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 * . We currently have no way to determine which nasid an IPI came
29 * from. Thus, xpc_IPI_send() does a remote AMO write followed by
30 * an IPI. The AMO indicates where data is to be pulled from, so
31 * after the IPI arrives, the remote partition checks the AMO word.
32 * The IPI can actually arrive before the AMO however, so other code
33 * must periodically check for this case. Also, remote AMO operations
34 * do not reliably time out. Thus we do a remote PIO read solely to
35 * know whether the remote partition is down and whether we should
36 * stop sending IPIs to it. This remote PIO read operation is set up
37 * in a special nofault region so SAL knows to ignore (and cleanup)
38 * any errors due to the remote AMO write, PIO read, and/or PIO
41 * If/when new hardware solves this IPI problem, we should abandon
42 * the current approach.
46 #include <linux/kernel.h>
47 #include <linux/module.h>
48 #include <linux/init.h>
49 #include <linux/cache.h>
50 #include <linux/interrupt.h>
51 #include <linux/delay.h>
52 #include <linux/reboot.h>
53 #include <linux/completion.h>
54 #include <linux/kdebug.h>
55 #include <linux/kthread.h>
56 #include <linux/uaccess.h>
57 #include <asm/sn/intr.h>
58 #include <asm/sn/sn_sal.h>
61 /* define two XPC debug device structures to be used with dev_dbg() et al */
63 struct device_driver xpc_dbg_name
= {
67 struct device xpc_part_dbg_subname
= {
68 .bus_id
= {0}, /* set to "part" at xpc_init() time */
69 .driver
= &xpc_dbg_name
72 struct device xpc_chan_dbg_subname
= {
73 .bus_id
= {0}, /* set to "chan" at xpc_init() time */
74 .driver
= &xpc_dbg_name
77 struct device
*xpc_part
= &xpc_part_dbg_subname
;
78 struct device
*xpc_chan
= &xpc_chan_dbg_subname
;
80 static int xpc_kdebug_ignore
;
82 /* systune related variables for /proc/sys directories */
84 static int xpc_hb_interval
= XPC_HB_DEFAULT_INTERVAL
;
85 static int xpc_hb_min_interval
= 1;
86 static int xpc_hb_max_interval
= 10;
88 static int xpc_hb_check_interval
= XPC_HB_CHECK_DEFAULT_INTERVAL
;
89 static int xpc_hb_check_min_interval
= 10;
90 static int xpc_hb_check_max_interval
= 120;
92 int xpc_disengage_request_timelimit
= XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT
;
93 static int xpc_disengage_request_min_timelimit
; /* = 0 */
94 static int xpc_disengage_request_max_timelimit
= 120;
96 static ctl_table xpc_sys_xpc_hb_dir
[] = {
98 .ctl_name
= CTL_UNNUMBERED
,
99 .procname
= "hb_interval",
100 .data
= &xpc_hb_interval
,
101 .maxlen
= sizeof(int),
103 .proc_handler
= &proc_dointvec_minmax
,
104 .strategy
= &sysctl_intvec
,
105 .extra1
= &xpc_hb_min_interval
,
106 .extra2
= &xpc_hb_max_interval
},
108 .ctl_name
= CTL_UNNUMBERED
,
109 .procname
= "hb_check_interval",
110 .data
= &xpc_hb_check_interval
,
111 .maxlen
= sizeof(int),
113 .proc_handler
= &proc_dointvec_minmax
,
114 .strategy
= &sysctl_intvec
,
115 .extra1
= &xpc_hb_check_min_interval
,
116 .extra2
= &xpc_hb_check_max_interval
},
119 static ctl_table xpc_sys_xpc_dir
[] = {
121 .ctl_name
= CTL_UNNUMBERED
,
124 .child
= xpc_sys_xpc_hb_dir
},
126 .ctl_name
= CTL_UNNUMBERED
,
127 .procname
= "disengage_request_timelimit",
128 .data
= &xpc_disengage_request_timelimit
,
129 .maxlen
= sizeof(int),
131 .proc_handler
= &proc_dointvec_minmax
,
132 .strategy
= &sysctl_intvec
,
133 .extra1
= &xpc_disengage_request_min_timelimit
,
134 .extra2
= &xpc_disengage_request_max_timelimit
},
137 static ctl_table xpc_sys_dir
[] = {
139 .ctl_name
= CTL_UNNUMBERED
,
142 .child
= xpc_sys_xpc_dir
},
145 static struct ctl_table_header
*xpc_sysctl
;
147 /* non-zero if any remote partition disengage request was timed out */
148 int xpc_disengage_request_timedout
;
150 /* #of IRQs received */
151 static atomic_t xpc_act_IRQ_rcvd
;
153 /* IRQ handler notifies this wait queue on receipt of an IRQ */
154 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq
);
156 static unsigned long xpc_hb_check_timeout
;
158 /* notification that the xpc_hb_checker thread has exited */
159 static DECLARE_COMPLETION(xpc_hb_checker_exited
);
161 /* notification that the xpc_discovery thread has exited */
162 static DECLARE_COMPLETION(xpc_discovery_exited
);
164 static struct timer_list xpc_hb_timer
;
166 static void xpc_kthread_waitmsgs(struct xpc_partition
*, struct xpc_channel
*);
168 static int xpc_system_reboot(struct notifier_block
*, unsigned long, void *);
169 static struct notifier_block xpc_reboot_notifier
= {
170 .notifier_call
= xpc_system_reboot
,
173 static int xpc_system_die(struct notifier_block
*, unsigned long, void *);
174 static struct notifier_block xpc_die_notifier
= {
175 .notifier_call
= xpc_system_die
,
178 enum xp_retval (*xpc_rsvd_page_init
) (struct xpc_rsvd_page
*rp
);
179 enum xp_retval (*xpc_make_first_contact
) (struct xpc_partition
*part
);
180 u64 (*xpc_get_IPI_flags
) (struct xpc_partition
*part
);
181 struct xpc_msg
*(*xpc_get_deliverable_msg
) (struct xpc_channel
*ch
);
182 enum xp_retval (*xpc_setup_infrastructure
) (struct xpc_partition
*part
);
183 void (*xpc_teardown_infrastructure
) (struct xpc_partition
*part
);
187 * Timer function to enforce the timelimit on the partition disengage request.
