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 <asm/sn/intr.h>
59 #include <asm/sn/sn_sal.h>
60 #include <asm/kdebug.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
;
88 /* systune related variables for /proc/sys directories */
90 static int xpc_hb_interval
= XPC_HB_DEFAULT_INTERVAL
;
91 static int xpc_hb_min_interval
= 1;
92 static int xpc_hb_max_interval
= 10;
94 static int xpc_hb_check_interval
= XPC_HB_CHECK_DEFAULT_INTERVAL
;
95 static int xpc_hb_check_min_interval
= 10;
96 static int xpc_hb_check_max_interval
= 120;
98 int xpc_disengage_request_timelimit
= XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT
;
99 static int xpc_disengage_request_min_timelimit
= 0;
100 static int xpc_disengage_request_max_timelimit
= 120;
102 static ctl_table xpc_sys_xpc_hb_dir
[] = {
110 &proc_dointvec_minmax
,
113 &xpc_hb_min_interval
,
119 &xpc_hb_check_interval
,
123 &proc_dointvec_minmax
,
126 &xpc_hb_check_min_interval
,
127 &xpc_hb_check_max_interval
131 static ctl_table xpc_sys_xpc_dir
[] = {
142 "disengage_request_timelimit",
143 &xpc_disengage_request_timelimit
,
147 &proc_dointvec_minmax
,
150 &xpc_disengage_request_min_timelimit
,
151 &xpc_disengage_request_max_timelimit
155 static ctl_table xpc_sys_dir
[] = {
166 static struct ctl_table_header
*xpc_sysctl
;
168 /* non-zero if any remote partition disengage request was timed out */
169 int xpc_disengage_request_timedout
;
171 /* #of IRQs received */
172 static atomic_t xpc_act_IRQ_rcvd
;
174 /* IRQ handler notifies this wait queue on receipt of an IRQ */
175 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq
);
177 static unsigned long xpc_hb_check_timeout
;
179 /* notification that the xpc_hb_checker thread has exited */
180 static DECLARE_COMPLETION(xpc_hb_checker_exited
);
182 /* notification that the xpc_discovery thread has exited */
183 static DECLARE_COMPLETION(xpc_discovery_exited
);
186 static struct timer_list xpc_hb_timer
;
189 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
192 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
193 static struct notifier_block xpc_reboot_notifier
= {
194 .notifier_call
= xpc_system_reboot
,
197 static int xpc_system_die(struct notifier_block
*, unsigned long, void *);
198 static struct notifier_block xpc_die_notifier
= {
199 .notifier_call
= xpc_system_die
,
204 * Timer function to enforce the timelimit on the partition disengage request.
207 xpc_timeout_partition_disengage_request(unsigned long data
)
209 struct xpc_partition
*part
= (struct xpc_partition
*) data
;
212 DBUG_ON(jiffies
< part
->disengage_request_timeout
);
214 (void) xpc_partition_disengaged(part
);
216 DBUG_ON(part
->disengage_request_timeout
!= 0);
217 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part
)) != 0);
222 * Notify the heartbeat check thread that an IRQ has been received.
225 xpc_act_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
227 atomic_inc(&xpc_act_IRQ_rcvd
);
228 wake_up_interruptible(&xpc_act_IRQ_wq
);
234 * Timer to produce the heartbeat. The timer structures function is
235 * already set when this is initially called. A tunable is used to
236 * specify when the next timeout should occur.
239 xpc_hb_beater(unsigned long dummy
)
241 xpc_vars
->heartbeat
++;
243 if (jiffies
>= xpc_hb_check_timeout
) {
244 wake_up_interruptible(&xpc_act_IRQ_wq
);
247 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
248 add_timer(&xpc_hb_timer
);
253 * This thread is responsible for nearly all of the partition
254 * activation/deactivation.
257 xpc_hb_checker(void *ignore
)
259 int last_IRQ_count
= 0;
264 /* this thread was marked active by xpc_hb_init() */
266 daemonize(XPC_HB_CHECK_THREAD_NAME
);
268 set_cpus_allowed(current
, cpumask_of_cpu(XPC_HB_CHECK_CPU
));
270 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
272 while (!(volatile int) xpc_exiting
) {
274 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
276 (int) (xpc_hb_check_timeout
- jiffies
),
277 atomic_read(&xpc_act_IRQ_rcvd
) - last_IRQ_count
);
280 /* checking of remote heartbeats is skewed by IRQ handling */
281 if (jiffies
>= xpc_hb_check_timeout
) {
282 dev_dbg(xpc_part
, "checking remote heartbeats\n");
283 xpc_check_remote_hb();
286 * We need to periodically recheck to ensure no
287 * IPI/AMO pairs have been missed. That check
288 * must always reset xpc_hb_check_timeout.
