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) 2004-2009 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
;
154 /* notification that the xpc_hb_checker thread has exited */
155 static DECLARE_COMPLETION(xpc_hb_checker_exited
);
157 /* notification that the xpc_discovery thread has exited */
158 static DECLARE_COMPLETION(xpc_discovery_exited
);
160 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
162 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
163 static struct notifier_block xpc_reboot_notifier
= {
164 .notifier_call
= xpc_system_reboot
,
167 static int xpc_system_die(struct notifier_block
*, unsigned long, void *);
168 static struct notifier_block xpc_die_notifier
= {
169 .notifier_call
= xpc_system_die
,
172 int (*xpc_setup_partitions_sn
) (void);
173 void (*xpc_teardown_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
);
179 void (*xpc_allow_hb
) (short partid
);
180 void (*xpc_disallow_hb
) (short partid
);
181 void (*xpc_disallow_all_hbs
) (void);
182 void (*xpc_heartbeat_init
) (void);
183 void (*xpc_heartbeat_exit
) (void);
184 void (*xpc_increment_heartbeat
) (void);
185 void (*xpc_offline_heartbeat
) (void);
186 void (*xpc_online_heartbeat
) (void);
187 enum xp_retval (*xpc_get_remote_heartbeat
) (struct xpc_partition
*part
);
189 enum xp_retval (*xpc_make_first_contact
) (struct xpc_partition
*part
);
190 void (*xpc_notify_senders_of_disconnect
) (struct xpc_channel
*ch
);
191 u64 (*xpc_get_chctl_all_flags
) (struct xpc_partition
*part
);
192 enum xp_retval (*xpc_setup_msg_structures
) (struct xpc_channel
*ch
);
193 void (*xpc_teardown_msg_structures
) (struct xpc_channel
*ch
);
194 void (*xpc_process_msg_chctl_flags
) (struct xpc_partition
*part
, int ch_number
);
195 int (*xpc_n_of_deliverable_payloads
) (struct xpc_channel
*ch
);
196 void *(*xpc_get_deliverable_payload
) (struct xpc_channel
*ch
);
198 void (*xpc_request_partition_activation
) (struct xpc_rsvd_page
*remote_rp
,
199 unsigned long remote_rp_pa
,
201 void (*xpc_request_partition_reactivation
) (struct xpc_partition
*part
);
202 void (*xpc_request_partition_deactivation
) (struct xpc_partition
*part
);
203 void (*xpc_cancel_partition_deactivation_request
) (struct xpc_partition
*part
);
205 void (*xpc_process_activate_IRQ_rcvd
) (void);
206 enum xp_retval (*xpc_setup_ch_structures_sn
) (struct xpc_partition
*part
);
207 void (*xpc_teardown_ch_structures_sn
) (struct xpc_partition
*part
);
209 void (*xpc_indicate_partition_engaged
) (struct xpc_partition
*part
);
210 int (*xpc_partition_engaged
) (short partid
);
211 int (*xpc_any_partition_engaged
) (void);
212 void (*xpc_indicate_partition_disengaged
) (struct xpc_partition
*part
);
213 void (*xpc_assume_partition_disengaged
) (short partid
);
215 void (*xpc_send_chctl_closerequest
) (struct xpc_channel
*ch
,
216 unsigned long *irq_flags
);
217 void (*xpc_send_chctl_closereply
) (struct xpc_channel
*ch
,
218 unsigned long *irq_flags
);
219 void (*xpc_send_chctl_openrequest
) (struct xpc_channel
*ch
,
220 unsigned long *irq_flags
);
221 void (*xpc_send_chctl_openreply
) (struct xpc_channel
*ch
,
222 unsigned long *irq_flags
);
224 enum xp_retval (*xpc_save_remote_msgqueue_pa
) (struct xpc_channel
*ch
,
225 unsigned long msgqueue_pa
);
227 enum xp_retval (*xpc_send_payload
) (struct xpc_channel
*ch
, u32 flags
,
228 void *payload
, u16 payload_size
,
229 u8 notify_type
, xpc_notify_func func
,
231 void (*xpc_received_payload
) (struct xpc_channel
*ch
, void *payload
);
234 * Timer function to enforce the timelimit on the partition disengage.
237 xpc_timeout_partition_disengage(unsigned long data
)
239 struct xpc_partition
*part
= (struct xpc_partition
*)data
;
241 DBUG_ON(time_is_after_jiffies(part
->disengage_timeout
));
243 (void)xpc_partition_disengaged(part
);
245 DBUG_ON(part
->disengage_timeout
!= 0);
246 DBUG_ON(xpc_partition_engaged(XPC_PARTID(part
)));
250 * Timer to produce the heartbeat. The timer structures function is
251 * already set when this is initially called. A tunable is used to
252 * specify when the next timeout should occur.
