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 .init_name
= "", /* set to "part" at xpc_init() time */
63 .driver
= &xpc_dbg_name
66 struct device xpc_chan_dbg_subname
= {
67 .init_name
= "", /* 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 void (*xpc_teardown_partitions_sn
) (void);
175 enum xp_retval (*xpc_get_partition_rsvd_page_pa
) (void *buf
, u64
*cookie
,
176 unsigned long *rp_pa
,
178 int (*xpc_setup_rsvd_page_sn
) (struct xpc_rsvd_page
*rp
);
179 void (*xpc_heartbeat_init
) (void);
180 void (*xpc_heartbeat_exit
) (void);
181 void (*xpc_increment_heartbeat
) (void);
182 void (*xpc_offline_heartbeat
) (void);
183 void (*xpc_online_heartbeat
) (void);
184 enum xp_retval (*xpc_get_remote_heartbeat
) (struct xpc_partition
*part
);
186 enum xp_retval (*xpc_make_first_contact
) (struct xpc_partition
*part
);
187 void (*xpc_notify_senders_of_disconnect
) (struct xpc_channel
*ch
);
188 u64 (*xpc_get_chctl_all_flags
) (struct xpc_partition
*part
);
189 enum xp_retval (*xpc_setup_msg_structures
) (struct xpc_channel
*ch
);
190 void (*xpc_teardown_msg_structures
) (struct xpc_channel
*ch
);
191 void (*xpc_process_msg_chctl_flags
) (struct xpc_partition
*part
, int ch_number
);
192 int (*xpc_n_of_deliverable_payloads
) (struct xpc_channel
*ch
);
193 void *(*xpc_get_deliverable_payload
) (struct xpc_channel
*ch
);
195 void (*xpc_request_partition_activation
) (struct xpc_rsvd_page
*remote_rp
,
196 unsigned long remote_rp_pa
,
198 void (*xpc_request_partition_reactivation
) (struct xpc_partition
*part
);
199 void (*xpc_request_partition_deactivation
) (struct xpc_partition
*part
);
200 void (*xpc_cancel_partition_deactivation_request
) (struct xpc_partition
*part
);
202 void (*xpc_process_activate_IRQ_rcvd
) (void);
203 enum xp_retval (*xpc_setup_ch_structures_sn
) (struct xpc_partition
*part
);
204 void (*xpc_teardown_ch_structures_sn
) (struct xpc_partition
*part
);
206 void (*xpc_indicate_partition_engaged
) (struct xpc_partition
*part
);
207 int (*xpc_partition_engaged
) (short partid
);
208 int (*xpc_any_partition_engaged
) (void);
209 void (*xpc_indicate_partition_disengaged
) (struct xpc_partition
*part
);
210 void (*xpc_assume_partition_disengaged
) (short partid
);
212 void (*xpc_send_chctl_closerequest
) (struct xpc_channel
*ch
,
213 unsigned long *irq_flags
);
214 void (*xpc_send_chctl_closereply
) (struct xpc_channel
*ch
,
215 unsigned long *irq_flags
);
216 void (*xpc_send_chctl_openrequest
) (struct xpc_channel
*ch
,
217 unsigned long *irq_flags
);
218 void (*xpc_send_chctl_openreply
) (struct xpc_channel
*ch
,
219 unsigned long *irq_flags
);
221 enum xp_retval (*xpc_save_remote_msgqueue_pa
) (struct xpc_channel
*ch
,
222 unsigned long msgqueue_pa
);
224 enum xp_retval (*xpc_send_payload
) (struct xpc_channel
*ch
, u32 flags
,
225 void *payload
, u16 payload_size
,
226 u8 notify_type
, xpc_notify_func func
,
228 void (*xpc_received_payload
) (struct xpc_channel
*ch
, void *payload
);
231 * Timer function to enforce the timelimit on the partition disengage.
234 xpc_timeout_partition_disengage(unsigned long data
)
236 struct xpc_partition
*part
= (struct xpc_partition
*)data
;
238 DBUG_ON(time_is_after_jiffies(part
->disengage_timeout
));
240 (void)xpc_partition_disengaged(part
);
242 DBUG_ON(part
->disengage_timeout
!= 0);
243 DBUG_ON(xpc_partition_engaged(XPC_PARTID(part
)));
247 * Timer to produce the heartbeat. The timer structures function is
248 * already set when this is initially called. A tunable is used to
249 * specify when the next timeout should occur.
