workqueue: Make worker_attach/detach_pool() update worker->pool
[linux/fpc-iii.git] / drivers / misc / sgi-xp / xpc_main.c
blob0c775d6fcf590d1b1691926d144237ff02c960e1
1 /*
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
4 * for more details.
6 * Copyright (c) 2004-2009 Silicon Graphics, Inc. All Rights Reserved.
7 */
9 /*
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.
26 * Caveats:
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/slab.h>
48 #include <linux/sysctl.h>
49 #include <linux/device.h>
50 #include <linux/delay.h>
51 #include <linux/reboot.h>
52 #include <linux/kdebug.h>
53 #include <linux/kthread.h>
54 #include "xpc.h"
56 #ifdef CONFIG_X86_64
57 #include <asm/traps.h>
58 #endif
60 /* define two XPC debug device structures to be used with dev_dbg() et al */
62 struct device_driver xpc_dbg_name = {
63 .name = "xpc"
66 struct device xpc_part_dbg_subname = {
67 .init_name = "", /* set to "part" at xpc_init() time */
68 .driver = &xpc_dbg_name
71 struct device xpc_chan_dbg_subname = {
72 .init_name = "", /* set to "chan" at xpc_init() time */
73 .driver = &xpc_dbg_name
76 struct device *xpc_part = &xpc_part_dbg_subname;
77 struct device *xpc_chan = &xpc_chan_dbg_subname;
79 static int xpc_kdebug_ignore;
81 /* systune related variables for /proc/sys directories */
83 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
84 static int xpc_hb_min_interval = 1;
85 static int xpc_hb_max_interval = 10;
87 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
88 static int xpc_hb_check_min_interval = 10;
89 static int xpc_hb_check_max_interval = 120;
91 int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
92 static int xpc_disengage_min_timelimit; /* = 0 */
93 static int xpc_disengage_max_timelimit = 120;
95 static struct ctl_table xpc_sys_xpc_hb_dir[] = {
97 .procname = "hb_interval",
98 .data = &xpc_hb_interval,
99 .maxlen = sizeof(int),
100 .mode = 0644,
101 .proc_handler = proc_dointvec_minmax,
102 .extra1 = &xpc_hb_min_interval,
103 .extra2 = &xpc_hb_max_interval},
105 .procname = "hb_check_interval",
106 .data = &xpc_hb_check_interval,
107 .maxlen = sizeof(int),
108 .mode = 0644,
109 .proc_handler = proc_dointvec_minmax,
110 .extra1 = &xpc_hb_check_min_interval,
111 .extra2 = &xpc_hb_check_max_interval},
114 static struct ctl_table xpc_sys_xpc_dir[] = {
116 .procname = "hb",
117 .mode = 0555,
118 .child = xpc_sys_xpc_hb_dir},
120 .procname = "disengage_timelimit",
121 .data = &xpc_disengage_timelimit,
122 .maxlen = sizeof(int),
123 .mode = 0644,
124 .proc_handler = proc_dointvec_minmax,
125 .extra1 = &xpc_disengage_min_timelimit,
126 .extra2 = &xpc_disengage_max_timelimit},
129 static struct ctl_table xpc_sys_dir[] = {
131 .procname = "xpc",
132 .mode = 0555,
133 .child = xpc_sys_xpc_dir},
136 static struct ctl_table_header *xpc_sysctl;
138 /* non-zero if any remote partition disengage was timed out */
139 int xpc_disengage_timedout;
141 /* #of activate IRQs received and not yet processed */
142 int xpc_activate_IRQ_rcvd;
143 DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
145 /* IRQ handler notifies this wait queue on receipt of an IRQ */
146 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
148 static unsigned long xpc_hb_check_timeout;
149 static struct timer_list xpc_hb_timer;
151 /* notification that the xpc_hb_checker thread has exited */
152 static DECLARE_COMPLETION(xpc_hb_checker_exited);
154 /* notification that the xpc_discovery thread has exited */
155 static DECLARE_COMPLETION(xpc_discovery_exited);
157 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
159 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
160 static struct notifier_block xpc_reboot_notifier = {
161 .notifier_call = xpc_system_reboot,
164 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
165 static struct notifier_block xpc_die_notifier = {
166 .notifier_call = xpc_system_die,
169 struct xpc_arch_operations xpc_arch_ops;
172 * Timer function to enforce the timelimit on the partition disengage.
174 static void
175 xpc_timeout_partition_disengage(struct timer_list *t)
177 struct xpc_partition *part = from_timer(part, t, disengage_timer);
179 DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
181 (void)xpc_partition_disengaged(part);
183 DBUG_ON(part->disengage_timeout != 0);
184 DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
188 * Timer to produce the heartbeat. The timer structures function is
189 * already set when this is initially called. A tunable is used to
190 * specify when the next timeout should occur.