294 /* check for outstanding IRQs */
295 new_IRQ_count
= atomic_read(&xpc_act_IRQ_rcvd
);
296 if (last_IRQ_count
< new_IRQ_count
|| force_IRQ
!= 0) {
299 dev_dbg(xpc_part
, "found an IRQ to process; will be "
300 "resetting xpc_hb_check_timeout\n");
302 last_IRQ_count
+= xpc_identify_act_IRQ_sender();
303 if (last_IRQ_count
< new_IRQ_count
) {
304 /* retry once to help avoid missing AMO */
305 (void) xpc_identify_act_IRQ_sender();
307 last_IRQ_count
= new_IRQ_count
;
309 xpc_hb_check_timeout
= jiffies
+
310 (xpc_hb_check_interval
* HZ
);
313 /* wait for IRQ or timeout */
314 (void) wait_event_interruptible(xpc_act_IRQ_wq
,
315 (last_IRQ_count
< atomic_read(&xpc_act_IRQ_rcvd
) ||
316 jiffies
>= xpc_hb_check_timeout
||
317 (volatile int) xpc_exiting
));
320 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
323 /* mark this thread as having exited */
324 complete(&xpc_hb_checker_exited
);
330 * This thread will attempt to discover other partitions to activate
331 * based on info provided by SAL. This new thread is short lived and
332 * will exit once discovery is complete.
335 xpc_initiate_discovery(void *ignore
)
337 daemonize(XPC_DISCOVERY_THREAD_NAME
);
341 dev_dbg(xpc_part
, "discovery thread is exiting\n");
343 /* mark this thread as having exited */
344 complete(&xpc_discovery_exited
);
350 * Establish first contact with the remote partititon. This involves pulling
351 * the XPC per partition variables from the remote partition and waiting for
352 * the remote partition to pull ours.
354 static enum xpc_retval
355 xpc_make_first_contact(struct xpc_partition
*part
)
360 while ((ret
= xpc_pull_remote_vars_part(part
)) != xpcSuccess
) {
361 if (ret
!= xpcRetry
) {
362 XPC_DEACTIVATE_PARTITION(part
, ret
);
366 dev_dbg(xpc_chan
, "waiting to make first contact with "
367 "partition %d\n", XPC_PARTID(part
));
369 /* wait a 1/4 of a second or so */
370 (void) msleep_interruptible(250);
372 if (part
->act_state
== XPC_P_DEACTIVATING
) {
377 return xpc_mark_partition_active(part
);
382 * The first kthread assigned to a newly activated partition is the one
383 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
384 * that kthread until the partition is brought down, at which time that kthread
385 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
386 * that XPC has dismantled all communication infrastructure for the associated
387 * partition.) This kthread becomes the channel manager for that partition.
389 * Each active partition has a channel manager, who, besides connecting and
390 * disconnecting channels, will ensure that each of the partition's connected
391 * channels has the required number of assigned kthreads to get the work done.
394 xpc_channel_mgr(struct xpc_partition
*part
)
396 while (part
->act_state
!= XPC_P_DEACTIVATING
||
397 atomic_read(&part
->nchannels_active
) > 0 ||
398 !xpc_partition_disengaged(part
)) {
400 xpc_process_channel_activity(part
);
404 * Wait until we've been requested to activate kthreads or
405 * all of the channel's message queues have been torn down or
406 * a signal is pending.
408 * The channel_mgr_requests is set to 1 after being awakened,
409 * This is done to prevent the channel mgr from making one pass
410 * through the loop for each request, since he will
411 * be servicing all the requests in one pass. The reason it's
412 * set to 1 instead of 0 is so that other kthreads will know
413 * that the channel mgr is running and won't bother trying to
416 atomic_dec(&part
->channel_mgr_requests
);
417 (void) wait_event_interruptible(part
->channel_mgr_wq
,
418 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
419 (volatile u64
) part
->local_IPI_amo
!= 0 ||
420 ((volatile u8
) part
->act_state
==
421 XPC_P_DEACTIVATING
&&
422 atomic_read(&part
->nchannels_active
) == 0 &&
423 xpc_partition_disengaged(part
))));
424 atomic_set(&part
->channel_mgr_requests
, 1);
426 // >>> Does it need to wakeup periodically as well? In case we
427 // >>> miscalculated the #of kthreads to wakeup or create?
