2 * Copyright (C) 2012 Cray, Inc.
4 * Copyright (c) 2014, Intel Corporation.
6 * Author: Nic Henke <nic@cray.com>
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #if defined(GNILND_USE_RCA)
26 #include <rsms/rs_sm_states.h>
28 /* Advance all timeouts by nap_time seconds. */
30 kgnilnd_bump_timeouts(__u32 nap_time, char *reason)
39 LCONSOLE_INFO("%s: bumping all timeouts by %ds\n", reason, nap_time);
41 LASSERTF(GNILND_IS_QUIESCED, "gnilnd not quiesced %d != %d\n",
42 atomic_read(&kgnilnd_data.kgn_nquiesce),
43 atomic_read(&kgnilnd_data.kgn_nthreads));
45 /* requiring that the threads are paused ensures a couple of things:
46 * - combined code paths for stack reset and quiesce event as stack reset
47 * runs with the threads paused
48 * - prevents traffic to the Gemini during a quiesce period
49 * - reduces the locking requirements
52 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
53 list_for_each_entry(peer, &kgnilnd_data.kgn_peers[i], gnp_list) {
55 /* we can reconnect again at any time */
56 peer->gnp_reconnect_time = jiffies;
57 /* reset now that network is healthy */
58 peer->gnp_reconnect_interval = 0;
59 /* tell LNet dude is still alive */
60 kgnilnd_peer_alive(peer);
62 list_for_each_entry(tx, &peer->gnp_tx_queue, tx_list) {
63 tx->tx_qtime = jiffies;
66 list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
67 unsigned long timeout;
69 timeout = cfs_time_seconds(conn->gnc_timeout);
71 /* bump last_rx/last_rx_cq on all conns - including
72 * closed ones, this will have the effect of
73 * bumping the purgatory timers for those */
74 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
76 /* we don't timeout based on old gnc_last_tx, so
77 * we'll back it up and schedule the conn to trigger
79 conn->gnc_last_tx = jiffies - timeout;
80 if (conn->gnc_state != GNILND_CONN_DONE)
81 kgnilnd_schedule_conn(conn);
86 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
87 dev = &kgnilnd_data.kgn_devices[i];
88 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
89 list_for_each_entry(dgram, &dev->gnd_dgrams[i], gndg_list) {
90 dgram->gndg_post_time = jiffies;
96 /* Quiesce or wake up the stack. The caller must hold the kgn_quiesce_sem semaphore
97 * on entry, which holds off any pending stack shutdown. */
99 kgnilnd_quiesce_wait(char *reason)
103 if (kgnilnd_data.kgn_quiesce_trigger) {
104 unsigned long quiesce_deadline, quiesce_to;
107 /* morning sunshine */
108 spin_lock(&kgnilnd_data.kgn_reaper_lock);
109 wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
110 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
112 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
113 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
115 wake_up_all(&dev->gnd_waitq);
116 wake_up_all(&dev->gnd_dgram_waitq);
117 wake_up_all(&dev->gnd_dgping_waitq);
120 kgnilnd_wakeup_rca_thread();
122 /* we'll wait for 10x the timeout for the threads to pause */
123 quiesce_to = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * 10);
124 quiesce_deadline = (long) jiffies + quiesce_to;
126 /* wait for everyone to check-in as quiesced */
128 while (!GNILND_IS_QUIESCED) {
130 LCONSOLE((((i) & (-i)) == i) ? D_WARNING : D_NET,
131 "%s: Waiting for %d threads to pause\n",
133 atomic_read(&kgnilnd_data.kgn_nthreads) -
134 atomic_read(&kgnilnd_data.kgn_nquiesce));
135 CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);
136 cfs_pause(cfs_time_seconds(1 * i));
138 LASSERTF(quiesce_deadline > jiffies,
139 "couldn't quiesce threads in %lu seconds, falling over now\n",
140 cfs_duration_sec(quiesce_to));
143 LCONSOLE_WARN("%s: All threads paused!\n", reason);
144 /* XXX Nic: Is there a set of counters we can grab here to
145 * ensure that there is no traffic until quiesce is over ?*/
149 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
150 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
151 kgnilnd_schedule_dgram(dev);
154 /* wait for everyone to check-in as running - they will be spinning
