINVAL);
2086 /* This one we size differently - overflows are possible and it needs to be
2087 * sized based on machine size */
2088 rrc = kgnilnd_cq_create(dev->gnd_handle,
2089 *kgnilnd_tunables.kgn_fma_cq_size,
2090 0, kgnilnd_device_callback, dev->gnd_id,
2091 &dev->gnd_rcv_fma_cqh);
2092 if (rrc != GNI_RC_SUCCESS) {
2093 CERROR("Can't create fma cq size %d for device %d (%d)\n",
2094 *kgnilnd_tunables.kgn_fma_cq_size, dev->gnd_id, rrc);
2095 GOTO(failed, rc = -EINVAL);
2098 rrc = kgnilnd_register_smdd_buf(dev);
2099 if (rrc != GNI_RC_SUCCESS) {
2100 GOTO(failed, rc = -EINVAL);
2106 kgnilnd_dev_fini(dev);
2111 kgnilnd_dev_fini(kgn_device_t *dev)
2116 /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
2117 LASSERTF(list_empty(&dev->gnd_ready_conns) &&
2118 list_empty(&dev->gnd_map_tx) &&
2119 list_empty(&dev->gnd_rdmaq),
2120 "dev 0x%p ready_conns %d@0x%p map_tx %d@0x%p rdmaq %d@0x%p\n",
2121 dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
2122 kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
2123 kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
2125 /* These should follow from tearing down all connections */
2126 LASSERTF(dev->gnd_map_nphys == 0 && dev->gnd_map_physnop == 0,
2127 "%d physical mappings of %d pages still mapped\n",
2128 dev->gnd_map_nphys, dev->gnd_map_physnop);
2130 LASSERTF(dev->gnd_map_nvirt == 0 && dev->gnd_map_virtnob == 0,
2131 "%d virtual mappings of "LPU64" bytes still mapped\n",
2132 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
2134 LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
2135 atomic_read(&dev->gnd_n_mdd_held) == 0 &&
2136 atomic64_read(&dev->gnd_nbytes_map) == 0,
2137 "%d SMSG mappings of %ld bytes still mapped or held %d\n",
2138 atomic_read(&dev->gnd_n_mdd),
2139 atomic64_read(&dev->gnd_nbytes_map), atomic_read(&dev->gnd_n_mdd_held));
2141 LASSERT(list_empty(&dev->gnd_map_list));
2143 /* What other assertions needed to ensure all connections torn down ? */
2145 /* check all counters == 0 (EP, MDD, etc) */
2147 /* if we are resetting due to quiese (stack reset), don't check
2149 LASSERTF(kgnilnd_data.kgn_quiesce_trigger ||
2150 atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
2151 "tried to shutdown with threads active\n");
2153 if (dev->gnd_smdd_hold_buf) {
2154 rrc = kgnilnd_deregister_smdd_buf(dev);
2155 LASSERTF(rrc == GNI_RC_SUCCESS,
2156 "bad rc from deregistion of sMDD buffer: %d\n", rrc);
2157 dev->gnd_smdd_hold_buf = NULL;
2160 if (dev->gnd_rcv_fma_cqh) {
2161 rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
2162 LASSERTF(rrc == GNI_RC_SUCCESS,
2163 "bad rc from gni_cq_destroy on rcv_fma_cqh: %d\n", rrc);
2164 dev->gnd_rcv_fma_cqh = NULL;
2167 if (dev->gnd_snd_rdma_cqh) {
2168 rrc = kgnilnd_cq_destroy(dev->gnd_snd_rdma_cqh);
2169 LASSERTF(rrc == GNI_RC_SUCCESS,
2170 "bad rc from gni_cq_destroy on send_rdma_cqh: %d\n", rrc);
2171 dev->gnd_snd_rdma_cqh = NULL;
2174 if (dev->gnd_snd_fma_cqh) {
2175 rrc = kgnilnd_cq_destroy(dev->gnd_snd_fma_cqh);
2176 LASSERTF(rrc == GNI_RC_SUCCESS,
2177 "bad rc from gni_cq_destroy on snd_fma_cqh: %d\n", rrc);
2178 dev->gnd_snd_fma_cqh = NULL;
2181 if (dev->gnd_err_handle) {
2182 rrc = kgnilnd_release_errors(dev->gnd_err_handle);
2183 LASSERTF(rrc == GNI_RC_SUCCESS,
2184 "bad rc from gni_release_errors: %d\n", rrc);
2185 dev->gnd_err_handle = NULL;
2188 if (dev->gnd_domain) {
2189 rrc = kgnilnd_cdm_destroy(dev->gnd_domain);
2190 LASSERTF(rrc == GNI_RC_SUCCESS,
2191 "bad rc from gni_cdm_destroy: %d\n", rrc);