433 * When XPC HB determines that a partition has come up, it will create a new
434 * kthread and that kthread will call this function to attempt to set up the
435 * basic infrastructure used for Cross Partition Communication with the newly
438 * The kthread that was created by XPC HB and which setup the XPC
439 * infrastructure will remain assigned to the partition until the partition
440 * goes down. At which time the kthread will teardown the XPC infrastructure
443 * XPC HB will put the remote partition's XPC per partition specific variables
444 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
445 * calling xpc_partition_up().
448 xpc_partition_up(struct xpc_partition
*part
)
450 DBUG_ON(part
->channels
!= NULL
);
452 dev_dbg(xpc_chan
, "activating partition %d\n", XPC_PARTID(part
));
454 if (xpc_setup_infrastructure(part
) != xpcSuccess
) {
459 * The kthread that XPC HB called us with will become the
460 * channel manager for this partition. It will not return
461 * back to XPC HB until the partition's XPC infrastructure
462 * has been dismantled.
465 (void) xpc_part_ref(part
); /* this will always succeed */
467 if (xpc_make_first_contact(part
) == xpcSuccess
) {
468 xpc_channel_mgr(part
);
471 xpc_part_deref(part
);
473 xpc_teardown_infrastructure(part
);
478 xpc_activating(void *__partid
)
480 partid_t partid
= (u64
) __partid
;
481 struct xpc_partition
*part
= &xpc_partitions
[partid
];
482 unsigned long irq_flags
;
483 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
487 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
489 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
491 if (part
->act_state
== XPC_P_DEACTIVATING
) {
492 part
->act_state
= XPC_P_INACTIVE
;
493 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
494 part
->remote_rp_pa
= 0;
498 /* indicate the thread is activating */
499 DBUG_ON(part
->act_state
!= XPC_P_ACTIVATION_REQ
);
500 part
->act_state
= XPC_P_ACTIVATING
;
502 XPC_SET_REASON(part
, 0, 0);
503 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
505 dev_dbg(xpc_part
, "bringing partition %d up\n", partid
);
507 daemonize("xpc%02d", partid
);
510 * This thread needs to run at a realtime priority to prevent a
511 * significant performance degradation.
513 ret
= sched_setscheduler(current
, SCHED_FIFO
, ¶m
);
515 dev_warn(xpc_part
, "unable to set pid %d to a realtime "
516 "priority, ret=%d\n", current
->pid
, ret
);
519 /* allow this thread and its children to run on any CPU */
520 set_cpus_allowed(current
, CPU_MASK_ALL
);
523 * Register the remote partition's AMOs with SAL so it can handle
524 * and cleanup errors within that address range should the remote
525 * partition go down. We don't unregister this range because it is
526 * difficult to tell when outstanding writes to the remote partition
527 * are finished and thus when it is safe to unregister. This should
528 * not result in wasted space in the SAL xp_addr_region table because
529 * we should get the same page for remote_amos_page_pa after module
530 * reloads and system reboots.
532 if (sn_register_xp_addr_region(part
->remote_amos_page_pa
,
534 dev_warn(xpc_part
, "xpc_partition_up(%d) failed to register "
535 "xp_addr region\n", partid
);
537 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
538 part
->act_state
= XPC_P_INACTIVE
;
539 XPC_SET_REASON(part
, xpcPhysAddrRegFailed
, __LINE__
);
540 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
541 part
->remote_rp_pa
= 0;
545 xpc_allow_hb(partid
, xpc_vars
);
546 xpc_IPI_send_activated(part
);
550 * xpc_partition_up() holds this thread and marks this partition as
551 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
553 (void) xpc_partition_up(part
);
555 xpc_disallow_hb(partid
, xpc_vars
);
556 xpc_mark_partition_inactive(part
);
558 if (part
->reason
== xpcReactivating
) {
559 /* interrupting ourselves results in activating partition */
560 xpc_IPI_send_reactivate(part
);
568 xpc_activate_partition(struct xpc_partition
*part
)
570 partid_t partid
= XPC_PARTID(part
);
571 unsigned long irq_flags
;
575 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
577 DBUG_ON(part
->act_state
!= XPC_P_INACTIVE
);
579 part
->act_state
= XPC_P_ACTIVATION_REQ
;
580 XPC_SET_REASON(part
, xpcCloneKThread
, __LINE__
);
582 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
584 pid
= kernel_thread(xpc_activating
, (void *) ((u64
) partid
), 0);
586 if (unlikely(pid
<= 0)) {
587 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
588 part
->act_state
= XPC_P_INACTIVE
;
589 XPC_SET_REASON(part
, xpcCloneKThreadFailed
, __LINE__
);
590 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
596 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
597 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
598 * than one partition, we use an AMO_t structure per partition to indicate
599 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
600 * associated kthread to handle it.