155 * and looking, so no need to poke any waitq */
157 while (atomic_read(&kgnilnd_data.kgn_nquiesce) > 0) {
159 LCONSOLE((((i) & (-i)) == i) ? D_WARNING : D_NET,
160 "%s: Waiting for %d threads to wake up\n",
162 atomic_read(&kgnilnd_data.kgn_nquiesce));
163 cfs_pause(cfs_time_seconds(1 * i));
166 LCONSOLE_WARN("%s: All threads awake!\n", reason);
170 /* Reset the stack. */
172 kgnilnd_reset_stack(void)
176 kgn_peer_t *peer, *peerN;
178 char *reason = "critical hardware error";
180 unsigned long start, end;
183 /* Race with del_peer and its atomics */
184 CFS_RACE(CFS_FAIL_GNI_RACE_RESET);
186 if (kgnilnd_data.kgn_init != GNILND_INIT_ALL) {
187 CERROR("can't reset the stack, gnilnd is not initialized\n");
191 /* First make sure we are not already quiesced - we panic if so,
192 * as that could leave software in a bad state */
193 LASSERTF(kgnilnd_data.kgn_quiesce_trigger == GNILND_QUIESCE_IDLE,
194 "can't reset the stack, already doing so: trigger %d\n",
195 kgnilnd_data.kgn_quiesce_trigger);
197 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_RESET);
199 /* wake up the dgram waitq thread - but after trigger set to make sure it
200 * goes into quiesce */
201 CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);
202 /* same for scheduler that is dropping state transitiosn */
203 CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
204 CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
206 kgnilnd_quiesce_wait(reason);
210 kgnilnd_data.kgn_in_reset = 1;
211 kgnilnd_data.kgn_nresets++;
212 LCONSOLE_WARN("%s: resetting all resources (count %d)\n",
213 reason, kgnilnd_data.kgn_nresets);
215 for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
216 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
217 rc = kgnilnd_cancel_net_dgrams(net);
218 LASSERTF(rc == 0, "couldn't cleanup datagrams: %d\n", rc);
222 /* error -ENOTRECOVERABLE is stack reset */
223 kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_CONN, -ENOTRECOVERABLE);
225 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
226 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
227 kgnilnd_cancel_wc_dgrams(dev);
228 kgnilnd_wait_for_canceled_dgrams(dev);
231 /* manually do some conn processing ala kgnilnd_process_conns */
232 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
233 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
237 /* go find all the closed conns that need to be nuked - the
238 * scheduler thread isn't running to do this for us */
240 CDEBUG(D_NET, "will try to clear up %d ready_conns\n",
241 kgnilnd_count_list(&dev->gnd_ready_conns));
243 /* use while/list_first_entry loop to ensure we can handle any
244 * DESTROY_EP conns added from kgnilnd_complete_closed_conn */
245 while (!list_empty(&dev->gnd_ready_conns)) {
246 conn = list_first_entry(&dev->gnd_ready_conns,
247 kgn_conn_t, gnc_schedlist);
248 conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
250 LASSERTF(conn_sched != GNILND_CONN_IDLE &&
251 conn_sched != GNILND_CONN_PROCESS,
252 "conn %p on ready list but in bad state: %d\n",
255 list_del_init(&conn->gnc_schedlist);
257 if (conn->gnc_state == GNILND_CONN_CLOSING) {
258 /* bump to CLOSED to fake out send of CLOSE */
259 conn->gnc_state = GNILND_CONN_CLOSED;
260 conn->gnc_close_sent = 1;
263 if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
264 kgnilnd_destroy_conn_ep(conn);
266 kgnilnd_complete_closed_conn(conn);
269 /* there really shouldn't be any other states here -
270 * they would have been cleared out in the del_peer_or_conn or the dgram
272 * there is an LASSERTF in kgnilnd_complete_closed_conn that will take
273 * care of catching anything else for us */
275 kgnilnd_schedule_process_conn(conn, -1);
277 kgnilnd_conn_decref(conn);
281 /* don't let the little weasily purgatory conns hide from us */
282 for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
283 list_for_each_entry_safe(peer, peerN, &kgnilnd_data.kgn_peers[i], gnp_list) {
