sysctl: Drop & in front of every proc_handler.
[linux-2.6.git] / drivers / misc / sgi-xp / xpc_main.c
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.
5  *
6  * Copyright (c) 2004-2009 Silicon Graphics, Inc.  All Rights Reserved.
7  */
8
9 /*
10  * Cross Partition Communication (XPC) support - standard version.
11  *
12  *      XPC provides a message passing capability that crosses partition
13  *      boundaries. This module is made up of two parts:
14  *
15  *          partition   This part detects the presence/absence of other
16  *                      partitions. It provides a heartbeat and monitors
17  *                      the heartbeats of other partitions.
18  *
19  *          channel     This part manages the channels and sends/receives
20  *                      messages across them to/from other partitions.
21  *
22  *      There are a couple of additional functions residing in XP, which
23  *      provide an interface to XPC for its users.
24  *
25  *
26  *      Caveats:
27  *
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.
40  *
41  *          If/when new hardware solves this IPI problem, we should abandon
42  *          the current approach.
43  *
44  */
45
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>
53 #include "xpc.h"
54
55 /* define two XPC debug device structures to be used with dev_dbg() et al */
56
57 struct device_driver xpc_dbg_name = {
58         .name = "xpc"
59 };
60
61 struct device xpc_part_dbg_subname = {
62         .init_name = "",        /* set to "part" at xpc_init() time */
63         .driver = &xpc_dbg_name
64 };
65
66 struct device xpc_chan_dbg_subname = {
67         .init_name = "",        /* set to "chan" at xpc_init() time */
68         .driver = &xpc_dbg_name
69 };
70
71 struct device *xpc_part = &xpc_part_dbg_subname;
72 struct device *xpc_chan = &xpc_chan_dbg_subname;
73
74 static int xpc_kdebug_ignore;
75
76 /* systune related variables for /proc/sys directories */
77
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;
81
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;
85
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;
89
90 static ctl_table xpc_sys_xpc_hb_dir[] = {
91         {
92          .procname = "hb_interval",
93          .data = &xpc_hb_interval,
94          .maxlen = sizeof(int),
95          .mode = 0644,
96          .proc_handler = proc_dointvec_minmax,
97          .extra1 = &xpc_hb_min_interval,
98          .extra2 = &xpc_hb_max_interval},
99         {
100          .procname = "hb_check_interval",
101          .data = &xpc_hb_check_interval,
102          .maxlen = sizeof(int),
103          .mode = 0644,
104          .proc_handler = proc_dointvec_minmax,
105          .extra1 = &xpc_hb_check_min_interval,
106          .extra2 = &xpc_hb_check_max_interval},
107         {}
108 };
109 static ctl_table xpc_sys_xpc_dir[] = {
110         {
111          .procname = "hb",
112          .mode = 0555,
113          .child = xpc_sys_xpc_hb_dir},
114         {
115          .procname = "disengage_timelimit",
116          .data = &xpc_disengage_timelimit,
117          .maxlen = sizeof(int),
118          .mode = 0644,
119          .proc_handler = proc_dointvec_minmax,
120          .extra1 = &xpc_disengage_min_timelimit,
121          .extra2 = &xpc_disengage_max_timelimit},
122         {}
123 };
124 static ctl_table xpc_sys_dir[] = {
125         {
126          .procname = "xpc",
127          .mode = 0555,
128          .child = xpc_sys_xpc_dir},
129         {}
130 };
131 static struct ctl_table_header *xpc_sysctl;
132
133 /* non-zero if any remote partition disengage was timed out */
134 int xpc_disengage_timedout;
135
136 /* #of activate IRQs received and not yet processed */
137 int xpc_activate_IRQ_rcvd;
138 DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
139
140 /* IRQ handler notifies this wait queue on receipt of an IRQ */
141 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
142
143 static unsigned long xpc_hb_check_timeout;
144 static struct timer_list xpc_hb_timer;
145
146 /* notification that the xpc_hb_checker thread has exited */
147 static DECLARE_COMPLETION(xpc_hb_checker_exited);
148
149 /* notification that the xpc_discovery thread has exited */
150 static DECLARE_COMPLETION(xpc_discovery_exited);
151
152 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
153
154 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
155 static struct notifier_block xpc_reboot_notifier = {
156         .notifier_call = xpc_system_reboot,
157 };
158
159 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
160 static struct notifier_block xpc_die_notifier = {
161         .notifier_call = xpc_system_die,
162 };
163
164 struct xpc_arch_operations xpc_arch_ops;
165
166 /*
167  * Timer function to enforce the timelimit on the partition disengage.
168  */
169 static void
170 xpc_timeout_partition_disengage(unsigned long data)
171 {
172         struct xpc_partition *part = (struct xpc_partition *)data;
173
174         DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
175
176         (void)xpc_partition_disengaged(part);
177
178         DBUG_ON(part->disengage_timeout != 0);
179         DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
180 }
181
182 /*
183  * Timer to produce the heartbeat.  The timer structures function is
184  * already set when this is initially called.  A tunable is used to
185  * specify when the next timeout should occur.
