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/delay.h>
56 #include <linux/reboot.h>
57 #include <linux/completion.h>
58 #include <linux/kdebug.h>
59 #include <asm/sn/intr.h>
60 #include <asm/sn/sn_sal.h>
61 #include <asm/uaccess.h>
62 #include <asm/sn/xpc.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 static int xpc_kdebug_ignore;
86 static int xpc_kdebug_entered;
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[] = {
105 .ctl_name = CTL_UNNUMBERED,
106 .procname = "hb_interval",
107 .data = &xpc_hb_interval,
108 .maxlen = sizeof(int),
110 .proc_handler = &proc_dointvec_minmax,
111 .strategy = &sysctl_intvec,
112 .extra1 = &xpc_hb_min_interval,
113 .extra2 = &xpc_hb_max_interval
116 .ctl_name = CTL_UNNUMBERED,
117 .procname = "hb_check_interval",
118 .data = &xpc_hb_check_interval,
119 .maxlen = sizeof(int),
121 .proc_handler = &proc_dointvec_minmax,
122 .strategy = &sysctl_intvec,
123 .extra1 = &xpc_hb_check_min_interval,
124 .extra2 = &xpc_hb_check_max_interval
128 static ctl_table xpc_sys_xpc_dir[] = {
130 .ctl_name = CTL_UNNUMBERED,
133 .child = xpc_sys_xpc_hb_dir
136 .ctl_name = CTL_UNNUMBERED,
137 .procname = "disengage_request_timelimit",
138 .data = &xpc_disengage_request_timelimit,
139 .maxlen = sizeof(int),
141 .proc_handler = &proc_dointvec_minmax,
142 .strategy = &sysctl_intvec,
143 .extra1 = &xpc_disengage_request_min_timelimit,
144 .extra2 = &xpc_disengage_request_max_timelimit
148 static ctl_table xpc_sys_dir[] = {
150 .ctl_name = CTL_UNNUMBERED,
153 .child = xpc_sys_xpc_dir
157 static struct ctl_table_header *xpc_sysctl;
159 /* non-zero if any remote partition disengage request was timed out */
160 int xpc_disengage_request_timedout;
162 /* #of IRQs received */
163 static atomic_t xpc_act_IRQ_rcvd;
165 /* IRQ handler notifies this wait queue on receipt of an IRQ */
166 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq);
168 static unsigned long xpc_hb_check_timeout;
170 /* notification that the xpc_hb_checker thread has exited */
171 static DECLARE_COMPLETION(xpc_hb_checker_exited);
173 /* notification that the xpc_discovery thread has exited */
174 static DECLARE_COMPLETION(xpc_discovery_exited);
177 static struct timer_list xpc_hb_timer;
180 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
183 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
184 static struct notifier_block xpc_reboot_notifier = {
185 .notifier_call = xpc_system_reboot,
188 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
189 static struct notifier_block xpc_die_notifier = {
190 .notifier_call = xpc_system_die,
195 * Timer function to enforce the timelimit on the partition disengage request.
198 xpc_timeout_partition_disengage_request(unsigned long data)
200 struct xpc_partition *part = (struct xpc_partition *) data;
203 DBUG_ON(jiffies < part->disengage_request_timeout);
205 (void) xpc_partition_disengaged(part);
207 DBUG_ON(part->disengage_request_timeout != 0);
208 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part)) != 0);
213 * Notify the heartbeat check thread that an IRQ has been received.
216 xpc_act_IRQ_handler(int irq, void *dev_id)
218 atomic_inc(&xpc_act_IRQ_rcvd);
219 wake_up_interruptible(&xpc_act_IRQ_wq);
225 * Timer to produce the heartbeat. The timer structures function is
226 * already set when this is initially called. A tunable is used to
227 * specify when the next timeout should occur.
230 xpc_hb_beater(unsigned long dummy)
232 xpc_vars->heartbeat++;
234 if (jiffies >= xpc_hb_check_timeout) {
235 wake_up_interruptible(&xpc_act_IRQ_wq);
238 xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
239 add_timer(&xpc_hb_timer);
244 * This thread is responsible for nearly all of the partition
245 * activation/deactivation.