192 static void
193 xpc_hb_beater(struct timer_list *unused)
195 xpc_arch_ops.increment_heartbeat();
197 if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
198 wake_up_interruptible(&xpc_activate_IRQ_wq);
200 xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
201 add_timer(&xpc_hb_timer);
204 static void
205 xpc_start_hb_beater(void)
207 xpc_arch_ops.heartbeat_init();
208 timer_setup(&xpc_hb_timer, xpc_hb_beater, 0);
209 xpc_hb_beater(0);
212 static void
213 xpc_stop_hb_beater(void)
215 del_timer_sync(&xpc_hb_timer);
216 xpc_arch_ops.heartbeat_exit();
220 * At periodic intervals, scan through all active partitions and ensure
221 * their heartbeat is still active. If not, the partition is deactivated.
223 static void
224 xpc_check_remote_hb(void)
226 struct xpc_partition *part;
227 short partid;
228 enum xp_retval ret;
230 for (partid = 0; partid < xp_max_npartitions; partid++) {
232 if (xpc_exiting)
233 break;
235 if (partid == xp_partition_id)
236 continue;
238 part = &xpc_partitions[partid];
240 if (part->act_state == XPC_P_AS_INACTIVE ||
241 part->act_state == XPC_P_AS_DEACTIVATING) {
242 continue;
245 ret = xpc_arch_ops.get_remote_heartbeat(part);
246 if (ret != xpSuccess)
247 XPC_DEACTIVATE_PARTITION(part, ret);
252 * This thread is responsible for nearly all of the partition
253 * activation/deactivation.
255 static int
256 xpc_hb_checker(void *ignore)
258 int force_IRQ = 0;
260 /* this thread was marked active by xpc_hb_init() */
262 set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
264 /* set our heartbeating to other partitions into motion */
265 xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
266 xpc_start_hb_beater();
268 while (!xpc_exiting) {
270 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
271 "been received\n",
272 (int)(xpc_hb_check_timeout - jiffies),
273 xpc_activate_IRQ_rcvd);
275 /* checking of remote heartbeats is skewed by IRQ handling */
276 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
277 xpc_hb_check_timeout = jiffies +
278 (xpc_hb_check_interval * HZ);
280 dev_dbg(xpc_part, "checking remote heartbeats\n");
281 xpc_check_remote_hb();
284 * On sn2 we need to periodically recheck to ensure no
285 * IRQ/amo pairs have been missed.
287 if (is_shub())
288 force_IRQ = 1;
291 /* check for outstanding IRQs */
292 if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
293 force_IRQ = 0;
294 dev_dbg(xpc_part, "processing activate IRQs "
295 "received\n");
296 xpc_arch_ops.process_activate_IRQ_rcvd();
299 /* wait for IRQ or timeout */
300 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
301 (time_is_before_eq_jiffies(
302 xpc_hb_check_timeout) ||
303 xpc_activate_IRQ_rcvd > 0 ||
304 xpc_exiting));
307 xpc_stop_hb_beater();
309 dev_dbg(xpc_part, "heartbeat checker is exiting\n");
311 /* mark this thread as having exited */
312 complete(&xpc_hb_checker_exited);
313 return 0;
317 * This thread will attempt to discover other partitions to activate
318 * based on info provided by SAL. This new thread is short lived and
319 * will exit once discovery is complete.
321 static int
322 xpc_initiate_discovery(void *ignore)
324 xpc_discovery();
326 dev_dbg(xpc_part, "discovery thread is exiting\n");
328 /* mark this thread as having exited */
329 complete(&xpc_discovery_exited);
330 return 0;
334 * The first kthread assigned to a newly activated partition is the one
335 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
336 * that kthread until the partition is brought down, at which time that kthread
337 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
338 * that XPC has dismantled all communication infrastructure for the associated
339 * partition.) This kthread becomes the channel manager for that partition.
341 * Each active partition has a channel manager, who, besides connecting and
342 * disconnecting channels, will ensure that each of the partition's connected
343 * channels has the required number of assigned kthreads to get the work done.
345 static void
346 xpc_channel_mgr(struct xpc_partition *part)
348 while (part->act_state != XPC_P_AS_DEACTIVATING ||
349 atomic_read(&part->nchannels_active) > 0 ||
350 !xpc_partition_disengaged(part)) {
352 xpc_process_sent_chctl_flags(part);
355 * Wait until we've been requested to activate kthreads or
356 * all of the channel's message queues have been torn down or
357 * a signal is pending.
359 * The channel_mgr_requests is set to 1 after being awakened,
360 * This is done to prevent the channel mgr from making one pass
361 * through the loop for each request, since he will
362 * be servicing all the requests in one pass. The reason it's
363 * set to 1 instead of 0 is so that other kthreads will know
364 * that the channel mgr is running and won't bother trying to
365 * wake him up.
