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 const 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,
38 kgn_data_t kgnilnd_data;
41 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
43 struct task_struct *thrd;
45 thrd = kthread_run(fn, arg, "%s_%02d", name, id);
49 atomic_inc(&kgnilnd_data.kgn_nthreads);
53 /* bind scheduler threads to cpus */
55 kgnilnd_start_sd_threads(void)
59 struct task_struct *task;
61 for_each_online_cpu(cpu) {
62 /* don't bind to cpu 0 - all interrupts are processed here */
66 task = kthread_create(kgnilnd_scheduler, (void *)((long)i),
67 "%s_%02d", "kgnilnd_sd", i);
69 kthread_bind(task, cpu);
70 wake_up_process(task);
72 CERROR("Can't spawn gnilnd scheduler[%d] %ld\n", i,
76 atomic_inc(&kgnilnd_data.kgn_nthreads);
78 if (++i >= *kgnilnd_tunables.kgn_sched_threads) {
86 /* needs write_lock on kgn_peer_conn_lock */
88 kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
90 kgn_conn_t *conn, *cnxt;
94 loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
96 list_for_each_entry_safe(conn, cnxt, &peer->gnp_conns, gnc_list) {
97 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
103 if (conn->gnc_device != newconn->gnc_device)
106 /* This is a two connection loopback - one talking to the other */
108 newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
109 newconn->gnc_peer_connstamp == conn->gnc_my_connstamp) {
110 CDEBUG(D_NET, "skipping prune of %p, "
111 "loopback and matching stamps"
112 " connstamp %llu(%llu)"
113 " peerstamp %llu(%llu)\n",
114 conn, newconn->gnc_my_connstamp,
115 conn->gnc_peer_connstamp,
116 newconn->gnc_peer_connstamp,
117 conn->gnc_my_connstamp);
121 if (conn->gnc_peerstamp != newconn->gnc_peerstamp) {
122 LASSERTF(conn->gnc_peerstamp < newconn->gnc_peerstamp,
123 "conn 0x%p peerstamp %llu >= "
124 "newconn 0x%p peerstamp %llu\n",
125 conn, conn->gnc_peerstamp,
126 newconn, newconn->gnc_peerstamp);
128 CDEBUG(D_NET, "Closing stale conn nid: %s "
129 " peerstamp:%#llx(%#llx)\n",
130 libcfs_nid2str(peer->gnp_nid),
131 conn->gnc_peerstamp, newconn->gnc_peerstamp);
134 LASSERTF(conn->gnc_peer_connstamp < newconn->gnc_peer_connstamp,
135 "conn 0x%p peer_connstamp %llu >= "
136 "newconn 0x%p peer_connstamp %llu\n",
137 conn, conn->gnc_peer_connstamp,
138 newconn, newconn->gnc_peer_connstamp);
140 CDEBUG(D_NET, "Closing stale conn nid: %s"
141 " connstamp:%llu(%llu)\n",
142 libcfs_nid2str(peer->gnp_nid),
143 conn->gnc_peer_connstamp, newconn->gnc_peer_connstamp);
147 kgnilnd_close_conn_locked(conn, -ESTALE);
151 CWARN("Closed %d stale conns to %s\n", count, libcfs_nid2str(peer->gnp_nid));
158 kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn)
164 loopback = peer->gnp_nid == peer->gnp_net->gnn_ni->ni_nid;
166 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
167 CDEBUG(D_NET, "checking conn 0x%p for peer %s"
168 " lo %d new %llu existing %llu"
169 " new peer %llu existing peer %llu"
170 " new dev %p existing dev %p\n",
171 conn, libcfs_nid2str(peer->gnp_nid),
173 newconn->gnc_peerstamp, conn->gnc_peerstamp,
174 newconn->gnc_peer_connstamp, conn->gnc_peer_connstamp,
175 newconn->gnc_device, conn->gnc_device);
177 /* conn is in the process of closing */
178 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
181 /* 'newconn' is from an earlier version of 'peer'!!! */
182 if (newconn->gnc_peerstamp < conn->gnc_peerstamp)
185 /* 'conn' is from an earlier version of 'peer': it will be
186 * removed when we cull stale conns later on... */
187 if (newconn->gnc_peerstamp > conn->gnc_peerstamp)
190 /* Different devices are OK */
191 if (conn->gnc_device != newconn->gnc_device)
194 /* It's me connecting to myself */
196 newconn->gnc_my_connstamp == conn->gnc_peer_connstamp &&
197 newconn->gnc_peer_connstamp == conn->gnc_my_connstamp)
200 /* 'newconn' is an earlier connection from 'peer'!!! */
201 if (newconn->gnc_peer_connstamp < conn->gnc_peer_connstamp)
204 /* 'conn' is an earlier connection from 'peer': it will be
205 * removed when we cull stale conns later on... */
206 if (newconn->gnc_peer_connstamp > conn->gnc_peer_connstamp)
209 /* 'newconn' has the SAME connection stamp; 'peer' isn't
210 * playing the game... */
218 kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
224 LASSERT (!in_interrupt());
225 atomic_inc(&kgnilnd_data.kgn_nconns);
227 /* divide by 2 to allow for complete reset and immediate reconnect */
228 if (atomic_read(&kgnilnd_data.kgn_nconns) >= GNILND_MAX_CQID/2) {
229 CERROR("Too many conn are live: %d > %d\n",
230 atomic_read(&kgnilnd_data.kgn_nconns), GNILND_MAX_CQID/2);
231 atomic_dec(&kgnilnd_data.kgn_nconns);
235 LIBCFS_ALLOC(conn, sizeof(*conn));
237 atomic_dec(&kgnilnd_data.kgn_nconns);
241 conn->gnc_tx_ref_table =
242 kgnilnd_vzalloc(GNILND_MAX_MSG_ID * sizeof(void *));
243 if (conn->gnc_tx_ref_table == NULL) {
244 CERROR("Can't allocate conn tx_ref_table\n");
245 GOTO(failed, rc = -ENOMEM);
248 mutex_init(&conn->gnc_smsg_mutex);
249 mutex_init(&conn->gnc_rdma_mutex);
250 atomic_set(&conn->gnc_refcount, 1);
251 atomic_set(&conn->gnc_reaper_noop, 0);
252 atomic_set(&conn->gnc_sched_noop, 0);
253 atomic_set(&conn->gnc_tx_in_use, 0);
254 INIT_LIST_HEAD(&conn->gnc_list);
255 INIT_LIST_HEAD(&conn->gnc_hashlist);
256 INIT_LIST_HEAD(&conn->gnc_schedlist);
257 INIT_LIST_HEAD(&conn->gnc_fmaq);
258 INIT_LIST_HEAD(&conn->gnc_mdd_list);
259 INIT_LIST_HEAD(&conn->gnc_delaylist);
260 spin_lock_init(&conn->gnc_list_lock);
261 spin_lock_init(&conn->gnc_tx_lock);
262 conn->gnc_magic = GNILND_CONN_MAGIC;
264 /* set tx id to nearly the end to make sure we find wrapping
266 conn->gnc_next_tx = (int) GNILND_MAX_MSG_ID - 10;
268 /* if this fails, we have conflicts and MAX_TX is too large */
269 BUILD_BUG_ON(GNILND_MAX_MSG_ID >= GNILND_MSGID_CLOSE);
271 /* get a new unique CQ id for this conn */
272 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
273 conn->gnc_my_connstamp = kgnilnd_data.kgn_connstamp++;
274 conn->gnc_cqid = kgnilnd_get_cqid_locked();
275 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
277 if (conn->gnc_cqid == 0) {
278 CERROR("Could not allocate unique CQ ID for conn 0x%p\n", conn);
279 GOTO(failed, rc = -E2BIG);
282 CDEBUG(D_NET, "alloc cqid %u for conn 0x%p\n",
283 conn->gnc_cqid, conn);
285 /* need to be set before gnc_ephandle to allow kgnilnd_destroy_conn_ep to
287 conn->gnc_device = dev;
289 conn->gnc_timeout = max(*kgnilnd_tunables.kgn_timeout,
291 kgnilnd_update_reaper_timeout(conn->gnc_timeout);
293 /* this is the ep_handle for doing SMSG & BTE */
294 mutex_lock(&dev->gnd_cq_mutex);
295 rrc = kgnilnd_ep_create(dev->gnd_handle, dev->gnd_snd_fma_cqh,
296 &conn->gnc_ephandle);
297 mutex_unlock(&dev->gnd_cq_mutex);
298 if (rrc != GNI_RC_SUCCESS)
299 GOTO(failed, rc = -ENETDOWN);
301 CDEBUG(D_NET, "created conn 0x%p ep_hndl 0x%p\n",
302 conn, conn->gnc_ephandle);
304 /* add ref for EP canceling */
305 kgnilnd_conn_addref(conn);
306 atomic_inc(&dev->gnd_neps);
312 atomic_dec(&kgnilnd_data.kgn_nconns);
313 kgnilnd_vfree(conn->gnc_tx_ref_table,
314 GNILND_MAX_MSG_ID * sizeof(void *));
315 LIBCFS_FREE(conn, sizeof(*conn));
319 /* needs to be called with kgn_peer_conn_lock held (read or write) */
321 kgnilnd_find_conn_locked(kgn_peer_t *peer)
323 kgn_conn_t *conn = NULL;
325 /* if we are in reset, this conn is going to die soon */
326 if (unlikely(kgnilnd_data.kgn_in_reset)) {
330 /* just return the first ESTABLISHED connection */
331 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
332 /* kgnilnd_finish_connect doesn't put connections on the
333 * peer list until they are actually established */
334 LASSERTF(conn->gnc_state >= GNILND_CONN_ESTABLISHED,
335 "found conn %p state %s on peer %p (%s)\n",
336 conn, kgnilnd_conn_state2str(conn), peer,
337 libcfs_nid2str(peer->gnp_nid));
338 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
346 /* needs write_lock on kgn_peer_conn_lock held */
348 kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer) {
350 kgn_device_t *dev = peer->gnp_net->gnn_dev;
353 conn = kgnilnd_find_conn_locked(peer);
359 /* if the peer was previously connecting, check if we should
360 * trigger another connection attempt yet. */
361 if (time_before(jiffies, peer->gnp_reconnect_time)) {
365 /* This check prevents us from creating a new connection to a peer while we are
366 * still in the process of closing an existing connection to the peer.
