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 * Copyright (c) 2004-2005 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 * . We currently have no way to determine which nasid an IPI came
30 * from. Thus, xpc_IPI_send() does a remote AMO write followed by
31 * an IPI. The AMO indicates where data is to be pulled from, so
32 * after the IPI arrives, the remote partition checks the AMO word.
33 * The IPI can actually arrive before the AMO however, so other code
34 * must periodically check for this case. Also, remote AMO operations
35 * do not reliably time out. Thus we do a remote PIO read solely to
36 * know whether the remote partition is down and whether we should
37 * stop sending IPIs to it. This remote PIO read operation is set up
38 * in a special nofault region so SAL knows to ignore (and cleanup)
39 * any errors due to the remote AMO write, PIO read, and/or PIO
42 * If/when new hardware solves this IPI problem, we should abandon
43 * the current approach.
48 #include <linux/kernel.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51 #include <linux/sched.h>
52 #include <linux/syscalls.h>
53 #include <linux/cache.h>
54 #include <linux/interrupt.h>
55 #include <linux/slab.h>
56 #include <linux/delay.h>
57 #include <linux/reboot.h>
58 #include <asm/sn/intr.h>
59 #include <asm/sn/sn_sal.h>
60 #include <asm/kdebug.h>
61 #include <asm/uaccess.h>
65 /* define two XPC debug device structures to be used with dev_dbg() et al */
67 struct device_driver xpc_dbg_name = {
71 struct device xpc_part_dbg_subname = {
72 .bus_id = {0}, /* set to "part" at xpc_init() time */
73 .driver = &xpc_dbg_name
76 struct device xpc_chan_dbg_subname = {
77 .bus_id = {0}, /* set to "chan" at xpc_init() time */
78 .driver = &xpc_dbg_name
81 struct device *xpc_part = &xpc_part_dbg_subname;
82 struct device *xpc_chan = &xpc_chan_dbg_subname;
85 /* systune related variables for /proc/sys directories */
87 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
88 static int xpc_hb_min_interval = 1;
89 static int xpc_hb_max_interval = 10;
91 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
92 static int xpc_hb_check_min_interval = 10;
93 static int xpc_hb_check_max_interval = 120;
95 int xpc_disengage_request_timelimit = XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT;
96 static int xpc_disengage_request_min_timelimit = 0;
97 static int xpc_disengage_request_max_timelimit = 120;
99 static ctl_table xpc_sys_xpc_hb_dir[] = {
107 &proc_dointvec_minmax,
110 &xpc_hb_min_interval,
116 &xpc_hb_check_interval,
120 &proc_dointvec_minmax,
123 &xpc_hb_check_min_interval,
124 &xpc_hb_check_max_interval
128 static ctl_table xpc_sys_xpc_dir[] = {
139 "disengage_request_timelimit",
140 &xpc_disengage_request_timelimit,
144 &proc_dointvec_minmax,
147 &xpc_disengage_request_min_timelimit,
148 &xpc_disengage_request_max_timelimit
152 static ctl_table xpc_sys_dir[] = {
163 static struct ctl_table_header *xpc_sysctl;
165 /* non-zero if any remote partition disengage request was timed out */
166 int xpc_disengage_request_timedout;
168 /* #of IRQs received */
169 static atomic_t xpc_act_IRQ_rcvd;
171 /* IRQ handler notifies this wait queue on receipt of an IRQ */
172 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq);
174 static unsigned long xpc_hb_check_timeout;
176 /* notification that the xpc_hb_checker thread has exited */
177 static DECLARE_MUTEX_LOCKED(xpc_hb_checker_exited);
179 /* notification that the xpc_discovery thread has exited */
180 static DECLARE_MUTEX_LOCKED(xpc_discovery_exited);
183 static struct timer_list xpc_hb_timer;
186 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
189 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
190 static struct notifier_block xpc_reboot_notifier = {
191 .notifier_call = xpc_system_reboot,
194 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
195 static struct notifier_block xpc_die_notifier = {
196 .notifier_call = xpc_system_die,
201 * Timer function to enforce the timelimit on the partition disengage request.
204 xpc_timeout_partition_disengage_request(unsigned long data)
206 struct xpc_partition *part = (struct xpc_partition *) data;
209 DBUG_ON(jiffies < part->disengage_request_timeout);
211 (void) xpc_partition_disengaged(part);
213 DBUG_ON(part->disengage_request_timeout != 0);
214 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part)) != 0);
219 * Notify the heartbeat check thread that an IRQ has been received.
