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
,
114 &xpc_hb_min_interval
,
120 &xpc_hb_check_interval
,
124 &proc_dointvec_minmax
,
127 &xpc_hb_check_min_interval
,
128 &xpc_hb_check_max_interval
132 static ctl_table xpc_sys_xpc_dir
[] = {
143 "disengage_request_timelimit",
144 &xpc_disengage_request_timelimit
,
148 &proc_dointvec_minmax
,
151 &xpc_disengage_request_min_timelimit
,
152 &xpc_disengage_request_max_timelimit
156 static ctl_table xpc_sys_dir
[] = {
167 static struct ctl_table_header
*xpc_sysctl
;
169 /* non-zero if any remote partition disengage request was timed out */
170 int xpc_disengage_request_timedout
;
172 /* #of IRQs received */
173 static atomic_t xpc_act_IRQ_rcvd
;
175 /* IRQ handler notifies this wait queue on receipt of an IRQ */
176 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq
);
178 static unsigned long xpc_hb_check_timeout
;
180 /* notification that the xpc_hb_checker thread has exited */
181 static DECLARE_COMPLETION(xpc_hb_checker_exited
);
183 /* notification that the xpc_discovery thread has exited */
184 static DECLARE_COMPLETION(xpc_discovery_exited
);
187 static struct timer_list xpc_hb_timer
;
190 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
193 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
194 static struct notifier_block xpc_reboot_notifier
= {
195 .notifier_call
= xpc_system_reboot
,
198 static int xpc_system_die(struct notifier_block
*, unsigned long, void *);
199 static struct notifier_block xpc_die_notifier
= {
200 .notifier_call
= xpc_system_die
,
205 * Timer function to enforce the timelimit on the partition disengage request.
208 xpc_timeout_partition_disengage_request(unsigned long data
)
210 struct xpc_partition
*part
= (struct xpc_partition
*) data
;
213 DBUG_ON(jiffies
< part
->disengage_request_timeout
);
215 (void) xpc_partition_disengaged(part
);
217 DBUG_ON(part
->disengage_request_timeout
!= 0);
218 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part
)) != 0);
223 * Notify the heartbeat check thread that an IRQ has been received.
226 xpc_act_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
228 atomic_inc(&xpc_act_IRQ_rcvd
);
229 wake_up_interruptible(&xpc_act_IRQ_wq
);
235 * Timer to produce the heartbeat. The timer structures function is
236 * already set when this is initially called. A tunable is used to
237 * specify when the next timeout should occur.
240 xpc_hb_beater(unsigned long dummy
)
242 xpc_vars
->heartbeat
++;
244 if (jiffies
>= xpc_hb_check_timeout
) {
245 wake_up_interruptible(&xpc_act_IRQ_wq
);
248 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
249 add_timer(&xpc_hb_timer
);
254 * This thread is responsible for nearly all of the partition
255 * activation/deactivation.
258 xpc_hb_checker(void *ignore
)
260 int last_IRQ_count
= 0;
265 /* this thread was marked active by xpc_hb_init() */
267 daemonize(XPC_HB_CHECK_THREAD_NAME
);
269 set_cpus_allowed(current
, cpumask_of_cpu(XPC_HB_CHECK_CPU
));
271 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
273 while (!(volatile int) xpc_exiting
) {
275 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
277 (int) (xpc_hb_check_timeout
- jiffies
),
278 atomic_read(&xpc_act_IRQ_rcvd
) - last_IRQ_count
);
281 /* checking of remote heartbeats is skewed by IRQ handling */
282 if (jiffies
>= xpc_hb_check_timeout
) {
283 dev_dbg(xpc_part
, "checking remote heartbeats\n");
284 xpc_check_remote_hb();
287 * We need to periodically recheck to ensure no
288 * IPI/AMO pairs have been missed. That check
289 * must always reset xpc_hb_check_timeout.