2192 dev->gnd_domain = NULL;
2195 if (kgnilnd_data.kgn_sock)
2196 sock_release(kgnilnd_data.kgn_sock);
2201 int kgnilnd_base_startup(void)
2204 int pkmem = atomic_read(&libcfs_kmemory);
2208 struct task_struct *thrd;
2210 #if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
2211 /* limit how much memory can be allocated for fma blocks in
2212 * instances where many nodes need to reconnects at the same time */
2215 kgnilnd_data.free_pages_limit = si.totalram/4;
2220 LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
2221 "init %d\n", kgnilnd_data.kgn_init);
2223 /* zero pointers, flags etc */
2224 memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));
2225 kgnilnd_check_kgni_version();
2227 /* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
2228 * a unique (for all time) connstamp so we can uniquely identify
2229 * the sender. The connstamp is an incrementing counter
2230 * initialised with seconds + microseconds at startup time. So we
2231 * rely on NOT creating connections more frequently on average than
2232 * 1MHz to ensure we don't use old connstamps when we reboot. */
2233 do_gettimeofday(&tv);
2234 kgnilnd_data.kgn_connstamp =
2235 kgnilnd_data.kgn_peerstamp =
2236 (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2238 init_rwsem(&kgnilnd_data.kgn_net_rw_sem);
2240 for (i = 0; i < GNILND_MAXDEVS; i++) {
2241 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2244 INIT_LIST_HEAD(&dev->gnd_ready_conns);
2245 INIT_LIST_HEAD(&dev->gnd_map_tx);
2246 INIT_LIST_HEAD(&dev->gnd_fma_buffs);
2247 mutex_init(&dev->gnd_cq_mutex);
2248 mutex_init(&dev->gnd_fmablk_mutex);
2249 spin_lock_init(&dev->gnd_fmablk_lock);
2250 init_waitqueue_head(&dev->gnd_waitq);
2251 init_waitqueue_head(&dev->gnd_dgram_waitq);
2252 init_waitqueue_head(&dev->gnd_dgping_waitq);
2253 spin_lock_init(&dev->gnd_lock);
2254 INIT_LIST_HEAD(&dev->gnd_map_list);
2255 spin_lock_init(&dev->gnd_map_lock);
2256 atomic_set(&dev->gnd_nfmablk, 0);
2257 atomic_set(&dev->gnd_fmablk_vers, 1);
2258 atomic_set(&dev->gnd_neps, 0);
2259 atomic_set(&dev->gnd_canceled_dgrams, 0);
2260 INIT_LIST_HEAD(&dev->gnd_connd_peers);
2261 spin_lock_init(&dev->gnd_connd_lock);
2262 spin_lock_init(&dev->gnd_dgram_lock);
2263 spin_lock_init(&dev->gnd_rdmaq_lock);
2264 INIT_LIST_HEAD(&dev->gnd_rdmaq);
2265 init_rwsem(&dev->gnd_conn_sem);
2267 /* alloc & setup nid based dgram table */
2268 LIBCFS_ALLOC(dev->gnd_dgrams,
2269 sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
2271 if (dev->gnd_dgrams == NULL)
2272 GOTO(failed, rc = -ENOMEM);
2274 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2275 INIT_LIST_HEAD(&dev->gnd_dgrams[i]);
2277 atomic_set(&dev->gnd_ndgrams, 0);
2278 atomic_set(&dev->gnd_nwcdgrams, 0);
2279 /* setup timer for RDMAQ processing */
2280 setup_timer(&dev->gnd_rdmaq_timer, kgnilnd_schedule_device_timer,
2281 (unsigned long)dev);
2283 /* setup timer for mapping processing */
2284 setup_timer(&dev->gnd_map_timer, kgnilnd_schedule_device_timer,
2285 (unsigned long)dev);
2289 /* CQID 0 isn't allowed, set to MAX_MSG_ID - 1 to check for conflicts early */
2290 kgnilnd_data.kgn_next_cqid = GNILND_MAX_MSG_ID - 1;
2291 kgnilnd_data.kgn_new_min_timeout = *kgnilnd_tunables.kgn_timeout;
2292 init_waitqueue_head(&kgnilnd_data.kgn_reaper_waitq);
2293 init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
2294 spin_lock_init(&kgnilnd_data.kgn_reaper_lock);
2296 mutex_init(&kgnilnd_data.kgn_quiesce_mutex);
2297 atomic_set(&kgnilnd_data.kgn_nquiesce, 0);
