2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Copyright (c) 2004-2005 Silicon Graphics, Inc. All Rights Reserved.
11 * Cross Partition Communication (XPC) support - standard version.
13 * XPC provides a message passing capability that crosses partition
14 * boundaries. This module is made up of two parts:
16 * partition This part detects the presence/absence of other
17 * partitions. It provides a heartbeat and monitors
18 * the heartbeats of other partitions.
20 * channel This part manages the channels and sends/receives
21 * messages across them to/from other partitions.
23 * There are a couple of additional functions residing in XP, which
24 * provide an interface to XPC for its users.
29 * . We currently have no way to determine which nasid an IPI came
30 * from. Thus, xpc_IPI_send() does a remote AMO write followed by
31 * an IPI. The AMO indicates where data is to be pulled from, so
32 * after the IPI arrives, the remote partition checks the AMO word.
33 * The IPI can actually arrive before the AMO however, so other code
34 * must periodically check for this case. Also, remote AMO operations
35 * do not reliably time out. Thus we do a remote PIO read solely to
36 * know whether the remote partition is down and whether we should
37 * stop sending IPIs to it. This remote PIO read operation is set up
38 * in a special nofault region so SAL knows to ignore (and cleanup)
39 * any errors due to the remote AMO write, PIO read, and/or PIO
42 * If/when new hardware solves this IPI problem, we should abandon
43 * the current approach.
48 #include <linux/kernel.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
51 #include <linux/sched.h>
52 #include <linux/syscalls.h>
53 #include <linux/cache.h>
54 #include <linux/interrupt.h>
55 #include <linux/slab.h>
56 #include <linux/delay.h>
57 #include <linux/reboot.h>
58 #include <asm/sn/intr.h>
59 #include <asm/sn/sn_sal.h>
60 #include <asm/uaccess.h>
64 /* define two XPC debug device structures to be used with dev_dbg() et al */
66 struct device_driver xpc_dbg_name
= {
70 struct device xpc_part_dbg_subname
= {
71 .bus_id
= {0}, /* set to "part" at xpc_init() time */
72 .driver
= &xpc_dbg_name
75 struct device xpc_chan_dbg_subname
= {
76 .bus_id
= {0}, /* set to "chan" at xpc_init() time */
77 .driver
= &xpc_dbg_name
80 struct device
*xpc_part
= &xpc_part_dbg_subname
;
81 struct device
*xpc_chan
= &xpc_chan_dbg_subname
;
84 /* systune related variables for /proc/sys directories */
86 static int xpc_hb_interval
= XPC_HB_DEFAULT_INTERVAL
;
87 static int xpc_hb_min_interval
= 1;
88 static int xpc_hb_max_interval
= 10;
90 static int xpc_hb_check_interval
= XPC_HB_CHECK_DEFAULT_INTERVAL
;
91 static int xpc_hb_check_min_interval
= 10;
92 static int xpc_hb_check_max_interval
= 120;
94 static ctl_table xpc_sys_xpc_hb_dir
[] = {
102 &proc_dointvec_minmax
,
105 &xpc_hb_min_interval
,
111 &xpc_hb_check_interval
,
115 &proc_dointvec_minmax
,
118 &xpc_hb_check_min_interval
,
119 &xpc_hb_check_max_interval
123 static ctl_table xpc_sys_xpc_dir
[] = {
134 static ctl_table xpc_sys_dir
[] = {
145 static struct ctl_table_header
*xpc_sysctl
;
148 /* #of IRQs received */
149 static atomic_t xpc_act_IRQ_rcvd
;
151 /* IRQ handler notifies this wait queue on receipt of an IRQ */
152 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq
);
154 static unsigned long xpc_hb_check_timeout
;
156 /* used as an indication of when the xpc_hb_checker thread is inactive */
157 static DECLARE_MUTEX_LOCKED(xpc_hb_checker_inactive
);
159 /* used as an indication of when the xpc_discovery thread is inactive */
160 static DECLARE_MUTEX_LOCKED(xpc_discovery_inactive
);
163 static struct timer_list xpc_hb_timer
;
166 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
169 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
170 static struct notifier_block xpc_reboot_notifier
= {
171 .notifier_call
= xpc_system_reboot
,
176 * Timer function to enforce the timelimit on the partition disengage request.