295 /* check for outstanding IRQs */
296 new_IRQ_count
= atomic_read(&xpc_act_IRQ_rcvd
);
297 if (last_IRQ_count
< new_IRQ_count
|| force_IRQ
!= 0) {
300 dev_dbg(xpc_part
, "found an IRQ to process; will be "
301 "resetting xpc_hb_check_timeout\n");
303 last_IRQ_count
+= xpc_identify_act_IRQ_sender();
304 if (last_IRQ_count
< new_IRQ_count
) {
305 /* retry once to help avoid missing AMO */
306 (void) xpc_identify_act_IRQ_sender();
308 last_IRQ_count
= new_IRQ_count
;
310 xpc_hb_check_timeout
= jiffies
+
311 (xpc_hb_check_interval
* HZ
);
314 /* wait for IRQ or timeout */
315 (void) wait_event_interruptible(xpc_act_IRQ_wq
,
316 (last_IRQ_count
< atomic_read(&xpc_act_IRQ_rcvd
) ||
317 jiffies
>= xpc_hb_check_timeout
||
318 (volatile int) xpc_exiting
));
321 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
324 /* mark this thread as having exited */
325 complete(&xpc_hb_checker_exited
);
331 * This thread will attempt to discover other partitions to activate
332 * based on info provided by SAL. This new thread is short lived and
333 * will exit once discovery is complete.
336 xpc_initiate_discovery(void *ignore
)
338 daemonize(XPC_DISCOVERY_THREAD_NAME
);
342 dev_dbg(xpc_part
, "discovery thread is exiting\n");
344 /* mark this thread as having exited */
345 complete(&xpc_discovery_exited
);
351 * Establish first contact with the remote partititon. This involves pulling
352 * the XPC per partition variables from the remote partition and waiting for
353 * the remote partition to pull ours.
355 static enum xpc_retval
356 xpc_make_first_contact(struct xpc_partition
*part
)
361 while ((ret
= xpc_pull_remote_vars_part(part
)) != xpcSuccess
) {
362 if (ret
!= xpcRetry
) {
363 XPC_DEACTIVATE_PARTITION(part
, ret
);
367 dev_dbg(xpc_chan
, "waiting to make first contact with "
368 "partition %d\n", XPC_PARTID(part
));
370 /* wait a 1/4 of a second or so */
371 (void) msleep_interruptible(250);
373 if (part
->act_state
== XPC_P_DEACTIVATING
) {
378 return xpc_mark_partition_active(part
);
383 * The first kthread assigned to a newly activated partition is the one
384 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
385 * that kthread until the partition is brought down, at which time that kthread
386 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
387 * that XPC has dismantled all communication infrastructure for the associated
388 * partition.) This kthread becomes the channel manager for that partition.
390 * Each active partition has a channel manager, who, besides connecting and
391 * disconnecting channels, will ensure that each of the partition's connected
392 * channels has the required number of assigned kthreads to get the work done.
395 xpc_channel_mgr(struct xpc_partition
*part
)
397 while (part
->act_state
!= XPC_P_DEACTIVATING
||
398 atomic_read(&part
->nchannels_active
) > 0 ||
399 !xpc_partition_disengaged(part
)) {
401 xpc_process_channel_activity(part
);
405 * Wait until we've been requested to activate kthreads or
406 * all of the channel's message queues have been torn down or
407 * a signal is pending.
409 * The channel_mgr_requests is set to 1 after being awakened,
410 * This is done to prevent the channel mgr from making one pass
411 * through the loop for each request, since he will
412 * be servicing all the requests in one pass. The reason it's
413 * set to 1 instead of 0 is so that other kthreads will know
414 * that the channel mgr is running and won't bother trying to
417 atomic_dec(&part
->channel_mgr_requests
);
418 (void) wait_event_interruptible(part
->channel_mgr_wq
,
419 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
420 (volatile u64
) part
->local_IPI_amo
!= 0 ||
421 ((volatile u8
) part
->act_state
==
422 XPC_P_DEACTIVATING
&&
423 atomic_read(&part
->nchannels_active
) == 0 &&
424 xpc_partition_disengaged(part
))));
425 atomic_set(&part
->channel_mgr_requests
, 1);
427 // >>> Does it need to wakeup periodically as well? In case we
428 // >>> miscalculated the #of kthreads to wakeup or create?
434 * When XPC HB determines that a partition has come up, it will create a new
435 * kthread and that kthread will call this function to attempt to set up the
436 * basic infrastructure used for Cross Partition Communication with the newly
439 * The kthread that was created by XPC HB and which setup the XPC
440 * infrastructure will remain assigned to the partition until the partition
441 * goes down. At which time the kthread will teardown the XPC infrastructure
444 * XPC HB will put the remote partition's XPC per partition specific variables
445 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
446 * calling xpc_partition_up().