222 xpc_act_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
224 atomic_inc(&xpc_act_IRQ_rcvd);
225 wake_up_interruptible(&xpc_act_IRQ_wq);
231 * Timer to produce the heartbeat. The timer structures function is
232 * already set when this is initially called. A tunable is used to
233 * specify when the next timeout should occur.
236 xpc_hb_beater(unsigned long dummy)
238 xpc_vars->heartbeat++;
240 if (jiffies >= xpc_hb_check_timeout) {
241 wake_up_interruptible(&xpc_act_IRQ_wq);
244 xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
245 add_timer(&xpc_hb_timer);
250 * This thread is responsible for nearly all of the partition
251 * activation/deactivation.
254 xpc_hb_checker(void *ignore)
256 int last_IRQ_count = 0;
261 /* this thread was marked active by xpc_hb_init() */
263 daemonize(XPC_HB_CHECK_THREAD_NAME);
265 set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU));
267 xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
269 while (!(volatile int) xpc_exiting) {
271 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
273 (int) (xpc_hb_check_timeout - jiffies),
274 atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count);
277 /* checking of remote heartbeats is skewed by IRQ handling */
278 if (jiffies >= xpc_hb_check_timeout) {
279 dev_dbg(xpc_part, "checking remote heartbeats\n");
280 xpc_check_remote_hb();
283 * We need to periodically recheck to ensure no
284 * IPI/AMO pairs have been missed. That check
285 * must always reset xpc_hb_check_timeout.
291 /* check for outstanding IRQs */
292 new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd);
293 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
296 dev_dbg(xpc_part, "found an IRQ to process; will be "
297 "resetting xpc_hb_check_timeout\n");
299 last_IRQ_count += xpc_identify_act_IRQ_sender();
300 if (last_IRQ_count < new_IRQ_count) {
301 /* retry once to help avoid missing AMO */
302 (void) xpc_identify_act_IRQ_sender();
304 last_IRQ_count = new_IRQ_count;
306 xpc_hb_check_timeout = jiffies +
307 (xpc_hb_check_interval * HZ);
310 /* wait for IRQ or timeout */
311 (void) wait_event_interruptible(xpc_act_IRQ_wq,
312 (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) ||
313 jiffies >= xpc_hb_check_timeout ||
314 (volatile int) xpc_exiting));
317 dev_dbg(xpc_part, "heartbeat checker is exiting\n");
320 /* mark this thread as having exited */
321 up(&xpc_hb_checker_exited);
327 * This thread will attempt to discover other partitions to activate
328 * based on info provided by SAL. This new thread is short lived and
329 * will exit once discovery is complete.
332 xpc_initiate_discovery(void *ignore)
334 daemonize(XPC_DISCOVERY_THREAD_NAME);
338 dev_dbg(xpc_part, "discovery thread is exiting\n");
340 /* mark this thread as having exited */
341 up(&xpc_discovery_exited);
347 * Establish first contact with the remote partititon. This involves pulling
348 * the XPC per partition variables from the remote partition and waiting for
349 * the remote partition to pull ours.
351 static enum xpc_retval
352 xpc_make_first_contact(struct xpc_partition *part)
357 while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) {
358 if (ret != xpcRetry) {
359 XPC_DEACTIVATE_PARTITION(part, ret);
363 dev_dbg(xpc_chan, "waiting to make first contact with "
364 "partition %d\n", XPC_PARTID(part));
366 /* wait a 1/4 of a second or so */
367 (void) msleep_interruptible(250);
369 if (part->act_state == XPC_P_DEACTIVATING) {
374 return xpc_mark_partition_active(part);
379 * The first kthread assigned to a newly activated partition is the one
380 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
381 * that kthread until the partition is brought down, at which time that kthread
382 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
383 * that XPC has dismantled all communication infrastructure for the associated
384 * partition.) This kthread becomes the channel manager for that partition.
386 * Each active partition has a channel manager, who, besides connecting and
387 * disconnecting channels, will ensure that each of the partition's connected
388 * channels has the required number of assigned kthreads to get the work done.
391 xpc_channel_mgr(struct xpc_partition *part)
393 while (part->act_state != XPC_P_DEACTIVATING ||
394 atomic_read(&part->nchannels_active) > 0 ||
395 !xpc_partition_disengaged(part)) {
397 xpc_process_channel_activity(part);
401 * Wait until we've been requested to activate kthreads or
402 * all of the channel's message queues have been torn down or
403 * a signal is pending.