602 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
603 * running on other partitions.
605 * Noteworthy Arguments:
607 * irq - Interrupt ReQuest number. NOT USED.
609 * dev_id - partid of IPI's potential sender.
611 * regs - processor's context before the processor entered
612 * interrupt code. NOT USED.
615 xpc_notify_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
617 partid_t partid
= (partid_t
) (u64
) dev_id
;
618 struct xpc_partition
*part
= &xpc_partitions
[partid
];
621 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
623 if (xpc_part_ref(part
)) {
624 xpc_check_for_channel_activity(part
);
626 xpc_part_deref(part
);
633 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
634 * because the write to their associated IPI amo completed after the IRQ/IPI
638 xpc_dropped_IPI_check(struct xpc_partition
*part
)
640 if (xpc_part_ref(part
)) {
641 xpc_check_for_channel_activity(part
);
643 part
->dropped_IPI_timer
.expires
= jiffies
+
644 XPC_P_DROPPED_IPI_WAIT
;
645 add_timer(&part
->dropped_IPI_timer
);
646 xpc_part_deref(part
);
652 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
654 int idle
= atomic_read(&ch
->kthreads_idle
);
655 int assigned
= atomic_read(&ch
->kthreads_assigned
);
659 DBUG_ON(needed
<= 0);
662 wakeup
= (needed
> idle
) ? idle
: needed
;
665 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
666 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
668 /* only wakeup the requested number of kthreads */
669 wake_up_nr(&ch
->idle_wq
, wakeup
);
676 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
677 needed
= ch
->kthreads_assigned_limit
- assigned
;
678 // >>>should never be less than 0
684 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
685 needed
, ch
->partid
, ch
->number
);
687 xpc_create_kthreads(ch
, needed
);
692 * This function is where XPC's kthreads wait for messages to deliver.
695 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
698 /* deliver messages to their intended recipients */
700 while ((volatile s64
) ch
->w_local_GP
.get
<
701 (volatile s64
) ch
->w_remote_GP
.put
&&
702 !((volatile u32
) ch
->flags
&
703 XPC_C_DISCONNECTING
)) {
707 if (atomic_inc_return(&ch
->kthreads_idle
) >
708 ch
->kthreads_idle_limit
) {
709 /* too many idle kthreads on this channel */
710 atomic_dec(&ch
->kthreads_idle
);
714 dev_dbg(xpc_chan
, "idle kthread calling "
715 "wait_event_interruptible_exclusive()\n");
717 (void) wait_event_interruptible_exclusive(ch
->idle_wq
,
718 ((volatile s64
) ch
->w_local_GP
.get
<
719 (volatile s64
) ch
->w_remote_GP
.put
||
720 ((volatile u32
) ch
->flags
&
721 XPC_C_DISCONNECTING
)));
723 atomic_dec(&ch
->kthreads_idle
);
725 } while (!((volatile u32
) ch
->flags
& XPC_C_DISCONNECTING
));
730 xpc_daemonize_kthread(void *args
)
732 partid_t partid
= XPC_UNPACK_ARG1(args
);
733 u16 ch_number
= XPC_UNPACK_ARG2(args
);
734 struct xpc_partition
*part
= &xpc_partitions
[partid
];
735 struct xpc_channel
*ch
;
737 unsigned long irq_flags
;
740 daemonize("xpc%02dc%d", partid
, ch_number
);
742 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
745 ch
= &part
->channels
[ch_number
];
747 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
749 /* let registerer know that connection has been established */
751 spin_lock_irqsave(&ch
->lock
, irq_flags
);
752 if (!(ch
->flags
& XPC_C_CONNECTEDCALLOUT
)) {
753 ch
->flags
|= XPC_C_CONNECTEDCALLOUT
;
754 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
756 xpc_connected_callout(ch
);
758 spin_lock_irqsave(&ch
->lock
, irq_flags
);
759 ch
->flags
|= XPC_C_CONNECTEDCALLOUT_MADE
;
760 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
763 * It is possible that while the callout was being
764 * made that the remote partition sent some messages.
