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) 2004-2006 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 <linux/completion.h>
59 #include <asm/sn/intr.h>
60 #include <asm/sn/sn_sal.h>
61 #include <asm/kdebug.h>
62 #include <asm/uaccess.h>
63 #include <asm/sn/xpc.h>
66 /* define two XPC debug device structures to be used with dev_dbg() et al */
68 struct device_driver xpc_dbg_name = {
72 struct device xpc_part_dbg_subname = {
73 .bus_id = {0}, /* set to "part" at xpc_init() time */
74 .driver = &xpc_dbg_name
77 struct device xpc_chan_dbg_subname = {
78 .bus_id = {0}, /* set to "chan" at xpc_init() time */
79 .driver = &xpc_dbg_name
82 struct device *xpc_part = &xpc_part_dbg_subname;
83 struct device *xpc_chan = &xpc_chan_dbg_subname;
86 static int xpc_kdebug_ignore;
89 /* systune related variables for /proc/sys directories */
91 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
92 static int xpc_hb_min_interval = 1;
93 static int xpc_hb_max_interval = 10;
95 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
96 static int xpc_hb_check_min_interval = 10;
97 static int xpc_hb_check_max_interval = 120;
99 int xpc_disengage_request_timelimit = XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT;
100 static int xpc_disengage_request_min_timelimit = 0;
101 static int xpc_disengage_request_max_timelimit = 120;
103 static ctl_table xpc_sys_xpc_hb_dir[] = {
111 &proc_dointvec_minmax,
115 &xpc_hb_min_interval,
121 &xpc_hb_check_interval,
125 &proc_dointvec_minmax,
129 &xpc_hb_check_min_interval,
130 &xpc_hb_check_max_interval
134 static ctl_table xpc_sys_xpc_dir[] = {
145 "disengage_request_timelimit",
146 &xpc_disengage_request_timelimit,
150 &proc_dointvec_minmax,
154 &xpc_disengage_request_min_timelimit,
155 &xpc_disengage_request_max_timelimit
159 static ctl_table xpc_sys_dir[] = {
170 static struct ctl_table_header *xpc_sysctl;
172 /* non-zero if any remote partition disengage request was timed out */
173 int xpc_disengage_request_timedout;
175 /* #of IRQs received */
176 static atomic_t xpc_act_IRQ_rcvd;
178 /* IRQ handler notifies this wait queue on receipt of an IRQ */
179 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq);
181 static unsigned long xpc_hb_check_timeout;
183 /* notification that the xpc_hb_checker thread has exited */
184 static DECLARE_COMPLETION(xpc_hb_checker_exited);
186 /* notification that the xpc_discovery thread has exited */
187 static DECLARE_COMPLETION(xpc_discovery_exited);
190 static struct timer_list xpc_hb_timer;
193 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
196 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
197 static struct notifier_block xpc_reboot_notifier = {
198 .notifier_call = xpc_system_reboot,
201 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
202 static struct notifier_block xpc_die_notifier = {
203 .notifier_call = xpc_system_die,
208 * Timer function to enforce the timelimit on the partition disengage request.
211 xpc_timeout_partition_disengage_request(unsigned long data)
213 struct xpc_partition *part = (struct xpc_partition *) data;
216 DBUG_ON(jiffies < part->disengage_request_timeout);
218 (void) xpc_partition_disengaged(part);
220 DBUG_ON(part->disengage_request_timeout != 0);
221 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part)) != 0);
226 * Notify the heartbeat check thread that an IRQ has been received.
229 xpc_act_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
231 atomic_inc(&xpc_act_IRQ_rcvd);
232 wake_up_interruptible(&xpc_act_IRQ_wq);
238 * Timer to produce the heartbeat. The timer structures function is
239 * already set when this is initially called. A tunable is used to
240 * specify when the next timeout should occur.
243 xpc_hb_beater(unsigned long dummy)
245 xpc_vars->heartbeat++;
247 if (jiffies >= xpc_hb_check_timeout) {
248 wake_up_interruptible(&xpc_act_IRQ_wq);
251 xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
252 add_timer(&xpc_hb_timer);
257 * This thread is responsible for nearly all of the partition
258 * activation/deactivation.