2298 atomic_set(&kgnilnd_data.kgn_npending_conns, 0);
2299 atomic_set(&kgnilnd_data.kgn_npending_unlink, 0);
2300 atomic_set(&kgnilnd_data.kgn_npending_detach, 0);
2301 atomic_set(&kgnilnd_data.kgn_rev_offset, 0);
2302 atomic_set(&kgnilnd_data.kgn_rev_length, 0);
2303 atomic_set(&kgnilnd_data.kgn_rev_copy_buff, 0);
2305 /* OK to call kgnilnd_api_shutdown() to cleanup now */
2306 kgnilnd_data.kgn_init = GNILND_INIT_DATA;
2307 try_module_get(THIS_MODULE);
2309 rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);
2311 LIBCFS_ALLOC(kgnilnd_data.kgn_peers,
2312 sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
2314 if (kgnilnd_data.kgn_peers == NULL)
2315 GOTO(failed, rc = -ENOMEM);
2317 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2318 INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
2321 LIBCFS_ALLOC(kgnilnd_data.kgn_conns,
2322 sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
2324 if (kgnilnd_data.kgn_conns == NULL)
2325 GOTO(failed, rc = -ENOMEM);
2327 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2328 INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
2331 LIBCFS_ALLOC(kgnilnd_data.kgn_nets,
2332 sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size);
2334 if (kgnilnd_data.kgn_nets == NULL)
2335 GOTO(failed, rc = -ENOMEM);
2337 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2338 INIT_LIST_HEAD(&kgnilnd_data.kgn_nets[i]);
2341 kgnilnd_data.kgn_mbox_cache =
2342 kmem_cache_create("kgn_mbox_block", KMALLOC_MAX_SIZE, 0,
2343 SLAB_HWCACHE_ALIGN, NULL);
2344 if (kgnilnd_data.kgn_mbox_cache == NULL) {
2345 CERROR("Can't create slab for physical mbox blocks\n");
2346 GOTO(failed, rc = -ENOMEM);
2349 kgnilnd_data.kgn_rx_cache =
2350 kmem_cache_create("kgn_rx_t", sizeof(kgn_rx_t), 0, 0, NULL);
2351 if (kgnilnd_data.kgn_rx_cache == NULL) {
2352 CERROR("Can't create slab for kgn_rx_t descriptors\n");
2353 GOTO(failed, rc = -ENOMEM);
2356 kgnilnd_data.kgn_tx_cache =
2357 kmem_cache_create("kgn_tx_t", sizeof(kgn_tx_t), 0, 0, NULL);
2358 if (kgnilnd_data.kgn_tx_cache == NULL) {
2359 CERROR("Can't create slab for kgn_tx_t\n");
2360 GOTO(failed, rc = -ENOMEM);
2363 kgnilnd_data.kgn_tx_phys_cache =
2364 kmem_cache_create("kgn_tx_phys",
2365 LNET_MAX_IOV * sizeof(gni_mem_segment_t),
2367 if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
2368 CERROR("Can't create slab for kgn_tx_phys\n");
2369 GOTO(failed, rc = -ENOMEM);
2372 kgnilnd_data.kgn_dgram_cache =
2373 kmem_cache_create("kgn_dgram_t", sizeof(kgn_dgram_t), 0, 0, NULL);
2374 if (kgnilnd_data.kgn_dgram_cache == NULL) {
2375 CERROR("Can't create slab for outgoing datagrams\n");
2376 GOTO(failed, rc = -ENOMEM);
2379 /* allocate a MAX_IOV array of page pointers for each cpu */
2380 kgnilnd_data.kgn_cksum_map_pages = kmalloc(num_possible_cpus() * sizeof (struct page *),
2382 if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
2383 CERROR("Can't allocate vmap cksum pages\n");
2384 GOTO(failed, rc = -ENOMEM);
2386 kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
2387 memset(kgnilnd_data.kgn_cksum_map_pages, 0,
2388 kgnilnd_data.kgn_cksum_npages * sizeof (struct page *));
2390 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2391 kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(LNET_MAX_IOV * sizeof (struct page *),
2393 if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
2394 CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
2395 GOTO(failed, rc = -ENOMEM);
2399 LASSERT(kgnilnd_data.kgn_ndevs == 0);