449 xpc_partition_up(struct xpc_partition
*part
)
451 DBUG_ON(part
->channels
!= NULL
);
453 dev_dbg(xpc_chan
, "activating partition %d\n", XPC_PARTID(part
));
455 if (xpc_setup_infrastructure(part
) != xpcSuccess
) {
460 * The kthread that XPC HB called us with will become the
461 * channel manager for this partition. It will not return
462 * back to XPC HB until the partition's XPC infrastructure
463 * has been dismantled.
466 (void) xpc_part_ref(part
); /* this will always succeed */
468 if (xpc_make_first_contact(part
) == xpcSuccess
) {
469 xpc_channel_mgr(part
);
472 xpc_part_deref(part
);
474 xpc_teardown_infrastructure(part
);
479 xpc_activating(void *__partid
)
481 partid_t partid
= (u64
) __partid
;
482 struct xpc_partition
*part
= &xpc_partitions
[partid
];
483 unsigned long irq_flags
;
484 struct sched_param param
= { sched_priority
: MAX_RT_PRIO
- 1 };
488 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
490 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
492 if (part
->act_state
== XPC_P_DEACTIVATING
) {
493 part
->act_state
= XPC_P_INACTIVE
;
494 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
495 part
->remote_rp_pa
= 0;
499 /* indicate the thread is activating */
500 DBUG_ON(part
->act_state
!= XPC_P_ACTIVATION_REQ
);
501 part
->act_state
= XPC_P_ACTIVATING
;
503 XPC_SET_REASON(part
, 0, 0);
504 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
506 dev_dbg(xpc_part
, "bringing partition %d up\n", partid
);
508 daemonize("xpc%02d", partid
);
511 * This thread needs to run at a realtime priority to prevent a
512 * significant performance degradation.
514 ret
= sched_setscheduler(current
, SCHED_FIFO
, ¶m
);
516 dev_warn(xpc_part
, "unable to set pid %d to a realtime "
517 "priority, ret=%d\n", current
->pid
, ret
);
520 /* allow this thread and its children to run on any CPU */
521 set_cpus_allowed(current
, CPU_MASK_ALL
);
524 * Register the remote partition's AMOs with SAL so it can handle
525 * and cleanup errors within that address range should the remote
526 * partition go down. We don't unregister this range because it is
527 * difficult to tell when outstanding writes to the remote partition
528 * are finished and thus when it is safe to unregister. This should
529 * not result in wasted space in the SAL xp_addr_region table because
530 * we should get the same page for remote_amos_page_pa after module
531 * reloads and system reboots.
533 if (sn_register_xp_addr_region(part
->remote_amos_page_pa
,
535 dev_warn(xpc_part
, "xpc_partition_up(%d) failed to register "
536 "xp_addr region\n", partid
);
538 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
539 part
->act_state
= XPC_P_INACTIVE
;
540 XPC_SET_REASON(part
, xpcPhysAddrRegFailed
, __LINE__
);
541 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
542 part
->remote_rp_pa
= 0;
546 xpc_allow_hb(partid
, xpc_vars
);
547 xpc_IPI_send_activated(part
);
551 * xpc_partition_up() holds this thread and marks this partition as
552 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
554 (void) xpc_partition_up(part
);
556 xpc_disallow_hb(partid
, xpc_vars
);
557 xpc_mark_partition_inactive(part
);
559 if (part
->reason
== xpcReactivating
) {
560 /* interrupting ourselves results in activating partition */
561 xpc_IPI_send_reactivate(part
);
569 xpc_activate_partition(struct xpc_partition
*part
)
571 partid_t partid
= XPC_PARTID(part
);
572 unsigned long irq_flags
;
576 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
578 DBUG_ON(part
->act_state
!= XPC_P_INACTIVE
);
580 part
->act_state
= XPC_P_ACTIVATION_REQ
;
581 XPC_SET_REASON(part
, xpcCloneKThread
, __LINE__
);
583 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
585 pid
= kernel_thread(xpc_activating
, (void *) ((u64
) partid
), 0);
587 if (unlikely(pid
<= 0)) {
588 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
589 part
->act_state
= XPC_P_INACTIVE
;
590 XPC_SET_REASON(part
, xpcCloneKThreadFailed
, __LINE__
);
591 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
597 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
598 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
599 * than one partition, we use an AMO_t structure per partition to indicate
600 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
601 * associated kthread to handle it.