179 xpc_timeout_partition_disengage_request(unsigned long data
)
181 struct xpc_partition
*part
= (struct xpc_partition
*) data
;
184 DBUG_ON(XPC_TICKS
< part
->disengage_request_timeout
);
186 (void) xpc_partition_disengaged(part
);
188 DBUG_ON(part
->disengage_request_timeout
!= 0);
189 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part
)) != 0);
194 * Notify the heartbeat check thread that an IRQ has been received.
197 xpc_act_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
199 atomic_inc(&xpc_act_IRQ_rcvd
);
200 wake_up_interruptible(&xpc_act_IRQ_wq
);
206 * Timer to produce the heartbeat. The timer structures function is
207 * already set when this is initially called. A tunable is used to
208 * specify when the next timeout should occur.
211 xpc_hb_beater(unsigned long dummy
)
213 xpc_vars
->heartbeat
++;
215 if (jiffies
>= xpc_hb_check_timeout
) {
216 wake_up_interruptible(&xpc_act_IRQ_wq
);
219 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
220 add_timer(&xpc_hb_timer
);
225 * This thread is responsible for nearly all of the partition
226 * activation/deactivation.
229 xpc_hb_checker(void *ignore
)
231 int last_IRQ_count
= 0;
236 /* this thread was marked active by xpc_hb_init() */
238 daemonize(XPC_HB_CHECK_THREAD_NAME
);
240 set_cpus_allowed(current
, cpumask_of_cpu(XPC_HB_CHECK_CPU
));
242 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
244 while (!(volatile int) xpc_exiting
) {
246 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
248 (int) (xpc_hb_check_timeout
- jiffies
),
249 atomic_read(&xpc_act_IRQ_rcvd
) - last_IRQ_count
);
252 /* checking of remote heartbeats is skewed by IRQ handling */
253 if (jiffies
>= xpc_hb_check_timeout
) {
254 dev_dbg(xpc_part
, "checking remote heartbeats\n");
255 xpc_check_remote_hb();
258 * We need to periodically recheck to ensure no
259 * IPI/AMO pairs have been missed. That check
260 * must always reset xpc_hb_check_timeout.
266 /* check for outstanding IRQs */
267 new_IRQ_count
= atomic_read(&xpc_act_IRQ_rcvd
);
268 if (last_IRQ_count
< new_IRQ_count
|| force_IRQ
!= 0) {
271 dev_dbg(xpc_part
, "found an IRQ to process; will be "
272 "resetting xpc_hb_check_timeout\n");
274 last_IRQ_count
+= xpc_identify_act_IRQ_sender();
275 if (last_IRQ_count
< new_IRQ_count
) {
276 /* retry once to help avoid missing AMO */
277 (void) xpc_identify_act_IRQ_sender();
279 last_IRQ_count
= new_IRQ_count
;
281 xpc_hb_check_timeout
= jiffies
+
282 (xpc_hb_check_interval
* HZ
);
285 /* wait for IRQ or timeout */
286 (void) wait_event_interruptible(xpc_act_IRQ_wq
,
287 (last_IRQ_count
< atomic_read(&xpc_act_IRQ_rcvd
) ||
288 jiffies
>= xpc_hb_check_timeout
||
289 (volatile int) xpc_exiting
));
292 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
295 /* mark this thread as inactive */
296 up(&xpc_hb_checker_inactive
);
302 * This thread will attempt to discover other partitions to activate
303 * based on info provided by SAL. This new thread is short lived and
304 * will exit once discovery is complete.
307 xpc_initiate_discovery(void *ignore
)
309 daemonize(XPC_DISCOVERY_THREAD_NAME
);
313 dev_dbg(xpc_part
, "discovery thread is exiting\n");
315 /* mark this thread as inactive */
316 up(&xpc_discovery_inactive
);
322 * Establish first contact with the remote partititon. This involves pulling
323 * the XPC per partition variables from the remote partition and waiting for
324 * the remote partition to pull ours.
326 static enum xpc_retval
327 xpc_make_first_contact(struct xpc_partition
*part
)
332 while ((ret
= xpc_pull_remote_vars_part(part
)) != xpcSuccess
) {
333 if (ret
!= xpcRetry
) {
334 XPC_DEACTIVATE_PARTITION(part
, ret
);
338 dev_dbg(xpc_chan
, "waiting to make first contact with "
339 "partition %d\n", XPC_PARTID(part
));
341 /* wait a 1/4 of a second or so */
342 (void) msleep_interruptible(250);
344 if (part
->act_state
== XPC_P_DEACTIVATING
) {
349 return xpc_mark_partition_active(part
);
354 * The first kthread assigned to a newly activated partition is the one
355 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
356 * that kthread until the partition is brought down, at which time that kthread
357 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
358 * that XPC has dismantled all communication infrastructure for the associated
359 * partition.) This kthread becomes the channel manager for that partition.