368 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
369 if (conn->gnc_ephandle != NULL) {
370 CDEBUG(D_NET, "Not connecting non-null ephandle found peer 0x%p->%s\n", peer,
371 libcfs_nid2str(peer->gnp_nid));
376 if (peer->gnp_connecting != GNILND_PEER_IDLE) {
377 /* if we are not connecting, fire up a new connection */
378 /* or if we are anything but IDLE DONT start a new connection */
382 CDEBUG(D_NET, "starting connect to %s\n",
383 libcfs_nid2str(peer->gnp_nid));
384 peer->gnp_connecting = GNILND_PEER_CONNECT;
385 kgnilnd_peer_addref(peer); /* extra ref for connd */
387 spin_lock(&dev->gnd_connd_lock);
388 list_add_tail(&peer->gnp_connd_list, &dev->gnd_connd_peers);
389 spin_unlock(&dev->gnd_connd_lock);
391 kgnilnd_schedule_dgram(dev);
392 CDEBUG(D_NETTRACE, "scheduling new connect\n");
397 /* Caller is responsible for deciding if/when to call this */
399 kgnilnd_destroy_conn_ep(kgn_conn_t *conn)
402 gni_ep_handle_t tmp_ep;
404 /* only if we actually initialized it,
405 * then set NULL to tell kgnilnd_destroy_conn to leave it alone */
407 tmp_ep = xchg(&conn->gnc_ephandle, NULL);
408 if (tmp_ep != NULL) {
409 /* we never re-use the EP, so unbind is not needed */
410 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
411 rrc = kgnilnd_ep_destroy(tmp_ep);
413 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
415 /* if this fails, it could hork up kgni smsg retransmit and others
416 * since we could free the SMSG mbox memory, etc. */
417 LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d conn 0x%p ep 0x%p\n",
418 rrc, conn, conn->gnc_ephandle);
420 atomic_dec(&conn->gnc_device->gnd_neps);
422 /* clear out count added in kgnilnd_close_conn_locked
423 * conn will have a peer once it hits finish_connect, where it
424 * is the first spot we'll mark it ESTABLISHED as well */
425 if (conn->gnc_peer) {
426 kgnilnd_admin_decref(conn->gnc_peer->gnp_dirty_eps);
429 /* drop ref for EP */
430 kgnilnd_conn_decref(conn);
435 kgnilnd_destroy_conn(kgn_conn_t *conn)
437 LASSERTF(!in_interrupt() &&
438 !conn->gnc_scheduled &&
439 !conn->gnc_in_purgatory &&
440 conn->gnc_ephandle == NULL &&
441 list_empty(&conn->gnc_list) &&
442 list_empty(&conn->gnc_hashlist) &&
443 list_empty(&conn->gnc_schedlist) &&
444 list_empty(&conn->gnc_mdd_list) &&
445 list_empty(&conn->gnc_delaylist) &&
446 conn->gnc_magic == GNILND_CONN_MAGIC,
447 "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d/%d\n",
448 conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
450 !!in_interrupt(), conn->gnc_scheduled,
451 conn->gnc_in_purgatory,
454 list_empty(&conn->gnc_list),
455 list_empty(&conn->gnc_hashlist),
456 list_empty(&conn->gnc_schedlist),
457 list_empty(&conn->gnc_mdd_list),
458 list_empty(&conn->gnc_delaylist));
460 /* Tripping these is especially bad, as it means we have items on the
461 * lists that didn't keep their refcount on the connection - or
462 * somebody evil released their own */
463 LASSERTF(list_empty(&conn->gnc_fmaq) &&
464 atomic_read(&conn->gnc_nlive_fma) == 0 &&
465 atomic_read(&conn->gnc_nlive_rdma) == 0,
466 "conn 0x%p fmaq %d@0x%p nfma %d nrdma %d\n",
467 conn, kgnilnd_count_list(&conn->gnc_fmaq), &conn->gnc_fmaq,
468 atomic_read(&conn->gnc_nlive_fma), atomic_read(&conn->gnc_nlive_rdma));
470 CDEBUG(D_NET, "destroying conn %p ephandle %p error %d\n",
471 conn, conn->gnc_ephandle, conn->gnc_error);
473 /* We are freeing this memory remove the magic value from the connection */
476 /* if there is an FMA blk left here, we'll tear it down */
477 if (conn->gnc_fma_blk) {
478 if (conn->gnc_peer) {
479 kgn_mbox_info_t *mbox;
480 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
481 mbox->mbx_prev_nid = conn->gnc_peer->gnp_nid;
483 kgnilnd_release_mbox(conn, 0);
486 if (conn->gnc_peer != NULL)
487 kgnilnd_peer_decref(conn->gnc_peer);
489 if (conn->gnc_tx_ref_table != NULL) {
490 kgnilnd_vfree(conn->gnc_tx_ref_table,
491 GNILND_MAX_MSG_ID * sizeof(void *));
494 LIBCFS_FREE(conn, sizeof(*conn));
495 atomic_dec(&kgnilnd_data.kgn_nconns);
498 /* peer_alive and peer_notify done in the style of the o2iblnd */
500 kgnilnd_peer_alive(kgn_peer_t *peer)
502 time64_t now = ktime_get_seconds();
504 set_mb(peer->gnp_last_alive, now);
508 kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive)
519 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DONT_NOTIFY))
522 /* Tell LNet we are giving ups on this peer - but only
523 * if it isn't already reconnected or trying to reconnect */
524 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
526 /* use kgnilnd_find_conn_locked to avoid any conns in the process of being nuked
528 * don't tell LNet if we are in reset - we assume that everyone will be able to
529 * reconnect just fine
531 conn = kgnilnd_find_conn_locked(peer);
533 CDEBUG(D_NETTRACE, "peer 0x%p->%s ting %d conn 0x%p, rst %d error %d\n",
534 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting, conn,
535 kgnilnd_data.kgn_in_reset, error);
537 if (((peer->gnp_connecting == GNILND_PEER_IDLE) &&
539 (!kgnilnd_data.kgn_in_reset) &&
540 (!kgnilnd_conn_clean_errno(error))) || alive) {
544 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
547 /* short circuit if we dont need to notify Lnet */
551 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
554 /* dont do this if this fails since LNET is in shutdown or something else
557 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
558 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
559 /* if gnn_shutdown set for any net shutdown is in progress just return */
560 if (net->gnn_shutdown) {
561 up_read(&kgnilnd_data.kgn_net_rw_sem);
569 /* shutdown in progress most likely */
570 up_read(&kgnilnd_data.kgn_net_rw_sem);
574 CFS_ALLOC_PTR_ARRAY(nets, nnets);
577 up_read(&kgnilnd_data.kgn_net_rw_sem);
578 CERROR("Failed to allocate nets[%d]\n", nnets);
583 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
584 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
586 kgnilnd_net_addref(net);
590 up_read(&kgnilnd_data.kgn_net_rw_sem);
592 for (i = 0; i < nnets; i++) {
597 peer_nid = kgnilnd_lnd2lnetnid(net->gnn_ni->ni_nid,
600 CDEBUG(D_NET, "peer 0x%p->%s last_alive %lld (%llds ago)\n",
601 peer, libcfs_nid2str(peer_nid), peer->gnp_last_alive,
602 ktime_get_seconds() - peer->gnp_last_alive);
604 lnet_notify(net->gnn_ni, peer_nid, alive, true,
605 peer->gnp_last_alive);
607 kgnilnd_net_decref(net);
610 CFS_FREE_PTR_ARRAY(nets, nnets);
614 /* need write_lock on kgn_peer_conn_lock */
616 kgnilnd_close_conn_locked(kgn_conn_t *conn, int error)
618 kgn_peer_t *peer = conn->gnc_peer;
621 LASSERT(!in_interrupt());
623 /* store error for tx completion */
624 conn->gnc_error = error;
625 peer->gnp_last_errno = error;
627 /* use real error from peer if possible */
628 if (error == -ECONNRESET) {
629 error = conn->gnc_peer_error;
632 /* if we NETERROR, make sure it is rate limited */
633 if (!kgnilnd_conn_clean_errno(error) &&
634 peer->gnp_state != GNILND_PEER_DOWN) {
635 CNETERR("closing conn to %s: error %d\n",
636 libcfs_nid2str(peer->gnp_nid), error);
638 CDEBUG(D_NET, "closing conn to %s: error %d\n",
639 libcfs_nid2str(peer->gnp_nid), error);
642 LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
643 "conn %p to %s with bogus state %s\n", conn,
644 libcfs_nid2str(conn->gnc_peer->gnp_nid),
645 kgnilnd_conn_state2str(conn));
646 LASSERT(!list_empty(&conn->gnc_hashlist));
647 LASSERT(!list_empty(&conn->gnc_list));
650 /* mark peer count here so any place the EP gets destroyed will
651 * open up the peer count so that a new ESTABLISHED conn is then free
652 * to send new messages -- sending before the previous EPs are destroyed
653 * could end up with messages on the network for the old conn _after_
654 * the new conn and break the mbox safety protocol */
655 kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
657 /* Remove from conn hash table: no new callbacks */
658 list_del_init(&conn->gnc_hashlist);
659 kgnilnd_data.kgn_conn_version++;
660 kgnilnd_conn_decref(conn);
662 /* if we are in reset, go right to CLOSED as there is no scheduler
663 * thread to move from CLOSING to CLOSED */
664 if (unlikely(kgnilnd_data.kgn_in_reset)) {
665 conn->gnc_state = GNILND_CONN_CLOSED;
667 conn->gnc_state = GNILND_CONN_CLOSING;
670 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RDMA_CQ_ERROR)) {
671 msleep_interruptible(MSEC_PER_SEC);
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));
736 /* We shouldnt be on the delay list, the conn can
737 * get added to this list during a retransmit, and retransmits
738 * only occur within scheduler threads.
740 LASSERT(list_empty(&conn->gnc_delaylist));
742 /* we've sent the close, start nuking */
743 if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
744 kgnilnd_schedule_conn(conn);
746 if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
747 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
748 "done, Attempting to recover conn 0x%p "
749 "scheduled %d function: %s line: %d\n", conn,
750 conn->gnc_scheduled, conn->gnc_sched_caller,
751 conn->gnc_sched_line);
755 /* we don't use lists to track things that we can get out of the
758 /* need to hold locks for tx_list_state, sampling it is too racy:
759 * - the lock actually protects tx != NULL, but we can't take the proper
760 * lock until we check tx_list_state, which would be too late and
761 * we could have the TX change under us.
762 * gnd_rdmaq_lock and gnd_lock and not used together, so taking both
764 spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
765 spin_lock(&conn->gnc_device->gnd_lock);
767 for (nrdma = 0; nrdma < GNILND_MAX_MSG_ID; nrdma++) {
768 tx = conn->gnc_tx_ref_table[nrdma];
771 /* only print the first error and if not CLOSE, we often don't see
772 * CQ events for that by the time we get here... and really don't care */
773 if (nlive || tx->tx_msg.gnm_type == GNILND_MSG_CLOSE)
774 tx->tx_state |= GNILND_TX_QUIET_ERROR;
776 GNIDBG_TX(D_NET, tx, "cleaning up on close, nlive %d", nlive);
778 /* don't worry about gnc_lock here as nobody else should be
779 * touching this conn */
780 kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
781 list_add_tail(&tx->tx_list, &sinners);
784 spin_unlock(&conn->gnc_device->gnd_lock);
785 spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
787 /* nobody should have marked this as needing scheduling after
788 * we called close - so only ref should be us handling it */
789 if (conn->gnc_scheduled != GNILND_CONN_PROCESS) {
790 CDEBUG(D_NETERROR, "Error someone scheduled us after we were "
791 "done, Attempting to recover conn 0x%p "
792 "scheduled %d function %s line: %d\n", conn,
793 conn->gnc_scheduled, conn->gnc_sched_caller,
794 conn->gnc_sched_line);
796 /* now reset a few to actual counters... */
797 nrdma = atomic_read(&conn->gnc_nlive_rdma);
798 nq_rdma = atomic_read(&conn->gnc_nq_rdma);
800 if (!list_empty(&sinners)) {
801 list_for_each_entry_safe(tx, txn, &sinners, tx_list) {
802 /* clear tx_list to make tx_add_list_locked happy */
803 list_del_init(&tx->tx_list);
804 /* The error codes determine if we hold onto the MDD */
805 kgnilnd_tx_done(tx, conn->gnc_error);
809 logmsg = (nlive + nrdma + nq_rdma);
812 int level = conn->gnc_peer->gnp_state == GNILND_PEER_UP ?