252 xpc_hb_beater(unsigned long dummy
)
254 xpc_increment_heartbeat();
256 if (time_is_before_eq_jiffies(xpc_hb_check_timeout
))
257 wake_up_interruptible(&xpc_activate_IRQ_wq
);
259 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
260 add_timer(&xpc_hb_timer
);
264 xpc_start_hb_beater(void)
266 xpc_heartbeat_init();
267 init_timer(&xpc_hb_timer
);
268 xpc_hb_timer
.function
= xpc_hb_beater
;
273 xpc_stop_hb_beater(void)
275 del_timer_sync(&xpc_hb_timer
);
276 xpc_heartbeat_exit();
280 * At periodic intervals, scan through all active partitions and ensure
281 * their heartbeat is still active. If not, the partition is deactivated.
284 xpc_check_remote_hb(void)
286 struct xpc_partition
*part
;
290 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
295 if (partid
== xp_partition_id
)
298 part
= &xpc_partitions
[partid
];
300 if (part
->act_state
== XPC_P_AS_INACTIVE
||
301 part
->act_state
== XPC_P_AS_DEACTIVATING
) {
305 ret
= xpc_get_remote_heartbeat(part
);
306 if (ret
!= xpSuccess
)
307 XPC_DEACTIVATE_PARTITION(part
, ret
);
312 * This thread is responsible for nearly all of the partition
313 * activation/deactivation.
316 xpc_hb_checker(void *ignore
)
320 /* this thread was marked active by xpc_hb_init() */
322 set_cpus_allowed_ptr(current
, cpumask_of(XPC_HB_CHECK_CPU
));
324 /* set our heartbeating to other partitions into motion */
325 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
326 xpc_start_hb_beater();
328 while (!xpc_exiting
) {
330 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
332 (int)(xpc_hb_check_timeout
- jiffies
),
333 xpc_activate_IRQ_rcvd
);
335 /* checking of remote heartbeats is skewed by IRQ handling */
336 if (time_is_before_eq_jiffies(xpc_hb_check_timeout
)) {
337 xpc_hb_check_timeout
= jiffies
+
338 (xpc_hb_check_interval
* HZ
);
340 dev_dbg(xpc_part
, "checking remote heartbeats\n");
341 xpc_check_remote_hb();
344 * On sn2 we need to periodically recheck to ensure no
345 * IRQ/amo pairs have been missed.
351 /* check for outstanding IRQs */
352 if (xpc_activate_IRQ_rcvd
> 0 || force_IRQ
!= 0) {
354 dev_dbg(xpc_part
, "processing activate IRQs "
356 xpc_process_activate_IRQ_rcvd();
359 /* wait for IRQ or timeout */
360 (void)wait_event_interruptible(xpc_activate_IRQ_wq
,
361 (time_is_before_eq_jiffies(
362 xpc_hb_check_timeout
) ||
363 xpc_activate_IRQ_rcvd
> 0 ||
367 xpc_stop_hb_beater();
369 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
371 /* mark this thread as having exited */
372 complete(&xpc_hb_checker_exited
);
377 * This thread will attempt to discover other partitions to activate
378 * based on info provided by SAL. This new thread is short lived and
379 * will exit once discovery is complete.
382 xpc_initiate_discovery(void *ignore
)
386 dev_dbg(xpc_part
, "discovery thread is exiting\n");
388 /* mark this thread as having exited */
389 complete(&xpc_discovery_exited
);
394 * The first kthread assigned to a newly activated partition is the one
395 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
396 * that kthread until the partition is brought down, at which time that kthread
397 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
398 * that XPC has dismantled all communication infrastructure for the associated
399 * partition.) This kthread becomes the channel manager for that partition.
401 * Each active partition has a channel manager, who, besides connecting and
402 * disconnecting channels, will ensure that each of the partition's connected
403 * channels has the required number of assigned kthreads to get the work done.
406 xpc_channel_mgr(struct xpc_partition
*part
)
408 while (part
->act_state
!= XPC_P_AS_DEACTIVATING
||
409 atomic_read(&part
->nchannels_active
) > 0 ||
410 !xpc_partition_disengaged(part
)) {
412 xpc_process_sent_chctl_flags(part
);
415 * Wait until we've been requested to activate kthreads or
416 * all of the channel's message queues have been torn down or
417 * a signal is pending.