361 * Each active partition has a channel manager, who, besides connecting and
362 * disconnecting channels, will ensure that each of the partition's connected
363 * channels has the required number of assigned kthreads to get the work done.
366 xpc_channel_mgr(struct xpc_partition
*part
)
368 while (part
->act_state
!= XPC_P_DEACTIVATING
||
369 atomic_read(&part
->nchannels_active
) > 0 ||
370 !xpc_partition_disengaged(part
)) {
372 xpc_process_channel_activity(part
);
376 * Wait until we've been requested to activate kthreads or
377 * all of the channel's message queues have been torn down or
378 * a signal is pending.
380 * The channel_mgr_requests is set to 1 after being awakened,
381 * This is done to prevent the channel mgr from making one pass
382 * through the loop for each request, since he will
383 * be servicing all the requests in one pass. The reason it's
384 * set to 1 instead of 0 is so that other kthreads will know
385 * that the channel mgr is running and won't bother trying to
388 atomic_dec(&part
->channel_mgr_requests
);
389 (void) wait_event_interruptible(part
->channel_mgr_wq
,
390 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
391 (volatile u64
) part
->local_IPI_amo
!= 0 ||
392 ((volatile u8
) part
->act_state
==
393 XPC_P_DEACTIVATING
&&
394 atomic_read(&part
->nchannels_active
) == 0 &&
395 xpc_partition_disengaged(part
))));
396 atomic_set(&part
->channel_mgr_requests
, 1);
398 // >>> Does it need to wakeup periodically as well? In case we
399 // >>> miscalculated the #of kthreads to wakeup or create?
405 * When XPC HB determines that a partition has come up, it will create a new
406 * kthread and that kthread will call this function to attempt to set up the
407 * basic infrastructure used for Cross Partition Communication with the newly
410 * The kthread that was created by XPC HB and which setup the XPC
411 * infrastructure will remain assigned to the partition until the partition
412 * goes down. At which time the kthread will teardown the XPC infrastructure
415 * XPC HB will put the remote partition's XPC per partition specific variables
416 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
417 * calling xpc_partition_up().
420 xpc_partition_up(struct xpc_partition
*part
)
422 DBUG_ON(part
->channels
!= NULL
);
424 dev_dbg(xpc_chan
, "activating partition %d\n", XPC_PARTID(part
));
426 if (xpc_setup_infrastructure(part
) != xpcSuccess
) {
431 * The kthread that XPC HB called us with will become the
432 * channel manager for this partition. It will not return
433 * back to XPC HB until the partition's XPC infrastructure
434 * has been dismantled.
437 (void) xpc_part_ref(part
); /* this will always succeed */
439 if (xpc_make_first_contact(part
) == xpcSuccess
) {
440 xpc_channel_mgr(part
);
443 xpc_part_deref(part
);
445 xpc_teardown_infrastructure(part
);
450 xpc_activating(void *__partid
)
452 partid_t partid
= (u64
) __partid
;
453 struct xpc_partition
*part
= &xpc_partitions
[partid
];
454 unsigned long irq_flags
;
455 struct sched_param param
= { sched_priority
: MAX_RT_PRIO
- 1 };
459 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
461 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
463 if (part
->act_state
== XPC_P_DEACTIVATING
) {
464 part
->act_state
= XPC_P_INACTIVE
;
465 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
466 part
->remote_rp_pa
= 0;
470 /* indicate the thread is activating */
471 DBUG_ON(part
->act_state
!= XPC_P_ACTIVATION_REQ
);
472 part
->act_state
= XPC_P_ACTIVATING
;
474 XPC_SET_REASON(part
, 0, 0);
475 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
477 dev_dbg(xpc_part
, "bringing partition %d up\n", partid
);
479 daemonize("xpc%02d", partid
);
482 * This thread needs to run at a realtime priority to prevent a
483 * significant performance degradation.