186  */
187 static void
188 xpc_hb_beater(unsigned long dummy)
189 {
190         xpc_arch_ops.increment_heartbeat();
191
192         if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
193                 wake_up_interruptible(&xpc_activate_IRQ_wq);
194
195         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
196         add_timer(&xpc_hb_timer);
197 }
198
199 static void
200 xpc_start_hb_beater(void)
201 {
202         xpc_arch_ops.heartbeat_init();
203         init_timer(&xpc_hb_timer);
204         xpc_hb_timer.function = xpc_hb_beater;
205         xpc_hb_beater(0);
206 }
207
208 static void
209 xpc_stop_hb_beater(void)
210 {
211         del_timer_sync(&xpc_hb_timer);
212         xpc_arch_ops.heartbeat_exit();
213 }
214
215 /*
216  * At periodic intervals, scan through all active partitions and ensure
217  * their heartbeat is still active.  If not, the partition is deactivated.
218  */
219 static void
220 xpc_check_remote_hb(void)
221 {
222         struct xpc_partition *part;
223         short partid;
224         enum xp_retval ret;
225
226         for (partid = 0; partid < xp_max_npartitions; partid++) {
227
228                 if (xpc_exiting)
229                         break;
230
231                 if (partid == xp_partition_id)
232                         continue;
233
234                 part = &xpc_partitions[partid];
235
236                 if (part->act_state == XPC_P_AS_INACTIVE ||
237                     part->act_state == XPC_P_AS_DEACTIVATING) {
238                         continue;
239                 }
240
241                 ret = xpc_arch_ops.get_remote_heartbeat(part);
242                 if (ret != xpSuccess)
243                         XPC_DEACTIVATE_PARTITION(part, ret);
244         }
245 }
246
247 /*
248  * This thread is responsible for nearly all of the partition
249  * activation/deactivation.
250  */
251 static int
252 xpc_hb_checker(void *ignore)
253 {
254         int force_IRQ = 0;
255
256         /* this thread was marked active by xpc_hb_init() */
257
258         set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
259
260         /* set our heartbeating to other partitions into motion */
261         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
262         xpc_start_hb_beater();
263
264         while (!xpc_exiting) {
265
266                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
267                         "been received\n",
268                         (int)(xpc_hb_check_timeout - jiffies),
269                         xpc_activate_IRQ_rcvd);
270
271                 /* checking of remote heartbeats is skewed by IRQ handling */
272                 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
273                         xpc_hb_check_timeout = jiffies +
274                             (xpc_hb_check_interval * HZ);
275
276                         dev_dbg(xpc_part, "checking remote heartbeats\n");
277                         xpc_check_remote_hb();
278
279                         /*
280                          * On sn2 we need to periodically recheck to ensure no
281                          * IRQ/amo pairs have been missed.
282                          */
283                         if (is_shub())
284                                 force_IRQ = 1;
285                 }
286
287                 /* check for outstanding IRQs */
288                 if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
289                         force_IRQ = 0;
290                         dev_dbg(xpc_part, "processing activate IRQs "
291                                 "received\n");
292                         xpc_arch_ops.process_activate_IRQ_rcvd();
293                 }
294
295                 /* wait for IRQ or timeout */
296                 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
297                                                (time_is_before_eq_jiffies(
298                                                 xpc_hb_check_timeout) ||
299                                                 xpc_activate_IRQ_rcvd > 0 ||
300                                                 xpc_exiting));
301         }
302
303         xpc_stop_hb_beater();
304
305         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
306
307         /* mark this thread as having exited */
308         complete(&xpc_hb_checker_exited);
309         return 0;
310 }
311
312 /*
313  * This thread will attempt to discover other partitions to activate
314  * based on info provided by SAL. This new thread is short lived and
315  * will exit once discovery is complete.
316  */
317 static int
318 xpc_initiate_discovery(void *ignore)
319 {
320         xpc_discovery();
321
322         dev_dbg(xpc_part, "discovery thread is exiting\n");
323
324         /* mark this thread as having exited */
325         complete(&xpc_discovery_exited);
326         return 0;
327 }
328
329 /*
330  * The first kthread assigned to a newly activated partition is the one
331  * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
332  * that kthread until the partition is brought down, at which time that kthread
333  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
334  * that XPC has dismantled all communication infrastructure for the associated
335  * partition.) This kthread becomes the channel manager for that partition.
336  *
337  * Each active partition has a channel manager, who, besides connecting and
338  * disconnecting channels, will ensure that each of the partition's connected
339  * channels has the required number of assigned kthreads to get the work done.
340  */
341 static void
342 xpc_channel_mgr(struct xpc_partition *part)
343 {
344         while (part->act_state != XPC_P_AS_DEACTIVATING ||
345                atomic_read(&part->nchannels_active) > 0 ||
346                !xpc_partition_disengaged(part)) {
347
348                 xpc_process_sent_chctl_flags(part);
349
350                 /*
351                  * Wait until we've been requested to activate kthreads or
352                  * all of the channel's message queues have been torn down or
353                  * a signal is pending.
354                  *
355                  * The channel_mgr_requests is set to 1 after being awakened,
356                  * This is done to prevent the channel mgr from making one pass
357                  * through the loop for each request, since he will
358                  * be servicing all the requests in one pass. The reason it's
359                  * set to 1 instead of 0 is so that other kthreads will know
360                  * that the channel mgr is running and won't bother trying to
361                  * wake him up.
