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) 2004-2008 Silicon Graphics, Inc. All Rights Reserved.
10 * Cross Partition Communication (XPC) support - standard version.
12 * XPC provides a message passing capability that crosses partition
13 * boundaries. This module is made up of two parts:
15 * partition This part detects the presence/absence of other
16 * partitions. It provides a heartbeat and monitors
17 * the heartbeats of other partitions.
19 * channel This part manages the channels and sends/receives
20 * messages across them to/from other partitions.
22 * There are a couple of additional functions residing in XP, which
23 * provide an interface to XPC for its users.
28 * . Currently on sn2, we have no way to determine which nasid an IRQ
29 * came from. Thus, xpc_send_IRQ_sn2() does a remote amo write
30 * followed by an IPI. The amo indicates where data is to be pulled
31 * from, so after the IPI arrives, the remote partition checks the amo
32 * word. The IPI can actually arrive before the amo however, so other
33 * code must periodically check for this case. Also, remote amo
34 * operations do not reliably time out. Thus we do a remote PIO read
35 * solely to know whether the remote partition is down and whether we
36 * should stop sending IPIs to it. This remote PIO read operation is
37 * set up in a special nofault region so SAL knows to ignore (and
38 * cleanup) any errors due to the remote amo write, PIO read, and/or
39 * PIO write operations.
41 * If/when new hardware solves this IPI problem, we should abandon
42 * the current approach.
46 #include <linux/module.h>
47 #include <linux/sysctl.h>
48 #include <linux/device.h>
49 #include <linux/delay.h>
50 #include <linux/reboot.h>
51 #include <linux/kdebug.h>
52 #include <linux/kthread.h>
55 /* define two XPC debug device structures to be used with dev_dbg() et al */
57 struct device_driver xpc_dbg_name
= {
61 struct device xpc_part_dbg_subname
= {
62 .bus_id
= {0}, /* set to "part" at xpc_init() time */
63 .driver
= &xpc_dbg_name
66 struct device xpc_chan_dbg_subname
= {
67 .bus_id
= {0}, /* set to "chan" at xpc_init() time */
68 .driver
= &xpc_dbg_name
71 struct device
*xpc_part
= &xpc_part_dbg_subname
;
72 struct device
*xpc_chan
= &xpc_chan_dbg_subname
;
74 static int xpc_kdebug_ignore
;
76 /* systune related variables for /proc/sys directories */
78 static int xpc_hb_interval
= XPC_HB_DEFAULT_INTERVAL
;
79 static int xpc_hb_min_interval
= 1;
80 static int xpc_hb_max_interval
= 10;
82 static int xpc_hb_check_interval
= XPC_HB_CHECK_DEFAULT_INTERVAL
;
83 static int xpc_hb_check_min_interval
= 10;
84 static int xpc_hb_check_max_interval
= 120;
86 int xpc_disengage_timelimit
= XPC_DISENGAGE_DEFAULT_TIMELIMIT
;
87 static int xpc_disengage_min_timelimit
; /* = 0 */
88 static int xpc_disengage_max_timelimit
= 120;
90 static ctl_table xpc_sys_xpc_hb_dir
[] = {
92 .ctl_name
= CTL_UNNUMBERED
,
93 .procname
= "hb_interval",
94 .data
= &xpc_hb_interval
,
95 .maxlen
= sizeof(int),
97 .proc_handler
= &proc_dointvec_minmax
,
98 .strategy
= &sysctl_intvec
,
99 .extra1
= &xpc_hb_min_interval
,
100 .extra2
= &xpc_hb_max_interval
},
102 .ctl_name
= CTL_UNNUMBERED
,
103 .procname
= "hb_check_interval",
104 .data
= &xpc_hb_check_interval
,
105 .maxlen
= sizeof(int),
107 .proc_handler
= &proc_dointvec_minmax
,
108 .strategy
= &sysctl_intvec
,
109 .extra1
= &xpc_hb_check_min_interval
,
110 .extra2
= &xpc_hb_check_max_interval
},
113 static ctl_table xpc_sys_xpc_dir
[] = {
115 .ctl_name
= CTL_UNNUMBERED
,
118 .child
= xpc_sys_xpc_hb_dir
},
120 .ctl_name
= CTL_UNNUMBERED
,
121 .procname
= "disengage_timelimit",
122 .data
= &xpc_disengage_timelimit
,
123 .maxlen
= sizeof(int),
125 .proc_handler
= &proc_dointvec_minmax
,
126 .strategy
= &sysctl_intvec
,
127 .extra1
= &xpc_disengage_min_timelimit
,
128 .extra2
= &xpc_disengage_max_timelimit
},
131 static ctl_table xpc_sys_dir
[] = {
133 .ctl_name
= CTL_UNNUMBERED
,
136 .child
= xpc_sys_xpc_dir
},
139 static struct ctl_table_header
*xpc_sysctl
;
141 /* non-zero if any remote partition disengage was timed out */
142 int xpc_disengage_timedout
;
144 /* #of activate IRQs received and not yet processed */
145 int xpc_activate_IRQ_rcvd
;