603 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
604 * running on other partitions.
606 * Noteworthy Arguments:
608 * irq - Interrupt ReQuest number. NOT USED.
610 * dev_id - partid of IPI's potential sender.
612 * regs - processor's context before the processor entered
613 * interrupt code. NOT USED.
616 xpc_notify_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
618 partid_t partid
= (partid_t
) (u64
) dev_id
;
619 struct xpc_partition
*part
= &xpc_partitions
[partid
];
622 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
624 if (xpc_part_ref(part
)) {
625 xpc_check_for_channel_activity(part
);
627 xpc_part_deref(part
);
634 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
635 * because the write to their associated IPI amo completed after the IRQ/IPI
639 xpc_dropped_IPI_check(struct xpc_partition
*part
)
641 if (xpc_part_ref(part
)) {
642 xpc_check_for_channel_activity(part
);
644 part
->dropped_IPI_timer
.expires
= jiffies
+
645 XPC_P_DROPPED_IPI_WAIT
;
646 add_timer(&part
->dropped_IPI_timer
);
647 xpc_part_deref(part
);
653 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
655 int idle
= atomic_read(&ch
->kthreads_idle
);
656 int assigned
= atomic_read(&ch
->kthreads_assigned
);
660 DBUG_ON(needed
<= 0);
663 wakeup
= (needed
> idle
) ? idle
: needed
;
666 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
667 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
669 /* only wakeup the requested number of kthreads */
670 wake_up_nr(&ch
->idle_wq
, wakeup
);
677 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
678 needed
= ch
->kthreads_assigned_limit
- assigned
;
679 // >>>should never be less than 0
685 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
686 needed
, ch
->partid
, ch
->number
);
688 xpc_create_kthreads(ch
, needed
);
693 * This function is where XPC's kthreads wait for messages to deliver.
696 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
699 /* deliver messages to their intended recipients */
701 while ((volatile s64
) ch
->w_local_GP
.get
<
702 (volatile s64
) ch
->w_remote_GP
.put
&&
703 !((volatile u32
) ch
->flags
&
704 XPC_C_DISCONNECTING
)) {
708 if (atomic_inc_return(&ch
->kthreads_idle
) >
709 ch
->kthreads_idle_limit
) {
710 /* too many idle kthreads on this channel */
711 atomic_dec(&ch
->kthreads_idle
);
715 dev_dbg(xpc_chan
, "idle kthread calling "
716 "wait_event_interruptible_exclusive()\n");
718 (void) wait_event_interruptible_exclusive(ch
->idle_wq
,
719 ((volatile s64
) ch
->w_local_GP
.get
<
720 (volatile s64
) ch
->w_remote_GP
.put
||
721 ((volatile u32
) ch
->flags
&
722 XPC_C_DISCONNECTING
)));
724 atomic_dec(&ch
->kthreads_idle
);
726 } while (!((volatile u32
) ch
->flags
& XPC_C_DISCONNECTING
));
731 xpc_daemonize_kthread(void *args
)
733 partid_t partid
= XPC_UNPACK_ARG1(args
);
734 u16 ch_number
= XPC_UNPACK_ARG2(args
);
735 struct xpc_partition
*part
= &xpc_partitions
[partid
];
736 struct xpc_channel
*ch
;
738 unsigned long irq_flags
;
741 daemonize("xpc%02dc%d", partid
, ch_number
);
743 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
746 ch
= &part
->channels
[ch_number
];
748 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
750 /* let registerer know that connection has been established */
752 spin_lock_irqsave(&ch
->lock
, irq_flags
);
753 if (!(ch
->flags
& XPC_C_CONNECTCALLOUT
)) {
754 ch
->flags
|= XPC_C_CONNECTCALLOUT
;
755 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
757 xpc_connected_callout(ch
);
760 * It is possible that while the callout was being
761 * made that the remote partition sent some messages.
762 * If that is the case, we may need to activate
763 * additional kthreads to help deliver them. We only
764 * need one less than total #of messages to deliver.