2401 /* Use all available GNI devices */
2402 for (i = 0; i < GNILND_MAXDEVS; i++) {
2403 dev = &kgnilnd_data.kgn_devices[kgnilnd_data.kgn_ndevs];
2405 rc = kgnilnd_dev_init(dev);
2407 /* Increment here so base_shutdown cleans it up */
2408 kgnilnd_data.kgn_ndevs++;
2410 rc = kgnilnd_allocate_phys_fmablk(dev);
2416 if (kgnilnd_data.kgn_ndevs == 0) {
2417 CERROR("Can't initialise any GNI devices\n");
2418 GOTO(failed, rc = -ENODEV);
2421 rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
2423 CERROR("Can't spawn gnilnd reaper: %d\n", rc);
2427 rc = kgnilnd_start_rca_thread();
2429 CERROR("Can't spawn gnilnd rca: %d\n", rc);
2434 * Start ruhroh thread. We can't use kgnilnd_thread_start() because
2435 * we don't want this thread included in kgnilnd_data.kgn_nthreads
2436 * count. This thread controls quiesce, so it mustn't
2439 thrd = kthread_run(kgnilnd_ruhroh_thread, NULL, "%s_%02d", "kgnilnd_rr", 0);
2442 CERROR("Can't spawn gnilnd ruhroh thread: %d\n", rc);
2446 /* threads will load balance across devs as they are available */
2447 if (*kgnilnd_tunables.kgn_thread_affinity) {
2448 rc = kgnilnd_start_sd_threads();
2452 for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
2453 rc = kgnilnd_thread_start(kgnilnd_scheduler,
2457 CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
2464 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2465 dev = &kgnilnd_data.kgn_devices[i];
2466 rc = kgnilnd_thread_start(kgnilnd_dgram_mover, dev,
2467 "kgnilnd_dg", dev->gnd_id);
2469 CERROR("Can't spawn gnilnd dgram_mover[%d]: %d\n",
2474 rc = kgnilnd_thread_start(kgnilnd_dgram_waitq, dev,
2475 "kgnilnd_dgn", dev->gnd_id);
2477 CERROR("Can't spawn gnilnd dgram_waitq[%d]: %d\n",
2482 rc = kgnilnd_setup_wildcard_dgram(dev);
2485 CERROR("Can't create wildcard dgrams[%d]: %d\n",
2491 /* flag everything initialised */
2492 kgnilnd_data.kgn_init = GNILND_INIT_ALL;
2493 /*****************************************************/
2495 CDEBUG(D_MALLOC, "initial kmem %d\n", pkmem);
2499 kgnilnd_base_shutdown();
2500 kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2505 kgnilnd_base_shutdown(void)
2510 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};
2512 kgnilnd_data.kgn_wc_kill = 1;
2514 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2515 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2516 kgnilnd_cancel_wc_dgrams(dev);
2517 kgnilnd_cancel_dgrams(dev);
2518 kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2519 kgnilnd_wait_for_canceled_dgrams(dev);
2522 /* We need to verify there are no conns left before we let the threads
2523 * shut down otherwise we could clean up the peers but still have
2524 * some outstanding conns due to orphaned datagram conns that are
2528 while (atomic_read(&kgnilnd_data.kgn_nconns) != 0) {
2531 for(j = 0; j < kgnilnd_data.kgn_ndevs; ++j) {
2532 kgn_device_t *dev = &kgnilnd_data.kgn_devices[j];
2533 kgnilnd_schedule_device(dev);
2536 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2537 "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
2538 set_current_state(TASK_UNINTERRUPTIBLE);
2539 schedule_timeout(cfs_time_seconds(1));
2541 /* Peer state all cleaned up BEFORE setting shutdown, so threads don't
2542 * have to worry about shutdown races. NB connections may be created
2543 * while there are still active connds, but these will be temporary
2544 * since peer creation always fails after the listener has started to
2546 * all peers should have been cleared out on the nets */