367 atomic_dec(&part->channel_mgr_requests);
368 (void)wait_event_interruptible(part->channel_mgr_wq,
369 (atomic_read(&part->channel_mgr_requests) > 0 ||
370 part->chctl.all_flags != 0 ||
371 (part->act_state == XPC_P_AS_DEACTIVATING &&
372 atomic_read(&part->nchannels_active) == 0 &&
373 xpc_partition_disengaged(part))));
374 atomic_set(&part->channel_mgr_requests, 1);
379 * Guarantee that the kzalloc'd memory is cacheline aligned.
381 void *
382 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
384 /* see if kzalloc will give us cachline aligned memory by default */
385 *base = kzalloc(size, flags);
386 if (*base == NULL)
387 return NULL;
389 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
390 return *base;
392 kfree(*base);
394 /* nope, we'll have to do it ourselves */
395 *base = kzalloc(size + L1_CACHE_BYTES, flags);
396 if (*base == NULL)
397 return NULL;
399 return (void *)L1_CACHE_ALIGN((u64)*base);
403 * Setup the channel structures necessary to support XPartition Communication
404 * between the specified remote partition and the local one.
406 static enum xp_retval
407 xpc_setup_ch_structures(struct xpc_partition *part)
409 enum xp_retval ret;
410 int ch_number;
411 struct xpc_channel *ch;
412 short partid = XPC_PARTID(part);
415 * Allocate all of the channel structures as a contiguous chunk of
416 * memory.
418 DBUG_ON(part->channels != NULL);
419 part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
420 GFP_KERNEL);
421 if (part->channels == NULL) {
422 dev_err(xpc_chan, "can't get memory for channels\n");
423 return xpNoMemory;
426 /* allocate the remote open and close args */
428 part->remote_openclose_args =
429 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
430 GFP_KERNEL, &part->
431 remote_openclose_args_base);
432 if (part->remote_openclose_args == NULL) {
433 dev_err(xpc_chan, "can't get memory for remote connect args\n");
434 ret = xpNoMemory;
435 goto out_1;
438 part->chctl.all_flags = 0;
439 spin_lock_init(&part->chctl_lock);
441 atomic_set(&part->channel_mgr_requests, 1);
442 init_waitqueue_head(&part->channel_mgr_wq);
444 part->nchannels = XPC_MAX_NCHANNELS;
446 atomic_set(&part->nchannels_active, 0);
447 atomic_set(&part->nchannels_engaged, 0);
449 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
450 ch = &part->channels[ch_number];
452 ch->partid = partid;
453 ch->number = ch_number;
454 ch->flags = XPC_C_DISCONNECTED;
456 atomic_set(&ch->kthreads_assigned, 0);
457 atomic_set(&ch->kthreads_idle, 0);
458 atomic_set(&ch->kthreads_active, 0);
460 atomic_set(&ch->references, 0);
461 atomic_set(&ch->n_to_notify, 0);
463 spin_lock_init(&ch->lock);
464 init_completion(&ch->wdisconnect_wait);
466 atomic_set(&ch->n_on_msg_allocate_wq, 0);
467 init_waitqueue_head(&ch->msg_allocate_wq);
468 init_waitqueue_head(&ch->idle_wq);
471 ret = xpc_arch_ops.setup_ch_structures(part);
472 if (ret != xpSuccess)
473 goto out_2;
476 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
477 * we're declaring that this partition is ready to go.
479 part->setup_state = XPC_P_SS_SETUP;
481 return xpSuccess;
483 /* setup of ch structures failed */
484 out_2:
485 kfree(part->remote_openclose_args_base);
486 part->remote_openclose_args = NULL;
487 out_1:
488 kfree(part->channels);
489 part->channels = NULL;
490 return ret;
494 * Teardown the channel structures necessary to support XPartition Communication
495 * between the specified remote partition and the local one.
497 static void
498 xpc_teardown_ch_structures(struct xpc_partition *part)
500 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
501 DBUG_ON(atomic_read(&part->nchannels_active) != 0);
504 * Make this partition inaccessible to local processes by marking it
505 * as no longer setup. Then wait before proceeding with the teardown
506 * until all existing references cease.
508 DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
509 part->setup_state = XPC_P_SS_WTEARDOWN;
511 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
513 /* now we can begin tearing down the infrastructure */
515 xpc_arch_ops.teardown_ch_structures(part);
517 kfree(part->remote_openclose_args_base);
518 part->remote_openclose_args = NULL;
519 kfree(part->channels);
520 part->channels = NULL;
522 part->setup_state = XPC_P_SS_TORNDOWN;
526 * When XPC HB determines that a partition has come up, it will create a new
527 * kthread and that kthread will call this function to attempt to set up the
528 * basic infrastructure used for Cross Partition Communication with the newly
529 * upped partition.
531 * The kthread that was created by XPC HB and which setup the XPC
532 * infrastructure will remain assigned to the partition becoming the channel
533 * manager for that partition until the partition is deactivating, at which
534 * time the kthread will teardown the XPC infrastructure and then exit.