261 xpc_hb_checker(void *ignore)
263 int last_IRQ_count = 0;
268 /* this thread was marked active by xpc_hb_init() */
270 daemonize(XPC_HB_CHECK_THREAD_NAME);
272 set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU));
274 xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
276 while (!(volatile int) xpc_exiting) {
278 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
280 (int) (xpc_hb_check_timeout - jiffies),
281 atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count);
284 /* checking of remote heartbeats is skewed by IRQ handling */
285 if (jiffies >= xpc_hb_check_timeout) {
286 dev_dbg(xpc_part, "checking remote heartbeats\n");
287 xpc_check_remote_hb();
290 * We need to periodically recheck to ensure no
291 * IPI/AMO pairs have been missed. That check
292 * must always reset xpc_hb_check_timeout.
298 /* check for outstanding IRQs */
299 new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd);
300 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
303 dev_dbg(xpc_part, "found an IRQ to process; will be "
304 "resetting xpc_hb_check_timeout\n");
306 last_IRQ_count += xpc_identify_act_IRQ_sender();
307 if (last_IRQ_count < new_IRQ_count) {
308 /* retry once to help avoid missing AMO */
309 (void) xpc_identify_act_IRQ_sender();
311 last_IRQ_count = new_IRQ_count;
313 xpc_hb_check_timeout = jiffies +
314 (xpc_hb_check_interval * HZ);
317 /* wait for IRQ or timeout */
318 (void) wait_event_interruptible(xpc_act_IRQ_wq,
319 (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) ||
320 jiffies >= xpc_hb_check_timeout ||
321 (volatile int) xpc_exiting));
324 dev_dbg(xpc_part, "heartbeat checker is exiting\n");
327 /* mark this thread as having exited */
328 complete(&xpc_hb_checker_exited);
334 * This thread will attempt to discover other partitions to activate
335 * based on info provided by SAL. This new thread is short lived and
336 * will exit once discovery is complete.
339 xpc_initiate_discovery(void *ignore)
341 daemonize(XPC_DISCOVERY_THREAD_NAME);
345 dev_dbg(xpc_part, "discovery thread is exiting\n");
347 /* mark this thread as having exited */
348 complete(&xpc_discovery_exited);
354 * Establish first contact with the remote partititon. This involves pulling
355 * the XPC per partition variables from the remote partition and waiting for
356 * the remote partition to pull ours.
358 static enum xpc_retval
359 xpc_make_first_contact(struct xpc_partition *part)
364 while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) {
365 if (ret != xpcRetry) {
366 XPC_DEACTIVATE_PARTITION(part, ret);
370 dev_dbg(xpc_chan, "waiting to make first contact with "
371 "partition %d\n", XPC_PARTID(part));
373 /* wait a 1/4 of a second or so */
374 (void) msleep_interruptible(250);
376 if (part->act_state == XPC_P_DEACTIVATING) {
381 return xpc_mark_partition_active(part);
386 * The first kthread assigned to a newly activated partition is the one
387 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
388 * that kthread until the partition is brought down, at which time that kthread
389 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
390 * that XPC has dismantled all communication infrastructure for the associated
391 * partition.) This kthread becomes the channel manager for that partition.
393 * Each active partition has a channel manager, who, besides connecting and
394 * disconnecting channels, will ensure that each of the partition's connected
395 * channels has the required number of assigned kthreads to get the work done.
398 xpc_channel_mgr(struct xpc_partition *part)
400 while (part->act_state != XPC_P_DEACTIVATING ||
401 atomic_read(&part->nchannels_active) > 0 ||
402 !xpc_partition_disengaged(part)) {
404 xpc_process_channel_activity(part);
408 * Wait until we've been requested to activate kthreads or
409 * all of the channel's message queues have been torn down or
410 * a signal is pending.
412 * The channel_mgr_requests is set to 1 after being awakened,
413 * This is done to prevent the channel mgr from making one pass
414 * through the loop for each request, since he will
415 * be servicing all the requests in one pass. The reason it's
416 * set to 1 instead of 0 is so that other kthreads will know
417 * that the channel mgr is running and won't bother trying to
420 atomic_dec(&part->channel_mgr_requests);
421 (void) wait_event_interruptible(part->channel_mgr_wq,
422 (atomic_read(&part->channel_mgr_requests) > 0 ||
423 (volatile u64) part->local_IPI_amo != 0 ||
424 ((volatile u8) part->act_state ==
425 XPC_P_DEACTIVATING &&
426 atomic_read(&part->nchannels_active) == 0 &&
427 xpc_partition_disengaged(part))));
428 atomic_set(&part->channel_mgr_requests, 1);
430 // >>> Does it need to wakeup periodically as well? In case we
431 // >>> miscalculated the #of kthreads to wakeup or create?