146 DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock
);
148 /* IRQ handler notifies this wait queue on receipt of an IRQ */
149 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq
);
151 static unsigned long xpc_hb_check_timeout
;
152 static struct timer_list xpc_hb_timer
;
153 void *xpc_heartbeating_to_mask
;
155 /* notification that the xpc_hb_checker thread has exited */
156 static DECLARE_COMPLETION(xpc_hb_checker_exited
);
158 /* notification that the xpc_discovery thread has exited */
159 static DECLARE_COMPLETION(xpc_discovery_exited
);
161 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
163 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
164 static struct notifier_block xpc_reboot_notifier
= {
165 .notifier_call
= xpc_system_reboot
,
168 static int xpc_system_die(struct notifier_block
*, unsigned long, void *);
169 static struct notifier_block xpc_die_notifier
= {
170 .notifier_call
= xpc_system_die
,
173 int (*xpc_setup_partitions_sn
) (void);
174 enum xp_retval (*xpc_get_partition_rsvd_page_pa
) (void *buf
, u64
*cookie
,
175 unsigned long *rp_pa
,
177 int (*xpc_setup_rsvd_page_sn
) (struct xpc_rsvd_page
*rp
);
178 void (*xpc_heartbeat_init
) (void);
179 void (*xpc_heartbeat_exit
) (void);
180 void (*xpc_increment_heartbeat
) (void);
181 void (*xpc_offline_heartbeat
) (void);
182 void (*xpc_online_heartbeat
) (void);
183 enum xp_retval (*xpc_get_remote_heartbeat
) (struct xpc_partition
*part
);
185 enum xp_retval (*xpc_make_first_contact
) (struct xpc_partition
*part
);
186 void (*xpc_notify_senders_of_disconnect
) (struct xpc_channel
*ch
);
187 u64 (*xpc_get_chctl_all_flags
) (struct xpc_partition
*part
);
188 enum xp_retval (*xpc_setup_msg_structures
) (struct xpc_channel
*ch
);
189 void (*xpc_teardown_msg_structures
) (struct xpc_channel
*ch
);
190 void (*xpc_process_msg_chctl_flags
) (struct xpc_partition
*part
, int ch_number
);
191 int (*xpc_n_of_deliverable_payloads
) (struct xpc_channel
*ch
);
192 void *(*xpc_get_deliverable_payload
) (struct xpc_channel
*ch
);
194 void (*xpc_request_partition_activation
) (struct xpc_rsvd_page
*remote_rp
,
195 unsigned long remote_rp_pa
,
197 void (*xpc_request_partition_reactivation
) (struct xpc_partition
*part
);
198 void (*xpc_request_partition_deactivation
) (struct xpc_partition
*part
);
199 void (*xpc_cancel_partition_deactivation_request
) (struct xpc_partition
*part
);
201 void (*xpc_process_activate_IRQ_rcvd
) (void);
202 enum xp_retval (*xpc_setup_ch_structures_sn
) (struct xpc_partition
*part
);
203 void (*xpc_teardown_ch_structures_sn
) (struct xpc_partition
*part
);
205 void (*xpc_indicate_partition_engaged
) (struct xpc_partition
*part
);
206 int (*xpc_partition_engaged
) (short partid
);
207 int (*xpc_any_partition_engaged
) (void);
208 void (*xpc_indicate_partition_disengaged
) (struct xpc_partition
*part
);
209 void (*xpc_assume_partition_disengaged
) (short partid
);
211 void (*xpc_send_chctl_closerequest
) (struct xpc_channel
*ch
,
212 unsigned long *irq_flags
);
213 void (*xpc_send_chctl_closereply
) (struct xpc_channel
*ch
,
214 unsigned long *irq_flags
);
215 void (*xpc_send_chctl_openrequest
) (struct xpc_channel
*ch
,
216 unsigned long *irq_flags
);
217 void (*xpc_send_chctl_openreply
) (struct xpc_channel
*ch
,
218 unsigned long *irq_flags
);
220 void (*xpc_save_remote_msgqueue_pa
) (struct xpc_channel
*ch
,
221 unsigned long msgqueue_pa
);
223 enum xp_retval (*xpc_send_payload
) (struct xpc_channel
*ch
, u32 flags
,
224 void *payload
, u16 payload_size
,
225 u8 notify_type
, xpc_notify_func func
,
227 void (*xpc_received_payload
) (struct xpc_channel
*ch
, void *payload
);
230 * Timer function to enforce the timelimit on the partition disengage.
233 xpc_timeout_partition_disengage(unsigned long data
)
235 struct xpc_partition
*part
= (struct xpc_partition
*)data
;
237 DBUG_ON(time_is_after_jiffies(part
->disengage_timeout
));
239 (void)xpc_partition_disengaged(part
);
241 DBUG_ON(part
->disengage_timeout
!= 0);
242 DBUG_ON(xpc_partition_engaged(XPC_PARTID(part
)));
246 * Timer to produce the heartbeat. The timer structures function is
247 * already set when this is initially called. A tunable is used to
248 * specify when the next timeout should occur.
