2 * Copyright (C) 2012 Cray, Inc.
4 * Copyright (c) 2013, 2017, 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. */
28 struct lnet_lnd the_kgnilnd = {
30 .lnd_startup = kgnilnd_startup,
31 .lnd_shutdown = kgnilnd_shutdown,
32 .lnd_ctl = kgnilnd_ctl,
33 .lnd_send = kgnilnd_send,
34 .lnd_recv = kgnilnd_recv,
35 .lnd_eager_recv = kgnilnd_eager_recv,
36 .lnd_query = kgnilnd_query,
39 kgn_data_t kgnilnd_data;
42 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
44 struct task_struct *thrd;
46 thrd = kthread_run(fn, arg, "%s_%02d", name, id);
50 atomic_inc(&kgnilnd_data.kgn_nthreads);
54 /* bind scheduler threads to cpus */
56 kgnilnd_start_sd_threads(void)
60 struct task_struct *task;
62 for_each_online_cpu(cpu) {
63 /* don't bind to cpu 0 - all interrupts are processed here */
67 task = kthread_create(kgnilnd_scheduler, (void *)((long)i),
68 "%s_%02d", "kgnilnd_sd", i);
70 kthread_bind(task, cpu);
71 wake_up_process(task);
73 CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i,
77 atomic_inc(&kgnilnd_data.kgn_nthreads);
79 if (++i >= *kgnilnd_tunables.kgn_sched_threads) {
87 /* needs write_lock on kgn_peer_conn_lock */
89 kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
92 struct list_head *ctmp, *cnxt;
96 loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
98 list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
99 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
101 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
107 if (conn->gnc_device != newconn->gnc_device)
110 /* This is a two connection loopback - one talking to the other */
112 newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
113 newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) {
114 CDEBUG(D_NET, "skipping prune of %p, "
115 "loopback and matching stamps"
116 " connstamp %llu(%llu)"
117 " peerstamp %llu(%llu)\n",
118 conn, newconn->gnc_my_connstamp,
119 conn->gnc_peer_connstamp,
120 newconn->gnc_peer_connstamp,
121 conn->gnc_my_connstamp);
125 if (conn->gnc_peerstamp != newconn->gnc_peerstamp) {
126 LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp,
127 "conn 0x%p peerstamp %llu >= "
128 "newconn 0x%p peerstamp %llu\n",
129 conn, conn->gnc_peerstamp,
130 newconn, newconn->gnc_peerstamp);
132 CDEBUG(D_NET, "Closing stale conn nid: %s "
133 " peerstamp:%#llx(%#llx)\n",
134 libcfs_nid2str(peer->gnp_nid),
135 conn->gnc_peerstamp, newconn->gnc_peerstamp);
138 LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp,
139 "conn 0x%p peer_connstamp %llu >= "
140 "newconn 0x%p peer_connstamp %llu\n",
141 conn, conn->gnc_peer_connstamp,
142 newconn, newconn->gnc_peer_connstamp);
144 CDEBUG(D_NET, "Closing stale conn nid: %s"
145 " connstamp:%llu(%llu)\n",
146 libcfs_nid2str(peer->gnp_nid),
147 conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
151 kgnilnd_close_conn_locked(conn, -ESTALE);
155 CWARN("Closed %d stale conns to %s\n", count, libcfs_nid2str(peer->gnp_nid));
162 kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
165 struct list_head *tmp;
169 loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
171 list_for_each(tmp, &peer->gnp_conns) {
172 conn = list_entry(tmp, kgn_conn_t, gnc_list);
173 CDEBUG(D_NET, "checking conn 0x%p for peer %s"
174 " lo %d new %llu existing %llu"
175 " new peer %llu existing peer %llu"
176 " new dev %p existing dev %p\n",
177 conn, libcfs_nid2str(peer->gnp_nid),
179 newconn->gnc_peerstamp, conn->gnc_peerstamp,
180 newconn->gnc_peer_connstamp, conn->gnc_peer_connstamp,
181 newconn->gnc_device, conn->gnc_device);
183 /* conn is in the process of closing */
184 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
187 /* 'newconn' is from an earlier version of 'peer'!!! */
188 if (newconn->gnc_peerstamp < conn->gnc_peerstamp)
191 /* 'conn' is from an earlier version of 'peer': it will be
192 * removed when we cull stale conns later on... */
193 if (newconn->gnc_peerstamp > conn->gnc_peerstamp)
196 /* Different devices are OK */
197 if (conn->gnc_device != newconn->gnc_device)
200 /* It's me connecting to myself */
202 newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
203 newconn->gnc_peer_connstamp == conn->gnc_my_connstamp)
206 /* 'newconn' is an earlier connection from 'peer'!!! */
207 if (newconn->gnc_peer_connstamp < conn->gnc_peer_connstamp)
210 /* 'conn' is an earlier connection from 'peer': it will be
211 * removed when we cull stale conns later on... */
212 if (newconn->gnc_peer_connstamp > conn->gnc_peer_connstamp)
215 /* 'newconn' has the SAME connection stamp; 'peer' isn't
216 * playing the game... */
224 kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
230 LASSERT (!in_interrupt());
231 atomic_inc(&kgnilnd_data.kgn_nconns);
233 /* divide by 2 to allow for complete reset and immediate reconnect */
234 if (atomic_read(&kgnilnd_data.kgn_nconns) >= GNILND_MAX_CQID/2) {
235 CERROR("Too many conn are live: %d > %d\n",
236 atomic_read(&kgnilnd_data.kgn_nconns), GNILND_MAX_CQID/2);
237 atomic_dec(&kgnilnd_data.kgn_nconns);
241 LIBCFS_ALLOC(conn, sizeof(*conn));
243 atomic_dec(&kgnilnd_data.kgn_nconns);
247 conn->gnc_tx_ref_table =
248 kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *));
249 if (conn->gnc_tx_ref_table == NULL) {
250 CERROR("Can't allocate conn tx_ref_table\n");
251 GOTO(failed, rc = -ENOMEM);
254 mutex_init(&conn->gnc_smsg_mutex);
255 mutex_init(&conn->gnc_rdma_mutex);
256 atomic_set(&conn->gnc_refcount, 1);
257 atomic_set(&conn->gnc_reaper_noop, 0);
258 atomic_set(&conn->gnc_sched_noop, 0);
259 atomic_set(&conn->gnc_tx_in_use, 0);
260 INIT_LIST_HEAD(&conn->gnc_list);
261 INIT_LIST_HEAD(&conn->gnc_hashlist);
262 INIT_LIST_HEAD(&conn->gnc_schedlist);
263 INIT_LIST_HEAD(&conn->gnc_fmaq);
264 INIT_LIST_HEAD(&conn->gnc_mdd_list);
265 INIT_LIST_HEAD(&conn->gnc_delaylist);
266 spin_lock_init(&conn->gnc_list_lock);
267 spin_lock_init(&conn->gnc_tx_lock);
268 conn->gnc_magic = GNILND_CONN_MAGIC;
270 /* set tx id to nearly the end to make sure we find wrapping
272 conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10;
274 /* if this fails, we have conflicts and MAX_TX is too large */
275 CLASSERT(GNILND_MAX_MSG_ID < GNILND_MSGID_CLOSE);
277 /* get a new unique CQ id for this conn */
278 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
279 conn->gnc_my_connstamp = kgnilnd_data.kgn_connstamp++;
280 conn->gnc_cqid = kgnilnd_get_cqid_locked();
281 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
283 if (conn->gnc_cqid == 0) {
284 CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
285 GOTO(failed, rc = -E2BIG);
288 CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
289 conn->gnc_cqid, conn);
291 /* need to be set before gnc_ephandle to allow kgnilnd_destroy_conn_ep to
293 conn->gnc_device = dev;
295 conn->gnc_timeout = MAX(*kgnilnd_tunables.kgn_timeout,
297 kgnilnd_update_reaper_timeout(conn->gnc_timeout);
299 /* this is the ep_handle for doing SMSG & BTE */
300 mutex_lock(&dev->gnd_cq_mutex);
301 rrc = kgnilnd_ep_create(dev->gnd_handle, dev->gnd_snd_fma_cqh,
302 &conn->gnc_ephandle);
303 mutex_unlock(&dev->gnd_cq_mutex);
304 if (rrc != GNI_RC_SUCCESS)
305 GOTO(failed, rc = -ENETDOWN);
307 CDEBUG(D_NET, "created conn 0x%p ep_hndl 0x%p\n",
308 conn, conn->gnc_ephandle);
310 /* add ref for EP canceling */
311 kgnilnd_conn_addref(conn);
312 atomic_inc(&dev->gnd_neps);
318 atomic_dec(&kgnilnd_data.kgn_nconns);
319 kgnilnd_vfree(conn->gnc_tx_ref_table,
320 GNILND_MAX_MSG_ID * sizeof(void *));
321 LIBCFS_FREE(conn, sizeof(*conn));
325 /* needs to be called with kgn_peer_conn_lock held (read or write) */
327 kgnilnd_find_conn_locked(kgn_peer_t *peer)
329 kgn_conn_t *conn = NULL;
331 /* if we are in reset, this conn is going to die soon */
332 if (unlikely(kgnilnd_data.kgn_in_reset)) {
336 /* just return the first ESTABLISHED connection */
337 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
338 /* kgnilnd_finish_connect doesn't put connections on the
339 * peer list until they are actually established */
340 LASSERTF(conn->gnc_state >= GNILND_CONN_ESTABLISHED,
341 "found conn %p state %s on peer %p (%s)\n",
342 conn, kgnilnd_conn_state2str(conn), peer,
343 libcfs_nid2str(peer->gnp_nid));
344 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
352 /* needs write_lock on kgn_peer_conn_lock held */
354 kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer) {
356 kgn_device_t *dev = peer->gnp_net->gnn_dev;
359 conn = kgnilnd_find_conn_locked(peer);
365 /* if the peer was previously connecting, check if we should
366 * trigger another connection attempt yet. */
367 if (time_before(jiffies, peer->gnp_reconnect_time)) {
371 /* This check prevents us from creating a new connection to a peer while we are
372 * still in the process of closing an existing connection to the peer.