437 * When XPC HB determines that a partition has come up, it will create a new
438 * kthread and that kthread will call this function to attempt to set up the
439 * basic infrastructure used for Cross Partition Communication with the newly
442 * The kthread that was created by XPC HB and which setup the XPC
443 * infrastructure will remain assigned to the partition until the partition
444 * goes down. At which time the kthread will teardown the XPC infrastructure
447 * XPC HB will put the remote partition's XPC per partition specific variables
448 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
449 * calling xpc_partition_up().
452 xpc_partition_up(struct xpc_partition *part)
454 DBUG_ON(part->channels != NULL);
456 dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part));
458 if (xpc_setup_infrastructure(part) != xpcSuccess) {
463 * The kthread that XPC HB called us with will become the
464 * channel manager for this partition. It will not return
465 * back to XPC HB until the partition's XPC infrastructure
466 * has been dismantled.
469 (void) xpc_part_ref(part); /* this will always succeed */
471 if (xpc_make_first_contact(part) == xpcSuccess) {
472 xpc_channel_mgr(part);
475 xpc_part_deref(part);
477 xpc_teardown_infrastructure(part);
482 xpc_activating(void *__partid)
484 partid_t partid = (u64) __partid;
485 struct xpc_partition *part = &xpc_partitions[partid];
486 unsigned long irq_flags;
487 struct sched_param param = { sched_priority: MAX_RT_PRIO - 1 };
491 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
493 spin_lock_irqsave(&part->act_lock, irq_flags);
495 if (part->act_state == XPC_P_DEACTIVATING) {
496 part->act_state = XPC_P_INACTIVE;
497 spin_unlock_irqrestore(&part->act_lock, irq_flags);
498 part->remote_rp_pa = 0;
502 /* indicate the thread is activating */
503 DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
504 part->act_state = XPC_P_ACTIVATING;
506 XPC_SET_REASON(part, 0, 0);
507 spin_unlock_irqrestore(&part->act_lock, irq_flags);
509 dev_dbg(xpc_part, "bringing partition %d up\n", partid);
511 daemonize("xpc%02d", partid);
514 * This thread needs to run at a realtime priority to prevent a
515 * significant performance degradation.
517 ret = sched_setscheduler(current, SCHED_FIFO, ¶m);
519 dev_warn(xpc_part, "unable to set pid %d to a realtime "
520 "priority, ret=%d\n", current->pid, ret);
523 /* allow this thread and its children to run on any CPU */
524 set_cpus_allowed(current, CPU_MASK_ALL);
527 * Register the remote partition's AMOs with SAL so it can handle
528 * and cleanup errors within that address range should the remote
529 * partition go down. We don't unregister this range because it is
530 * difficult to tell when outstanding writes to the remote partition
531 * are finished and thus when it is safe to unregister. This should
532 * not result in wasted space in the SAL xp_addr_region table because
533 * we should get the same page for remote_amos_page_pa after module
534 * reloads and system reboots.
536 if (sn_register_xp_addr_region(part->remote_amos_page_pa,
538 dev_warn(xpc_part, "xpc_partition_up(%d) failed to register "
539 "xp_addr region\n", partid);
541 spin_lock_irqsave(&part->act_lock, irq_flags);
542 part->act_state = XPC_P_INACTIVE;
543 XPC_SET_REASON(part, xpcPhysAddrRegFailed, __LINE__);
544 spin_unlock_irqrestore(&part->act_lock, irq_flags);
545 part->remote_rp_pa = 0;
549 xpc_allow_hb(partid, xpc_vars);
550 xpc_IPI_send_activated(part);
554 * xpc_partition_up() holds this thread and marks this partition as
555 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
557 (void) xpc_partition_up(part);
559 xpc_disallow_hb(partid, xpc_vars);
560 xpc_mark_partition_inactive(part);
562 if (part->reason == xpcReactivating) {
563 /* interrupting ourselves results in activating partition */
564 xpc_IPI_send_reactivate(part);
572 xpc_activate_partition(struct xpc_partition *part)
574 partid_t partid = XPC_PARTID(part);
575 unsigned long irq_flags;
579 spin_lock_irqsave(&part->act_lock, irq_flags);
581 DBUG_ON(part->act_state != XPC_P_INACTIVE);
583 part->act_state = XPC_P_ACTIVATION_REQ;
584 XPC_SET_REASON(part, xpcCloneKThread, __LINE__);
586 spin_unlock_irqrestore(&part->act_lock, irq_flags);
588 pid = kernel_thread(xpc_activating, (void *) ((u64) partid), 0);
590 if (unlikely(pid <= 0)) {
591 spin_lock_irqsave(&part->act_lock, irq_flags);
592 part->act_state = XPC_P_INACTIVE;
593 XPC_SET_REASON(part, xpcCloneKThreadFailed, __LINE__);
594 spin_unlock_irqrestore(&part->act_lock, irq_flags);
600 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
601 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
602 * than one partition, we use an AMO_t structure per partition to indicate
603 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
604 * associated kthread to handle it.