419 * The channel_mgr_requests is set to 1 after being awakened,
420 * This is done to prevent the channel mgr from making one pass
421 * through the loop for each request, since he will
422 * be servicing all the requests in one pass. The reason it's
423 * set to 1 instead of 0 is so that other kthreads will know
424 * that the channel mgr is running and won't bother trying to
427 atomic_dec(&part
->channel_mgr_requests
);
428 (void)wait_event_interruptible(part
->channel_mgr_wq
,
429 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
430 part
->chctl
.all_flags
!= 0 ||
431 (part
->act_state
== XPC_P_AS_DEACTIVATING
&&
432 atomic_read(&part
->nchannels_active
) == 0 &&
433 xpc_partition_disengaged(part
))));
434 atomic_set(&part
->channel_mgr_requests
, 1);
439 * Guarantee that the kzalloc'd memory is cacheline aligned.
442 xpc_kzalloc_cacheline_aligned(size_t size
, gfp_t flags
, void **base
)
444 /* see if kzalloc will give us cachline aligned memory by default */
445 *base
= kzalloc(size
, flags
);
449 if ((u64
)*base
== L1_CACHE_ALIGN((u64
)*base
))
454 /* nope, we'll have to do it ourselves */
455 *base
= kzalloc(size
+ L1_CACHE_BYTES
, flags
);
459 return (void *)L1_CACHE_ALIGN((u64
)*base
);
463 * Setup the channel structures necessary to support XPartition Communication
464 * between the specified remote partition and the local one.
466 static enum xp_retval
467 xpc_setup_ch_structures(struct xpc_partition
*part
)
471 struct xpc_channel
*ch
;
472 short partid
= XPC_PARTID(part
);
475 * Allocate all of the channel structures as a contiguous chunk of
478 DBUG_ON(part
->channels
!= NULL
);
479 part
->channels
= kzalloc(sizeof(struct xpc_channel
) * XPC_MAX_NCHANNELS
,
481 if (part
->channels
== NULL
) {
482 dev_err(xpc_chan
, "can't get memory for channels\n");
486 /* allocate the remote open and close args */
488 part
->remote_openclose_args
=
489 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE
,
491 remote_openclose_args_base
);
492 if (part
->remote_openclose_args
== NULL
) {
493 dev_err(xpc_chan
, "can't get memory for remote connect args\n");
498 part
->chctl
.all_flags
= 0;
499 spin_lock_init(&part
->chctl_lock
);
501 atomic_set(&part
->channel_mgr_requests
, 1);
502 init_waitqueue_head(&part
->channel_mgr_wq
);
504 part
->nchannels
= XPC_MAX_NCHANNELS
;
506 atomic_set(&part
->nchannels_active
, 0);
507 atomic_set(&part
->nchannels_engaged
, 0);
509 for (ch_number
= 0; ch_number
< part
->nchannels
; ch_number
++) {
510 ch
= &part
->channels
[ch_number
];
513 ch
->number
= ch_number
;
514 ch
->flags
= XPC_C_DISCONNECTED
;
516 atomic_set(&ch
->kthreads_assigned
, 0);
517 atomic_set(&ch
->kthreads_idle
, 0);
518 atomic_set(&ch
->kthreads_active
, 0);
520 atomic_set(&ch
->references
, 0);
521 atomic_set(&ch
->n_to_notify
, 0);
523 spin_lock_init(&ch
->lock
);
524 init_completion(&ch
->wdisconnect_wait
);
526 atomic_set(&ch
->n_on_msg_allocate_wq
, 0);
527 init_waitqueue_head(&ch
->msg_allocate_wq
);
528 init_waitqueue_head(&ch
->idle_wq
);
531 ret
= xpc_setup_ch_structures_sn(part
);
532 if (ret
!= xpSuccess
)
536 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
537 * we're declaring that this partition is ready to go.
539 part
->setup_state
= XPC_P_SS_SETUP
;
543 /* setup of ch structures failed */
545 kfree(part
->remote_openclose_args_base
);
546 part
->remote_openclose_args
= NULL
;
548 kfree(part
->channels
);
549 part
->channels
= NULL
;
554 * Teardown the channel structures necessary to support XPartition Communication
555 * between the specified remote partition and the local one.
558 xpc_teardown_ch_structures(struct xpc_partition
*part
)
560 DBUG_ON(atomic_read(&part
->nchannels_engaged
) != 0);
561 DBUG_ON(atomic_read(&part
->nchannels_active
) != 0);
564 * Make this partition inaccessible to local processes by marking it
565 * as no longer setup. Then wait before proceeding with the teardown
566 * until all existing references cease.