251 xpc_hb_beater(unsigned long dummy
)
253 xpc_increment_heartbeat();
255 if (time_is_before_eq_jiffies(xpc_hb_check_timeout
))
256 wake_up_interruptible(&xpc_activate_IRQ_wq
);
258 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
259 add_timer(&xpc_hb_timer
);
263 xpc_start_hb_beater(void)
265 xpc_heartbeat_init();
266 init_timer(&xpc_hb_timer
);
267 xpc_hb_timer
.function
= xpc_hb_beater
;
272 xpc_stop_hb_beater(void)
274 del_timer_sync(&xpc_hb_timer
);
275 xpc_heartbeat_exit();
279 * At periodic intervals, scan through all active partitions and ensure
280 * their heartbeat is still active. If not, the partition is deactivated.
283 xpc_check_remote_hb(void)
285 struct xpc_partition
*part
;
289 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
294 if (partid
== xp_partition_id
)
297 part
= &xpc_partitions
[partid
];
299 if (part
->act_state
== XPC_P_AS_INACTIVE
||
300 part
->act_state
== XPC_P_AS_DEACTIVATING
) {
304 ret
= xpc_get_remote_heartbeat(part
);
305 if (ret
!= xpSuccess
)
306 XPC_DEACTIVATE_PARTITION(part
, ret
);
311 * This thread is responsible for nearly all of the partition
312 * activation/deactivation.
315 xpc_hb_checker(void *ignore
)
319 /* this thread was marked active by xpc_hb_init() */
321 set_cpus_allowed_ptr(current
, &cpumask_of_cpu(XPC_HB_CHECK_CPU
));
323 /* set our heartbeating to other partitions into motion */
324 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
325 xpc_start_hb_beater();
327 while (!xpc_exiting
) {
329 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
331 (int)(xpc_hb_check_timeout
- jiffies
),
332 xpc_activate_IRQ_rcvd
);
334 /* checking of remote heartbeats is skewed by IRQ handling */
335 if (time_is_before_eq_jiffies(xpc_hb_check_timeout
)) {
336 xpc_hb_check_timeout
= jiffies
+
337 (xpc_hb_check_interval
* HZ
);
339 dev_dbg(xpc_part
, "checking remote heartbeats\n");
340 xpc_check_remote_hb();
343 * On sn2 we need to periodically recheck to ensure no
344 * IRQ/amo pairs have been missed.
350 /* check for outstanding IRQs */
351 if (xpc_activate_IRQ_rcvd
> 0 || force_IRQ
!= 0) {
353 dev_dbg(xpc_part
, "processing activate IRQs "
355 xpc_process_activate_IRQ_rcvd();
358 /* wait for IRQ or timeout */
359 (void)wait_event_interruptible(xpc_activate_IRQ_wq
,
360 (time_is_before_eq_jiffies(
361 xpc_hb_check_timeout
) ||
362 xpc_activate_IRQ_rcvd
> 0 ||
366 xpc_stop_hb_beater();
368 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
370 /* mark this thread as having exited */
371 complete(&xpc_hb_checker_exited
);
376 * This thread will attempt to discover other partitions to activate
377 * based on info provided by SAL. This new thread is short lived and
378 * will exit once discovery is complete.
381 xpc_initiate_discovery(void *ignore
)
385 dev_dbg(xpc_part
, "discovery thread is exiting\n");
387 /* mark this thread as having exited */
388 complete(&xpc_discovery_exited
);
393 * The first kthread assigned to a newly activated partition is the one
394 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
395 * that kthread until the partition is brought down, at which time that kthread
396 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
397 * that XPC has dismantled all communication infrastructure for the associated
398 * partition.) This kthread becomes the channel manager for that partition.
400 * Each active partition has a channel manager, who, besides connecting and
401 * disconnecting channels, will ensure that each of the partition's connected
402 * channels has the required number of assigned kthreads to get the work done.
405 xpc_channel_mgr(struct xpc_partition
*part
)
407 while (part
->act_state
!= XPC_P_AS_DEACTIVATING
||
408 atomic_read(&part
->nchannels_active
) > 0 ||
409 !xpc_partition_disengaged(part
)) {
411 xpc_process_sent_chctl_flags(part
);
414 * Wait until we've been requested to activate kthreads or
415 * all of the channel's message queues have been torn down or
416 * a signal is pending.
418 * The channel_mgr_requests is set to 1 after being awakened,
419 * This is done to prevent the channel mgr from making one pass
420 * through the loop for each request, since he will
421 * be servicing all the requests in one pass. The reason it's
422 * set to 1 instead of 0 is so that other kthreads will know
423 * that the channel mgr is running and won't bother trying to
426 atomic_dec(&part
->channel_mgr_requests
);
427 (void)wait_event_interruptible(part
->channel_mgr_wq
,
428 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
429 part
->chctl
.all_flags
!= 0 ||
430 (part
->act_state
== XPC_P_AS_DEACTIVATING
&&
431 atomic_read(&part
->nchannels_active
) == 0 &&
432 xpc_partition_disengaged(part
))));
433 atomic_set(&part
->channel_mgr_requests
, 1);
438 * Guarantee that the kzalloc'd memory is cacheline aligned.
441 xpc_kzalloc_cacheline_aligned(size_t size
, gfp_t flags
, void **base
)
443 /* see if kzalloc will give us cachline aligned memory by default */
444 *base
= kzalloc(size
, flags
);
448 if ((u64
)*base
== L1_CACHE_ALIGN((u64
)*base
))
453 /* nope, we'll have to do it ourselves */
454 *base
= kzalloc(size
+ L1_CACHE_BYTES
, flags
);
458 return (void *)L1_CACHE_ALIGN((u64
)*base
);
462 * Setup the channel structures necessary to support XPartition Communication
463 * between the specified remote partition and the local one.