814 CDEBUG(level, "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);
821 kgnilnd_destroy_conn_ep(conn);
823 /* Bug 765042 - race this with completing a new conn to same peer - we need
824 * finish_connect to detach purgatory before we can do it ourselves here */
825 CFS_RACE(CFS_FAIL_GNI_FINISH_PURG);
827 /* now it is safe to remove from peer list - anyone looking at
828 * gnp_conns now is free to unlink if not on purgatory */
829 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
831 conn->gnc_state = GNILND_CONN_DONE;
833 /* Decrement counter if we are marked by del_conn_or_peers for closing
835 if (conn->gnc_needs_closing)
836 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_conns);
838 /* Remove from peer's list of valid connections if its not in purgatory */
839 if (!conn->gnc_in_purgatory) {
840 list_del_init(&conn->gnc_list);
841 /* Lose peers reference on the conn */
842 kgnilnd_conn_decref(conn);
845 /* NB - only unlinking if we set pending in del_peer_locked from admin or
847 if (kgnilnd_peer_active(conn->gnc_peer) &&
848 conn->gnc_peer->gnp_pending_unlink &&
849 kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
850 kgnilnd_unlink_peer_locked(conn->gnc_peer);
853 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
855 /* I'm telling Mommy! - use peer_error if they initiated close */
856 kgnilnd_peer_notify(conn->gnc_peer,
857 conn->gnc_error == -ECONNRESET ?
858 conn->gnc_peer_error : conn->gnc_error, 0);
864 kgnilnd_set_conn_params(kgn_dgram_t *dgram)
866 kgn_conn_t *conn = dgram->gndg_conn;
867 kgn_connreq_t *connreq = &dgram->gndg_conn_in;
868 kgn_gniparams_t *rem_param = &connreq->gncr_gnparams;
871 gni_smsg_attr_t *remote = &connreq->gncr_gnparams.gnpr_smsg_attr;
873 /* set timeout vals in conn early so we can use them for the NAK */
875 /* use max of the requested and our timeout, peer will do the same */
876 conn->gnc_timeout = max(conn->gnc_timeout, connreq->gncr_timeout);
878 /* only ep_bind really mucks around with the CQ */
879 /* only ep bind if we are not connecting to ourself and the dstnid is not a wildcard. this check
880 * is necessary as you can only bind an ep once and we must make sure we dont bind when already bound.
882 if (connreq->gncr_dstnid != LNET_NID_ANY && dgram->gndg_conn_out.gncr_dstnid != connreq->gncr_srcnid) {
883 mutex_lock(&conn->gnc_device->gnd_cq_mutex);
884 rrc = kgnilnd_ep_bind(conn->gnc_ephandle,
885 connreq->gncr_gnparams.gnpr_host_id,
887 mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
888 if (rrc != GNI_RC_SUCCESS) {
894 rrc = kgnilnd_ep_set_eventdata(conn->gnc_ephandle, conn->gnc_cqid,
895 connreq->gncr_gnparams.gnpr_cqid);
896 if (rrc != GNI_RC_SUCCESS) {
901 /* Initialize SMSG */
902 rrc = kgnilnd_smsg_init(conn->gnc_ephandle, &conn->gnpr_smsg_attr,
903 &connreq->gncr_gnparams.gnpr_smsg_attr);
904 if (unlikely(rrc == GNI_RC_INVALID_PARAM)) {
905 gni_smsg_attr_t *local = &conn->gnpr_smsg_attr;
906 /* help folks figure out if there is a tunable off, etc. */
907 LCONSOLE_ERROR("SMSG attribute mismatch. Data from local/remote:"
908 " type %d/%d msg_maxsize %u/%u"
909 " mbox_maxcredit %u/%u. Please check kgni"
910 " logs for further data\n",
911 local->msg_type, remote->msg_type,
912 local->msg_maxsize, remote->msg_maxsize,
913 local->mbox_maxcredit, remote->mbox_maxcredit);
915 if (rrc != GNI_RC_SUCCESS) {
920 /* log this for help in debuggin SMSG buffer re-use */
921 CDEBUG(D_NET, "conn %p src %s dst %s smsg %p acquired"
922 " local cqid %u SMSG %p->%u hndl %#llx.%#llx"
923 " remote cqid %u SMSG %p->%u hndl %#llx.%#llx\n",
924 conn, libcfs_nid2str(connreq->gncr_srcnid),
925 libcfs_nid2str(connreq->gncr_dstnid),
926 &conn->gnpr_smsg_attr,
928 conn->gnpr_smsg_attr.msg_buffer,
929 conn->gnpr_smsg_attr.mbox_offset,
930 conn->gnpr_smsg_attr.mem_hndl.qword1,
931 conn->gnpr_smsg_attr.mem_hndl.qword2,
932 rem_param->gnpr_cqid,
933 rem_param->gnpr_smsg_attr.msg_buffer,
934 rem_param->gnpr_smsg_attr.mbox_offset,
935 rem_param->gnpr_smsg_attr.mem_hndl.qword1,
936 rem_param->gnpr_smsg_attr.mem_hndl.qword2);
938 conn->gnc_peerstamp = connreq->gncr_peerstamp;
939 conn->gnc_peer_connstamp = connreq->gncr_connstamp;
940 conn->remote_mbox_addr = (void *)((char *)remote->msg_buffer + remote->mbox_offset);
942 /* We update the reaper timeout once we have a valid conn and timeout */
943 kgnilnd_update_reaper_timeout(GNILND_TO2KA(conn->gnc_timeout));
948 rrc = kgnilnd_ep_unbind(conn->gnc_ephandle);
949 /* not sure I can just let this fly */
950 LASSERTF(rrc == GNI_RC_SUCCESS,
951 "bad rc from gni_ep_unbind trying to cleanup: %d\n", rrc);
954 LASSERTF(rc != 0, "SOFTWARE BUG: rc == 0\n");
955 CERROR("Error setting connection params from %s: %d\n",
956 libcfs_nid2str(connreq->gncr_srcnid), rc);
960 /* needs down_read on kgn_net_rw_sem held from before this call until
961 * after the write_lock on kgn_peer_conn_lock - this ensures we stay sane
962 * with kgnilnd_shutdown - it'll get the sem and set shutdown, then get the
963 * kgn_peer_conn_lock to start del_peer'ing. If we hold the sem until after
964 * kgn_peer_conn_lock is held, we guarantee that nobody calls
965 * kgnilnd_add_peer_locked without checking gnn_shutdown */
967 kgnilnd_create_peer_safe(kgn_peer_t **peerp,
975 LASSERT(nid != LNET_NID_ANY);
977 /* We dont pass the net around in the dgram anymore so here is where we find it
978 * this will work unless its in shutdown or the nid has a net that is invalid.
979 * Either way error code needs to be returned in that case.
981 * If the net passed in is not NULL then we can use it, this alleviates looking it
982 * when the calling function has access to the data.
985 rc = kgnilnd_find_net(nid, &net);
989 /* find net adds a reference on the net if we are not using
990 * it we must do it manually so the net references are
991 * correct when tearing down the net
993 kgnilnd_net_addref(net);
996 LIBCFS_ALLOC(peer, sizeof(*peer));
998 kgnilnd_net_decref(net);
1001 peer->gnp_nid = nid;
1002 peer->gnp_state = node_state;
1004 /* translate from nid to nic addr & store */
1005 rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(nid), 1, &peer->gnp_host_id);
1007 kgnilnd_net_decref(net);
1008 LIBCFS_FREE(peer, sizeof(*peer));
1011 CDEBUG(D_NET, "peer 0x%p->%s -> NIC 0x%x\n", peer,
1012 libcfs_nid2str(nid), peer->gnp_host_id);
1014 atomic_set(&peer->gnp_refcount, 1); /* 1 ref for caller */
1015 atomic_set(&peer->gnp_dirty_eps, 0);
1017 INIT_LIST_HEAD(&peer->gnp_list);
1018 INIT_LIST_HEAD(&peer->gnp_connd_list);
1019 INIT_LIST_HEAD(&peer->gnp_conns);
1020 INIT_LIST_HEAD(&peer->gnp_tx_queue);
1022 /* the first reconnect should happen immediately, so we leave
1023 * gnp_reconnect_interval set to 0 */
1025 LASSERTF(net != NULL, "peer 0x%p->%s with NULL net\n",
1026 peer, libcfs_nid2str(nid));
1028 /* must have kgn_net_rw_sem held for this... */
1029 if (net->gnn_shutdown) {
1030 /* shutdown has started already */
1031 kgnilnd_net_decref(net);
1032 LIBCFS_FREE(peer, sizeof(*peer));
1036 peer->gnp_net = net;
1038 atomic_inc(&kgnilnd_data.kgn_npeers);
1045 kgnilnd_destroy_peer(kgn_peer_t *peer)
1047 CDEBUG(D_NET, "peer %s %p deleted\n",
1048 libcfs_nid2str(peer->gnp_nid), peer);
1049 LASSERTF(atomic_read(&peer->gnp_refcount) == 0,
1050 "peer 0x%p->%s refs %d\n",
1051 peer, libcfs_nid2str(peer->gnp_nid),
1052 atomic_read(&peer->gnp_refcount));
1053 LASSERTF(atomic_read(&peer->gnp_dirty_eps) == 0,
1054 "peer 0x%p->%s dirty eps %d\n",
1055 peer, libcfs_nid2str(peer->gnp_nid),
1056 atomic_read(&peer->gnp_dirty_eps));
1057 LASSERTF(peer->gnp_net != NULL, "peer %p (%s) with NULL net\n",
1058 peer, libcfs_nid2str(peer->gnp_nid));
1059 LASSERTF(!kgnilnd_peer_active(peer),
1061 peer, libcfs_nid2str(peer->gnp_nid));
1062 LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE || peer->gnp_connecting == GNILND_PEER_KILL,
1063 "peer 0x%p->%s, connecting %d\n",
1064 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1065 LASSERTF(list_empty(&peer->gnp_conns),
1067 peer, libcfs_nid2str(peer->gnp_nid));
1068 LASSERTF(list_empty(&peer->gnp_tx_queue),
1070 peer, libcfs_nid2str(peer->gnp_nid));
1071 LASSERTF(list_empty(&peer->gnp_connd_list),
1073 peer, libcfs_nid2str(peer->gnp_nid));
1075 /* NB a peer's connections keep a reference on their peer until
1076 * they are destroyed, so we can be assured that _all_ state to do
1077 * with this peer has been cleaned up when its refcount drops to
1080 atomic_dec(&kgnilnd_data.kgn_npeers);
1081 kgnilnd_net_decref(peer->gnp_net);
1083 LIBCFS_FREE(peer, sizeof(*peer));
1086 /* the conn might not have made it all the way through to a connected
1087 * state - but we need to purgatory any conn that a remote peer might
1088 * have seen through a posted dgram as well */
1090 kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer)
1092 kgn_mbox_info_t *mbox = NULL;
1095 /* NB - the caller should own conn by removing him from the
1096 * scheduler thread when finishing the close */
1098 LASSERTF(peer != NULL, "conn %p with NULL peer\n", conn);
1100 /* If this is still true, need to add the calls to unlink back in and
1101 * figure out how to close the hole on loopback conns */
1102 LASSERTF(kgnilnd_peer_active(peer), "can't use inactive peer %s (%p)"
1103 " we'll never recover the resources\n",
1104 libcfs_nid2str(peer->gnp_nid), peer);
1106 CDEBUG(D_NET, "conn %p peer %p dev %p\n", conn, peer,
1109 LASSERTF(conn->gnc_in_purgatory == 0,
1110 "Conn already in purgatory\n");
1111 conn->gnc_in_purgatory = 1;
1113 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1114 mbox->mbx_prev_purg_nid = peer->gnp_nid;
1115 mbox->mbx_add_purgatory = jiffies;
1116 kgnilnd_release_mbox(conn, 1);
1118 LASSERTF(list_empty(&conn->gnc_mdd_list),
1119 "conn 0x%p->%s with active purgatory hold MDD %d\n",
1120 conn, libcfs_nid2str(peer->gnp_nid),
1121 kgnilnd_count_list(&conn->gnc_mdd_list));
1126 /* Instead of detaching everything from purgatory here we just mark the conn as needing
1127 * detach, when the reaper checks the conn the next time it will detach it.