568 DBUG_ON(part
->setup_state
!= XPC_P_SS_SETUP
);
569 part
->setup_state
= XPC_P_SS_WTEARDOWN
;
571 wait_event(part
->teardown_wq
, (atomic_read(&part
->references
) == 0));
573 /* now we can begin tearing down the infrastructure */
575 xpc_teardown_ch_structures_sn(part
);
577 kfree(part
->remote_openclose_args_base
);
578 part
->remote_openclose_args
= NULL
;
579 kfree(part
->channels
);
580 part
->channels
= NULL
;
582 part
->setup_state
= XPC_P_SS_TORNDOWN
;
586 * When XPC HB determines that a partition has come up, it will create a new
587 * kthread and that kthread will call this function to attempt to set up the
588 * basic infrastructure used for Cross Partition Communication with the newly
591 * The kthread that was created by XPC HB and which setup the XPC
592 * infrastructure will remain assigned to the partition becoming the channel
593 * manager for that partition until the partition is deactivating, at which
594 * time the kthread will teardown the XPC infrastructure and then exit.
597 xpc_activating(void *__partid
)
599 short partid
= (u64
)__partid
;
600 struct xpc_partition
*part
= &xpc_partitions
[partid
];
601 unsigned long irq_flags
;
603 DBUG_ON(partid
< 0 || partid
>= xp_max_npartitions
);
605 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
607 if (part
->act_state
== XPC_P_AS_DEACTIVATING
) {
608 part
->act_state
= XPC_P_AS_INACTIVE
;
609 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
610 part
->remote_rp_pa
= 0;
614 /* indicate the thread is activating */
615 DBUG_ON(part
->act_state
!= XPC_P_AS_ACTIVATION_REQ
);
616 part
->act_state
= XPC_P_AS_ACTIVATING
;
618 XPC_SET_REASON(part
, 0, 0);
619 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
621 dev_dbg(xpc_part
, "activating partition %d\n", partid
);
623 xpc_allow_hb(partid
);
625 if (xpc_setup_ch_structures(part
) == xpSuccess
) {
626 (void)xpc_part_ref(part
); /* this will always succeed */
628 if (xpc_make_first_contact(part
) == xpSuccess
) {
629 xpc_mark_partition_active(part
);
630 xpc_channel_mgr(part
);
631 /* won't return until partition is deactivating */
634 xpc_part_deref(part
);
635 xpc_teardown_ch_structures(part
);
638 xpc_disallow_hb(partid
);
639 xpc_mark_partition_inactive(part
);
641 if (part
->reason
== xpReactivating
) {
642 /* interrupting ourselves results in activating partition */
643 xpc_request_partition_reactivation(part
);
650 xpc_activate_partition(struct xpc_partition
*part
)
652 short partid
= XPC_PARTID(part
);
653 unsigned long irq_flags
;
654 struct task_struct
*kthread
;
656 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
658 DBUG_ON(part
->act_state
!= XPC_P_AS_INACTIVE
);
660 part
->act_state
= XPC_P_AS_ACTIVATION_REQ
;
661 XPC_SET_REASON(part
, xpCloneKThread
, __LINE__
);
663 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
665 kthread
= kthread_run(xpc_activating
, (void *)((u64
)partid
), "xpc%02d",
667 if (IS_ERR(kthread
)) {
668 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
669 part
->act_state
= XPC_P_AS_INACTIVE
;
670 XPC_SET_REASON(part
, xpCloneKThreadFailed
, __LINE__
);
671 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
676 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
678 int idle
= atomic_read(&ch
->kthreads_idle
);
679 int assigned
= atomic_read(&ch
->kthreads_assigned
);
682 DBUG_ON(needed
<= 0);
685 wakeup
= (needed
> idle
) ? idle
: needed
;
688 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
689 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
691 /* only wakeup the requested number of kthreads */
692 wake_up_nr(&ch
->idle_wq
, wakeup
);
698 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
699 needed
= ch
->kthreads_assigned_limit
- assigned
;
704 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
705 needed
, ch
->partid
, ch
->number
);
707 xpc_create_kthreads(ch
, needed
, 0);
711 * This function is where XPC's kthreads wait for messages to deliver.