465 static enum xp_retval
466 xpc_setup_ch_structures(struct xpc_partition
*part
)
470 struct xpc_channel
*ch
;
471 short partid
= XPC_PARTID(part
);
474 * Allocate all of the channel structures as a contiguous chunk of
477 DBUG_ON(part
->channels
!= NULL
);
478 part
->channels
= kzalloc(sizeof(struct xpc_channel
) * XPC_MAX_NCHANNELS
,
480 if (part
->channels
== NULL
) {
481 dev_err(xpc_chan
, "can't get memory for channels\n");
485 /* allocate the remote open and close args */
487 part
->remote_openclose_args
=
488 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE
,
490 remote_openclose_args_base
);
491 if (part
->remote_openclose_args
== NULL
) {
492 dev_err(xpc_chan
, "can't get memory for remote connect args\n");
497 part
->chctl
.all_flags
= 0;
498 spin_lock_init(&part
->chctl_lock
);
500 atomic_set(&part
->channel_mgr_requests
, 1);
501 init_waitqueue_head(&part
->channel_mgr_wq
);
503 part
->nchannels
= XPC_MAX_NCHANNELS
;
505 atomic_set(&part
->nchannels_active
, 0);
506 atomic_set(&part
->nchannels_engaged
, 0);
508 for (ch_number
= 0; ch_number
< part
->nchannels
; ch_number
++) {
509 ch
= &part
->channels
[ch_number
];
512 ch
->number
= ch_number
;
513 ch
->flags
= XPC_C_DISCONNECTED
;
515 atomic_set(&ch
->kthreads_assigned
, 0);
516 atomic_set(&ch
->kthreads_idle
, 0);
517 atomic_set(&ch
->kthreads_active
, 0);
519 atomic_set(&ch
->references
, 0);
520 atomic_set(&ch
->n_to_notify
, 0);
522 spin_lock_init(&ch
->lock
);
523 init_completion(&ch
->wdisconnect_wait
);
525 atomic_set(&ch
->n_on_msg_allocate_wq
, 0);
526 init_waitqueue_head(&ch
->msg_allocate_wq
);
527 init_waitqueue_head(&ch
->idle_wq
);
530 ret
= xpc_setup_ch_structures_sn(part
);
531 if (ret
!= xpSuccess
)
535 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
536 * we're declaring that this partition is ready to go.
538 part
->setup_state
= XPC_P_SS_SETUP
;
542 /* setup of ch structures failed */
544 kfree(part
->remote_openclose_args_base
);
545 part
->remote_openclose_args
= NULL
;
547 kfree(part
->channels
);
548 part
->channels
= NULL
;
553 * Teardown the channel structures necessary to support XPartition Communication
554 * between the specified remote partition and the local one.
557 xpc_teardown_ch_structures(struct xpc_partition
*part
)
559 DBUG_ON(atomic_read(&part
->nchannels_engaged
) != 0);
560 DBUG_ON(atomic_read(&part
->nchannels_active
) != 0);
563 * Make this partition inaccessible to local processes by marking it
564 * as no longer setup. Then wait before proceeding with the teardown
565 * until all existing references cease.
567 DBUG_ON(part
->setup_state
!= XPC_P_SS_SETUP
);
568 part
->setup_state
= XPC_P_SS_WTEARDOWN
;
570 wait_event(part
->teardown_wq
, (atomic_read(&part
->references
) == 0));
572 /* now we can begin tearing down the infrastructure */
574 xpc_teardown_ch_structures_sn(part
);
576 kfree(part
->remote_openclose_args_base
);
577 part
->remote_openclose_args
= NULL
;
578 kfree(part
->channels
);
579 part
->channels
= NULL
;
581 part
->setup_state
= XPC_P_SS_TORNDOWN
;
585 * When XPC HB determines that a partition has come up, it will create a new
586 * kthread and that kthread will call this function to attempt to set up the
587 * basic infrastructure used for Cross Partition Communication with the newly
590 * The kthread that was created by XPC HB and which setup the XPC
591 * infrastructure will remain assigned to the partition becoming the channel
592 * manager for that partition until the partition is deactivating, at which
593 * time the kthread will teardown the XPC infrastructure and then exit.