248 xpc_hb_checker(void *ignore)
250 int last_IRQ_count = 0;
255 /* this thread was marked active by xpc_hb_init() */
257 daemonize(XPC_HB_CHECK_THREAD_NAME);
259 set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU));
261 xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
263 while (!(volatile int) xpc_exiting) {
265 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
267 (int) (xpc_hb_check_timeout - jiffies),
268 atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count);
271 /* checking of remote heartbeats is skewed by IRQ handling */
272 if (jiffies >= xpc_hb_check_timeout) {
273 dev_dbg(xpc_part, "checking remote heartbeats\n");
274 xpc_check_remote_hb();
277 * We need to periodically recheck to ensure no
278 * IPI/AMO pairs have been missed. That check
279 * must always reset xpc_hb_check_timeout.
285 /* check for outstanding IRQs */
286 new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd);
287 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
290 dev_dbg(xpc_part, "found an IRQ to process; will be "
291 "resetting xpc_hb_check_timeout\n");
293 last_IRQ_count += xpc_identify_act_IRQ_sender();
294 if (last_IRQ_count < new_IRQ_count) {
295 /* retry once to help avoid missing AMO */
296 (void) xpc_identify_act_IRQ_sender();
298 last_IRQ_count = new_IRQ_count;
300 xpc_hb_check_timeout = jiffies +
301 (xpc_hb_check_interval * HZ);
304 /* wait for IRQ or timeout */
305 (void) wait_event_interruptible(xpc_act_IRQ_wq,
306 (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) ||
307 jiffies >= xpc_hb_check_timeout ||
308 (volatile int) xpc_exiting));
311 dev_dbg(xpc_part, "heartbeat checker is exiting\n");
314 /* mark this thread as having exited */
315 complete(&xpc_hb_checker_exited);
321 * This thread will attempt to discover other partitions to activate
322 * based on info provided by SAL. This new thread is short lived and
323 * will exit once discovery is complete.
326 xpc_initiate_discovery(void *ignore)
328 daemonize(XPC_DISCOVERY_THREAD_NAME);
332 dev_dbg(xpc_part, "discovery thread is exiting\n");
334 /* mark this thread as having exited */
335 complete(&xpc_discovery_exited);
341 * Establish first contact with the remote partititon. This involves pulling
342 * the XPC per partition variables from the remote partition and waiting for
343 * the remote partition to pull ours.
345 static enum xpc_retval
346 xpc_make_first_contact(struct xpc_partition *part)
351 while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) {
352 if (ret != xpcRetry) {
353 XPC_DEACTIVATE_PARTITION(part, ret);
357 dev_dbg(xpc_chan, "waiting to make first contact with "
358 "partition %d\n", XPC_PARTID(part));
360 /* wait a 1/4 of a second or so */
361 (void) msleep_interruptible(250);
363 if (part->act_state == XPC_P_DEACTIVATING) {
368 return xpc_mark_partition_active(part);
373 * The first kthread assigned to a newly activated partition is the one
374 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
375 * that kthread until the partition is brought down, at which time that kthread
376 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
377 * that XPC has dismantled all communication infrastructure for the associated
378 * partition.) This kthread becomes the channel manager for that partition.
380 * Each active partition has a channel manager, who, besides connecting and
381 * disconnecting channels, will ensure that each of the partition's connected
382 * channels has the required number of assigned kthreads to get the work done.
385 xpc_channel_mgr(struct xpc_partition *part)
387 while (part->act_state != XPC_P_DEACTIVATING ||
388 atomic_read(&part->nchannels_active) > 0 ||
389 !xpc_partition_disengaged(part)) {
391 xpc_process_channel_activity(part);
395 * Wait until we've been requested to activate kthreads or
396 * all of the channel's message queues have been torn down or
397 * a signal is pending.
399 * The channel_mgr_requests is set to 1 after being awakened,
400 * This is done to prevent the channel mgr from making one pass
401 * through the loop for each request, since he will
402 * be servicing all the requests in one pass. The reason it's
403 * set to 1 instead of 0 is so that other kthreads will know
404 * that the channel mgr is running and won't bother trying to
407 atomic_dec(&part->channel_mgr_requests);
408 (void) wait_event_interruptible(part->channel_mgr_wq,
409 (atomic_read(&part->channel_mgr_requests) > 0 ||
410 (volatile u64) part->local_IPI_amo != 0 ||
411 ((volatile u8) part->act_state ==
412 XPC_P_DEACTIVATING &&
413 atomic_read(&part->nchannels_active) == 0 &&
414 xpc_partition_disengaged(part))));
415 atomic_set(&part->channel_mgr_requests, 1);
417 // >>> Does it need to wakeup periodically as well? In case we
418 // >>> miscalculated the #of kthreads to wakeup or create?