536 static int
537 xpc_activating(void *__partid)
539 short partid = (u64)__partid;
540 struct xpc_partition *part = &xpc_partitions[partid];
541 unsigned long irq_flags;
543 DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
545 spin_lock_irqsave(&part->act_lock, irq_flags);
547 if (part->act_state == XPC_P_AS_DEACTIVATING) {
548 part->act_state = XPC_P_AS_INACTIVE;
549 spin_unlock_irqrestore(&part->act_lock, irq_flags);
550 part->remote_rp_pa = 0;
551 return 0;
554 /* indicate the thread is activating */
555 DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
556 part->act_state = XPC_P_AS_ACTIVATING;
558 XPC_SET_REASON(part, 0, 0);
559 spin_unlock_irqrestore(&part->act_lock, irq_flags);
561 dev_dbg(xpc_part, "activating partition %d\n", partid);
563 xpc_arch_ops.allow_hb(partid);
565 if (xpc_setup_ch_structures(part) == xpSuccess) {
566 (void)xpc_part_ref(part); /* this will always succeed */
568 if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
569 xpc_mark_partition_active(part);
570 xpc_channel_mgr(part);
571 /* won't return until partition is deactivating */
574 xpc_part_deref(part);
575 xpc_teardown_ch_structures(part);
578 xpc_arch_ops.disallow_hb(partid);
579 xpc_mark_partition_inactive(part);
581 if (part->reason == xpReactivating) {
582 /* interrupting ourselves results in activating partition */
583 xpc_arch_ops.request_partition_reactivation(part);
586 return 0;
589 void
590 xpc_activate_partition(struct xpc_partition *part)
592 short partid = XPC_PARTID(part);
593 unsigned long irq_flags;
594 struct task_struct *kthread;
596 spin_lock_irqsave(&part->act_lock, irq_flags);
598 DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
600 part->act_state = XPC_P_AS_ACTIVATION_REQ;
601 XPC_SET_REASON(part, xpCloneKThread, __LINE__);
603 spin_unlock_irqrestore(&part->act_lock, irq_flags);
605 kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
606 partid);
607 if (IS_ERR(kthread)) {
608 spin_lock_irqsave(&part->act_lock, irq_flags);
609 part->act_state = XPC_P_AS_INACTIVE;
610 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
611 spin_unlock_irqrestore(&part->act_lock, irq_flags);
615 void
616 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
618 int idle = atomic_read(&ch->kthreads_idle);
619 int assigned = atomic_read(&ch->kthreads_assigned);
620 int wakeup;
622 DBUG_ON(needed <= 0);
624 if (idle > 0) {
625 wakeup = (needed > idle) ? idle : needed;
626 needed -= wakeup;
628 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
629 "channel=%d\n", wakeup, ch->partid, ch->number);
631 /* only wakeup the requested number of kthreads */
632 wake_up_nr(&ch->idle_wq, wakeup);
635 if (needed <= 0)
636 return;
638 if (needed + assigned > ch->kthreads_assigned_limit) {
639 needed = ch->kthreads_assigned_limit - assigned;
640 if (needed <= 0)
641 return;
644 dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
645 needed, ch->partid, ch->number);
647 xpc_create_kthreads(ch, needed, 0);
651 * This function is where XPC's kthreads wait for messages to deliver.
653 static void
654 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
656 int (*n_of_deliverable_payloads) (struct xpc_channel *) =
657 xpc_arch_ops.n_of_deliverable_payloads;
659 do {
660 /* deliver messages to their intended recipients */
662 while (n_of_deliverable_payloads(ch) > 0 &&
663 !(ch->flags & XPC_C_DISCONNECTING)) {
664 xpc_deliver_payload(ch);
667 if (atomic_inc_return(&ch->kthreads_idle) >
668 ch->kthreads_idle_limit) {
669 /* too many idle kthreads on this channel */
670 atomic_dec(&ch->kthreads_idle);
671 break;
674 dev_dbg(xpc_chan, "idle kthread calling "
675 "wait_event_interruptible_exclusive()\n");
677 (void)wait_event_interruptible_exclusive(ch->idle_wq,
678 (n_of_deliverable_payloads(ch) > 0 ||
679 (ch->flags & XPC_C_DISCONNECTING)));
681 atomic_dec(&ch->kthreads_idle);
683 } while (!(ch->flags & XPC_C_DISCONNECTING));
686 static int
687 xpc_kthread_start(void *args)
689 short partid = XPC_UNPACK_ARG1(args);
690 u16 ch_number = XPC_UNPACK_ARG2(args);
691 struct xpc_partition *part = &xpc_partitions[partid];
692 struct xpc_channel *ch;
693 int n_needed;
694 unsigned long irq_flags;
695 int (*n_of_deliverable_payloads) (struct xpc_channel *) =
696 xpc_arch_ops.n_of_deliverable_payloads;
698 dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
699 partid, ch_number);
701 ch = &part->channels[ch_number];
703 if (!(ch->flags & XPC_C_DISCONNECTING)) {
705 /* let registerer know that connection has been established */
707 spin_lock_irqsave(&ch->lock, irq_flags);
708 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
709 ch->flags |= XPC_C_CONNECTEDCALLOUT;
710 spin_unlock_irqrestore(&ch->lock, irq_flags);
712 xpc_connected_callout(ch);
714 spin_lock_irqsave(&ch->lock, irq_flags);
715 ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
716 spin_unlock_irqrestore(&ch->lock, irq_flags);
719 * It is possible that while the callout was being
720 * made that the remote partition sent some messages.