255 xpc_hb_beater(unsigned long dummy
)
257 xpc_increment_heartbeat();
259 if (time_is_before_eq_jiffies(xpc_hb_check_timeout
))
260 wake_up_interruptible(&xpc_activate_IRQ_wq
);
262 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
263 add_timer(&xpc_hb_timer
);
267 xpc_start_hb_beater(void)
269 xpc_heartbeat_init();
270 init_timer(&xpc_hb_timer
);
271 xpc_hb_timer
.function
= xpc_hb_beater
;
276 xpc_stop_hb_beater(void)
278 del_timer_sync(&xpc_hb_timer
);
279 xpc_heartbeat_exit();
283 * At periodic intervals, scan through all active partitions and ensure
284 * their heartbeat is still active. If not, the partition is deactivated.
287 xpc_check_remote_hb(void)
289 struct xpc_partition
*part
;
293 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
298 if (partid
== xp_partition_id
)
301 part
= &xpc_partitions
[partid
];
303 if (part
->act_state
== XPC_P_AS_INACTIVE
||
304 part
->act_state
== XPC_P_AS_DEACTIVATING
) {
308 ret
= xpc_get_remote_heartbeat(part
);
309 if (ret
!= xpSuccess
)
310 XPC_DEACTIVATE_PARTITION(part
, ret
);
315 * This thread is responsible for nearly all of the partition
316 * activation/deactivation.
319 xpc_hb_checker(void *ignore
)
323 /* this thread was marked active by xpc_hb_init() */
325 set_cpus_allowed_ptr(current
, cpumask_of(XPC_HB_CHECK_CPU
));
327 /* set our heartbeating to other partitions into motion */
328 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
329 xpc_start_hb_beater();
331 while (!xpc_exiting
) {
333 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
335 (int)(xpc_hb_check_timeout
- jiffies
),
336 xpc_activate_IRQ_rcvd
);
338 /* checking of remote heartbeats is skewed by IRQ handling */
339 if (time_is_before_eq_jiffies(xpc_hb_check_timeout
)) {
340 xpc_hb_check_timeout
= jiffies
+
341 (xpc_hb_check_interval
* HZ
);
343 dev_dbg(xpc_part
, "checking remote heartbeats\n");
344 xpc_check_remote_hb();
347 * On sn2 we need to periodically recheck to ensure no
348 * IRQ/amo pairs have been missed.
354 /* check for outstanding IRQs */
355 if (xpc_activate_IRQ_rcvd
> 0 || force_IRQ
!= 0) {
357 dev_dbg(xpc_part
, "processing activate IRQs "
359 xpc_process_activate_IRQ_rcvd();
362 /* wait for IRQ or timeout */
363 (void)wait_event_interruptible(xpc_activate_IRQ_wq
,
364 (time_is_before_eq_jiffies(
365 xpc_hb_check_timeout
) ||
366 xpc_activate_IRQ_rcvd
> 0 ||
370 xpc_stop_hb_beater();
372 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
374 /* mark this thread as having exited */
375 complete(&xpc_hb_checker_exited
);
380 * This thread will attempt to discover other partitions to activate
381 * based on info provided by SAL. This new thread is short lived and
382 * will exit once discovery is complete.
385 xpc_initiate_discovery(void *ignore
)
389 dev_dbg(xpc_part
, "discovery thread is exiting\n");
391 /* mark this thread as having exited */
392 complete(&xpc_discovery_exited
);
397 * The first kthread assigned to a newly activated partition is the one
398 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
399 * that kthread until the partition is brought down, at which time that kthread
400 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
401 * that XPC has dismantled all communication infrastructure for the associated
402 * partition.) This kthread becomes the channel manager for that partition.
404 * Each active partition has a channel manager, who, besides connecting and
405 * disconnecting channels, will ensure that each of the partition's connected
406 * channels has the required number of assigned kthreads to get the work done.
409 xpc_channel_mgr(struct xpc_partition
*part
)
411 while (part
->act_state
!= XPC_P_AS_DEACTIVATING
||
412 atomic_read(&part
->nchannels_active
) > 0 ||
413 !xpc_partition_disengaged(part
)) {
415 xpc_process_sent_chctl_flags(part
);
418 * Wait until we've been requested to activate kthreads or
419 * all of the channel's message queues have been torn down or
420 * a signal is pending.