485 ret
= sched_setscheduler(current
, SCHED_FIFO
, ¶m
);
487 dev_warn(xpc_part
, "unable to set pid %d to a realtime "
488 "priority, ret=%d\n", current
->pid
, ret
);
491 /* allow this thread and its children to run on any CPU */
492 set_cpus_allowed(current
, CPU_MASK_ALL
);
495 * Register the remote partition's AMOs with SAL so it can handle
496 * and cleanup errors within that address range should the remote
497 * partition go down. We don't unregister this range because it is
498 * difficult to tell when outstanding writes to the remote partition
499 * are finished and thus when it is safe to unregister. This should
500 * not result in wasted space in the SAL xp_addr_region table because
501 * we should get the same page for remote_amos_page_pa after module
502 * reloads and system reboots.
504 if (sn_register_xp_addr_region(part
->remote_amos_page_pa
,
506 dev_warn(xpc_part
, "xpc_partition_up(%d) failed to register "
507 "xp_addr region\n", partid
);
509 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
510 part
->act_state
= XPC_P_INACTIVE
;
511 XPC_SET_REASON(part
, xpcPhysAddrRegFailed
, __LINE__
);
512 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
513 part
->remote_rp_pa
= 0;
517 xpc_allow_hb(partid
, xpc_vars
);
518 xpc_IPI_send_activated(part
);
522 * xpc_partition_up() holds this thread and marks this partition as
523 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
525 (void) xpc_partition_up(part
);
527 xpc_disallow_hb(partid
, xpc_vars
);
528 xpc_mark_partition_inactive(part
);
530 if (part
->reason
== xpcReactivating
) {
531 /* interrupting ourselves results in activating partition */
532 xpc_IPI_send_reactivate(part
);
540 xpc_activate_partition(struct xpc_partition
*part
)
542 partid_t partid
= XPC_PARTID(part
);
543 unsigned long irq_flags
;
547 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
549 pid
= kernel_thread(xpc_activating
, (void *) ((u64
) partid
), 0);
551 DBUG_ON(part
->act_state
!= XPC_P_INACTIVE
);
554 part
->act_state
= XPC_P_ACTIVATION_REQ
;
555 XPC_SET_REASON(part
, xpcCloneKThread
, __LINE__
);
557 XPC_SET_REASON(part
, xpcCloneKThreadFailed
, __LINE__
);
560 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
565 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
566 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
567 * than one partition, we use an AMO_t structure per partition to indicate
568 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
569 * associated kthread to handle it.
571 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
572 * running on other partitions.
574 * Noteworthy Arguments:
576 * irq - Interrupt ReQuest number. NOT USED.
578 * dev_id - partid of IPI's potential sender.
580 * regs - processor's context before the processor entered
581 * interrupt code. NOT USED.
584 xpc_notify_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
586 partid_t partid
= (partid_t
) (u64
) dev_id
;
587 struct xpc_partition
*part
= &xpc_partitions
[partid
];
590 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
592 if (xpc_part_ref(part
)) {
593 xpc_check_for_channel_activity(part
);
595 xpc_part_deref(part
);
602 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
603 * because the write to their associated IPI amo completed after the IRQ/IPI
607 xpc_dropped_IPI_check(struct xpc_partition
*part
)
609 if (xpc_part_ref(part
)) {
610 xpc_check_for_channel_activity(part
);
612 part
->dropped_IPI_timer
.expires
= jiffies
+
613 XPC_P_DROPPED_IPI_WAIT
;
614 add_timer(&part
->dropped_IPI_timer
);
615 xpc_part_deref(part
);
621 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
623 int idle
= atomic_read(&ch
->kthreads_idle
);
624 int assigned
= atomic_read(&ch
->kthreads_assigned
);
628 DBUG_ON(needed
<= 0);
631 wakeup
= (needed
> idle
) ? idle
: needed
;
634 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
635 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
637 /* only wakeup the requested number of kthreads */
638 wake_up_nr(&ch
->idle_wq
, wakeup
);
645 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
646 needed
= ch
->kthreads_assigned_limit
- assigned
;
647 // >>>should never be less than 0
653 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
654 needed
, ch
->partid
, ch
->number
);
656 xpc_create_kthreads(ch
, needed
);
661 * This function is where XPC's kthreads wait for messages to deliver.