721 * If that is the case, we may need to activate
722 * additional kthreads to help deliver them. We only
723 * need one less than total #of messages to deliver.
725 n_needed = n_of_deliverable_payloads(ch) - 1;
726 if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
727 xpc_activate_kthreads(ch, n_needed);
729 } else {
730 spin_unlock_irqrestore(&ch->lock, irq_flags);
733 xpc_kthread_waitmsgs(part, ch);
736 /* let registerer know that connection is disconnecting */
738 spin_lock_irqsave(&ch->lock, irq_flags);
739 if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
740 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
741 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
742 spin_unlock_irqrestore(&ch->lock, irq_flags);
744 xpc_disconnect_callout(ch, xpDisconnecting);
746 spin_lock_irqsave(&ch->lock, irq_flags);
747 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
749 spin_unlock_irqrestore(&ch->lock, irq_flags);
751 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
752 atomic_dec_return(&part->nchannels_engaged) == 0) {
753 xpc_arch_ops.indicate_partition_disengaged(part);
756 xpc_msgqueue_deref(ch);
758 dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
759 partid, ch_number);
761 xpc_part_deref(part);
762 return 0;
766 * For each partition that XPC has established communications with, there is
767 * a minimum of one kernel thread assigned to perform any operation that
768 * may potentially sleep or block (basically the callouts to the asynchronous
769 * functions registered via xpc_connect()).
771 * Additional kthreads are created and destroyed by XPC as the workload
772 * demands.
774 * A kthread is assigned to one of the active channels that exists for a given
775 * partition.
777 void
778 xpc_create_kthreads(struct xpc_channel *ch, int needed,
779 int ignore_disconnecting)
781 unsigned long irq_flags;
782 u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
783 struct xpc_partition *part = &xpc_partitions[ch->partid];
784 struct task_struct *kthread;
785 void (*indicate_partition_disengaged) (struct xpc_partition *) =
786 xpc_arch_ops.indicate_partition_disengaged;
788 while (needed-- > 0) {
791 * The following is done on behalf of the newly created
792 * kthread. That kthread is responsible for doing the
793 * counterpart to the following before it exits.
795 if (ignore_disconnecting) {
796 if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
797 /* kthreads assigned had gone to zero */
798 BUG_ON(!(ch->flags &
799 XPC_C_DISCONNECTINGCALLOUT_MADE));
800 break;
803 } else if (ch->flags & XPC_C_DISCONNECTING) {
804 break;
806 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
807 atomic_inc_return(&part->nchannels_engaged) == 1) {
808 xpc_arch_ops.indicate_partition_engaged(part);
810 (void)xpc_part_ref(part);
811 xpc_msgqueue_ref(ch);
813 kthread = kthread_run(xpc_kthread_start, (void *)args,
814 "xpc%02dc%d", ch->partid, ch->number);
815 if (IS_ERR(kthread)) {
816 /* the fork failed */
819 * NOTE: if (ignore_disconnecting &&
820 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
821 * then we'll deadlock if all other kthreads assigned
822 * to this channel are blocked in the channel's
823 * registerer, because the only thing that will unblock
824 * them is the xpDisconnecting callout that this
825 * failed kthread_run() would have made.
828 if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
829 atomic_dec_return(&part->nchannels_engaged) == 0) {
830 indicate_partition_disengaged(part);
832 xpc_msgqueue_deref(ch);
833 xpc_part_deref(part);
835 if (atomic_read(&ch->kthreads_assigned) <
836 ch->kthreads_idle_limit) {
838 * Flag this as an error only if we have an
839 * insufficient #of kthreads for the channel
840 * to function.