405 * The channel_mgr_requests is set to 1 after being awakened,
406 * This is done to prevent the channel mgr from making one pass
407 * through the loop for each request, since he will
408 * be servicing all the requests in one pass. The reason it's
409 * set to 1 instead of 0 is so that other kthreads will know
410 * that the channel mgr is running and won't bother trying to
413 atomic_dec(&part->channel_mgr_requests);
414 (void) wait_event_interruptible(part->channel_mgr_wq,
415 (atomic_read(&part->channel_mgr_requests) > 0 ||
416 (volatile u64) part->local_IPI_amo != 0 ||
417 ((volatile u8) part->act_state ==
418 XPC_P_DEACTIVATING &&
419 atomic_read(&part->nchannels_active) == 0 &&
420 xpc_partition_disengaged(part))));
421 atomic_set(&part->channel_mgr_requests, 1);
423 // >>> Does it need to wakeup periodically as well? In case we
424 // >>> miscalculated the #of kthreads to wakeup or create?
430 * When XPC HB determines that a partition has come up, it will create a new
431 * kthread and that kthread will call this function to attempt to set up the
432 * basic infrastructure used for Cross Partition Communication with the newly
435 * The kthread that was created by XPC HB and which setup the XPC
436 * infrastructure will remain assigned to the partition until the partition
437 * goes down. At which time the kthread will teardown the XPC infrastructure
440 * XPC HB will put the remote partition's XPC per partition specific variables
441 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
442 * calling xpc_partition_up().
445 xpc_partition_up(struct xpc_partition *part)
447 DBUG_ON(part->channels != NULL);
449 dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part));
451 if (xpc_setup_infrastructure(part) != xpcSuccess) {
456 * The kthread that XPC HB called us with will become the
457 * channel manager for this partition. It will not return
458 * back to XPC HB until the partition's XPC infrastructure
459 * has been dismantled.
462 (void) xpc_part_ref(part); /* this will always succeed */
464 if (xpc_make_first_contact(part) == xpcSuccess) {
465 xpc_channel_mgr(part);
468 xpc_part_deref(part);
470 xpc_teardown_infrastructure(part);
475 xpc_activating(void *__partid)
477 partid_t partid = (u64) __partid;
478 struct xpc_partition *part = &xpc_partitions[partid];
479 unsigned long irq_flags;
480 struct sched_param param = { sched_priority: MAX_RT_PRIO - 1 };
484 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
486 spin_lock_irqsave(&part->act_lock, irq_flags);
488 if (part->act_state == XPC_P_DEACTIVATING) {
489 part->act_state = XPC_P_INACTIVE;
490 spin_unlock_irqrestore(&part->act_lock, irq_flags);
491 part->remote_rp_pa = 0;
495 /* indicate the thread is activating */
496 DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
497 part->act_state = XPC_P_ACTIVATING;
499 XPC_SET_REASON(part, 0, 0);
500 spin_unlock_irqrestore(&part->act_lock, irq_flags);
502 dev_dbg(xpc_part, "bringing partition %d up\n", partid);
504 daemonize("xpc%02d", partid);
507 * This thread needs to run at a realtime priority to prevent a
508 * significant performance degradation.
510 ret = sched_setscheduler(current, SCHED_FIFO, ¶m);
512 dev_warn(xpc_part, "unable to set pid %d to a realtime "
513 "priority, ret=%d\n", current->pid, ret);
516 /* allow this thread and its children to run on any CPU */
517 set_cpus_allowed(current, CPU_MASK_ALL);
520 * Register the remote partition's AMOs with SAL so it can handle
521 * and cleanup errors within that address range should the remote
522 * partition go down. We don't unregister this range because it is
523 * difficult to tell when outstanding writes to the remote partition
524 * are finished and thus when it is safe to unregister. This should
525 * not result in wasted space in the SAL xp_addr_region table because
526 * we should get the same page for remote_amos_page_pa after module
527 * reloads and system reboots.
