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 int xpc_disengage_request_timelimit
= XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT
;
95 static int xpc_disengage_request_min_timelimit
= 0;
96 static int xpc_disengage_request_max_timelimit
= 120;
98 static ctl_table xpc_sys_xpc_hb_dir
[] = {
106 &proc_dointvec_minmax
,
109 &xpc_hb_min_interval
,
115 &xpc_hb_check_interval
,
119 &proc_dointvec_minmax
,
122 &xpc_hb_check_min_interval
,
123 &xpc_hb_check_max_interval
127 static ctl_table xpc_sys_xpc_dir
[] = {
138 "disengage_request_timelimit",
139 &xpc_disengage_request_timelimit
,
143 &proc_dointvec_minmax
,
146 &xpc_disengage_request_min_timelimit
,
147 &xpc_disengage_request_max_timelimit
151 static ctl_table xpc_sys_dir
[] = {
162 static struct ctl_table_header
*xpc_sysctl
;
165 /* #of IRQs received */
166 static atomic_t xpc_act_IRQ_rcvd
;
168 /* IRQ handler notifies this wait queue on receipt of an IRQ */
169 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq
);
171 static unsigned long xpc_hb_check_timeout
;
173 /* notification that the xpc_hb_checker thread has exited */
174 static DECLARE_MUTEX_LOCKED(xpc_hb_checker_exited
);
176 /* notification that the xpc_discovery thread has exited */
177 static DECLARE_MUTEX_LOCKED(xpc_discovery_exited
);
180 static struct timer_list xpc_hb_timer
;
183 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
186 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
187 static struct notifier_block xpc_reboot_notifier
= {
188 .notifier_call
= xpc_system_reboot
,
193 * Timer function to enforce the timelimit on the partition disengage request.
196 xpc_timeout_partition_disengage_request(unsigned long data
)
198 struct xpc_partition
*part
= (struct xpc_partition
*) data
;
201 DBUG_ON(jiffies
< part
->disengage_request_timeout
);
203 (void) xpc_partition_disengaged(part
);
205 DBUG_ON(part
->disengage_request_timeout
!= 0);
206 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part
)) != 0);
211 * Notify the heartbeat check thread that an IRQ has been received.
214 xpc_act_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
216 atomic_inc(&xpc_act_IRQ_rcvd
);
217 wake_up_interruptible(&xpc_act_IRQ_wq
);
223 * Timer to produce the heartbeat. The timer structures function is
224 * already set when this is initially called. A tunable is used to
225 * specify when the next timeout should occur.
228 xpc_hb_beater(unsigned long dummy
)
230 xpc_vars
->heartbeat
++;
232 if (jiffies
>= xpc_hb_check_timeout
) {
233 wake_up_interruptible(&xpc_act_IRQ_wq
);
236 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
237 add_timer(&xpc_hb_timer
);
242 * This thread is responsible for nearly all of the partition
243 * activation/deactivation.
246 xpc_hb_checker(void *ignore
)
248 int last_IRQ_count
= 0;
253 /* this thread was marked active by xpc_hb_init() */
255 daemonize(XPC_HB_CHECK_THREAD_NAME
);
257 set_cpus_allowed(current
, cpumask_of_cpu(XPC_HB_CHECK_CPU
));
259 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
261 while (!(volatile int) xpc_exiting
) {
263 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
265 (int) (xpc_hb_check_timeout
- jiffies
),
266 atomic_read(&xpc_act_IRQ_rcvd
) - last_IRQ_count
);
269 /* checking of remote heartbeats is skewed by IRQ handling */
270 if (jiffies
>= xpc_hb_check_timeout
) {
271 dev_dbg(xpc_part
, "checking remote heartbeats\n");
272 xpc_check_remote_hb();
275 * We need to periodically recheck to ensure no
276 * IPI/AMO pairs have been missed. That check
277 * must always reset xpc_hb_check_timeout.