842 spin_lock_irqsave(&ch->lock, irq_flags);
843 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
844 &irq_flags);
845 spin_unlock_irqrestore(&ch->lock, irq_flags);
847 break;
852 void
853 xpc_disconnect_wait(int ch_number)
855 unsigned long irq_flags;
856 short partid;
857 struct xpc_partition *part;
858 struct xpc_channel *ch;
859 int wakeup_channel_mgr;
861 /* now wait for all callouts to the caller's function to cease */
862 for (partid = 0; partid < xp_max_npartitions; partid++) {
863 part = &xpc_partitions[partid];
865 if (!xpc_part_ref(part))
866 continue;
868 ch = &part->channels[ch_number];
870 if (!(ch->flags & XPC_C_WDISCONNECT)) {
871 xpc_part_deref(part);
872 continue;
875 wait_for_completion(&ch->wdisconnect_wait);
877 spin_lock_irqsave(&ch->lock, irq_flags);
878 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
879 wakeup_channel_mgr = 0;
881 if (ch->delayed_chctl_flags) {
882 if (part->act_state != XPC_P_AS_DEACTIVATING) {
883 spin_lock(&part->chctl_lock);
884 part->chctl.flags[ch->number] |=
885 ch->delayed_chctl_flags;
886 spin_unlock(&part->chctl_lock);
887 wakeup_channel_mgr = 1;
889 ch->delayed_chctl_flags = 0;
892 ch->flags &= ~XPC_C_WDISCONNECT;
893 spin_unlock_irqrestore(&ch->lock, irq_flags);
895 if (wakeup_channel_mgr)
896 xpc_wakeup_channel_mgr(part);
898 xpc_part_deref(part);
902 static int
903 xpc_setup_partitions(void)
905 short partid;
906 struct xpc_partition *part;
908 xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
909 xp_max_npartitions, GFP_KERNEL);
910 if (xpc_partitions == NULL) {
911 dev_err(xpc_part, "can't get memory for partition structure\n");
912 return -ENOMEM;
916 * The first few fields of each entry of xpc_partitions[] need to
917 * be initialized now so that calls to xpc_connect() and
918 * xpc_disconnect() can be made prior to the activation of any remote
919 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
920 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
921 * PARTITION HAS BEEN ACTIVATED.
923 for (partid = 0; partid < xp_max_npartitions; partid++) {
924 part = &xpc_partitions[partid];
926 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
928 part->activate_IRQ_rcvd = 0;
929 spin_lock_init(&part->act_lock);
930 part->act_state = XPC_P_AS_INACTIVE;
931 XPC_SET_REASON(part, 0, 0);
933 timer_setup(&part->disengage_timer,
934 xpc_timeout_partition_disengage, 0);
936 part->setup_state = XPC_P_SS_UNSET;
937 init_waitqueue_head(&part->teardown_wq);
938 atomic_set(&part->references, 0);
941 return xpc_arch_ops.setup_partitions();
944 static void
945 xpc_teardown_partitions(void)
947 xpc_arch_ops.teardown_partitions();
948 kfree(xpc_partitions);
951 static void
952 xpc_do_exit(enum xp_retval reason)
954 short partid;
955 int active_part_count, printed_waiting_msg = 0;
956 struct xpc_partition *part;
957 unsigned long printmsg_time, disengage_timeout = 0;
959 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
960 DBUG_ON(xpc_exiting == 1);
963 * Let the heartbeat checker thread and the discovery thread
964 * (if one is running) know that they should exit. Also wake up
965 * the heartbeat checker thread in case it's sleeping.
967 xpc_exiting = 1;
968 wake_up_interruptible(&xpc_activate_IRQ_wq);
970 /* wait for the discovery thread to exit */
971 wait_for_completion(&xpc_discovery_exited);
973 /* wait for the heartbeat checker thread to exit */
974 wait_for_completion(&xpc_hb_checker_exited);
976 /* sleep for a 1/3 of a second or so */
977 (void)msleep_interruptible(300);
979 /* wait for all partitions to become inactive */
981 printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
982 xpc_disengage_timedout = 0;
984 do {
985 active_part_count = 0;
987 for (partid = 0; partid < xp_max_npartitions; partid++) {
988 part = &xpc_partitions[partid];
990 if (xpc_partition_disengaged(part) &&
991 part->act_state == XPC_P_AS_INACTIVE) {
992 continue;
995 active_part_count++;
997 XPC_DEACTIVATE_PARTITION(part, reason);
999 if (part->disengage_timeout > disengage_timeout)
1000 disengage_timeout = part->disengage_timeout;
1003 if (xpc_arch_ops.any_partition_engaged()) {
1004 if (time_is_before_jiffies(printmsg_time)) {
1005 dev_info(xpc_part, "waiting for remote "
1006 "partitions to deactivate, timeout in "
1007 "%ld seconds\n", (disengage_timeout -
1008 jiffies) / HZ);
1009 printmsg_time = jiffies +
1010 (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
1011 printed_waiting_msg = 1;
1014 } else if (active_part_count > 0) {
1015 if (printed_waiting_msg) {
1016 dev_info(xpc_part, "waiting for local partition"
1017 " to deactivate\n");
1018 printed_waiting_msg = 0;
1021 } else {
1022 if (!xpc_disengage_timedout) {
1023 dev_info(xpc_part, "all partitions have "
1024 "deactivated\n");
1026 break;
1029 /* sleep for a 1/3 of a second or so */
1030 (void)msleep_interruptible(300);
1032 } while (1);
1034 DBUG_ON(xpc_arch_ops.any_partition_engaged());
1036 xpc_teardown_rsvd_page();
1038 if (reason == xpUnloading) {
1039 (void)unregister_die_notifier(&xpc_die_notifier);
1040 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1043 /* clear the interface to XPC's functions */
1044 xpc_clear_interface();
1046 if (xpc_sysctl)
1047 unregister_sysctl_table(xpc_sysctl);
1049 xpc_teardown_partitions();
1051 if (is_shub())
1052 xpc_exit_sn2();
1053 else if (is_uv())
1054 xpc_exit_uv();
1058 * This function is called when the system is being rebooted.
