2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * (C) Copyright 2020 Hewlett Packard Enterprise Development LP
7 * Copyright (c) 2004-2009 Silicon Graphics, Inc. All Rights Reserved.
11 * Cross Partition Communication (XPC) support - standard version.
13 * XPC provides a message passing capability that crosses partition
14 * boundaries. This module is made up of two parts:
16 * partition This part detects the presence/absence of other
17 * partitions. It provides a heartbeat and monitors
18 * the heartbeats of other partitions.
20 * channel This part manages the channels and sends/receives
21 * messages across them to/from other partitions.
23 * There are a couple of additional functions residing in XP, which
24 * provide an interface to XPC for its users.
29 * . Currently on sn2, we have no way to determine which nasid an IRQ
30 * came from. Thus, xpc_send_IRQ_sn2() does a remote amo write
31 * followed by an IPI. The amo indicates where data is to be pulled
32 * from, so after the IPI arrives, the remote partition checks the amo
33 * word. The IPI can actually arrive before the amo however, so other
34 * code must periodically check for this case. Also, remote amo
35 * operations do not reliably time out. Thus we do a remote PIO read
36 * solely to know whether the remote partition is down and whether we
37 * should stop sending IPIs to it. This remote PIO read operation is
38 * set up in a special nofault region so SAL knows to ignore (and
39 * cleanup) any errors due to the remote amo write, PIO read, and/or
40 * PIO write operations.
42 * If/when new hardware solves this IPI problem, we should abandon
43 * the current approach.
47 #include <linux/module.h>
48 #include <linux/slab.h>
49 #include <linux/sysctl.h>
50 #include <linux/device.h>
51 #include <linux/delay.h>
52 #include <linux/reboot.h>
53 #include <linux/kdebug.h>
54 #include <linux/kthread.h>
58 #include <asm/traps.h>
61 /* define two XPC debug device structures to be used with dev_dbg() et al */
63 static struct device_driver xpc_dbg_name = {
67 static struct device xpc_part_dbg_subname = {
68 .init_name = "", /* set to "part" at xpc_init() time */
69 .driver = &xpc_dbg_name
72 static struct device xpc_chan_dbg_subname = {
73 .init_name = "", /* set to "chan" at xpc_init() time */
74 .driver = &xpc_dbg_name
77 struct device *xpc_part = &xpc_part_dbg_subname;
78 struct device *xpc_chan = &xpc_chan_dbg_subname;
80 static int xpc_kdebug_ignore;
82 /* systune related variables for /proc/sys directories */
84 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
85 static int xpc_hb_min_interval = 1;
86 static int xpc_hb_max_interval = 10;
88 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
89 static int xpc_hb_check_min_interval = 10;
90 static int xpc_hb_check_max_interval = 120;
92 int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
93 static int xpc_disengage_min_timelimit; /* = 0 */
94 static int xpc_disengage_max_timelimit = 120;
96 static struct ctl_table xpc_sys_xpc_hb_dir[] = {
98 .procname = "hb_interval",
99 .data = &xpc_hb_interval,
100 .maxlen = sizeof(int),
102 .proc_handler = proc_dointvec_minmax,
103 .extra1 = &xpc_hb_min_interval,
104 .extra2 = &xpc_hb_max_interval},
106 .procname = "hb_check_interval",
107 .data = &xpc_hb_check_interval,
108 .maxlen = sizeof(int),
110 .proc_handler = proc_dointvec_minmax,
111 .extra1 = &xpc_hb_check_min_interval,
112 .extra2 = &xpc_hb_check_max_interval},
115 static struct ctl_table xpc_sys_xpc_dir[] = {
119 .child = xpc_sys_xpc_hb_dir},
121 .procname = "disengage_timelimit",
122 .data = &xpc_disengage_timelimit,
123 .maxlen = sizeof(int),
125 .proc_handler = proc_dointvec_minmax,
126 .extra1 = &xpc_disengage_min_timelimit,
127 .extra2 = &xpc_disengage_max_timelimit},
130 static struct ctl_table xpc_sys_dir[] = {
134 .child = xpc_sys_xpc_dir},
137 static struct ctl_table_header *xpc_sysctl;
139 /* non-zero if any remote partition disengage was timed out */
140 int xpc_disengage_timedout;
142 /* #of activate IRQs received and not yet processed */
143 int xpc_activate_IRQ_rcvd;
144 DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
146 /* IRQ handler notifies this wait queue on receipt of an IRQ */
147 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
149 static unsigned long xpc_hb_check_timeout;
150 static struct timer_list xpc_hb_timer;
152 /* notification that the xpc_hb_checker thread has exited */
153 static DECLARE_COMPLETION(xpc_hb_checker_exited);
155 /* notification that the xpc_discovery thread has exited */
156 static DECLARE_COMPLETION(xpc_discovery_exited);
158 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
160 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
161 static struct notifier_block xpc_reboot_notifier = {
162 .notifier_call = xpc_system_reboot,
165 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
166 static struct notifier_block xpc_die_notifier = {
167 .notifier_call = xpc_system_die,
170 struct xpc_arch_operations xpc_arch_ops;
173 * Timer function to enforce the timelimit on the partition disengage.