1128 * Calling function requires write_lock held on kgn_peer_conn_lock
1131 kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer) {
1134 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
1135 if (conn->gnc_in_purgatory && !conn->gnc_needs_detach) {
1136 conn->gnc_needs_detach = 1;
1137 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_detach);
1142 /* Calling function needs a write_lock held on kgn_peer_conn_lock */
1144 kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list)
1146 kgn_mbox_info_t *mbox = NULL;
1148 /* if needed, add the conn purgatory data to the list passed in */
1149 if (conn->gnc_in_purgatory) {
1150 CDEBUG(D_NET, "peer %p->%s purg_conn %p@%s mdd_list #tx %d\n",
1151 conn->gnc_peer, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1152 conn, kgnilnd_conn_state2str(conn),
1153 kgnilnd_count_list(&conn->gnc_mdd_list));
1155 mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
1156 mbox->mbx_detach_of_purgatory = jiffies;
1158 /* conn->gnc_list is the entry point on peer->gnp_conns, so detaching it
1159 * here removes it from the list of 'valid' peer connections.
1160 * We put the current conn onto a list of conns to call kgnilnd_release_purgatory_locked()
1161 * and as such the caller of kgnilnd_detach_purgatory_locked() now owns that conn, since its not
1162 * on the peer's conn_list anymore.
1165 list_del_init(&conn->gnc_list);
1167 /* NB - only unlinking if we set pending in del_peer_locked from admin or
1169 if (kgnilnd_peer_active(conn->gnc_peer) &&
1170 conn->gnc_peer->gnp_pending_unlink &&
1171 kgnilnd_can_unlink_peer_locked(conn->gnc_peer)) {
1172 kgnilnd_unlink_peer_locked(conn->gnc_peer);
1174 /* The reaper will not call detach unless the conn is fully through kgnilnd_complete_closed_conn.
1175 * If the conn is not in a DONE state somehow we are attempting to detach even though
1176 * the conn has not been fully cleaned up. If we detach while the conn is still closing
1177 * we will end up with an orphaned connection that has valid ep_handle, that is not on a
1181 LASSERTF(conn->gnc_state == GNILND_CONN_DONE, "Conn in invalid state %p@%s \n",
1182 conn, kgnilnd_conn_state2str(conn));
1184 /* move from peer to the delayed release list */
1185 list_add_tail(&conn->gnc_list, conn_list);
1190 kgnilnd_release_purgatory_list(struct list_head *conn_list)
1193 kgn_conn_t *conn, *connN;
1194 kgn_mdd_purgatory_t *gmp, *gmpN;
1196 list_for_each_entry_safe(conn, connN, conn_list, gnc_list) {
1197 dev = conn->gnc_device;
1199 kgnilnd_release_mbox(conn, -1);
1200 conn->gnc_in_purgatory = 0;
1202 list_del_init(&conn->gnc_list);
1204 /* gnc_needs_detach is set in kgnilnd_del_conn_or_peer. It is used to keep track
1205 * of conns that have been marked for detach by kgnilnd_del_conn_or_peer.
1206 * The function uses kgn_npending_detach to verify the conn has
1207 * actually been detached.
1210 if (conn->gnc_needs_detach)
1211 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_detach);
1213 /* if this guy is really dead (we are doing release from reaper),
1214 * make sure we tell LNet - if this is from other context,
1215 * the checks in the function will prevent an errant
1217 kgnilnd_peer_notify(conn->gnc_peer, conn->gnc_error, 0);
1219 list_for_each_entry_safe(gmp, gmpN, &conn->gnc_mdd_list,
1222 "dev %p releasing held mdd %#llx.%#llx\n",
1223 conn->gnc_device, gmp->gmp_map_key.qword1,
1224 gmp->gmp_map_key.qword2);
1226 atomic_dec(&dev->gnd_n_mdd_held);
1227 kgnilnd_mem_mdd_release(conn->gnc_device->gnd_handle,
1229 /* ignoring the return code - if kgni/ghal can't find it
1230 * it must be released already */
1232 list_del_init(&gmp->gmp_list);
1233 LIBCFS_FREE(gmp, sizeof(*gmp));
1235 /* lose conn ref for purgatory */
1236 kgnilnd_conn_decref(conn);
1240 /* needs write_lock on kgnilnd_data.kgn_peer_conn_lock held */
1242 kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer)
1246 current_to = peer->gnp_reconnect_interval;
1248 /* we'll try to reconnect fast the first time, then back-off */
1249 if (current_to == 0) {
1250 peer->gnp_reconnect_time = jiffies - 1;
1251 current_to = *kgnilnd_tunables.kgn_min_reconnect_interval;
1253 peer->gnp_reconnect_time = jiffies + cfs_time_seconds(current_to);
1254 /* add 50% of min timeout & retry */
1255 current_to += *kgnilnd_tunables.kgn_min_reconnect_interval / 2;
1258 current_to = min(current_to,
1259 *kgnilnd_tunables.kgn_max_reconnect_interval);
1261 peer->gnp_reconnect_interval = current_to;
1262 CDEBUG(D_NET, "peer %s can reconnect at %lu interval %lu\n",
1263 libcfs_nid2str(peer->gnp_nid), peer->gnp_reconnect_time,
1264 peer->gnp_reconnect_interval);
1267 /* needs kgnilnd_data.kgn_peer_conn_lock held */
1269 kgnilnd_find_peer_locked(lnet_nid_t nid)
1271 struct list_head *peer_list = kgnilnd_nid2peerlist(nid);
1274 /* Chopping nid down to only NIDADDR using LNET_NIDADDR so we only
1275 * have a single peer per device instead of a peer per nid/net combo.
1278 list_for_each_entry(peer, peer_list, gnp_list) {
1279 if (LNET_NIDADDR(nid) != LNET_NIDADDR(peer->gnp_nid))
1282 CDEBUG(D_NET, "got peer [%p] -> %s c %d (%d)\n",
1283 peer, libcfs_nid2str(nid),
1284 peer->gnp_connecting,
1285 atomic_read(&peer->gnp_refcount));
1291 /* need write_lock on kgn_peer_conn_lock */
1293 kgnilnd_unlink_peer_locked(kgn_peer_t *peer)
1295 LASSERTF(list_empty(&peer->gnp_conns),
1297 peer, libcfs_nid2str(peer->gnp_nid));
1298 LASSERTF(list_empty(&peer->gnp_tx_queue),
1300 peer, libcfs_nid2str(peer->gnp_nid));
1301 LASSERTF(kgnilnd_peer_active(peer),
1303 peer, libcfs_nid2str(peer->gnp_nid));
1304 CDEBUG(D_NET, "unlinking peer 0x%p->%s\n",
1305 peer, libcfs_nid2str(peer->gnp_nid));
1307 list_del_init(&peer->gnp_list);
1308 kgnilnd_data.kgn_peer_version++;
1309 kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
1310 /* lose peerlist's ref */
1311 kgnilnd_peer_decref(peer);
1315 kgnilnd_get_peer_info(int index,
1316 kgn_peer_t **found_peer,
1317 lnet_nid_t *id, __u32 *nic_addr,
1318 int *refcount, int *connecting)
1324 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1326 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1327 list_for_each_entry(peer, &kgnilnd_data.kgn_peers[i], gnp_list) {
1331 CDEBUG(D_NET, "found peer %p (%s) at index %d\n",
1332 peer, libcfs_nid2str(peer->gnp_nid), index);
1335 *id = peer->gnp_nid;
1336 *nic_addr = peer->gnp_host_id;
1337 *refcount = atomic_read(&peer->gnp_refcount);
1338 *connecting = peer->gnp_connecting;
1345 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1347 CDEBUG(D_NET, "no gni peer at index %d\n", index);
1351 /* requires write_lock on kgn_peer_conn_lock held */
1353 kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp)
1355 kgn_peer_t *peer, *peer2;
1357 LASSERTF(new_stub_peer != NULL, "bad stub peer for nid %s\n",
1358 libcfs_nid2str(nid));
1360 peer2 = kgnilnd_find_peer_locked(nid);
1361 if (peer2 != NULL) {
1362 /* A peer was created during the lock transition, so drop
1363 * the new one we created */
1364 kgnilnd_peer_decref(new_stub_peer);
1367 peer = new_stub_peer;
1368 /* peer table takes existing ref on peer */
1370 LASSERTF(!kgnilnd_peer_active(peer),
1371 "peer 0x%p->%s already in peer table\n",
1372 peer, libcfs_nid2str(peer->gnp_nid));
1373 list_add_tail(&peer->gnp_list,
1374 kgnilnd_nid2peerlist(nid));
1375 kgnilnd_data.kgn_peer_version++;
1378 LASSERTF(peer->gnp_net != NULL, "peer 0x%p->%s with NULL net\n",
1379 peer, libcfs_nid2str(peer->gnp_nid));
1384 kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp)
1391 if (nid == LNET_NID_ANY)
1394 node_state = kgnilnd_get_node_state(LNET_NIDADDR(nid));
1396 /* NB - this will not block during normal operations -
1397 * the only writer of this is in the startup/shutdown path. */
1398 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1403 rc = kgnilnd_create_peer_safe(&peer, nid, net, node_state);
1405 up_read(&kgnilnd_data.kgn_net_rw_sem);
1409 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1410 up_read(&kgnilnd_data.kgn_net_rw_sem);
1412 kgnilnd_add_peer_locked(nid, peer, peerp);
1414 CDEBUG(D_NET, "peer 0x%p->%s connecting %d\n",
1415 peerp, libcfs_nid2str((*peerp)->gnp_nid),
1416 (*peerp)->gnp_connecting);
1418 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1422 /* needs write_lock on kgn_peer_conn_lock */
1424 kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies)
1428 /* we do care about state of gnp_connecting - we could be between
1429 * reconnect attempts, so try to find the dgram and cancel the TX
1430 * anyways. If we are in the process of posting DONT do anything;
1431 * once it fails or succeeds we can nuke the connect attempt.
1432 * We have no idea where in kgnilnd_post_dgram we are so we cant
1433 * attempt to cancel until the function is done.