362                  */
363                 atomic_dec(&part->channel_mgr_requests);
364                 (void)wait_event_interruptible(part->channel_mgr_wq,
365                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
366                                  part->chctl.all_flags != 0 ||
367                                  (part->act_state == XPC_P_AS_DEACTIVATING &&
368                                  atomic_read(&part->nchannels_active) == 0 &&
369                                  xpc_partition_disengaged(part))));
370                 atomic_set(&part->channel_mgr_requests, 1);
371         }
372 }
373
374 /*
375  * Guarantee that the kzalloc'd memory is cacheline aligned.
376  */
377 void *
378 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
379 {
380         /* see if kzalloc will give us cachline aligned memory by default */
381         *base = kzalloc(size, flags);
382         if (*base == NULL)
383                 return NULL;
384
385         if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
386                 return *base;
387
388         kfree(*base);
389
390         /* nope, we'll have to do it ourselves */
391         *base = kzalloc(size + L1_CACHE_BYTES, flags);
392         if (*base == NULL)
393                 return NULL;
394
395         return (void *)L1_CACHE_ALIGN((u64)*base);
396 }
397
398 /*
399  * Setup the channel structures necessary to support XPartition Communication
400  * between the specified remote partition and the local one.
401  */
402 static enum xp_retval
403 xpc_setup_ch_structures(struct xpc_partition *part)
404 {
405         enum xp_retval ret;
406         int ch_number;
407         struct xpc_channel *ch;
408         short partid = XPC_PARTID(part);
409
410         /*
411          * Allocate all of the channel structures as a contiguous chunk of
412          * memory.
413          */
414         DBUG_ON(part->channels != NULL);
415         part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
416                                  GFP_KERNEL);
417         if (part->channels == NULL) {
418                 dev_err(xpc_chan, "can't get memory for channels\n");
419                 return xpNoMemory;
420         }
421
422         /* allocate the remote open and close args */
423
424         part->remote_openclose_args =
425             xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
426                                           GFP_KERNEL, &part->
427                                           remote_openclose_args_base);
428         if (part->remote_openclose_args == NULL) {
429                 dev_err(xpc_chan, "can't get memory for remote connect args\n");
430                 ret = xpNoMemory;
431                 goto out_1;
432         }
433
434         part->chctl.all_flags = 0;
435         spin_lock_init(&part->chctl_lock);
436
437         atomic_set(&part->channel_mgr_requests, 1);
438         init_waitqueue_head(&part->channel_mgr_wq);
439
440         part->nchannels = XPC_MAX_NCHANNELS;
441
442         atomic_set(&part->nchannels_active, 0);
443         atomic_set(&part->nchannels_engaged, 0);
444
445         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
446                 ch = &part->channels[ch_number];
447
448                 ch->partid = partid;
449                 ch->number = ch_number;
450                 ch->flags = XPC_C_DISCONNECTED;
451
452                 atomic_set(&ch->kthreads_assigned, 0);
453                 atomic_set(&ch->kthreads_idle, 0);
454                 atomic_set(&ch->kthreads_active, 0);
455
456                 atomic_set(&ch->references, 0);
457                 atomic_set(&ch->n_to_notify, 0);
458
459                 spin_lock_init(&ch->lock);
460                 init_completion(&ch->wdisconnect_wait);
461
462                 atomic_set(&ch->n_on_msg_allocate_wq, 0);
463                 init_waitqueue_head(&ch->msg_allocate_wq);
464                 init_waitqueue_head(&ch->idle_wq);
465         }
466
467         ret = xpc_arch_ops.setup_ch_structures(part);
468         if (ret != xpSuccess)
469                 goto out_2;
470
471         /*
472          * With the setting of the partition setup_state to XPC_P_SS_SETUP,
473          * we're declaring that this partition is ready to go.
474          */
475         part->setup_state = XPC_P_SS_SETUP;
476
477         return xpSuccess;
478
479         /* setup of ch structures failed */
480 out_2:
481         kfree(part->remote_openclose_args_base);
482         part->remote_openclose_args = NULL;
483 out_1:
484         kfree(part->channels);
485         part->channels = NULL;
486         return ret;
487 }
488
489 /*
490  * Teardown the channel structures necessary to support XPartition Communication
491  * between the specified remote partition and the local one.
492  */
493 static void
494 xpc_teardown_ch_structures(struct xpc_partition *part)
495 {
496         DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
497         DBUG_ON(atomic_read(&part->nchannels_active) != 0);
498
499         /*
500          * Make this partition inaccessible to local processes by marking it
501          * as no longer setup. Then wait before proceeding with the teardown
502          * until all existing references cease.
503          */
504         DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
505         part->setup_state = XPC_P_SS_WTEARDOWN;
506
507         wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
508
509         /* now we can begin tearing down the infrastructure */
510
511         xpc_arch_ops.teardown_ch_structures(part);
512
513         kfree(part->remote_openclose_args_base);
514         part->remote_openclose_args = NULL;
515         kfree(part->channels);
516         part->channels = NULL;
517
518         part->setup_state = XPC_P_SS_TORNDOWN;
519 }
520
521 /*
522  * When XPC HB determines that a partition has come up, it will create a new
523  * kthread and that kthread will call this function to attempt to set up the
524  * basic infrastructure used for Cross Partition Communication with the newly
525  * upped partition.