374 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
375 if (conn->gnc_ephandle != NULL) {
376 CDEBUG(D_NET, "Not connecting non-null ephandle found peer 0x%p->%s\n", peer,
377 libcfs_nid2str(peer->gnp_nid));
382 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
383 /* if we are not connecting, fire up a new connection */
384 /* or if we are anything but IDLE DONT start a new connection */
388 CDEBUG(D_NET, "starting connect to %s\n",
389 libcfs_nid2str(peer->gnp_nid));
390 peer->gnp_connecting = GNILND_PEER_CONNECT;
391 kgnilnd_peer_addref(peer); /* extra ref for connd */
393 spin_lock(&dev->gnd_connd_lock);
394 list_add_tail(&peer->gnp_connd_list, &dev->gnd_connd_peers);
395 spin_unlock(&dev->gnd_connd_lock);
397 kgnilnd_schedule_dgram(dev);
398 CDEBUG(D_NETTRACE, "scheduling new connect\n");
403 /* Caller is responsible for deciding if/when to call this */
405 kgnilnd_destroy_conn_ep(kgn_conn_t *conn)
408 gni_ep_handle_t tmp_ep;
410 /* only if we actually initialized it,
411 * then set NULL to tell kgnilnd_destroy_conn to leave it alone */
413 tmp_ep = xchg(&conn->gnc_ephandle, NULL);
414 if (tmp_ep != NULL) {
415 /* we never re-use the EP, so unbind is not needed */
416 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
417 rrc = kgnilnd_ep_destroy(tmp_ep);
419 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
421 /* if this fails, it could hork up kgni smsg retransmit and others
422 * since we could free the SMSG mbox memory, etc. */
423 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d conn 0x%p ep 0x%p\n",
424 rrc, conn, conn->gnc_ephandle);
426 atomic_dec(&conn->gnc_device->gnd_neps);
428 /* clear out count added in kgnilnd_close_conn_locked
429 * conn will have a peer once it hits finish_connect, where it
430 * is the first spot we'll mark it ESTABLISHED as well */
431 if (conn->gnc_peer) {
432 kgnilnd_admin_decref(conn->gnc_peer->gnp_dirty_eps);
435 /* drop ref for EP */
436 kgnilnd_conn_decref(conn);
441 kgnilnd_destroy_conn(kgn_conn_t *conn)
443 LASSERTF(!in_interrupt() &&
444 !conn->gnc_scheduled &&
445 !conn->gnc_in_purgatory &&
446 conn->gnc_ephandle == NULL &&
447 list_empty(&conn->gnc_list) &&
448 list_empty(&conn->gnc_hashlist) &&
449 list_empty(&conn->gnc_schedlist) &&
450 list_empty(&conn->gnc_mdd_list) &&
451 list_empty(&conn->gnc_delaylist) &&
452 conn->gnc_magic == GNILND_CONN_MAGIC,
453 "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d/%d\n",
454 conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
456 !!in_interrupt(), conn->gnc_scheduled,
457 conn->gnc_in_purgatory,
460 list_empty(&conn->gnc_list),
461 list_empty(&conn->gnc_hashlist),
462 list_empty(&conn->gnc_schedlist),
463 list_empty(&conn->gnc_mdd_list),
464 list_empty(&conn->gnc_delaylist));
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 kgnilnd_vfree(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_state != GNILND_PEER_DOWN) {
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 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
675 msleep_interruptible(MSEC_PER_SEC);
678 /* leave on peer->gnp_conns to make sure we don't let the reaper
679 * or others try to unlink this peer until the conn is fully
680 * processed for closing */
682 if (kgnilnd_check_purgatory_conn(conn)) {
683 kgnilnd_add_purgatory_locked(conn, conn->gnc_peer);
686 /* Reset RX timeout to ensure we wait for an incoming CLOSE
687 * for the full timeout. If we get a CLOSE we know the
688 * peer has stopped all RDMA. Otherwise if we wait for
689 * the full timeout we can also be sure all RDMA has stopped. */
690 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
693 /* schedule sending CLOSE - if we are in quiesce, this adds to
694 * gnd_ready_conns and allows us to find it in quiesce processing */
695 kgnilnd_schedule_conn(conn);
701 kgnilnd_close_conn(kgn_conn_t *conn, int error)
703 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
704 /* need to check the state here - this call is racy and we don't
705 * know the state until after the lock is grabbed */
706 if (conn->gnc_state == GNILND_CONN_ESTABLISHED) {
707 kgnilnd_close_conn_locked(conn, error);
709 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
713 kgnilnd_complete_closed_conn(kgn_conn_t *conn)
723 /* Dump log on cksum error - wait until complete phase to let
724 * RX of error happen */
725 if (*kgnilnd_tunables.kgn_checksum_dump &&
726 (conn != NULL && conn->gnc_peer_error == -ENOKEY)) {
727 libcfs_debug_dumplog();
730 /* _CLOSED set in kgnilnd_process_fmaq once we decide to
731 * send the CLOSE or not */
732 LASSERTF(conn->gnc_state == GNILND_CONN_CLOSED,
733 "conn 0x%p->%s with bad state %s\n",
734 conn, conn->gnc_peer ?
735 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
737 kgnilnd_conn_state2str(conn));
739 LASSERT(list_empty(&conn->gnc_hashlist));
740 /* We shouldnt be on the delay list, the conn can
741 * get added to this list during a retransmit, and retransmits
742 * only occur within scheduler threads.
744 LASSERT(list_empty(&conn->gnc_delaylist));
746 /* we've sent the close, start nuking */
747 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
748 kgnilnd_schedule_conn(conn);
750 if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
751 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
752 "done, Attempting to recover conn 0x%p "
753 "scheduled %d function: %s line: %d\n", conn,
754 conn->gnc_scheduled, conn->gnc_sched_caller,
755 conn->gnc_sched_line);
759 /* we don't use lists to track things that we can get out of the
762 /* need to hold locks for tx_list_state, sampling it is too racy:
763 * - the lock actually protects tx != NULL, but we can't take the proper
764 * lock until we check tx_list_state, which would be too late and
765 * we could have the TX change under us.
766 * gnd_rdmaq_lock and gnd_lock and not used together, so taking both
768 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
769 spin_lock(&conn->gnc_device->gnd_lock);
771 for (nrdma = 0; nrdma < GNILND_MAX_MSG_ID; nrdma++) {
772 tx = conn->gnc_tx_ref_table[nrdma];
775 /* only print the first error and if not CLOSE, we often don't see
776 * CQ events for that by the time we get here... and really don't care */
777 if (nlive || tx->tx_msg.gnm_type == GNILND_MSG_CLOSE)
778 tx->tx_state |= GNILND_TX_QUIET_ERROR;
780 GNIDBG_TX(D_NET, tx, "cleaning up on close, nlive %d", nlive);
782 /* don't worry about gnc_lock here as nobody else should be
783 * touching this conn */
784 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
785 list_add_tail(&tx->tx_list, &sinners);
788 spin_unlock(&conn->gnc_device->gnd_lock);
789 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
791 /* nobody should have marked this as needing scheduling after
792 * we called close - so only ref should be us handling it */
793 if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
794 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
795 "done, Attempting to recover conn 0x%p "
796 "scheduled %d function %s line: %d\n", conn,
797 conn->gnc_scheduled, conn->gnc_sched_caller,
798 conn->gnc_sched_line);
800 /* now reset a few to actual counters... */
801 nrdma = atomic_read(&conn->gnc_nlive_rdma);
802 nq_rdma = atomic_read(&conn->gnc_nq_rdma);
804 if (!list_empty(&sinners)) {
805 list_for_each_entry_safe(tx, txn, &sinners, tx_list) {
806 /* clear tx_list to make tx_add_list_locked happy */
807 list_del_init(&tx->tx_list);
808 /* The error codes determine if we hold onto the MDD */
809 kgnilnd_tx_done(tx, conn->gnc_error);
813 logmsg = (nlive + nrdma + nq_rdma);
816 int level = conn->gnc_peer->gnp_state == GNILND_PEER_UP ?
818 CDEBUG(level, "Closed conn 0x%p->%s (errno %d,"
819 " peer errno %d): canceled %d TX, %d/%d RDMA\n",
820 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
821 conn->gnc_error, conn->gnc_peer_error,
822 nlive, nq_rdma, nrdma);
825 kgnilnd_destroy_conn_ep(conn);
827 /* Bug 765042 - race this with completing a new conn to same peer - we need
828 * finish_connect to detach purgatory before we can do it ourselves here */
829 CFS_RACE(CFS_FAIL_GNI_FINISH_PURG);
831 /* now it is safe to remove from peer list - anyone looking at
832 * gnp_conns now is free to unlink if not on purgatory */
833 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
835 conn->gnc_state = GNILND_CONN_DONE;
837 /* Decrement counter if we are marked by del_conn_or_peers for closing
839 if (conn->gnc_needs_closing)
840 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_conns);
842 /* Remove from peer's list of valid connections if its not in purgatory */
843 if (!conn->gnc_in_purgatory) {
844 list_del_init(&conn->gnc_list);
845 /* Lose peers reference on the conn */
846 kgnilnd_conn_decref(conn);
849 /* NB - only unlinking if we set pending in del_peer_locked from admin or
851 if (kgnilnd_peer_active(conn->gnc_peer) &&
852 conn->gnc_peer->gnp_pending_unlink &&
853 kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
854 kgnilnd_unlink_peer_locked(conn->gnc_peer);
857 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
859 /* I'm telling Mommy! - use peer_error if they initiated close */
860 kgnilnd_peer_notify(conn->gnc_peer,
861 conn->gnc_error == -ECONNRESET ?
862 conn->gnc_peer_error : conn->gnc_error, 0);
868 kgnilnd_set_conn_params(kgn_dgram_t *dgram)
870 kgn_conn_t *conn = dgram->gndg_conn;
871 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
872 kgn_gniparams_t *rem_param = &connreq->gncr_gnparams;
875 gni_smsg_attr_t *remote = &connreq->gncr_gnparams.gnpr_smsg_attr;
877 /* set timeout vals in conn early so we can use them for the NAK */
879 /* use max of the requested and our timeout, peer will do the same */
880 conn->gnc_timeout = MAX(conn->gnc_timeout, connreq->gncr_timeout);
882 /* only ep_bind really mucks around with the CQ */
883 /* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check
884 * is necessary as you can only bind an ep once and we must make sure we dont bind when already bound.