176 xpc_timeout_partition_disengage(struct timer_list *t)
178 struct xpc_partition *part = from_timer(part, t, disengage_timer);
180 DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
182 (void)xpc_partition_disengaged(part);
184 DBUG_ON(part->disengage_timeout != 0);
185 DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
189 * Timer to produce the heartbeat. The timer structures function is
190 * already set when this is initially called. A tunable is used to
191 * specify when the next timeout should occur.
194 xpc_hb_beater(struct timer_list *unused)
196 xpc_arch_ops.increment_heartbeat();
198 if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
199 wake_up_interruptible(&xpc_activate_IRQ_wq);
201 xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
202 add_timer(&xpc_hb_timer);
206 xpc_start_hb_beater(void)
208 xpc_arch_ops.heartbeat_init();
209 timer_setup(&xpc_hb_timer, xpc_hb_beater, 0);
214 xpc_stop_hb_beater(void)
216 del_timer_sync(&xpc_hb_timer);
217 xpc_arch_ops.heartbeat_exit();
221 * At periodic intervals, scan through all active partitions and ensure
222 * their heartbeat is still active. If not, the partition is deactivated.
225 xpc_check_remote_hb(void)
227 struct xpc_partition *part;
231 for (partid = 0; partid < xp_max_npartitions; partid++) {
236 if (partid == xp_partition_id)
239 part = &xpc_partitions[partid];
241 if (part->act_state == XPC_P_AS_INACTIVE ||
242 part->act_state == XPC_P_AS_DEACTIVATING) {
246 ret = xpc_arch_ops.get_remote_heartbeat(part);
247 if (ret != xpSuccess)
248 XPC_DEACTIVATE_PARTITION(part, ret);
253 * This thread is responsible for nearly all of the partition
254 * activation/deactivation.
257 xpc_hb_checker(void *ignore)
261 /* this thread was marked active by xpc_hb_init() */
263 set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
265 /* set our heartbeating to other partitions into motion */
266 xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
267 xpc_start_hb_beater();
269 while (!xpc_exiting) {
271 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
273 (int)(xpc_hb_check_timeout - jiffies),
274 xpc_activate_IRQ_rcvd);
276 /* checking of remote heartbeats is skewed by IRQ handling */
277 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
278 xpc_hb_check_timeout = jiffies +
279 (xpc_hb_check_interval * HZ);
281 dev_dbg(xpc_part, "checking remote heartbeats\n");
282 xpc_check_remote_hb();
285 /* check for outstanding IRQs */
286 if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
288 dev_dbg(xpc_part, "processing activate IRQs "
290 xpc_arch_ops.process_activate_IRQ_rcvd();
293 /* wait for IRQ or timeout */
294 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
295 (time_is_before_eq_jiffies(
296 xpc_hb_check_timeout) ||
297 xpc_activate_IRQ_rcvd > 0 ||
301 xpc_stop_hb_beater();
303 dev_dbg(xpc_part, "heartbeat checker is exiting\n");
305 /* mark this thread as having exited */
306 complete(&xpc_hb_checker_exited);
311 * This thread will attempt to discover other partitions to activate
312 * based on info provided by SAL. This new thread is short lived and
313 * will exit once discovery is complete.
316 xpc_initiate_discovery(void *ignore)
320 dev_dbg(xpc_part, "discovery thread is exiting\n");
322 /* mark this thread as having exited */
323 complete(&xpc_discovery_exited);
328 * The first kthread assigned to a newly activated partition is the one
329 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
330 * that kthread until the partition is brought down, at which time that kthread
331 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
332 * that XPC has dismantled all communication infrastructure for the associated
333 * partition.) This kthread becomes the channel manager for that partition.