526  *
527  * The kthread that was created by XPC HB and which setup the XPC
528  * infrastructure will remain assigned to the partition becoming the channel
529  * manager for that partition until the partition is deactivating, at which
530  * time the kthread will teardown the XPC infrastructure and then exit.
531  */
532 static int
533 xpc_activating(void *__partid)
534 {
535         short partid = (u64)__partid;
536         struct xpc_partition *part = &xpc_partitions[partid];
537         unsigned long irq_flags;
538
539         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
540
541         spin_lock_irqsave(&part->act_lock, irq_flags);
542
543         if (part->act_state == XPC_P_AS_DEACTIVATING) {
544                 part->act_state = XPC_P_AS_INACTIVE;
545                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
546                 part->remote_rp_pa = 0;
547                 return 0;
548         }
549
550         /* indicate the thread is activating */
551         DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
552         part->act_state = XPC_P_AS_ACTIVATING;
553
554         XPC_SET_REASON(part, 0, 0);
555         spin_unlock_irqrestore(&part->act_lock, irq_flags);
556
557         dev_dbg(xpc_part, "activating partition %d\n", partid);
558
559         xpc_arch_ops.allow_hb(partid);
560
561         if (xpc_setup_ch_structures(part) == xpSuccess) {
562                 (void)xpc_part_ref(part);       /* this will always succeed */
563
564                 if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
565                         xpc_mark_partition_active(part);
566                         xpc_channel_mgr(part);
567                         /* won't return until partition is deactivating */
568                 }
569
570                 xpc_part_deref(part);
571                 xpc_teardown_ch_structures(part);
572         }
573
574         xpc_arch_ops.disallow_hb(partid);
575         xpc_mark_partition_inactive(part);
576
577         if (part->reason == xpReactivating) {
578                 /* interrupting ourselves results in activating partition */
579                 xpc_arch_ops.request_partition_reactivation(part);
580         }
581
582         return 0;
583 }
584
585 void
586 xpc_activate_partition(struct xpc_partition *part)
587 {
588         short partid = XPC_PARTID(part);
589         unsigned long irq_flags;
590         struct task_struct *kthread;
591
592         spin_lock_irqsave(&part->act_lock, irq_flags);
593
594         DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
595
596         part->act_state = XPC_P_AS_ACTIVATION_REQ;
597         XPC_SET_REASON(part, xpCloneKThread, __LINE__);
598
599         spin_unlock_irqrestore(&part->act_lock, irq_flags);
600
601         kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
602                               partid);
603         if (IS_ERR(kthread)) {
604                 spin_lock_irqsave(&part->act_lock, irq_flags);
605                 part->act_state = XPC_P_AS_INACTIVE;
606                 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
607                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
608         }
609 }
610
611 void
612 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
613 {
614         int idle = atomic_read(&ch->kthreads_idle);
615         int assigned = atomic_read(&ch->kthreads_assigned);
616         int wakeup;
617
618         DBUG_ON(needed <= 0);
619
620         if (idle > 0) {
621                 wakeup = (needed > idle) ? idle : needed;
622                 needed -= wakeup;
623
624                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
625                         "channel=%d\n", wakeup, ch->partid, ch->number);
626
627                 /* only wakeup the requested number of kthreads */
628                 wake_up_nr(&ch->idle_wq, wakeup);
629         }
630
631         if (needed <= 0)
632                 return;
633
634         if (needed + assigned > ch->kthreads_assigned_limit) {
635                 needed = ch->kthreads_assigned_limit - assigned;
636                 if (needed <= 0)
637                         return;
638         }
639
640         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
641                 needed, ch->partid, ch->number);
642
643         xpc_create_kthreads(ch, needed, 0);
644 }
645
646 /*
647  * This function is where XPC's kthreads wait for messages to deliver.
648  */
649 static void
650 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
651 {
652         int (*n_of_deliverable_payloads) (struct xpc_channel *) =
653                 xpc_arch_ops.n_of_deliverable_payloads;
654
655         do {
656                 /* deliver messages to their intended recipients */
657
658                 while (n_of_deliverable_payloads(ch) > 0 &&
659                        !(ch->flags & XPC_C_DISCONNECTING)) {
660                         xpc_deliver_payload(ch);
661                 }
662
663                 if (atomic_inc_return(&ch->kthreads_idle) >
664                     ch->kthreads_idle_limit) {
665                         /* too many idle kthreads on this channel */
666                         atomic_dec(&ch->kthreads_idle);
667                         break;
668                 }
669
670                 dev_dbg(xpc_chan, "idle kthread calling "
671                         "wait_event_interruptible_exclusive()\n");
672
673                 (void)wait_event_interruptible_exclusive(ch->idle_wq,
674                                 (n_of_deliverable_payloads(ch) > 0 ||
675                                  (ch->flags & XPC_C_DISCONNECTING)));
676
677                 atomic_dec(&ch->kthreads_idle);
678
679         } while (!(ch->flags & XPC_C_DISCONNECTING));
680 }
681
682 static int
683 xpc_kthread_start(void *args)
684 {
685         short partid = XPC_UNPACK_ARG1(args);
686         u16 ch_number = XPC_UNPACK_ARG2(args);
687         struct xpc_partition *part = &xpc_partitions[partid];
688         struct xpc_channel *ch;
689         int n_needed;
690         unsigned long irq_flags;
691         int (*n_of_deliverable_payloads) (struct xpc_channel *) =
692                 xpc_arch_ops.n_of_deliverable_payloads;
693
694         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
695                 partid, ch_number);
696
697         ch = &part->channels[ch_number];
698
699         if (!(ch->flags & XPC_C_DISCONNECTING)) {
700
701                 /* let registerer know that connection has been established */
702
703                 spin_lock_irqsave(&ch->lock, irq_flags);
704                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
705                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
706                         spin_unlock_irqrestore(&ch->lock, irq_flags);
707
708                         xpc_connected_callout(ch);
709
710                         spin_lock_irqsave(&ch->lock, irq_flags);
711                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
712                         spin_unlock_irqrestore(&ch->lock, irq_flags);
713
714                         /*
715                          * It is possible that while the callout was being
716                          * made that the remote partition sent some messages.