1436 /* make sure peer isn't in process of connecting or waiting for connect*/
1437 spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1438 if (!(list_empty(&peer->gnp_connd_list))) {
1439 list_del_init(&peer->gnp_connd_list);
1440 /* remove connd ref */
1441 kgnilnd_peer_decref(peer);
1443 spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
1445 if (peer->gnp_connecting == GNILND_PEER_POSTING || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
1446 peer->gnp_connecting = GNILND_PEER_NEEDS_DEATH;
1447 /* We are in process of posting right now the xchg set it up for us to
1448 * cancel the connect so we are finished for now */
1450 /* no need for exchange we have the peer lock and its ready for us to nuke */
1451 LASSERTF(peer->gnp_connecting != GNILND_PEER_POSTING,
1452 "Peer in invalid state 0x%p->%s, connecting %d\n",
1453 peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
1454 peer->gnp_connecting = GNILND_PEER_IDLE;
1455 set_mb(peer->gnp_last_dgram_errno, -ETIMEDOUT);
1456 kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
1460 /* The least we can do is nuke the tx's no matter what.... */
1461 list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
1462 kgnilnd_tx_del_state_locked(tx, peer, NULL,
1464 list_add_tail(&tx->tx_list, zombies);
1468 /* needs write_lock on kgn_peer_conn_lock */
1470 kgnilnd_del_peer_locked(kgn_peer_t *peer, int error)
1472 /* this peer could be passive and only held for purgatory,
1473 * take a ref to ensure it doesn't disappear in this function */
1474 kgnilnd_peer_addref(peer);
1476 CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
1478 /* if purgatory release cleared it out, don't try again */
1479 if (kgnilnd_peer_active(peer)) {
1480 /* always do this to allow kgnilnd_start_connect and
1481 * kgnilnd_finish_connect to catch this before they
1482 * wrap up their operations */
1483 if (kgnilnd_can_unlink_peer_locked(peer)) {
1484 /* already released purgatory, so only active
1486 kgnilnd_unlink_peer_locked(peer);
1488 kgnilnd_close_peer_conns_locked(peer, error);
1489 /* peer unlinks itself when last conn is closed */
1493 /* we are done, release back to the wild */
1494 kgnilnd_peer_decref(peer);
1498 kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command,
1502 LIST_HEAD (zombies);
1503 kgn_peer_t *peer, *pnxt;
1509 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1511 if (nid != LNET_NID_ANY)
1512 lo = hi = kgnilnd_nid2peerlist(nid) - kgnilnd_data.kgn_peers;
1515 hi = *kgnilnd_tunables.kgn_peer_hash_size - 1;
1516 /* wildcards always succeed */
1520 for (i = lo; i <= hi; i++) {
1521 list_for_each_entry_safe(peer, pnxt, &kgnilnd_data.kgn_peers[i],
1523 LASSERTF(peer->gnp_net != NULL,
1524 "peer %p (%s) with NULL net\n",
1525 peer, libcfs_nid2str(peer->gnp_nid));
1527 if (net != NULL && peer->gnp_net != net)
1530 if (!(nid == LNET_NID_ANY || LNET_NIDADDR(peer->gnp_nid) == LNET_NIDADDR(nid)))
1533 /* In both cases, we want to stop any in-flight
1534 * connect attempts */
1535 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1538 case GNILND_DEL_CONN:
1539 kgnilnd_close_peer_conns_locked(peer, error);
1541 case GNILND_DEL_PEER:
1542 peer->gnp_pending_unlink = 1;
1543 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_unlink);
1544 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1545 kgnilnd_del_peer_locked(peer, error);
1547 case GNILND_CLEAR_PURGATORY:
1548 /* Mark everything ready for detach reaper will cleanup
1549 * once we release the kgn_peer_conn_lock
1551 kgnilnd_mark_for_detach_purgatory_all_locked(peer);
1552 peer->gnp_last_errno = -EISCONN;
1553 /* clear reconnect so he can reconnect soon */
1554 peer->gnp_reconnect_time = 0;
1555 peer->gnp_reconnect_interval = 0;
1558 CERROR("bad command %d\n", command);
1561 /* we matched something */
1566 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1569 kgnilnd_txlist_done(&zombies, error);
1571 /* This function does not return until the commands it initiated have completed,
1572 * since they have to work there way through the other threads. In the case of shutdown
1573 * threads are not woken up until after this call is initiated so we cannot wait, we just
1574 * need to return. The same applies for stack reset we shouldnt wait as the reset thread
1578 CFS_RACE(CFS_FAIL_GNI_RACE_RESET);
1580 if (error == -ENOTRECOVERABLE || error == -ESHUTDOWN) {
1584 wait_var_event_warning(&kgnilnd_data,
1585 !atomic_read(&kgnilnd_data.kgn_npending_conns) &&
1586 !atomic_read(&kgnilnd_data.kgn_npending_detach) &&
1587 !atomic_read(&kgnilnd_data.kgn_npending_unlink),
1588 "Waiting on %d peers %d closes %d detaches\n",
1589 atomic_read(&kgnilnd_data.kgn_npending_unlink),
1590 atomic_read(&kgnilnd_data.kgn_npending_conns),
1591 atomic_read(&kgnilnd_data.kgn_npending_detach));
1597 kgnilnd_get_conn_by_idx(int index)
1604 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
1605 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1606 list_for_each_entry(peer, &kgnilnd_data.kgn_peers[i], gnp_list) {
1607 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
1608 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1614 CDEBUG(D_NET, "++conn[%p] -> %s (%d)\n", conn,
1615 libcfs_nid2str(conn->gnc_peer->gnp_nid),
1616 atomic_read(&conn->gnc_refcount));
1617 kgnilnd_conn_addref(conn);
1618 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1622 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1629 kgnilnd_get_conn_info(kgn_peer_t *peer,
1630 int *device_id, __u64 *peerstamp,
1631 int *tx_seq, int *rx_seq,
1632 int *fmaq_len, int *nfma, int *nrdma)
1637 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1639 conn = kgnilnd_find_conn_locked(peer);
1645 *device_id = conn->gnc_device->gnd_host_id;
1646 *peerstamp = conn->gnc_peerstamp;
1647 *tx_seq = atomic_read(&conn->gnc_tx_seq);
1648 *rx_seq = atomic_read(&conn->gnc_rx_seq);
1649 *fmaq_len = kgnilnd_count_list(&conn->gnc_fmaq);
1650 *nfma = atomic_read(&conn->gnc_nlive_fma);
1651 *nrdma = atomic_read(&conn->gnc_nlive_rdma);
1653 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1657 /* needs write_lock on kgn_peer_conn_lock */
1659 kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why)
1662 struct list_head *ctmp, *cnxt;
1665 list_for_each_safe(ctmp, cnxt, &peer->gnp_conns) {
1666 conn = list_entry(ctmp, kgn_conn_t, gnc_list);
1668 if (conn->gnc_state != GNILND_CONN_ESTABLISHED)
1672 /* we mark gnc_needs closing and increment kgn_npending_conns so that
1673 * kgnilnd_del_conn_or_peer can wait on the other threads closing
1674 * and cleaning up the connection.
1676 if (!conn->gnc_needs_closing) {
1677 conn->gnc_needs_closing = 1;
1678 kgnilnd_admin_addref(kgnilnd_data.kgn_npending_conns);
1680 kgnilnd_close_conn_locked(conn, why);
1686 kgnilnd_report_node_state(lnet_nid_t nid, int down)
1689 kgn_peer_t *peer, *new_peer;
1692 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1693 peer = kgnilnd_find_peer_locked(nid);
1700 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1702 /* Don't add a peer for node up events */
1703 if (down == GNILND_PEER_UP)
1706 /* find any valid net - we don't care which one... */
1707 down_read(&kgnilnd_data.kgn_net_rw_sem);
1708 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
1709 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i],
1719 up_read(&kgnilnd_data.kgn_net_rw_sem);
1722 CNETERR("Could not find a net for nid %lld\n", nid);
1726 /* The nid passed in does not yet contain the net portion.
1727 * Let's build it up now
1729 nid = LNET_MKNID(LNET_NIDNET(net->gnn_ni->ni_nid), nid);
1730 rc = kgnilnd_add_peer(net, nid, &new_peer);
1733 CNETERR("Could not add peer for nid %lld, rc %d\n",
1738 write_lock(&kgnilnd_data.kgn_peer_conn_lock);
1739 peer = kgnilnd_find_peer_locked(nid);
1742 CNETERR("Could not find peer for nid %lld\n", nid);
1743 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1748 peer->gnp_state = down;
1750 if (down == GNILND_PEER_DOWN) {
1753 peer->gnp_down_event_time = jiffies;
1754 kgnilnd_cancel_peer_connect_locked(peer, &zombies);
1755 conn = kgnilnd_find_conn_locked(peer);
1758 kgnilnd_close_conn_locked(conn, -ENETRESET);
1761 peer->gnp_up_event_time = jiffies;
1764 write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1766 if (down == GNILND_PEER_DOWN) {
1767 /* using ENETRESET so we don't get messages from
1770 kgnilnd_txlist_done(&zombies, -ENETRESET);
1771 kgnilnd_peer_notify(peer, -ECONNRESET, 0);
1772 LCONSOLE_INFO("Received down event for nid %d\n",
1780 kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1782 struct libcfs_ioctl_data *data = arg;
1783 kgn_net_t *net = ni->ni_data;
1786 LASSERT(ni == net->gnn_ni);
1789 case IOC_LIBCFS_GET_PEER: {
1791 kgn_peer_t *peer = NULL;
1793 __u64 peerstamp = 0;
1794 int peer_refcount = 0, peer_connecting = 0;
1796 int tx_seq = 0, rx_seq = 0;
1797 int fmaq_len = 0, nfma = 0, nrdma = 0;
1799 rc = kgnilnd_get_peer_info(data->ioc_count, &peer,
1800 &nid, &nic_addr, &peer_refcount,
1806 /* LNET_MKNID is used to mask from lnet the multiplexing/demultiplexing of connections and peers
1807 * LNET assumes a conn and peer per net, the LNET_MKNID/LNET_NIDADDR allows us to let Lnet see what it
1808 * wants to see instead of the underlying network that is being used to send the data
1810 data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(nid));
1811 data->ioc_flags = peer_connecting;
1812 data->ioc_count = peer_refcount;
1814 rc = kgnilnd_get_conn_info(peer, &device_id, &peerstamp,
1815 &tx_seq, &rx_seq, &fmaq_len,
1818 /* This is allowable - a persistent peer could not
1819 * have a connection */
1821 /* flag to indicate we are not connected -
1822 * need to print as such */
1823 data->ioc_flags |= (1<<16);
1827 data->ioc_net = device_id;
1828 data->ioc_u64[0] = peerstamp;
1829 data->ioc_u32[0] = fmaq_len;
1830 data->ioc_u32[1] = nfma;
1831 data->ioc_u32[2] = tx_seq;
1832 data->ioc_u32[3] = rx_seq;
1833 data->ioc_u32[4] = nrdma;
1837 case IOC_LIBCFS_ADD_PEER: {
1838 /* just dummy value to allow using common interface */
1840 rc = kgnilnd_add_peer(net, data->ioc_nid, &peer);
1843 case IOC_LIBCFS_DEL_PEER: {
1844 /* NULL is passed in so it affects all peers in existence without regard to network
1845 * as the peer may not exist on the network LNET believes it to be on.
1847 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1848 GNILND_DEL_PEER, -EUCLEAN);
1851 case IOC_LIBCFS_GET_CONN: {
1852 kgn_conn_t *conn = kgnilnd_get_conn_by_idx(data->ioc_count);
1858 /* LNET_MKNID is used to build the correct address based on what LNET wants to see instead of
1859 * the generic connection that is used to send the data
1861 data->ioc_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), LNET_NIDADDR(conn->gnc_peer->gnp_nid));
1862 data->ioc_u32[0] = conn->gnc_device->gnd_id;
1863 kgnilnd_conn_decref(conn);
1867 case IOC_LIBCFS_CLOSE_CONNECTION: {
1868 /* use error = -ENETRESET to indicate it was lctl disconnect */
1869 /* NULL is passed in so it affects all the nets as the connection is virtual
1870 * and may not exist on the network LNET believes it to be on.
1872 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1873 GNILND_DEL_CONN, -ENETRESET);
1876 case IOC_LIBCFS_PUSH_CONNECTION: {
1877 /* we use this to flush purgatory */
1878 rc = kgnilnd_del_conn_or_peer(NULL, data->ioc_nid,
1879 GNILND_CLEAR_PURGATORY, -EUCLEAN);
1882 case IOC_LIBCFS_REGISTER_MYNID: {
1883 /* Ignore if this is a noop */
1884 if (data->ioc_nid == ni->ni_nid) {
1887 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1888 libcfs_nid2str(data->ioc_nid),
1889 libcfs_nid2str(ni->ni_nid));
1900 kgnilnd_dev_init(kgn_device_t *dev)
1904 unsigned int cq_size;
1907 /* size of these CQs should be able to accommodate the outgoing
1908 * RDMA and SMSG transactions. Since we really don't know what we
1909 * really need here, we'll take credits * 2 * 3 to allow a bunch.