283 /* check for outstanding IRQs */
284 new_IRQ_count
= atomic_read(&xpc_act_IRQ_rcvd
);
285 if (last_IRQ_count
< new_IRQ_count
|| force_IRQ
!= 0) {
288 dev_dbg(xpc_part
, "found an IRQ to process; will be "
289 "resetting xpc_hb_check_timeout\n");
291 last_IRQ_count
+= xpc_identify_act_IRQ_sender();
292 if (last_IRQ_count
< new_IRQ_count
) {
293 /* retry once to help avoid missing AMO */
294 (void) xpc_identify_act_IRQ_sender();
296 last_IRQ_count
= new_IRQ_count
;
298 xpc_hb_check_timeout
= jiffies
+
299 (xpc_hb_check_interval
* HZ
);
302 /* wait for IRQ or timeout */
303 (void) wait_event_interruptible(xpc_act_IRQ_wq
,
304 (last_IRQ_count
< atomic_read(&xpc_act_IRQ_rcvd
) ||
305 jiffies
>= xpc_hb_check_timeout
||
306 (volatile int) xpc_exiting
));
309 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
312 /* mark this thread as having exited */
313 up(&xpc_hb_checker_exited
);
319 * This thread will attempt to discover other partitions to activate
320 * based on info provided by SAL. This new thread is short lived and
321 * will exit once discovery is complete.
324 xpc_initiate_discovery(void *ignore
)
326 daemonize(XPC_DISCOVERY_THREAD_NAME
);
330 dev_dbg(xpc_part
, "discovery thread is exiting\n");
332 /* mark this thread as having exited */
333 up(&xpc_discovery_exited
);
339 * Establish first contact with the remote partititon. This involves pulling
340 * the XPC per partition variables from the remote partition and waiting for
341 * the remote partition to pull ours.
343 static enum xpc_retval
344 xpc_make_first_contact(struct xpc_partition
*part
)
349 while ((ret
= xpc_pull_remote_vars_part(part
)) != xpcSuccess
) {
350 if (ret
!= xpcRetry
) {
351 XPC_DEACTIVATE_PARTITION(part
, ret
);
355 dev_dbg(xpc_chan
, "waiting to make first contact with "
356 "partition %d\n", XPC_PARTID(part
));
358 /* wait a 1/4 of a second or so */
359 (void) msleep_interruptible(250);
361 if (part
->act_state
== XPC_P_DEACTIVATING
) {
366 return xpc_mark_partition_active(part
);
371 * The first kthread assigned to a newly activated partition is the one
372 * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
373 * that kthread until the partition is brought down, at which time that kthread
374 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
375 * that XPC has dismantled all communication infrastructure for the associated
376 * partition.) This kthread becomes the channel manager for that partition.
378 * Each active partition has a channel manager, who, besides connecting and
379 * disconnecting channels, will ensure that each of the partition's connected
380 * channels has the required number of assigned kthreads to get the work done.
383 xpc_channel_mgr(struct xpc_partition
*part
)
385 while (part
->act_state
!= XPC_P_DEACTIVATING
||
386 atomic_read(&part
->nchannels_active
) > 0 ||
387 !xpc_partition_disengaged(part
)) {
389 xpc_process_channel_activity(part
);
393 * Wait until we've been requested to activate kthreads or
394 * all of the channel's message queues have been torn down or
395 * a signal is pending.
397 * The channel_mgr_requests is set to 1 after being awakened,
398 * This is done to prevent the channel mgr from making one pass
399 * through the loop for each request, since he will
400 * be servicing all the requests in one pass. The reason it's
401 * set to 1 instead of 0 is so that other kthreads will know
402 * that the channel mgr is running and won't bother trying to
405 atomic_dec(&part
->channel_mgr_requests
);
406 (void) wait_event_interruptible(part
->channel_mgr_wq
,
407 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
408 (volatile u64
) part
->local_IPI_amo
!= 0 ||
409 ((volatile u8
) part
->act_state
==
410 XPC_P_DEACTIVATING
&&
411 atomic_read(&part
->nchannels_active
) == 0 &&
412 xpc_partition_disengaged(part
))));
413 atomic_set(&part
->channel_mgr_requests
, 1);
415 // >>> Does it need to wakeup periodically as well? In case we
416 // >>> miscalculated the #of kthreads to wakeup or create?
422 * When XPC HB determines that a partition has come up, it will create a new
423 * kthread and that kthread will call this function to attempt to set up the
424 * basic infrastructure used for Cross Partition Communication with the newly
427 * The kthread that was created by XPC HB and which setup the XPC
428 * infrastructure will remain assigned to the partition until the partition
429 * goes down. At which time the kthread will teardown the XPC infrastructure
432 * XPC HB will put the remote partition's XPC per partition specific variables
433 * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
434 * calling xpc_partition_up().