190 xpc_timeout_partition_disengage_request(unsigned long data
)
192 struct xpc_partition
*part
= (struct xpc_partition
*)data
;
194 DBUG_ON(time_before(jiffies
, part
->disengage_request_timeout
));
196 (void)xpc_partition_disengaged(part
);
198 DBUG_ON(part
->disengage_request_timeout
!= 0);
199 DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part
)) != 0);
203 * Notify the heartbeat check thread that an IRQ has been received.
206 xpc_act_IRQ_handler(int irq
, void *dev_id
)
208 atomic_inc(&xpc_act_IRQ_rcvd
);
209 wake_up_interruptible(&xpc_act_IRQ_wq
);
214 * Timer to produce the heartbeat. The timer structures function is
215 * already set when this is initially called. A tunable is used to
216 * specify when the next timeout should occur.
219 xpc_hb_beater(unsigned long dummy
)
221 xpc_vars
->heartbeat
++;
223 if (time_after_eq(jiffies
, xpc_hb_check_timeout
))
224 wake_up_interruptible(&xpc_act_IRQ_wq
);
226 xpc_hb_timer
.expires
= jiffies
+ (xpc_hb_interval
* HZ
);
227 add_timer(&xpc_hb_timer
);
231 * This thread is responsible for nearly all of the partition
232 * activation/deactivation.
235 xpc_hb_checker(void *ignore
)
237 int last_IRQ_count
= 0;
241 /* this thread was marked active by xpc_hb_init() */
243 set_cpus_allowed_ptr(current
, &cpumask_of_cpu(XPC_HB_CHECK_CPU
));
245 /* set our heartbeating to other partitions into motion */
246 xpc_hb_check_timeout
= jiffies
+ (xpc_hb_check_interval
* HZ
);
249 while (!xpc_exiting
) {
251 dev_dbg(xpc_part
, "woke up with %d ticks rem; %d IRQs have "
253 (int)(xpc_hb_check_timeout
- jiffies
),
254 atomic_read(&xpc_act_IRQ_rcvd
) - last_IRQ_count
);
256 /* checking of remote heartbeats is skewed by IRQ handling */
257 if (time_after_eq(jiffies
, xpc_hb_check_timeout
)) {
258 dev_dbg(xpc_part
, "checking remote heartbeats\n");
259 xpc_check_remote_hb();
262 * We need to periodically recheck to ensure no
263 * IPI/AMO pairs have been missed. That check
264 * must always reset xpc_hb_check_timeout.
269 /* check for outstanding IRQs */
270 new_IRQ_count
= atomic_read(&xpc_act_IRQ_rcvd
);
271 if (last_IRQ_count
< new_IRQ_count
|| force_IRQ
!= 0) {
274 dev_dbg(xpc_part
, "found an IRQ to process; will be "
275 "resetting xpc_hb_check_timeout\n");
277 last_IRQ_count
+= xpc_identify_act_IRQ_sender();
278 if (last_IRQ_count
< new_IRQ_count
) {
279 /* retry once to help avoid missing AMO */
280 (void)xpc_identify_act_IRQ_sender();
282 last_IRQ_count
= new_IRQ_count
;
284 xpc_hb_check_timeout
= jiffies
+
285 (xpc_hb_check_interval
* HZ
);
288 /* wait for IRQ or timeout */
289 (void)wait_event_interruptible(xpc_act_IRQ_wq
,
291 atomic_read(&xpc_act_IRQ_rcvd
)
292 || time_after_eq(jiffies
,
293 xpc_hb_check_timeout
) ||
297 dev_dbg(xpc_part
, "heartbeat checker is exiting\n");
299 /* mark this thread as having exited */
300 complete(&xpc_hb_checker_exited
);
305 * This thread will attempt to discover other partitions to activate
306 * based on info provided by SAL. This new thread is short lived and
307 * will exit once discovery is complete.