717                          * If that is the case, we may need to activate
718                          * additional kthreads to help deliver them. We only
719                          * need one less than total #of messages to deliver.
720                          */
721                         n_needed = n_of_deliverable_payloads(ch) - 1;
722                         if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
723                                 xpc_activate_kthreads(ch, n_needed);
724
725                 } else {
726                         spin_unlock_irqrestore(&ch->lock, irq_flags);
727                 }
728
729                 xpc_kthread_waitmsgs(part, ch);
730         }
731
732         /* let registerer know that connection is disconnecting */
733
734         spin_lock_irqsave(&ch->lock, irq_flags);
735         if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
736             !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
737                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
738                 spin_unlock_irqrestore(&ch->lock, irq_flags);
739
740                 xpc_disconnect_callout(ch, xpDisconnecting);
741
742                 spin_lock_irqsave(&ch->lock, irq_flags);
743                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
744         }
745         spin_unlock_irqrestore(&ch->lock, irq_flags);
746
747         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
748             atomic_dec_return(&part->nchannels_engaged) == 0) {
749                 xpc_arch_ops.indicate_partition_disengaged(part);
750         }
751
752         xpc_msgqueue_deref(ch);
753
754         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
755                 partid, ch_number);
756
757         xpc_part_deref(part);
758         return 0;
759 }
760
761 /*
762  * For each partition that XPC has established communications with, there is
763  * a minimum of one kernel thread assigned to perform any operation that
764  * may potentially sleep or block (basically the callouts to the asynchronous
765  * functions registered via xpc_connect()).
766  *
767  * Additional kthreads are created and destroyed by XPC as the workload
768  * demands.
769  *
770  * A kthread is assigned to one of the active channels that exists for a given
771  * partition.
772  */
773 void
774 xpc_create_kthreads(struct xpc_channel *ch, int needed,
775                     int ignore_disconnecting)
776 {
777         unsigned long irq_flags;
778         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
779         struct xpc_partition *part = &xpc_partitions[ch->partid];
780         struct task_struct *kthread;
781         void (*indicate_partition_disengaged) (struct xpc_partition *) =
782                 xpc_arch_ops.indicate_partition_disengaged;
783
784         while (needed-- > 0) {
785
786                 /*
787                  * The following is done on behalf of the newly created
788                  * kthread. That kthread is responsible for doing the
789                  * counterpart to the following before it exits.
790                  */
791                 if (ignore_disconnecting) {
792                         if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
793                                 /* kthreads assigned had gone to zero */
794                                 BUG_ON(!(ch->flags &
795                                          XPC_C_DISCONNECTINGCALLOUT_MADE));
796                                 break;
797                         }
798
799                 } else if (ch->flags & XPC_C_DISCONNECTING) {
800                         break;
801
802                 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
803                            atomic_inc_return(&part->nchannels_engaged) == 1) {
804                         xpc_arch_ops.indicate_partition_engaged(part);
805                 }
806                 (void)xpc_part_ref(part);
807                 xpc_msgqueue_ref(ch);
808
809                 kthread = kthread_run(xpc_kthread_start, (void *)args,
810                                       "xpc%02dc%d", ch->partid, ch->number);
811                 if (IS_ERR(kthread)) {
812                         /* the fork failed */
813
814                         /*
815                          * NOTE: if (ignore_disconnecting &&
816                          * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
817                          * then we'll deadlock if all other kthreads assigned
818                          * to this channel are blocked in the channel's
819                          * registerer, because the only thing that will unblock
820                          * them is the xpDisconnecting callout that this
821                          * failed kthread_run() would have made.
822                          */
823
824                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
825                             atomic_dec_return(&part->nchannels_engaged) == 0) {
826                                 indicate_partition_disengaged(part);
827                         }
828                         xpc_msgqueue_deref(ch);
829                         xpc_part_deref(part);
830
831                         if (atomic_read(&ch->kthreads_assigned) <
832                             ch->kthreads_idle_limit) {
833                                 /*
834                                  * Flag this as an error only if we have an
835                                  * insufficient #of kthreads for the channel
836                                  * to function.