422 * The channel_mgr_requests is set to 1 after being awakened,
423 * This is done to prevent the channel mgr from making one pass
424 * through the loop for each request, since he will
425 * be servicing all the requests in one pass. The reason it's
426 * set to 1 instead of 0 is so that other kthreads will know
427 * that the channel mgr is running and won't bother trying to
430 atomic_dec(&part
->channel_mgr_requests
);
431 (void)wait_event_interruptible(part
->channel_mgr_wq
,
432 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
433 part
->chctl
.all_flags
!= 0 ||
434 (part
->act_state
== XPC_P_AS_DEACTIVATING
&&
435 atomic_read(&part
->nchannels_active
) == 0 &&
436 xpc_partition_disengaged(part
))));
437 atomic_set(&part
->channel_mgr_requests
, 1);
442 * Guarantee that the kzalloc'd memory is cacheline aligned.
445 xpc_kzalloc_cacheline_aligned(size_t size
, gfp_t flags
, void **base
)
447 /* see if kzalloc will give us cachline aligned memory by default */
448 *base
= kzalloc(size
, flags
);
452 if ((u64
)*base
== L1_CACHE_ALIGN((u64
)*base
))
457 /* nope, we'll have to do it ourselves */
458 *base
= kzalloc(size
+ L1_CACHE_BYTES
, flags
);
462 return (void *)L1_CACHE_ALIGN((u64
)*base
);
466 * Setup the channel structures necessary to support XPartition Communication
467 * between the specified remote partition and the local one.
469 static enum xp_retval
470 xpc_setup_ch_structures(struct xpc_partition
*part
)
474 struct xpc_channel
*ch
;
475 short partid
= XPC_PARTID(part
);
478 * Allocate all of the channel structures as a contiguous chunk of
481 DBUG_ON(part
->channels
!= NULL
);
482 part
->channels
= kzalloc(sizeof(struct xpc_channel
) * XPC_MAX_NCHANNELS
,
484 if (part
->channels
== NULL
) {
485 dev_err(xpc_chan
, "can't get memory for channels\n");
489 /* allocate the remote open and close args */
491 part
->remote_openclose_args
=
492 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE
,
494 remote_openclose_args_base
);
495 if (part
->remote_openclose_args
== NULL
) {
496 dev_err(xpc_chan
, "can't get memory for remote connect args\n");
501 part
->chctl
.all_flags
= 0;
502 spin_lock_init(&part
->chctl_lock
);
504 atomic_set(&part
->channel_mgr_requests
, 1);
505 init_waitqueue_head(&part
->channel_mgr_wq
);
507 part
->nchannels
= XPC_MAX_NCHANNELS
;
509 atomic_set(&part
->nchannels_active
, 0);
510 atomic_set(&part
->nchannels_engaged
, 0);
512 for (ch_number
= 0; ch_number
< part
->nchannels
; ch_number
++) {
513 ch
= &part
->channels
[ch_number
];
516 ch
->number
= ch_number
;
517 ch
->flags
= XPC_C_DISCONNECTED
;
519 atomic_set(&ch
->kthreads_assigned
, 0);
520 atomic_set(&ch
->kthreads_idle
, 0);
521 atomic_set(&ch
->kthreads_active
, 0);
523 atomic_set(&ch
->references
, 0);
524 atomic_set(&ch
->n_to_notify
, 0);
526 spin_lock_init(&ch
->lock
);
527 init_completion(&ch
->wdisconnect_wait
);
529 atomic_set(&ch
->n_on_msg_allocate_wq
, 0);
530 init_waitqueue_head(&ch
->msg_allocate_wq
);
531 init_waitqueue_head(&ch
->idle_wq
);
534 ret
= xpc_setup_ch_structures_sn(part
);
535 if (ret
!= xpSuccess
)
539 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
540 * we're declaring that this partition is ready to go.
542 part
->setup_state
= XPC_P_SS_SETUP
;
546 /* setup of ch structures failed */
548 kfree(part
->remote_openclose_args_base
);
549 part
->remote_openclose_args
= NULL
;
551 kfree(part
->channels
);
552 part
->channels
= NULL
;
557 * Teardown the channel structures necessary to support XPartition Communication
558 * between the specified remote partition and the local one.
561 xpc_teardown_ch_structures(struct xpc_partition
*part
)
563 DBUG_ON(atomic_read(&part
->nchannels_engaged
) != 0);
564 DBUG_ON(atomic_read(&part
->nchannels_active
) != 0);
567 * Make this partition inaccessible to local processes by marking it
568 * as no longer setup. Then wait before proceeding with the teardown
569 * until all existing references cease.