766 n_needed
= ch
->w_remote_GP
.put
- ch
->w_local_GP
.get
- 1;
768 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
769 xpc_activate_kthreads(ch
, n_needed
);
772 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
775 xpc_kthread_waitmsgs(part
, ch
);
778 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0) {
779 spin_lock_irqsave(&ch
->lock
, irq_flags
);
780 if ((ch
->flags
& XPC_C_CONNECTCALLOUT
) &&
781 !(ch
->flags
& XPC_C_DISCONNECTCALLOUT
)) {
782 ch
->flags
|= XPC_C_DISCONNECTCALLOUT
;
783 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
785 xpc_disconnect_callout(ch
, xpcDisconnecting
);
787 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
789 if (atomic_dec_return(&part
->nchannels_engaged
) == 0) {
790 xpc_mark_partition_disengaged(part
);
791 xpc_IPI_send_disengage(part
);
796 xpc_msgqueue_deref(ch
);
798 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
801 xpc_part_deref(part
);
807 * For each partition that XPC has established communications with, there is
808 * a minimum of one kernel thread assigned to perform any operation that
809 * may potentially sleep or block (basically the callouts to the asynchronous
810 * functions registered via xpc_connect()).
812 * Additional kthreads are created and destroyed by XPC as the workload
815 * A kthread is assigned to one of the active channels that exists for a given
819 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
)
821 unsigned long irq_flags
;
823 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
824 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
827 while (needed
-- > 0) {
830 * The following is done on behalf of the newly created
831 * kthread. That kthread is responsible for doing the
832 * counterpart to the following before it exits.
834 (void) xpc_part_ref(part
);
835 xpc_msgqueue_ref(ch
);
836 if (atomic_inc_return(&ch
->kthreads_assigned
) == 1 &&
837 atomic_inc_return(&part
->nchannels_engaged
) == 1) {
838 xpc_mark_partition_engaged(part
);
841 pid
= kernel_thread(xpc_daemonize_kthread
, (void *) args
, 0);
843 /* the fork failed */
844 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
845 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
846 xpc_mark_partition_disengaged(part
);
847 xpc_IPI_send_disengage(part
);
849 xpc_msgqueue_deref(ch
);
850 xpc_part_deref(part
);
852 if (atomic_read(&ch
->kthreads_assigned
) <
853 ch
->kthreads_idle_limit
) {
855 * Flag this as an error only if we have an
856 * insufficient #of kthreads for the channel
859 * No xpc_msgqueue_ref() is needed here since
860 * the channel mgr is doing this.
862 spin_lock_irqsave(&ch
->lock
, irq_flags
);
863 XPC_DISCONNECT_CHANNEL(ch
, xpcLackOfResources
,
865 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
870 ch
->kthreads_created
++; // >>> temporary debug only!!!
876 xpc_disconnect_wait(int ch_number
)
878 unsigned long irq_flags
;
880 struct xpc_partition
*part
;
881 struct xpc_channel
*ch
;
882 int wakeup_channel_mgr
;
885 /* now wait for all callouts to the caller's function to cease */
886 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
887 part
= &xpc_partitions
[partid
];
889 if (!xpc_part_ref(part
)) {
893 ch
= &part
->channels
[ch_number
];
895 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
896 xpc_part_deref(part
);
900 wait_for_completion(&ch
->wdisconnect_wait
);
902 spin_lock_irqsave(&ch
->lock
, irq_flags
);
903 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
904 wakeup_channel_mgr
= 0;
906 if (ch
->delayed_IPI_flags
) {
907 if (part
->act_state
!= XPC_P_DEACTIVATING
) {
908 spin_lock(&part
->IPI_lock
);
909 XPC_SET_IPI_FLAGS(part
->local_IPI_amo
,
910 ch
->number
, ch
->delayed_IPI_flags
);
911 spin_unlock(&part
->IPI_lock
);
912 wakeup_channel_mgr
= 1;
914 ch
->delayed_IPI_flags
= 0;
917 ch
->flags
&= ~XPC_C_WDISCONNECT
;
918 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
920 if (wakeup_channel_mgr
) {
921 xpc_wakeup_channel_mgr(part
);
924 xpc_part_deref(part
);
930 xpc_do_exit(enum xpc_retval reason
)
933 int active_part_count
, printed_waiting_msg
= 0;
934 struct xpc_partition
*part
;
935 unsigned long printmsg_time
, disengage_request_timeout
= 0;
938 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
939 DBUG_ON(xpc_exiting
== 1);
942 * Let the heartbeat checker thread and the discovery thread
943 * (if one is running) know that they should exit. Also wake up
944 * the heartbeat checker thread in case it's sleeping.