529 if (sn_register_xp_addr_region(part->remote_amos_page_pa,
531 dev_warn(xpc_part, "xpc_partition_up(%d) failed to register "
532 "xp_addr region\n", partid);
534 spin_lock_irqsave(&part->act_lock, irq_flags);
535 part->act_state = XPC_P_INACTIVE;
536 XPC_SET_REASON(part, xpcPhysAddrRegFailed, __LINE__);
537 spin_unlock_irqrestore(&part->act_lock, irq_flags);
538 part->remote_rp_pa = 0;
542 xpc_allow_hb(partid, xpc_vars);
543 xpc_IPI_send_activated(part);
547 * xpc_partition_up() holds this thread and marks this partition as
548 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
550 (void) xpc_partition_up(part);
552 xpc_disallow_hb(partid, xpc_vars);
553 xpc_mark_partition_inactive(part);
555 if (part->reason == xpcReactivating) {
556 /* interrupting ourselves results in activating partition */
557 xpc_IPI_send_reactivate(part);
565 xpc_activate_partition(struct xpc_partition *part)
567 partid_t partid = XPC_PARTID(part);
568 unsigned long irq_flags;
572 spin_lock_irqsave(&part->act_lock, irq_flags);
574 pid = kernel_thread(xpc_activating, (void *) ((u64) partid), 0);
576 DBUG_ON(part->act_state != XPC_P_INACTIVE);
579 part->act_state = XPC_P_ACTIVATION_REQ;
580 XPC_SET_REASON(part, xpcCloneKThread, __LINE__);
582 XPC_SET_REASON(part, xpcCloneKThreadFailed, __LINE__);
585 spin_unlock_irqrestore(&part->act_lock, irq_flags);
590 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
591 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
592 * than one partition, we use an AMO_t structure per partition to indicate
593 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
594 * associated kthread to handle it.
596 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
597 * running on other partitions.
599 * Noteworthy Arguments:
601 * irq - Interrupt ReQuest number. NOT USED.
603 * dev_id - partid of IPI's potential sender.
605 * regs - processor's context before the processor entered
606 * interrupt code. NOT USED.
609 xpc_notify_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
611 partid_t partid = (partid_t) (u64) dev_id;
612 struct xpc_partition *part = &xpc_partitions[partid];
615 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
617 if (xpc_part_ref(part)) {
618 xpc_check_for_channel_activity(part);
620 xpc_part_deref(part);
627 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
628 * because the write to their associated IPI amo completed after the IRQ/IPI
632 xpc_dropped_IPI_check(struct xpc_partition *part)
634 if (xpc_part_ref(part)) {
635 xpc_check_for_channel_activity(part);
637 part->dropped_IPI_timer.expires = jiffies +
638 XPC_P_DROPPED_IPI_WAIT;
639 add_timer(&part->dropped_IPI_timer);
640 xpc_part_deref(part);
646 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
648 int idle = atomic_read(&ch->kthreads_idle);
649 int assigned = atomic_read(&ch->kthreads_assigned);
653 DBUG_ON(needed <= 0);
656 wakeup = (needed > idle) ? idle : needed;
659 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
660 "channel=%d\n", wakeup, ch->partid, ch->number);
662 /* only wakeup the requested number of kthreads */
663 wake_up_nr(&ch->idle_wq, wakeup);
670 if (needed + assigned > ch->kthreads_assigned_limit) {
671 needed = ch->kthreads_assigned_limit - assigned;
672 // >>>should never be less than 0
678 dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
679 needed, ch->partid, ch->number);
681 xpc_create_kthreads(ch, needed);
686 * This function is where XPC's kthreads wait for messages to deliver.
689 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
692 /* deliver messages to their intended recipients */
694 while ((volatile s64) ch->w_local_GP.get <
695 (volatile s64) ch->w_remote_GP.put &&
696 !((volatile u32) ch->flags &
697 XPC_C_DISCONNECTING)) {
701 if (atomic_inc_return(&ch->kthreads_idle) >
702 ch->kthreads_idle_limit) {
703 /* too many idle kthreads on this channel */
704 atomic_dec(&ch->kthreads_idle);
708 dev_dbg(xpc_chan, "idle kthread calling "
709 "wait_event_interruptible_exclusive()\n");
711 (void) wait_event_interruptible_exclusive(ch->idle_wq,
712 ((volatile s64) ch->w_local_GP.get <
713 (volatile s64) ch->w_remote_GP.put ||
714 ((volatile u32) ch->flags &
715 XPC_C_DISCONNECTING)));
717 atomic_dec(&ch->kthreads_idle);
719 } while (!((volatile u32) ch->flags & XPC_C_DISCONNECTING));
724 xpc_daemonize_kthread(void *args)
726 partid_t partid = XPC_UNPACK_ARG1(args);
727 u16 ch_number = XPC_UNPACK_ARG2(args);
728 struct xpc_partition *part = &xpc_partitions[partid];
729 struct xpc_channel *ch;
731 unsigned long irq_flags;
734 daemonize("xpc%02dc%d", partid, ch_number);
736 dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
739 ch = &part->channels[ch_number];
741 if (!(ch->flags & XPC_C_DISCONNECTING)) {
743 /* let registerer know that connection has been established */
745 spin_lock_irqsave(&ch->lock, irq_flags);
746 if (!(ch->flags & XPC_C_CONNECTCALLOUT)) {
747 ch->flags |= XPC_C_CONNECTCALLOUT;
748 spin_unlock_irqrestore(&ch->lock, irq_flags);
750 xpc_connected_callout(ch);
753 * It is possible that while the callout was being
754 * made that the remote partition sent some messages.