886 if (connreq->gncr_dstnid != LNET_NID_ANY && dgram->gndg_conn_out.gncr_dstnid != connreq->gncr_srcnid) {
887 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
888 rrc = kgnilnd_ep_bind(conn->gnc_ephandle,
889 connreq->gncr_gnparams.gnpr_host_id,
891 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
892 if (rrc != GNI_RC_SUCCESS) {
898 rrc = kgnilnd_ep_set_eventdata(conn->gnc_ephandle, conn->gnc_cqid,
899 connreq->gncr_gnparams.gnpr_cqid);
900 if (rrc != GNI_RC_SUCCESS) {
905 /* Initialize SMSG */
906 rrc = kgnilnd_smsg_init(conn->gnc_ephandle, &conn->gnpr_smsg_attr,
907 &connreq->gncr_gnparams.gnpr_smsg_attr);
908 if (unlikely(rrc == GNI_RC_INVALID_PARAM)) {
909 gni_smsg_attr_t *local = &conn->gnpr_smsg_attr;
910 /* help folks figure out if there is a tunable off, etc. */
911 LCONSOLE_ERROR("SMSG attribute mismatch. Data from local/remote:"
912 " type %d/%d msg_maxsize %u/%u"
913 " mbox_maxcredit %u/%u. Please check kgni"
914 " logs for further data\n",
915 local->msg_type, remote->msg_type,
916 local->msg_maxsize, remote->msg_maxsize,
917 local->mbox_maxcredit, remote->mbox_maxcredit);
919 if (rrc != GNI_RC_SUCCESS) {
924 /* log this for help in debuggin SMSG buffer re-use */
925 CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired"
926 " local cqid %u SMSG %p->%u hndl %#llx.%#llx"
927 " remote cqid %u SMSG %p->%u hndl %#llx.%#llx\n",
928 conn, libcfs_nid2str(connreq->gncr_srcnid),
929 libcfs_nid2str(connreq->gncr_dstnid),
930 &conn->gnpr_smsg_attr,
932 conn->gnpr_smsg_attr.msg_buffer,
933 conn->gnpr_smsg_attr.mbox_offset,
934 conn->gnpr_smsg_attr.mem_hndl.qword1,
935 conn->gnpr_smsg_attr.mem_hndl.qword2,
936 rem_param->gnpr_cqid,
937 rem_param->gnpr_smsg_attr.msg_buffer,
938 rem_param->gnpr_smsg_attr.mbox_offset,
939 rem_param->gnpr_smsg_attr.mem_hndl.qword1,
940 rem_param->gnpr_smsg_attr.mem_hndl.qword2);
942 conn->gnc_peerstamp = connreq->gncr_peerstamp;
943 conn->gnc_peer_connstamp = connreq->gncr_connstamp;
944 conn->remote_mbox_addr = (void *)((char *)remote->msg_buffer + remote->mbox_offset);
946 /* We update the reaper timeout once we have a valid conn and timeout */
947 kgnilnd_update_reaper_timeout(GNILND_TO2KA(conn->gnc_timeout));
952 rrc = kgnilnd_ep_unbind(conn->gnc_ephandle);
953 /* not sure I can just let this fly */
954 LASSERTF(rrc == GNI_RC_SUCCESS,
955 "bad rc from gni_ep_unbind trying to cleanup: %d\n", rrc);
958 LASSERTF(rc != 0, "SOFTWARE BUG: rc == 0\n");
959 CERROR("Error setting connection params from %s: %d\n",
960 libcfs_nid2str(connreq->gncr_srcnid), rc);
964 /* needs down_read on kgn_net_rw_sem held from before this call until
965 * after the write_lock on kgn_peer_conn_lock - this ensures we stay sane
966 * with kgnilnd_shutdown - it'll get the sem and set shutdown, then get the
967 * kgn_peer_conn_lock to start del_peer'ing. If we hold the sem until after
968 * kgn_peer_conn_lock is held, we guarantee that nobody calls
969 * kgnilnd_add_peer_locked without checking gnn_shutdown */
971 kgnilnd_create_peer_safe(kgn_peer_t **peerp,
979 LASSERT(nid != LNET_NID_ANY);
981 /* We dont pass the net around in the dgram anymore so here is where we find it
982 * this will work unless its in shutdown or the nid has a net that is invalid.
983 * Either way error code needs to be returned in that case.
985 * If the net passed in is not NULL then we can use it, this alleviates looking it
986 * when the calling function has access to the data.
989 rc = kgnilnd_find_net(nid, &net);
993 /* find net adds a reference on the net if we are not using
994 * it we must do it manually so the net references are
995 * correct when tearing down the net
997 kgnilnd_net_addref(net);
1000 LIBCFS_ALLOC(peer, sizeof(*peer));
1002 kgnilnd_net_decref(net);
1005 peer->gnp_nid = nid;
1006 peer->gnp_state = node_state;
1008 /* translate from nid to nic addr & store */
1009 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
1011 kgnilnd_net_decref(net);
1012 LIBCFS_FREE(peer, sizeof(*peer));
1015 CDEBUG(D_NET, "peer 0x%p->%s -> NIC 0x%x\n", peer,
1016 libcfs_nid2str(nid), peer->gnp_host_id);
1018 atomic_set(&peer->gnp_refcount, 1); /* 1 ref for caller */
1019 atomic_set(&peer->gnp_dirty_eps, 0);
1021 INIT_LIST_HEAD(&peer->gnp_list);
1022 INIT_LIST_HEAD(&peer->gnp_connd_list);
1023 INIT_LIST_HEAD(&peer->gnp_conns);
1024 INIT_LIST_HEAD(&peer->gnp_tx_queue);
1026 /* the first reconnect should happen immediately, so we leave
1027 * gnp_reconnect_interval set to 0 */
1029 LASSERTF(net != NULL, "peer 0x%p->%s with NULL net\n",
1030 peer, libcfs_nid2str(nid));
1032 /* must have kgn_net_rw_sem held for this... */
1033 if (net->gnn_shutdown) {
1034 /* shutdown has started already */
1035 kgnilnd_net_decref(net);
1036 LIBCFS_FREE(peer, sizeof(*peer));
1040 peer->gnp_net = net;
1042 atomic_inc(&kgnilnd_data.kgn_npeers);
1049 kgnilnd_destroy_peer(kgn_peer_t *peer)
1051 CDEBUG(D_NET, "peer %s %p deleted\n",
1052 libcfs_nid2str(peer->gnp_nid), peer);
1053 LASSERTF(atomic_read(&peer->gnp_refcount) == 0,
1054 "peer 0x%p->%s refs %d\n",
1055 peer, libcfs_nid2str(peer->gnp_nid),
1056 atomic_read(&peer->gnp_refcount));
1057 LASSERTF(atomic_read(&peer->gnp_dirty_eps) == 0,
1058 "peer 0x%p->%s dirty eps %d\n",
1059 peer, libcfs_nid2str(peer->gnp_nid),
1060 atomic_read(&peer->gnp_dirty_eps));
1061 LASSERTF(peer->gnp_net != NULL, "peer %p (%s) with NULL net\n",
1062 peer, libcfs_nid2str(peer->gnp_nid));
1063 LASSERTF(!kgnilnd_peer_active(peer),
1065 peer, libcfs_nid2str(peer->gnp_nid));
1066 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE || peer->gnp_connecting == GNILND_PEER_KILL,
1067 "peer 0x%p->%s, connecting %d\n",
1068 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1069 LASSERTF(list_empty(&peer->gnp_conns),
1071 peer, libcfs_nid2str(peer->gnp_nid));
1072 LASSERTF(list_empty(&peer->gnp_tx_queue),
1074 peer, libcfs_nid2str(peer->gnp_nid));
1075 LASSERTF(list_empty(&peer->gnp_connd_list),
1077 peer, libcfs_nid2str(peer->gnp_nid));
1079 /* NB a peer's connections keep a reference on their peer until
1080 * they are destroyed, so we can be assured that _all_ state to do
1081 * with this peer has been cleaned up when its refcount drops to
1084 atomic_dec(&kgnilnd_data.kgn_npeers);
1085 kgnilnd_net_decref(peer->gnp_net);
1087 LIBCFS_FREE(peer, sizeof(*peer));
1090 /* the conn might not have made it all the way through to a connected
1091 * state - but we need to purgatory any conn that a remote peer might
1092 * have seen through a posted dgram as well */
1094 kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer)
1096 kgn_mbox_info_t *mbox = NULL;
1099 /* NB - the caller should own conn by removing him from the
1100 * scheduler thread when finishing the close */
1102 LASSERTF(peer != NULL, "conn %p with NULL peer\n", conn);
1104 /* If this is still true, need to add the calls to unlink back in and
1105 * figure out how to close the hole on loopback conns */
1106 LASSERTF(kgnilnd_peer_active(peer), "can't use inactive peer %s (%p)"
1107 " we'll never recover the resources\n",
1108 libcfs_nid2str(peer->gnp_nid), peer);
1110 CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
1113 LASSERTF(conn->gnc_in_purgatory == 0,
1114 "Conn already in purgatory\n");
1115 conn->gnc_in_purgatory = 1;
1117 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1118 mbox->mbx_prev_purg_nid = peer->gnp_nid;
1119 mbox->mbx_add_purgatory = jiffies;
1120 kgnilnd_release_mbox(conn, 1);
1122 LASSERTF(list_empty(&conn->gnc_mdd_list),
1123 "conn 0x%p->%s with active purgatory hold MDD %d\n",
1124 conn, libcfs_nid2str(peer->gnp_nid),
1125 kgnilnd_count_list(&conn->gnc_mdd_list));
1130 /* Instead of detaching everything from purgatory here we just mark the conn as needing
1131 * detach, when the reaper checks the conn the next time it will detach it.
1132 * Calling function requires write_lock held on kgn_peer_conn_lock
1135 kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer) {
1138 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
1139 if (conn->gnc_in_purgatory && !conn->gnc_needs_detach) {
1140 conn->gnc_needs_detach = 1;
1141 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_detach);
1146 /* Calling function needs a write_lock held on kgn_peer_conn_lock */
1148 kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list)
1150 kgn_mbox_info_t *mbox = NULL;
1152 /* if needed, add the conn purgatory data to the list passed in */
1153 if (conn->gnc_in_purgatory) {
1154 CDEBUG(D_NET, "peer %p->%s purg_conn %p@%s mdd_list #tx %d\n",
1155 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1156 conn, kgnilnd_conn_state2str(conn),
1157 kgnilnd_count_list(&conn->gnc_mdd_list));
1159 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1160 mbox->mbx_detach_of_purgatory = jiffies;
1162 /* conn->gnc_list is the entry point on peer->gnp_conns, so detaching it
1163 * here removes it from the list of 'valid' peer connections.
1164 * We put the current conn onto a list of conns to call kgnilnd_release_purgatory_locked()
1165 * and as such the caller of kgnilnd_detach_purgatory_locked() now owns that conn, since its not
1166 * on the peer's conn_list anymore.
1169 list_del_init(&conn->gnc_list);
1171 /* NB - only unlinking if we set pending in del_peer_locked from admin or
1173 if (kgnilnd_peer_active(conn->gnc_peer) &&
1174 conn->gnc_peer->gnp_pending_unlink &&
1175 kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
1176 kgnilnd_unlink_peer_locked(conn->gnc_peer);
1178 /* The reaper will not call detach unless the conn is fully through kgnilnd_complete_closed_conn.