335 * Each active partition has a channel manager, who, besides connecting and
336 * disconnecting channels, will ensure that each of the partition's connected
337 * channels has the required number of assigned kthreads to get the work done.
340 xpc_channel_mgr(struct xpc_partition *part)
342 while (part->act_state != XPC_P_AS_DEACTIVATING ||
343 atomic_read(&part->nchannels_active) > 0 ||
344 !xpc_partition_disengaged(part)) {
346 xpc_process_sent_chctl_flags(part);
349 * Wait until we've been requested to activate kthreads or
350 * all of the channel's message queues have been torn down or
351 * a signal is pending.
353 * The channel_mgr_requests is set to 1 after being awakened,
354 * This is done to prevent the channel mgr from making one pass
355 * through the loop for each request, since he will
356 * be servicing all the requests in one pass. The reason it's
357 * set to 1 instead of 0 is so that other kthreads will know
358 * that the channel mgr is running and won't bother trying to
361 atomic_dec(&part->channel_mgr_requests);
362 (void)wait_event_interruptible(part->channel_mgr_wq,
363 (atomic_read(&part->channel_mgr_requests) > 0 ||
364 part->chctl.all_flags != 0 ||
365 (part->act_state == XPC_P_AS_DEACTIVATING &&
366 atomic_read(&part->nchannels_active) == 0 &&
367 xpc_partition_disengaged(part))));
368 atomic_set(&part->channel_mgr_requests, 1);
373 * Guarantee that the kzalloc'd memory is cacheline aligned.
376 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
378 /* see if kzalloc will give us cachline aligned memory by default */
379 *base = kzalloc(size, flags);
383 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
388 /* nope, we'll have to do it ourselves */
389 *base = kzalloc(size + L1_CACHE_BYTES, flags);
393 return (void *)L1_CACHE_ALIGN((u64)*base);
397 * Setup the channel structures necessary to support XPartition Communication
398 * between the specified remote partition and the local one.
400 static enum xp_retval
401 xpc_setup_ch_structures(struct xpc_partition *part)
405 struct xpc_channel *ch;
406 short partid = XPC_PARTID(part);
409 * Allocate all of the channel structures as a contiguous chunk of
412 DBUG_ON(part->channels != NULL);
413 part->channels = kcalloc(XPC_MAX_NCHANNELS,
414 sizeof(struct xpc_channel),
416 if (part->channels == NULL) {
417 dev_err(xpc_chan, "can't get memory for channels\n");
421 /* allocate the remote open and close args */
423 part->remote_openclose_args =
424 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
426 remote_openclose_args_base);
427 if (part->remote_openclose_args == NULL) {
428 dev_err(xpc_chan, "can't get memory for remote connect args\n");
433 part->chctl.all_flags = 0;
434 spin_lock_init(&part->chctl_lock);
436 atomic_set(&part->channel_mgr_requests, 1);
437 init_waitqueue_head(&part->channel_mgr_wq);
439 part->nchannels = XPC_MAX_NCHANNELS;
441 atomic_set(&part->nchannels_active, 0);
442 atomic_set(&part->nchannels_engaged, 0);
444 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
445 ch = &part->channels[ch_number];
448 ch->number = ch_number;
449 ch->flags = XPC_C_DISCONNECTED;
451 atomic_set(&ch->kthreads_assigned, 0);
452 atomic_set(&ch->kthreads_idle, 0);
453 atomic_set(&ch->kthreads_active, 0);
455 atomic_set(&ch->references, 0);
456 atomic_set(&ch->n_to_notify, 0);
458 spin_lock_init(&ch->lock);
459 init_completion(&ch->wdisconnect_wait);
461 atomic_set(&ch->n_on_msg_allocate_wq, 0);
462 init_waitqueue_head(&ch->msg_allocate_wq);
463 init_waitqueue_head(&ch->idle_wq);
466 ret = xpc_arch_ops.setup_ch_structures(part);
467 if (ret != xpSuccess)
471 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
472 * we're declaring that this partition is ready to go.
474 part->setup_state = XPC_P_SS_SETUP;
478 /* setup of ch structures failed */
480 kfree(part->remote_openclose_args_base);
481 part->remote_openclose_args = NULL;
483 kfree(part->channels);
484 part->channels = NULL;
489 * Teardown the channel structures necessary to support XPartition Communication
490 * between the specified remote partition and the local one.