310 xpc_initiate_discovery(void *ignore
)
314 dev_dbg(xpc_part
, "discovery thread is exiting\n");
316 /* mark this thread as having exited */
317 complete(&xpc_discovery_exited
);
322 * The first kthread assigned to a newly activated partition is the one
323 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
324 * that kthread until the partition is brought down, at which time that kthread
325 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
326 * that XPC has dismantled all communication infrastructure for the associated
327 * partition.) This kthread becomes the channel manager for that partition.
329 * Each active partition has a channel manager, who, besides connecting and
330 * disconnecting channels, will ensure that each of the partition's connected
331 * channels has the required number of assigned kthreads to get the work done.
334 xpc_channel_mgr(struct xpc_partition
*part
)
336 while (part
->act_state
!= XPC_P_DEACTIVATING
||
337 atomic_read(&part
->nchannels_active
) > 0 ||
338 !xpc_partition_disengaged(part
)) {
340 xpc_process_channel_activity(part
);
343 * Wait until we've been requested to activate kthreads or
344 * all of the channel's message queues have been torn down or
345 * a signal is pending.
347 * The channel_mgr_requests is set to 1 after being awakened,
348 * This is done to prevent the channel mgr from making one pass
349 * through the loop for each request, since he will
350 * be servicing all the requests in one pass. The reason it's
351 * set to 1 instead of 0 is so that other kthreads will know
352 * that the channel mgr is running and won't bother trying to
355 atomic_dec(&part
->channel_mgr_requests
);
356 (void)wait_event_interruptible(part
->channel_mgr_wq
,
357 (atomic_read(&part
->channel_mgr_requests
) > 0 ||
358 part
->local_IPI_amo
!= 0 ||
359 (part
->act_state
== XPC_P_DEACTIVATING
&&
360 atomic_read(&part
->nchannels_active
) == 0 &&
361 xpc_partition_disengaged(part
))));
362 atomic_set(&part
->channel_mgr_requests
, 1);
367 * When XPC HB determines that a partition has come up, it will create a new
368 * kthread and that kthread will call this function to attempt to set up the
369 * basic infrastructure used for Cross Partition Communication with the newly
372 * The kthread that was created by XPC HB and which setup the XPC
373 * infrastructure will remain assigned to the partition becoming the channel
374 * manager for that partition until the partition is deactivating, at which
375 * time the kthread will teardown the XPC infrastructure and then exit.
378 xpc_activating(void *__partid
)
380 short partid
= (u64
)__partid
;
381 struct xpc_partition
*part
= &xpc_partitions
[partid
];
382 unsigned long irq_flags
;
384 DBUG_ON(partid
< 0 || partid
>= xp_max_npartitions
);
386 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
388 if (part
->act_state
== XPC_P_DEACTIVATING
) {
389 part
->act_state
= XPC_P_INACTIVE
;
390 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
391 part
->remote_rp_pa
= 0;
395 /* indicate the thread is activating */
396 DBUG_ON(part
->act_state
!= XPC_P_ACTIVATION_REQ
);
397 part
->act_state
= XPC_P_ACTIVATING
;
399 XPC_SET_REASON(part
, 0, 0);
400 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
402 dev_dbg(xpc_part
, "activating partition %d\n", partid
);
405 * Register the remote partition's AMOs with SAL so it can handle
406 * and cleanup errors within that address range should the remote
407 * partition go down. We don't unregister this range because it is
408 * difficult to tell when outstanding writes to the remote partition
409 * are finished and thus when it is safe to unregister. This should
410 * not result in wasted space in the SAL xp_addr_region table because
411 * we should get the same page for remote_amos_page_pa after module
412 * reloads and system reboots.
414 if (sn_register_xp_addr_region(part
->remote_amos_page_pa
,
416 dev_warn(xpc_part
, "xpc_activating(%d) failed to register "
417 "xp_addr region\n", partid
);
419 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
420 part
->act_state
= XPC_P_INACTIVE
;
421 XPC_SET_REASON(part
, xpPhysAddrRegFailed
, __LINE__
);
422 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
423 part
->remote_rp_pa
= 0;
427 xpc_allow_hb(partid
, xpc_vars
);
428 xpc_IPI_send_activated(part
);
430 if (xpc_setup_infrastructure(part
) == xpSuccess
) {
431 (void)xpc_part_ref(part
); /* this will always succeed */
433 if (xpc_make_first_contact(part
) == xpSuccess
) {
434 xpc_mark_partition_active(part
);
435 xpc_channel_mgr(part
);
436 /* won't return until partition is deactivating */
439 xpc_part_deref(part
);
440 xpc_teardown_infrastructure(part
);
443 xpc_disallow_hb(partid
, xpc_vars
);
444 xpc_mark_partition_inactive(part
);
446 if (part
->reason
== xpReactivating
) {
447 /* interrupting ourselves results in activating partition */
448 xpc_IPI_send_reactivate(part
);
455 xpc_activate_partition(struct xpc_partition
*part
)
457 short partid
= XPC_PARTID(part
);
458 unsigned long irq_flags
;
459 struct task_struct
*kthread
;
461 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
463 DBUG_ON(part
->act_state
!= XPC_P_INACTIVE
);
465 part
->act_state
= XPC_P_ACTIVATION_REQ
;
466 XPC_SET_REASON(part
, xpCloneKThread
, __LINE__
);
468 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
470 kthread
= kthread_run(xpc_activating
, (void *)((u64
)partid
), "xpc%02d",
472 if (IS_ERR(kthread
)) {
473 spin_lock_irqsave(&part
->act_lock
, irq_flags
);
474 part
->act_state
= XPC_P_INACTIVE
;
475 XPC_SET_REASON(part
, xpCloneKThreadFailed
, __LINE__
);
476 spin_unlock_irqrestore(&part
->act_lock
, irq_flags
);
481 * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
482 * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
483 * than one partition, we use an AMO_t structure per partition to indicate
484 * whether a partition has sent an IPI or not. If it has, then wake up the
485 * associated kthread to handle it.