765 * If that is the case, we may need to activate
766 * additional kthreads to help deliver them. We only
767 * need one less than total #of messages to deliver.
769 n_needed
= ch
->w_remote_GP
.put
- ch
->w_local_GP
.get
- 1;
771 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
772 xpc_activate_kthreads(ch
, n_needed
);
775 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
778 xpc_kthread_waitmsgs(part
, ch
);
781 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0) {
782 spin_lock_irqsave(&ch
->lock
, irq_flags
);
783 if ((ch
->flags
& XPC_C_CONNECTEDCALLOUT_MADE
) &&
784 !(ch
->flags
& XPC_C_DISCONNECTINGCALLOUT
)) {
785 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT
;
786 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
788 xpc_disconnect_callout(ch
, xpcDisconnecting
);
790 spin_lock_irqsave(&ch
->lock
, irq_flags
);
791 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT_MADE
;
793 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
794 if (atomic_dec_return(&part
->nchannels_engaged
) == 0) {
795 xpc_mark_partition_disengaged(part
);
796 xpc_IPI_send_disengage(part
);
801 xpc_msgqueue_deref(ch
);
803 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
806 xpc_part_deref(part
);
812 * For each partition that XPC has established communications with, there is
813 * a minimum of one kernel thread assigned to perform any operation that
814 * may potentially sleep or block (basically the callouts to the asynchronous
815 * functions registered via xpc_connect()).
817 * Additional kthreads are created and destroyed by XPC as the workload
820 * A kthread is assigned to one of the active channels that exists for a given
824 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
)
826 unsigned long irq_flags
;
828 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
829 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
832 while (needed
-- > 0) {
835 * The following is done on behalf of the newly created
836 * kthread. That kthread is responsible for doing the
837 * counterpart to the following before it exits.
839 (void) xpc_part_ref(part
);
840 xpc_msgqueue_ref(ch
);
841 if (atomic_inc_return(&ch
->kthreads_assigned
) == 1 &&
842 atomic_inc_return(&part
->nchannels_engaged
) == 1) {
843 xpc_mark_partition_engaged(part
);
846 pid
= kernel_thread(xpc_daemonize_kthread
, (void *) args
, 0);
848 /* the fork failed */
849 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
850 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
851 xpc_mark_partition_disengaged(part
);
852 xpc_IPI_send_disengage(part
);
854 xpc_msgqueue_deref(ch
);
855 xpc_part_deref(part
);
857 if (atomic_read(&ch
->kthreads_assigned
) <
858 ch
->kthreads_idle_limit
) {
860 * Flag this as an error only if we have an
861 * insufficient #of kthreads for the channel
864 * No xpc_msgqueue_ref() is needed here since
865 * the channel mgr is doing this.
867 spin_lock_irqsave(&ch
->lock
, irq_flags
);
868 XPC_DISCONNECT_CHANNEL(ch
, xpcLackOfResources
,
870 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
875 ch
->kthreads_created
++; // >>> temporary debug only!!!
881 xpc_disconnect_wait(int ch_number
)
883 unsigned long irq_flags
;
885 struct xpc_partition
*part
;
886 struct xpc_channel
*ch
;
887 int wakeup_channel_mgr
;
890 /* now wait for all callouts to the caller's function to cease */
891 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
892 part
= &xpc_partitions
[partid
];
894 if (!xpc_part_ref(part
)) {
898 ch
= &part
->channels
[ch_number
];
900 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
901 xpc_part_deref(part
);
905 wait_for_completion(&ch
->wdisconnect_wait
);
907 spin_lock_irqsave(&ch
->lock
, irq_flags
);
908 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
909 wakeup_channel_mgr
= 0;
911 if (ch
->delayed_IPI_flags
) {
912 if (part
->act_state
!= XPC_P_DEACTIVATING
) {
913 spin_lock(&part
->IPI_lock
);
914 XPC_SET_IPI_FLAGS(part
->local_IPI_amo
,
915 ch
->number
, ch
->delayed_IPI_flags
);
916 spin_unlock(&part
->IPI_lock
);
917 wakeup_channel_mgr
= 1;
919 ch
->delayed_IPI_flags
= 0;
922 ch
->flags
&= ~XPC_C_WDISCONNECT
;
923 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
925 if (wakeup_channel_mgr
) {
926 xpc_wakeup_channel_mgr(part
);
929 xpc_part_deref(part
);
935 xpc_do_exit(enum xpc_retval reason
)
938 int active_part_count
, printed_waiting_msg
= 0;
939 struct xpc_partition
*part
;
940 unsigned long printmsg_time
, disengage_request_timeout
= 0;
943 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
944 DBUG_ON(xpc_exiting
== 1);
947 * Let the heartbeat checker thread and the discovery thread
948 * (if one is running) know that they should exit. Also wake up
949 * the heartbeat checker thread in case it's sleeping.