1179 * If the conn is not in a DONE state somehow we are attempting to detach even though
1180 * the conn has not been fully cleaned up. If we detach while the conn is still closing
1181 * we will end up with an orphaned connection that has valid ep_handle, that is not on a
1185 LASSERTF(conn->gnc_state == GNILND_CONN_DONE, "Conn in invalid state %p@%s \n",
1186 conn, kgnilnd_conn_state2str(conn));
1188 /* move from peer to the delayed release list */
1189 list_add_tail(&conn->gnc_list, conn_list);
1194 kgnilnd_release_purgatory_list(struct list_head *conn_list)
1197 kgn_conn_t *conn, *connN;
1198 kgn_mdd_purgatory_t *gmp, *gmpN;
1200 list_for_each_entry_safe(conn, connN, conn_list, gnc_list) {
1201 dev = conn->gnc_device;
1203 kgnilnd_release_mbox(conn, -1);
1204 conn->gnc_in_purgatory = 0;
1206 list_del_init(&conn->gnc_list);
1208 /* gnc_needs_detach is set in kgnilnd_del_conn_or_peer. It is used to keep track
1209 * of conns that have been marked for detach by kgnilnd_del_conn_or_peer.
1210 * The function uses kgn_npending_detach to verify the conn has
1211 * actually been detached.
1214 if (conn->gnc_needs_detach)
1215 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_detach);
1217 /* if this guy is really dead (we are doing release from reaper),
1218 * make sure we tell LNet - if this is from other context,
1219 * the checks in the function will prevent an errant
1221 kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0);
1223 list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
1226 "dev %p releasing held mdd %#llx.%#llx\n",
1227 conn->gnc_device, gmp->gmp_map_key.qword1,
1228 gmp->gmp_map_key.qword2);
1230 atomic_dec(&dev->gnd_n_mdd_held);
1231 kgnilnd_mem_mdd_release(conn->gnc_device->gnd_handle,
1233 /* ignoring the return code - if kgni/ghal can't find it
1234 * it must be released already */
1236 list_del_init(&gmp->gmp_list);
1237 LIBCFS_FREE(gmp, sizeof(*gmp));
1239 /* lose conn ref for purgatory */
1240 kgnilnd_conn_decref(conn);
1244 /* needs write_lock on kgnilnd_data.kgn_peer_conn_lock held */
1246 kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer)
1250 current_to = peer->gnp_reconnect_interval;
1252 /* we'll try to reconnect fast the first time, then back-off */
1253 if (current_to == 0) {
1254 peer->gnp_reconnect_time = jiffies - 1;
1255 current_to = *kgnilnd_tunables.kgn_min_reconnect_interval;
1257 peer->gnp_reconnect_time = jiffies + cfs_time_seconds(current_to);
1258 /* add 50% of min timeout & retry */
1259 current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2;
1262 current_to = MIN(current_to,
1263 *kgnilnd_tunables.kgn_max_reconnect_interval);
1265 peer->gnp_reconnect_interval = current_to;
1266 CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n",
1267 libcfs_nid2str(peer->gnp_nid), peer->gnp_reconnect_time,
1268 peer->gnp_reconnect_interval);
1271 /* needs kgnilnd_data.kgn_peer_conn_lock held */
1273 kgnilnd_find_peer_locked(lnet_nid_t nid)
1275 struct list_head *peer_list = kgnilnd_nid2peerlist(nid);
1278 /* Chopping nid down to only NIDADDR using LNET_NIDADDR so we only
1279 * have a single peer per device instead of a peer per nid/net combo.
1282 list_for_each_entry(peer, peer_list, gnp_list) {
1283 if (LNET_NIDADDR(nid) != LNET_NIDADDR(peer->gnp_nid))
1286 CDEBUG(D_NET, "got peer [%p] -> %s c %d (%d)\n",
1287 peer, libcfs_nid2str(nid),
1288 peer->gnp_connecting,
1289 atomic_read(&peer->gnp_refcount));
1295 /* need write_lock on kgn_peer_conn_lock */
1297 kgnilnd_unlink_peer_locked(kgn_peer_t *peer)
1299 LASSERTF(list_empty(&peer->gnp_conns),
1301 peer, libcfs_nid2str(peer->gnp_nid));
1302 LASSERTF(list_empty(&peer->gnp_tx_queue),
1304 peer, libcfs_nid2str(peer->gnp_nid));
1305 LASSERTF(kgnilnd_peer_active(peer),
1307 peer, libcfs_nid2str(peer->gnp_nid));
1308 CDEBUG(D_NET, "unlinking peer 0x%p->%s\n",
1309 peer, libcfs_nid2str(peer->gnp_nid));
1311 list_del_init(&peer->gnp_list);
1312 kgnilnd_data.kgn_peer_version++;
1313 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1314 /* lose peerlist's ref */
1315 kgnilnd_peer_decref(peer);
1319 kgnilnd_get_peer_info(int index,
1320 kgn_peer_t **found_peer,
1321 lnet_nid_t *id, __u32 *nic_addr,
1322 int *refcount, int *connecting)
1324 struct list_head *ptmp;
1329 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1331 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1333 list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
1334 peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1339 CDEBUG(D_NET, "found peer %p (%s) at index %d\n",
1340 peer, libcfs_nid2str(peer->gnp_nid), index);
1343 *id = peer->gnp_nid;
1344 *nic_addr = peer->gnp_host_id;
1345 *refcount = atomic_read(&peer->gnp_refcount);
1346 *connecting = peer->gnp_connecting;
1353 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1355 CDEBUG(D_NET, "no gni peer at index %d\n", index);
1359 /* requires write_lock on kgn_peer_conn_lock held */
1361 kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp)
1363 kgn_peer_t *peer, *peer2;
1365 LASSERTF(new_stub_peer != NULL, "bad stub peer for nid %s\n",
1366 libcfs_nid2str(nid));
1368 peer2 = kgnilnd_find_peer_locked(nid);
1369 if (peer2 != NULL) {
1370 /* A peer was created during the lock transition, so drop
1371 * the new one we created */
1372 kgnilnd_peer_decref(new_stub_peer);
1375 peer = new_stub_peer;
1376 /* peer table takes existing ref on peer */
1378 LASSERTF(!kgnilnd_peer_active(peer),
1379 "peer 0x%p->%s already in peer table\n",
1380 peer, libcfs_nid2str(peer->gnp_nid));
1381 list_add_tail(&peer->gnp_list,
1382 kgnilnd_nid2peerlist(nid));
1383 kgnilnd_data.kgn_peer_version++;
1386 LASSERTF(peer->gnp_net != NULL, "peer 0x%p->%s with NULL net\n",
1387 peer, libcfs_nid2str(peer->gnp_nid));
1392 kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp)
1399 if (nid == LNET_NID_ANY)
1402 node_state = kgnilnd_get_node_state(LNET_NIDADDR(nid));
1404 /* NB - this will not block during normal operations -
1405 * the only writer of this is in the startup/shutdown path. */
1406 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1411 rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
1413 up_read(&kgnilnd_data.kgn_net_rw_sem);
1417 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1418 up_read(&kgnilnd_data.kgn_net_rw_sem);
1420 kgnilnd_add_peer_locked(nid, peer, peerp);
1422 CDEBUG(D_NET, "peer 0x%p->%s connecting %d\n",
1423 peerp, libcfs_nid2str((*peerp)->gnp_nid),
1424 (*peerp)->gnp_connecting);
1426 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1430 /* needs write_lock on kgn_peer_conn_lock */
1432 kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies)
1436 /* we do care about state of gnp_connecting - we could be between
1437 * reconnect attempts, so try to find the dgram and cancel the TX
1438 * anyways. If we are in the process of posting DONT do anything;
1439 * once it fails or succeeds we can nuke the connect attempt.
1440 * We have no idea where in kgnilnd_post_dgram we are so we cant
1441 * attempt to cancel until the function is done.
1444 /* make sure peer isn't in process of connecting or waiting for connect*/
1445 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1446 if (!(list_empty(&peer->gnp_connd_list))) {
1447 list_del_init(&peer->gnp_connd_list);
1448 /* remove connd ref */
1449 kgnilnd_peer_decref(peer);
1451 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1453 if (peer->gnp_connecting == GNILND_PEER_POSTING || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1454 peer->gnp_connecting = GNILND_PEER_NEEDS_DEATH;
1455 /* We are in process of posting right now the xchg set it up for us to
1456 * cancel the connect so we are finished for now */
1458 /* no need for exchange we have the peer lock and its ready for us to nuke */
1459 LASSERTF(peer->gnp_connecting != GNILND_PEER_POSTING,
1460 "Peer in invalid state 0x%p->%s, connecting %d\n",
1461 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1462 peer->gnp_connecting = GNILND_PEER_IDLE;
1463 set_mb(peer->gnp_last_dgram_errno, -ETIMEDOUT);
1464 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
1468 /* The least we can do is nuke the tx's no matter what.... */
1469 list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1470 kgnilnd_tx_del_state_locked(tx, peer, NULL,
1472 list_add_tail(&tx->tx_list, zombies);
1476 /* needs write_lock on kgn_peer_conn_lock */
1478 kgnilnd_del_peer_locked(kgn_peer_t *peer, int error)
1480 /* this peer could be passive and only held for purgatory,
1481 * take a ref to ensure it doesn't disappear in this function */
1482 kgnilnd_peer_addref(peer);
1484 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1486 /* if purgatory release cleared it out, don't try again */
1487 if (kgnilnd_peer_active(peer)) {
1488 /* always do this to allow kgnilnd_start_connect and
1489 * kgnilnd_finish_connect to catch this before they
1490 * wrap up their operations */
1491 if (kgnilnd_can_unlink_peer_locked(peer)) {
1492 /* already released purgatory, so only active
1494 kgnilnd_unlink_peer_locked(peer);
1496 kgnilnd_close_peer_conns_locked(peer, error);
1497 /* peer unlinks itself when last conn is closed */
1501 /* we are done, release back to the wild */
1502 kgnilnd_peer_decref(peer);
1506 kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
1510 LIST_HEAD (zombies);
1511 struct list_head *ptmp, *pnxt;
1518 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1520 if (nid != LNET_NID_ANY)
1521 lo = hi = kgnilnd_nid2peerlist(nid) - kgnilnd_data.kgn_peers;
1524 hi = *kgnilnd_tunables.kgn_peer_hash_size - 1;
1525 /* wildcards always succeed */
1529 for (i = lo; i <= hi; i++) {
1530 list_for_each_safe(ptmp, pnxt, &kgnilnd_data.kgn_peers[i]) {
1531 peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1533 LASSERTF(peer->gnp_net != NULL,
1534 "peer %p (%s) with NULL net\n",
1535 peer, libcfs_nid2str(peer->gnp_nid));
1537 if (net != NULL && peer->gnp_net != net)
1540 if (!(nid == LNET_NID_ANY || LNET_NIDADDR(peer->gnp_nid) == LNET_NIDADDR(nid)))
1543 /* In both cases, we want to stop any in-flight
1544 * connect attempts */
1545 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1548 case GNILND_DEL_CONN:
1549 kgnilnd_close_peer_conns_locked(peer, error);
1551 case GNILND_DEL_PEER:
1552 peer->gnp_pending_unlink = 1;
1553 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_unlink);
1554 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1555 kgnilnd_del_peer_locked(peer, error);
1557 case GNILND_CLEAR_PURGATORY:
1558 /* Mark everything ready for detach reaper will cleanup
1559 * once we release the kgn_peer_conn_lock
1561 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1562 peer->gnp_last_errno = -EISCONN;
1563 /* clear reconnect so he can reconnect soon */
1564 peer->gnp_reconnect_time = 0;
1565 peer->gnp_reconnect_interval = 0;
1568 CERROR("bad command %d\n", command);
1571 /* we matched something */
1576 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1579 kgnilnd_txlist_done(&zombies, error);
1581 /* This function does not return until the commands it initiated have completed,
1582 * since they have to work there way through the other threads. In the case of shutdown
1583 * threads are not woken up until after this call is initiated so we cannot wait, we just
1584 * need to return. The same applies for stack reset we shouldnt wait as the reset thread
1588 CFS_RACE(CFS_FAIL_GNI_RACE_RESET);
1590 if (error == -ENOTRECOVERABLE || error == -ESHUTDOWN) {
1595 while (atomic_read(&kgnilnd_data.kgn_npending_conns) ||
1596 atomic_read(&kgnilnd_data.kgn_npending_detach) ||
1597 atomic_read(&kgnilnd_data.kgn_npending_unlink)) {
1599 set_current_state(TASK_UNINTERRUPTIBLE);
1600 schedule_timeout(cfs_time_seconds(1));
1603 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, "Waiting on %d peers %d closes %d detaches\n",
1604 atomic_read(&kgnilnd_data.kgn_npending_unlink),
1605 atomic_read(&kgnilnd_data.kgn_npending_conns),
1606 atomic_read(&kgnilnd_data.kgn_npending_detach));
1613 kgnilnd_get_conn_by_idx(int index)
1616 struct list_head *ptmp;
1618 struct list_head *ctmp;
1622 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1623 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1624 list_for_each(ptmp, &kgnilnd_data.kgn_peers[i]) {
1626 peer = list_entry(ptmp, kgn_peer_t, gnp_list);
1628 list_for_each(ctmp, &peer->gnp_conns) {
1629 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1631 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1637 CDEBUG(D_NET, "++conn[%p] -> %s (%d)\n", conn,
1638 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1639 atomic_read(&conn->gnc_refcount));
1640 kgnilnd_conn_addref(conn);
1641 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1645 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1652 kgnilnd_get_conn_info(kgn_peer_t *peer,
1653 int *device_id, __u64 *peerstamp,
1654 int *tx_seq, int *rx_seq,
1655 int *fmaq_len, int *nfma, int *nrdma)
1660 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1662 conn = kgnilnd_find_conn_locked(peer);
1668 *device_id = conn->gnc_device->gnd_host_id;
1669 *peerstamp = conn->gnc_peerstamp;
1670 *tx_seq = atomic_read(&conn->gnc_tx_seq);
1671 *rx_seq = atomic_read(&conn->gnc_rx_seq);
1672 *fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq);
1673 *nfma = atomic_read(&conn->gnc_nlive_fma);
1674 *nrdma = atomic_read(&conn->gnc_nlive_rdma);
1676 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1680 /* needs write_lock on kgn_peer_conn_lock */
1682 kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why)
1685 struct list_head *ctmp, *cnxt;
1688 list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
1689 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1691 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1695 /* we mark gnc_needs closing and increment kgn_npending_conns so that
1696 * kgnilnd_del_conn_or_peer can wait on the other threads closing
1697 * and cleaning up the connection.
1699 if (!conn->gnc_needs_closing) {
1700 conn->gnc_needs_closing = 1;
1701 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_conns);
1703 kgnilnd_close_conn_locked(conn, why);
1709 kgnilnd_report_node_state(lnet_nid_t nid, int down)
1712 kgn_peer_t *peer, *new_peer;
1715 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1716 peer = kgnilnd_find_peer_locked(nid);
1723 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1725 /* Don't add a peer for node up events */
1726 if (down == GNILND_PEER_UP)
1729 /* find any valid net - we don't care which one... */
1730 down_read(&kgnilnd_data.kgn_net_rw_sem);
1731 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
1732 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i],
1742 up_read(&kgnilnd_data.kgn_net_rw_sem);
1745 CNETERR("Could not find a net for nid %lld\n", nid);
1749 /* The nid passed in does not yet contain the net portion.
1750 * Let's build it up now
1752 nid = LNET_MKNID(LNET_NIDNET(net->gnn_ni->ni_nid), nid);
1753 rc = kgnilnd_add_peer(net, nid, &new_peer);
1756 CNETERR("Could not add peer for nid %lld, rc %d\n",
1761 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1762 peer = kgnilnd_find_peer_locked(nid);
1765 CNETERR("Could not find peer for nid %lld\n", nid);
1766 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1771 peer->gnp_state = down;
1773 if (down == GNILND_PEER_DOWN) {
1776 peer->gnp_down_event_time = jiffies;
1777 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1778 conn = kgnilnd_find_conn_locked(peer);
1781 kgnilnd_close_conn_locked(conn, -ENETRESET);
1784 peer->gnp_up_event_time = jiffies;
1787 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1789 if (down == GNILND_PEER_DOWN) {
1790 /* using ENETRESET so we don't get messages from
1793 kgnilnd_txlist_done(&zombies, -ENETRESET);
1794 kgnilnd_peer_notify(peer, -ECONNRESET, 0);
1795 LCONSOLE_INFO("Received down event for nid %d\n",
1803 kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1805 struct libcfs_ioctl_data *data = arg;
1806 kgn_net_t *net = ni->ni_data;
1809 LASSERT(ni == net->gnn_ni);
1812 case IOC_LIBCFS_GET_PEER: {
1814 kgn_peer_t *peer = NULL;
1816 __u64 peerstamp = 0;
1817 int peer_refcount = 0, peer_connecting = 0;
1819 int tx_seq = 0, rx_seq = 0;
1820 int fmaq_len = 0, nfma = 0, nrdma = 0;
1822 rc = kgnilnd_get_peer_info(data->ioc_count, &peer,
1823 &nid, &nic_addr, &peer_refcount,
1829 /* LNET_MKNID is used to mask from lnet the multiplexing/demultiplexing of connections and peers
1830 * LNET assumes a conn and peer per net, the LNET_MKNID/LNET_NIDADDR allows us to let Lnet see what it
1831 * wants to see instead of the underlying network that is being used to send the data
1833 data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(nid));
1834 data->ioc_flags = peer_connecting;
1835 data->ioc_count = peer_refcount;
1837 rc = kgnilnd_get_conn_info(peer, &device_id, &peerstamp,
1838 &tx_seq, &rx_seq, &fmaq_len,
1841 /* This is allowable - a persistent peer could not
1842 * have a connection */
1844 /* flag to indicate we are not connected -
1845 * need to print as such */
1846 data->ioc_flags |= (1<<16);
1850 data->ioc_net = device_id;
1851 data->ioc_u64[0] = peerstamp;
1852 data->ioc_u32[0] = fmaq_len;
1853 data->ioc_u32[1] = nfma;
1854 data->ioc_u32[2] = tx_seq;
1855 data->ioc_u32[3] = rx_seq;
1856 data->ioc_u32[4] = nrdma;
1860 case IOC_LIBCFS_ADD_PEER: {
1861 /* just dummy value to allow using common interface */
1863 rc = kgnilnd_add_peer(net, data->ioc_nid, &peer);
1866 case IOC_LIBCFS_DEL_PEER: {
1867 /* NULL is passed in so it affects all peers in existence without regard to network
1868 * as the peer may not exist on the network LNET believes it to be on.
1870 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1871 GNILND_DEL_PEER, -EUCLEAN);
1874 case IOC_LIBCFS_GET_CONN: {
1875 kgn_conn_t *conn = kgnilnd_get_conn_by_idx(data->ioc_count);
1881 /* LNET_MKNID is used to build the correct address based on what LNET wants to see instead of
1882 * the generic connection that is used to send the data
1884 data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(conn->gnc_peer->gnp_nid));
1885 data->ioc_u32[0] = conn->gnc_device->gnd_id;
1886 kgnilnd_conn_decref(conn);
1890 case IOC_LIBCFS_CLOSE_CONNECTION: {
1891 /* use error = -ENETRESET to indicate it was lctl disconnect */
1892 /* NULL is passed in so it affects all the nets as the connection is virtual
1893 * and may not exist on the network LNET believes it to be on.
1895 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1896 GNILND_DEL_CONN, -ENETRESET);
1899 case IOC_LIBCFS_PUSH_CONNECTION: {
1900 /* we use this to flush purgatory */
1901 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1902 GNILND_CLEAR_PURGATORY, -EUCLEAN);
1905 case IOC_LIBCFS_REGISTER_MYNID: {
1906 /* Ignore if this is a noop */
1907 if (data->ioc_nid == ni->ni_nid) {
1910 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1911 libcfs_nid2str(data->ioc_nid),
1912 libcfs_nid2str(ni->ni_nid));
1923 kgnilnd_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1925 kgn_net_t *net = ni->ni_data;
1927 kgn_peer_t *peer = NULL;
1928 kgn_conn_t *conn = NULL;
1929 struct lnet_process_id id = {
1931 .pid = LNET_PID_LUSTRE,
1935 /* I expect to find him, so only take a read lock */
1936 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1937 peer = kgnilnd_find_peer_locked(nid);
1939 /* LIE if in a quiesce - we will update the timeouts after,
1940 * but we don't want sends failing during it */
1941 if (kgnilnd_data.kgn_quiesce_trigger) {
1942 *when = ktime_get_seconds();
1943 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1947 /* Update to best guess, might refine on later checks */
1948 *when = peer->gnp_last_alive;
1950 /* we have a peer, how about a conn? */
1951 conn = kgnilnd_find_conn_locked(peer);
1954 /* if there is no conn, check peer last errno to see if clean disconnect
1955 * - if it was, we lie to LNet because we believe a TX would complete
1957 if (kgnilnd_conn_clean_errno(peer->gnp_last_errno)) {
1958 *when = ktime_get_seconds();
1960 /* we still want to fire a TX and new conn in this case */
1962 /* gnp_last_alive is valid, run for the hills */
1963 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1967 /* if we get here, either we have no peer or no conn for him, so fire off
1968 * new TX to trigger conn setup */
1969 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1971 /* if we couldn't find him, we'll fire up a TX and get connected -
1972 * if we don't do this, after ni_peer_timeout, LNet will declare him dead.