487 * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
488 * running on other partitions.
490 * Noteworthy Arguments:
492 * irq - Interrupt ReQuest number. NOT USED.
494 * dev_id - partid of IPI's potential sender.
497 xpc_notify_IRQ_handler(int irq
, void *dev_id
)
499 short partid
= (short)(u64
)dev_id
;
500 struct xpc_partition
*part
= &xpc_partitions
[partid
];
502 DBUG_ON(partid
< 0 || partid
>= xp_max_npartitions
);
504 if (xpc_part_ref(part
)) {
505 xpc_check_for_channel_activity(part
);
507 xpc_part_deref(part
);
513 * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
514 * because the write to their associated IPI amo completed after the IRQ/IPI
518 xpc_dropped_IPI_check(struct xpc_partition
*part
)
520 if (xpc_part_ref(part
)) {
521 xpc_check_for_channel_activity(part
);
523 part
->dropped_IPI_timer
.expires
= jiffies
+
524 XPC_P_DROPPED_IPI_WAIT_INTERVAL
;
525 add_timer(&part
->dropped_IPI_timer
);
526 xpc_part_deref(part
);
531 xpc_activate_kthreads(struct xpc_channel
*ch
, int needed
)
533 int idle
= atomic_read(&ch
->kthreads_idle
);
534 int assigned
= atomic_read(&ch
->kthreads_assigned
);
537 DBUG_ON(needed
<= 0);
540 wakeup
= (needed
> idle
) ? idle
: needed
;
543 dev_dbg(xpc_chan
, "wakeup %d idle kthreads, partid=%d, "
544 "channel=%d\n", wakeup
, ch
->partid
, ch
->number
);
546 /* only wakeup the requested number of kthreads */
547 wake_up_nr(&ch
->idle_wq
, wakeup
);
553 if (needed
+ assigned
> ch
->kthreads_assigned_limit
) {
554 needed
= ch
->kthreads_assigned_limit
- assigned
;
559 dev_dbg(xpc_chan
, "create %d new kthreads, partid=%d, channel=%d\n",
560 needed
, ch
->partid
, ch
->number
);
562 xpc_create_kthreads(ch
, needed
, 0);
566 * This function is where XPC's kthreads wait for messages to deliver.
569 xpc_kthread_waitmsgs(struct xpc_partition
*part
, struct xpc_channel
*ch
)
572 /* deliver messages to their intended recipients */
574 while (ch
->w_local_GP
.get
< ch
->w_remote_GP
.put
&&
575 !(ch
->flags
& XPC_C_DISCONNECTING
)) {
579 if (atomic_inc_return(&ch
->kthreads_idle
) >
580 ch
->kthreads_idle_limit
) {
581 /* too many idle kthreads on this channel */
582 atomic_dec(&ch
->kthreads_idle
);
586 dev_dbg(xpc_chan
, "idle kthread calling "
587 "wait_event_interruptible_exclusive()\n");
589 (void)wait_event_interruptible_exclusive(ch
->idle_wq
,
590 (ch
->w_local_GP
.get
< ch
->w_remote_GP
.put
||
591 (ch
->flags
& XPC_C_DISCONNECTING
)));
593 atomic_dec(&ch
->kthreads_idle
);
595 } while (!(ch
->flags
& XPC_C_DISCONNECTING
));
599 xpc_kthread_start(void *args
)
601 short partid
= XPC_UNPACK_ARG1(args
);
602 u16 ch_number
= XPC_UNPACK_ARG2(args
);
603 struct xpc_partition
*part
= &xpc_partitions
[partid
];
604 struct xpc_channel
*ch
;
606 unsigned long irq_flags
;
608 dev_dbg(xpc_chan
, "kthread starting, partid=%d, channel=%d\n",
611 ch
= &part
->channels
[ch_number
];
613 if (!(ch
->flags
& XPC_C_DISCONNECTING
)) {
615 /* let registerer know that connection has been established */
617 spin_lock_irqsave(&ch
->lock
, irq_flags
);
618 if (!(ch
->flags
& XPC_C_CONNECTEDCALLOUT
)) {
619 ch
->flags
|= XPC_C_CONNECTEDCALLOUT
;
620 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
622 xpc_connected_callout(ch
);
624 spin_lock_irqsave(&ch
->lock
, irq_flags
);
625 ch
->flags
|= XPC_C_CONNECTEDCALLOUT_MADE
;
626 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
629 * It is possible that while the callout was being
630 * made that the remote partition sent some messages.