424 * When XPC HB determines that a partition has come up, it will create a new
425 * kthread and that kthread will call this function to attempt to set up the
426 * basic infrastructure used for Cross Partition Communication with the newly
429 * The kthread that was created by XPC HB and which setup the XPC
430 * infrastructure will remain assigned to the partition until the partition
431 * goes down. At which time the kthread will teardown the XPC infrastructure
434 * XPC HB will put the remote partition's XPC per partition specific variables
435 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
436 * calling xpc_partition_up().
439 xpc_partition_up(struct xpc_partition *part)
441 DBUG_ON(part->channels != NULL);
443 dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part));
445 if (xpc_setup_infrastructure(part) != xpcSuccess) {
450 * The kthread that XPC HB called us with will become the
451 * channel manager for this partition. It will not return
452 * back to XPC HB until the partition's XPC infrastructure
453 * has been dismantled.
456 (void) xpc_part_ref(part); /* this will always succeed */
458 if (xpc_make_first_contact(part) == xpcSuccess) {
459 xpc_channel_mgr(part);
462 xpc_part_deref(part);
464 xpc_teardown_infrastructure(part);
469 xpc_activating(void *__partid)
471 partid_t partid = (u64) __partid;
472 struct xpc_partition *part = &xpc_partitions[partid];
473 unsigned long irq_flags;
474 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
478 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
480 spin_lock_irqsave(&part->act_lock, irq_flags);
482 if (part->act_state == XPC_P_DEACTIVATING) {
483 part->act_state = XPC_P_INACTIVE;
484 spin_unlock_irqrestore(&part->act_lock, irq_flags);
485 part->remote_rp_pa = 0;
489 /* indicate the thread is activating */
490 DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
491 part->act_state = XPC_P_ACTIVATING;
493 XPC_SET_REASON(part, 0, 0);
494 spin_unlock_irqrestore(&part->act_lock, irq_flags);
496 dev_dbg(xpc_part, "bringing partition %d up\n", partid);
498 daemonize("xpc%02d", partid);
501 * This thread needs to run at a realtime priority to prevent a
502 * significant performance degradation.
504 ret = sched_setscheduler(current, SCHED_FIFO, ¶m);
506 dev_warn(xpc_part, "unable to set pid %d to a realtime "
507 "priority, ret=%d\n", current->pid, ret);
510 /* allow this thread and its children to run on any CPU */
511 set_cpus_allowed(current, CPU_MASK_ALL);
514 * Register the remote partition's AMOs with SAL so it can handle
515 * and cleanup errors within that address range should the remote
516 * partition go down. We don't unregister this range because it is
517 * difficult to tell when outstanding writes to the remote partition
518 * are finished and thus when it is safe to unregister. This should
519 * not result in wasted space in the SAL xp_addr_region table because
520 * we should get the same page for remote_amos_page_pa after module
521 * reloads and system reboots.
523 if (sn_register_xp_addr_region(part->remote_amos_page_pa,
525 dev_warn(xpc_part, "xpc_partition_up(%d) failed to register "
526 "xp_addr region\n", partid);
528 spin_lock_irqsave(&part->act_lock, irq_flags);
529 part->act_state = XPC_P_INACTIVE;
530 XPC_SET_REASON(part, xpcPhysAddrRegFailed, __LINE__);
531 spin_unlock_irqrestore(&part->act_lock, irq_flags);
532 part->remote_rp_pa = 0;
536 xpc_allow_hb(partid, xpc_vars);
537 xpc_IPI_send_activated(part);
541 * xpc_partition_up() holds this thread and marks this partition as
542 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
544 (void) xpc_partition_up(part);
546 xpc_disallow_hb(partid, xpc_vars);
547 xpc_mark_partition_inactive(part);
549 if (part->reason == xpcReactivating) {
550 /* interrupting ourselves results in activating partition */
551 xpc_IPI_send_reactivate(part);
559 xpc_activate_partition(struct xpc_partition *part)
561 partid_t partid = XPC_PARTID(part);
562 unsigned long irq_flags;
566 spin_lock_irqsave(&part->act_lock, irq_flags);
568 DBUG_ON(part->act_state != XPC_P_INACTIVE);
570 part->act_state = XPC_P_ACTIVATION_REQ;
571 XPC_SET_REASON(part, xpcCloneKThread, __LINE__);
573 spin_unlock_irqrestore(&part->act_lock, irq_flags);
575 pid = kernel_thread(xpc_activating, (void *) ((u64) partid), 0);
577 if (unlikely(pid <= 0)) {
578 spin_lock_irqsave(&part->act_lock, irq_flags);
579 part->act_state = XPC_P_INACTIVE;
580 XPC_SET_REASON(part, xpcCloneKThreadFailed, __LINE__);
581 spin_unlock_irqrestore(&part->act_lock, irq_flags);
587 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
588 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
589 * than one partition, we use an AMO_t structure per partition to indicate
590 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
591 * associated kthread to handle it.