755 * If that is the case, we may need to activate
756 * additional kthreads to help deliver them. We only
757 * need one less than total #of messages to deliver.
759 n_needed = ch->w_remote_GP.put - ch->w_local_GP.get - 1;
761 !(ch->flags & XPC_C_DISCONNECTING)) {
762 xpc_activate_kthreads(ch, n_needed);
765 spin_unlock_irqrestore(&ch->lock, irq_flags);
768 xpc_kthread_waitmsgs(part, ch);
771 if (atomic_dec_return(&ch->kthreads_assigned) == 0) {
772 spin_lock_irqsave(&ch->lock, irq_flags);
773 if ((ch->flags & XPC_C_CONNECTCALLOUT) &&
774 !(ch->flags & XPC_C_DISCONNECTCALLOUT)) {
775 ch->flags |= XPC_C_DISCONNECTCALLOUT;
776 spin_unlock_irqrestore(&ch->lock, irq_flags);
778 xpc_disconnect_callout(ch, xpcDisconnecting);
780 spin_unlock_irqrestore(&ch->lock, irq_flags);
782 if (atomic_dec_return(&part->nchannels_engaged) == 0) {
783 xpc_mark_partition_disengaged(part);
784 xpc_IPI_send_disengage(part);
789 xpc_msgqueue_deref(ch);
791 dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
794 xpc_part_deref(part);
800 * For each partition that XPC has established communications with, there is
801 * a minimum of one kernel thread assigned to perform any operation that
802 * may potentially sleep or block (basically the callouts to the asynchronous
803 * functions registered via xpc_connect()).
805 * Additional kthreads are created and destroyed by XPC as the workload
808 * A kthread is assigned to one of the active channels that exists for a given
812 xpc_create_kthreads(struct xpc_channel *ch, int needed)
814 unsigned long irq_flags;
816 u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
817 struct xpc_partition *part = &xpc_partitions[ch->partid];
820 while (needed-- > 0) {
823 * The following is done on behalf of the newly created
824 * kthread. That kthread is responsible for doing the
825 * counterpart to the following before it exits.
827 (void) xpc_part_ref(part);
828 xpc_msgqueue_ref(ch);
829 if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
830 atomic_inc_return(&part->nchannels_engaged) == 1) {
831 xpc_mark_partition_engaged(part);
834 pid = kernel_thread(xpc_daemonize_kthread, (void *) args, 0);
836 /* the fork failed */
837 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
838 atomic_dec_return(&part->nchannels_engaged) == 0) {
839 xpc_mark_partition_disengaged(part);
840 xpc_IPI_send_disengage(part);
842 xpc_msgqueue_deref(ch);
843 xpc_part_deref(part);
845 if (atomic_read(&ch->kthreads_assigned) <
846 ch->kthreads_idle_limit) {
848 * Flag this as an error only if we have an
849 * insufficient #of kthreads for the channel
852 * No xpc_msgqueue_ref() is needed here since
853 * the channel mgr is doing this.
855 spin_lock_irqsave(&ch->lock, irq_flags);
856 XPC_DISCONNECT_CHANNEL(ch, xpcLackOfResources,
858 spin_unlock_irqrestore(&ch->lock, irq_flags);
863 ch->kthreads_created++; // >>> temporary debug only!!!