664 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
667 /* deliver messages to their intended recipients */
669 while ((volatile s64
) ch
->w_local_GP
.get
<
670 (volatile s64
) ch
->w_remote_GP
.put
&&
671 !((volatile u32
) ch
->flags
&
672 XPC_C_DISCONNECTING
)) {
676 if (atomic_inc_return(&ch
->kthreads_idle
) >
677 ch
->kthreads_idle_limit
) {
678 /* too many idle kthreads on this channel */
679 atomic_dec(&ch
->kthreads_idle
);
683 dev_dbg(xpc_chan
, "idle kthread calling "
684 "wait_event_interruptible_exclusive()\n");
686 (void) wait_event_interruptible_exclusive(ch
->idle_wq
,
687 ((volatile s64
) ch
->w_local_GP
.get
<
688 (volatile s64
) ch
->w_remote_GP
.put
||
689 ((volatile u32
) ch
->flags
&
690 XPC_C_DISCONNECTING
)));
692 atomic_dec(&ch
->kthreads_idle
);
694 } while (!((volatile u32
) ch
->flags
& XPC_C_DISCONNECTING
));
699 xpc_daemonize_kthread(void *args
)
701 partid_t partid
= XPC_UNPACK_ARG1(args
);
702 u16 ch_number
= XPC_UNPACK_ARG2(args
);
703 struct xpc_partition
*part
= &xpc_partitions
[partid
];
704 struct xpc_channel
*ch
;
708 daemonize("xpc%02dc%d", partid
, ch_number
);
710 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
713 ch
= &part
->channels
[ch_number
];
715 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
716 DBUG_ON(!(ch
->flags
& XPC_C_CONNECTED
));
718 /* let registerer know that connection has been established */
720 if (atomic_read(&ch
->kthreads_assigned
) == 1) {
721 xpc_connected_callout(ch
);
724 * It is possible that while the callout was being
725 * made that the remote partition sent some messages.
726 * If that is the case, we may need to activate
727 * additional kthreads to help deliver them. We only
728 * need one less than total #of messages to deliver.
730 n_needed
= ch
->w_remote_GP
.put
- ch
->w_local_GP
.get
- 1;
732 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
733 xpc_activate_kthreads(ch
, n_needed
);
737 xpc_kthread_waitmsgs(part
, ch
);
740 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0) {
741 if (ch
->flags
& XPC_C_CONNECTCALLOUT
) {
742 xpc_disconnecting_callout(ch
);
744 if (atomic_dec_return(&part
->nchannels_engaged
) == 0) {
745 xpc_mark_partition_disengaged(part
);
746 xpc_IPI_send_disengage(part
);
751 xpc_msgqueue_deref(ch
);
753 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
756 xpc_part_deref(part
);
762 * For each partition that XPC has established communications with, there is
763 * a minimum of one kernel thread assigned to perform any operation that
764 * may potentially sleep or block (basically the callouts to the asynchronous
765 * functions registered via xpc_connect()).
767 * Additional kthreads are created and destroyed by XPC as the workload
770 * A kthread is assigned to one of the active channels that exists for a given
774 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
)
776 unsigned long irq_flags
;
778 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
779 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
782 while (needed
-- > 0) {
783 pid
= kernel_thread(xpc_daemonize_kthread
, (void *) args
, 0);
785 /* the fork failed */
787 if (atomic_read(&ch
->kthreads_assigned
) <
788 ch
->kthreads_idle_limit
) {
790 * Flag this as an error only if we have an
791 * insufficient #of kthreads for the channel
794 * No xpc_msgqueue_ref() is needed here since
795 * the channel mgr is doing this.
797 spin_lock_irqsave(&ch
->lock
, irq_flags
);
798 XPC_DISCONNECT_CHANNEL(ch
, xpcLackOfResources
,
800 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
806 * The following is done on behalf of the newly created
807 * kthread. That kthread is responsible for doing the
808 * counterpart to the following before it exits.
810 (void) xpc_part_ref(part
);
811 xpc_msgqueue_ref(ch
);
812 if (atomic_inc_return(&ch
->kthreads_assigned
) == 1) {
813 if (atomic_inc_return(&part
->nchannels_engaged
) == 1) {
814 xpc_mark_partition_engaged(part
);
817 ch
->kthreads_created
++; // >>> temporary debug only!!!