593 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
594 * running on other partitions.
596 * Noteworthy Arguments:
598 * irq - Interrupt ReQuest number. NOT USED.
600 * dev_id - partid of IPI's potential sender.
603 xpc_notify_IRQ_handler(int irq, void *dev_id)
605 partid_t partid = (partid_t) (u64) dev_id;
606 struct xpc_partition *part = &xpc_partitions[partid];
609 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
611 if (xpc_part_ref(part)) {
612 xpc_check_for_channel_activity(part);
614 xpc_part_deref(part);
621 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
622 * because the write to their associated IPI amo completed after the IRQ/IPI
626 xpc_dropped_IPI_check(struct xpc_partition *part)
628 if (xpc_part_ref(part)) {
629 xpc_check_for_channel_activity(part);
631 part->dropped_IPI_timer.expires = jiffies +
632 XPC_P_DROPPED_IPI_WAIT;
633 add_timer(&part->dropped_IPI_timer);
634 xpc_part_deref(part);
640 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
642 int idle = atomic_read(&ch->kthreads_idle);
643 int assigned = atomic_read(&ch->kthreads_assigned);
647 DBUG_ON(needed <= 0);
650 wakeup = (needed > idle) ? idle : needed;
653 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
654 "channel=%d\n", wakeup, ch->partid, ch->number);
656 /* only wakeup the requested number of kthreads */
657 wake_up_nr(&ch->idle_wq, wakeup);
664 if (needed + assigned > ch->kthreads_assigned_limit) {
665 needed = ch->kthreads_assigned_limit - assigned;
666 // >>>should never be less than 0
672 dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
673 needed, ch->partid, ch->number);
675 xpc_create_kthreads(ch, needed, 0);
680 * This function is where XPC's kthreads wait for messages to deliver.
683 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
686 /* deliver messages to their intended recipients */
688 while ((volatile s64) ch->w_local_GP.get <
689 (volatile s64) ch->w_remote_GP.put &&
690 !((volatile u32) ch->flags &
691 XPC_C_DISCONNECTING)) {
695 if (atomic_inc_return(&ch->kthreads_idle) >
696 ch->kthreads_idle_limit) {
697 /* too many idle kthreads on this channel */
698 atomic_dec(&ch->kthreads_idle);
702 dev_dbg(xpc_chan, "idle kthread calling "
703 "wait_event_interruptible_exclusive()\n");
705 (void) wait_event_interruptible_exclusive(ch->idle_wq,
706 ((volatile s64) ch->w_local_GP.get <
707 (volatile s64) ch->w_remote_GP.put ||
708 ((volatile u32) ch->flags &
709 XPC_C_DISCONNECTING)));
711 atomic_dec(&ch->kthreads_idle);
713 } while (!((volatile u32) ch->flags & XPC_C_DISCONNECTING));
718 xpc_daemonize_kthread(void *args)
720 partid_t partid = XPC_UNPACK_ARG1(args);
721 u16 ch_number = XPC_UNPACK_ARG2(args);
722 struct xpc_partition *part = &xpc_partitions[partid];
723 struct xpc_channel *ch;
725 unsigned long irq_flags;
728 daemonize("xpc%02dc%d", partid, ch_number);
730 dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
733 ch = &part->channels[ch_number];
735 if (!(ch->flags & XPC_C_DISCONNECTING)) {
737 /* let registerer know that connection has been established */
739 spin_lock_irqsave(&ch->lock, irq_flags);
740 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
741 ch->flags |= XPC_C_CONNECTEDCALLOUT;
742 spin_unlock_irqrestore(&ch->lock, irq_flags);
744 xpc_connected_callout(ch);
746 spin_lock_irqsave(&ch->lock, irq_flags);
747 ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
748 spin_unlock_irqrestore(&ch->lock, irq_flags);
751 * It is possible that while the callout was being
752 * made that the remote partition sent some messages.
753 * If that is the case, we may need to activate
754 * additional kthreads to help deliver them. We only
755 * need one less than total #of messages to deliver.