284 kgn_conn_t *conn, *connN;
286 list_for_each_entry_safe(conn, connN, &peer->gnp_conns, gnc_list) {
287 kgnilnd_detach_purgatory_locked(conn, &souls);
292 CDEBUG(D_NET, "about to release %d purgatory entries\n",
293 kgnilnd_count_list(&souls));
295 kgnilnd_release_purgatory_list(&souls);
297 /* validate we are now clean */
298 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
299 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
301 /* now all the cons/mboxes should be cleaned up, including purgatory
302 * so go through and release the MDDs for our persistent PHYS fma_blks
304 kgnilnd_unmap_fma_blocks(dev);
306 LASSERTF(atomic_read(&dev->gnd_nfmablk) == 0,
307 "reset failed: fma blocks still live %d\n",
308 atomic_read(&dev->gnd_nfmablk));
310 LASSERTF(atomic_read(&dev->gnd_neps) == 0,
311 "reset failed: EP handles still live %d\n",
312 atomic_read(&dev->gnd_neps));
315 LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
316 "reset failed: conns left %d\n",
317 atomic_read(&kgnilnd_data.kgn_nconns));
319 /* fine to have peers left - they are waiting for new conns
320 * but should not be holding any open HW resources */
322 /* like the last part of kgnilnd_base_shutdown() */
324 CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);
326 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
327 kgnilnd_dev_fini(&kgnilnd_data.kgn_devices[i]);
330 /* no need to free and recreate the TX descriptors
331 * we nuked all the ones that could be using HW resources in
332 * kgnilnd_close_matching_conns and asserted it worked in
333 * kgnilnd_dev_fini */
335 /* At this point, all HW is torn down, start to reset */
337 /* only reset our known devs */
338 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
339 kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
340 rc = kgnilnd_dev_init(dev);
341 LASSERTF(rc == 0, "dev_init failed for dev %d\n", i);
342 kgnilnd_map_phys_fmablk(dev);
343 LASSERTF(rc == 0, "map_phys_fmablk failed for dev %d\n", i);
344 rc = kgnilnd_setup_wildcard_dgram(dev);
345 LASSERTF(rc == 0, "couldnt setup datagrams on dev %d: %d\n",
349 /* Now the fun restarts... - release the hounds! */
352 seconds = cfs_duration_sec((long)end - start);
353 kgnilnd_bump_timeouts(seconds, reason);
355 kgnilnd_data.kgn_in_reset = 0;
356 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
357 kgnilnd_quiesce_wait(reason);
358 LCONSOLE_WARN("%s reset of all hardware resources\n",
359 rc ? "failed" : "successful");
364 /* A thread that handles quiece and reset hardware events.
365 * We do the same thing regardless of which device reported the event. */
367 kgnilnd_ruhroh_thread(void *arg)
373 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
374 kgnilnd_data.kgn_ruhroh_running = 1;
378 /* Block until there's a request.. A reset request could come in
379 * while we're handling a quiesce one, or vice versa.
380 * Keep processing requests until there are none.*/
381 prepare_to_wait(&kgnilnd_data.kgn_ruhroh_waitq, &wait, TASK_INTERRUPTIBLE);
382 while (!(kgnilnd_data.kgn_ruhroh_shutdown ||
383 kgnilnd_data.kgn_needs_reset || kgnilnd_data.kgn_needs_pause))
385 finish_wait(&kgnilnd_data.kgn_ruhroh_waitq, &wait);
387 /* Exit if the driver is shutting down. */
388 if (kgnilnd_data.kgn_ruhroh_shutdown)
391 /* Serialize with driver startup and shutdown. */
392 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
394 CDEBUG(D_NET, "trigger %d reset %d to_bump %d pause %d\n",
395 kgnilnd_data.kgn_quiesce_trigger,
396 kgnilnd_data.kgn_needs_reset,
397 kgnilnd_data.kgn_bump_info_rdy,
398 kgnilnd_data.kgn_needs_pause);
400 /* Do we need to do a pause/quiesce? */
401 if (kgnilnd_data.kgn_needs_pause) {
403 /* Pause all other kgnilnd threads. */
404 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_HW_QUIESCE);
405 kgnilnd_quiesce_wait("hardware quiesce flag");
407 /* If the hardware quiesce flag is set, wait for it to clear.