493 xpc_teardown_ch_structures(struct xpc_partition *part)
495 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
496 DBUG_ON(atomic_read(&part->nchannels_active) != 0);
499 * Make this partition inaccessible to local processes by marking it
500 * as no longer setup. Then wait before proceeding with the teardown
501 * until all existing references cease.
503 DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
504 part->setup_state = XPC_P_SS_WTEARDOWN;
506 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
508 /* now we can begin tearing down the infrastructure */
510 xpc_arch_ops.teardown_ch_structures(part);
512 kfree(part->remote_openclose_args_base);
513 part->remote_openclose_args = NULL;
514 kfree(part->channels);
515 part->channels = NULL;
517 part->setup_state = XPC_P_SS_TORNDOWN;
521 * When XPC HB determines that a partition has come up, it will create a new
522 * kthread and that kthread will call this function to attempt to set up the
523 * basic infrastructure used for Cross Partition Communication with the newly
526 * The kthread that was created by XPC HB and which setup the XPC
527 * infrastructure will remain assigned to the partition becoming the channel
528 * manager for that partition until the partition is deactivating, at which
529 * time the kthread will teardown the XPC infrastructure and then exit.
532 xpc_activating(void *__partid)
534 short partid = (u64)__partid;
535 struct xpc_partition *part = &xpc_partitions[partid];
536 unsigned long irq_flags;
538 DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
540 spin_lock_irqsave(&part->act_lock, irq_flags);
542 if (part->act_state == XPC_P_AS_DEACTIVATING) {
543 part->act_state = XPC_P_AS_INACTIVE;
544 spin_unlock_irqrestore(&part->act_lock, irq_flags);
545 part->remote_rp_pa = 0;
549 /* indicate the thread is activating */
550 DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
551 part->act_state = XPC_P_AS_ACTIVATING;
553 XPC_SET_REASON(part, 0, 0);
554 spin_unlock_irqrestore(&part->act_lock, irq_flags);
556 dev_dbg(xpc_part, "activating partition %d\n", partid);
558 xpc_arch_ops.allow_hb(partid);
560 if (xpc_setup_ch_structures(part) == xpSuccess) {
561 (void)xpc_part_ref(part); /* this will always succeed */
563 if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
564 xpc_mark_partition_active(part);
565 xpc_channel_mgr(part);
566 /* won't return until partition is deactivating */
569 xpc_part_deref(part);
570 xpc_teardown_ch_structures(part);
573 xpc_arch_ops.disallow_hb(partid);
574 xpc_mark_partition_inactive(part);
576 if (part->reason == xpReactivating) {
577 /* interrupting ourselves results in activating partition */
578 xpc_arch_ops.request_partition_reactivation(part);
585 xpc_activate_partition(struct xpc_partition *part)
587 short partid = XPC_PARTID(part);
588 unsigned long irq_flags;
589 struct task_struct *kthread;
591 spin_lock_irqsave(&part->act_lock, irq_flags);
593 DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
595 part->act_state = XPC_P_AS_ACTIVATION_REQ;
596 XPC_SET_REASON(part, xpCloneKThread, __LINE__);
598 spin_unlock_irqrestore(&part->act_lock, irq_flags);
600 kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
602 if (IS_ERR(kthread)) {
603 spin_lock_irqsave(&part->act_lock, irq_flags);
604 part->act_state = XPC_P_AS_INACTIVE;
605 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
606 spin_unlock_irqrestore(&part->act_lock, irq_flags);
611 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
613 int idle = atomic_read(&ch->kthreads_idle);
614 int assigned = atomic_read(&ch->kthreads_assigned);
617 DBUG_ON(needed <= 0);
620 wakeup = (needed > idle) ? idle : needed;
623 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
624 "channel=%d\n", wakeup, ch->partid, ch->number);
626 /* only wakeup the requested number of kthreads */
627 wake_up_nr(&ch->idle_wq, wakeup);
633 if (needed + assigned > ch->kthreads_assigned_limit) {
634 needed = ch->kthreads_assigned_limit - assigned;
639 dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
640 needed, ch->partid, ch->number);
642 xpc_create_kthreads(ch, needed, 0);
646 * This function is where XPC's kthreads wait for messages to deliver.