606 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
607 * running on other partitions.
609 * Noteworthy Arguments:
611 * irq - Interrupt ReQuest number. NOT USED.
613 * dev_id - partid of IPI's potential sender.
615 * regs - processor's context before the processor entered
616 * interrupt code. NOT USED.
619 xpc_notify_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
621 partid_t partid = (partid_t) (u64) dev_id;
622 struct xpc_partition *part = &xpc_partitions[partid];
625 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
627 if (xpc_part_ref(part)) {
628 xpc_check_for_channel_activity(part);
630 xpc_part_deref(part);
637 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
638 * because the write to their associated IPI amo completed after the IRQ/IPI
642 xpc_dropped_IPI_check(struct xpc_partition *part)
644 if (xpc_part_ref(part)) {
645 xpc_check_for_channel_activity(part);
647 part->dropped_IPI_timer.expires = jiffies +
648 XPC_P_DROPPED_IPI_WAIT;
649 add_timer(&part->dropped_IPI_timer);
650 xpc_part_deref(part);
656 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
658 int idle = atomic_read(&ch->kthreads_idle);
659 int assigned = atomic_read(&ch->kthreads_assigned);
663 DBUG_ON(needed <= 0);
666 wakeup = (needed > idle) ? idle : needed;
669 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
670 "channel=%d\n", wakeup, ch->partid, ch->number);
672 /* only wakeup the requested number of kthreads */
673 wake_up_nr(&ch->idle_wq, wakeup);
680 if (needed + assigned > ch->kthreads_assigned_limit) {
681 needed = ch->kthreads_assigned_limit - assigned;
682 // >>>should never be less than 0
688 dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
689 needed, ch->partid, ch->number);
691 xpc_create_kthreads(ch, needed);
696 * This function is where XPC's kthreads wait for messages to deliver.
699 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
702 /* deliver messages to their intended recipients */
704 while ((volatile s64) ch->w_local_GP.get <
705 (volatile s64) ch->w_remote_GP.put &&
706 !((volatile u32) ch->flags &
707 XPC_C_DISCONNECTING)) {
711 if (atomic_inc_return(&ch->kthreads_idle) >
712 ch->kthreads_idle_limit) {
713 /* too many idle kthreads on this channel */
714 atomic_dec(&ch->kthreads_idle);
718 dev_dbg(xpc_chan, "idle kthread calling "
719 "wait_event_interruptible_exclusive()\n");
721 (void) wait_event_interruptible_exclusive(ch->idle_wq,
722 ((volatile s64) ch->w_local_GP.get <
723 (volatile s64) ch->w_remote_GP.put ||
724 ((volatile u32) ch->flags &
725 XPC_C_DISCONNECTING)));
727 atomic_dec(&ch->kthreads_idle);
729 } while (!((volatile u32) ch->flags & XPC_C_DISCONNECTING));
734 xpc_daemonize_kthread(void *args)
736 partid_t partid = XPC_UNPACK_ARG1(args);
737 u16 ch_number = XPC_UNPACK_ARG2(args);
738 struct xpc_partition *part = &xpc_partitions[partid];
739 struct xpc_channel *ch;
741 unsigned long irq_flags;
744 daemonize("xpc%02dc%d", partid, ch_number);
746 dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
749 ch = &part->channels[ch_number];
751 if (!(ch->flags & XPC_C_DISCONNECTING)) {
753 /* let registerer know that connection has been established */
755 spin_lock_irqsave(&ch->lock, irq_flags);
756 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
757 ch->flags |= XPC_C_CONNECTEDCALLOUT;
758 spin_unlock_irqrestore(&ch->lock, irq_flags);
760 xpc_connected_callout(ch);
762 spin_lock_irqsave(&ch->lock, irq_flags);
763 ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
764 spin_unlock_irqrestore(&ch->lock, irq_flags);
767 * It is possible that while the callout was being
768 * made that the remote partition sent some messages.