571 DBUG_ON(part
->setup_state
!= XPC_P_SS_SETUP
);
572 part
->setup_state
= XPC_P_SS_WTEARDOWN
;
574 wait_event(part
->teardown_wq
, (atomic_read(&part
->references
) == 0));
576 /* now we can begin tearing down the infrastructure */
578 xpc_teardown_ch_structures_sn(part
);
580 kfree(part
->remote_openclose_args_base
);
581 part
->remote_openclose_args
= NULL
;
582 kfree(part
->channels
);
583 part
->channels
= NULL
;
585 part
->setup_state
= XPC_P_SS_TORNDOWN
;
589 * When XPC HB determines that a partition has come up, it will create a new
590 * kthread and that kthread will call this function to attempt to set up the
591 * basic infrastructure used for Cross Partition Communication with the newly
594 * The kthread that was created by XPC HB and which setup the XPC
595 * infrastructure will remain assigned to the partition becoming the channel
596 * manager for that partition until the partition is deactivating, at which
597 * time the kthread will teardown the XPC infrastructure and then exit.
600 xpc_activating(void *__partid
)
602 short partid
= (u64
)__partid
;
603 struct xpc_partition
*part
= &xpc_partitions
[partid
];
604 unsigned long irq_flags
;
606 DBUG_ON(partid
< 0 || partid
>= xp_max_npartitions
);
608 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
610 if (part
->act_state
== XPC_P_AS_DEACTIVATING
) {
611 part
->act_state
= XPC_P_AS_INACTIVE
;
612 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
613 part
->remote_rp_pa
= 0;
617 /* indicate the thread is activating */
618 DBUG_ON(part
->act_state
!= XPC_P_AS_ACTIVATION_REQ
);
619 part
->act_state
= XPC_P_AS_ACTIVATING
;
621 XPC_SET_REASON(part
, 0, 0);
622 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
624 dev_dbg(xpc_part
, "activating partition %d\n", partid
);
626 xpc_allow_hb(partid
);
628 if (xpc_setup_ch_structures(part
) == xpSuccess
) {
629 (void)xpc_part_ref(part
); /* this will always succeed */
631 if (xpc_make_first_contact(part
) == xpSuccess
) {
632 xpc_mark_partition_active(part
);
633 xpc_channel_mgr(part
);
634 /* won't return until partition is deactivating */
637 xpc_part_deref(part
);
638 xpc_teardown_ch_structures(part
);
641 xpc_disallow_hb(partid
);
642 xpc_mark_partition_inactive(part
);
644 if (part
->reason
== xpReactivating
) {
645 /* interrupting ourselves results in activating partition */
646 xpc_request_partition_reactivation(part
);
653 xpc_activate_partition(struct xpc_partition
*part
)
655 short partid
= XPC_PARTID(part
);
656 unsigned long irq_flags
;
657 struct task_struct
*kthread
;
659 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
661 DBUG_ON(part
->act_state
!= XPC_P_AS_INACTIVE
);
663 part
->act_state
= XPC_P_AS_ACTIVATION_REQ
;
664 XPC_SET_REASON(part
, xpCloneKThread
, __LINE__
);
666 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
668 kthread
= kthread_run(xpc_activating
, (void *)((u64
)partid
), "xpc%02d",
670 if (IS_ERR(kthread
)) {
671 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
672 part
->act_state
= XPC_P_AS_INACTIVE
;
673 XPC_SET_REASON(part
, xpCloneKThreadFailed
, __LINE__
);
674 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
679 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
681 int idle
= atomic_read(&ch
->kthreads_idle
);
682 int assigned
= atomic_read(&ch
->kthreads_assigned
);
685 DBUG_ON(needed
<= 0);
688 wakeup
= (needed
> idle
) ? idle
: needed
;
691 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
692 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
694 /* only wakeup the requested number of kthreads */
695 wake_up_nr(&ch
->idle_wq
, wakeup
);
701 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
702 needed
= ch
->kthreads_assigned_limit
- assigned
;
707 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
708 needed
, ch
->partid
, ch
->number
);
710 xpc_create_kthreads(ch
, needed
, 0);
714 * This function is where XPC's kthreads wait for messages to deliver.