649 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
651 int (*n_of_deliverable_payloads) (struct xpc_channel *) =
652 xpc_arch_ops.n_of_deliverable_payloads;
655 /* deliver messages to their intended recipients */
657 while (n_of_deliverable_payloads(ch) > 0 &&
658 !(ch->flags & XPC_C_DISCONNECTING)) {
659 xpc_deliver_payload(ch);
662 if (atomic_inc_return(&ch->kthreads_idle) >
663 ch->kthreads_idle_limit) {
664 /* too many idle kthreads on this channel */
665 atomic_dec(&ch->kthreads_idle);
669 dev_dbg(xpc_chan, "idle kthread calling "
670 "wait_event_interruptible_exclusive()\n");
672 (void)wait_event_interruptible_exclusive(ch->idle_wq,
673 (n_of_deliverable_payloads(ch) > 0 ||
674 (ch->flags & XPC_C_DISCONNECTING)));
676 atomic_dec(&ch->kthreads_idle);
678 } while (!(ch->flags & XPC_C_DISCONNECTING));
682 xpc_kthread_start(void *args)
684 short partid = XPC_UNPACK_ARG1(args);
685 u16 ch_number = XPC_UNPACK_ARG2(args);
686 struct xpc_partition *part = &xpc_partitions[partid];
687 struct xpc_channel *ch;
689 unsigned long irq_flags;
690 int (*n_of_deliverable_payloads) (struct xpc_channel *) =
691 xpc_arch_ops.n_of_deliverable_payloads;
693 dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
696 ch = &part->channels[ch_number];
698 if (!(ch->flags & XPC_C_DISCONNECTING)) {
700 /* let registerer know that connection has been established */
702 spin_lock_irqsave(&ch->lock, irq_flags);
703 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
704 ch->flags |= XPC_C_CONNECTEDCALLOUT;
705 spin_unlock_irqrestore(&ch->lock, irq_flags);
707 xpc_connected_callout(ch);
709 spin_lock_irqsave(&ch->lock, irq_flags);
710 ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
711 spin_unlock_irqrestore(&ch->lock, irq_flags);
714 * It is possible that while the callout was being
715 * made that the remote partition sent some messages.
716 * If that is the case, we may need to activate
717 * additional kthreads to help deliver them. We only
718 * need one less than total #of messages to deliver.
720 n_needed = n_of_deliverable_payloads(ch) - 1;
721 if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
722 xpc_activate_kthreads(ch, n_needed);
725 spin_unlock_irqrestore(&ch->lock, irq_flags);
728 xpc_kthread_waitmsgs(part, ch);
731 /* let registerer know that connection is disconnecting */
733 spin_lock_irqsave(&ch->lock, irq_flags);
734 if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
735 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
736 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
737 spin_unlock_irqrestore(&ch->lock, irq_flags);
739 xpc_disconnect_callout(ch, xpDisconnecting);
741 spin_lock_irqsave(&ch->lock, irq_flags);
742 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
744 spin_unlock_irqrestore(&ch->lock, irq_flags);
746 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
747 atomic_dec_return(&part->nchannels_engaged) == 0) {
748 xpc_arch_ops.indicate_partition_disengaged(part);
751 xpc_msgqueue_deref(ch);
753 dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
756 xpc_part_deref(part);
761 * For each partition that XPC has established communications with, there is
762 * a minimum of one kernel thread assigned to perform any operation that
763 * may potentially sleep or block (basically the callouts to the asynchronous
764 * functions registered via xpc_connect()).
766 * Additional kthreads are created and destroyed by XPC as the workload
769 * A kthread is assigned to one of the active channels that exists for a given
773 xpc_create_kthreads(struct xpc_channel *ch, int needed,
774 int ignore_disconnecting)
776 unsigned long irq_flags;
777 u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
778 struct xpc_partition *part = &xpc_partitions[ch->partid];
779 struct task_struct *kthread;
780 void (*indicate_partition_disengaged) (struct xpc_partition *) =
781 xpc_arch_ops.indicate_partition_disengaged;
783 while (needed-- > 0) {
786 * The following is done on behalf of the newly created
787 * kthread. That kthread is responsible for doing the
788 * counterpart to the following before it exits.
790 if (ignore_disconnecting) {
791 if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
792 /* kthreads assigned had gone to zero */
794 XPC_C_DISCONNECTINGCALLOUT_MADE));
798 } else if (ch->flags & XPC_C_DISCONNECTING) {
801 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
802 atomic_inc_return(&part->nchannels_engaged) == 1) {
803 xpc_arch_ops.indicate_partition_engaged(part);
805 (void)xpc_part_ref(part);
806 xpc_msgqueue_ref(ch);
808 kthread = kthread_run(xpc_kthread_start, (void *)args,
809 "xpc%02dc%d", ch->partid, ch->number);
810 if (IS_ERR(kthread)) {
811 /* the fork failed */
814 * NOTE: if (ignore_disconnecting &&
815 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
816 * then we'll deadlock if all other kthreads assigned
817 * to this channel are blocked in the channel's
818 * registerer, because the only thing that will unblock
819 * them is the xpDisconnecting callout that this
820 * failed kthread_run() would have made.