947 wake_up_interruptible(&xpc_act_IRQ_wq
);
949 /* ignore all incoming interrupts */
950 free_irq(SGI_XPC_ACTIVATE
, NULL
);
952 /* wait for the discovery thread to exit */
953 wait_for_completion(&xpc_discovery_exited
);
955 /* wait for the heartbeat checker thread to exit */
956 wait_for_completion(&xpc_hb_checker_exited
);
959 /* sleep for a 1/3 of a second or so */
960 (void) msleep_interruptible(300);
963 /* wait for all partitions to become inactive */
965 printmsg_time
= jiffies
+ (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
966 xpc_disengage_request_timedout
= 0;
969 active_part_count
= 0;
971 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
972 part
= &xpc_partitions
[partid
];
974 if (xpc_partition_disengaged(part
) &&
975 part
->act_state
== XPC_P_INACTIVE
) {
981 XPC_DEACTIVATE_PARTITION(part
, reason
);
983 if (part
->disengage_request_timeout
>
984 disengage_request_timeout
) {
985 disengage_request_timeout
=
986 part
->disengage_request_timeout
;
990 if (xpc_partition_engaged(-1UL)) {
991 if (time_after(jiffies
, printmsg_time
)) {
992 dev_info(xpc_part
, "waiting for remote "
993 "partitions to disengage, timeout in "
995 (disengage_request_timeout
- jiffies
)
997 printmsg_time
= jiffies
+
998 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
999 printed_waiting_msg
= 1;
1002 } else if (active_part_count
> 0) {
1003 if (printed_waiting_msg
) {
1004 dev_info(xpc_part
, "waiting for local partition"
1006 printed_waiting_msg
= 0;
1010 if (!xpc_disengage_request_timedout
) {
1011 dev_info(xpc_part
, "all partitions have "
1017 /* sleep for a 1/3 of a second or so */
1018 (void) msleep_interruptible(300);
1022 DBUG_ON(xpc_partition_engaged(-1UL));
1025 /* indicate to others that our reserved page is uninitialized */
1026 xpc_rsvd_page
->vars_pa
= 0;
1028 /* now it's time to eliminate our heartbeat */
1029 del_timer_sync(&xpc_hb_timer
);
1030 DBUG_ON(xpc_vars
->heartbeating_to_mask
!= 0);
1032 if (reason
== xpcUnloading
) {
1033 /* take ourselves off of the reboot_notifier_list */
1034 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1036 /* take ourselves off of the die_notifier list */
1037 (void) unregister_die_notifier(&xpc_die_notifier
);
1040 /* close down protections for IPI operations */
1041 xpc_restrict_IPI_ops();
1044 /* clear the interface to XPC's functions */
1045 xpc_clear_interface();
1048 unregister_sysctl_table(xpc_sysctl
);
1054 * This function is called when the system is being rebooted.
1057 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1059 enum xpc_retval reason
;
1064 reason
= xpcSystemReboot
;
1067 reason
= xpcSystemHalt
;
1070 reason
= xpcSystemPoweroff
;
1073 reason
= xpcSystemGoingDown
;
1076 xpc_do_exit(reason
);
1082 * Notify other partitions to disengage from all references to our memory.
1085 xpc_die_disengage(void)
1087 struct xpc_partition
*part
;
1089 unsigned long engaged
;
1090 long time
, printmsg_time
, disengage_request_timeout
;
1093 /* keep xpc_hb_checker thread from doing anything (just in case) */
1096 xpc_vars
->heartbeating_to_mask
= 0; /* indicate we're deactivated */
1098 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1099 part
= &xpc_partitions
[partid
];
1101 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part
->
1102 remote_vars_version
)) {
1104 /* just in case it was left set by an earlier XPC */
1105 xpc_clear_partition_engaged(1UL << partid
);
1109 if (xpc_partition_engaged(1UL << partid
) ||
1110 part
->act_state
!= XPC_P_INACTIVE
) {
1111 xpc_request_partition_disengage(part
);
1112 xpc_mark_partition_disengaged(part
);
1113 xpc_IPI_send_disengage(part
);
1118 printmsg_time
= time
+
1119 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* sn_rtc_cycles_per_second
);
1120 disengage_request_timeout
= time
+
1121 (xpc_disengage_request_timelimit
* sn_rtc_cycles_per_second
);
1123 /* wait for all other partitions to disengage from us */
1126 engaged
= xpc_partition_engaged(-1UL);
1128 dev_info(xpc_part
, "all partitions have disengaged\n");
1133 if (time
>= disengage_request_timeout
) {
1134 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1135 if (engaged
& (1UL << partid
)) {
1136 dev_info(xpc_part
, "disengage from "
1137 "remote partition %d timed "
1144 if (time
>= printmsg_time
) {
1145 dev_info(xpc_part
, "waiting for remote partitions to "
1146 "disengage, timeout in %ld seconds\n",
1147 (disengage_request_timeout
- time
) /
1148 sn_rtc_cycles_per_second
);
1149 printmsg_time
= time
+
1150 (XPC_DISENGAGE_PRINTMSG_INTERVAL
*
1151 sn_rtc_cycles_per_second
);
1158 * This function is called when the system is being restarted or halted due
1159 * to some sort of system failure. If this is the case we need to notify the
1160 * other partitions to disengage from all references to our memory.