952 wake_up_interruptible(&xpc_act_IRQ_wq
);
954 /* ignore all incoming interrupts */
955 free_irq(SGI_XPC_ACTIVATE
, NULL
);
957 /* wait for the discovery thread to exit */
958 wait_for_completion(&xpc_discovery_exited
);
960 /* wait for the heartbeat checker thread to exit */
961 wait_for_completion(&xpc_hb_checker_exited
);
964 /* sleep for a 1/3 of a second or so */
965 (void) msleep_interruptible(300);
968 /* wait for all partitions to become inactive */
970 printmsg_time
= jiffies
+ (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
971 xpc_disengage_request_timedout
= 0;
974 active_part_count
= 0;
976 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
977 part
= &xpc_partitions
[partid
];
979 if (xpc_partition_disengaged(part
) &&
980 part
->act_state
== XPC_P_INACTIVE
) {
986 XPC_DEACTIVATE_PARTITION(part
, reason
);
988 if (part
->disengage_request_timeout
>
989 disengage_request_timeout
) {
990 disengage_request_timeout
=
991 part
->disengage_request_timeout
;
995 if (xpc_partition_engaged(-1UL)) {
996 if (time_after(jiffies
, printmsg_time
)) {
997 dev_info(xpc_part
, "waiting for remote "
998 "partitions to disengage, timeout in "
1000 (disengage_request_timeout
- jiffies
)
1002 printmsg_time
= jiffies
+
1003 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
1004 printed_waiting_msg
= 1;
1007 } else if (active_part_count
> 0) {
1008 if (printed_waiting_msg
) {
1009 dev_info(xpc_part
, "waiting for local partition"
1011 printed_waiting_msg
= 0;
1015 if (!xpc_disengage_request_timedout
) {
1016 dev_info(xpc_part
, "all partitions have "
1022 /* sleep for a 1/3 of a second or so */
1023 (void) msleep_interruptible(300);
1027 DBUG_ON(xpc_partition_engaged(-1UL));
1030 /* indicate to others that our reserved page is uninitialized */
1031 xpc_rsvd_page
->vars_pa
= 0;
1033 /* now it's time to eliminate our heartbeat */
1034 del_timer_sync(&xpc_hb_timer
);
1035 DBUG_ON(xpc_vars
->heartbeating_to_mask
!= 0);
1037 if (reason
== xpcUnloading
) {
1038 /* take ourselves off of the reboot_notifier_list */
1039 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1041 /* take ourselves off of the die_notifier list */
1042 (void) unregister_die_notifier(&xpc_die_notifier
);
1045 /* close down protections for IPI operations */
1046 xpc_restrict_IPI_ops();
1049 /* clear the interface to XPC's functions */
1050 xpc_clear_interface();
1053 unregister_sysctl_table(xpc_sysctl
);
1056 kfree(xpc_remote_copy_buffer_base
);
1061 * This function is called when the system is being rebooted.
1064 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1066 enum xpc_retval reason
;
1071 reason
= xpcSystemReboot
;
1074 reason
= xpcSystemHalt
;
1077 reason
= xpcSystemPoweroff
;
1080 reason
= xpcSystemGoingDown
;
1083 xpc_do_exit(reason
);
1089 * Notify other partitions to disengage from all references to our memory.