823 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
824 atomic_dec_return(&part->nchannels_engaged) == 0) {
825 indicate_partition_disengaged(part);
827 xpc_msgqueue_deref(ch);
828 xpc_part_deref(part);
830 if (atomic_read(&ch->kthreads_assigned) <
831 ch->kthreads_idle_limit) {
833 * Flag this as an error only if we have an
834 * insufficient #of kthreads for the channel
837 spin_lock_irqsave(&ch->lock, irq_flags);
838 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
840 spin_unlock_irqrestore(&ch->lock, irq_flags);
848 xpc_disconnect_wait(int ch_number)
850 unsigned long irq_flags;
852 struct xpc_partition *part;
853 struct xpc_channel *ch;
854 int wakeup_channel_mgr;
856 /* now wait for all callouts to the caller's function to cease */
857 for (partid = 0; partid < xp_max_npartitions; partid++) {
858 part = &xpc_partitions[partid];
860 if (!xpc_part_ref(part))
863 ch = &part->channels[ch_number];
865 if (!(ch->flags & XPC_C_WDISCONNECT)) {
866 xpc_part_deref(part);
870 wait_for_completion(&ch->wdisconnect_wait);
872 spin_lock_irqsave(&ch->lock, irq_flags);
873 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
874 wakeup_channel_mgr = 0;
876 if (ch->delayed_chctl_flags) {
877 if (part->act_state != XPC_P_AS_DEACTIVATING) {
878 spin_lock(&part->chctl_lock);
879 part->chctl.flags[ch->number] |=
880 ch->delayed_chctl_flags;
881 spin_unlock(&part->chctl_lock);
882 wakeup_channel_mgr = 1;
884 ch->delayed_chctl_flags = 0;
887 ch->flags &= ~XPC_C_WDISCONNECT;
888 spin_unlock_irqrestore(&ch->lock, irq_flags);
890 if (wakeup_channel_mgr)
891 xpc_wakeup_channel_mgr(part);
893 xpc_part_deref(part);
898 xpc_setup_partitions(void)
901 struct xpc_partition *part;
903 xpc_partitions = kcalloc(xp_max_npartitions,
904 sizeof(struct xpc_partition),
906 if (xpc_partitions == NULL) {
907 dev_err(xpc_part, "can't get memory for partition structure\n");
912 * The first few fields of each entry of xpc_partitions[] need to
913 * be initialized now so that calls to xpc_connect() and
914 * xpc_disconnect() can be made prior to the activation of any remote
915 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
916 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
917 * PARTITION HAS BEEN ACTIVATED.
919 for (partid = 0; partid < xp_max_npartitions; partid++) {
920 part = &xpc_partitions[partid];
922 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
924 part->activate_IRQ_rcvd = 0;
925 spin_lock_init(&part->act_lock);
926 part->act_state = XPC_P_AS_INACTIVE;
927 XPC_SET_REASON(part, 0, 0);
929 timer_setup(&part->disengage_timer,
930 xpc_timeout_partition_disengage, 0);
932 part->setup_state = XPC_P_SS_UNSET;
933 init_waitqueue_head(&part->teardown_wq);
934 atomic_set(&part->references, 0);
937 return xpc_arch_ops.setup_partitions();
941 xpc_teardown_partitions(void)
943 xpc_arch_ops.teardown_partitions();
944 kfree(xpc_partitions);
948 xpc_do_exit(enum xp_retval reason)
951 int active_part_count, printed_waiting_msg = 0;
952 struct xpc_partition *part;
953 unsigned long printmsg_time, disengage_timeout = 0;
955 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
956 DBUG_ON(xpc_exiting == 1);
959 * Let the heartbeat checker thread and the discovery thread
960 * (if one is running) know that they should exit. Also wake up
961 * the heartbeat checker thread in case it's sleeping.