437 xpc_partition_up(struct xpc_partition
*part
)
439 DBUG_ON(part
->channels
!= NULL
);
441 dev_dbg(xpc_chan
, "activating partition %d\n", XPC_PARTID(part
));
443 if (xpc_setup_infrastructure(part
) != xpcSuccess
) {
448 * The kthread that XPC HB called us with will become the
449 * channel manager for this partition. It will not return
450 * back to XPC HB until the partition's XPC infrastructure
451 * has been dismantled.
454 (void) xpc_part_ref(part
); /* this will always succeed */
456 if (xpc_make_first_contact(part
) == xpcSuccess
) {
457 xpc_channel_mgr(part
);
460 xpc_part_deref(part
);
462 xpc_teardown_infrastructure(part
);
467 xpc_activating(void *__partid
)
469 partid_t partid
= (u64
) __partid
;
470 struct xpc_partition
*part
= &xpc_partitions
[partid
];
471 unsigned long irq_flags
;
472 struct sched_param param
= { sched_priority
: MAX_RT_PRIO
- 1 };
476 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
478 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
480 if (part
->act_state
== XPC_P_DEACTIVATING
) {
481 part
->act_state
= XPC_P_INACTIVE
;
482 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
483 part
->remote_rp_pa
= 0;
487 /* indicate the thread is activating */
488 DBUG_ON(part
->act_state
!= XPC_P_ACTIVATION_REQ
);
489 part
->act_state
= XPC_P_ACTIVATING
;
491 XPC_SET_REASON(part
, 0, 0);
492 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
494 dev_dbg(xpc_part
, "bringing partition %d up\n", partid
);
496 daemonize("xpc%02d", partid
);
499 * This thread needs to run at a realtime priority to prevent a
500 * significant performance degradation.
502 ret
= sched_setscheduler(current
, SCHED_FIFO
, ¶m
);
504 dev_warn(xpc_part
, "unable to set pid %d to a realtime "
505 "priority, ret=%d\n", current
->pid
, ret
);
508 /* allow this thread and its children to run on any CPU */
509 set_cpus_allowed(current
, CPU_MASK_ALL
);
512 * Register the remote partition's AMOs with SAL so it can handle
513 * and cleanup errors within that address range should the remote
514 * partition go down. We don't unregister this range because it is
515 * difficult to tell when outstanding writes to the remote partition
516 * are finished and thus when it is safe to unregister. This should
517 * not result in wasted space in the SAL xp_addr_region table because
518 * we should get the same page for remote_amos_page_pa after module
519 * reloads and system reboots.
521 if (sn_register_xp_addr_region(part
->remote_amos_page_pa
,
523 dev_warn(xpc_part
, "xpc_partition_up(%d) failed to register "
524 "xp_addr region\n", partid
);
526 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
527 part
->act_state
= XPC_P_INACTIVE
;
528 XPC_SET_REASON(part
, xpcPhysAddrRegFailed
, __LINE__
);
529 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
530 part
->remote_rp_pa
= 0;
534 xpc_allow_hb(partid
, xpc_vars
);
535 xpc_IPI_send_activated(part
);
539 * xpc_partition_up() holds this thread and marks this partition as
540 * XPC_P_ACTIVE by calling xpc_hb_mark_active().
542 (void) xpc_partition_up(part
);
544 xpc_disallow_hb(partid
, xpc_vars
);
545 xpc_mark_partition_inactive(part
);
547 if (part
->reason
== xpcReactivating
) {
548 /* interrupting ourselves results in activating partition */
549 xpc_IPI_send_reactivate(part
);
557 xpc_activate_partition(struct xpc_partition
*part
)
559 partid_t partid
= XPC_PARTID(part
);
560 unsigned long irq_flags
;
564 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
566 pid
= kernel_thread(xpc_activating
, (void *) ((u64
) partid
), 0);
568 DBUG_ON(part
->act_state
!= XPC_P_INACTIVE
);
571 part
->act_state
= XPC_P_ACTIVATION_REQ
;
572 XPC_SET_REASON(part
, xpcCloneKThread
, __LINE__
);
574 XPC_SET_REASON(part
, xpcCloneKThreadFailed
, __LINE__
);
577 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
582 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
583 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
584 * than one partition, we use an AMO_t structure per partition to indicate
585 * whether a partition has sent an IPI or not. >>> If it has, then wake up the
586 * associated kthread to handle it.