1973 * So really we treat kgnilnd_query as a bit of a 'connect now' type
1974 * event because it'll only do this when it wants to send
1976 * Use a real TX for this to get the proper gnp_tx_queue behavior, etc
1977 * normally we'd use kgnilnd_send_ctlmsg for this, but we don't really
1978 * care that this goes out quickly since we already know we need a new conn
1980 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
1983 tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, ni->ni_nid);
1985 kgnilnd_launch_tx(tx, net, &id);
1988 CDEBUG(D_NETTRACE, "peer 0x%p->%s when %lld\n", peer,
1989 libcfs_nid2str(nid), *when);
1994 kgnilnd_dev_init(kgn_device_t *dev)
1998 unsigned int cq_size;
2001 /* size of these CQs should be able to accommodate the outgoing
2002 * RDMA and SMSG transactions. Since we really don't know what we
2003 * really need here, we'll take credits * 2 * 3 to allow a bunch.
2004 * We need to dig into this more with the performance work. */
2005 cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
2007 rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
2008 *kgnilnd_tunables.kgn_pkey, 0,
2010 if (rrc != GNI_RC_SUCCESS) {
2011 CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
2012 GOTO(failed, rc = -ENODEV);
2015 rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
2016 &dev->gnd_host_id, &dev->gnd_handle);
2017 if (rrc != GNI_RC_SUCCESS) {
2018 CERROR("Can't attach CDM to device %d (%d)\n",
2020 GOTO(failed, rc = -ENODEV);
2023 /* a bit gross, but not much we can do - Aries Sim doesn't have
2024 * hardcoded NIC/NID that we can use */
2025 rc = kgnilnd_setup_nic_translation(dev->gnd_host_id);
2027 GOTO(failed, rc = -ENODEV);
2029 /* only dev 0 gets the errors - no need to reset the stack twice
2030 * - this works because we have a single PTAG, if we had more
2031 * then we'd need to have multiple handlers */
2032 if (dev->gnd_id == 0) {
2033 rrc = kgnilnd_subscribe_errors(dev->gnd_handle,
2034 GNI_ERRMASK_CRITICAL |
2035 GNI_ERRMASK_UNKNOWN_TRANSACTION,
2036 0, NULL, kgnilnd_critical_error,
2037 &dev->gnd_err_handle);
2038 if (rrc != GNI_RC_SUCCESS) {
2039 CERROR("Can't subscribe for errors on device %d: rc %d\n",
2041 GOTO(failed, rc = -ENODEV);
2044 rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
2045 kgnilnd_quiesce_end_callback);
2046 if (rc != GNI_RC_SUCCESS) {
2047 CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
2049 GOTO(failed, rc = -ENODEV);
2053 rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
2055 /* log messages during startup */
2056 if (kgnilnd_data.kgn_init < GNILND_INIT_ALL) {
2057 CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
2058 dev->gnd_host_id, rc);
2060 GOTO(failed, rc = -ESRCH);
2062 CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);
2064 rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits,
2065 0, kgnilnd_device_callback,
2066 dev->gnd_id, &dev->gnd_snd_rdma_cqh);
2067 if (rrc != GNI_RC_SUCCESS) {
2068 CERROR("Can't create rdma send cq size %u for device "
2069 "%d (%d)\n", cq_size, dev->gnd_id, rrc);
2070 GOTO(failed, rc = -EINVAL);
2073 rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
2074 0, kgnilnd_device_callback, dev->gnd_id,
2075 &dev->gnd_snd_fma_cqh);
2076 if (rrc != GNI_RC_SUCCESS) {
2077 CERROR("Can't create fma send cq size %u for device %d (%d)\n",
2078 cq_size, dev->gnd_id, rrc);
2079 GOTO(failed, rc = -EINVAL);
2082 /* This one we size differently - overflows are possible and it needs to be
2083 * sized based on machine size */
2084 rrc = kgnilnd_cq_create(dev->gnd_handle,
2085 *kgnilnd_tunables.kgn_fma_cq_size,
2086 0, kgnilnd_device_callback, dev->gnd_id,
2087 &dev->gnd_rcv_fma_cqh);
2088 if (rrc != GNI_RC_SUCCESS) {
2089 CERROR("Can't create fma cq size %d for device %d (%d)\n",
2090 *kgnilnd_tunables.kgn_fma_cq_size, dev->gnd_id, rrc);
2091 GOTO(failed, rc = -EINVAL);
2094 rrc = kgnilnd_register_smdd_buf(dev);
2095 if (rrc != GNI_RC_SUCCESS) {
2096 GOTO(failed, rc = -EINVAL);
2102 kgnilnd_dev_fini(dev);
2107 kgnilnd_dev_fini(kgn_device_t *dev)
2112 /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
2113 LASSERTF(list_empty(&dev->gnd_ready_conns) &&
2114 list_empty(&dev->gnd_map_tx) &&
2115 list_empty(&dev->gnd_rdmaq) &&
2116 list_empty(&dev->gnd_delay_conns),
2117 "dev 0x%p ready_conns %d@0x%p delay_conns %d@0x%p"
2118 "map_tx %d@0x%p rdmaq %d@0x%p\n",
2119 dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
2120 kgnilnd_count_list(&dev->gnd_delay_conns), &dev->gnd_delay_conns,
2121 kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
2122 kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
2124 /* These should follow from tearing down all connections */
2125 LASSERTF(dev->gnd_map_nphys == 0 && dev->gnd_map_physnop == 0,
2126 "%d physical mappings of %d pages still mapped\n",
2127 dev->gnd_map_nphys, dev->gnd_map_physnop);
2129 LASSERTF(dev->gnd_map_nvirt == 0 && dev->gnd_map_virtnob == 0,
2130 "%d virtual mappings of %llu bytes still mapped\n",
2131 dev->gnd_map_nvirt, dev->gnd_map_virtnob);
2133 LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
2134 atomic_read(&dev->gnd_n_mdd_held) == 0 &&
2135 atomic64_read(&dev->gnd_nbytes_map) == 0,
2136 "%d SMSG mappings of %ld bytes still mapped or held %d\n",
2137 atomic_read(&dev->gnd_n_mdd),
2138 atomic64_read(&dev->gnd_nbytes_map), atomic_read(&dev->gnd_n_mdd_held));
2140 LASSERT(list_empty(&dev->gnd_map_list));
2142 /* What other assertions needed to ensure all connections torn down ? */
2144 /* check all counters == 0 (EP, MDD, etc) */
2146 /* if we are resetting due to quiese (stack reset), don't check
2148 LASSERTF(kgnilnd_data.kgn_quiesce_trigger ||
2149 atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
2150 "tried to shutdown with threads active\n");
2152 if (dev->gnd_smdd_hold_buf) {
2153 rrc = kgnilnd_deregister_smdd_buf(dev);
2154 LASSERTF(rrc == GNI_RC_SUCCESS,
2155 "bad rc from deregistion of sMDD buffer: %d\n", rrc);
2156 dev->gnd_smdd_hold_buf = NULL;
2159 if (dev->gnd_rcv_fma_cqh) {
2160 rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
2161 LASSERTF(rrc == GNI_RC_SUCCESS,
2162 "bad rc from gni_cq_destroy on rcv_fma_cqh: %d\n", rrc);
2163 dev->gnd_rcv_fma_cqh = NULL;
2166 if (dev->gnd_snd_rdma_cqh) {
2167 rrc = kgnilnd_cq_destroy(dev->gnd_snd_rdma_cqh);
2168 LASSERTF(rrc == GNI_RC_SUCCESS,
2169 "bad rc from gni_cq_destroy on send_rdma_cqh: %d\n", rrc);
2170 dev->gnd_snd_rdma_cqh = NULL;
2173 if (dev->gnd_snd_fma_cqh) {
2174 rrc = kgnilnd_cq_destroy(dev->gnd_snd_fma_cqh);
2175 LASSERTF(rrc == GNI_RC_SUCCESS,
2176 "bad rc from gni_cq_destroy on snd_fma_cqh: %d\n", rrc);
2177 dev->gnd_snd_fma_cqh = NULL;
2180 if (dev->gnd_err_handle) {
2181 rrc = kgnilnd_release_errors(dev->gnd_err_handle);
2182 LASSERTF(rrc == GNI_RC_SUCCESS,
2183 "bad rc from gni_release_errors: %d\n", rrc);
2184 dev->gnd_err_handle = NULL;
2187 if (dev->gnd_domain) {
2188 rrc = kgnilnd_cdm_destroy(dev->gnd_domain);
2189 LASSERTF(rrc == GNI_RC_SUCCESS,
2190 "bad rc from gni_cdm_destroy: %d\n", rrc);
2191 dev->gnd_domain = NULL;
2197 int kgnilnd_base_startup(void)
2200 int pkmem = atomic_read(&libcfs_kmemory);
2204 struct task_struct *thrd;
2206 #if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
2207 /* limit how much memory can be allocated for fma blocks in
2208 * instances where many nodes need to reconnects at the same time */
2211 kgnilnd_data.free_pages_limit = si.totalram/4;
2216 LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
2217 "init %d\n", kgnilnd_data.kgn_init);
2219 /* zero pointers, flags etc */
2220 memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));
2221 kgnilnd_check_kgni_version();
2223 /* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
2224 * a unique (for all time) connstamp so we can uniquely identify
2225 * the sender. The connstamp is an incrementing counter
2226 * initialised with seconds + microseconds at startup time. So we
2227 * rely on NOT creating connections more frequently on average than
2228 * 1MHz to ensure we don't use old connstamps when we reboot. */
2229 do_gettimeofday(&tv);
2230 kgnilnd_data.kgn_connstamp =
2231 kgnilnd_data.kgn_peerstamp =
2232 (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2234 init_rwsem(&kgnilnd_data.kgn_net_rw_sem);
2236 for (i = 0; i < GNILND_MAXDEVS; i++) {
2237 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2240 INIT_LIST_HEAD(&dev->gnd_ready_conns);
2241 INIT_LIST_HEAD(&dev->gnd_delay_conns);
2242 INIT_LIST_HEAD(&dev->gnd_map_tx);
2243 INIT_LIST_HEAD(&dev->gnd_fma_buffs);
2244 mutex_init(&dev->gnd_cq_mutex);
2245 mutex_init(&dev->gnd_fmablk_mutex);
2246 spin_lock_init(&dev->gnd_fmablk_lock);
2247 init_waitqueue_head(&dev->gnd_waitq);
2248 init_waitqueue_head(&dev->gnd_dgram_waitq);
2249 init_waitqueue_head(&dev->gnd_dgping_waitq);
2250 spin_lock_init(&dev->gnd_lock);
2251 INIT_LIST_HEAD(&dev->gnd_map_list);
2252 spin_lock_init(&dev->gnd_map_lock);
2253 atomic_set(&dev->gnd_nfmablk, 0);
2254 atomic_set(&dev->gnd_fmablk_vers, 1);
2255 atomic_set(&dev->gnd_neps, 0);
2256 atomic_set(&dev->gnd_canceled_dgrams, 0);
2257 INIT_LIST_HEAD(&dev->gnd_connd_peers);
2258 spin_lock_init(&dev->gnd_connd_lock);
2259 spin_lock_init(&dev->gnd_dgram_lock);
2260 spin_lock_init(&dev->gnd_rdmaq_lock);
2261 INIT_LIST_HEAD(&dev->gnd_rdmaq);
2262 init_rwsem(&dev->gnd_conn_sem);
2264 /* alloc & setup nid based dgram table */
2265 LIBCFS_ALLOC(dev->gnd_dgrams,