837                                  */
838                                 spin_lock_irqsave(&ch->lock, irq_flags);
839                                 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
840                                                        &irq_flags);
841                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
842                         }
843                         break;
844                 }
845         }
846 }
847
848 void
849 xpc_disconnect_wait(int ch_number)
850 {
851         unsigned long irq_flags;
852         short partid;
853         struct xpc_partition *part;
854         struct xpc_channel *ch;
855         int wakeup_channel_mgr;
856
857         /* now wait for all callouts to the caller's function to cease */
858         for (partid = 0; partid < xp_max_npartitions; partid++) {
859                 part = &xpc_partitions[partid];
860
861                 if (!xpc_part_ref(part))
862                         continue;
863
864                 ch = &part->channels[ch_number];
865
866                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
867                         xpc_part_deref(part);
868                         continue;
869                 }
870
871                 wait_for_completion(&ch->wdisconnect_wait);
872
873                 spin_lock_irqsave(&ch->lock, irq_flags);
874                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
875                 wakeup_channel_mgr = 0;
876
877                 if (ch->delayed_chctl_flags) {
878                         if (part->act_state != XPC_P_AS_DEACTIVATING) {
879                                 spin_lock(&part->chctl_lock);
880                                 part->chctl.flags[ch->number] |=
881                                     ch->delayed_chctl_flags;
882                                 spin_unlock(&part->chctl_lock);
883                                 wakeup_channel_mgr = 1;
884                         }
885                         ch->delayed_chctl_flags = 0;
886                 }
887
888                 ch->flags &= ~XPC_C_WDISCONNECT;
889                 spin_unlock_irqrestore(&ch->lock, irq_flags);
890
891                 if (wakeup_channel_mgr)
892                         xpc_wakeup_channel_mgr(part);
893
894                 xpc_part_deref(part);
895         }
896 }
897
898 static int
899 xpc_setup_partitions(void)
900 {
901         short partid;
902         struct xpc_partition *part;
903
904         xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
905                                  xp_max_npartitions, GFP_KERNEL);
906         if (xpc_partitions == NULL) {
907                 dev_err(xpc_part, "can't get memory for partition structure\n");
908                 return -ENOMEM;
909         }
910
911         /*
912          * The first few fields of each entry of xpc_partitions[] need to
913          * be initialized now so that calls to xpc_connect() and
914          * xpc_disconnect() can be made prior to the activation of any remote
915          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
916          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
917          * PARTITION HAS BEEN ACTIVATED.
918          */
919         for (partid = 0; partid < xp_max_npartitions; partid++) {
920                 part = &xpc_partitions[partid];
921
922                 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
923
924                 part->activate_IRQ_rcvd = 0;
925                 spin_lock_init(&part->act_lock);
926                 part->act_state = XPC_P_AS_INACTIVE;
927                 XPC_SET_REASON(part, 0, 0);
928
929                 init_timer(&part->disengage_timer);
930                 part->disengage_timer.function =
931                     xpc_timeout_partition_disengage;
932                 part->disengage_timer.data = (unsigned long)part;
933
934                 part->setup_state = XPC_P_SS_UNSET;
935                 init_waitqueue_head(&part->teardown_wq);
936                 atomic_set(&part->references, 0);
937         }
938
939         return xpc_arch_ops.setup_partitions();
940 }
941
942 static void
943 xpc_teardown_partitions(void)
944 {
945         xpc_arch_ops.teardown_partitions();
946         kfree(xpc_partitions);
947 }
948
949 static void
950 xpc_do_exit(enum xp_retval reason)
951 {
952         short partid;
953         int active_part_count, printed_waiting_msg = 0;
954         struct xpc_partition *part;
955         unsigned long printmsg_time, disengage_timeout = 0;
956
957         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
958         DBUG_ON(xpc_exiting == 1);
959
960         /*
961          * Let the heartbeat checker thread and the discovery thread
962          * (if one is running) know that they should exit. Also wake up
963          * the heartbeat checker thread in case it's sleeping.
964          */
965         xpc_exiting = 1;
966         wake_up_interruptible(&xpc_activate_IRQ_wq);
967
968         /* wait for the discovery thread to exit */
969         wait_for_completion(&xpc_discovery_exited);
970
971         /* wait for the heartbeat checker thread to exit */
972         wait_for_completion(&xpc_hb_checker_exited);
973
974         /* sleep for a 1/3 of a second or so */
975         (void)msleep_interruptible(300);
976
977         /* wait for all partitions to become inactive */
978
979         printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
980         xpc_disengage_timedout = 0;
981
982         do {
983                 active_part_count = 0;
984
985                 for (partid = 0; partid < xp_max_npartitions; partid++) {
986                         part = &xpc_partitions[partid];
987
988                         if (xpc_partition_disengaged(part) &&
989                             part->act_state == XPC_P_AS_INACTIVE) {
990                                 continue;
991                         }
992
993                         active_part_count++;
994
995                         XPC_DEACTIVATE_PARTITION(part, reason);
996
997                         if (part->disengage_timeout > disengage_timeout)
998                                 disengage_timeout = part->disengage_timeout;
999                 }
1000
1001                 if (xpc_arch_ops.any_partition_engaged()) {
1002                         if (time_is_before_jiffies(printmsg_time)) {
1003                                 dev_info(xpc_part, "waiting for remote "
1004                                          "partitions to deactivate, timeout in "
1005                                          "%ld seconds\n", (disengage_timeout -
1006                                          jiffies) / HZ);
1007                                 printmsg_time = jiffies +
1008                                     (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
1009                                 printed_waiting_msg = 1;
1010                         }
1011
1012                 } else if (active_part_count > 0) {
1013                         if (printed_waiting_msg) {
1014                                 dev_info(xpc_part, "waiting for local partition"
1015                                          " to deactivate\n");
1016                                 printed_waiting_msg = 0;
1017                         }
1018
1019                 } else {
1020                         if (!xpc_disengage_timedout) {
1021                                 dev_info(xpc_part, "all partitions have "
1022                                          "deactivated\n");
1023                         }
1024                         break;
1025                 }
1026
1027                 /* sleep for a 1/3 of a second or so */
1028                 (void)msleep_interruptible(300);
1029
1030         } while (1);
1031
1032         DBUG_ON(xpc_arch_ops.any_partition_engaged());
1033
1034         xpc_teardown_rsvd_page();
1035
1036         if (reason == xpUnloading) {
1037                 (void)unregister_die_notifier(&xpc_die_notifier);
1038                 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1039         }
1040
1041         /* clear the interface to XPC's functions */
1042         xpc_clear_interface();
1043
1044         if (xpc_sysctl)
1045                 unregister_sysctl_table(xpc_sysctl);
1046
1047         xpc_teardown_partitions();
1048
1049         if (is_shub())
1050                 xpc_exit_sn2();
1051         else if (is_uv())
1052                 xpc_exit_uv();
1053 }
1054
1055 /*
1056  * This function is called when the system is being rebooted.