1060 static int
1061 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1063 enum xp_retval reason;
1065 switch (event) {
1066 case SYS_RESTART:
1067 reason = xpSystemReboot;
1068 break;
1069 case SYS_HALT:
1070 reason = xpSystemHalt;
1071 break;
1072 case SYS_POWER_OFF:
1073 reason = xpSystemPoweroff;
1074 break;
1075 default:
1076 reason = xpSystemGoingDown;
1079 xpc_do_exit(reason);
1080 return NOTIFY_DONE;
1083 /* Used to only allow one cpu to complete disconnect */
1084 static unsigned int xpc_die_disconnecting;
1087 * Notify other partitions to deactivate from us by first disengaging from all
1088 * references to our memory.
1090 static void
1091 xpc_die_deactivate(void)
1093 struct xpc_partition *part;
1094 short partid;
1095 int any_engaged;
1096 long keep_waiting;
1097 long wait_to_print;
1099 if (cmpxchg(&xpc_die_disconnecting, 0, 1))
1100 return;
1102 /* keep xpc_hb_checker thread from doing anything (just in case) */
1103 xpc_exiting = 1;
1105 xpc_arch_ops.disallow_all_hbs(); /*indicate we're deactivated */
1107 for (partid = 0; partid < xp_max_npartitions; partid++) {
1108 part = &xpc_partitions[partid];
1110 if (xpc_arch_ops.partition_engaged(partid) ||
1111 part->act_state != XPC_P_AS_INACTIVE) {
1112 xpc_arch_ops.request_partition_deactivation(part);
1113 xpc_arch_ops.indicate_partition_disengaged(part);
1118 * Though we requested that all other partitions deactivate from us,
1119 * we only wait until they've all disengaged or we've reached the
1120 * defined timelimit.
1122 * Given that one iteration through the following while-loop takes
1123 * approximately 200 microseconds, calculate the #of loops to take
1124 * before bailing and the #of loops before printing a waiting message.
1126 keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1127 wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1129 while (1) {
1130 any_engaged = xpc_arch_ops.any_partition_engaged();
1131 if (!any_engaged) {
1132 dev_info(xpc_part, "all partitions have deactivated\n");
1133 break;
1136 if (!keep_waiting--) {
1137 for (partid = 0; partid < xp_max_npartitions;
1138 partid++) {
1139 if (xpc_arch_ops.partition_engaged(partid)) {
1140 dev_info(xpc_part, "deactivate from "
1141 "remote partition %d timed "
1142 "out\n", partid);
1145 break;
1148 if (!wait_to_print--) {
1149 dev_info(xpc_part, "waiting for remote partitions to "
1150 "deactivate, timeout in %ld seconds\n",
1151 keep_waiting / (1000 * 5));
1152 wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1153 1000 * 5;
1156 udelay(200);
1161 * This function is called when the system is being restarted or halted due
1162 * to some sort of system failure. If this is the case we need to notify the
1163 * other partitions to disengage from all references to our memory.
1164 * This function can also be called when our heartbeater could be offlined
1165 * for a time. In this case we need to notify other partitions to not worry
1166 * about the lack of a heartbeat.
1168 static int
1169 xpc_system_die(struct notifier_block *nb, unsigned long event, void *_die_args)
1171 #ifdef CONFIG_IA64 /* !!! temporary kludge */
1172 switch (event) {
1173 case DIE_MACHINE_RESTART:
1174 case DIE_MACHINE_HALT:
1175 xpc_die_deactivate();
1176 break;
1178 case DIE_KDEBUG_ENTER:
1179 /* Should lack of heartbeat be ignored by other partitions? */
1180 if (!xpc_kdebug_ignore)
1181 break;
1183 /* fall through */
1184 case DIE_MCA_MONARCH_ENTER:
1185 case DIE_INIT_MONARCH_ENTER:
1186 xpc_arch_ops.offline_heartbeat();
1187 break;
1189 case DIE_KDEBUG_LEAVE:
1190 /* Is lack of heartbeat being ignored by other partitions? */
1191 if (!xpc_kdebug_ignore)
1192 break;
1194 /* fall through */
1195 case DIE_MCA_MONARCH_LEAVE:
1196 case DIE_INIT_MONARCH_LEAVE:
1197 xpc_arch_ops.online_heartbeat();
1198 break;
1200 #else
1201 struct die_args *die_args = _die_args;
1203 switch (event) {
1204 case DIE_TRAP:
1205 if (die_args->trapnr == X86_TRAP_DF)
1206 xpc_die_deactivate();
1208 if (((die_args->trapnr == X86_TRAP_MF) ||
1209 (die_args->trapnr == X86_TRAP_XF)) &&
1210 !user_mode(die_args->regs))
1211 xpc_die_deactivate();
1213 break;
1214 case DIE_INT3:
1215 case DIE_DEBUG:
1216 break;
1217 case DIE_OOPS:
1218 case DIE_GPF:
1219 default:
1220 xpc_die_deactivate();
1222 #endif
1224 return NOTIFY_DONE;
1227 int __init
1228 xpc_init(void)
1230 int ret;
1231 struct task_struct *kthread;
1233 dev_set_name(xpc_part, "part");
1234 dev_set_name(xpc_chan, "chan");
1236 if (is_shub()) {
1238 * The ia64-sn2 architecture supports at most 64 partitions.