2266 sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
2268 if (dev->gnd_dgrams == NULL)
2269 GOTO(failed, rc = -ENOMEM);
2271 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2272 INIT_LIST_HEAD(&dev->gnd_dgrams[i]);
2274 atomic_set(&dev->gnd_ndgrams, 0);
2275 atomic_set(&dev->gnd_nwcdgrams, 0);
2276 /* setup timer for RDMAQ processing */
2277 setup_timer(&dev->gnd_rdmaq_timer, kgnilnd_schedule_device_timer,
2278 (unsigned long)dev);
2280 /* setup timer for mapping processing */
2281 setup_timer(&dev->gnd_map_timer, kgnilnd_schedule_device_timer,
2282 (unsigned long)dev);
2286 /* CQID 0 isn't allowed, set to MAX_MSG_ID - 1 to check for conflicts early */
2287 kgnilnd_data.kgn_next_cqid = GNILND_MAX_MSG_ID - 1;
2288 kgnilnd_data.kgn_new_min_timeout = *kgnilnd_tunables.kgn_timeout;
2289 init_waitqueue_head(&kgnilnd_data.kgn_reaper_waitq);
2290 init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
2291 spin_lock_init(&kgnilnd_data.kgn_reaper_lock);
2293 mutex_init(&kgnilnd_data.kgn_quiesce_mutex);
2294 atomic_set(&kgnilnd_data.kgn_nquiesce, 0);
2295 atomic_set(&kgnilnd_data.kgn_npending_conns, 0);
2296 atomic_set(&kgnilnd_data.kgn_npending_unlink, 0);
2297 atomic_set(&kgnilnd_data.kgn_npending_detach, 0);
2298 atomic_set(&kgnilnd_data.kgn_rev_offset, 0);
2299 atomic_set(&kgnilnd_data.kgn_rev_length, 0);
2300 atomic_set(&kgnilnd_data.kgn_rev_copy_buff, 0);
2302 /* OK to call kgnilnd_api_shutdown() to cleanup now */
2303 kgnilnd_data.kgn_init = GNILND_INIT_DATA;
2304 try_module_get(THIS_MODULE);
2306 rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);
2308 LIBCFS_ALLOC(kgnilnd_data.kgn_peers,
2309 sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
2311 if (kgnilnd_data.kgn_peers == NULL)
2312 GOTO(failed, rc = -ENOMEM);
2314 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2315 INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
2318 LIBCFS_ALLOC(kgnilnd_data.kgn_conns,
2319 sizeof(struct list_head) * *kgnilnd_tunables.kgn_peer_hash_size);
2321 if (kgnilnd_data.kgn_conns == NULL)
2322 GOTO(failed, rc = -ENOMEM);
2324 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2325 INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
2328 LIBCFS_ALLOC(kgnilnd_data.kgn_nets,
2329 sizeof(struct list_head) * *kgnilnd_tunables.kgn_net_hash_size);
2331 if (kgnilnd_data.kgn_nets == NULL)
2332 GOTO(failed, rc = -ENOMEM);
2334 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2335 INIT_LIST_HEAD(&kgnilnd_data.kgn_nets[i]);
2338 kgnilnd_data.kgn_mbox_cache =
2339 kmem_cache_create("kgn_mbox_block", GNILND_MBOX_SIZE, 0,
2340 SLAB_HWCACHE_ALIGN, NULL);
2341 if (kgnilnd_data.kgn_mbox_cache == NULL) {
2342 CERROR("Can't create slab for physical mbox blocks\n");
2343 GOTO(failed, rc = -ENOMEM);
2346 kgnilnd_data.kgn_rx_cache =
2347 kmem_cache_create("kgn_rx_t", sizeof(kgn_rx_t), 0, 0, NULL);
2348 if (kgnilnd_data.kgn_rx_cache == NULL) {
2349 CERROR("Can't create slab for kgn_rx_t descriptors\n");
2350 GOTO(failed, rc = -ENOMEM);
2353 kgnilnd_data.kgn_tx_cache =
2354 kmem_cache_create("kgn_tx_t", sizeof(kgn_tx_t), 0, 0, NULL);
2355 if (kgnilnd_data.kgn_tx_cache == NULL) {
2356 CERROR("Can't create slab for kgn_tx_t\n");
2357 GOTO(failed, rc = -ENOMEM);
2360 kgnilnd_data.kgn_tx_phys_cache =
2361 kmem_cache_create("kgn_tx_phys",
2362 LNET_MAX_IOV * sizeof(gni_mem_segment_t),
2364 if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
2365 CERROR("Can't create slab for kgn_tx_phys\n");
2366 GOTO(failed, rc = -ENOMEM);
2369 kgnilnd_data.kgn_dgram_cache =
2370 kmem_cache_create("kgn_dgram_t", sizeof(kgn_dgram_t), 0, 0, NULL);
2371 if (kgnilnd_data.kgn_dgram_cache == NULL) {
2372 CERROR("Can't create slab for outgoing datagrams\n");
2373 GOTO(failed, rc = -ENOMEM);
2376 /* allocate a MAX_IOV array of page pointers for each cpu */
2377 kgnilnd_data.kgn_cksum_map_pages = kmalloc(num_possible_cpus() * sizeof (struct page *),
2379 if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
2380 CERROR("Can't allocate vmap cksum pages\n");
2381 GOTO(failed, rc = -ENOMEM);
2383 kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
2384 memset(kgnilnd_data.kgn_cksum_map_pages, 0,
2385 kgnilnd_data.kgn_cksum_npages * sizeof (struct page *));
2387 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2388 kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(LNET_MAX_IOV * sizeof (struct page *),
2390 if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
2391 CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
2392 GOTO(failed, rc = -ENOMEM);
2396 LASSERT(kgnilnd_data.kgn_ndevs == 0);
2398 /* Use all available GNI devices */
2399 for (i = 0; i < GNILND_MAXDEVS; i++) {
2400 dev = &kgnilnd_data.kgn_devices[kgnilnd_data.kgn_ndevs];
2402 rc = kgnilnd_dev_init(dev);
2404 /* Increment here so base_shutdown cleans it up */
2405 kgnilnd_data.kgn_ndevs++;
2407 rc = kgnilnd_allocate_phys_fmablk(dev);
2413 if (kgnilnd_data.kgn_ndevs == 0) {
2414 CERROR("Can't initialise any GNI devices\n");
2415 GOTO(failed, rc = -ENODEV);
2418 rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
2420 CERROR("Can't spawn gnilnd reaper: %d\n", rc);
2424 rc = kgnilnd_start_rca_thread();
2426 CERROR("Can't spawn gnilnd rca: %d\n", rc);
2431 * Start ruhroh thread. We can't use kgnilnd_thread_start() because
2432 * we don't want this thread included in kgnilnd_data.kgn_nthreads
2433 * count. This thread controls quiesce, so it mustn't
2436 thrd = kthread_run(kgnilnd_ruhroh_thread, NULL, "%s_%02d", "kgnilnd_rr", 0);
2439 CERROR("Can't spawn gnilnd ruhroh thread: %d\n", rc);
2443 /* threads will load balance across devs as they are available */
2444 if (*kgnilnd_tunables.kgn_thread_affinity) {
2445 rc = kgnilnd_start_sd_threads();
2449 for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
2450 rc = kgnilnd_thread_start(kgnilnd_scheduler,
2454 CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
2461 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2462 dev = &kgnilnd_data.kgn_devices[i];
2463 rc = kgnilnd_thread_start(kgnilnd_dgram_mover, dev,
2464 "kgnilnd_dg", dev->gnd_id);
2466 CERROR("Can't spawn gnilnd dgram_mover[%d]: %d\n",
2471 rc = kgnilnd_thread_start(kgnilnd_dgram_waitq, dev,
2472 "kgnilnd_dgn", dev->gnd_id);
2474 CERROR("Can't spawn gnilnd dgram_waitq[%d]: %d\n",
2479 rc = kgnilnd_setup_wildcard_dgram(dev);
2482 CERROR("Can't create wildcard dgrams[%d]: %d\n",
2488 /* flag everything initialised */
2489 kgnilnd_data.kgn_init = GNILND_INIT_ALL;
2490 /*****************************************************/
2492 CDEBUG(D_MALLOC, "initial kmem %d\n", pkmem);
2496 kgnilnd_base_shutdown();
2497 kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2502 kgnilnd_base_shutdown(void)
2507 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};
2509 kgnilnd_data.kgn_wc_kill = 1;
2511 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2512 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2513 kgnilnd_cancel_wc_dgrams(dev);
2514 kgnilnd_cancel_dgrams(dev);
2515 kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2516 kgnilnd_wait_for_canceled_dgrams(dev);
2519 /* We need to verify there are no conns left before we let the threads
2520 * shut down otherwise we could clean up the peers but still have
2521 * some outstanding conns due to orphaned datagram conns that are
2525 while (atomic_read(&kgnilnd_data.kgn_nconns) != 0) {
2528 for(j = 0; j < kgnilnd_data.kgn_ndevs; ++j) {
2529 kgn_device_t *dev = &kgnilnd_data.kgn_devices[j];
2530 kgnilnd_schedule_device(dev);
2533 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2534 "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
2535 set_current_state(TASK_UNINTERRUPTIBLE);
2536 schedule_timeout(cfs_time_seconds(1));
2538 /* Peer state all cleaned up BEFORE setting shutdown, so threads don't
2539 * have to worry about shutdown races. NB connections may be created
2540 * while there are still active connds, but these will be temporary
2541 * since peer creation always fails after the listener has started to
2543 * all peers should have been cleared out on the nets */
2544 LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2545 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2547 /* Wait for the ruhroh thread to shut down. */
2548 kgnilnd_data.kgn_ruhroh_shutdown = 1;
2549 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
2551 while (kgnilnd_data.kgn_ruhroh_running != 0) {
2553 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2554 "Waiting for ruhroh thread to terminate\n");
2555 set_current_state(TASK_UNINTERRUPTIBLE);
2556 schedule_timeout(cfs_time_seconds(1));
2559 /* Flag threads to terminate */
2560 kgnilnd_data.kgn_shutdown = 1;
2562 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2563 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2565 /* should clear all the MDDs */
2566 kgnilnd_unmap_fma_blocks(dev);
2568 kgnilnd_schedule_device(dev);
2569 wake_up_all(&dev->gnd_dgram_waitq);
2570 wake_up_all(&dev->gnd_dgping_waitq);
2571 LASSERT(list_empty(&dev->gnd_connd_peers));
2574 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2575 wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
2576 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2578 if (atomic_read(&kgnilnd_data.kgn_nthreads))
2579 kgnilnd_wakeup_rca_thread();
2581 /* Wait for threads to exit */
2583 while (atomic_read(&kgnilnd_data.kgn_nthreads) != 0) {