769 * If that is the case, we may need to activate
770 * additional kthreads to help deliver them. We only
771 * need one less than total #of messages to deliver.
773 n_needed = ch->w_remote_GP.put - ch->w_local_GP.get - 1;
775 !(ch->flags & XPC_C_DISCONNECTING)) {
776 xpc_activate_kthreads(ch, n_needed);
779 spin_unlock_irqrestore(&ch->lock, irq_flags);
782 xpc_kthread_waitmsgs(part, ch);
785 if (atomic_dec_return(&ch->kthreads_assigned) == 0) {
786 spin_lock_irqsave(&ch->lock, irq_flags);
787 if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
788 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
789 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
790 spin_unlock_irqrestore(&ch->lock, irq_flags);
792 xpc_disconnect_callout(ch, xpcDisconnecting);
794 spin_lock_irqsave(&ch->lock, irq_flags);
795 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
797 spin_unlock_irqrestore(&ch->lock, irq_flags);
798 if (atomic_dec_return(&part->nchannels_engaged) == 0) {
799 xpc_mark_partition_disengaged(part);
800 xpc_IPI_send_disengage(part);
805 xpc_msgqueue_deref(ch);
807 dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
810 xpc_part_deref(part);
816 * For each partition that XPC has established communications with, there is
817 * a minimum of one kernel thread assigned to perform any operation that
818 * may potentially sleep or block (basically the callouts to the asynchronous
819 * functions registered via xpc_connect()).
821 * Additional kthreads are created and destroyed by XPC as the workload
824 * A kthread is assigned to one of the active channels that exists for a given
828 xpc_create_kthreads(struct xpc_channel *ch, int needed)
830 unsigned long irq_flags;
832 u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
833 struct xpc_partition *part = &xpc_partitions[ch->partid];
836 while (needed-- > 0) {
839 * The following is done on behalf of the newly created
840 * kthread. That kthread is responsible for doing the
841 * counterpart to the following before it exits.
843 (void) xpc_part_ref(part);
844 xpc_msgqueue_ref(ch);
845 if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
846 atomic_inc_return(&part->nchannels_engaged) == 1) {
847 xpc_mark_partition_engaged(part);
850 pid = kernel_thread(xpc_daemonize_kthread, (void *) args, 0);
852 /* the fork failed */
853 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
854 atomic_dec_return(&part->nchannels_engaged) == 0) {
855 xpc_mark_partition_disengaged(part);
856 xpc_IPI_send_disengage(part);
858 xpc_msgqueue_deref(ch);
859 xpc_part_deref(part);
861 if (atomic_read(&ch->kthreads_assigned) <
862 ch->kthreads_idle_limit) {
864 * Flag this as an error only if we have an
865 * insufficient #of kthreads for the channel
868 * No xpc_msgqueue_ref() is needed here since
869 * the channel mgr is doing this.
871 spin_lock_irqsave(&ch->lock, irq_flags);
872 XPC_DISCONNECT_CHANNEL(ch, xpcLackOfResources,
874 spin_unlock_irqrestore(&ch->lock, irq_flags);
879 ch->kthreads_created++; // >>> temporary debug only!!!
885 xpc_disconnect_wait(int ch_number)
887 unsigned long irq_flags;
889 struct xpc_partition *part;
890 struct xpc_channel *ch;
891 int wakeup_channel_mgr;
894 /* now wait for all callouts to the caller's function to cease */
895 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
896 part = &xpc_partitions[partid];
898 if (!xpc_part_ref(part)) {
902 ch = &part->channels[ch_number];
904 if (!(ch->flags & XPC_C_WDISCONNECT)) {
905 xpc_part_deref(part);
909 wait_for_completion(&ch->wdisconnect_wait);
911 spin_lock_irqsave(&ch->lock, irq_flags);
912 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
913 wakeup_channel_mgr = 0;
915 if (ch->delayed_IPI_flags) {
916 if (part->act_state != XPC_P_DEACTIVATING) {
917 spin_lock(&part->IPI_lock);
918 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
919 ch->number, ch->delayed_IPI_flags);
920 spin_unlock(&part->IPI_lock);
921 wakeup_channel_mgr = 1;
923 ch->delayed_IPI_flags = 0;
926 ch->flags &= ~XPC_C_WDISCONNECT;
927 spin_unlock_irqrestore(&ch->lock, irq_flags);
929 if (wakeup_channel_mgr) {
930 xpc_wakeup_channel_mgr(part);
933 xpc_part_deref(part);
939 xpc_do_exit(enum xpc_retval reason)
942 int active_part_count, printed_waiting_msg = 0;
943 struct xpc_partition *part;
944 unsigned long printmsg_time, disengage_request_timeout = 0;
947 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
948 DBUG_ON(xpc_exiting == 1);
951 * Let the heartbeat checker thread and the discovery thread
952 * (if one is running) know that they should exit. Also wake up
953 * the heartbeat checker thread in case it's sleeping.