714 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
717 /* deliver messages to their intended recipients */
719 while (xpc_n_of_deliverable_payloads(ch
) > 0 &&
720 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
721 xpc_deliver_payload(ch
);
724 if (atomic_inc_return(&ch
->kthreads_idle
) >
725 ch
->kthreads_idle_limit
) {
726 /* too many idle kthreads on this channel */
727 atomic_dec(&ch
->kthreads_idle
);
731 dev_dbg(xpc_chan
, "idle kthread calling "
732 "wait_event_interruptible_exclusive()\n");
734 (void)wait_event_interruptible_exclusive(ch
->idle_wq
,
735 (xpc_n_of_deliverable_payloads(ch
) > 0 ||
736 (ch
->flags
& XPC_C_DISCONNECTING
)));
738 atomic_dec(&ch
->kthreads_idle
);
740 } while (!(ch
->flags
& XPC_C_DISCONNECTING
));
744 xpc_kthread_start(void *args
)
746 short partid
= XPC_UNPACK_ARG1(args
);
747 u16 ch_number
= XPC_UNPACK_ARG2(args
);
748 struct xpc_partition
*part
= &xpc_partitions
[partid
];
749 struct xpc_channel
*ch
;
751 unsigned long irq_flags
;
753 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
756 ch
= &part
->channels
[ch_number
];
758 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
760 /* let registerer know that connection has been established */
762 spin_lock_irqsave(&ch
->lock
, irq_flags
);
763 if (!(ch
->flags
& XPC_C_CONNECTEDCALLOUT
)) {
764 ch
->flags
|= XPC_C_CONNECTEDCALLOUT
;
765 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
767 xpc_connected_callout(ch
);
769 spin_lock_irqsave(&ch
->lock
, irq_flags
);
770 ch
->flags
|= XPC_C_CONNECTEDCALLOUT_MADE
;
771 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
774 * It is possible that while the callout was being
775 * made that the remote partition sent some messages.
776 * If that is the case, we may need to activate
777 * additional kthreads to help deliver them. We only
778 * need one less than total #of messages to deliver.
780 n_needed
= xpc_n_of_deliverable_payloads(ch
) - 1;
781 if (n_needed
> 0 && !(ch
->flags
& XPC_C_DISCONNECTING
))
782 xpc_activate_kthreads(ch
, n_needed
);
785 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
788 xpc_kthread_waitmsgs(part
, ch
);
791 /* let registerer know that connection is disconnecting */
793 spin_lock_irqsave(&ch
->lock
, irq_flags
);
794 if ((ch
->flags
& XPC_C_CONNECTEDCALLOUT_MADE
) &&
795 !(ch
->flags
& XPC_C_DISCONNECTINGCALLOUT
)) {
796 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT
;
797 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
799 xpc_disconnect_callout(ch
, xpDisconnecting
);
801 spin_lock_irqsave(&ch
->lock
, irq_flags
);
802 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT_MADE
;
804 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
806 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
807 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
808 xpc_indicate_partition_disengaged(part
);
811 xpc_msgqueue_deref(ch
);
813 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
816 xpc_part_deref(part
);
821 * For each partition that XPC has established communications with, there is
822 * a minimum of one kernel thread assigned to perform any operation that
823 * may potentially sleep or block (basically the callouts to the asynchronous
824 * functions registered via xpc_connect()).
826 * Additional kthreads are created and destroyed by XPC as the workload
829 * A kthread is assigned to one of the active channels that exists for a given
833 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
,
834 int ignore_disconnecting
)
836 unsigned long irq_flags
;
837 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
838 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
839 struct task_struct
*kthread
;
841 while (needed
-- > 0) {
844 * The following is done on behalf of the newly created
845 * kthread. That kthread is responsible for doing the
846 * counterpart to the following before it exits.
848 if (ignore_disconnecting
) {
849 if (!atomic_inc_not_zero(&ch
->kthreads_assigned
)) {
850 /* kthreads assigned had gone to zero */
852 XPC_C_DISCONNECTINGCALLOUT_MADE
));
856 } else if (ch
->flags
& XPC_C_DISCONNECTING
) {
859 } else if (atomic_inc_return(&ch
->kthreads_assigned
) == 1 &&
860 atomic_inc_return(&part
->nchannels_engaged
) == 1) {
861 xpc_indicate_partition_engaged(part
);
863 (void)xpc_part_ref(part
);
864 xpc_msgqueue_ref(ch
);
866 kthread
= kthread_run(xpc_kthread_start
, (void *)args
,
867 "xpc%02dc%d", ch
->partid
, ch
->number
);
868 if (IS_ERR(kthread
)) {
869 /* the fork failed */
872 * NOTE: if (ignore_disconnecting &&
873 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
874 * then we'll deadlock if all other kthreads assigned
875 * to this channel are blocked in the channel's
876 * registerer, because the only thing that will unblock
877 * them is the xpDisconnecting callout that this
878 * failed kthread_run() would have made.