717 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
720 /* deliver messages to their intended recipients */
722 while (xpc_n_of_deliverable_payloads(ch
) > 0 &&
723 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
724 xpc_deliver_payload(ch
);
727 if (atomic_inc_return(&ch
->kthreads_idle
) >
728 ch
->kthreads_idle_limit
) {
729 /* too many idle kthreads on this channel */
730 atomic_dec(&ch
->kthreads_idle
);
734 dev_dbg(xpc_chan
, "idle kthread calling "
735 "wait_event_interruptible_exclusive()\n");
737 (void)wait_event_interruptible_exclusive(ch
->idle_wq
,
738 (xpc_n_of_deliverable_payloads(ch
) > 0 ||
739 (ch
->flags
& XPC_C_DISCONNECTING
)));
741 atomic_dec(&ch
->kthreads_idle
);
743 } while (!(ch
->flags
& XPC_C_DISCONNECTING
));
747 xpc_kthread_start(void *args
)
749 short partid
= XPC_UNPACK_ARG1(args
);
750 u16 ch_number
= XPC_UNPACK_ARG2(args
);
751 struct xpc_partition
*part
= &xpc_partitions
[partid
];
752 struct xpc_channel
*ch
;
754 unsigned long irq_flags
;
756 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
759 ch
= &part
->channels
[ch_number
];
761 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
763 /* let registerer know that connection has been established */
765 spin_lock_irqsave(&ch
->lock
, irq_flags
);
766 if (!(ch
->flags
& XPC_C_CONNECTEDCALLOUT
)) {
767 ch
->flags
|= XPC_C_CONNECTEDCALLOUT
;
768 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
770 xpc_connected_callout(ch
);
772 spin_lock_irqsave(&ch
->lock
, irq_flags
);
773 ch
->flags
|= XPC_C_CONNECTEDCALLOUT_MADE
;
774 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
777 * It is possible that while the callout was being
778 * made that the remote partition sent some messages.
779 * If that is the case, we may need to activate
780 * additional kthreads to help deliver them. We only
781 * need one less than total #of messages to deliver.
783 n_needed
= xpc_n_of_deliverable_payloads(ch
) - 1;
784 if (n_needed
> 0 && !(ch
->flags
& XPC_C_DISCONNECTING
))
785 xpc_activate_kthreads(ch
, n_needed
);
788 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
791 xpc_kthread_waitmsgs(part
, ch
);
794 /* let registerer know that connection is disconnecting */
796 spin_lock_irqsave(&ch
->lock
, irq_flags
);
797 if ((ch
->flags
& XPC_C_CONNECTEDCALLOUT_MADE
) &&
798 !(ch
->flags
& XPC_C_DISCONNECTINGCALLOUT
)) {
799 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT
;
800 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
802 xpc_disconnect_callout(ch
, xpDisconnecting
);
804 spin_lock_irqsave(&ch
->lock
, irq_flags
);
805 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT_MADE
;
807 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
809 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
810 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
811 xpc_indicate_partition_disengaged(part
);
814 xpc_msgqueue_deref(ch
);
816 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
819 xpc_part_deref(part
);
824 * For each partition that XPC has established communications with, there is
825 * a minimum of one kernel thread assigned to perform any operation that
826 * may potentially sleep or block (basically the callouts to the asynchronous
827 * functions registered via xpc_connect()).
829 * Additional kthreads are created and destroyed by XPC as the workload
832 * A kthread is assigned to one of the active channels that exists for a given
836 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
,
837 int ignore_disconnecting
)
839 unsigned long irq_flags
;
840 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
841 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
842 struct task_struct
*kthread
;
844 while (needed
-- > 0) {
847 * The following is done on behalf of the newly created
848 * kthread. That kthread is responsible for doing the
849 * counterpart to the following before it exits.
851 if (ignore_disconnecting
) {
852 if (!atomic_inc_not_zero(&ch
->kthreads_assigned
)) {
853 /* kthreads assigned had gone to zero */
855 XPC_C_DISCONNECTINGCALLOUT_MADE
));
859 } else if (ch
->flags
& XPC_C_DISCONNECTING
) {
862 } else if (atomic_inc_return(&ch
->kthreads_assigned
) == 1 &&
863 atomic_inc_return(&part
->nchannels_engaged
) == 1) {
864 xpc_indicate_partition_engaged(part
);
866 (void)xpc_part_ref(part
);
867 xpc_msgqueue_ref(ch
);
869 kthread
= kthread_run(xpc_kthread_start
, (void *)args
,
870 "xpc%02dc%d", ch
->partid
, ch
->number
);
871 if (IS_ERR(kthread
)) {
872 /* the fork failed */
875 * NOTE: if (ignore_disconnecting &&
876 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
877 * then we'll deadlock if all other kthreads assigned
878 * to this channel are blocked in the channel's
879 * registerer, because the only thing that will unblock
880 * them is the xpDisconnecting callout that this
881 * failed kthread_run() would have made.