757 n_needed = ch->w_remote_GP.put - ch->w_local_GP.get - 1;
759 !(ch->flags & XPC_C_DISCONNECTING)) {
760 xpc_activate_kthreads(ch, n_needed);
763 spin_unlock_irqrestore(&ch->lock, irq_flags);
766 xpc_kthread_waitmsgs(part, ch);
769 /* let registerer know that connection is disconnecting */
771 spin_lock_irqsave(&ch->lock, irq_flags);
772 if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
773 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
774 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
775 spin_unlock_irqrestore(&ch->lock, irq_flags);
777 xpc_disconnect_callout(ch, xpcDisconnecting);
779 spin_lock_irqsave(&ch->lock, irq_flags);
780 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
782 spin_unlock_irqrestore(&ch->lock, irq_flags);
784 if (atomic_dec_return(&ch->kthreads_assigned) == 0) {
785 if (atomic_dec_return(&part->nchannels_engaged) == 0) {
786 xpc_mark_partition_disengaged(part);
787 xpc_IPI_send_disengage(part);
791 xpc_msgqueue_deref(ch);
793 dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
796 xpc_part_deref(part);
802 * For each partition that XPC has established communications with, there is
803 * a minimum of one kernel thread assigned to perform any operation that
804 * may potentially sleep or block (basically the callouts to the asynchronous
805 * functions registered via xpc_connect()).
807 * Additional kthreads are created and destroyed by XPC as the workload
810 * A kthread is assigned to one of the active channels that exists for a given
814 xpc_create_kthreads(struct xpc_channel *ch, int needed,
815 int ignore_disconnecting)
817 unsigned long irq_flags;
819 u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
820 struct xpc_partition *part = &xpc_partitions[ch->partid];
823 while (needed-- > 0) {
826 * The following is done on behalf of the newly created
827 * kthread. That kthread is responsible for doing the
828 * counterpart to the following before it exits.
830 if (ignore_disconnecting) {
831 if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
832 /* kthreads assigned had gone to zero */
834 XPC_C_DISCONNECTINGCALLOUT_MADE));
838 } else if (ch->flags & XPC_C_DISCONNECTING) {
841 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1) {
842 if (atomic_inc_return(&part->nchannels_engaged) == 1)
843 xpc_mark_partition_engaged(part);
845 (void) xpc_part_ref(part);
846 xpc_msgqueue_ref(ch);
848 pid = kernel_thread(xpc_daemonize_kthread, (void *) args, 0);
850 /* the fork failed */
853 * NOTE: if (ignore_disconnecting &&
854 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
855 * then we'll deadlock if all other kthreads assigned
856 * to this channel are blocked in the channel's
857 * registerer, because the only thing that will unblock
858 * them is the xpcDisconnecting callout that this
859 * failed kernel_thread would have made.
862 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
863 atomic_dec_return(&part->nchannels_engaged) == 0) {
864 xpc_mark_partition_disengaged(part);
865 xpc_IPI_send_disengage(part);
867 xpc_msgqueue_deref(ch);
868 xpc_part_deref(part);
870 if (atomic_read(&ch->kthreads_assigned) <
871 ch->kthreads_idle_limit) {
873 * Flag this as an error only if we have an
874 * insufficient #of kthreads for the channel
877 spin_lock_irqsave(&ch->lock, irq_flags);
878 XPC_DISCONNECT_CHANNEL(ch, xpcLackOfResources,
880 spin_unlock_irqrestore(&ch->lock, irq_flags);
885 ch->kthreads_created++; // >>> temporary debug only!!!
891 xpc_disconnect_wait(int ch_number)
893 unsigned long irq_flags;
895 struct xpc_partition *part;
896 struct xpc_channel *ch;
897 int wakeup_channel_mgr;
900 /* now wait for all callouts to the caller's function to cease */
901 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
902 part = &xpc_partitions[partid];
904 if (!xpc_part_ref(part)) {
908 ch = &part->channels[ch_number];
910 if (!(ch->flags & XPC_C_WDISCONNECT)) {
911 xpc_part_deref(part);
915 wait_for_completion(&ch->wdisconnect_wait);
917 spin_lock_irqsave(&ch->lock, irq_flags);
918 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
919 wakeup_channel_mgr = 0;
921 if (ch->delayed_IPI_flags) {
922 if (part->act_state != XPC_P_DEACTIVATING) {
923 spin_lock(&part->IPI_lock);
924 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
925 ch->number, ch->delayed_IPI_flags);
926 spin_unlock(&part->IPI_lock);
927 wakeup_channel_mgr = 1;
929 ch->delayed_IPI_flags = 0;
932 ch->flags &= ~XPC_C_WDISCONNECT;
933 spin_unlock_irqrestore(&ch->lock, irq_flags);
935 if (wakeup_channel_mgr) {
936 xpc_wakeup_channel_mgr(part);
939 xpc_part_deref(part);
945 xpc_do_exit(enum xpc_retval reason)
948 int active_part_count, printed_waiting_msg = 0;
949 struct xpc_partition *part;
950 unsigned long printmsg_time, disengage_request_timeout = 0;
953 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
954 DBUG_ON(xpc_exiting == 1);
957 * Let the heartbeat checker thread and the discovery thread
958 * (if one is running) know that they should exit. Also wake up
959 * the heartbeat checker thread in case it's sleeping.