881 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
882 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
883 xpc_indicate_partition_disengaged(part
);
885 xpc_msgqueue_deref(ch
);
886 xpc_part_deref(part
);
888 if (atomic_read(&ch
->kthreads_assigned
) <
889 ch
->kthreads_idle_limit
) {
891 * Flag this as an error only if we have an
892 * insufficient #of kthreads for the channel
895 spin_lock_irqsave(&ch
->lock
, irq_flags
);
896 XPC_DISCONNECT_CHANNEL(ch
, xpLackOfResources
,
898 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
906 xpc_disconnect_wait(int ch_number
)
908 unsigned long irq_flags
;
910 struct xpc_partition
*part
;
911 struct xpc_channel
*ch
;
912 int wakeup_channel_mgr
;
914 /* now wait for all callouts to the caller's function to cease */
915 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
916 part
= &xpc_partitions
[partid
];
918 if (!xpc_part_ref(part
))
921 ch
= &part
->channels
[ch_number
];
923 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
924 xpc_part_deref(part
);
928 wait_for_completion(&ch
->wdisconnect_wait
);
930 spin_lock_irqsave(&ch
->lock
, irq_flags
);
931 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
932 wakeup_channel_mgr
= 0;
934 if (ch
->delayed_chctl_flags
) {
935 if (part
->act_state
!= XPC_P_AS_DEACTIVATING
) {
936 spin_lock(&part
->chctl_lock
);
937 part
->chctl
.flags
[ch
->number
] |=
938 ch
->delayed_chctl_flags
;
939 spin_unlock(&part
->chctl_lock
);
940 wakeup_channel_mgr
= 1;
942 ch
->delayed_chctl_flags
= 0;
945 ch
->flags
&= ~XPC_C_WDISCONNECT
;
946 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
948 if (wakeup_channel_mgr
)
949 xpc_wakeup_channel_mgr(part
);
951 xpc_part_deref(part
);
956 xpc_setup_partitions(void)
959 struct xpc_partition
*part
;
961 xpc_partitions
= kzalloc(sizeof(struct xpc_partition
) *
962 xp_max_npartitions
, GFP_KERNEL
);
963 if (xpc_partitions
== NULL
) {
964 dev_err(xpc_part
, "can't get memory for partition structure\n");
969 * The first few fields of each entry of xpc_partitions[] need to
970 * be initialized now so that calls to xpc_connect() and
971 * xpc_disconnect() can be made prior to the activation of any remote
972 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
973 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
974 * PARTITION HAS BEEN ACTIVATED.
976 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
977 part
= &xpc_partitions
[partid
];
979 DBUG_ON((u64
)part
!= L1_CACHE_ALIGN((u64
)part
));
981 part
->activate_IRQ_rcvd
= 0;
982 spin_lock_init(&part
->act_lock
);
983 part
->act_state
= XPC_P_AS_INACTIVE
;
984 XPC_SET_REASON(part
, 0, 0);
986 init_timer(&part
->disengage_timer
);
987 part
->disengage_timer
.function
=
988 xpc_timeout_partition_disengage
;
989 part
->disengage_timer
.data
= (unsigned long)part
;
991 part
->setup_state
= XPC_P_SS_UNSET
;
992 init_waitqueue_head(&part
->teardown_wq
);
993 atomic_set(&part
->references
, 0);
996 return xpc_setup_partitions_sn();
1000 xpc_teardown_partitions(void)
1002 xpc_teardown_partitions_sn();
1003 kfree(xpc_partitions
);
1007 xpc_do_exit(enum xp_retval reason
)
1010 int active_part_count
, printed_waiting_msg
= 0;
1011 struct xpc_partition
*part
;
1012 unsigned long printmsg_time
, disengage_timeout
= 0;
1014 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
1015 DBUG_ON(xpc_exiting
== 1);
1018 * Let the heartbeat checker thread and the discovery thread
1019 * (if one is running) know that they should exit. Also wake up
1020 * the heartbeat checker thread in case it's sleeping.