596 xpc_activating(void *__partid
)
598 short partid
= (u64
)__partid
;
599 struct xpc_partition
*part
= &xpc_partitions
[partid
];
600 unsigned long irq_flags
;
602 DBUG_ON(partid
< 0 || partid
>= xp_max_npartitions
);
604 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
606 if (part
->act_state
== XPC_P_AS_DEACTIVATING
) {
607 part
->act_state
= XPC_P_AS_INACTIVE
;
608 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
609 part
->remote_rp_pa
= 0;
613 /* indicate the thread is activating */
614 DBUG_ON(part
->act_state
!= XPC_P_AS_ACTIVATION_REQ
);
615 part
->act_state
= XPC_P_AS_ACTIVATING
;
617 XPC_SET_REASON(part
, 0, 0);
618 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
620 dev_dbg(xpc_part
, "activating partition %d\n", partid
);
622 xpc_allow_hb(partid
);
624 if (xpc_setup_ch_structures(part
) == xpSuccess
) {
625 (void)xpc_part_ref(part
); /* this will always succeed */
627 if (xpc_make_first_contact(part
) == xpSuccess
) {
628 xpc_mark_partition_active(part
);
629 xpc_channel_mgr(part
);
630 /* won't return until partition is deactivating */
633 xpc_part_deref(part
);
634 xpc_teardown_ch_structures(part
);
637 xpc_disallow_hb(partid
);
638 xpc_mark_partition_inactive(part
);
640 if (part
->reason
== xpReactivating
) {
641 /* interrupting ourselves results in activating partition */
642 xpc_request_partition_reactivation(part
);
649 xpc_activate_partition(struct xpc_partition
*part
)
651 short partid
= XPC_PARTID(part
);
652 unsigned long irq_flags
;
653 struct task_struct
*kthread
;
655 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
657 DBUG_ON(part
->act_state
!= XPC_P_AS_INACTIVE
);
659 part
->act_state
= XPC_P_AS_ACTIVATION_REQ
;
660 XPC_SET_REASON(part
, xpCloneKThread
, __LINE__
);
662 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
664 kthread
= kthread_run(xpc_activating
, (void *)((u64
)partid
), "xpc%02d",
666 if (IS_ERR(kthread
)) {
667 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
668 part
->act_state
= XPC_P_AS_INACTIVE
;
669 XPC_SET_REASON(part
, xpCloneKThreadFailed
, __LINE__
);
670 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
675 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
677 int idle
= atomic_read(&ch
->kthreads_idle
);
678 int assigned
= atomic_read(&ch
->kthreads_assigned
);
681 DBUG_ON(needed
<= 0);
684 wakeup
= (needed
> idle
) ? idle
: needed
;
687 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
688 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
690 /* only wakeup the requested number of kthreads */
691 wake_up_nr(&ch
->idle_wq
, wakeup
);
697 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
698 needed
= ch
->kthreads_assigned_limit
- assigned
;
703 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
704 needed
, ch
->partid
, ch
->number
);
706 xpc_create_kthreads(ch
, needed
, 0);
710 * This function is where XPC's kthreads wait for messages to deliver.
713 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
716 /* deliver messages to their intended recipients */
718 while (xpc_n_of_deliverable_payloads(ch
) > 0 &&
719 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
720 xpc_deliver_payload(ch
);
723 if (atomic_inc_return(&ch
->kthreads_idle
) >
724 ch
->kthreads_idle_limit
) {
725 /* too many idle kthreads on this channel */
726 atomic_dec(&ch
->kthreads_idle
);
730 dev_dbg(xpc_chan
, "idle kthread calling "
731 "wait_event_interruptible_exclusive()\n");
733 (void)wait_event_interruptible_exclusive(ch
->idle_wq
,
734 (xpc_n_of_deliverable_payloads(ch
) > 0 ||
735 (ch
->flags
& XPC_C_DISCONNECTING
)));
737 atomic_dec(&ch
->kthreads_idle
);
739 } while (!(ch
->flags
& XPC_C_DISCONNECTING
));
743 xpc_kthread_start(void *args
)
745 short partid
= XPC_UNPACK_ARG1(args
);
746 u16 ch_number
= XPC_UNPACK_ARG2(args
);
747 struct xpc_partition
*part
= &xpc_partitions
[partid
];
748 struct xpc_channel
*ch
;
750 unsigned long irq_flags
;
752 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
755 ch
= &part
->channels
[ch_number
];
757 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
759 /* let registerer know that connection has been established */
761 spin_lock_irqsave(&ch
->lock
, irq_flags
);
762 if (!(ch
->flags
& XPC_C_CONNECTEDCALLOUT
)) {
763 ch
->flags
|= XPC_C_CONNECTEDCALLOUT
;
764 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
766 xpc_connected_callout(ch
);
768 spin_lock_irqsave(&ch
->lock
, irq_flags
);
769 ch
->flags
|= XPC_C_CONNECTEDCALLOUT_MADE
;
770 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
773 * It is possible that while the callout was being
774 * made that the remote partition sent some messages.
775 * If that is the case, we may need to activate
776 * additional kthreads to help deliver them. We only
777 * need one less than total #of messages to deliver.
779 n_needed
= xpc_n_of_deliverable_payloads(ch
) - 1;
780 if (n_needed
> 0 && !(ch
->flags
& XPC_C_DISCONNECTING
))
781 xpc_activate_kthreads(ch
, n_needed
);
784 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
787 xpc_kthread_waitmsgs(part
, ch
);
790 /* let registerer know that connection is disconnecting */
792 spin_lock_irqsave(&ch
->lock
, irq_flags
);
793 if ((ch
->flags
& XPC_C_CONNECTEDCALLOUT_MADE
) &&
794 !(ch
->flags
& XPC_C_DISCONNECTINGCALLOUT
)) {
795 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT
;
796 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
798 xpc_disconnect_callout(ch
, xpDisconnecting
);
800 spin_lock_irqsave(&ch
->lock
, irq_flags
);
801 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT_MADE
;
803 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
805 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
806 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
807 xpc_indicate_partition_disengaged(part
);
810 xpc_msgqueue_deref(ch
);
812 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
815 xpc_part_deref(part
);
820 * For each partition that XPC has established communications with, there is
821 * a minimum of one kernel thread assigned to perform any operation that
822 * may potentially sleep or block (basically the callouts to the asynchronous
823 * functions registered via xpc_connect()).