962 wake_up_interruptible(&xpc_act_IRQ_wq);
964 /* ignore all incoming interrupts */
965 free_irq(SGI_XPC_ACTIVATE, NULL);
967 /* wait for the discovery thread to exit */
968 wait_for_completion(&xpc_discovery_exited);
970 /* wait for the heartbeat checker thread to exit */
971 wait_for_completion(&xpc_hb_checker_exited);
974 /* sleep for a 1/3 of a second or so */
975 (void) msleep_interruptible(300);
978 /* wait for all partitions to become inactive */
980 printmsg_time = jiffies + (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
981 xpc_disengage_request_timedout = 0;
984 active_part_count = 0;
986 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
987 part = &xpc_partitions[partid];
989 if (xpc_partition_disengaged(part) &&
990 part->act_state == XPC_P_INACTIVE) {
996 XPC_DEACTIVATE_PARTITION(part, reason);
998 if (part->disengage_request_timeout >
999 disengage_request_timeout) {
1000 disengage_request_timeout =
1001 part->disengage_request_timeout;
1005 if (xpc_partition_engaged(-1UL)) {
1006 if (time_after(jiffies, printmsg_time)) {
1007 dev_info(xpc_part, "waiting for remote "
1008 "partitions to disengage, timeout in "
1010 (disengage_request_timeout - jiffies)
1012 printmsg_time = jiffies +
1013 (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
1014 printed_waiting_msg = 1;
1017 } else if (active_part_count > 0) {
1018 if (printed_waiting_msg) {
1019 dev_info(xpc_part, "waiting for local partition"
1021 printed_waiting_msg = 0;
1025 if (!xpc_disengage_request_timedout) {
1026 dev_info(xpc_part, "all partitions have "
1032 /* sleep for a 1/3 of a second or so */
1033 (void) msleep_interruptible(300);
1037 DBUG_ON(xpc_partition_engaged(-1UL));
1040 /* indicate to others that our reserved page is uninitialized */
1041 xpc_rsvd_page->vars_pa = 0;
1043 /* now it's time to eliminate our heartbeat */
1044 del_timer_sync(&xpc_hb_timer);
1045 DBUG_ON(xpc_vars->heartbeating_to_mask != 0);
1047 if (reason == xpcUnloading) {
1048 /* take ourselves off of the reboot_notifier_list */
1049 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1051 /* take ourselves off of the die_notifier list */
1052 (void) unregister_die_notifier(&xpc_die_notifier);
1055 /* close down protections for IPI operations */
1056 xpc_restrict_IPI_ops();
1059 /* clear the interface to XPC's functions */
1060 xpc_clear_interface();
1063 unregister_sysctl_table(xpc_sysctl);
1066 kfree(xpc_remote_copy_buffer_base);
1071 * This function is called when the system is being rebooted.
1074 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1076 enum xpc_retval reason;
1081 reason = xpcSystemReboot;
1084 reason = xpcSystemHalt;
1087 reason = xpcSystemPoweroff;
1090 reason = xpcSystemGoingDown;
1093 xpc_do_exit(reason);
1099 * Notify other partitions to disengage from all references to our memory.