1239 * And the inability to unregister remote amos restricts us
1240 * further to only support exactly 64 partitions on this
1241 * architecture, no less.
1243 if (xp_max_npartitions != 64) {
1244 dev_err(xpc_part, "max #of partitions not set to 64\n");
1245 ret = -EINVAL;
1246 } else {
1247 ret = xpc_init_sn2();
1250 } else if (is_uv()) {
1251 ret = xpc_init_uv();
1253 } else {
1254 ret = -ENODEV;
1257 if (ret != 0)
1258 return ret;
1260 ret = xpc_setup_partitions();
1261 if (ret != 0) {
1262 dev_err(xpc_part, "can't get memory for partition structure\n");
1263 goto out_1;
1266 xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1269 * Fill the partition reserved page with the information needed by
1270 * other partitions to discover we are alive and establish initial
1271 * communications.
1273 ret = xpc_setup_rsvd_page();
1274 if (ret != 0) {
1275 dev_err(xpc_part, "can't setup our reserved page\n");
1276 goto out_2;
1279 /* add ourselves to the reboot_notifier_list */
1280 ret = register_reboot_notifier(&xpc_reboot_notifier);
1281 if (ret != 0)
1282 dev_warn(xpc_part, "can't register reboot notifier\n");
1284 /* add ourselves to the die_notifier list */
1285 ret = register_die_notifier(&xpc_die_notifier);
1286 if (ret != 0)
1287 dev_warn(xpc_part, "can't register die notifier\n");
1290 * The real work-horse behind xpc. This processes incoming
1291 * interrupts and monitors remote heartbeats.
1293 kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1294 if (IS_ERR(kthread)) {
1295 dev_err(xpc_part, "failed while forking hb check thread\n");
1296 ret = -EBUSY;
1297 goto out_3;
1301 * Startup a thread that will attempt to discover other partitions to
1302 * activate based on info provided by SAL. This new thread is short
1303 * lived and will exit once discovery is complete.
1305 kthread = kthread_run(xpc_initiate_discovery, NULL,
1306 XPC_DISCOVERY_THREAD_NAME);
1307 if (IS_ERR(kthread)) {
1308 dev_err(xpc_part, "failed while forking discovery thread\n");
1310 /* mark this new thread as a non-starter */
1311 complete(&xpc_discovery_exited);
1313 xpc_do_exit(xpUnloading);
1314 return -EBUSY;
1317 /* set the interface to point at XPC's functions */
1318 xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1319 xpc_initiate_send, xpc_initiate_send_notify,
1320 xpc_initiate_received, xpc_initiate_partid_to_nasids);
1322 return 0;
1324 /* initialization was not successful */
1325 out_3:
1326 xpc_teardown_rsvd_page();
1328 (void)unregister_die_notifier(&xpc_die_notifier);
1329 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1330 out_2:
1331 if (xpc_sysctl)
1332 unregister_sysctl_table(xpc_sysctl);
1334 xpc_teardown_partitions();
1335 out_1:
1336 if (is_shub())
1337 xpc_exit_sn2();
1338 else if (is_uv())
1339 xpc_exit_uv();
1340 return ret;
1343 module_init(xpc_init);
1345 void __exit
1346 xpc_exit(void)
1348 xpc_do_exit(xpUnloading);
1351 module_exit(xpc_exit);
1353 MODULE_AUTHOR("Silicon Graphics, Inc.");
1354 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1355 MODULE_LICENSE("GPL");
1357 module_param(xpc_hb_interval, int, 0);
1358 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1359 "heartbeat increments.");
1361 module_param(xpc_hb_check_interval, int, 0);
1362 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1363 "heartbeat checks.");
1365 module_param(xpc_disengage_timelimit, int, 0);
1366 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1367 "for disengage to complete.");
1369 module_param(xpc_kdebug_ignore, int, 0);
1370 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1371 "other partitions when dropping into kdebug.");