825 * Additional kthreads are created and destroyed by XPC as the workload
828 * A kthread is assigned to one of the active channels that exists for a given
832 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
,
833 int ignore_disconnecting
)
835 unsigned long irq_flags
;
836 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
837 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
838 struct task_struct
*kthread
;
840 while (needed
-- > 0) {
843 * The following is done on behalf of the newly created
844 * kthread. That kthread is responsible for doing the
845 * counterpart to the following before it exits.
847 if (ignore_disconnecting
) {
848 if (!atomic_inc_not_zero(&ch
->kthreads_assigned
)) {
849 /* kthreads assigned had gone to zero */
851 XPC_C_DISCONNECTINGCALLOUT_MADE
));
855 } else if (ch
->flags
& XPC_C_DISCONNECTING
) {
858 } else if (atomic_inc_return(&ch
->kthreads_assigned
) == 1 &&
859 atomic_inc_return(&part
->nchannels_engaged
) == 1) {
860 xpc_indicate_partition_engaged(part
);
862 (void)xpc_part_ref(part
);
863 xpc_msgqueue_ref(ch
);
865 kthread
= kthread_run(xpc_kthread_start
, (void *)args
,
866 "xpc%02dc%d", ch
->partid
, ch
->number
);
867 if (IS_ERR(kthread
)) {
868 /* the fork failed */
871 * NOTE: if (ignore_disconnecting &&
872 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
873 * then we'll deadlock if all other kthreads assigned
874 * to this channel are blocked in the channel's
875 * registerer, because the only thing that will unblock
876 * them is the xpDisconnecting callout that this
877 * failed kthread_run() would have made.
880 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
881 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
882 xpc_indicate_partition_disengaged(part
);
884 xpc_msgqueue_deref(ch
);
885 xpc_part_deref(part
);
887 if (atomic_read(&ch
->kthreads_assigned
) <
888 ch
->kthreads_idle_limit
) {
890 * Flag this as an error only if we have an
891 * insufficient #of kthreads for the channel
894 spin_lock_irqsave(&ch
->lock
, irq_flags
);
895 XPC_DISCONNECT_CHANNEL(ch
, xpLackOfResources
,
897 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
905 xpc_disconnect_wait(int ch_number
)
907 unsigned long irq_flags
;
909 struct xpc_partition
*part
;
910 struct xpc_channel
*ch
;
911 int wakeup_channel_mgr
;
913 /* now wait for all callouts to the caller's function to cease */
914 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
915 part
= &xpc_partitions
[partid
];
917 if (!xpc_part_ref(part
))
920 ch
= &part
->channels
[ch_number
];
922 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
923 xpc_part_deref(part
);
927 wait_for_completion(&ch
->wdisconnect_wait
);
929 spin_lock_irqsave(&ch
->lock
, irq_flags
);
930 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
931 wakeup_channel_mgr
= 0;
933 if (ch
->delayed_chctl_flags
) {
934 if (part
->act_state
!= XPC_P_AS_DEACTIVATING
) {
935 spin_lock(&part
->chctl_lock
);
936 part
->chctl
.flags
[ch
->number
] |=
937 ch
->delayed_chctl_flags
;
938 spin_unlock(&part
->chctl_lock
);
939 wakeup_channel_mgr
= 1;
941 ch
->delayed_chctl_flags
= 0;
944 ch
->flags
&= ~XPC_C_WDISCONNECT
;
945 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
947 if (wakeup_channel_mgr
)
948 xpc_wakeup_channel_mgr(part
);
950 xpc_part_deref(part
);
955 xpc_setup_partitions(void)
958 struct xpc_partition
*part
;
960 xpc_partitions
= kzalloc(sizeof(struct xpc_partition
) *
961 xp_max_npartitions
, GFP_KERNEL
);
962 if (xpc_partitions
== NULL
) {
963 dev_err(xpc_part
, "can't get memory for partition structure\n");
968 * The first few fields of each entry of xpc_partitions[] need to
969 * be initialized now so that calls to xpc_connect() and
970 * xpc_disconnect() can be made prior to the activation of any remote
971 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
972 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
973 * PARTITION HAS BEEN ACTIVATED.
975 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
976 part
= &xpc_partitions
[partid
];
978 DBUG_ON((u64
)part
!= L1_CACHE_ALIGN((u64
)part
));
980 part
->activate_IRQ_rcvd
= 0;
981 spin_lock_init(&part
->act_lock
);
982 part
->act_state
= XPC_P_AS_INACTIVE
;
983 XPC_SET_REASON(part
, 0, 0);
985 init_timer(&part
->disengage_timer
);
986 part
->disengage_timer
.function
=
987 xpc_timeout_partition_disengage
;
988 part
->disengage_timer
.data
= (unsigned long)part
;
990 part
->setup_state
= XPC_P_SS_UNSET
;
991 init_waitqueue_head(&part
->teardown_wq
);
992 atomic_set(&part
->references
, 0);
995 return xpc_setup_partitions_sn();
999 xpc_teardown_partitions(void)
1001 kfree(xpc_partitions
);
1005 xpc_do_exit(enum xp_retval reason
)
1008 int active_part_count
, printed_waiting_msg
= 0;
1009 struct xpc_partition
*part
;
1010 unsigned long printmsg_time
, disengage_timeout
= 0;
1012 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
1013 DBUG_ON(xpc_exiting
== 1);
1016 * Let the heartbeat checker thread and the discovery thread
1017 * (if one is running) know that they should exit. Also wake up
1018 * the heartbeat checker thread in case it's sleeping.