1102 xpc_die_disengage(void)
1104 struct xpc_partition *part;
1106 unsigned long engaged;
1107 long time, printmsg_time, disengage_request_timeout;
1110 /* keep xpc_hb_checker thread from doing anything (just in case) */
1113 xpc_vars->heartbeating_to_mask = 0; /* indicate we're deactivated */
1115 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1116 part = &xpc_partitions[partid];
1118 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part->
1119 remote_vars_version)) {
1121 /* just in case it was left set by an earlier XPC */
1122 xpc_clear_partition_engaged(1UL << partid);
1126 if (xpc_partition_engaged(1UL << partid) ||
1127 part->act_state != XPC_P_INACTIVE) {
1128 xpc_request_partition_disengage(part);
1129 xpc_mark_partition_disengaged(part);
1130 xpc_IPI_send_disengage(part);
1135 printmsg_time = time +
1136 (XPC_DISENGAGE_PRINTMSG_INTERVAL * sn_rtc_cycles_per_second);
1137 disengage_request_timeout = time +
1138 (xpc_disengage_request_timelimit * sn_rtc_cycles_per_second);
1140 /* wait for all other partitions to disengage from us */
1143 engaged = xpc_partition_engaged(-1UL);
1145 dev_info(xpc_part, "all partitions have disengaged\n");
1150 if (time >= disengage_request_timeout) {
1151 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1152 if (engaged & (1UL << partid)) {
1153 dev_info(xpc_part, "disengage from "
1154 "remote partition %d timed "
1161 if (time >= printmsg_time) {
1162 dev_info(xpc_part, "waiting for remote partitions to "
1163 "disengage, timeout in %ld seconds\n",
1164 (disengage_request_timeout - time) /
1165 sn_rtc_cycles_per_second);
1166 printmsg_time = time +
1167 (XPC_DISENGAGE_PRINTMSG_INTERVAL *
1168 sn_rtc_cycles_per_second);
1175 * This function is called when the system is being restarted or halted due
1176 * to some sort of system failure. If this is the case we need to notify the
1177 * other partitions to disengage from all references to our memory.
1178 * This function can also be called when our heartbeater could be offlined
1179 * for a time. In this case we need to notify other partitions to not worry
1180 * about the lack of a heartbeat.
1183 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1186 case DIE_MACHINE_RESTART:
1187 case DIE_MACHINE_HALT:
1188 xpc_die_disengage();
1191 case DIE_KDEBUG_ENTER:
1192 xpc_kdebug_entered = 1;
1194 /* Should lack of heartbeat be ignored by other partitions? */
1195 if (!xpc_kdebug_ignore) {
1199 case DIE_MCA_MONARCH_ENTER:
1200 case DIE_INIT_MONARCH_ENTER:
1201 xpc_vars->heartbeat++;
1202 xpc_vars->heartbeat_offline = 1;
1205 case DIE_KDEBUG_LEAVE:
1206 xpc_kdebug_entered = 0;
1208 /* Is lack of heartbeat being ignored by other partitions? */
1209 if (!xpc_kdebug_ignore) {
1213 case DIE_MCA_MONARCH_LEAVE:
1214 case DIE_INIT_MONARCH_LEAVE:
1215 xpc_vars->heartbeat++;
1216 xpc_vars->heartbeat_offline = 0;
1225 xpc_kdebug_force_disengage(void)
1227 if (!xpc_kdebug_entered) {
1228 /* not being called from kdebug */
1232 xpc_vars->heartbeat_offline = 0;
1233 xpc_die_disengage();
1236 EXPORT_SYMBOL_GPL(xpc_kdebug_force_disengage);
1244 struct xpc_partition *part;
1249 if (!ia64_platform_is("sn2")) {
1254 buf_size = max(XPC_RP_VARS_SIZE,
1255 XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES);
1256 xpc_remote_copy_buffer = xpc_kmalloc_cacheline_aligned(buf_size,
1257 GFP_KERNEL, &xpc_remote_copy_buffer_base);
1258 if (xpc_remote_copy_buffer == NULL)
1261 snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1262 snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1264 xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1267 * The first few fields of each entry of xpc_partitions[] need to
1268 * be initialized now so that calls to xpc_connect() and
1269 * xpc_disconnect() can be made prior to the activation of any remote
1270 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1271 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1272 * PARTITION HAS BEEN ACTIVATED.
1274 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1275 part = &xpc_partitions[partid];
1277 DBUG_ON((u64) part != L1_CACHE_ALIGN((u64) part));
1279 part->act_IRQ_rcvd = 0;
1280 spin_lock_init(&part->act_lock);
1281 part->act_state = XPC_P_INACTIVE;
1282 XPC_SET_REASON(part, 0, 0);
1284 init_timer(&part->disengage_request_timer);
1285 part->disengage_request_timer.function =
1286 xpc_timeout_partition_disengage_request;
1287 part->disengage_request_timer.data = (unsigned long) part;
1289 part->setup_state = XPC_P_UNSET;
1290 init_waitqueue_head(&part->teardown_wq);
1291 atomic_set(&part->references, 0);
1295 * Open up protections for IPI operations (and AMO operations on
1296 * Shub 1.1 systems).