408 * This should happen relatively quickly, so we wait for it.
409 * This will hold up the eventd thread, but on everything but
410 * the simulator, this is ok-- there is one thread per core.
412 * Handle (possibly multiple) quiesce events while we wait. The
413 * memory barrier ensures that the core doesn't start fetching
414 * kgn_bump_info_rdy before it fetches kgn_needs_pause, and
415 * matches the second mb in kgnilnd_quiesce_end_callback(). */
417 while (kgnilnd_hw_in_quiesce() || kgnilnd_data.kgn_bump_info_rdy) {
420 LCONSOLE((((i) & (-i)) == i) ? D_WARNING : D_NET,
421 "Waiting for hardware quiesce flag to clear\n");
422 cfs_pause(cfs_time_seconds(1 * i));
424 /* If we got a quiesce event with bump info, DO THE BUMP!. */
425 if (kgnilnd_data.kgn_bump_info_rdy) {
426 /* reset console rate limiting for each event */
429 /* Make sure the core doesn't start fetching
430 * kgni_quiesce_seconds until after it sees
431 * kgn_bump_info_rdy set. This is the match to the
432 * first mb in kgnilnd_quiesce_end_callback(). */
434 (void) kgnilnd_bump_timeouts(kgnilnd_data.kgn_quiesce_secs,
435 "hardware quiesce callback");
436 set_mb(kgnilnd_data.kgn_quiesce_secs, 0);
437 set_mb(kgnilnd_data.kgn_bump_info_rdy, 0);
441 /* Reset the kgn_needs_pause flag before coming out of
442 * the pause. This ordering avoids a race with the
443 * setting of this flag in kgnilnd_pause_threads(). */
444 set_mb(kgnilnd_data.kgn_needs_pause, 0);
446 /* ok, let the kids back into the pool */
447 set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
448 kgnilnd_quiesce_wait("hardware quiesce");
451 /* Do a stack reset if needed. */
452 if (kgnilnd_data.kgn_needs_reset) {
453 kgnilnd_reset_stack();
454 set_mb(kgnilnd_data.kgn_needs_reset, 0);
457 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
460 kgnilnd_data.kgn_ruhroh_running = 0;
464 /* Set pause request flag. Any functions that
465 * call this one are responsible for ensuring that
466 * variables they set up are visible on other cores before
467 * this flag setting. This executes in interrupt or kernel
470 kgnilnd_pause_threads(void)
472 /* only device 0 gets the handle, see kgnilnd_dev_init */
473 kgn_device_t *dev = &kgnilnd_data.kgn_devices[0];
474 LASSERTF(dev != NULL, "dev 0 is NULL\n");
476 /* If we're currently in a pause triggered by the pause flag,
477 * there's no need to set it again. We clear the kgn_needs_pause
478 * flag before we reset kgn_quiesce_trigger to avoid a race. The
479 * read memory barrier matches the setmb() on the trigger in
480 * kgnilnd_ruhroh_task(). */
482 if (!(kgnilnd_data.kgn_quiesce_trigger == GNILND_QUIESCE_HW_QUIESCE &&
483 GNILND_IS_QUIESCED)) {
484 CDEBUG(D_NET, "requesting thread pause\n");
486 kgnilnd_data.kgn_needs_pause = 1;
488 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
490 CDEBUG(D_NET, "thread pause already underway\n");
494 /* Return non-zero if the GNI hardware quiesce flag is set */
496 kgnilnd_hw_in_quiesce(void)
498 /* only device 0 gets the handle, see kgnilnd_dev_init */
499 kgn_device_t *dev0 = &kgnilnd_data.kgn_devices[0];
501 LASSERTF(dev0 != NULL, "dev 0 is NULL\n");
504 return kgnilnd_get_quiesce_status(dev0->gnd_handle) != 0;
508 /* If the GNI hardware quiesce flag is set, initiate our pause and
509 * return non-zero. Also return non-zero if the stack is shutting down. */
511 kgnilnd_check_hw_quiesce(void)
513 if (likely(!kgnilnd_hw_in_quiesce()))
516 if (!kgnilnd_data.kgn_ruhroh_shutdown) {
517 CDEBUG(D_NET, "initiating thread pause\n");
518 kgnilnd_pause_threads();
520 CDEBUG(D_NET, "thread pause bypassed because of shutdown\n");
526 /* Callback from kngi with the quiesce duration. This executes
527 * in interrupt context. */
529 kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs)
531 /* only device 0 gets the handle, see kgnilnd_dev_init */
532 kgn_device_t *dev = &kgnilnd_data.kgn_devices[0];
533 LASSERTF(dev != NULL, "dev 0 is NULL\n");
535 if (!kgnilnd_data.kgn_ruhroh_shutdown) {
537 CDEBUG(D_NET, "requesting timeout bump by "LPD64" msecs\n", msecs);
539 /* Save the bump interval and request the bump.