1021 wake_up_interruptible(&xpc_activate_IRQ_wq
);
1023 /* wait for the discovery thread to exit */
1024 wait_for_completion(&xpc_discovery_exited
);
1026 /* wait for the heartbeat checker thread to exit */
1027 wait_for_completion(&xpc_hb_checker_exited
);
1029 /* sleep for a 1/3 of a second or so */
1030 (void)msleep_interruptible(300);
1032 /* wait for all partitions to become inactive */
1034 printmsg_time
= jiffies
+ (XPC_DEACTIVATE_PRINTMSG_INTERVAL
* HZ
);
1035 xpc_disengage_timedout
= 0;
1038 active_part_count
= 0;
1040 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1041 part
= &xpc_partitions
[partid
];
1043 if (xpc_partition_disengaged(part
) &&
1044 part
->act_state
== XPC_P_AS_INACTIVE
) {
1048 active_part_count
++;
1050 XPC_DEACTIVATE_PARTITION(part
, reason
);
1052 if (part
->disengage_timeout
> disengage_timeout
)
1053 disengage_timeout
= part
->disengage_timeout
;
1056 if (xpc_any_partition_engaged()) {
1057 if (time_is_before_jiffies(printmsg_time
)) {
1058 dev_info(xpc_part
, "waiting for remote "
1059 "partitions to deactivate, timeout in "
1060 "%ld seconds\n", (disengage_timeout
-
1062 printmsg_time
= jiffies
+
1063 (XPC_DEACTIVATE_PRINTMSG_INTERVAL
* HZ
);
1064 printed_waiting_msg
= 1;
1067 } else if (active_part_count
> 0) {
1068 if (printed_waiting_msg
) {
1069 dev_info(xpc_part
, "waiting for local partition"
1070 " to deactivate\n");
1071 printed_waiting_msg
= 0;
1075 if (!xpc_disengage_timedout
) {
1076 dev_info(xpc_part
, "all partitions have "
1082 /* sleep for a 1/3 of a second or so */
1083 (void)msleep_interruptible(300);
1087 DBUG_ON(xpc_any_partition_engaged());
1088 DBUG_ON(xpc_any_hbs_allowed() != 0);
1090 xpc_teardown_rsvd_page();
1092 if (reason
== xpUnloading
) {
1093 (void)unregister_die_notifier(&xpc_die_notifier
);
1094 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1097 /* clear the interface to XPC's functions */
1098 xpc_clear_interface();
1101 unregister_sysctl_table(xpc_sysctl
);
1103 xpc_teardown_partitions();
1112 * This function is called when the system is being rebooted.
1115 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1117 enum xp_retval reason
;
1121 reason
= xpSystemReboot
;
1124 reason
= xpSystemHalt
;
1127 reason
= xpSystemPoweroff
;
1130 reason
= xpSystemGoingDown
;
1133 xpc_do_exit(reason
);
1138 * Notify other partitions to deactivate from us by first disengaging from all
1139 * references to our memory.
1142 xpc_die_deactivate(void)
1144 struct xpc_partition
*part
;
1150 /* keep xpc_hb_checker thread from doing anything (just in case) */
1153 xpc_disallow_all_hbs(); /*indicate we're deactivated */
1155 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1156 part
= &xpc_partitions
[partid
];
1158 if (xpc_partition_engaged(partid
) ||
1159 part
->act_state
!= XPC_P_AS_INACTIVE
) {
1160 xpc_request_partition_deactivation(part
);
1161 xpc_indicate_partition_disengaged(part
);
1166 * Though we requested that all other partitions deactivate from us,
1167 * we only wait until they've all disengaged or we've reached the
1168 * defined timelimit.
1170 * Given that one iteration through the following while-loop takes
1171 * approximately 200 microseconds, calculate the #of loops to take
1172 * before bailing and the #of loops before printing a waiting message.
1174 keep_waiting
= xpc_disengage_timelimit
* 1000 * 5;
1175 wait_to_print
= XPC_DEACTIVATE_PRINTMSG_INTERVAL
* 1000 * 5;
1178 any_engaged
= xpc_any_partition_engaged();
1180 dev_info(xpc_part
, "all partitions have deactivated\n");
1184 if (!keep_waiting
--) {
1185 for (partid
= 0; partid
< xp_max_npartitions
;
1187 if (xpc_partition_engaged(partid
)) {
1188 dev_info(xpc_part
, "deactivate from "
1189 "remote partition %d timed "
1196 if (!wait_to_print
--) {
1197 dev_info(xpc_part
, "waiting for remote partitions to "
1198 "deactivate, timeout in %ld seconds\n",
1199 keep_waiting
/ (1000 * 5));
1200 wait_to_print
= XPC_DEACTIVATE_PRINTMSG_INTERVAL
*
1209 * This function is called when the system is being restarted or halted due
1210 * to some sort of system failure. If this is the case we need to notify the
1211 * other partitions to disengage from all references to our memory.