1910 * We need to dig into this more with the performance work. */
1911 cq_size = *kgnilnd_tunables.kgn_credits * 2 * 3;
1913 rrc = kgnilnd_cdm_create(dev->gnd_id, *kgnilnd_tunables.kgn_ptag,
1914 *kgnilnd_tunables.kgn_pkey, 0,
1916 if (rrc != GNI_RC_SUCCESS) {
1917 CERROR("Can't create CDM %d (%d)\n", dev->gnd_id, rrc);
1918 GOTO(failed, rc = -ENODEV);
1921 rrc = kgnilnd_cdm_attach(dev->gnd_domain, dev->gnd_id,
1922 &dev->gnd_host_id, &dev->gnd_handle);
1923 if (rrc != GNI_RC_SUCCESS) {
1924 CERROR("Can't attach CDM to device %d (%d)\n",
1926 GOTO(failed, rc = -ENODEV);
1929 /* a bit gross, but not much we can do - Aries Sim doesn't have
1930 * hardcoded NIC/NID that we can use */
1931 rc = kgnilnd_setup_nic_translation(dev->gnd_host_id);
1933 GOTO(failed, rc = -ENODEV);
1935 /* only dev 0 gets the errors - no need to reset the stack twice
1936 * - this works because we have a single PTAG, if we had more
1937 * then we'd need to have multiple handlers */
1938 if (dev->gnd_id == 0) {
1939 rrc = kgnilnd_subscribe_errors(dev->gnd_handle,
1940 GNI_ERRMASK_CRITICAL |
1941 GNI_ERRMASK_UNKNOWN_TRANSACTION,
1942 0, NULL, kgnilnd_critical_error,
1943 &dev->gnd_err_handle);
1944 if (rrc != GNI_RC_SUCCESS) {
1945 CERROR("Can't subscribe for errors on device %d: rc %d\n",
1947 GOTO(failed, rc = -ENODEV);
1950 rc = kgnilnd_set_quiesce_callback(dev->gnd_handle,
1951 kgnilnd_quiesce_end_callback);
1952 if (rc != GNI_RC_SUCCESS) {
1953 CERROR("Can't subscribe for quiesce callback on device %d: rc %d\n",
1955 GOTO(failed, rc = -ENODEV);
1959 rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
1961 /* log messages during startup */
1962 if (kgnilnd_data.kgn_init < GNILND_INIT_ALL) {
1963 CERROR("couldn't translate host_id 0x%x to nid. rc %d\n",
1964 dev->gnd_host_id, rc);
1966 GOTO(failed, rc = -ESRCH);
1968 CDEBUG(D_NET, "NIC %x -> NID %d\n", dev->gnd_host_id, dev->gnd_nid);
1970 rrc = kgnilnd_cq_create(dev->gnd_handle, *kgnilnd_tunables.kgn_credits,
1971 0, kgnilnd_device_callback,
1972 dev->gnd_id, &dev->gnd_snd_rdma_cqh);
1973 if (rrc != GNI_RC_SUCCESS) {
1974 CERROR("Can't create rdma send cq size %u for device "
1975 "%d (%d)\n", cq_size, dev->gnd_id, rrc);
1976 GOTO(failed, rc = -EINVAL);
1979 rrc = kgnilnd_cq_create(dev->gnd_handle, cq_size,
1980 0, kgnilnd_device_callback, dev->gnd_id,
1981 &dev->gnd_snd_fma_cqh);
1982 if (rrc != GNI_RC_SUCCESS) {
1983 CERROR("Can't create fma send cq size %u for device %d (%d)\n",
1984 cq_size, dev->gnd_id, rrc);
1985 GOTO(failed, rc = -EINVAL);
1988 /* This one we size differently - overflows are possible and it needs to be
1989 * sized based on machine size */
1990 rrc = kgnilnd_cq_create(dev->gnd_handle,
1991 *kgnilnd_tunables.kgn_fma_cq_size,
1992 0, kgnilnd_device_callback, dev->gnd_id,
1993 &dev->gnd_rcv_fma_cqh);
1994 if (rrc != GNI_RC_SUCCESS) {
1995 CERROR("Can't create fma cq size %d for device %d (%d)\n",
1996 *kgnilnd_tunables.kgn_fma_cq_size, dev->gnd_id, rrc);
1997 GOTO(failed, rc = -EINVAL);
2000 rrc = kgnilnd_register_smdd_buf(dev);
2001 if (rrc != GNI_RC_SUCCESS) {
2002 GOTO(failed, rc = -EINVAL);
2008 kgnilnd_dev_fini(dev);
2013 kgnilnd_dev_fini(kgn_device_t *dev)
2018 /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
2019 LASSERTF(list_empty(&dev->gnd_ready_conns) &&
2020 list_empty(&dev->gnd_map_tx) &&
2021 list_empty(&dev->gnd_rdmaq) &&
2022 list_empty(&dev->gnd_delay_conns),
2023 "dev 0x%p ready_conns %d@0x%p delay_conns %d@0x%p"
2024 "map_tx %d@0x%p rdmaq %d@0x%p\n",
2025 dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
2026 kgnilnd_count_list(&dev->gnd_delay_conns), &dev->gnd_delay_conns,
2027 kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
2028 kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
2030 /* These should follow from tearing down all connections */
2031 LASSERTF(dev->gnd_map_nphys == 0 && dev->gnd_map_physnop == 0,
2032 "%d physical mappings of %d pages still mapped\n",
2033 dev->gnd_map_nphys, dev->gnd_map_physnop);
2035 LASSERTF(atomic_read(&dev->gnd_n_mdd) == 0 &&
2036 atomic_read(&dev->gnd_n_mdd_held) == 0 &&
2037 atomic64_read(&dev->gnd_nbytes_map) == 0,
2038 "%d SMSG mappings of %lld bytes still mapped or held %d\n",
2039 atomic_read(&dev->gnd_n_mdd),
2040 (u64)atomic64_read(&dev->gnd_nbytes_map),
2041 atomic_read(&dev->gnd_n_mdd_held));
2043 LASSERT(list_empty(&dev->gnd_map_list));
2045 /* What other assertions needed to ensure all connections torn down ? */
2047 /* check all counters == 0 (EP, MDD, etc) */
2049 /* if we are resetting due to quiese (stack reset), don't check
2051 LASSERTF(kgnilnd_data.kgn_quiesce_trigger ||
2052 atomic_read(&kgnilnd_data.kgn_nthreads) == 0,
2053 "tried to shutdown with threads active\n");
2055 if (dev->gnd_smdd_hold_buf) {
2056 rrc = kgnilnd_deregister_smdd_buf(dev);
2057 LASSERTF(rrc == GNI_RC_SUCCESS,
2058 "bad rc from deregistion of sMDD buffer: %d\n", rrc);
2059 dev->gnd_smdd_hold_buf = NULL;
2062 if (dev->gnd_rcv_fma_cqh) {
2063 rrc = kgnilnd_cq_destroy(dev->gnd_rcv_fma_cqh);
2064 LASSERTF(rrc == GNI_RC_SUCCESS,
2065 "bad rc from gni_cq_destroy on rcv_fma_cqh: %d\n", rrc);
2066 dev->gnd_rcv_fma_cqh = NULL;
2069 if (dev->gnd_snd_rdma_cqh) {
2070 rrc = kgnilnd_cq_destroy(dev->gnd_snd_rdma_cqh);
2071 LASSERTF(rrc == GNI_RC_SUCCESS,
2072 "bad rc from gni_cq_destroy on send_rdma_cqh: %d\n", rrc);
2073 dev->gnd_snd_rdma_cqh = NULL;
2076 if (dev->gnd_snd_fma_cqh) {
2077 rrc = kgnilnd_cq_destroy(dev->gnd_snd_fma_cqh);
2078 LASSERTF(rrc == GNI_RC_SUCCESS,
2079 "bad rc from gni_cq_destroy on snd_fma_cqh: %d\n", rrc);
2080 dev->gnd_snd_fma_cqh = NULL;
2083 if (dev->gnd_err_handle) {
2084 rrc = kgnilnd_release_errors(dev->gnd_err_handle);
2085 LASSERTF(rrc == GNI_RC_SUCCESS,
2086 "bad rc from gni_release_errors: %d\n", rrc);
2087 dev->gnd_err_handle = NULL;
2090 if (dev->gnd_domain) {
2091 rrc = kgnilnd_cdm_destroy(dev->gnd_domain);
2092 LASSERTF(rrc == GNI_RC_SUCCESS,
2093 "bad rc from gni_cdm_destroy: %d\n", rrc);
2094 dev->gnd_domain = NULL;
2100 int kgnilnd_base_startup(void)
2102 struct timespec64 ts;
2103 long long pkmem = libcfs_kmem_read();
2107 struct task_struct *thrd;
2109 #if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
2110 /* limit how much memory can be allocated for fma blocks in
2111 * instances where many nodes need to reconnects at the same time */
2114 kgnilnd_data.free_pages_limit = si.totalram/4;
2119 LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_NOTHING,
2120 "init %d\n", kgnilnd_data.kgn_init);
2122 /* zero pointers, flags etc */
2123 memset(&kgnilnd_data, 0, sizeof(kgnilnd_data));
2124 kgnilnd_check_kgni_version();
2126 /* CAVEAT EMPTOR: Every 'Fma' message includes the sender's NID and
2127 * a unique (for all time) connstamp so we can uniquely identify
2128 * the sender. The connstamp is an incrementing counter
2129 * initialised with seconds + microseconds at startup time. So we
2130 * rely on NOT creating connections more frequently on average than
2131 * 1MHz to ensure we don't use old connstamps when we reboot. */
2132 ktime_get_ts64(&ts);
2133 kgnilnd_data.kgn_connstamp =
2134 kgnilnd_data.kgn_peerstamp =
2135 (ts.tv_sec * 1000000) + (ts.tv_nsec / 100);
2137 init_rwsem(&kgnilnd_data.kgn_net_rw_sem);
2139 for (i = 0; i < GNILND_MAXDEVS; i++) {
2140 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2143 INIT_LIST_HEAD(&dev->gnd_ready_conns);
2144 INIT_LIST_HEAD(&dev->gnd_delay_conns);
2145 INIT_LIST_HEAD(&dev->gnd_map_tx);
2146 INIT_LIST_HEAD(&dev->gnd_fma_buffs);
2147 mutex_init(&dev->gnd_cq_mutex);
2148 mutex_init(&dev->gnd_fmablk_mutex);
2149 spin_lock_init(&dev->gnd_fmablk_lock);
2150 init_waitqueue_head(&dev->gnd_waitq);
2151 init_waitqueue_head(&dev->gnd_dgram_waitq);
2152 init_waitqueue_head(&dev->gnd_dgping_waitq);
2153 spin_lock_init(&dev->gnd_lock);
2154 INIT_LIST_HEAD(&dev->gnd_map_list);
2155 spin_lock_init(&dev->gnd_map_lock);
2156 atomic_set(&dev->gnd_nfmablk, 0);
2157 atomic_set(&dev->gnd_fmablk_vers, 1);
2158 atomic_set(&dev->gnd_neps, 0);
2159 atomic_set(&dev->gnd_canceled_dgrams, 0);
2160 INIT_LIST_HEAD(&dev->gnd_connd_peers);
2161 spin_lock_init(&dev->gnd_connd_lock);
2162 spin_lock_init(&dev->gnd_dgram_lock);
2163 spin_lock_init(&dev->gnd_rdmaq_lock);
2164 INIT_LIST_HEAD(&dev->gnd_rdmaq);
2165 init_rwsem(&dev->gnd_conn_sem);
2167 /* alloc & setup nid based dgram table */
2168 CFS_ALLOC_PTR_ARRAY(dev->gnd_dgrams,
2169 *kgnilnd_tunables.kgn_peer_hash_size);
2171 if (dev->gnd_dgrams == NULL)
2172 GOTO(failed, rc = -ENOMEM);
2174 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2175 INIT_LIST_HEAD(&dev->gnd_dgrams[i]);
2177 atomic_set(&dev->gnd_ndgrams, 0);
2178 atomic_set(&dev->gnd_nwcdgrams, 0);
2179 /* setup timer for RDMAQ processing */
2180 cfs_timer_setup(&dev->gnd_rdmaq_timer,
2181 kgnilnd_schedule_device_timer,
2182 (unsigned long)dev, 0);
2184 /* setup timer for mapping processing */
2185 cfs_timer_setup(&dev->gnd_map_timer,
2186 kgnilnd_schedule_device_timer,