1057  */
1058 static int
1059 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1060 {
1061         enum xp_retval reason;
1062
1063         switch (event) {
1064         case SYS_RESTART:
1065                 reason = xpSystemReboot;
1066                 break;
1067         case SYS_HALT:
1068                 reason = xpSystemHalt;
1069                 break;
1070         case SYS_POWER_OFF:
1071                 reason = xpSystemPoweroff;
1072                 break;
1073         default:
1074                 reason = xpSystemGoingDown;
1075         }
1076
1077         xpc_do_exit(reason);
1078         return NOTIFY_DONE;
1079 }
1080
1081 /*
1082  * Notify other partitions to deactivate from us by first disengaging from all
1083  * references to our memory.
1084  */
1085 static void
1086 xpc_die_deactivate(void)
1087 {
1088         struct xpc_partition *part;
1089         short partid;
1090         int any_engaged;
1091         long keep_waiting;
1092         long wait_to_print;
1093
1094         /* keep xpc_hb_checker thread from doing anything (just in case) */
1095         xpc_exiting = 1;
1096
1097         xpc_arch_ops.disallow_all_hbs();   /*indicate we're deactivated */
1098
1099         for (partid = 0; partid < xp_max_npartitions; partid++) {
1100                 part = &xpc_partitions[partid];
1101
1102                 if (xpc_arch_ops.partition_engaged(partid) ||
1103                     part->act_state != XPC_P_AS_INACTIVE) {
1104                         xpc_arch_ops.request_partition_deactivation(part);
1105                         xpc_arch_ops.indicate_partition_disengaged(part);
1106                 }
1107         }
1108
1109         /*
1110          * Though we requested that all other partitions deactivate from us,
1111          * we only wait until they've all disengaged or we've reached the
1112          * defined timelimit.
1113          *
1114          * Given that one iteration through the following while-loop takes
1115          * approximately 200 microseconds, calculate the #of loops to take
1116          * before bailing and the #of loops before printing a waiting message.
1117          */
1118         keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1119         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1120
1121         while (1) {
1122                 any_engaged = xpc_arch_ops.any_partition_engaged();
1123                 if (!any_engaged) {
1124                         dev_info(xpc_part, "all partitions have deactivated\n");
1125                         break;
1126                 }
1127
1128                 if (!keep_waiting--) {
1129                         for (partid = 0; partid < xp_max_npartitions;
1130                              partid++) {
1131                                 if (xpc_arch_ops.partition_engaged(partid)) {
1132                                         dev_info(xpc_part, "deactivate from "
1133                                                  "remote partition %d timed "
1134                                                  "out\n", partid);
1135                                 }
1136                         }
1137                         break;
1138                 }
1139
1140                 if (!wait_to_print--) {
1141                         dev_info(xpc_part, "waiting for remote partitions to "
1142                                  "deactivate, timeout in %ld seconds\n",
1143                                  keep_waiting / (1000 * 5));
1144                         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1145                             1000 * 5;
1146                 }
1147
1148                 udelay(200);
1149         }
1150 }
1151
1152 /*
1153  * This function is called when the system is being restarted or halted due
1154  * to some sort of system failure. If this is the case we need to notify the
1155  * other partitions to disengage from all references to our memory.
1156  * This function can also be called when our heartbeater could be offlined
1157  * for a time. In this case we need to notify other partitions to not worry
1158  * about the lack of a heartbeat.