588 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
589 * running on other partitions.
591 * Noteworthy Arguments:
593 * irq - Interrupt ReQuest number. NOT USED.
595 * dev_id - partid of IPI's potential sender.
597 * regs - processor's context before the processor entered
598 * interrupt code. NOT USED.
601 xpc_notify_IRQ_handler(int irq
, void *dev_id
, struct pt_regs
*regs
)
603 partid_t partid
= (partid_t
) (u64
) dev_id
;
604 struct xpc_partition
*part
= &xpc_partitions
[partid
];
607 DBUG_ON(partid
<= 0 || partid
>= XP_MAX_PARTITIONS
);
609 if (xpc_part_ref(part
)) {
610 xpc_check_for_channel_activity(part
);
612 xpc_part_deref(part
);
619 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
620 * because the write to their associated IPI amo completed after the IRQ/IPI
624 xpc_dropped_IPI_check(struct xpc_partition
*part
)
626 if (xpc_part_ref(part
)) {
627 xpc_check_for_channel_activity(part
);
629 part
->dropped_IPI_timer
.expires
= jiffies
+
630 XPC_P_DROPPED_IPI_WAIT
;
631 add_timer(&part
->dropped_IPI_timer
);
632 xpc_part_deref(part
);
638 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
640 int idle
= atomic_read(&ch
->kthreads_idle
);
641 int assigned
= atomic_read(&ch
->kthreads_assigned
);
645 DBUG_ON(needed
<= 0);
648 wakeup
= (needed
> idle
) ? idle
: needed
;
651 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
652 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
654 /* only wakeup the requested number of kthreads */
655 wake_up_nr(&ch
->idle_wq
, wakeup
);
662 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
663 needed
= ch
->kthreads_assigned_limit
- assigned
;
664 // >>>should never be less than 0
670 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
671 needed
, ch
->partid
, ch
->number
);
673 xpc_create_kthreads(ch
, needed
);
678 * This function is where XPC's kthreads wait for messages to deliver.
681 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
684 /* deliver messages to their intended recipients */
686 while ((volatile s64
) ch
->w_local_GP
.get
<
687 (volatile s64
) ch
->w_remote_GP
.put
&&
688 !((volatile u32
) ch
->flags
&
689 XPC_C_DISCONNECTING
)) {
693 if (atomic_inc_return(&ch
->kthreads_idle
) >
694 ch
->kthreads_idle_limit
) {
695 /* too many idle kthreads on this channel */
696 atomic_dec(&ch
->kthreads_idle
);
700 dev_dbg(xpc_chan
, "idle kthread calling "
701 "wait_event_interruptible_exclusive()\n");
703 (void) wait_event_interruptible_exclusive(ch
->idle_wq
,
704 ((volatile s64
) ch
->w_local_GP
.get
<
705 (volatile s64
) ch
->w_remote_GP
.put
||
706 ((volatile u32
) ch
->flags
&
707 XPC_C_DISCONNECTING
)));
709 atomic_dec(&ch
->kthreads_idle
);
711 } while (!((volatile u32
) ch
->flags
& XPC_C_DISCONNECTING
));
716 xpc_daemonize_kthread(void *args
)
718 partid_t partid
= XPC_UNPACK_ARG1(args
);
719 u16 ch_number
= XPC_UNPACK_ARG2(args
);
720 struct xpc_partition
*part
= &xpc_partitions
[partid
];
721 struct xpc_channel
*ch
;
723 unsigned long irq_flags
;
726 daemonize("xpc%02dc%d", partid
, ch_number
);
728 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
731 ch
= &part
->channels
[ch_number
];
733 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
735 /* let registerer know that connection has been established */
737 spin_lock_irqsave(&ch
->lock
, irq_flags
);
738 if (!(ch
->flags
& XPC_C_CONNECTCALLOUT
)) {
739 ch
->flags
|= XPC_C_CONNECTCALLOUT
;
740 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
742 xpc_connected_callout(ch
);
745 * It is possible that while the callout was being
746 * made that the remote partition sent some messages.
747 * If that is the case, we may need to activate
748 * additional kthreads to help deliver them. We only
749 * need one less than total #of messages to deliver.