2187 (unsigned long)dev, 0);
2191 /* CQID 0 isn't allowed, set to MAX_MSG_ID - 1 to check for conflicts early */
2192 kgnilnd_data.kgn_next_cqid = GNILND_MAX_MSG_ID - 1;
2193 kgnilnd_data.kgn_new_min_timeout = *kgnilnd_tunables.kgn_timeout;
2194 init_waitqueue_head(&kgnilnd_data.kgn_reaper_waitq);
2195 init_waitqueue_head(&kgnilnd_data.kgn_ruhroh_waitq);
2196 spin_lock_init(&kgnilnd_data.kgn_reaper_lock);
2198 mutex_init(&kgnilnd_data.kgn_quiesce_mutex);
2199 atomic_set(&kgnilnd_data.kgn_nquiesce, 0);
2200 atomic_set(&kgnilnd_data.kgn_npending_conns, 0);
2201 atomic_set(&kgnilnd_data.kgn_npending_unlink, 0);
2202 atomic_set(&kgnilnd_data.kgn_npending_detach, 0);
2203 atomic_set(&kgnilnd_data.kgn_rev_offset, 0);
2204 atomic_set(&kgnilnd_data.kgn_rev_length, 0);
2205 atomic_set(&kgnilnd_data.kgn_rev_copy_buff, 0);
2207 /* OK to call kgnilnd_api_shutdown() to cleanup now */
2208 kgnilnd_data.kgn_init = GNILND_INIT_DATA;
2209 if (!try_module_get(THIS_MODULE))
2210 GOTO(failed, rc = -ENOENT);
2212 rwlock_init(&kgnilnd_data.kgn_peer_conn_lock);
2214 CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_peers,
2215 *kgnilnd_tunables.kgn_peer_hash_size);
2217 if (kgnilnd_data.kgn_peers == NULL)
2218 GOTO(failed, rc = -ENOMEM);
2220 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2221 INIT_LIST_HEAD(&kgnilnd_data.kgn_peers[i]);
2224 CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_conns,
2225 *kgnilnd_tunables.kgn_peer_hash_size);
2227 if (kgnilnd_data.kgn_conns == NULL)
2228 GOTO(failed, rc = -ENOMEM);
2230 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
2231 INIT_LIST_HEAD(&kgnilnd_data.kgn_conns[i]);
2234 CFS_ALLOC_PTR_ARRAY(kgnilnd_data.kgn_nets,
2235 *kgnilnd_tunables.kgn_net_hash_size);
2237 if (kgnilnd_data.kgn_nets == NULL)
2238 GOTO(failed, rc = -ENOMEM);
2240 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2241 INIT_LIST_HEAD(&kgnilnd_data.kgn_nets[i]);
2244 kgnilnd_data.kgn_mbox_cache =
2245 kmem_cache_create("kgn_mbox_block", GNILND_MBOX_SIZE, 0,
2246 SLAB_HWCACHE_ALIGN, NULL);
2247 if (kgnilnd_data.kgn_mbox_cache == NULL) {
2248 CERROR("Can't create slab for physical mbox blocks\n");
2249 GOTO(failed, rc = -ENOMEM);
2252 kgnilnd_data.kgn_rx_cache =
2253 kmem_cache_create("kgn_rx_t", sizeof(kgn_rx_t), 0, 0, NULL);
2254 if (kgnilnd_data.kgn_rx_cache == NULL) {
2255 CERROR("Can't create slab for kgn_rx_t descriptors\n");
2256 GOTO(failed, rc = -ENOMEM);
2259 kgnilnd_data.kgn_tx_cache =
2260 kmem_cache_create("kgn_tx_t", sizeof(kgn_tx_t), 0, 0, NULL);
2261 if (kgnilnd_data.kgn_tx_cache == NULL) {
2262 CERROR("Can't create slab for kgn_tx_t\n");
2263 GOTO(failed, rc = -ENOMEM);
2266 kgnilnd_data.kgn_tx_phys_cache =
2267 kmem_cache_create("kgn_tx_phys",
2268 LNET_MAX_IOV * sizeof(gni_mem_segment_t),
2270 if (kgnilnd_data.kgn_tx_phys_cache == NULL) {
2271 CERROR("Can't create slab for kgn_tx_phys\n");
2272 GOTO(failed, rc = -ENOMEM);
2275 kgnilnd_data.kgn_dgram_cache =
2276 kmem_cache_create("kgn_dgram_t", sizeof(kgn_dgram_t), 0, 0, NULL);
2277 if (kgnilnd_data.kgn_dgram_cache == NULL) {
2278 CERROR("Can't create slab for outgoing datagrams\n");
2279 GOTO(failed, rc = -ENOMEM);
2282 /* allocate a MAX_IOV array of page pointers for each cpu */
2283 kgnilnd_data.kgn_cksum_map_pages = kmalloc(num_possible_cpus() * sizeof (struct page *),
2285 if (kgnilnd_data.kgn_cksum_map_pages == NULL) {
2286 CERROR("Can't allocate vmap cksum pages\n");
2287 GOTO(failed, rc = -ENOMEM);
2289 kgnilnd_data.kgn_cksum_npages = num_possible_cpus();
2290 memset(kgnilnd_data.kgn_cksum_map_pages, 0,
2291 kgnilnd_data.kgn_cksum_npages * sizeof (struct page *));
2293 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2294 kgnilnd_data.kgn_cksum_map_pages[i] = kmalloc(LNET_MAX_IOV * sizeof (struct page *),
2296 if (kgnilnd_data.kgn_cksum_map_pages[i] == NULL) {
2297 CERROR("Can't allocate vmap cksum pages for cpu %d\n", i);
2298 GOTO(failed, rc = -ENOMEM);
2302 LASSERT(kgnilnd_data.kgn_ndevs == 0);
2304 /* Use all available GNI devices */
2305 for (i = 0; i < GNILND_MAXDEVS; i++) {
2306 dev = &kgnilnd_data.kgn_devices[kgnilnd_data.kgn_ndevs];
2308 rc = kgnilnd_dev_init(dev);
2310 /* Increment here so base_shutdown cleans it up */
2311 kgnilnd_data.kgn_ndevs++;
2313 rc = kgnilnd_allocate_phys_fmablk(dev);
2319 if (kgnilnd_data.kgn_ndevs == 0) {
2320 CERROR("Can't initialise any GNI devices\n");
2321 GOTO(failed, rc = -ENODEV);
2324 rc = kgnilnd_thread_start(kgnilnd_reaper, NULL, "kgnilnd_rpr", 0);
2326 CERROR("Can't spawn gnilnd reaper: %d\n", rc);
2330 rc = kgnilnd_start_rca_thread();
2332 CERROR("Can't spawn gnilnd rca: %d\n", rc);
2337 * Start ruhroh thread. We can't use kgnilnd_thread_start() because
2338 * we don't want this thread included in kgnilnd_data.kgn_nthreads
2339 * count. This thread controls quiesce, so it mustn't
2342 thrd = kthread_run(kgnilnd_ruhroh_thread, NULL, "%s_%02d", "kgnilnd_rr", 0);
2345 CERROR("Can't spawn gnilnd ruhroh thread: %d\n", rc);
2349 /* threads will load balance across devs as they are available */
2350 if (*kgnilnd_tunables.kgn_thread_affinity) {
2351 rc = kgnilnd_start_sd_threads();
2355 for (i = 0; i < *kgnilnd_tunables.kgn_sched_threads; i++) {
2356 rc = kgnilnd_thread_start(kgnilnd_scheduler,
2360 CERROR("Can't spawn gnilnd scheduler[%d]: %d\n",
2367 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2368 dev = &kgnilnd_data.kgn_devices[i];
2369 rc = kgnilnd_thread_start(kgnilnd_dgram_mover, dev,
2370 "kgnilnd_dg", dev->gnd_id);
2372 CERROR("Can't spawn gnilnd dgram_mover[%d]: %d\n",
2377 rc = kgnilnd_thread_start(kgnilnd_dgram_waitq, dev,
2378 "kgnilnd_dgn", dev->gnd_id);
2380 CERROR("Can't spawn gnilnd dgram_waitq[%d]: %d\n",
2385 rc = kgnilnd_setup_wildcard_dgram(dev);
2388 CERROR("Can't create wildcard dgrams[%d]: %d\n",
2394 /* flag everything initialised */
2395 kgnilnd_data.kgn_init = GNILND_INIT_ALL;
2396 /*****************************************************/
2398 CDEBUG(D_MALLOC, "initial kmem %lld\n", pkmem);
2402 kgnilnd_base_shutdown();
2403 kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2408 kgnilnd_base_shutdown(void)
2413 while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_SHUTDOWN, 1)) {};
2415 kgnilnd_data.kgn_wc_kill = 1;
2417 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2418 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2419 kgnilnd_cancel_wc_dgrams(dev);
2420 kgnilnd_cancel_dgrams(dev);
2421 kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2422 kgnilnd_wait_for_canceled_dgrams(dev);
2425 /* We need to verify there are no conns left before we let the threads
2426 * shut down otherwise we could clean up the peers but still have
2427 * some outstanding conns due to orphaned datagram conns that are
2431 while (atomic_read(&kgnilnd_data.kgn_nconns) != 0) {
2434 for(j = 0; j < kgnilnd_data.kgn_ndevs; ++j) {
2435 kgn_device_t *dev = &kgnilnd_data.kgn_devices[j];
2436 kgnilnd_schedule_device(dev);
2439 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2440 "Waiting for conns to be cleaned up %d\n",atomic_read(&kgnilnd_data.kgn_nconns));
2441 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2443 /* Peer state all cleaned up BEFORE setting shutdown, so threads don't
2444 * have to worry about shutdown races. NB connections may be created
2445 * while there are still active connds, but these will be temporary
2446 * since peer creation always fails after the listener has started to
2448 * all peers should have been cleared out on the nets */
2449 LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2450 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2452 /* Wait for the ruhroh thread to shut down. */
2453 kgnilnd_data.kgn_ruhroh_shutdown = 1;
2454 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
2456 while (kgnilnd_data.kgn_ruhroh_running != 0) {
2458 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2459 "Waiting for ruhroh thread to terminate\n");
2460 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2463 /* Flag threads to terminate */
2464 kgnilnd_data.kgn_shutdown = 1;
2466 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2467 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2469 /* should clear all the MDDs */
2470 kgnilnd_unmap_fma_blocks(dev);
2472 kgnilnd_schedule_device(dev);
2473 wake_up(&dev->gnd_dgram_waitq);
2474 wake_up(&dev->gnd_dgping_waitq);
2475 LASSERT(list_empty(&dev->gnd_connd_peers));
2478 spin_lock(&kgnilnd_data.kgn_reaper_lock);
2479 wake_up(&kgnilnd_data.kgn_reaper_waitq);
2480 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
2482 if (atomic_read(&kgnilnd_data.kgn_nthreads))
2483 kgnilnd_wakeup_rca_thread();
2485 /* Wait for threads to exit */
2487 while (atomic_read(&kgnilnd_data.kgn_nthreads) != 0) {
2489 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2490 "Waiting for %d threads to terminate\n",
2491 atomic_read(&kgnilnd_data.kgn_nthreads));
2492 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2495 LASSERTF(atomic_read(&kgnilnd_data.kgn_npeers) == 0,