1161 * This function can also be called when our heartbeater could be offlined
1162 * for a time. In this case we need to notify other partitions to not worry
1163 * about the lack of a heartbeat.
1166 xpc_system_die(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1169 case DIE_MACHINE_RESTART
:
1170 case DIE_MACHINE_HALT
:
1171 xpc_die_disengage();
1174 case DIE_KDEBUG_ENTER
:
1175 /* Should lack of heartbeat be ignored by other partitions? */
1176 if (!xpc_kdebug_ignore
) {
1180 case DIE_MCA_MONARCH_ENTER
:
1181 case DIE_INIT_MONARCH_ENTER
:
1182 xpc_vars
->heartbeat
++;
1183 xpc_vars
->heartbeat_offline
= 1;
1186 case DIE_KDEBUG_LEAVE
:
1187 /* Is lack of heartbeat being ignored by other partitions? */
1188 if (!xpc_kdebug_ignore
) {
1192 case DIE_MCA_MONARCH_LEAVE
:
1193 case DIE_INIT_MONARCH_LEAVE
:
1194 xpc_vars
->heartbeat
++;
1195 xpc_vars
->heartbeat_offline
= 0;
1208 struct xpc_partition
*part
;
1212 if (!ia64_platform_is("sn2")) {
1217 * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng
1218 * various portions of a partition's reserved page. Its size is based
1219 * on the size of the reserved page header and part_nasids mask. So we
1220 * need to ensure that the other items will fit as well.
1222 if (XPC_RP_VARS_SIZE
> XPC_RP_HEADER_SIZE
+ XP_NASID_MASK_BYTES
) {
1223 dev_err(xpc_part
, "xpc_remote_copy_buffer is not big enough\n");
1226 DBUG_ON((u64
) xpc_remote_copy_buffer
!=
1227 L1_CACHE_ALIGN((u64
) xpc_remote_copy_buffer
));
1229 snprintf(xpc_part
->bus_id
, BUS_ID_SIZE
, "part");
1230 snprintf(xpc_chan
->bus_id
, BUS_ID_SIZE
, "chan");
1232 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
, 1);
1235 * The first few fields of each entry of xpc_partitions[] need to
1236 * be initialized now so that calls to xpc_connect() and
1237 * xpc_disconnect() can be made prior to the activation of any remote
1238 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1239 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1240 * PARTITION HAS BEEN ACTIVATED.
1242 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1243 part
= &xpc_partitions
[partid
];
1245 DBUG_ON((u64
) part
!= L1_CACHE_ALIGN((u64
) part
));
1247 part
->act_IRQ_rcvd
= 0;
1248 spin_lock_init(&part
->act_lock
);
1249 part
->act_state
= XPC_P_INACTIVE
;
1250 XPC_SET_REASON(part
, 0, 0);
1252 init_timer(&part
->disengage_request_timer
);
1253 part
->disengage_request_timer
.function
=
1254 xpc_timeout_partition_disengage_request
;
1255 part
->disengage_request_timer
.data
= (unsigned long) part
;
1257 part
->setup_state
= XPC_P_UNSET
;
1258 init_waitqueue_head(&part
->teardown_wq
);
1259 atomic_set(&part
->references
, 0);
1263 * Open up protections for IPI operations (and AMO operations on
1264 * Shub 1.1 systems).