751 n_needed
= ch
->w_remote_GP
.put
- ch
->w_local_GP
.get
- 1;
753 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
754 xpc_activate_kthreads(ch
, n_needed
);
757 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
760 xpc_kthread_waitmsgs(part
, ch
);
763 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0) {
764 spin_lock_irqsave(&ch
->lock
, irq_flags
);
765 if ((ch
->flags
& XPC_C_CONNECTCALLOUT
) &&
766 !(ch
->flags
& XPC_C_DISCONNECTCALLOUT
)) {
767 ch
->flags
|= XPC_C_DISCONNECTCALLOUT
;
768 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
770 xpc_disconnecting_callout(ch
);
772 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
774 if (atomic_dec_return(&part
->nchannels_engaged
) == 0) {
775 xpc_mark_partition_disengaged(part
);
776 xpc_IPI_send_disengage(part
);
781 xpc_msgqueue_deref(ch
);
783 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
786 xpc_part_deref(part
);
792 * For each partition that XPC has established communications with, there is
793 * a minimum of one kernel thread assigned to perform any operation that
794 * may potentially sleep or block (basically the callouts to the asynchronous
795 * functions registered via xpc_connect()).
797 * Additional kthreads are created and destroyed by XPC as the workload
800 * A kthread is assigned to one of the active channels that exists for a given
804 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
)
806 unsigned long irq_flags
;
808 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
809 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
812 while (needed
-- > 0) {
815 * The following is done on behalf of the newly created
816 * kthread. That kthread is responsible for doing the
817 * counterpart to the following before it exits.
819 (void) xpc_part_ref(part
);
820 xpc_msgqueue_ref(ch
);
821 if (atomic_inc_return(&ch
->kthreads_assigned
) == 1 &&
822 atomic_inc_return(&part
->nchannels_engaged
) == 1) {
823 xpc_mark_partition_engaged(part
);
826 pid
= kernel_thread(xpc_daemonize_kthread
, (void *) args
, 0);
828 /* the fork failed */
829 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
830 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
831 xpc_mark_partition_disengaged(part
);
832 xpc_IPI_send_disengage(part
);
834 xpc_msgqueue_deref(ch
);
835 xpc_part_deref(part
);
837 if (atomic_read(&ch
->kthreads_assigned
) <
838 ch
->kthreads_idle_limit
) {
840 * Flag this as an error only if we have an
841 * insufficient #of kthreads for the channel
844 * No xpc_msgqueue_ref() is needed here since
845 * the channel mgr is doing this.
847 spin_lock_irqsave(&ch
->lock
, irq_flags
);
848 XPC_DISCONNECT_CHANNEL(ch
, xpcLackOfResources
,
850 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
855 ch
->kthreads_created
++; // >>> temporary debug only!!!
861 xpc_disconnect_wait(int ch_number
)
863 unsigned long irq_flags
;
865 struct xpc_partition
*part
;
866 struct xpc_channel
*ch
;
867 int wakeup_channel_mgr
;
870 /* now wait for all callouts to the caller's function to cease */
871 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
872 part
= &xpc_partitions
[partid
];
874 if (!xpc_part_ref(part
)) {
878 ch
= &part
->channels
[ch_number
];
880 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
881 xpc_part_deref(part
);
885 (void) down(&ch
->wdisconnect_sema
);
887 spin_lock_irqsave(&ch
->lock
, irq_flags
);
888 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
889 wakeup_channel_mgr
= 0;
891 if (ch
->delayed_IPI_flags
) {
892 if (part
->act_state
!= XPC_P_DEACTIVATING
) {
893 spin_lock(&part
->IPI_lock
);
894 XPC_SET_IPI_FLAGS(part
->local_IPI_amo
,
895 ch
->number
, ch
->delayed_IPI_flags
);
896 spin_unlock(&part
->IPI_lock
);
897 wakeup_channel_mgr
= 1;
899 ch
->delayed_IPI_flags
= 0;
902 ch
->flags
&= ~XPC_C_WDISCONNECT
;
903 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
905 if (wakeup_channel_mgr
) {
906 xpc_wakeup_channel_mgr(part
);
909 xpc_part_deref(part
);
915 xpc_do_exit(enum xpc_retval reason
)
918 int active_part_count
;
919 struct xpc_partition
*part
;
920 unsigned long printmsg_time
;
923 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
924 DBUG_ON(xpc_exiting
== 1);
927 * Let the heartbeat checker thread and the discovery thread
928 * (if one is running) know that they should exit. Also wake up
929 * the heartbeat checker thread in case it's sleeping.