869 xpc_disconnect_wait(int ch_number)
871 unsigned long irq_flags;
873 struct xpc_partition *part;
874 struct xpc_channel *ch;
875 int wakeup_channel_mgr;
878 /* now wait for all callouts to the caller's function to cease */
879 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
880 part = &xpc_partitions[partid];
882 if (!xpc_part_ref(part)) {
886 ch = &part->channels[ch_number];
888 if (!(ch->flags & XPC_C_WDISCONNECT)) {
889 xpc_part_deref(part);
893 (void) down(&ch->wdisconnect_sema);
895 spin_lock_irqsave(&ch->lock, irq_flags);
896 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
897 wakeup_channel_mgr = 0;
899 if (ch->delayed_IPI_flags) {
900 if (part->act_state != XPC_P_DEACTIVATING) {
901 spin_lock(&part->IPI_lock);
902 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
903 ch->number, ch->delayed_IPI_flags);
904 spin_unlock(&part->IPI_lock);
905 wakeup_channel_mgr = 1;
907 ch->delayed_IPI_flags = 0;
910 ch->flags &= ~XPC_C_WDISCONNECT;
911 spin_unlock_irqrestore(&ch->lock, irq_flags);
913 if (wakeup_channel_mgr) {
914 xpc_wakeup_channel_mgr(part);
917 xpc_part_deref(part);
923 xpc_do_exit(enum xpc_retval reason)
926 int active_part_count, printed_waiting_msg = 0;
927 struct xpc_partition *part;
928 unsigned long printmsg_time, disengage_request_timeout = 0;
931 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
932 DBUG_ON(xpc_exiting == 1);
935 * Let the heartbeat checker thread and the discovery thread
936 * (if one is running) know that they should exit. Also wake up
937 * the heartbeat checker thread in case it's sleeping.
940 wake_up_interruptible(&xpc_act_IRQ_wq);
942 /* ignore all incoming interrupts */
943 free_irq(SGI_XPC_ACTIVATE, NULL);
945 /* wait for the discovery thread to exit */
946 down(&xpc_discovery_exited);
948 /* wait for the heartbeat checker thread to exit */
949 down(&xpc_hb_checker_exited);
952 /* sleep for a 1/3 of a second or so */
953 (void) msleep_interruptible(300);
956 /* wait for all partitions to become inactive */
958 printmsg_time = jiffies + (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
959 xpc_disengage_request_timedout = 0;
962 active_part_count = 0;
964 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
965 part = &xpc_partitions[partid];
967 if (xpc_partition_disengaged(part) &&
968 part->act_state == XPC_P_INACTIVE) {
974 XPC_DEACTIVATE_PARTITION(part, reason);
976 if (part->disengage_request_timeout >
977 disengage_request_timeout) {
978 disengage_request_timeout =
979 part->disengage_request_timeout;
983 if (xpc_partition_engaged(-1UL)) {
984 if (time_after(jiffies, printmsg_time)) {
985 dev_info(xpc_part, "waiting for remote "
986 "partitions to disengage, timeout in "
988 (disengage_request_timeout - jiffies)
990 printmsg_time = jiffies +
991 (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
992 printed_waiting_msg = 1;
995 } else if (active_part_count > 0) {
996 if (printed_waiting_msg) {
997 dev_info(xpc_part, "waiting for local partition"
999 printed_waiting_msg = 0;
1003 if (!xpc_disengage_request_timedout) {
1004 dev_info(xpc_part, "all partitions have "
1010 /* sleep for a 1/3 of a second or so */
1011 (void) msleep_interruptible(300);
1015 DBUG_ON(xpc_partition_engaged(-1UL));
1018 /* indicate to others that our reserved page is uninitialized */
1019 xpc_rsvd_page->vars_pa = 0;
1021 /* now it's time to eliminate our heartbeat */
1022 del_timer_sync(&xpc_hb_timer);
1023 DBUG_ON(xpc_vars->heartbeating_to_mask != 0);
1025 if (reason == xpcUnloading) {
1026 /* take ourselves off of the reboot_notifier_list */
1027 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1029 /* take ourselves off of the die_notifier list */
1030 (void) unregister_die_notifier(&xpc_die_notifier);
1033 /* close down protections for IPI operations */
1034 xpc_restrict_IPI_ops();
1037 /* clear the interface to XPC's functions */
1038 xpc_clear_interface();
1041 unregister_sysctl_table(xpc_sysctl);
1047 * Called when the system is about to be either restarted or halted.