631 * If that is the case, we may need to activate
632 * additional kthreads to help deliver them. We only
633 * need one less than total #of messages to deliver.
635 n_needed
= ch
->w_remote_GP
.put
- ch
->w_local_GP
.get
- 1;
636 if (n_needed
> 0 && !(ch
->flags
& XPC_C_DISCONNECTING
))
637 xpc_activate_kthreads(ch
, n_needed
);
640 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
643 xpc_kthread_waitmsgs(part
, ch
);
646 /* let registerer know that connection is disconnecting */
648 spin_lock_irqsave(&ch
->lock
, irq_flags
);
649 if ((ch
->flags
& XPC_C_CONNECTEDCALLOUT_MADE
) &&
650 !(ch
->flags
& XPC_C_DISCONNECTINGCALLOUT
)) {
651 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT
;
652 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
654 xpc_disconnect_callout(ch
, xpDisconnecting
);
656 spin_lock_irqsave(&ch
->lock
, irq_flags
);
657 ch
->flags
|= XPC_C_DISCONNECTINGCALLOUT_MADE
;
659 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
661 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0) {
662 if (atomic_dec_return(&part
->nchannels_engaged
) == 0) {
663 xpc_mark_partition_disengaged(part
);
664 xpc_IPI_send_disengage(part
);
668 xpc_msgqueue_deref(ch
);
670 dev_dbg(xpc_chan
, "kthread exiting, partid=%d, channel=%d\n",
673 xpc_part_deref(part
);
678 * For each partition that XPC has established communications with, there is
679 * a minimum of one kernel thread assigned to perform any operation that
680 * may potentially sleep or block (basically the callouts to the asynchronous
681 * functions registered via xpc_connect()).
683 * Additional kthreads are created and destroyed by XPC as the workload
686 * A kthread is assigned to one of the active channels that exists for a given
690 xpc_create_kthreads(struct xpc_channel
*ch
, int needed
,
691 int ignore_disconnecting
)
693 unsigned long irq_flags
;
694 u64 args
= XPC_PACK_ARGS(ch
->partid
, ch
->number
);
695 struct xpc_partition
*part
= &xpc_partitions
[ch
->partid
];
696 struct task_struct
*kthread
;
698 while (needed
-- > 0) {
701 * The following is done on behalf of the newly created
702 * kthread. That kthread is responsible for doing the
703 * counterpart to the following before it exits.
705 if (ignore_disconnecting
) {
706 if (!atomic_inc_not_zero(&ch
->kthreads_assigned
)) {
707 /* kthreads assigned had gone to zero */
709 XPC_C_DISCONNECTINGCALLOUT_MADE
));
713 } else if (ch
->flags
& XPC_C_DISCONNECTING
) {
716 } else if (atomic_inc_return(&ch
->kthreads_assigned
) == 1) {
717 if (atomic_inc_return(&part
->nchannels_engaged
) == 1)
718 xpc_mark_partition_engaged(part
);
720 (void)xpc_part_ref(part
);
721 xpc_msgqueue_ref(ch
);
723 kthread
= kthread_run(xpc_kthread_start
, (void *)args
,
724 "xpc%02dc%d", ch
->partid
, ch
->number
);
725 if (IS_ERR(kthread
)) {
726 /* the fork failed */
729 * NOTE: if (ignore_disconnecting &&
730 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
731 * then we'll deadlock if all other kthreads assigned
732 * to this channel are blocked in the channel's
733 * registerer, because the only thing that will unblock
734 * them is the xpDisconnecting callout that this
735 * failed kthread_run() would have made.
738 if (atomic_dec_return(&ch
->kthreads_assigned
) == 0 &&
739 atomic_dec_return(&part
->nchannels_engaged
) == 0) {
740 xpc_mark_partition_disengaged(part
);
741 xpc_IPI_send_disengage(part
);
743 xpc_msgqueue_deref(ch
);
744 xpc_part_deref(part
);
746 if (atomic_read(&ch
->kthreads_assigned
) <
747 ch
->kthreads_idle_limit
) {
749 * Flag this as an error only if we have an
750 * insufficient #of kthreads for the channel
753 spin_lock_irqsave(&ch
->lock
, irq_flags
);
754 XPC_DISCONNECT_CHANNEL(ch
, xpLackOfResources
,
756 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
764 xpc_disconnect_wait(int ch_number
)
766 unsigned long irq_flags
;
768 struct xpc_partition
*part
;
769 struct xpc_channel
*ch
;
770 int wakeup_channel_mgr
;
772 /* now wait for all callouts to the caller's function to cease */
773 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
774 part
= &xpc_partitions
[partid
];
776 if (!xpc_part_ref(part
))
779 ch
= &part
->channels
[ch_number
];
781 if (!(ch
->flags
& XPC_C_WDISCONNECT
)) {
782 xpc_part_deref(part
);
786 wait_for_completion(&ch
->wdisconnect_wait
);
788 spin_lock_irqsave(&ch
->lock
, irq_flags
);
789 DBUG_ON(!(ch
->flags
& XPC_C_DISCONNECTED
));
790 wakeup_channel_mgr
= 0;
792 if (ch
->delayed_IPI_flags
) {
793 if (part
->act_state
!= XPC_P_DEACTIVATING
) {
794 spin_lock(&part
->IPI_lock
);
795 XPC_SET_IPI_FLAGS(part
->local_IPI_amo
,
797 ch
->delayed_IPI_flags
);
798 spin_unlock(&part
->IPI_lock
);
799 wakeup_channel_mgr
= 1;
801 ch
->delayed_IPI_flags
= 0;
804 ch
->flags
&= ~XPC_C_WDISCONNECT
;
805 spin_unlock_irqrestore(&ch
->lock
, irq_flags
);
807 if (wakeup_channel_mgr
)
808 xpc_wakeup_channel_mgr(part
);
810 xpc_part_deref(part
);
815 xpc_do_exit(enum xp_retval reason
)
818 int active_part_count
, printed_waiting_msg
= 0;
819 struct xpc_partition
*part
;
820 unsigned long printmsg_time
, disengage_request_timeout
= 0;
822 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
823 DBUG_ON(xpc_exiting
== 1);
826 * Let the heartbeat checker thread and the discovery thread
827 * (if one is running) know that they should exit. Also wake up
828 * the heartbeat checker thread in case it's sleeping.