932 wake_up_interruptible(&xpc_act_IRQ_wq
);
934 /* ignore all incoming interrupts */
935 free_irq(SGI_XPC_ACTIVATE
, NULL
);
937 /* wait for the discovery thread to exit */
938 down(&xpc_discovery_exited
);
940 /* wait for the heartbeat checker thread to exit */
941 down(&xpc_hb_checker_exited
);
944 /* sleep for a 1/3 of a second or so */
945 (void) msleep_interruptible(300);
948 /* wait for all partitions to become inactive */
950 printmsg_time
= jiffies
;
953 active_part_count
= 0;
955 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
956 part
= &xpc_partitions
[partid
];
958 if (xpc_partition_disengaged(part
) &&
959 part
->act_state
== XPC_P_INACTIVE
) {
965 XPC_DEACTIVATE_PARTITION(part
, reason
);
968 if (active_part_count
== 0) {
972 if (jiffies
>= printmsg_time
) {
973 dev_info(xpc_part
, "waiting for partitions to "
974 "deactivate/disengage, active count=%d, remote "
975 "engaged=0x%lx\n", active_part_count
,
976 xpc_partition_engaged(1UL << partid
));
978 printmsg_time
= jiffies
+
979 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
982 /* sleep for a 1/3 of a second or so */
983 (void) msleep_interruptible(300);
987 DBUG_ON(xpc_partition_engaged(-1UL));
990 /* indicate to others that our reserved page is uninitialized */
991 xpc_rsvd_page
->vars_pa
= 0;
993 /* now it's time to eliminate our heartbeat */
994 del_timer_sync(&xpc_hb_timer
);
995 DBUG_ON(xpc_vars
->heartbeating_to_mask
!= 0);
997 /* take ourselves off of the reboot_notifier_list */
998 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1000 /* close down protections for IPI operations */
1001 xpc_restrict_IPI_ops();
1004 /* clear the interface to XPC's functions */
1005 xpc_clear_interface();
1008 unregister_sysctl_table(xpc_sysctl
);
1014 * This function is called when the system is being rebooted.
1017 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1019 enum xpc_retval reason
;
1024 reason
= xpcSystemReboot
;
1027 reason
= xpcSystemHalt
;
1030 reason
= xpcSystemPoweroff
;
1033 reason
= xpcSystemGoingDown
;
1036 xpc_do_exit(reason
);
1046 struct xpc_partition
*part
;
1050 if (!ia64_platform_is("sn2")) {
1055 * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng
1056 * various portions of a partition's reserved page. Its size is based
1057 * on the size of the reserved page header and part_nasids mask. So we
1058 * need to ensure that the other items will fit as well.
1060 if (XPC_RP_VARS_SIZE
> XPC_RP_HEADER_SIZE
+ XP_NASID_MASK_BYTES
) {
1061 dev_err(xpc_part
, "xpc_remote_copy_buffer is not big enough\n");
1064 DBUG_ON((u64
) xpc_remote_copy_buffer
!=
1065 L1_CACHE_ALIGN((u64
) xpc_remote_copy_buffer
));
1067 snprintf(xpc_part
->bus_id
, BUS_ID_SIZE
, "part");
1068 snprintf(xpc_chan
->bus_id
, BUS_ID_SIZE
, "chan");
1070 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
, 1);
1073 * The first few fields of each entry of xpc_partitions[] need to
1074 * be initialized now so that calls to xpc_connect() and
1075 * xpc_disconnect() can be made prior to the activation of any remote
1076 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1077 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1078 * PARTITION HAS BEEN ACTIVATED.
1080 for (partid
= 1; partid
< XP_MAX_PARTITIONS
; partid
++) {
1081 part
= &xpc_partitions
[partid
];
1083 DBUG_ON((u64
) part
!= L1_CACHE_ALIGN((u64
) part
));
1085 part
->act_IRQ_rcvd
= 0;
1086 spin_lock_init(&part
->act_lock
);
1087 part
->act_state
= XPC_P_INACTIVE
;
1088 XPC_SET_REASON(part
, 0, 0);
1090 init_timer(&part
->disengage_request_timer
);
1091 part
->disengage_request_timer
.function
=
1092 xpc_timeout_partition_disengage_request
;
1093 part
->disengage_request_timer
.data
= (unsigned long) part
;
1095 part
->setup_state
= XPC_P_UNSET
;
1096 init_waitqueue_head(&part
->teardown_wq
);
1097 atomic_set(&part
->references
, 0);
1101 * Open up protections for IPI operations (and AMO operations on
1102 * Shub 1.1 systems).