1266 xpc_allow_IPI_ops();
1269 * Interrupts being processed will increment this atomic variable and
1270 * awaken the heartbeat thread which will process the interrupts.
1272 atomic_set(&xpc_act_IRQ_rcvd
, 0);
1275 * This is safe to do before the xpc_hb_checker thread has started
1276 * because the handler releases a wait queue. If an interrupt is
1277 * received before the thread is waiting, it will not go to sleep,
1278 * but rather immediately process the interrupt.
1280 ret
= request_irq(SGI_XPC_ACTIVATE
, xpc_act_IRQ_handler
, 0,
1283 dev_err(xpc_part
, "can't register ACTIVATE IRQ handler, "
1284 "errno=%d\n", -ret
);
1286 xpc_restrict_IPI_ops();
1289 unregister_sysctl_table(xpc_sysctl
);
1295 * Fill the partition reserved page with the information needed by
1296 * other partitions to discover we are alive and establish initial
1299 xpc_rsvd_page
= xpc_rsvd_page_init();
1300 if (xpc_rsvd_page
== NULL
) {
1301 dev_err(xpc_part
, "could not setup our reserved page\n");
1303 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1304 xpc_restrict_IPI_ops();
1307 unregister_sysctl_table(xpc_sysctl
);
1313 /* add ourselves to the reboot_notifier_list */
1314 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1316 dev_warn(xpc_part
, "can't register reboot notifier\n");
1319 /* add ourselves to the die_notifier list (i.e., ia64die_chain) */
1320 ret
= register_die_notifier(&xpc_die_notifier
);
1322 dev_warn(xpc_part
, "can't register die notifier\n");
1327 * Set the beating to other partitions into motion. This is
1328 * the last requirement for other partitions' discovery to
1329 * initiate communications with us.
1331 init_timer(&xpc_hb_timer
);
1332 xpc_hb_timer
.function
= xpc_hb_beater
;
1337 * The real work-horse behind xpc. This processes incoming
1338 * interrupts and monitors remote heartbeats.
1340 pid
= kernel_thread(xpc_hb_checker
, NULL
, 0);
1342 dev_err(xpc_part
, "failed while forking hb check thread\n");
1344 /* indicate to others that our reserved page is uninitialized */
1345 xpc_rsvd_page
->vars_pa
= 0;
1347 /* take ourselves off of the reboot_notifier_list */
1348 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1350 /* take ourselves off of the die_notifier list */
1351 (void) unregister_die_notifier(&xpc_die_notifier
);
1353 del_timer_sync(&xpc_hb_timer
);
1354 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1355 xpc_restrict_IPI_ops();
1358 unregister_sysctl_table(xpc_sysctl
);
1365 * Startup a thread that will attempt to discover other partitions to
1366 * activate based on info provided by SAL. This new thread is short
1367 * lived and will exit once discovery is complete.
1369 pid
= kernel_thread(xpc_initiate_discovery
, NULL
, 0);
1371 dev_err(xpc_part
, "failed while forking discovery thread\n");
1373 /* mark this new thread as a non-starter */
1374 complete(&xpc_discovery_exited
);
1376 xpc_do_exit(xpcUnloading
);
1381 /* set the interface to point at XPC's functions */
1382 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1383 xpc_initiate_allocate
, xpc_initiate_send
,
1384 xpc_initiate_send_notify
, xpc_initiate_received
,
1385 xpc_initiate_partid_to_nasids
);
1389 module_init(xpc_init
);
1395 xpc_do_exit(xpcUnloading
);
1397 module_exit(xpc_exit
);
1400 MODULE_AUTHOR("Silicon Graphics, Inc.");
1401 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1402 MODULE_LICENSE("GPL");
1404 module_param(xpc_hb_interval
, int, 0);
1405 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1406 "heartbeat increments.");
1408 module_param(xpc_hb_check_interval
, int, 0);
1409 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1410 "heartbeat checks.");
1412 module_param(xpc_disengage_request_timelimit
, int, 0);
1413 MODULE_PARM_DESC(xpc_disengage_request_timelimit
, "Number of seconds to wait "
1414 "for disengage request to complete.");
1416 module_param(xpc_kdebug_ignore
, int, 0);
1417 MODULE_PARM_DESC(xpc_kdebug_ignore
, "Should lack of heartbeat be ignored by "
1418 "other partitions when dropping into kdebug.");