1023 wake_up_interruptible(&xpc_activate_IRQ_wq
);
1025 /* wait for the discovery thread to exit */
1026 wait_for_completion(&xpc_discovery_exited
);
1028 /* wait for the heartbeat checker thread to exit */
1029 wait_for_completion(&xpc_hb_checker_exited
);
1031 /* sleep for a 1/3 of a second or so */
1032 (void)msleep_interruptible(300);
1034 /* wait for all partitions to become inactive */
1036 printmsg_time
= jiffies
+ (XPC_DEACTIVATE_PRINTMSG_INTERVAL
* HZ
);
1037 xpc_disengage_timedout
= 0;
1040 active_part_count
= 0;
1042 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1043 part
= &xpc_partitions
[partid
];
1045 if (xpc_partition_disengaged(part
) &&
1046 part
->act_state
== XPC_P_AS_INACTIVE
) {
1050 active_part_count
++;
1052 XPC_DEACTIVATE_PARTITION(part
, reason
);
1054 if (part
->disengage_timeout
> disengage_timeout
)
1055 disengage_timeout
= part
->disengage_timeout
;
1058 if (xpc_any_partition_engaged()) {
1059 if (time_is_before_jiffies(printmsg_time
)) {
1060 dev_info(xpc_part
, "waiting for remote "
1061 "partitions to deactivate, timeout in "
1062 "%ld seconds\n", (disengage_timeout
-
1064 printmsg_time
= jiffies
+
1065 (XPC_DEACTIVATE_PRINTMSG_INTERVAL
* HZ
);
1066 printed_waiting_msg
= 1;
1069 } else if (active_part_count
> 0) {
1070 if (printed_waiting_msg
) {
1071 dev_info(xpc_part
, "waiting for local partition"
1072 " to deactivate\n");
1073 printed_waiting_msg
= 0;
1077 if (!xpc_disengage_timedout
) {
1078 dev_info(xpc_part
, "all partitions have "
1084 /* sleep for a 1/3 of a second or so */
1085 (void)msleep_interruptible(300);
1089 DBUG_ON(xpc_any_partition_engaged());
1090 DBUG_ON(xpc_any_hbs_allowed() != 0);
1092 xpc_teardown_rsvd_page();
1094 if (reason
== xpUnloading
) {
1095 (void)unregister_die_notifier(&xpc_die_notifier
);
1096 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1099 /* clear the interface to XPC's functions */
1100 xpc_clear_interface();
1103 unregister_sysctl_table(xpc_sysctl
);
1105 xpc_teardown_partitions();
1114 * This function is called when the system is being rebooted.
1117 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1119 enum xp_retval reason
;
1123 reason
= xpSystemReboot
;
1126 reason
= xpSystemHalt
;
1129 reason
= xpSystemPoweroff
;
1132 reason
= xpSystemGoingDown
;
1135 xpc_do_exit(reason
);
1140 * Notify other partitions to deactivate from us by first disengaging from all
1141 * references to our memory.
1144 xpc_die_deactivate(void)
1146 struct xpc_partition
*part
;
1152 /* keep xpc_hb_checker thread from doing anything (just in case) */
1155 xpc_disallow_all_hbs(); /*indicate we're deactivated */
1157 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1158 part
= &xpc_partitions
[partid
];
1160 if (xpc_partition_engaged(partid
) ||
1161 part
->act_state
!= XPC_P_AS_INACTIVE
) {
1162 xpc_request_partition_deactivation(part
);
1163 xpc_indicate_partition_disengaged(part
);
1168 * Though we requested that all other partitions deactivate from us,
1169 * we only wait until they've all disengaged or we've reached the
1170 * defined timelimit.
1172 * Given that one iteration through the following while-loop takes
1173 * approximately 200 microseconds, calculate the #of loops to take
1174 * before bailing and the #of loops before printing a waiting message.
1176 keep_waiting
= xpc_disengage_timelimit
* 1000 * 5;
1177 wait_to_print
= XPC_DEACTIVATE_PRINTMSG_INTERVAL
* 1000 * 5;
1180 any_engaged
= xpc_any_partition_engaged();
1182 dev_info(xpc_part
, "all partitions have deactivated\n");
1186 if (!keep_waiting
--) {
1187 for (partid
= 0; partid
< xp_max_npartitions
;
1189 if (xpc_partition_engaged(partid
)) {
1190 dev_info(xpc_part
, "deactivate from "
1191 "remote partition %d timed "
1198 if (!wait_to_print
--) {
1199 dev_info(xpc_part
, "waiting for remote partitions to "
1200 "deactivate, timeout in %ld seconds\n",
1201 keep_waiting
/ (1000 * 5));
1202 wait_to_print
= XPC_DEACTIVATE_PRINTMSG_INTERVAL
*
1211 * This function is called when the system is being restarted or halted due
1212 * to some sort of system failure. If this is the case we need to notify the
1213 * other partitions to disengage from all references to our memory.
1214 * This function can also be called when our heartbeater could be offlined
1215 * for a time. In this case we need to notify other partitions to not worry
1216 * about the lack of a heartbeat.