1050 xpc_die_disengage(void)
1052 struct xpc_partition *part;
1054 unsigned long engaged;
1055 long time, printmsg_time, disengage_request_timeout;
1058 /* keep xpc_hb_checker thread from doing anything (just in case) */
1061 xpc_vars->heartbeating_to_mask = 0; /* indicate we're deactivated */
1063 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1064 part = &xpc_partitions[partid];
1066 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part->
1067 remote_vars_version)) {
1069 /* just in case it was left set by an earlier XPC */
1070 xpc_clear_partition_engaged(1UL << partid);
1074 if (xpc_partition_engaged(1UL << partid) ||
1075 part->act_state != XPC_P_INACTIVE) {
1076 xpc_request_partition_disengage(part);
1077 xpc_mark_partition_disengaged(part);
1078 xpc_IPI_send_disengage(part);
1083 printmsg_time = time +
1084 (XPC_DISENGAGE_PRINTMSG_INTERVAL * sn_rtc_cycles_per_second);
1085 disengage_request_timeout = time +
1086 (xpc_disengage_request_timelimit * sn_rtc_cycles_per_second);
1088 /* wait for all other partitions to disengage from us */
1091 engaged = xpc_partition_engaged(-1UL);
1093 dev_info(xpc_part, "all partitions have disengaged\n");
1098 if (time >= disengage_request_timeout) {
1099 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1100 if (engaged & (1UL << partid)) {
1101 dev_info(xpc_part, "disengage from "
1102 "remote partition %d timed "
1109 if (time >= printmsg_time) {
1110 dev_info(xpc_part, "waiting for remote partitions to "
1111 "disengage, timeout in %ld seconds\n",
1112 (disengage_request_timeout - time) /
1113 sn_rtc_cycles_per_second);
1114 printmsg_time = time +
1115 (XPC_DISENGAGE_PRINTMSG_INTERVAL *
1116 sn_rtc_cycles_per_second);
1123 * This function is called when the system is being rebooted.
1126 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1128 enum xpc_retval reason;
1133 reason = xpcSystemReboot;
1136 reason = xpcSystemHalt;
1139 reason = xpcSystemPoweroff;
1142 reason = xpcSystemGoingDown;
1145 xpc_do_exit(reason);
1151 * This function is called when the system is being rebooted.
1154 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1157 case DIE_MACHINE_RESTART:
1158 case DIE_MACHINE_HALT:
1159 xpc_die_disengage();
1161 case DIE_MCA_MONARCH_ENTER:
1162 case DIE_INIT_MONARCH_ENTER:
1163 xpc_vars->heartbeat++;
1164 xpc_vars->heartbeat_offline = 1;
1166 case DIE_MCA_MONARCH_LEAVE:
1167 case DIE_INIT_MONARCH_LEAVE:
1168 xpc_vars->heartbeat++;
1169 xpc_vars->heartbeat_offline = 0;
1182 struct xpc_partition *part;
1186 if (!ia64_platform_is("sn2")) {
1191 * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng
1192 * various portions of a partition's reserved page. Its size is based
1193 * on the size of the reserved page header and part_nasids mask. So we
1194 * need to ensure that the other items will fit as well.
1196 if (XPC_RP_VARS_SIZE > XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES) {
1197 dev_err(xpc_part, "xpc_remote_copy_buffer is not big enough\n");
1200 DBUG_ON((u64) xpc_remote_copy_buffer !=
1201 L1_CACHE_ALIGN((u64) xpc_remote_copy_buffer));
1203 snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1204 snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1206 xpc_sysctl = register_sysctl_table(xpc_sys_dir, 1);
1209 * The first few fields of each entry of xpc_partitions[] need to
1210 * be initialized now so that calls to xpc_connect() and
1211 * xpc_disconnect() can be made prior to the activation of any remote
1212 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1213 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1214 * PARTITION HAS BEEN ACTIVATED.
1216 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1217 part = &xpc_partitions[partid];
1219 DBUG_ON((u64) part != L1_CACHE_ALIGN((u64) part));
1221 part->act_IRQ_rcvd = 0;
1222 spin_lock_init(&part->act_lock);
1223 part->act_state = XPC_P_INACTIVE;
1224 XPC_SET_REASON(part, 0, 0);
1226 init_timer(&part->disengage_request_timer);
1227 part->disengage_request_timer.function =
1228 xpc_timeout_partition_disengage_request;
1229 part->disengage_request_timer.data = (unsigned long) part;
1231 part->setup_state = XPC_P_UNSET;
1232 init_waitqueue_head(&part->teardown_wq);
1233 atomic_set(&part->references, 0);
1237 * Open up protections for IPI operations (and AMO operations on
1238 * Shub 1.1 systems).