831 wake_up_interruptible(&xpc_act_IRQ_wq
);
833 /* ignore all incoming interrupts */
834 free_irq(SGI_XPC_ACTIVATE
, NULL
);
836 /* wait for the discovery thread to exit */
837 wait_for_completion(&xpc_discovery_exited
);
839 /* wait for the heartbeat checker thread to exit */
840 wait_for_completion(&xpc_hb_checker_exited
);
842 /* sleep for a 1/3 of a second or so */
843 (void)msleep_interruptible(300);
845 /* wait for all partitions to become inactive */
847 printmsg_time
= jiffies
+ (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
848 xpc_disengage_request_timedout
= 0;
851 active_part_count
= 0;
853 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
854 part
= &xpc_partitions
[partid
];
856 if (xpc_partition_disengaged(part
) &&
857 part
->act_state
== XPC_P_INACTIVE
) {
863 XPC_DEACTIVATE_PARTITION(part
, reason
);
865 if (part
->disengage_request_timeout
>
866 disengage_request_timeout
) {
867 disengage_request_timeout
=
868 part
->disengage_request_timeout
;
872 if (xpc_partition_engaged(-1UL)) {
873 if (time_after(jiffies
, printmsg_time
)) {
874 dev_info(xpc_part
, "waiting for remote "
875 "partitions to disengage, timeout in "
877 (disengage_request_timeout
- jiffies
)
879 printmsg_time
= jiffies
+
880 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* HZ
);
881 printed_waiting_msg
= 1;
884 } else if (active_part_count
> 0) {
885 if (printed_waiting_msg
) {
886 dev_info(xpc_part
, "waiting for local partition"
888 printed_waiting_msg
= 0;
892 if (!xpc_disengage_request_timedout
) {
893 dev_info(xpc_part
, "all partitions have "
899 /* sleep for a 1/3 of a second or so */
900 (void)msleep_interruptible(300);
904 DBUG_ON(xpc_partition_engaged(-1UL));
906 /* indicate to others that our reserved page is uninitialized */
907 xpc_rsvd_page
->stamp
= ZERO_STAMP
;
909 /* now it's time to eliminate our heartbeat */
910 del_timer_sync(&xpc_hb_timer
);
911 DBUG_ON(xpc_vars
->heartbeating_to_mask
!= 0);
913 if (reason
== xpUnloading
) {
914 (void)unregister_die_notifier(&xpc_die_notifier
);
915 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
918 /* close down protections for IPI operations */
919 xpc_restrict_IPI_ops();
921 /* clear the interface to XPC's functions */
922 xpc_clear_interface();
925 unregister_sysctl_table(xpc_sysctl
);
927 kfree(xpc_partitions
);
928 kfree(xpc_remote_copy_buffer_base
);
932 * This function is called when the system is being rebooted.
935 xpc_system_reboot(struct notifier_block
*nb
, unsigned long event
, void *unused
)
937 enum xp_retval reason
;
941 reason
= xpSystemReboot
;
944 reason
= xpSystemHalt
;
947 reason
= xpSystemPoweroff
;
950 reason
= xpSystemGoingDown
;
958 * Notify other partitions to disengage from all references to our memory.