1104 xpc_allow_IPI_ops();
1107 * Interrupts being processed will increment this atomic variable and
1108 * awaken the heartbeat thread which will process the interrupts.
1110 atomic_set(&xpc_act_IRQ_rcvd
, 0);
1113 * This is safe to do before the xpc_hb_checker thread has started
1114 * because the handler releases a wait queue. If an interrupt is
1115 * received before the thread is waiting, it will not go to sleep,
1116 * but rather immediately process the interrupt.
1118 ret
= request_irq(SGI_XPC_ACTIVATE
, xpc_act_IRQ_handler
, 0,
1121 dev_err(xpc_part
, "can't register ACTIVATE IRQ handler, "
1122 "errno=%d\n", -ret
);
1124 xpc_restrict_IPI_ops();
1127 unregister_sysctl_table(xpc_sysctl
);
1133 * Fill the partition reserved page with the information needed by
1134 * other partitions to discover we are alive and establish initial
1137 xpc_rsvd_page
= xpc_rsvd_page_init();
1138 if (xpc_rsvd_page
== NULL
) {
1139 dev_err(xpc_part
, "could not setup our reserved page\n");
1141 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1142 xpc_restrict_IPI_ops();
1145 unregister_sysctl_table(xpc_sysctl
);
1151 /* add ourselves to the reboot_notifier_list */
1152 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1154 dev_warn(xpc_part
, "can't register reboot notifier\n");
1159 * Set the beating to other partitions into motion. This is
1160 * the last requirement for other partitions' discovery to
1161 * initiate communications with us.
1163 init_timer(&xpc_hb_timer
);
1164 xpc_hb_timer
.function
= xpc_hb_beater
;
1169 * The real work-horse behind xpc. This processes incoming
1170 * interrupts and monitors remote heartbeats.
1172 pid
= kernel_thread(xpc_hb_checker
, NULL
, 0);
1174 dev_err(xpc_part
, "failed while forking hb check thread\n");
1176 /* indicate to others that our reserved page is uninitialized */
1177 xpc_rsvd_page
->vars_pa
= 0;
1179 /* take ourselves off of the reboot_notifier_list */
1180 (void) unregister_reboot_notifier(&xpc_reboot_notifier
);
1182 del_timer_sync(&xpc_hb_timer
);
1183 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1184 xpc_restrict_IPI_ops();
1187 unregister_sysctl_table(xpc_sysctl
);
1194 * Startup a thread that will attempt to discover other partitions to
1195 * activate based on info provided by SAL. This new thread is short
1196 * lived and will exit once discovery is complete.
1198 pid
= kernel_thread(xpc_initiate_discovery
, NULL
, 0);
1200 dev_err(xpc_part
, "failed while forking discovery thread\n");
1202 /* mark this new thread as a non-starter */
1203 up(&xpc_discovery_exited
);
1205 xpc_do_exit(xpcUnloading
);
1210 /* set the interface to point at XPC's functions */
1211 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1212 xpc_initiate_allocate
, xpc_initiate_send
,
1213 xpc_initiate_send_notify
, xpc_initiate_received
,
1214 xpc_initiate_partid_to_nasids
);
1218 module_init(xpc_init
);
1224 xpc_do_exit(xpcUnloading
);
1226 module_exit(xpc_exit
);
1229 MODULE_AUTHOR("Silicon Graphics, Inc.");
1230 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1231 MODULE_LICENSE("GPL");
1233 module_param(xpc_hb_interval
, int, 0);
1234 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1235 "heartbeat increments.");
1237 module_param(xpc_hb_check_interval
, int, 0);
1238 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1239 "heartbeat checks.");
1241 module_param(xpc_disengage_request_timelimit
, int, 0);
1242 MODULE_PARM_DESC(xpc_disengage_request_timelimit
, "Number of seconds to wait "
1243 "for disengage request to complete.");