1298 xpc_allow_IPI_ops();
1301 * Interrupts being processed will increment this atomic variable and
1302 * awaken the heartbeat thread which will process the interrupts.
1304 atomic_set(&xpc_act_IRQ_rcvd, 0);
1307 * This is safe to do before the xpc_hb_checker thread has started
1308 * because the handler releases a wait queue. If an interrupt is
1309 * received before the thread is waiting, it will not go to sleep,
1310 * but rather immediately process the interrupt.
1312 ret = request_irq(SGI_XPC_ACTIVATE, xpc_act_IRQ_handler, 0,
1315 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
1316 "errno=%d\n", -ret);
1318 xpc_restrict_IPI_ops();
1321 unregister_sysctl_table(xpc_sysctl);
1324 kfree(xpc_remote_copy_buffer_base);
1329 * Fill the partition reserved page with the information needed by
1330 * other partitions to discover we are alive and establish initial
1333 xpc_rsvd_page = xpc_rsvd_page_init();
1334 if (xpc_rsvd_page == NULL) {
1335 dev_err(xpc_part, "could not setup our reserved page\n");
1337 free_irq(SGI_XPC_ACTIVATE, NULL);
1338 xpc_restrict_IPI_ops();
1341 unregister_sysctl_table(xpc_sysctl);
1344 kfree(xpc_remote_copy_buffer_base);
1349 /* add ourselves to the reboot_notifier_list */
1350 ret = register_reboot_notifier(&xpc_reboot_notifier);
1352 dev_warn(xpc_part, "can't register reboot notifier\n");
1355 /* add ourselves to the die_notifier list */
1356 ret = register_die_notifier(&xpc_die_notifier);
1358 dev_warn(xpc_part, "can't register die notifier\n");
1363 * Set the beating to other partitions into motion. This is
1364 * the last requirement for other partitions' discovery to
1365 * initiate communications with us.
1367 init_timer(&xpc_hb_timer);
1368 xpc_hb_timer.function = xpc_hb_beater;
1373 * The real work-horse behind xpc. This processes incoming
1374 * interrupts and monitors remote heartbeats.
1376 pid = kernel_thread(xpc_hb_checker, NULL, 0);
1378 dev_err(xpc_part, "failed while forking hb check thread\n");
1380 /* indicate to others that our reserved page is uninitialized */
1381 xpc_rsvd_page->vars_pa = 0;
1383 /* take ourselves off of the reboot_notifier_list */
1384 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1386 /* take ourselves off of the die_notifier list */
1387 (void) unregister_die_notifier(&xpc_die_notifier);
1389 del_timer_sync(&xpc_hb_timer);
1390 free_irq(SGI_XPC_ACTIVATE, NULL);
1391 xpc_restrict_IPI_ops();
1394 unregister_sysctl_table(xpc_sysctl);
1397 kfree(xpc_remote_copy_buffer_base);
1403 * Startup a thread that will attempt to discover other partitions to
1404 * activate based on info provided by SAL. This new thread is short
1405 * lived and will exit once discovery is complete.
1407 pid = kernel_thread(xpc_initiate_discovery, NULL, 0);
1409 dev_err(xpc_part, "failed while forking discovery thread\n");
1411 /* mark this new thread as a non-starter */
1412 complete(&xpc_discovery_exited);
1414 xpc_do_exit(xpcUnloading);
1419 /* set the interface to point at XPC's functions */
1420 xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1421 xpc_initiate_allocate, xpc_initiate_send,
1422 xpc_initiate_send_notify, xpc_initiate_received,
1423 xpc_initiate_partid_to_nasids);
1427 module_init(xpc_init);
1433 xpc_do_exit(xpcUnloading);
1435 module_exit(xpc_exit);
1438 MODULE_AUTHOR("Silicon Graphics, Inc.");
1439 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1440 MODULE_LICENSE("GPL");
1442 module_param(xpc_hb_interval, int, 0);
1443 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1444 "heartbeat increments.");
1446 module_param(xpc_hb_check_interval, int, 0);
1447 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1448 "heartbeat checks.");
1450 module_param(xpc_disengage_request_timelimit, int, 0);
1451 MODULE_PARM_DESC(xpc_disengage_request_timelimit, "Number of seconds to wait "
1452 "for disengage request to complete.");
1454 module_param(xpc_kdebug_ignore, int, 0);
1455 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1456 "other partitions when dropping into kdebug.");