2496 "peers left %d\n", atomic_read(&kgnilnd_data.kgn_npeers));
2498 if (kgnilnd_data.kgn_peers != NULL) {
2499 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2500 LASSERT(list_empty(&kgnilnd_data.kgn_peers[i]));
2502 CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_peers,
2503 *kgnilnd_tunables.kgn_peer_hash_size);
2506 down_write(&kgnilnd_data.kgn_net_rw_sem);
2507 if (kgnilnd_data.kgn_nets != NULL) {
2508 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++)
2509 LASSERT(list_empty(&kgnilnd_data.kgn_nets[i]));
2511 CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_nets,
2512 *kgnilnd_tunables.kgn_net_hash_size);
2514 up_write(&kgnilnd_data.kgn_net_rw_sem);
2516 LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
2517 "conns left %d\n", atomic_read(&kgnilnd_data.kgn_nconns));
2519 if (kgnilnd_data.kgn_conns != NULL) {
2520 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++)
2521 LASSERT(list_empty(&kgnilnd_data.kgn_conns[i]));
2523 CFS_FREE_PTR_ARRAY(kgnilnd_data.kgn_conns,
2524 *kgnilnd_tunables.kgn_peer_hash_size);
2527 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
2528 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
2529 kgnilnd_dev_fini(dev);
2531 LASSERTF(atomic_read(&dev->gnd_ndgrams) == 0,
2532 "dgrams left %d\n", atomic_read(&dev->gnd_ndgrams));
2534 if (dev->gnd_dgrams != NULL) {
2535 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size;
2537 LASSERT(list_empty(&dev->gnd_dgrams[i]));
2539 CFS_FREE_PTR_ARRAY(dev->gnd_dgrams,
2540 *kgnilnd_tunables.kgn_peer_hash_size);
2543 kgnilnd_free_phys_fmablk(dev);
2546 if (kgnilnd_data.kgn_mbox_cache != NULL)
2547 kmem_cache_destroy(kgnilnd_data.kgn_mbox_cache);
2549 if (kgnilnd_data.kgn_rx_cache != NULL)
2550 kmem_cache_destroy(kgnilnd_data.kgn_rx_cache);
2552 if (kgnilnd_data.kgn_tx_cache != NULL)
2553 kmem_cache_destroy(kgnilnd_data.kgn_tx_cache);
2555 if (kgnilnd_data.kgn_tx_phys_cache != NULL)
2556 kmem_cache_destroy(kgnilnd_data.kgn_tx_phys_cache);
2558 if (kgnilnd_data.kgn_dgram_cache != NULL)
2559 kmem_cache_destroy(kgnilnd_data.kgn_dgram_cache);
2561 if (kgnilnd_data.kgn_cksum_map_pages != NULL) {
2562 for (i = 0; i < kgnilnd_data.kgn_cksum_npages; i++) {
2563 if (kgnilnd_data.kgn_cksum_map_pages[i] != NULL) {
2564 kfree(kgnilnd_data.kgn_cksum_map_pages[i]);
2567 kfree(kgnilnd_data.kgn_cksum_map_pages);
2570 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2571 libcfs_kmem_read());
2573 kgnilnd_data.kgn_init = GNILND_INIT_NOTHING;
2574 module_put(THIS_MODULE);
2580 kgnilnd_startup(struct lnet_ni *ni)
2586 LASSERTF(ni->ni_net->net_lnd == &the_kgnilnd,
2587 "bad LND 0x%p != the_kgnilnd @ 0x%p\n",
2588 ni->ni_net->net_lnd, &the_kgnilnd);
2590 if (kgnilnd_data.kgn_init == GNILND_INIT_NOTHING) {
2591 rc = kgnilnd_base_startup();
2596 /* Serialize with shutdown. */
2597 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2599 LIBCFS_ALLOC(net, sizeof(*net));
2601 CERROR("could not allocate net for new interface instance\n");
2602 /* no need to cleanup the CDM... */
2603 GOTO(failed, rc = -ENOMEM);
2605 INIT_LIST_HEAD(&net->gnn_list);
2608 if (!ni->ni_net->net_tunables_set) {
2609 ni->ni_net->net_tunables.lct_max_tx_credits =
2610 *kgnilnd_tunables.kgn_credits;
2611 ni->ni_net->net_tunables.lct_peer_tx_credits =
2612 *kgnilnd_tunables.kgn_peer_credits;
2615 if (*kgnilnd_tunables.kgn_peer_health) {
2618 /* give this a bit of leeway - we don't have a hard timeout
2619 * as we only check timeouts periodically - see comment in kgnilnd_reaper */
2620 fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
2621 timeout = *kgnilnd_tunables.kgn_timeout + fudge;
2623 if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) {
2624 ni->ni_net->net_tunables.lct_peer_timeout =
2625 *kgnilnd_tunables.kgn_peer_timeout;
2626 } else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
2627 LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
2628 *kgnilnd_tunables.kgn_peer_timeout,
2631 LIBCFS_FREE(net, sizeof(*net));
2632 GOTO(failed, rc = -EINVAL);
2634 ni->ni_net->net_tunables.lct_peer_timeout = timeout;
2636 LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
2637 ni->ni_net->net_tunables.lct_peer_timeout);
2640 atomic_set(&net->gnn_refcount, 1);
2642 /* if we have multiple devices, spread the nets around */
2643 net->gnn_netnum = LNET_NETNUM(LNET_NIDNET(ni->ni_nid));
2645 devno = LNET_NIDNET(ni->ni_nid) % GNILND_MAXDEVS;
2646 net->gnn_dev = &kgnilnd_data.kgn_devices[devno];
2648 /* allocate a 'dummy' cdm for datagram use. We can only have a single
2649 * datagram between a nid:inst_id and nid2:inst_id. The fake cdm
2650 * give us additional inst_id to use, allowing the datagrams to flow
2651 * like rivers of honey and beer */
2653 /* the instance id for the cdm is the NETNUM offset by MAXDEVS -
2654 * ensuring we'll have a unique id */
2657 ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), net->gnn_dev->gnd_nid);
2658 CDEBUG(D_NET, "adding net %p nid=%s on dev %d \n",
2659 net, libcfs_nid2str(ni->ni_nid), net->gnn_dev->gnd_id);
2660 /* until the gnn_list is set, we need to cleanup ourselves as
2661 * kgnilnd_shutdown is just gonna get confused */
2663 down_write(&kgnilnd_data.kgn_net_rw_sem);
2664 list_add_tail(&net->gnn_list, kgnilnd_netnum2netlist(net->gnn_netnum));
2665 up_write(&kgnilnd_data.kgn_net_rw_sem);
2667 /* we need a separate thread to call probe_wait_by_id until
2668 * we get a function callback notifier from kgni */
2669 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2672 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2673 kgnilnd_shutdown(ni);
2678 kgnilnd_shutdown(struct lnet_ni *ni)
2680 kgn_net_t *net = ni->ni_data;
2685 CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);
2687 LASSERTF(kgnilnd_data.kgn_init == GNILND_INIT_ALL,
2688 "init %d\n", kgnilnd_data.kgn_init);
2690 /* Serialize with startup. */
2691 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
2692 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2693 libcfs_kmem_read());
2696 CERROR("got NULL net for ni %p\n", ni);
2697 GOTO(out, rc = -EINVAL);
2700 LASSERTF(ni == net->gnn_ni,
2701 "ni %p gnn_ni %p\n", net, net->gnn_ni);
2705 LASSERT(!net->gnn_shutdown);
2706 LASSERTF(atomic_read(&net->gnn_refcount) != 0,
2707 "net %p refcount %d\n",
2708 net, atomic_read(&net->gnn_refcount));
2710 if (!list_empty(&net->gnn_list)) {
2711 /* serialize with peer creation */
2712 down_write(&kgnilnd_data.kgn_net_rw_sem);
2713 net->gnn_shutdown = 1;
2714 up_write(&kgnilnd_data.kgn_net_rw_sem);
2716 kgnilnd_cancel_net_dgrams(net);
2718 kgnilnd_del_conn_or_peer(net, LNET_NID_ANY, GNILND_DEL_PEER, -ESHUTDOWN);
2720 /* if we are quiesced, need to wake up - we need those threads
2721 * alive to release peers, etc */
2722 if (GNILND_IS_QUIESCED) {
2723 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
2724 kgnilnd_quiesce_wait("shutdown");
2727 kgnilnd_wait_for_canceled_dgrams(net->gnn_dev);
2729 /* We wait until the nets ref's are 1, we will release final ref which is ours
2730 * this allows us to make sure everything else is done before we free the
2734 while (atomic_read(&net->gnn_refcount) != 1) {
2736 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2737 "Waiting for %d references to clear on net %d\n",
2738 atomic_read(&net->gnn_refcount),
2740 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2743 /* release ref from kgnilnd_startup */
2744 kgnilnd_net_decref(net);
2745 /* serialize with reaper and conn_task looping */
2746 down_write(&kgnilnd_data.kgn_net_rw_sem);
2747 list_del_init(&net->gnn_list);
2748 up_write(&kgnilnd_data.kgn_net_rw_sem);
2752 /* not locking, this can't race with writers */
2753 LASSERTF(atomic_read(&net->gnn_refcount) == 0,
2754 "net %p refcount %d\n",
2755 net, atomic_read(&net->gnn_refcount));
2756 LIBCFS_FREE(net, sizeof(*net));
2759 down_read(&kgnilnd_data.kgn_net_rw_sem);
2760 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
2761 if (!list_empty(&kgnilnd_data.kgn_nets[i])) {
2762 up_read(&kgnilnd_data.kgn_net_rw_sem);
2766 if (i == *kgnilnd_tunables.kgn_net_hash_size - 1) {
2767 up_read(&kgnilnd_data.kgn_net_rw_sem);
2768 kgnilnd_base_shutdown();
2771 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2772 libcfs_kmem_read());
2774 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
2778 static void __exit kgnilnd_exit(void)
2780 lnet_unregister_lnd(&the_kgnilnd);
2781 kgnilnd_proc_fini();
2782 kgnilnd_remove_sysctl();
2785 static int __init kgnilnd_init(void)
2789 rc = kgnilnd_tunables_init();
2793 LCONSOLE_INFO("Lustre: kgnilnd build version: "LUSTRE_VERSION_STRING"\n");
2795 kgnilnd_insert_sysctl();
2796 kgnilnd_proc_init();
2798 lnet_register_lnd(&the_kgnilnd);
2803 MODULE_AUTHOR("Cray, Inc. <nic@cray.com>");
2804 MODULE_DESCRIPTION("Gemini LNet Network Driver");
2805 MODULE_VERSION(LUSTRE_VERSION_STRING);
2806 MODULE_LICENSE("GPL");
2808 module_init(kgnilnd_init);
2809 module_exit(kgnilnd_exit);