2547 LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2548 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2550 /* Wait for the ruhroh thread to shut down. */
2551 kgnilnd_data.kgn_ruhroh_shutdown = 1;
2552 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
2554 while (kgnilnd_data.kgn_ruhroh_running != 0) {
2556 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2557 "Waiting for ruhroh thread to terminate\n");
2558 set_current_state(TASK_UNINTERRUPTIBLE);
2559 schedule_timeout(cfs_time_seconds(1));
2562 /* Flag threads to terminate */
2563 kgnilnd_data.kgn_shutdown = 1;
2565 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2566 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2568 /* should clear all the MDDs */
2569 kgnilnd_unmap_fma_blocks(dev);
2571 kgnilnd_schedule_device(dev);
2572 wake_up_all(&dev->gnd_dgram_waitq);
2573 wake_up_all(&dev->gnd_dgping_waitq);
2574 LASSERT(list_empty(&dev->gnd_connd_peers));
2577 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2578 wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
2579 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2581 if (atomic_read(&kgnilnd_data.kgn_nthreads))
2582 kgnilnd_wakeup_rca_thread();
2584 /* Wait for threads to exit */
2586 while (atomic_read(&kgnilnd_data.kgn_nthreads) != 0) {
2588 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2589 "Waiting for %d threads to terminate\n",
2590 atomic_read(&kgnilnd_data.kgn_nthreads));
2591 set_current_state(TASK_UNINTERRUPTIBLE);
2592 schedule_timeout(cfs_time_seconds(1));
2595 LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2596 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2598 if (kgnilnd_data.kgn_peers != NULL) {
2599 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2600 LASSERT(list_empty(&kgnilnd_data.kgn_peers[i]));
2602 LIBCFS_FREE(kgnilnd_data.kgn_peers,
2603 sizeof (struct list_head) *
2604 *kgnilnd_tunables.kgn_peer_hash_size);
2607 down_write(&kgnilnd_data.kgn_net_rw_sem);
2608 if (kgnilnd_data.kgn_nets != NULL) {
2609 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++)
2610 LASSERT(list_empty(&kgnilnd_data.kgn_nets[i]));
2612 LIBCFS_FREE(kgnilnd_data.kgn_nets,
2613 sizeof (struct list_head) *
2614 *kgnilnd_tunables.kgn_net_hash_size);
2616 up_write(&kgnilnd_data.kgn_net_rw_sem);
2618 LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
2619 "conns left %d\n", atomic_read(&kgnilnd_data.kgn_nconns));
2621 if (kgnilnd_data.kgn_conns != NULL) {
2622 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2623 LASSERT(list_empty(&kgnilnd_data.kgn_conns[i]));
2625 LIBCFS_FREE(kgnilnd_data.kgn_conns,
2626 sizeof (struct list_head) *
2627 *kgnilnd_tunables.kgn_peer_hash_size);
2630 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2631 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2632 kgnilnd_dev_fini(dev);
2634 LASSERTF(atomic_read(&dev->gnd_ndgrams) == 0,
2635 "dgrams left %d\n", atomic_read(&dev->gnd_ndgrams));
2637 if (dev->gnd_dgrams != NULL) {
2638 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2639 LASSERT(list_empty(&dev->gnd_dgrams[i]));
2641 LIBCFS_FREE(dev->gnd_dgrams,
2642 sizeof (struct list_head) *
2643 *kgnilnd_tunables.kgn_peer_hash_size);
2646 kgnilnd_free_phys_fmablk(dev);
2649 if (kgnilnd_data.kgn_mbox_cache != NULL)
2650 kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
2652 if (kgnilnd_data.kgn_rx_cache != NULL)
2653 kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);
2655 if (kgnilnd_data.kgn_tx_cache != NULL)