1219 xpc_system_die(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1221 #ifdef CONFIG_IA64 /* !!! temporary kludge */
1223 case DIE_MACHINE_RESTART
:
1224 case DIE_MACHINE_HALT
:
1225 xpc_die_deactivate();
1228 case DIE_KDEBUG_ENTER
:
1229 /* Should lack of heartbeat be ignored by other partitions? */
1230 if (!xpc_kdebug_ignore
)
1234 case DIE_MCA_MONARCH_ENTER
:
1235 case DIE_INIT_MONARCH_ENTER
:
1236 xpc_offline_heartbeat();
1239 case DIE_KDEBUG_LEAVE
:
1240 /* Is lack of heartbeat being ignored by other partitions? */
1241 if (!xpc_kdebug_ignore
)
1245 case DIE_MCA_MONARCH_LEAVE
:
1246 case DIE_INIT_MONARCH_LEAVE
:
1247 xpc_online_heartbeat();
1251 xpc_die_deactivate();
1261 struct task_struct
*kthread
;
1263 dev_set_name(xpc_part
, "part");
1264 dev_set_name(xpc_chan
, "chan");
1268 * The ia64-sn2 architecture supports at most 64 partitions.
1269 * And the inability to unregister remote amos restricts us
1270 * further to only support exactly 64 partitions on this
1271 * architecture, no less.
1273 if (xp_max_npartitions
!= 64) {
1274 dev_err(xpc_part
, "max #of partitions not set to 64\n");
1277 ret
= xpc_init_sn2();
1280 } else if (is_uv()) {
1281 ret
= xpc_init_uv();
1290 ret
= xpc_setup_partitions();
1292 dev_err(xpc_part
, "can't get memory for partition structure\n");
1296 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
);
1299 * Fill the partition reserved page with the information needed by
1300 * other partitions to discover we are alive and establish initial
1303 ret
= xpc_setup_rsvd_page();
1305 dev_err(xpc_part
, "can't setup our reserved page\n");
1309 /* add ourselves to the reboot_notifier_list */
1310 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1312 dev_warn(xpc_part
, "can't register reboot notifier\n");
1314 /* add ourselves to the die_notifier list */
1315 ret
= register_die_notifier(&xpc_die_notifier
);
1317 dev_warn(xpc_part
, "can't register die notifier\n");
1320 * The real work-horse behind xpc. This processes incoming
1321 * interrupts and monitors remote heartbeats.
1323 kthread
= kthread_run(xpc_hb_checker
, NULL
, XPC_HB_CHECK_THREAD_NAME
);
1324 if (IS_ERR(kthread
)) {
1325 dev_err(xpc_part
, "failed while forking hb check thread\n");
1331 * Startup a thread that will attempt to discover other partitions to
1332 * activate based on info provided by SAL. This new thread is short
1333 * lived and will exit once discovery is complete.
1335 kthread
= kthread_run(xpc_initiate_discovery
, NULL
,
1336 XPC_DISCOVERY_THREAD_NAME
);
1337 if (IS_ERR(kthread
)) {
1338 dev_err(xpc_part
, "failed while forking discovery thread\n");
1340 /* mark this new thread as a non-starter */
1341 complete(&xpc_discovery_exited
);
1343 xpc_do_exit(xpUnloading
);
1347 /* set the interface to point at XPC's functions */
1348 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1349 xpc_initiate_send
, xpc_initiate_send_notify
,
1350 xpc_initiate_received
, xpc_initiate_partid_to_nasids
);
1354 /* initialization was not successful */
1356 xpc_teardown_rsvd_page();
1358 (void)unregister_die_notifier(&xpc_die_notifier
);
1359 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1362 unregister_sysctl_table(xpc_sysctl
);
1364 xpc_teardown_partitions();
1373 module_init(xpc_init
);
1378 xpc_do_exit(xpUnloading
);
1381 module_exit(xpc_exit
);
1383 MODULE_AUTHOR("Silicon Graphics, Inc.");
1384 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1385 MODULE_LICENSE("GPL");
1387 module_param(xpc_hb_interval
, int, 0);
1388 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1389 "heartbeat increments.");
1391 module_param(xpc_hb_check_interval
, int, 0);
1392 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1393 "heartbeat checks.");
1395 module_param(xpc_disengage_timelimit
, int, 0);
1396 MODULE_PARM_DESC(xpc_disengage_timelimit
, "Number of seconds to wait "
1397 "for disengage to complete.");
1399 module_param(xpc_kdebug_ignore
, int, 0);
1400 MODULE_PARM_DESC(xpc_kdebug_ignore
, "Should lack of heartbeat be ignored by "
1401 "other partitions when dropping into kdebug.");