823 xpc_disconnect_wait(int ch_number
)
825 unsigned long irq_flags
;
827 struct xpc_partition
*part
;
828 struct xpc_channel
*ch
;
831 /* now wait for all callouts to the caller's function to cease */
832 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
833 part
= &xpc_partitions
[partid
];
835 if (xpc_part_ref(part
)) {
836 ch
= &part
->channels
[ch_number
];
838 if (ch
->flags
& XPC_C_WDISCONNECT
) {
839 if (!(ch
->flags
& XPC_C_DISCONNECTED
)) {
840 (void) down(&ch
->wdisconnect_sema
);
842 spin_lock_irqsave(&ch
->lock
, irq_flags
);
843 ch
->flags
&= ~XPC_C_WDISCONNECT
;
844 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
847 xpc_part_deref(part
);
854 xpc_do_exit(enum xpc_retval reason
)
857 int active_part_count
;
858 struct xpc_partition
*part
;
859 unsigned long printmsg_time
;
862 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
863 DBUG_ON(xpc_exiting
== 1);
866 * Let the heartbeat checker thread and the discovery thread
867 * (if one is running) know that they should exit. Also wake up
868 * the heartbeat checker thread in case it's sleeping.
871 wake_up_interruptible(&xpc_act_IRQ_wq
);
873 /* ignore all incoming interrupts */
874 free_irq(SGI_XPC_ACTIVATE
, NULL
);
876 /* wait for the discovery thread to mark itself inactive */
877 down(&xpc_discovery_inactive
);
879 /* wait for the heartbeat checker thread to mark itself inactive */
880 down(&xpc_hb_checker_inactive
);
883 /* sleep for a 1/3 of a second or so */
884 (void) msleep_interruptible(300);
887 /* wait for all partitions to become inactive */
889 printmsg_time
= jiffies
;
892 active_part_count
= 0;
894 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
895 part
= &xpc_partitions
[partid
];
896 if (xpc_partition_disengaged(part
) &&
897 part
->act_state
== XPC_P_INACTIVE
) {
903 XPC_DEACTIVATE_PARTITION(part
, reason
);
906 if (active_part_count
== 0) {
910 if (jiffies
>= printmsg_time
) {
911 dev_info(xpc_part
, "waiting for partitions to "
912 "deactivate/disengage, active count=%d, remote "
913 "engaged=0x%lx\n", active_part_count
,
914 xpc_partition_engaged(1UL << partid
));
916 printmsg_time
= jiffies
+
917 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
920 /* sleep for a 1/3 of a second or so */
921 (void) msleep_interruptible(300);
925 DBUG_ON(xpc_partition_engaged(-1UL));
928 /* indicate to others that our reserved page is uninitialized */
929 xpc_rsvd_page
->vars_pa
= 0;
931 /* now it's time to eliminate our heartbeat */
932 del_timer_sync(&xpc_hb_timer
);
933 DBUG_ON(xpc_vars
->heartbeating_to_mask
== 0);
935 /* take ourselves off of the reboot_notifier_list */
936 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
938 /* close down protections for IPI operations */
939 xpc_restrict_IPI_ops();
942 /* clear the interface to XPC's functions */
943 xpc_clear_interface();
946 unregister_sysctl_table(xpc_sysctl
);
952 * This function is called when the system is being rebooted.
955 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
957 enum xpc_retval reason
;
962 reason
= xpcSystemReboot
;
965 reason
= xpcSystemHalt
;
968 reason
= xpcSystemPoweroff
;
971 reason
= xpcSystemGoingDown
;
984 struct xpc_partition
*part
;
989 * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng
990 * both a partition's reserved page and its XPC variables. Its size was
991 * based on the size of a reserved page. So we need to ensure that the
992 * XPC variables will fit as well.
994 if (XPC_VARS_ALIGNED_SIZE
> XPC_RSVD_PAGE_ALIGNED_SIZE
) {
995 dev_err(xpc_part
, "xpc_remote_copy_buffer is not big enough\n");
998 DBUG_ON((u64
) xpc_remote_copy_buffer
!=
999 L1_CACHE_ALIGN((u64
) xpc_remote_copy_buffer
));
1001 snprintf(xpc_part
->bus_id
, BUS_ID_SIZE
, "part");
1002 snprintf(xpc_chan
->bus_id
, BUS_ID_SIZE
, "chan");
1004 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
, 1);
1007 * The first few fields of each entry of xpc_partitions[] need to
1008 * be initialized now so that calls to xpc_connect() and
1009 * xpc_disconnect() can be made prior to the activation of any remote
1010 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1011 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1012 * PARTITION HAS BEEN ACTIVATED.