1092 xpc_die_disengage(void)
1094 struct xpc_partition
*part
;
1096 unsigned long engaged
;
1097 long time
, printmsg_time
, disengage_request_timeout
;
1100 /* keep xpc_hb_checker thread from doing anything (just in case) */
1103 xpc_vars
->heartbeating_to_mask
= 0; /* indicate we're deactivated */
1105 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1106 part
= &xpc_partitions
[partid
];
1108 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part
->
1109 remote_vars_version
)) {
1111 /* just in case it was left set by an earlier XPC */
1112 xpc_clear_partition_engaged(1UL << partid
);
1116 if (xpc_partition_engaged(1UL << partid
) ||
1117 part
->act_state
!= XPC_P_INACTIVE
) {
1118 xpc_request_partition_disengage(part
);
1119 xpc_mark_partition_disengaged(part
);
1120 xpc_IPI_send_disengage(part
);
1125 printmsg_time
= time
+
1126 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* sn_rtc_cycles_per_second
);
1127 disengage_request_timeout
= time
+
1128 (xpc_disengage_request_timelimit
* sn_rtc_cycles_per_second
);
1130 /* wait for all other partitions to disengage from us */
1133 engaged
= xpc_partition_engaged(-1UL);
1135 dev_info(xpc_part
, "all partitions have disengaged\n");
1140 if (time
>= disengage_request_timeout
) {
1141 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1142 if (engaged
& (1UL << partid
)) {
1143 dev_info(xpc_part
, "disengage from "
1144 "remote partition %d timed "
1151 if (time
>= printmsg_time
) {
1152 dev_info(xpc_part
, "waiting for remote partitions to "
1153 "disengage, timeout in %ld seconds\n",
1154 (disengage_request_timeout
- time
) /
1155 sn_rtc_cycles_per_second
);
1156 printmsg_time
= time
+
1157 (XPC_DISENGAGE_PRINTMSG_INTERVAL
*
1158 sn_rtc_cycles_per_second
);
1165 * This function is called when the system is being restarted or halted due
1166 * to some sort of system failure. If this is the case we need to notify the
1167 * other partitions to disengage from all references to our memory.
1168 * This function can also be called when our heartbeater could be offlined
1169 * for a time. In this case we need to notify other partitions to not worry
1170 * about the lack of a heartbeat.
1173 xpc_system_die(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1176 case DIE_MACHINE_RESTART
:
1177 case DIE_MACHINE_HALT
:
1178 xpc_die_disengage();
1181 case DIE_KDEBUG_ENTER
:
1182 /* Should lack of heartbeat be ignored by other partitions? */
1183 if (!xpc_kdebug_ignore
) {
1187 case DIE_MCA_MONARCH_ENTER
:
1188 case DIE_INIT_MONARCH_ENTER
:
1189 xpc_vars
->heartbeat
++;
1190 xpc_vars
->heartbeat_offline
= 1;
1193 case DIE_KDEBUG_LEAVE
:
1194 /* Is lack of heartbeat being ignored by other partitions? */
1195 if (!xpc_kdebug_ignore
) {
1199 case DIE_MCA_MONARCH_LEAVE
:
1200 case DIE_INIT_MONARCH_LEAVE
:
1201 xpc_vars
->heartbeat
++;
1202 xpc_vars
->heartbeat_offline
= 0;
1215 struct xpc_partition
*part
;
1220 if (!ia64_platform_is("sn2")) {
1225 buf_size
= max(XPC_RP_VARS_SIZE
,
1226 XPC_RP_HEADER_SIZE
+ XP_NASID_MASK_BYTES
);
1227 xpc_remote_copy_buffer
= xpc_kmalloc_cacheline_aligned(buf_size
,
1228 GFP_KERNEL
, &xpc_remote_copy_buffer_base
);
1229 if (xpc_remote_copy_buffer
== NULL
)
1232 snprintf(xpc_part
->bus_id
, BUS_ID_SIZE
, "part");
1233 snprintf(xpc_chan
->bus_id
, BUS_ID_SIZE
, "chan");
1235 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
, 1);
1238 * The first few fields of each entry of xpc_partitions[] need to
1239 * be initialized now so that calls to xpc_connect() and
1240 * xpc_disconnect() can be made prior to the activation of any remote
1241 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1242 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1243 * PARTITION HAS BEEN ACTIVATED.
1245 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1246 part
= &xpc_partitions
[partid
];
1248 DBUG_ON((u64
) part
!= L1_CACHE_ALIGN((u64
) part
));
1250 part
->act_IRQ_rcvd
= 0;
1251 spin_lock_init(&part
->act_lock
);
1252 part
->act_state
= XPC_P_INACTIVE
;
1253 XPC_SET_REASON(part
, 0, 0);
1255 init_timer(&part
->disengage_request_timer
);
1256 part
->disengage_request_timer
.function
=
1257 xpc_timeout_partition_disengage_request
;
1258 part
->disengage_request_timer
.data
= (unsigned long) part
;
1260 part
->setup_state
= XPC_P_UNSET
;
1261 init_waitqueue_head(&part
->teardown_wq
);
1262 atomic_set(&part
->references
, 0);
1266 * Open up protections for IPI operations (and AMO operations on
1267 * Shub 1.1 systems).