1212 * This function can also be called when our heartbeater could be offlined
1213 * for a time. In this case we need to notify other partitions to not worry
1214 * about the lack of a heartbeat.
1217 xpc_system_die(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1219 #ifdef CONFIG_IA64 /* !!! temporary kludge */
1221 case DIE_MACHINE_RESTART
:
1222 case DIE_MACHINE_HALT
:
1223 xpc_die_deactivate();
1226 case DIE_KDEBUG_ENTER
:
1227 /* Should lack of heartbeat be ignored by other partitions? */
1228 if (!xpc_kdebug_ignore
)
1232 case DIE_MCA_MONARCH_ENTER
:
1233 case DIE_INIT_MONARCH_ENTER
:
1234 xpc_offline_heartbeat();
1237 case DIE_KDEBUG_LEAVE
:
1238 /* Is lack of heartbeat being ignored by other partitions? */
1239 if (!xpc_kdebug_ignore
)
1243 case DIE_MCA_MONARCH_LEAVE
:
1244 case DIE_INIT_MONARCH_LEAVE
:
1245 xpc_online_heartbeat();
1249 xpc_die_deactivate();
1259 struct task_struct
*kthread
;
1261 snprintf(xpc_part
->bus_id
, BUS_ID_SIZE
, "part");
1262 snprintf(xpc_chan
->bus_id
, BUS_ID_SIZE
, "chan");
1266 * The ia64-sn2 architecture supports at most 64 partitions.
1267 * And the inability to unregister remote amos restricts us
1268 * further to only support exactly 64 partitions on this
1269 * architecture, no less.
1271 if (xp_max_npartitions
!= 64) {
1272 dev_err(xpc_part
, "max #of partitions not set to 64\n");
1275 ret
= xpc_init_sn2();
1278 } else if (is_uv()) {
1279 ret
= xpc_init_uv();
1288 ret
= xpc_setup_partitions();
1290 dev_err(xpc_part
, "can't get memory for partition structure\n");
1294 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
);
1297 * Fill the partition reserved page with the information needed by
1298 * other partitions to discover we are alive and establish initial
1301 ret
= xpc_setup_rsvd_page();
1303 dev_err(xpc_part
, "can't setup our reserved page\n");
1307 /* add ourselves to the reboot_notifier_list */
1308 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1310 dev_warn(xpc_part
, "can't register reboot notifier\n");
1312 /* add ourselves to the die_notifier list */
1313 ret
= register_die_notifier(&xpc_die_notifier
);
1315 dev_warn(xpc_part
, "can't register die notifier\n");
1318 * The real work-horse behind xpc. This processes incoming
1319 * interrupts and monitors remote heartbeats.
1321 kthread
= kthread_run(xpc_hb_checker
, NULL
, XPC_HB_CHECK_THREAD_NAME
);
1322 if (IS_ERR(kthread
)) {
1323 dev_err(xpc_part
, "failed while forking hb check thread\n");
1329 * Startup a thread that will attempt to discover other partitions to
1330 * activate based on info provided by SAL. This new thread is short
1331 * lived and will exit once discovery is complete.
1333 kthread
= kthread_run(xpc_initiate_discovery
, NULL
,
1334 XPC_DISCOVERY_THREAD_NAME
);
1335 if (IS_ERR(kthread
)) {
1336 dev_err(xpc_part
, "failed while forking discovery thread\n");
1338 /* mark this new thread as a non-starter */
1339 complete(&xpc_discovery_exited
);
1341 xpc_do_exit(xpUnloading
);
1345 /* set the interface to point at XPC's functions */
1346 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1347 xpc_initiate_send
, xpc_initiate_send_notify
,
1348 xpc_initiate_received
, xpc_initiate_partid_to_nasids
);
1352 /* initialization was not successful */
1354 xpc_teardown_rsvd_page();
1356 (void)unregister_die_notifier(&xpc_die_notifier
);
1357 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1360 unregister_sysctl_table(xpc_sysctl
);
1362 xpc_teardown_partitions();
1371 module_init(xpc_init
);
1376 xpc_do_exit(xpUnloading
);
1379 module_exit(xpc_exit
);
1381 MODULE_AUTHOR("Silicon Graphics, Inc.");
1382 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1383 MODULE_LICENSE("GPL");
1385 module_param(xpc_hb_interval
, int, 0);
1386 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1387 "heartbeat increments.");
1389 module_param(xpc_hb_check_interval
, int, 0);
1390 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1391 "heartbeat checks.");
1393 module_param(xpc_disengage_timelimit
, int, 0);
1394 MODULE_PARM_DESC(xpc_disengage_timelimit
, "Number of seconds to wait "
1395 "for disengage to complete.");
1397 module_param(xpc_kdebug_ignore
, int, 0);
1398 MODULE_PARM_DESC(xpc_kdebug_ignore
, "Should lack of heartbeat be ignored by "
1399 "other partitions when dropping into kdebug.");