1240 xpc_allow_IPI_ops();
1243 * Interrupts being processed will increment this atomic variable and
1244 * awaken the heartbeat thread which will process the interrupts.
1246 atomic_set(&xpc_act_IRQ_rcvd, 0);
1249 * This is safe to do before the xpc_hb_checker thread has started
1250 * because the handler releases a wait queue. If an interrupt is
1251 * received before the thread is waiting, it will not go to sleep,
1252 * but rather immediately process the interrupt.
1254 ret = request_irq(SGI_XPC_ACTIVATE, xpc_act_IRQ_handler, 0,
1257 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
1258 "errno=%d\n", -ret);
1260 xpc_restrict_IPI_ops();
1263 unregister_sysctl_table(xpc_sysctl);
1269 * Fill the partition reserved page with the information needed by
1270 * other partitions to discover we are alive and establish initial
1273 xpc_rsvd_page = xpc_rsvd_page_init();
1274 if (xpc_rsvd_page == NULL) {
1275 dev_err(xpc_part, "could not setup our reserved page\n");
1277 free_irq(SGI_XPC_ACTIVATE, NULL);
1278 xpc_restrict_IPI_ops();
1281 unregister_sysctl_table(xpc_sysctl);
1287 /* add ourselves to the reboot_notifier_list */
1288 ret = register_reboot_notifier(&xpc_reboot_notifier);
1290 dev_warn(xpc_part, "can't register reboot notifier\n");
1293 /* add ourselves to the die_notifier list (i.e., ia64die_chain) */
1294 ret = register_die_notifier(&xpc_die_notifier);
1296 dev_warn(xpc_part, "can't register die notifier\n");
1301 * Set the beating to other partitions into motion. This is
1302 * the last requirement for other partitions' discovery to
1303 * initiate communications with us.
1305 init_timer(&xpc_hb_timer);
1306 xpc_hb_timer.function = xpc_hb_beater;
1311 * The real work-horse behind xpc. This processes incoming
1312 * interrupts and monitors remote heartbeats.
1314 pid = kernel_thread(xpc_hb_checker, NULL, 0);
1316 dev_err(xpc_part, "failed while forking hb check thread\n");
1318 /* indicate to others that our reserved page is uninitialized */
1319 xpc_rsvd_page->vars_pa = 0;
1321 /* take ourselves off of the reboot_notifier_list */
1322 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1324 /* take ourselves off of the die_notifier list */
1325 (void) unregister_die_notifier(&xpc_die_notifier);
1327 del_timer_sync(&xpc_hb_timer);
1328 free_irq(SGI_XPC_ACTIVATE, NULL);
1329 xpc_restrict_IPI_ops();
1332 unregister_sysctl_table(xpc_sysctl);
1339 * Startup a thread that will attempt to discover other partitions to
1340 * activate based on info provided by SAL. This new thread is short
1341 * lived and will exit once discovery is complete.
1343 pid = kernel_thread(xpc_initiate_discovery, NULL, 0);
1345 dev_err(xpc_part, "failed while forking discovery thread\n");
1347 /* mark this new thread as a non-starter */
1348 up(&xpc_discovery_exited);
1350 xpc_do_exit(xpcUnloading);
1355 /* set the interface to point at XPC's functions */
1356 xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1357 xpc_initiate_allocate, xpc_initiate_send,
1358 xpc_initiate_send_notify, xpc_initiate_received,
1359 xpc_initiate_partid_to_nasids);
1363 module_init(xpc_init);
1369 xpc_do_exit(xpcUnloading);
1371 module_exit(xpc_exit);
1374 MODULE_AUTHOR("Silicon Graphics, Inc.");
1375 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1376 MODULE_LICENSE("GPL");
1378 module_param(xpc_hb_interval, int, 0);
1379 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1380 "heartbeat increments.");
1382 module_param(xpc_hb_check_interval, int, 0);
1383 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1384 "heartbeat checks.");
1386 module_param(xpc_disengage_request_timelimit, int, 0);
1387 MODULE_PARM_DESC(xpc_disengage_request_timelimit, "Number of seconds to wait "
1388 "for disengage request to complete.");