964 wake_up_interruptible(&xpc_activate_IRQ_wq);
966 /* wait for the discovery thread to exit */
967 wait_for_completion(&xpc_discovery_exited);
969 /* wait for the heartbeat checker thread to exit */
970 wait_for_completion(&xpc_hb_checker_exited);
972 /* sleep for a 1/3 of a second or so */
973 (void)msleep_interruptible(300);
975 /* wait for all partitions to become inactive */
977 printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
978 xpc_disengage_timedout = 0;
981 active_part_count = 0;
983 for (partid = 0; partid < xp_max_npartitions; partid++) {
984 part = &xpc_partitions[partid];
986 if (xpc_partition_disengaged(part) &&
987 part->act_state == XPC_P_AS_INACTIVE) {
993 XPC_DEACTIVATE_PARTITION(part, reason);
995 if (part->disengage_timeout > disengage_timeout)
996 disengage_timeout = part->disengage_timeout;
999 if (xpc_arch_ops.any_partition_engaged()) {
1000 if (time_is_before_jiffies(printmsg_time)) {
1001 dev_info(xpc_part, "waiting for remote "
1002 "partitions to deactivate, timeout in "
1003 "%ld seconds\n", (disengage_timeout -
1005 printmsg_time = jiffies +
1006 (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
1007 printed_waiting_msg = 1;
1010 } else if (active_part_count > 0) {
1011 if (printed_waiting_msg) {
1012 dev_info(xpc_part, "waiting for local partition"
1013 " to deactivate\n");
1014 printed_waiting_msg = 0;
1018 if (!xpc_disengage_timedout) {
1019 dev_info(xpc_part, "all partitions have "
1025 /* sleep for a 1/3 of a second or so */
1026 (void)msleep_interruptible(300);
1030 DBUG_ON(xpc_arch_ops.any_partition_engaged());
1032 xpc_teardown_rsvd_page();
1034 if (reason == xpUnloading) {
1035 (void)unregister_die_notifier(&xpc_die_notifier);
1036 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1039 /* clear the interface to XPC's functions */
1040 xpc_clear_interface();
1043 unregister_sysctl_table(xpc_sysctl);
1045 xpc_teardown_partitions();
1052 * This function is called when the system is being rebooted.
1055 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1057 enum xp_retval reason;
1061 reason = xpSystemReboot;
1064 reason = xpSystemHalt;
1067 reason = xpSystemPoweroff;
1070 reason = xpSystemGoingDown;
1073 xpc_do_exit(reason);
1077 /* Used to only allow one cpu to complete disconnect */
1078 static unsigned int xpc_die_disconnecting;
1081 * Notify other partitions to deactivate from us by first disengaging from all
1082 * references to our memory.
1085 xpc_die_deactivate(void)
1087 struct xpc_partition *part;
1093 if (cmpxchg(&xpc_die_disconnecting, 0, 1))
1096 /* keep xpc_hb_checker thread from doing anything (just in case) */
1099 xpc_arch_ops.disallow_all_hbs(); /*indicate we're deactivated */
1101 for (partid = 0; partid < xp_max_npartitions; partid++) {
1102 part = &xpc_partitions[partid];
1104 if (xpc_arch_ops.partition_engaged(partid) ||
1105 part->act_state != XPC_P_AS_INACTIVE) {
1106 xpc_arch_ops.request_partition_deactivation(part);
1107 xpc_arch_ops.indicate_partition_disengaged(part);
1112 * Though we requested that all other partitions deactivate from us,
1113 * we only wait until they've all disengaged or we've reached the
1114 * defined timelimit.
1116 * Given that one iteration through the following while-loop takes
1117 * approximately 200 microseconds, calculate the #of loops to take
1118 * before bailing and the #of loops before printing a waiting message.
1120 keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1121 wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1124 any_engaged = xpc_arch_ops.any_partition_engaged();
1126 dev_info(xpc_part, "all partitions have deactivated\n");
1130 if (!keep_waiting--) {
1131 for (partid = 0; partid < xp_max_npartitions;
1133 if (xpc_arch_ops.partition_engaged(partid)) {
1134 dev_info(xpc_part, "deactivate from "
1135 "remote partition %d timed "
1142 if (!wait_to_print--) {
1143 dev_info(xpc_part, "waiting for remote partitions to "
1144 "deactivate, timeout in %ld seconds\n",
1145 keep_waiting / (1000 * 5));
1146 wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1155 * This function is called when the system is being restarted or halted due
1156 * to some sort of system failure. If this is the case we need to notify the
1157 * other partitions to disengage from all references to our memory.
1158 * This function can also be called when our heartbeater could be offlined
1159 * for a time. In this case we need to notify other partitions to not worry
1160 * about the lack of a heartbeat.