956 wake_up_interruptible(&xpc_act_IRQ_wq);
958 /* ignore all incoming interrupts */
959 free_irq(SGI_XPC_ACTIVATE, NULL);
961 /* wait for the discovery thread to exit */
962 wait_for_completion(&xpc_discovery_exited);
964 /* wait for the heartbeat checker thread to exit */
965 wait_for_completion(&xpc_hb_checker_exited);
968 /* sleep for a 1/3 of a second or so */
969 (void) msleep_interruptible(300);
972 /* wait for all partitions to become inactive */
974 printmsg_time = jiffies + (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
975 xpc_disengage_request_timedout = 0;
978 active_part_count = 0;
980 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
981 part = &xpc_partitions[partid];
983 if (xpc_partition_disengaged(part) &&
984 part->act_state == XPC_P_INACTIVE) {
990 XPC_DEACTIVATE_PARTITION(part, reason);
992 if (part->disengage_request_timeout >
993 disengage_request_timeout) {
994 disengage_request_timeout =
995 part->disengage_request_timeout;
999 if (xpc_partition_engaged(-1UL)) {
1000 if (time_after(jiffies, printmsg_time)) {
1001 dev_info(xpc_part, "waiting for remote "
1002 "partitions to disengage, timeout in "
1004 (disengage_request_timeout - jiffies)
1006 printmsg_time = jiffies +
1007 (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
1008 printed_waiting_msg = 1;
1011 } else if (active_part_count > 0) {
1012 if (printed_waiting_msg) {
1013 dev_info(xpc_part, "waiting for local partition"
1015 printed_waiting_msg = 0;
1019 if (!xpc_disengage_request_timedout) {
1020 dev_info(xpc_part, "all partitions have "
1026 /* sleep for a 1/3 of a second or so */
1027 (void) msleep_interruptible(300);
1031 DBUG_ON(xpc_partition_engaged(-1UL));
1034 /* indicate to others that our reserved page is uninitialized */
1035 xpc_rsvd_page->vars_pa = 0;
1037 /* now it's time to eliminate our heartbeat */
1038 del_timer_sync(&xpc_hb_timer);
1039 DBUG_ON(xpc_vars->heartbeating_to_mask != 0);
1041 if (reason == xpcUnloading) {
1042 /* take ourselves off of the reboot_notifier_list */
1043 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1045 /* take ourselves off of the die_notifier list */
1046 (void) unregister_die_notifier(&xpc_die_notifier);
1049 /* close down protections for IPI operations */
1050 xpc_restrict_IPI_ops();
1053 /* clear the interface to XPC's functions */
1054 xpc_clear_interface();
1057 unregister_sysctl_table(xpc_sysctl);
1060 kfree(xpc_remote_copy_buffer_base);
1065 * This function is called when the system is being rebooted.
1068 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1070 enum xpc_retval reason;
1075 reason = xpcSystemReboot;
1078 reason = xpcSystemHalt;
1081 reason = xpcSystemPoweroff;
1084 reason = xpcSystemGoingDown;
1087 xpc_do_exit(reason);
1093 * Notify other partitions to disengage from all references to our memory.