961 xpc_die_disengage(void)
963 struct xpc_partition
*part
;
965 unsigned long engaged
;
966 long time
, printmsg_time
, disengage_request_timeout
;
968 /* keep xpc_hb_checker thread from doing anything (just in case) */
971 xpc_vars
->heartbeating_to_mask
= 0; /* indicate we're deactivated */
973 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
974 part
= &xpc_partitions
[partid
];
976 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part
->
977 remote_vars_version
)) {
979 /* just in case it was left set by an earlier XPC */
980 xpc_clear_partition_engaged(1UL << partid
);
984 if (xpc_partition_engaged(1UL << partid
) ||
985 part
->act_state
!= XPC_P_INACTIVE
) {
986 xpc_request_partition_disengage(part
);
987 xpc_mark_partition_disengaged(part
);
988 xpc_IPI_send_disengage(part
);
993 printmsg_time
= time
+
994 (XPC_DISENGAGE_PRINTMSG_INTERVAL
* sn_rtc_cycles_per_second
);
995 disengage_request_timeout
= time
+
996 (xpc_disengage_request_timelimit
* sn_rtc_cycles_per_second
);
998 /* wait for all other partitions to disengage from us */
1001 engaged
= xpc_partition_engaged(-1UL);
1003 dev_info(xpc_part
, "all partitions have disengaged\n");
1008 if (time
>= disengage_request_timeout
) {
1009 for (partid
= 0; partid
< xp_max_npartitions
;
1011 if (engaged
& (1UL << partid
)) {
1012 dev_info(xpc_part
, "disengage from "
1013 "remote partition %d timed "
1020 if (time
>= printmsg_time
) {
1021 dev_info(xpc_part
, "waiting for remote partitions to "
1022 "disengage, timeout in %ld seconds\n",
1023 (disengage_request_timeout
- time
) /
1024 sn_rtc_cycles_per_second
);
1025 printmsg_time
= time
+
1026 (XPC_DISENGAGE_PRINTMSG_INTERVAL
*
1027 sn_rtc_cycles_per_second
);
1033 * This function is called when the system is being restarted or halted due
1034 * to some sort of system failure. If this is the case we need to notify the
1035 * other partitions to disengage from all references to our memory.
1036 * This function can also be called when our heartbeater could be offlined
1037 * for a time. In this case we need to notify other partitions to not worry
1038 * about the lack of a heartbeat.
1041 xpc_system_die(struct notifier_block
*nb
, unsigned long event
, void *unused
)
1044 case DIE_MACHINE_RESTART
:
1045 case DIE_MACHINE_HALT
:
1046 xpc_die_disengage();
1049 case DIE_KDEBUG_ENTER
:
1050 /* Should lack of heartbeat be ignored by other partitions? */
1051 if (!xpc_kdebug_ignore
)
1055 case DIE_MCA_MONARCH_ENTER
:
1056 case DIE_INIT_MONARCH_ENTER
:
1057 xpc_vars
->heartbeat
++;
1058 xpc_vars
->heartbeat_offline
= 1;
1061 case DIE_KDEBUG_LEAVE
:
1062 /* Is lack of heartbeat being ignored by other partitions? */
1063 if (!xpc_kdebug_ignore
)
1067 case DIE_MCA_MONARCH_LEAVE
:
1068 case DIE_INIT_MONARCH_LEAVE
:
1069 xpc_vars
->heartbeat
++;
1070 xpc_vars
->heartbeat_offline
= 0;
1082 struct xpc_partition
*part
;
1083 struct task_struct
*kthread
;
1088 * The ia64-sn2 architecture supports at most 64 partitions.
1089 * And the inability to unregister remote AMOs restricts us
1090 * further to only support exactly 64 partitions on this
1091 * architecture, no less.
1093 if (xp_max_npartitions
!= 64)
1098 } else if (is_uv()) {
1105 snprintf(xpc_part
->bus_id
, BUS_ID_SIZE
, "part");
1106 snprintf(xpc_chan
->bus_id
, BUS_ID_SIZE
, "chan");
1108 buf_size
= max(XPC_RP_VARS_SIZE
,
1109 XPC_RP_HEADER_SIZE
+ XP_NASID_MASK_BYTES
);
1110 xpc_remote_copy_buffer
= xpc_kmalloc_cacheline_aligned(buf_size
,
1112 &xpc_remote_copy_buffer_base
);
1113 if (xpc_remote_copy_buffer
== NULL
) {
1114 dev_err(xpc_part
, "can't get memory for remote copy buffer\n");
1118 xpc_partitions
= kzalloc(sizeof(struct xpc_partition
) *
1119 xp_max_npartitions
, GFP_KERNEL
);
1120 if (xpc_partitions
== NULL
) {
1121 dev_err(xpc_part
, "can't get memory for partition structure\n");
1127 * The first few fields of each entry of xpc_partitions[] need to
1128 * be initialized now so that calls to xpc_connect() and
1129 * xpc_disconnect() can be made prior to the activation of any remote
1130 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1131 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1132 * PARTITION HAS BEEN ACTIVATED.
1134 for (partid
= 0; partid
< xp_max_npartitions
; partid
++) {
1135 part
= &xpc_partitions
[partid
];
1137 DBUG_ON((u64
)part
!= L1_CACHE_ALIGN((u64
)part
));
1139 part
->act_IRQ_rcvd
= 0;
1140 spin_lock_init(&part
->act_lock
);
1141 part
->act_state
= XPC_P_INACTIVE
;
1142 XPC_SET_REASON(part
, 0, 0);
1144 init_timer(&part
->disengage_request_timer
);
1145 part
->disengage_request_timer
.function
=
1146 xpc_timeout_partition_disengage_request
;
1147 part
->disengage_request_timer
.data
= (unsigned long)part
;
1149 part
->setup_state
= XPC_P_UNSET
;
1150 init_waitqueue_head(&part
->teardown_wq
);
1151 atomic_set(&part
->references
, 0);
1154 xpc_sysctl
= register_sysctl_table(xpc_sys_dir
);
1157 * Open up protections for IPI operations (and AMO operations on
1158 * Shub 1.1 systems).