540 * The memory barrier ensures that the interval is in place before
541 * the bump flag can be seen (in case a core is already running the
542 * ruhroh task), and that the bump request flag in place before
543 * the pause request can be seen (to ensure a core doesn't miss the bump
545 /* If another callback occurred before the ruhroh task
546 * finished processing the first bump request, we'd over-write its info.
547 * Nic says that callbacks occur so slowly that this isn't an issue. */
548 set_mb(kgnilnd_data.kgn_quiesce_secs, msecs / MSEC_PER_SEC);
549 set_mb(kgnilnd_data.kgn_bump_info_rdy, 1);
550 kgnilnd_pause_threads();
552 CDEBUG(D_NET, "timeout bump bypassed because of shutdown\n");
557 kgnilnd_critical_error(struct gni_err *err_handle)
559 /* only device 0 gets the handle, see kgnilnd_dev_init */
560 kgn_device_t *dev = &kgnilnd_data.kgn_devices[0];
561 LASSERTF(dev != NULL, "dev 0 is NULL\n");
563 if (!kgnilnd_data.kgn_ruhroh_shutdown) {
564 CDEBUG(D_NET, "requesting stack reset\n");
565 kgnilnd_data.kgn_needs_reset = 1;
566 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
568 CDEBUG(D_NET, "stack reset bypassed because of shutdown\n");
572 #if defined(GNILND_USE_RCA)
573 #include <krca_lib.h>
575 /* RCA ticket is needed for krca_wakeup_wait_event() */
576 static krca_ticket_t rca_krt = KRCA_NULL_TICKET;
582 static struct rcadata rd[RCA_EVENTS] = {
583 {0, 0, ec_node_unavailable},
584 {0, 0, ec_node_available},
585 {0, 0, ec_node_failed}
588 /* thread for receiving rca events */
590 kgnilnd_rca(void *arg)
599 /* all gnilnd threads need to run fairly urgently */
600 set_user_nice(current, *kgnilnd_tunables.kgn_nice);
603 * Register our service with RCA and subscribe to events
606 rca_krt = KRCA_NULL_TICKET;
607 rc = krca_register(&rca_krt, RCA_SVCTYPE_GNILND, current->pid, 0);
609 CNETERR("krca_register(%x) returned %d\n", current->pid, rc);
613 for (i = 0; i < RCA_EVENTS; i++) {
616 rc = krca_subscribe(&rca_krt, rd[i].ec, RCA_RX_SVC_ANY,
619 if ((rc == -EINTR) && !retry_count) {
621 CNETERR("krca_subscribe returned %d - retrying\n", rc);
622 goto subscribe_retry;
626 CNETERR("rca subscription failed (%d)\n", rc);
630 rd[i].subscribed = 1;
633 while (!kgnilnd_data.kgn_shutdown) {
634 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
635 KGNILND_SPIN_QUIESCE;
637 /* wait here for a subscribed event */
638 rc = krca_wait_event(&rca_krt);
640 /* RCA return values:
641 * 0 indicates krca_wakeup_wait_event caused krca_wait_event
643 * -ERESTARTSYS indicates krca_wait_event returned because of a
645 * -ENOSPC indicates no space available to create an rcad_reg_t
646 * 1 indicates a message is waiting.