1163 xpc_system_die(struct notifier_block *nb, unsigned long event, void *_die_args)
1165 #ifdef CONFIG_IA64 /* !!! temporary kludge */
1167 case DIE_MACHINE_RESTART:
1168 case DIE_MACHINE_HALT:
1169 xpc_die_deactivate();
1172 case DIE_KDEBUG_ENTER:
1173 /* Should lack of heartbeat be ignored by other partitions? */
1174 if (!xpc_kdebug_ignore)
1178 case DIE_MCA_MONARCH_ENTER:
1179 case DIE_INIT_MONARCH_ENTER:
1180 xpc_arch_ops.offline_heartbeat();
1183 case DIE_KDEBUG_LEAVE:
1184 /* Is lack of heartbeat being ignored by other partitions? */
1185 if (!xpc_kdebug_ignore)
1189 case DIE_MCA_MONARCH_LEAVE:
1190 case DIE_INIT_MONARCH_LEAVE:
1191 xpc_arch_ops.online_heartbeat();
1195 struct die_args *die_args = _die_args;
1199 if (die_args->trapnr == X86_TRAP_DF)
1200 xpc_die_deactivate();
1202 if (((die_args->trapnr == X86_TRAP_MF) ||
1203 (die_args->trapnr == X86_TRAP_XF)) &&
1204 !user_mode(die_args->regs))
1205 xpc_die_deactivate();
1214 xpc_die_deactivate();
1225 struct task_struct *kthread;
1227 dev_set_name(xpc_part, "part");
1228 dev_set_name(xpc_chan, "chan");
1230 if (is_uv_system()) {
1231 ret = xpc_init_uv();
1240 ret = xpc_setup_partitions();
1242 dev_err(xpc_part, "can't get memory for partition structure\n");
1246 xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1249 * Fill the partition reserved page with the information needed by
1250 * other partitions to discover we are alive and establish initial
1253 ret = xpc_setup_rsvd_page();
1255 dev_err(xpc_part, "can't setup our reserved page\n");
1259 /* add ourselves to the reboot_notifier_list */
1260 ret = register_reboot_notifier(&xpc_reboot_notifier);
1262 dev_warn(xpc_part, "can't register reboot notifier\n");
1264 /* add ourselves to the die_notifier list */
1265 ret = register_die_notifier(&xpc_die_notifier);
1267 dev_warn(xpc_part, "can't register die notifier\n");
1270 * The real work-horse behind xpc. This processes incoming
1271 * interrupts and monitors remote heartbeats.
1273 kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1274 if (IS_ERR(kthread)) {
1275 dev_err(xpc_part, "failed while forking hb check thread\n");
1281 * Startup a thread that will attempt to discover other partitions to
1282 * activate based on info provided by SAL. This new thread is short
1283 * lived and will exit once discovery is complete.
1285 kthread = kthread_run(xpc_initiate_discovery, NULL,
1286 XPC_DISCOVERY_THREAD_NAME);
1287 if (IS_ERR(kthread)) {
1288 dev_err(xpc_part, "failed while forking discovery thread\n");
1290 /* mark this new thread as a non-starter */
1291 complete(&xpc_discovery_exited);
1293 xpc_do_exit(xpUnloading);
1297 /* set the interface to point at XPC's functions */
1298 xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1299 xpc_initiate_send, xpc_initiate_send_notify,
1300 xpc_initiate_received, xpc_initiate_partid_to_nasids);
1304 /* initialization was not successful */
1306 xpc_teardown_rsvd_page();
1308 (void)unregister_die_notifier(&xpc_die_notifier);
1309 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1312 unregister_sysctl_table(xpc_sysctl);
1314 xpc_teardown_partitions();
1321 module_init(xpc_init);
1326 xpc_do_exit(xpUnloading);
1329 module_exit(xpc_exit);
1331 MODULE_AUTHOR("Silicon Graphics, Inc.");
1332 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1333 MODULE_LICENSE("GPL");
1335 module_param(xpc_hb_interval, int, 0);
1336 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1337 "heartbeat increments.");
1339 module_param(xpc_hb_check_interval, int, 0);
1340 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1341 "heartbeat checks.");
1343 module_param(xpc_disengage_timelimit, int, 0);
1344 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1345 "for disengage to complete.");
1347 module_param(xpc_kdebug_ignore, int, 0);
1348 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1349 "other partitions when dropping into kdebug.");