884 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
885 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
886 xpc_indicate_partition_disengaged(part
);
888 xpc_msgqueue_deref(ch
);
889 xpc_part_deref(part
);
891 if (atomic_read(&ch
->kthreads_assigned
) <
892 ch
->kthreads_idle_limit
) {
894 * Flag this as an error only if we have an
895 * insufficient #of kthreads for the channel
898 spin_lock_irqsave(&ch
->lock
, irq_flags
);
899 XPC_DISCONNECT_CHANNEL(ch
, xpLackOfResources
,
901 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
909 xpc_disconnect_wait(int ch_number
)
911 unsigned long irq_flags
;
913 struct xpc_partition
*part
;
914 struct xpc_channel
*ch
;
915 int wakeup_channel_mgr
;
917 /* now wait for all callouts to the caller's function to cease */
918 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
919 part
= &xpc_partitions
[partid
];
921 if (!xpc_part_ref(part
))
924 ch
= &part
->channels
[ch_number
];
926 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
927 xpc_part_deref(part
);
931 wait_for_completion(&ch
->wdisconnect_wait
);
933 spin_lock_irqsave(&ch
->lock
, irq_flags
);
934 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
935 wakeup_channel_mgr
= 0;
937 if (ch
->delayed_chctl_flags
) {
938 if (part
->act_state
!= XPC_P_AS_DEACTIVATING
) {
939 spin_lock(&part
->chctl_lock
);
940 part
->chctl
.flags
[ch
->number
] |=
941 ch
->delayed_chctl_flags
;
942 spin_unlock(&part
->chctl_lock
);
943 wakeup_channel_mgr
= 1;
945 ch
->delayed_chctl_flags
= 0;
948 ch
->flags
&= ~XPC_C_WDISCONNECT
;
949 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
951 if (wakeup_channel_mgr
)
952 xpc_wakeup_channel_mgr(part
);
954 xpc_part_deref(part
);
959 xpc_setup_partitions(void)
962 struct xpc_partition
*part
;
964 xpc_partitions
= kzalloc(sizeof(struct xpc_partition
) *
965 xp_max_npartitions
, GFP_KERNEL
);
966 if (xpc_partitions
== NULL
) {
967 dev_err(xpc_part
, "can't get memory for partition structure\n");
972 * The first few fields of each entry of xpc_partitions[] need to
973 * be initialized now so that calls to xpc_connect() and
974 * xpc_disconnect() can be made prior to the activation of any remote
975 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
976 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
977 * PARTITION HAS BEEN ACTIVATED.
979 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
980 part
= &xpc_partitions
[partid
];
982 DBUG_ON((u64
)part
!= L1_CACHE_ALIGN((u64
)part
));
984 part
->activate_IRQ_rcvd
= 0;
985 spin_lock_init(&part
->act_lock
);
986 part
->act_state
= XPC_P_AS_INACTIVE
;
987 XPC_SET_REASON(part
, 0, 0);
989 init_timer(&part
->disengage_timer
);
990 part
->disengage_timer
.function
=
991 xpc_timeout_partition_disengage
;
992 part
->disengage_timer
.data
= (unsigned long)part
;
994 part
->setup_state
= XPC_P_SS_UNSET
;
995 init_waitqueue_head(&part
->teardown_wq
);
996 atomic_set(&part
->references
, 0);
999 return xpc_setup_partitions_sn();
1003 xpc_teardown_partitions(void)
1005 xpc_teardown_partitions_sn();
1006 kfree(xpc_partitions
);
1010 xpc_do_exit(enum xp_retval reason
)
1013 int active_part_count
, printed_waiting_msg
= 0;
1014 struct xpc_partition
*part
;
1015 unsigned long printmsg_time
, disengage_timeout
= 0;
1017 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
1018 DBUG_ON(xpc_exiting
== 1);
1021 * Let the heartbeat checker thread and the discovery thread
1022 * (if one is running) know that they should exit. Also wake up
1023 * the heartbeat checker thread in case it's sleeping.