1159  */
1160 static int
1161 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1162 {
1163 #ifdef CONFIG_IA64              /* !!! temporary kludge */
1164         switch (event) {
1165         case DIE_MACHINE_RESTART:
1166         case DIE_MACHINE_HALT:
1167                 xpc_die_deactivate();
1168                 break;
1169
1170         case DIE_KDEBUG_ENTER:
1171                 /* Should lack of heartbeat be ignored by other partitions? */
1172                 if (!xpc_kdebug_ignore)
1173                         break;
1174
1175                 /* fall through */
1176         case DIE_MCA_MONARCH_ENTER:
1177         case DIE_INIT_MONARCH_ENTER:
1178                 xpc_arch_ops.offline_heartbeat();
1179                 break;
1180
1181         case DIE_KDEBUG_LEAVE:
1182                 /* Is lack of heartbeat being ignored by other partitions? */
1183                 if (!xpc_kdebug_ignore)
1184                         break;
1185
1186                 /* fall through */
1187         case DIE_MCA_MONARCH_LEAVE:
1188         case DIE_INIT_MONARCH_LEAVE:
1189                 xpc_arch_ops.online_heartbeat();
1190                 break;
1191         }
1192 #else
1193         xpc_die_deactivate();
1194 #endif
1195
1196         return NOTIFY_DONE;
1197 }
1198
1199 int __init
1200 xpc_init(void)
1201 {
1202         int ret;
1203         struct task_struct *kthread;
1204
1205         dev_set_name(xpc_part, "part");
1206         dev_set_name(xpc_chan, "chan");
1207
1208         if (is_shub()) {
1209                 /*
1210                  * The ia64-sn2 architecture supports at most 64 partitions.
1211                  * And the inability to unregister remote amos restricts us
1212                  * further to only support exactly 64 partitions on this
1213                  * architecture, no less.
1214                  */
1215                 if (xp_max_npartitions != 64) {
1216                         dev_err(xpc_part, "max #of partitions not set to 64\n");
1217                         ret = -EINVAL;
1218                 } else {
1219                         ret = xpc_init_sn2();
1220                 }
1221
1222         } else if (is_uv()) {
1223                 ret = xpc_init_uv();
1224
1225         } else {
1226                 ret = -ENODEV;
1227         }
1228
1229         if (ret != 0)
1230                 return ret;
1231
1232         ret = xpc_setup_partitions();
1233         if (ret != 0) {
1234                 dev_err(xpc_part, "can't get memory for partition structure\n");
1235                 goto out_1;
1236         }
1237
1238         xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1239
1240         /*
1241          * Fill the partition reserved page with the information needed by
1242          * other partitions to discover we are alive and establish initial
1243          * communications.
1244          */
1245         ret = xpc_setup_rsvd_page();
1246         if (ret != 0) {
1247                 dev_err(xpc_part, "can't setup our reserved page\n");
1248                 goto out_2;
1249         }
1250
1251         /* add ourselves to the reboot_notifier_list */
1252         ret = register_reboot_notifier(&xpc_reboot_notifier);
1253         if (ret != 0)
1254                 dev_warn(xpc_part, "can't register reboot notifier\n");
1255
1256         /* add ourselves to the die_notifier list */
1257         ret = register_die_notifier(&xpc_die_notifier);
1258         if (ret != 0)
1259                 dev_warn(xpc_part, "can't register die notifier\n");
1260
1261         /*
1262          * The real work-horse behind xpc.  This processes incoming
1263          * interrupts and monitors remote heartbeats.
1264          */
1265         kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1266         if (IS_ERR(kthread)) {
1267                 dev_err(xpc_part, "failed while forking hb check thread\n");
1268                 ret = -EBUSY;
1269                 goto out_3;
1270         }
1271
1272         /*
1273          * Startup a thread that will attempt to discover other partitions to
1274          * activate based on info provided by SAL. This new thread is short
1275          * lived and will exit once discovery is complete.
1276          */
1277         kthread = kthread_run(xpc_initiate_discovery, NULL,
1278                               XPC_DISCOVERY_THREAD_NAME);
1279         if (IS_ERR(kthread)) {
1280                 dev_err(xpc_part, "failed while forking discovery thread\n");
1281
1282                 /* mark this new thread as a non-starter */
1283                 complete(&xpc_discovery_exited);
1284
1285                 xpc_do_exit(xpUnloading);
1286                 return -EBUSY;
1287         }
1288
1289         /* set the interface to point at XPC's functions */
1290         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1291                           xpc_initiate_send, xpc_initiate_send_notify,
1292                           xpc_initiate_received, xpc_initiate_partid_to_nasids);
1293
1294         return 0;
1295
1296         /* initialization was not successful */
1297 out_3:
1298         xpc_teardown_rsvd_page();
1299
1300         (void)unregister_die_notifier(&xpc_die_notifier);
1301         (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1302 out_2:
1303         if (xpc_sysctl)
1304                 unregister_sysctl_table(xpc_sysctl);
1305
1306         xpc_teardown_partitions();
1307 out_1:
1308         if (is_shub())
1309                 xpc_exit_sn2();
1310         else if (is_uv())
1311                 xpc_exit_uv();
1312         return ret;
1313 }
1314
1315 module_init(xpc_init);
1316
1317 void __exit
1318 xpc_exit(void)
1319 {
1320         xpc_do_exit(xpUnloading);
1321 }
1322
1323 module_exit(xpc_exit);
1324
1325 MODULE_AUTHOR("Silicon Graphics, Inc.");
1326 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1327 MODULE_LICENSE("GPL");
1328
1329 module_param(xpc_hb_interval, int, 0);
1330 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1331                  "heartbeat increments.");
1332
1333 module_param(xpc_hb_check_interval, int, 0);
1334 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1335                  "heartbeat checks.");
1336
1337 module_param(xpc_disengage_timelimit, int, 0);
1338 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1339                  "for disengage to complete.");
1340
1341 module_param(xpc_kdebug_ignore, int, 0);
1342 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1343                  "other partitions when dropping into kdebug.");