1269 xpc_allow_IPI_ops();
1272 * Interrupts being processed will increment this atomic variable and
1273 * awaken the heartbeat thread which will process the interrupts.
1275 atomic_set(&xpc_act_IRQ_rcvd
, 0);
1278 * This is safe to do before the xpc_hb_checker thread has started
1279 * because the handler releases a wait queue. If an interrupt is
1280 * received before the thread is waiting, it will not go to sleep,
1281 * but rather immediately process the interrupt.
1283 ret
= request_irq(SGI_XPC_ACTIVATE
, xpc_act_IRQ_handler
, 0,
1286 dev_err(xpc_part
, "can't register ACTIVATE IRQ handler, "
1287 "errno=%d\n", -ret
);
1289 xpc_restrict_IPI_ops();
1292 unregister_sysctl_table(xpc_sysctl
);
1295 kfree(xpc_remote_copy_buffer_base
);
1300 * Fill the partition reserved page with the information needed by
1301 * other partitions to discover we are alive and establish initial
1304 xpc_rsvd_page
= xpc_rsvd_page_init();
1305 if (xpc_rsvd_page
== NULL
) {
1306 dev_err(xpc_part
, "could not setup our reserved page\n");
1308 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1309 xpc_restrict_IPI_ops();
1312 unregister_sysctl_table(xpc_sysctl
);
1315 kfree(xpc_remote_copy_buffer_base
);
1320 /* add ourselves to the reboot_notifier_list */
1321 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1323 dev_warn(xpc_part
, "can't register reboot notifier\n");
1326 /* add ourselves to the die_notifier list (i.e., ia64die_chain) */
1327 ret
= register_die_notifier(&xpc_die_notifier
);
1329 dev_warn(xpc_part
, "can't register die notifier\n");
1334 * Set the beating to other partitions into motion. This is
1335 * the last requirement for other partitions' discovery to
1336 * initiate communications with us.
1338 init_timer(&xpc_hb_timer
);
1339 xpc_hb_timer
.function
= xpc_hb_beater
;
1344 * The real work-horse behind xpc. This processes incoming
1345 * interrupts and monitors remote heartbeats.
1347 pid
= kernel_thread(xpc_hb_checker
, NULL
, 0);
1349 dev_err(xpc_part
, "failed while forking hb check thread\n");
1351 /* indicate to others that our reserved page is uninitialized */
1352 xpc_rsvd_page
->vars_pa
= 0;
1354 /* take ourselves off of the reboot_notifier_list */
1355 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1357 /* take ourselves off of the die_notifier list */
1358 (void) unregister_die_notifier(&xpc_die_notifier
);
1360 del_timer_sync(&xpc_hb_timer
);
1361 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1362 xpc_restrict_IPI_ops();
1365 unregister_sysctl_table(xpc_sysctl
);
1368 kfree(xpc_remote_copy_buffer_base
);
1374 * Startup a thread that will attempt to discover other partitions to
1375 * activate based on info provided by SAL. This new thread is short
1376 * lived and will exit once discovery is complete.
1378 pid
= kernel_thread(xpc_initiate_discovery
, NULL
, 0);
1380 dev_err(xpc_part
, "failed while forking discovery thread\n");
1382 /* mark this new thread as a non-starter */
1383 complete(&xpc_discovery_exited
);
1385 xpc_do_exit(xpcUnloading
);
1390 /* set the interface to point at XPC's functions */
1391 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1392 xpc_initiate_allocate
, xpc_initiate_send
,
1393 xpc_initiate_send_notify
, xpc_initiate_received
,
1394 xpc_initiate_partid_to_nasids
);
1398 module_init(xpc_init
);
1404 xpc_do_exit(xpcUnloading
);
1406 module_exit(xpc_exit
);
1409 MODULE_AUTHOR("Silicon Graphics, Inc.");
1410 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1411 MODULE_LICENSE("GPL");
1413 module_param(xpc_hb_interval
, int, 0);
1414 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1415 "heartbeat increments.");
1417 module_param(xpc_hb_check_interval
, int, 0);
1418 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1419 "heartbeat checks.");
1421 module_param(xpc_disengage_request_timelimit
, int, 0);
1422 MODULE_PARM_DESC(xpc_disengage_request_timelimit
, "Number of seconds to wait "
1423 "for disengage request to complete.");
1425 module_param(xpc_kdebug_ignore
, int, 0);
1426 MODULE_PARM_DESC(xpc_kdebug_ignore
, "Should lack of heartbeat be ignored by "
1427 "other partitions when dropping into kdebug.");