1096 xpc_die_disengage(void)
1098 struct xpc_partition *part;
1100 unsigned long engaged;
1101 long time, printmsg_time, disengage_request_timeout;
1104 /* keep xpc_hb_checker thread from doing anything (just in case) */
1107 xpc_vars->heartbeating_to_mask = 0; /* indicate we're deactivated */
1109 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1110 part = &xpc_partitions[partid];
1112 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part->
1113 remote_vars_version)) {
1115 /* just in case it was left set by an earlier XPC */
1116 xpc_clear_partition_engaged(1UL << partid);
1120 if (xpc_partition_engaged(1UL << partid) ||
1121 part->act_state != XPC_P_INACTIVE) {
1122 xpc_request_partition_disengage(part);
1123 xpc_mark_partition_disengaged(part);
1124 xpc_IPI_send_disengage(part);
1129 printmsg_time = time +
1130 (XPC_DISENGAGE_PRINTMSG_INTERVAL * sn_rtc_cycles_per_second);
1131 disengage_request_timeout = time +
1132 (xpc_disengage_request_timelimit * sn_rtc_cycles_per_second);
1134 /* wait for all other partitions to disengage from us */
1137 engaged = xpc_partition_engaged(-1UL);
1139 dev_info(xpc_part, "all partitions have disengaged\n");
1144 if (time >= disengage_request_timeout) {
1145 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1146 if (engaged & (1UL << partid)) {
1147 dev_info(xpc_part, "disengage from "
1148 "remote partition %d timed "
1155 if (time >= printmsg_time) {
1156 dev_info(xpc_part, "waiting for remote partitions to "
1157 "disengage, timeout in %ld seconds\n",
1158 (disengage_request_timeout - time) /
1159 sn_rtc_cycles_per_second);
1160 printmsg_time = time +
1161 (XPC_DISENGAGE_PRINTMSG_INTERVAL *
1162 sn_rtc_cycles_per_second);
1169 * This function is called when the system is being restarted or halted due
1170 * to some sort of system failure. If this is the case we need to notify the
1171 * other partitions to disengage from all references to our memory.
1172 * This function can also be called when our heartbeater could be offlined
1173 * for a time. In this case we need to notify other partitions to not worry
1174 * about the lack of a heartbeat.
1177 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1180 case DIE_MACHINE_RESTART:
1181 case DIE_MACHINE_HALT:
1182 xpc_die_disengage();
1185 case DIE_KDEBUG_ENTER:
1186 /* Should lack of heartbeat be ignored by other partitions? */
1187 if (!xpc_kdebug_ignore) {
1191 case DIE_MCA_MONARCH_ENTER:
1192 case DIE_INIT_MONARCH_ENTER:
1193 xpc_vars->heartbeat++;
1194 xpc_vars->heartbeat_offline = 1;
1197 case DIE_KDEBUG_LEAVE:
1198 /* Is lack of heartbeat being ignored by other partitions? */
1199 if (!xpc_kdebug_ignore) {
1203 case DIE_MCA_MONARCH_LEAVE:
1204 case DIE_INIT_MONARCH_LEAVE:
1205 xpc_vars->heartbeat++;
1206 xpc_vars->heartbeat_offline = 0;
1219 struct xpc_partition *part;
1224 if (!ia64_platform_is("sn2")) {
1229 buf_size = max(XPC_RP_VARS_SIZE,
1230 XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES);
1231 xpc_remote_copy_buffer = xpc_kmalloc_cacheline_aligned(buf_size,
1232 GFP_KERNEL, &xpc_remote_copy_buffer_base);
1233 if (xpc_remote_copy_buffer == NULL)
1236 snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1237 snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1239 xpc_sysctl = register_sysctl_table(xpc_sys_dir, 1);
1242 * The first few fields of each entry of xpc_partitions[] need to
1243 * be initialized now so that calls to xpc_connect() and
1244 * xpc_disconnect() can be made prior to the activation of any remote
1245 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1246 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1247 * PARTITION HAS BEEN ACTIVATED.
1249 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1250 part = &xpc_partitions[partid];
1252 DBUG_ON((u64) part != L1_CACHE_ALIGN((u64) part));
1254 part->act_IRQ_rcvd = 0;
1255 spin_lock_init(&part->act_lock);
1256 part->act_state = XPC_P_INACTIVE;
1257 XPC_SET_REASON(part, 0, 0);
1259 init_timer(&part->disengage_request_timer);
1260 part->disengage_request_timer.function =
1261 xpc_timeout_partition_disengage_request;
1262 part->disengage_request_timer.data = (unsigned long) part;
1264 part->setup_state = XPC_P_UNSET;
1265 init_waitqueue_head(&part->teardown_wq);
1266 atomic_set(&part->references, 0);
1270 * Open up protections for IPI operations (and AMO operations on
1271 * Shub 1.1 systems).