2656 kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);
2658 if (kgnilnd_data.kgn_tx_phys_cache != NULL)
2659 kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
2661 if (kgnilnd_data.kgn_dgram_cache != NULL)
2662 kmem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
2664 if (kgnilnd_data.kgn_cksum_map_pages != NULL) {
2665 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2666 if (kgnilnd_data.kgn_cksum_map_pages[i] != NULL) {
2667 kfree(kgnilnd_data.kgn_cksum_map_pages[i]);
2670 kfree(kgnilnd_data.kgn_cksum_map_pages);
2673 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2674 atomic_read(&libcfs_kmemory));
2676 kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2677 module_put(THIS_MODULE);
2683 kgnilnd_startup(lnet_ni_t *ni)
2689 LASSERTF(ni->ni_lnd == &the_kgnilnd,
2690 "bad LND 0x%p != the_kgnilnd @ 0x%p\n",
2691 ni->ni_lnd, &the_kgnilnd);
2693 if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
2694 rc = kgnilnd_base_startup();
2699 /* Serialize with shutdown. */
2700 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2702 LIBCFS_ALLOC(net, sizeof(*net));
2704 CERROR("could not allocate net for new interface instance\n");
2705 /* no need to cleanup the CDM... */
2706 GOTO(failed, rc = -ENOMEM);
2708 INIT_LIST_HEAD(&net->gnn_list);
2711 ni->ni_maxtxcredits = *kgnilnd_tunables.kgn_credits;
2712 ni->ni_peertxcredits = *kgnilnd_tunables.kgn_peer_credits;
2714 if (*kgnilnd_tunables.kgn_peer_health) {
2717 /* give this a bit of leeway - we don't have a hard timeout
2718 * as we only check timeouts periodically - see comment in kgnilnd_reaper */
2719 fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
2720 timeout = *kgnilnd_tunables.kgn_timeout + fudge;
2722 if (*kgnilnd_tunables.kgn_peer_timeout >= timeout)
2723 ni->ni_peertimeout = *kgnilnd_tunables.kgn_peer_timeout;
2724 else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
2725 LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
2726 *kgnilnd_tunables.kgn_peer_timeout,
2729 LIBCFS_FREE(net, sizeof(*net));
2730 GOTO(failed, rc = -EINVAL);
2732 ni->ni_peertimeout = timeout;
2734 LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
2735 ni->ni_peertimeout);
2738 atomic_set(&net->gnn_refcount, 1);
2740 /* if we have multiple devices, spread the nets around */
2741 net->gnn_netnum = LNET_NETNUM(LNET_NIDNET(ni->ni_nid));
2743 devno = LNET_NIDNET(ni->ni_nid) % GNILND_MAXDEVS;
2744 net->gnn_dev = &kgnilnd_data.kgn_devices[devno];
2746 /* allocate a 'dummy' cdm for datagram use. We can only have a single
2747 * datagram between a nid:inst_id and nid2:inst_id. The fake cdm
2748 * give us additional inst_id to use, allowing the datagrams to flow
2749 * like rivers of honey and beer */
2751 /* the instance id for the cdm is the NETNUM offset by MAXDEVS -
2752 * ensuring we'll have a unique id */
2755 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), net->gnn_dev->gnd_nid);
2756 CDEBUG(D_NET, "adding net %p nid=%s on dev %d \n",
2757 net, libcfs_nid2str(ni->ni_nid), net->gnn_dev->gnd_id);
2758 /* until the gnn_list is set, we need to cleanup ourselves as
2759 * kgnilnd_shutdown is just gonna get confused */
2761 down_write(&kgnilnd_data.kgn_net_rw_sem);
2762 list_add_tail(&net->gnn_list, kgnilnd_netnum2netlist(net->gnn_netnum));
2763 up_write(&kgnilnd_data.kgn_net_rw_sem);
2765 /* we need a separate thread to call probe_wait_by_id until
2766 * we get a function callback notifier from kgni */