1026 wake_up_interruptible(&xpc_activate_IRQ_wq
);
1028 /* wait for the discovery thread to exit */
1029 wait_for_completion(&xpc_discovery_exited
);
1031 /* wait for the heartbeat checker thread to exit */
1032 wait_for_completion(&xpc_hb_checker_exited
);
1034 /* sleep for a 1/3 of a second or so */
1035 (void)msleep_interruptible(300);
1037 /* wait for all partitions to become inactive */
1039 printmsg_time
= jiffies
+ (XPC_DEACTIVATE_PRINTMSG_INTERVAL
* HZ
);
1040 xpc_disengage_timedout
= 0;
1043 active_part_count
= 0;
1045 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1046 part
= &xpc_partitions
[partid
];
1048 if (xpc_partition_disengaged(part
) &&
1049 part
->act_state
== XPC_P_AS_INACTIVE
) {
1053 active_part_count
++;
1055 XPC_DEACTIVATE_PARTITION(part
, reason
);
1057 if (part
->disengage_timeout
> disengage_timeout
)
1058 disengage_timeout
= part
->disengage_timeout
;
1061 if (xpc_any_partition_engaged()) {
1062 if (time_is_before_jiffies(printmsg_time
)) {
1063 dev_info(xpc_part
, "waiting for remote "
1064 "partitions to deactivate, timeout in "
1065 "%ld seconds\n", (disengage_timeout
-
1067 printmsg_time
= jiffies
+
1068 (XPC_DEACTIVATE_PRINTMSG_INTERVAL
* HZ
);
1069 printed_waiting_msg
= 1;
1072 } else if (active_part_count
> 0) {
1073 if (printed_waiting_msg
) {
1074 dev_info(xpc_part
, "waiting for local partition"
1075 " to deactivate\n");
1076 printed_waiting_msg
= 0;
1080 if (!xpc_disengage_timedout
) {
1081 dev_info(xpc_part
, "all partitions have "
1087 /* sleep for a 1/3 of a second or so */
1088 (void)msleep_interruptible(300);
1092 DBUG_ON(xpc_any_partition_engaged());
1094 xpc_teardown_rsvd_page();
1096 if (reason
== xpUnloading
) {
1097 (void)unregister_die_notifier(&xpc_die_notifier
);
1098 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1101 /* clear the interface to XPC's functions */
1102 xpc_clear_interface();
1105 unregister_sysctl_table(xpc_sysctl
);
1107 xpc_teardown_partitions();
1116 * This function is called when the system is being rebooted.
1119 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1121 enum xp_retval reason
;
1125 reason
= xpSystemReboot
;
1128 reason
= xpSystemHalt
;
1131 reason
= xpSystemPoweroff
;
1134 reason
= xpSystemGoingDown
;
1137 xpc_do_exit(reason
);
1142 * Notify other partitions to deactivate from us by first disengaging from all
1143 * references to our memory.
1146 xpc_die_deactivate(void)
1148 struct xpc_partition
*part
;
1154 /* keep xpc_hb_checker thread from doing anything (just in case) */
1157 xpc_disallow_all_hbs(); /*indicate we're deactivated */
1159 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1160 part
= &xpc_partitions
[partid
];
1162 if (xpc_partition_engaged(partid
) ||
1163 part
->act_state
!= XPC_P_AS_INACTIVE
) {
1164 xpc_request_partition_deactivation(part
);
1165 xpc_indicate_partition_disengaged(part
);
1170 * Though we requested that all other partitions deactivate from us,
1171 * we only wait until they've all disengaged or we've reached the
1172 * defined timelimit.
1174 * Given that one iteration through the following while-loop takes
1175 * approximately 200 microseconds, calculate the #of loops to take
1176 * before bailing and the #of loops before printing a waiting message.
1178 keep_waiting
= xpc_disengage_timelimit
* 1000 * 5;
1179 wait_to_print
= XPC_DEACTIVATE_PRINTMSG_INTERVAL
* 1000 * 5;
1182 any_engaged
= xpc_any_partition_engaged();
1184 dev_info(xpc_part
, "all partitions have deactivated\n");
1188 if (!keep_waiting
--) {
1189 for (partid
= 0; partid
< xp_max_npartitions
;
1191 if (xpc_partition_engaged(partid
)) {
1192 dev_info(xpc_part
, "deactivate from "
1193 "remote partition %d timed "
1200 if (!wait_to_print
--) {
1201 dev_info(xpc_part
, "waiting for remote partitions to "
1202 "deactivate, timeout in %ld seconds\n",
1203 keep_waiting
/ (1000 * 5));
1204 wait_to_print
= XPC_DEACTIVATE_PRINTMSG_INTERVAL
*
1213 * This function is called when the system is being restarted or halted due
1214 * to some sort of system failure. If this is the case we need to notify the
1215 * other partitions to disengage from all references to our memory.
1216 * This function can also be called when our heartbeater could be offlined
1217 * for a time. In this case we need to notify other partitions to not worry
1218 * about the lack of a heartbeat.