1273 xpc_allow_IPI_ops();
1276 * Interrupts being processed will increment this atomic variable and
1277 * awaken the heartbeat thread which will process the interrupts.
1279 atomic_set(&xpc_act_IRQ_rcvd, 0);
1282 * This is safe to do before the xpc_hb_checker thread has started
1283 * because the handler releases a wait queue. If an interrupt is
1284 * received before the thread is waiting, it will not go to sleep,
1285 * but rather immediately process the interrupt.
1287 ret = request_irq(SGI_XPC_ACTIVATE, xpc_act_IRQ_handler, 0,
1290 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
1291 "errno=%d\n", -ret);
1293 xpc_restrict_IPI_ops();
1296 unregister_sysctl_table(xpc_sysctl);
1299 kfree(xpc_remote_copy_buffer_base);
1304 * Fill the partition reserved page with the information needed by
1305 * other partitions to discover we are alive and establish initial
1308 xpc_rsvd_page = xpc_rsvd_page_init();
1309 if (xpc_rsvd_page == NULL) {
1310 dev_err(xpc_part, "could not setup our reserved page\n");
1312 free_irq(SGI_XPC_ACTIVATE, NULL);
1313 xpc_restrict_IPI_ops();
1316 unregister_sysctl_table(xpc_sysctl);
1319 kfree(xpc_remote_copy_buffer_base);
1324 /* add ourselves to the reboot_notifier_list */
1325 ret = register_reboot_notifier(&xpc_reboot_notifier);
1327 dev_warn(xpc_part, "can't register reboot notifier\n");
1330 /* add ourselves to the die_notifier list (i.e., ia64die_chain) */
1331 ret = register_die_notifier(&xpc_die_notifier);
1333 dev_warn(xpc_part, "can't register die notifier\n");
1338 * Set the beating to other partitions into motion. This is
1339 * the last requirement for other partitions' discovery to
1340 * initiate communications with us.
1342 init_timer(&xpc_hb_timer);
1343 xpc_hb_timer.function = xpc_hb_beater;
1348 * The real work-horse behind xpc. This processes incoming
1349 * interrupts and monitors remote heartbeats.
1351 pid = kernel_thread(xpc_hb_checker, NULL, 0);
1353 dev_err(xpc_part, "failed while forking hb check thread\n");
1355 /* indicate to others that our reserved page is uninitialized */
1356 xpc_rsvd_page->vars_pa = 0;
1358 /* take ourselves off of the reboot_notifier_list */
1359 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1361 /* take ourselves off of the die_notifier list */
1362 (void) unregister_die_notifier(&xpc_die_notifier);
1364 del_timer_sync(&xpc_hb_timer);
1365 free_irq(SGI_XPC_ACTIVATE, NULL);
1366 xpc_restrict_IPI_ops();
1369 unregister_sysctl_table(xpc_sysctl);
1372 kfree(xpc_remote_copy_buffer_base);
1378 * Startup a thread that will attempt to discover other partitions to
1379 * activate based on info provided by SAL. This new thread is short
1380 * lived and will exit once discovery is complete.
1382 pid = kernel_thread(xpc_initiate_discovery, NULL, 0);
1384 dev_err(xpc_part, "failed while forking discovery thread\n");
1386 /* mark this new thread as a non-starter */
1387 complete(&xpc_discovery_exited);
1389 xpc_do_exit(xpcUnloading);
1394 /* set the interface to point at XPC's functions */
1395 xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1396 xpc_initiate_allocate, xpc_initiate_send,
1397 xpc_initiate_send_notify, xpc_initiate_received,
1398 xpc_initiate_partid_to_nasids);
1402 module_init(xpc_init);
1408 xpc_do_exit(xpcUnloading);
1410 module_exit(xpc_exit);
1413 MODULE_AUTHOR("Silicon Graphics, Inc.");
1414 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1415 MODULE_LICENSE("GPL");
1417 module_param(xpc_hb_interval, int, 0);
1418 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1419 "heartbeat increments.");
1421 module_param(xpc_hb_check_interval, int, 0);
1422 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1423 "heartbeat checks.");
1425 module_param(xpc_disengage_request_timelimit, int, 0);
1426 MODULE_PARM_DESC(xpc_disengage_request_timelimit, "Number of seconds to wait "
1427 "for disengage request to complete.");
1429 module_param(xpc_kdebug_ignore, int, 0);
1430 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1431 "other partitions when dropping into kdebug.");