1160 xpc_allow_IPI_ops();
1163 * Interrupts being processed will increment this atomic variable and
1164 * awaken the heartbeat thread which will process the interrupts.
1166 atomic_set(&xpc_act_IRQ_rcvd
, 0);
1169 * This is safe to do before the xpc_hb_checker thread has started
1170 * because the handler releases a wait queue. If an interrupt is
1171 * received before the thread is waiting, it will not go to sleep,
1172 * but rather immediately process the interrupt.
1174 ret
= request_irq(SGI_XPC_ACTIVATE
, xpc_act_IRQ_handler
, 0,
1177 dev_err(xpc_part
, "can't register ACTIVATE IRQ handler, "
1178 "errno=%d\n", -ret
);
1184 * Fill the partition reserved page with the information needed by
1185 * other partitions to discover we are alive and establish initial
1188 xpc_rsvd_page
= xpc_setup_rsvd_page();
1189 if (xpc_rsvd_page
== NULL
) {
1190 dev_err(xpc_part
, "can't setup our reserved page\n");
1195 /* add ourselves to the reboot_notifier_list */
1196 ret
= register_reboot_notifier(&xpc_reboot_notifier
);
1198 dev_warn(xpc_part
, "can't register reboot notifier\n");
1200 /* add ourselves to the die_notifier list */
1201 ret
= register_die_notifier(&xpc_die_notifier
);
1203 dev_warn(xpc_part
, "can't register die notifier\n");
1205 init_timer(&xpc_hb_timer
);
1206 xpc_hb_timer
.function
= xpc_hb_beater
;
1209 * The real work-horse behind xpc. This processes incoming
1210 * interrupts and monitors remote heartbeats.
1212 kthread
= kthread_run(xpc_hb_checker
, NULL
, XPC_HB_CHECK_THREAD_NAME
);
1213 if (IS_ERR(kthread
)) {
1214 dev_err(xpc_part
, "failed while forking hb check thread\n");
1220 * Startup a thread that will attempt to discover other partitions to
1221 * activate based on info provided by SAL. This new thread is short
1222 * lived and will exit once discovery is complete.
1224 kthread
= kthread_run(xpc_initiate_discovery
, NULL
,
1225 XPC_DISCOVERY_THREAD_NAME
);
1226 if (IS_ERR(kthread
)) {
1227 dev_err(xpc_part
, "failed while forking discovery thread\n");
1229 /* mark this new thread as a non-starter */
1230 complete(&xpc_discovery_exited
);
1232 xpc_do_exit(xpUnloading
);
1236 /* set the interface to point at XPC's functions */
1237 xpc_set_interface(xpc_initiate_connect
, xpc_initiate_disconnect
,
1238 xpc_initiate_allocate
, xpc_initiate_send
,
1239 xpc_initiate_send_notify
, xpc_initiate_received
,
1240 xpc_initiate_partid_to_nasids
);
1244 /* initialization was not successful */
1246 /* indicate to others that our reserved page is uninitialized */
1247 xpc_rsvd_page
->stamp
= ZERO_STAMP
;
1249 del_timer_sync(&xpc_hb_timer
);
1250 (void)unregister_die_notifier(&xpc_die_notifier
);
1251 (void)unregister_reboot_notifier(&xpc_reboot_notifier
);
1253 free_irq(SGI_XPC_ACTIVATE
, NULL
);
1255 xpc_restrict_IPI_ops();
1257 unregister_sysctl_table(xpc_sysctl
);
1258 kfree(xpc_partitions
);
1260 kfree(xpc_remote_copy_buffer_base
);
1264 module_init(xpc_init
);
1269 xpc_do_exit(xpUnloading
);
1272 module_exit(xpc_exit
);
1274 MODULE_AUTHOR("Silicon Graphics, Inc.");
1275 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1276 MODULE_LICENSE("GPL");
1278 module_param(xpc_hb_interval
, int, 0);
1279 MODULE_PARM_DESC(xpc_hb_interval
, "Number of seconds between "
1280 "heartbeat increments.");
1282 module_param(xpc_hb_check_interval
, int, 0);
1283 MODULE_PARM_DESC(xpc_hb_check_interval
, "Number of seconds between "
1284 "heartbeat checks.");
1286 module_param(xpc_disengage_request_timelimit
, int, 0);
1287 MODULE_PARM_DESC(xpc_disengage_request_timelimit
, "Number of seconds to wait "
1288 "for disengage request to complete.");
1290 module_param(xpc_kdebug_ignore
, int, 0);
1291 MODULE_PARM_DESC(xpc_kdebug_ignore
, "Should lack of heartbeat be ignored by "
1292 "other partitions when dropping into kdebug.");