1014 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1015 part
= &xpc_partitions
[partid
];
1017 DBUG_ON((u64
) part
!= L1_CACHE_ALIGN((u64
) part
));
1019 part
->act_IRQ_rcvd
= 0;
1020 spin_lock_init(&part
->act_lock
);
1021 part
->act_state
= XPC_P_INACTIVE
;
1022 XPC_SET_REASON(part
, 0, 0);
1024 init_timer(&part
->disengage_request_timer
);
1025 part
->disengage_request_timer
.function
=
1026 xpc_timeout_partition_disengage_request
;
1027 part
->disengage_request_timer
.data
= (unsigned long) part
;
1029 part
->setup_state
= XPC_P_UNSET
;
1030 init_waitqueue_head(&part
->teardown_wq
);
1031 atomic_set(&part
->references
, 0);
1035 * Open up protections for IPI operations (and AMO operations on
1036 * Shub 1.1 systems).
1038 xpc_allow_IPI_ops();
1041 * Interrupts being processed will increment this atomic variable and
1042 * awaken the heartbeat thread which will process the interrupts.
1044 atomic_set(&xpc_act_IRQ_rcvd
, 0);
1047 * This is safe to do before the xpc_hb_checker thread has started
1048 * because the handler releases a wait queue. If an interrupt is
1049 * received before the thread is waiting, it will not go to sleep,
1050 * but rather immediately process the interrupt.
1052 ret
= request_irq(SGI_XPC_ACTIVATE
, xpc_act_IRQ_handler
, 0,
1055 dev_err(xpc_part
, "can't register ACTIVATE IRQ handler, "
1056 "errno=%d\n", -ret
);
1058 xpc_restrict_IPI_ops();
1061 unregister_sysctl_table(xpc_sysctl
);
1067 * Fill the partition reserved page with the information needed by
1068 * other partitions to discover we are alive and establish initial
1071 xpc_rsvd_page
= xpc_rsvd_page_init();
1072 if (xpc_rsvd_page
== NULL
) {
1073 dev_err(xpc_part
, "could not setup our reserved page\n");
1075 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1076 xpc_restrict_IPI_ops();
1079 unregister_sysctl_table(xpc_sysctl
);
1085 /* add ourselves to the reboot_notifier_list */
1086 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1088 dev_warn(xpc_part
, "can't register reboot notifier\n");
1093 * Set the beating to other partitions into motion. This is
1094 * the last requirement for other partitions' discovery to
1095 * initiate communications with us.
1097 init_timer(&xpc_hb_timer
);
1098 xpc_hb_timer
.function
= xpc_hb_beater
;
1103 * The real work-horse behind xpc. This processes incoming
1104 * interrupts and monitors remote heartbeats.
1106 pid
= kernel_thread(xpc_hb_checker
, NULL
, 0);
1108 dev_err(xpc_part
, "failed while forking hb check thread\n");
1110 /* indicate to others that our reserved page is uninitialized */
1111 xpc_rsvd_page
->vars_pa
= 0;
1113 /* take ourselves off of the reboot_notifier_list */
1114 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1116 del_timer_sync(&xpc_hb_timer
);
1117 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1118 xpc_restrict_IPI_ops();
1121 unregister_sysctl_table(xpc_sysctl
);
1128 * Startup a thread that will attempt to discover other partitions to
1129 * activate based on info provided by SAL. This new thread is short
1130 * lived and will exit once discovery is complete.
1132 pid
= kernel_thread(xpc_initiate_discovery
, NULL
, 0);
1134 dev_err(xpc_part
, "failed while forking discovery thread\n");
1136 /* mark this new thread as a non-starter */
1137 up(&xpc_discovery_inactive
);
1139 xpc_do_exit(xpcUnloading
);
1144 /* set the interface to point at XPC's functions */
1145 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1146 xpc_initiate_allocate
, xpc_initiate_send
,
1147 xpc_initiate_send_notify
, xpc_initiate_received
,
1148 xpc_initiate_partid_to_nasids
);
1152 module_init(xpc_init
);
1158 xpc_do_exit(xpcUnloading
);
1160 module_exit(xpc_exit
);
1163 MODULE_AUTHOR("Silicon Graphics, Inc.");
1164 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1165 MODULE_LICENSE("GPL");
1167 module_param(xpc_hb_interval
, int, 0);
1168 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1169 "heartbeat increments.");
1171 module_param(xpc_hb_check_interval
, int, 0);
1172 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1173 "heartbeat checks.");