652 if (krca_get_message(&rca_krt, &event) == 0) {
653 int node_down = GNILND_RCA_NODE_UNKNOWN;
657 /* Compute nodes don't care about other compute nodes
658 * so we don't need to create a peer.
660 if (GNILND_COMPUTE &&
661 !RSN_GET_FLD(event.ev_gen.svid_node.rs_node_flat,
666 /* Only care about compute and service nodes not GPUs */
667 if (RSN_GET_FLD(event.ev_gen.svid_node.rs_node_flat,
672 switch (event.ev_id) {
673 case ec_node_available:
674 CDEBUG(D_INFO, "ec_node_available\n");
675 node_down = GNILND_RCA_NODE_UP;
678 CDEBUG(D_INFO, "ec_node_failed\n");
679 if (event.ev_len > 0) {
681 "ec_node_failed ignored\n");
684 node_down = GNILND_RCA_NODE_DOWN;
686 case ec_node_unavailable:
687 state = RSN_GET_FLD(event.ev_gen.svid_node.rsn_intval, STATE);
689 CDEBUG(D_INFO, "ec_node_unavailable\n");
692 * Ignore overloaded ec_node_unavailable events
693 * generated by 'xtcli set_reserve'.
695 if (RS_GET_CS_STATE(state) == RS_CS_READY) {
696 CDEBUG(D_INFO, "ignoring "
697 "ec_node_unavailable event with"
698 " RS_CS_READY state\n");
701 node_down = GNILND_RCA_NODE_DOWN;
704 CDEBUG(D_INFO, "unknown event\n");
708 /* if we get an event we don't know about, just go ahead
709 * and wait for another event */
710 if (node_down == GNILND_RCA_NODE_UNKNOWN) {
714 nid = RSN_GET_FLD(event.ev_gen.svid_node.rs_node_flat,
716 CDEBUG(D_INFO,"kgnilnd_rca() reporting nid %d %s\n",
717 (int)nid, node_down ? "down" : "up");
718 kgnilnd_report_node_state(nid, node_down);
721 CNETERR("krca_get_message failed\n");
726 CDEBUG(D_INFO, "done\n");
728 for (i = 0; i < RCA_EVENTS; i++) {
729 if (rd[i].subscribed) {
730 rc = krca_unsubscribe(&rca_krt, rd[i].ticket);
733 CNETERR("rca unsubscribe failed (%d)\n", rc);
736 rd[i].subscribed = 0;
740 krca_unregister(&rca_krt);
741 kgnilnd_thread_fini();
747 kgnilnd_start_rca_thread(void)
749 return kgnilnd_thread_start(kgnilnd_rca, NULL, "kgnilnd_rca", 0);
753 kgnilnd_wakeup_rca_thread(void)
757 ret = krca_wakeup_wait_event(&rca_krt);
760 CDEBUG(D_ERROR, "krca_wakeup_wait_event failed\n");
765 kgnilnd_get_node_state(__u32 nid)
768 int rc = GNILND_RCA_NODE_UNKNOWN;
770 rs_node_array_t nlist;
771 rs_node_t *na = NULL;
773 if ((ret = krca_get_sysnodes(&nlist)) < 0) {
774 CDEBUG(D_NETERROR, "krca_get_sysnodes failed %d\n", ret);
780 for (i = 0; i < nlist.na_len; i++) {
781 if ((rca_nid_t)RSN_GET_FLD(na[i].rs_node_flat, NID) == nid) {
782 rc = RSN_GET_FLD(na[i].rs_node_flat, STATE) == RS_CS_READY ?
783 GNILND_RCA_NODE_UP : GNILND_RCA_NODE_DOWN;
790 CDEBUG(D_NET, "nid %d rc %d (0=up)\n", nid, rc);
794 #else /* GNILND_USE_RCA */
797 kgnilnd_start_rca_thread(void)
803 kgnilnd_wakeup_rca_thread(void)
808 kgnilnd_get_node_state(__u32 nid)
810 return GNILND_RCA_NODE_UP;
812 #endif /* GNILND_USE_RCA */