2585 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2586 "Waiting for %d threads to terminate\n",
2587 atomic_read(&kgnilnd_data.kgn_nthreads));
2588 set_current_state(TASK_UNINTERRUPTIBLE);
2589 schedule_timeout(cfs_time_seconds(1));
2592 LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2593 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2595 if (kgnilnd_data.kgn_peers != NULL) {
2596 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2597 LASSERT(list_empty(&kgnilnd_data.kgn_peers[i]));
2599 LIBCFS_FREE(kgnilnd_data.kgn_peers,
2600 sizeof (struct list_head) *
2601 *kgnilnd_tunables.kgn_peer_hash_size);
2604 down_write(&kgnilnd_data.kgn_net_rw_sem);
2605 if (kgnilnd_data.kgn_nets != NULL) {
2606 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++)
2607 LASSERT(list_empty(&kgnilnd_data.kgn_nets[i]));
2609 LIBCFS_FREE(kgnilnd_data.kgn_nets,
2610 sizeof (struct list_head) *
2611 *kgnilnd_tunables.kgn_net_hash_size);
2613 up_write(&kgnilnd_data.kgn_net_rw_sem);
2615 LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
2616 "conns left %d\n", atomic_read(&kgnilnd_data.kgn_nconns));
2618 if (kgnilnd_data.kgn_conns != NULL) {
2619 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2620 LASSERT(list_empty(&kgnilnd_data.kgn_conns[i]));
2622 LIBCFS_FREE(kgnilnd_data.kgn_conns,
2623 sizeof (struct list_head) *
2624 *kgnilnd_tunables.kgn_peer_hash_size);
2627 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2628 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2629 kgnilnd_dev_fini(dev);
2631 LASSERTF(atomic_read(&dev->gnd_ndgrams) == 0,
2632 "dgrams left %d\n", atomic_read(&dev->gnd_ndgrams));
2634 if (dev->gnd_dgrams != NULL) {
2635 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2636 LASSERT(list_empty(&dev->gnd_dgrams[i]));
2638 LIBCFS_FREE(dev->gnd_dgrams,
2639 sizeof (struct list_head) *
2640 *kgnilnd_tunables.kgn_peer_hash_size);
2643 kgnilnd_free_phys_fmablk(dev);
2646 if (kgnilnd_data.kgn_mbox_cache != NULL)
2647 kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
2649 if (kgnilnd_data.kgn_rx_cache != NULL)
2650 kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);
2652 if (kgnilnd_data.kgn_tx_cache != NULL)
2653 kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);
2655 if (kgnilnd_data.kgn_tx_phys_cache != NULL)
2656 kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
2658 if (kgnilnd_data.kgn_dgram_cache != NULL)
2659 kmem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
2661 if (kgnilnd_data.kgn_cksum_map_pages != NULL) {
2662 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2663 if (kgnilnd_data.kgn_cksum_map_pages[i] != NULL) {
2664 kfree(kgnilnd_data.kgn_cksum_map_pages[i]);
2667 kfree(kgnilnd_data.kgn_cksum_map_pages);
2670 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2671 atomic_read(&libcfs_kmemory));
2673 kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2674 module_put(THIS_MODULE);
2680 kgnilnd_startup(struct lnet_ni *ni)
2686 LASSERTF(ni->ni_net->net_lnd == &the_kgnilnd,
2687 "bad LND 0x%p != the_kgnilnd @ 0x%p\n",
2688 ni->ni_net->net_lnd, &the_kgnilnd);
2690 if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
2691 rc = kgnilnd_base_startup();
2696 /* Serialize with shutdown. */
2697 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2699 LIBCFS_ALLOC(net, sizeof(*net));
2701 CERROR("could not allocate net for new interface instance\n");
2702 /* no need to cleanup the CDM... */
2703 GOTO(failed, rc = -ENOMEM);
2705 INIT_LIST_HEAD(&net->gnn_list);
2708 if (!ni->ni_net->net_tunables_set) {
2709 ni->ni_net->net_tunables.lct_max_tx_credits =
2710 *kgnilnd_tunables.kgn_credits;
2711 ni->ni_net->net_tunables.lct_peer_tx_credits =
2712 *kgnilnd_tunables.kgn_peer_credits;
2715 if (*kgnilnd_tunables.kgn_peer_health) {
2718 /* give this a bit of leeway - we don't have a hard timeout
2719 * as we only check timeouts periodically - see comment in kgnilnd_reaper */
2720 fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
2721 timeout = *kgnilnd_tunables.kgn_timeout + fudge;
2723 if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) {
2724 ni->ni_net->net_tunables.lct_peer_timeout =
2725 *kgnilnd_tunables.kgn_peer_timeout;
2726 } else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
2727 LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
2728 *kgnilnd_tunables.kgn_peer_timeout,
2731 LIBCFS_FREE(net, sizeof(*net));
2732 GOTO(failed, rc = -EINVAL);
2734 ni->ni_net->net_tunables.lct_peer_timeout = timeout;
2736 LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
2737 ni->ni_net->net_tunables.lct_peer_timeout);
2740 atomic_set(&net->gnn_refcount, 1);
2742 /* if we have multiple devices, spread the nets around */
2743 net->gnn_netnum = LNET_NETNUM(LNET_NIDNET(ni->ni_nid));
2745 devno = LNET_NIDNET(ni->ni_nid) % GNILND_MAXDEVS;
2746 net->gnn_dev = &kgnilnd_data.kgn_devices[devno];
2748 /* allocate a 'dummy' cdm for datagram use. We can only have a single
2749 * datagram between a nid:inst_id and nid2:inst_id. The fake cdm
2750 * give us additional inst_id to use, allowing the datagrams to flow
2751 * like rivers of honey and beer */
2753 /* the instance id for the cdm is the NETNUM offset by MAXDEVS -
2754 * ensuring we'll have a unique id */
2757 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), net->gnn_dev->gnd_nid);
2758 CDEBUG(D_NET, "adding net %p nid=%s on dev %d \n",
2759 net, libcfs_nid2str(ni->ni_nid), net->gnn_dev->gnd_id);
2760 /* until the gnn_list is set, we need to cleanup ourselves as
2761 * kgnilnd_shutdown is just gonna get confused */
2763 down_write(&kgnilnd_data.kgn_net_rw_sem);
2764 list_add_tail(&net->gnn_list, kgnilnd_netnum2netlist(net->gnn_netnum));
2765 up_write(&kgnilnd_data.kgn_net_rw_sem);
2767 /* we need a separate thread to call probe_wait_by_id until
2768 * we get a function callback notifier from kgni */
2769 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2772 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2773 kgnilnd_shutdown(ni);
2778 kgnilnd_shutdown(struct lnet_ni *ni)
2780 kgn_net_t *net = ni->ni_data;
2785 CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);
2787 LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_ALL,
2788 "init %d\n", kgnilnd_data.kgn_init);
2790 /* Serialize with startup. */
2791 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2792 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2793 atomic_read(&libcfs_kmemory));
2796 CERROR("got NULL net for ni %p\n", ni);
2797 GOTO(out, rc = -EINVAL);
2800 LASSERTF(ni == net->gnn_ni,
2801 "ni %p gnn_ni %p\n", net, net->gnn_ni);
2805 LASSERT(!net->gnn_shutdown);
2806 LASSERTF(atomic_read(&net->gnn_refcount) != 0,
2807 "net %p refcount %d\n",
2808 net, atomic_read(&net->gnn_refcount));
2810 if (!list_empty(&net->gnn_list)) {
2811 /* serialize with peer creation */
2812 down_write(&kgnilnd_data.kgn_net_rw_sem);
2813 net->gnn_shutdown = 1;
2814 up_write(&kgnilnd_data.kgn_net_rw_sem);
2816 kgnilnd_cancel_net_dgrams(net);
2818 kgnilnd_del_conn_or_peer(net, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2820 /* if we are quiesced, need to wake up - we need those threads
2821 * alive to release peers, etc */
2822 if (GNILND_IS_QUIESCED) {
2823 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
2824 kgnilnd_quiesce_wait("shutdown");
2827 kgnilnd_wait_for_canceled_dgrams(net->gnn_dev);
2829 /* We wait until the nets ref's are 1, we will release final ref which is ours
2830 * this allows us to make sure everything else is done before we free the
2834 while (atomic_read(&net->gnn_refcount) != 1) {
2836 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2837 "Waiting for %d references to clear on net %d\n",
2838 atomic_read(&net->gnn_refcount),
2840 set_current_state(TASK_UNINTERRUPTIBLE);
2841 schedule_timeout(cfs_time_seconds(1));
2844 /* release ref from kgnilnd_startup */
2845 kgnilnd_net_decref(net);
2846 /* serialize with reaper and conn_task looping */
2847 down_write(&kgnilnd_data.kgn_net_rw_sem);
2848 list_del_init(&net->gnn_list);
2849 up_write(&kgnilnd_data.kgn_net_rw_sem);
2853 /* not locking, this can't race with writers */
2854 LASSERTF(atomic_read(&net->gnn_refcount) == 0,
2855 "net %p refcount %d\n",
2856 net, atomic_read(&net->gnn_refcount));
2857 LIBCFS_FREE(net, sizeof(*net));
2860 down_read(&kgnilnd_data.kgn_net_rw_sem);
2861 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2862 if (!list_empty(&kgnilnd_data.kgn_nets[i])) {
2863 up_read(&kgnilnd_data.kgn_net_rw_sem);
2867 if (i == *kgnilnd_tunables.kgn_net_hash_size - 1) {
2868 up_read(&kgnilnd_data.kgn_net_rw_sem);
2869 kgnilnd_base_shutdown();
2872 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2873 atomic_read(&libcfs_kmemory));
2875 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2879 static void __exit kgnilnd_exit(void)
2881 lnet_unregister_lnd(&the_kgnilnd);
2882 kgnilnd_proc_fini();
2883 kgnilnd_remove_sysctl();
2886 static int __init kgnilnd_init(void)
2890 rc = kgnilnd_tunables_init();
2894 printk(KERN_INFO "Lustre: kgnilnd build version: "KGNILND_BUILD_REV"\n");
2896 kgnilnd_insert_sysctl();
2897 kgnilnd_proc_init();
2899 lnet_register_lnd(&the_kgnilnd);
2904 MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
2905 MODULE_DESCRIPTION("Gemini LNet Network Driver");
2906 MODULE_VERSION(KGNILND_BUILD_REV);
2907 MODULE_LICENSE("GPL");
2909 module_init(kgnilnd_init);
2910 module_exit(kgnilnd_exit);