1221 xpc_system_die(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1223 #ifdef CONFIG_IA64 /* !!! temporary kludge */
1225 case DIE_MACHINE_RESTART
:
1226 case DIE_MACHINE_HALT
:
1227 xpc_die_deactivate();
1230 case DIE_KDEBUG_ENTER
:
1231 /* Should lack of heartbeat be ignored by other partitions? */
1232 if (!xpc_kdebug_ignore
)
1236 case DIE_MCA_MONARCH_ENTER
:
1237 case DIE_INIT_MONARCH_ENTER
:
1238 xpc_offline_heartbeat();
1241 case DIE_KDEBUG_LEAVE
:
1242 /* Is lack of heartbeat being ignored by other partitions? */
1243 if (!xpc_kdebug_ignore
)
1247 case DIE_MCA_MONARCH_LEAVE
:
1248 case DIE_INIT_MONARCH_LEAVE
:
1249 xpc_online_heartbeat();
1253 xpc_die_deactivate();
1263 struct task_struct
*kthread
;
1265 dev_set_name(xpc_part
, "part");
1266 dev_set_name(xpc_chan
, "chan");
1270 * The ia64-sn2 architecture supports at most 64 partitions.
1271 * And the inability to unregister remote amos restricts us
1272 * further to only support exactly 64 partitions on this
1273 * architecture, no less.
1275 if (xp_max_npartitions
!= 64) {
1276 dev_err(xpc_part
, "max #of partitions not set to 64\n");
1279 ret
= xpc_init_sn2();
1282 } else if (is_uv()) {
1283 ret
= xpc_init_uv();
1292 ret
= xpc_setup_partitions();
1294 dev_err(xpc_part
, "can't get memory for partition structure\n");
1298 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
);
1301 * Fill the partition reserved page with the information needed by
1302 * other partitions to discover we are alive and establish initial
1305 ret
= xpc_setup_rsvd_page();
1307 dev_err(xpc_part
, "can't setup our reserved page\n");
1311 /* add ourselves to the reboot_notifier_list */
1312 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1314 dev_warn(xpc_part
, "can't register reboot notifier\n");
1316 /* add ourselves to the die_notifier list */
1317 ret
= register_die_notifier(&xpc_die_notifier
);
1319 dev_warn(xpc_part
, "can't register die notifier\n");
1322 * The real work-horse behind xpc. This processes incoming
1323 * interrupts and monitors remote heartbeats.
1325 kthread
= kthread_run(xpc_hb_checker
, NULL
, XPC_HB_CHECK_THREAD_NAME
);
1326 if (IS_ERR(kthread
)) {
1327 dev_err(xpc_part
, "failed while forking hb check thread\n");
1333 * Startup a thread that will attempt to discover other partitions to
1334 * activate based on info provided by SAL. This new thread is short
1335 * lived and will exit once discovery is complete.
1337 kthread
= kthread_run(xpc_initiate_discovery
, NULL
,
1338 XPC_DISCOVERY_THREAD_NAME
);
1339 if (IS_ERR(kthread
)) {
1340 dev_err(xpc_part
, "failed while forking discovery thread\n");
1342 /* mark this new thread as a non-starter */
1343 complete(&xpc_discovery_exited
);
1345 xpc_do_exit(xpUnloading
);
1349 /* set the interface to point at XPC's functions */
1350 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1351 xpc_initiate_send
, xpc_initiate_send_notify
,
1352 xpc_initiate_received
, xpc_initiate_partid_to_nasids
);
1356 /* initialization was not successful */
1358 xpc_teardown_rsvd_page();
1360 (void)unregister_die_notifier(&xpc_die_notifier
);
1361 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1364 unregister_sysctl_table(xpc_sysctl
);
1366 xpc_teardown_partitions();
1375 module_init(xpc_init
);
1380 xpc_do_exit(xpUnloading
);
1383 module_exit(xpc_exit
);
1385 MODULE_AUTHOR("Silicon Graphics, Inc.");
1386 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1387 MODULE_LICENSE("GPL");
1389 module_param(xpc_hb_interval
, int, 0);
1390 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1391 "heartbeat increments.");
1393 module_param(xpc_hb_check_interval
, int, 0);
1394 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1395 "heartbeat checks.");
1397 module_param(xpc_disengage_timelimit
, int, 0);
1398 MODULE_PARM_DESC(xpc_disengage_timelimit
, "Number of seconds to wait "
1399 "for disengage to complete.");
1401 module_param(xpc_kdebug_ignore
, int, 0);
1402 MODULE_PARM_DESC(xpc_kdebug_ignore
, "Should lack of heartbeat be ignored by "
1403 "other partitions when dropping into kdebug.");