2767 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2770 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2771 kgnilnd_shutdown(ni);
2776 kgnilnd_shutdown(lnet_ni_t *ni)
2778 kgn_net_t *net = ni->ni_data;
2783 CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);
2785 LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_ALL,
2786 "init %d\n", kgnilnd_data.kgn_init);
2788 /* Serialize with startup. */
2789 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2790 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2791 atomic_read(&libcfs_kmemory));
2794 CERROR("got NULL net for ni %p\n", ni);
2795 GOTO(out, rc = -EINVAL);
2798 LASSERTF(ni == net->gnn_ni,
2799 "ni %p gnn_ni %p\n", net, net->gnn_ni);
2803 LASSERT(!net->gnn_shutdown);
2804 LASSERTF(atomic_read(&net->gnn_refcount) != 0,
2805 "net %p refcount %d\n",
2806 net, atomic_read(&net->gnn_refcount));
2808 if (!list_empty(&net->gnn_list)) {
2809 /* serialize with peer creation */
2810 down_write(&kgnilnd_data.kgn_net_rw_sem);
2811 net->gnn_shutdown = 1;
2812 up_write(&kgnilnd_data.kgn_net_rw_sem);
2814 kgnilnd_cancel_net_dgrams(net);
2816 kgnilnd_del_conn_or_peer(net, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2818 /* if we are quiesced, need to wake up - we need those threads
2819 * alive to release peers, etc */
2820 if (GNILND_IS_QUIESCED) {
2821 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
2822 kgnilnd_quiesce_wait("shutdown");
2825 kgnilnd_wait_for_canceled_dgrams(net->gnn_dev);
2827 /* We wait until the nets ref's are 1, we will release final ref which is ours
2828 * this allows us to make sure everything else is done before we free the
2832 while (atomic_read(&net->gnn_refcount) != 1) {
2834 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2835 "Waiting for %d references to clear on net %d\n",
2836 atomic_read(&net->gnn_refcount),
2838 set_current_state(TASK_UNINTERRUPTIBLE);
2839 schedule_timeout(cfs_time_seconds(1));
2842 /* release ref from kgnilnd_startup */
2843 kgnilnd_net_decref(net);
2844 /* serialize with reaper and conn_task looping */
2845 down_write(&kgnilnd_data.kgn_net_rw_sem);
2846 list_del_init(&net->gnn_list);
2847 up_write(&kgnilnd_data.kgn_net_rw_sem);
2851 /* not locking, this can't race with writers */
2852 LASSERTF(atomic_read(&net->gnn_refcount) == 0,
2853 "net %p refcount %d\n",
2854 net, atomic_read(&net->gnn_refcount));
2855 LIBCFS_FREE(net, sizeof(*net));
2858 down_read(&kgnilnd_data.kgn_net_rw_sem);
2859 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2860 if (!list_empty(&kgnilnd_data.kgn_nets[i])) {
2861 up_read(&kgnilnd_data.kgn_net_rw_sem);
2865 if (i == *kgnilnd_tunables.kgn_net_hash_size - 1) {
2866 up_read(&kgnilnd_data.kgn_net_rw_sem);
2867 kgnilnd_base_shutdown();
2870 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2871 atomic_read(&libcfs_kmemory));
2873 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2878 kgnilnd_module_fini(void)
2880 lnet_unregister_lnd(&the_kgnilnd);
2881 kgnilnd_proc_fini();
2882 kgnilnd_remove_sysctl();
2883 kgnilnd_tunables_fini();
2887 kgnilnd_module_init(void)
2891 rc = kgnilnd_tunables_init();
2895 printk(KERN_INFO "Lustre: kgnilnd build version: "KGNILND_BUILD_REV"\n");
2897 kgnilnd_insert_sysctl();
2898 kgnilnd_proc_init();
2900 lnet_register_lnd(&the_kgnilnd);
2905 MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
2906 MODULE_DESCRIPTION("Kernel Gemini LND v"KGNILND_BUILD_REV);
2907 MODULE_LICENSE("GPL");
2909 module_init(kgnilnd_module_init);
2910 module_exit(kgnilnd_module_fini);