ARM: 6214/2: driver for the character LCD found in ARM refdesigns
[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/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"
55
56 /* define two XPC debug device structures to be used with dev_dbg() et al */
57
58 struct device_driver xpc_dbg_name = {
59         .name = "xpc"
60 };
61
62 struct device xpc_part_dbg_subname = {
63         .init_name = "",        /* set to "part" at xpc_init() time */
64         .driver = &xpc_dbg_name
65 };
66
67 struct device xpc_chan_dbg_subname = {
68         .init_name = "",        /* set to "chan" at xpc_init() time */
69         .driver = &xpc_dbg_name
70 };
71
72 struct device *xpc_part = &xpc_part_dbg_subname;
73 struct device *xpc_chan = &xpc_chan_dbg_subname;
74
75 static int xpc_kdebug_ignore;
76
77 /* systune related variables for /proc/sys directories */
78
79 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
80 static int xpc_hb_min_interval = 1;
81 static int xpc_hb_max_interval = 10;
82
83 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
84 static int xpc_hb_check_min_interval = 10;
85 static int xpc_hb_check_max_interval = 120;
86
87 int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
88 static int xpc_disengage_min_timelimit; /* = 0 */
89 static int xpc_disengage_max_timelimit = 120;
90
91 static ctl_table xpc_sys_xpc_hb_dir[] = {
92         {
93          .procname = "hb_interval",
94          .data = &xpc_hb_interval,
95          .maxlen = sizeof(int),
96          .mode = 0644,
97          .proc_handler = proc_dointvec_minmax,
98          .extra1 = &xpc_hb_min_interval,
99          .extra2 = &xpc_hb_max_interval},
100         {
101          .procname = "hb_check_interval",
102          .data = &xpc_hb_check_interval,
103          .maxlen = sizeof(int),
104          .mode = 0644,
105          .proc_handler = proc_dointvec_minmax,
106          .extra1 = &xpc_hb_check_min_interval,
107          .extra2 = &xpc_hb_check_max_interval},
108         {}
109 };
110 static ctl_table xpc_sys_xpc_dir[] = {
111         {
112          .procname = "hb",
113          .mode = 0555,
114          .child = xpc_sys_xpc_hb_dir},
115         {
116          .procname = "disengage_timelimit",
117          .data = &xpc_disengage_timelimit,
118          .maxlen = sizeof(int),
119          .mode = 0644,
120          .proc_handler = proc_dointvec_minmax,
121          .extra1 = &xpc_disengage_min_timelimit,
122          .extra2 = &xpc_disengage_max_timelimit},
123         {}
124 };
125 static ctl_table xpc_sys_dir[] = {
126         {
127          .procname = "xpc",
128          .mode = 0555,
129          .child = xpc_sys_xpc_dir},
130         {}
131 };
132 static struct ctl_table_header *xpc_sysctl;
133
134 /* non-zero if any remote partition disengage was timed out */
135 int xpc_disengage_timedout;
136
137 /* #of activate IRQs received and not yet processed */
138 int xpc_activate_IRQ_rcvd;
139 DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
140
141 /* IRQ handler notifies this wait queue on receipt of an IRQ */
142 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
143
144 static unsigned long xpc_hb_check_timeout;
145 static struct timer_list xpc_hb_timer;
146
147 /* notification that the xpc_hb_checker thread has exited */
148 static DECLARE_COMPLETION(xpc_hb_checker_exited);
149
150 /* notification that the xpc_discovery thread has exited */
151 static DECLARE_COMPLETION(xpc_discovery_exited);
152
153 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
154
155 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
156 static struct notifier_block xpc_reboot_notifier = {
157         .notifier_call = xpc_system_reboot,
158 };
159
160 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
161 static struct notifier_block xpc_die_notifier = {
162         .notifier_call = xpc_system_die,
163 };
164
165 struct xpc_arch_operations xpc_arch_ops;
166
167 /*
168  * Timer function to enforce the timelimit on the partition disengage.
169  */
170 static void
171 xpc_timeout_partition_disengage(unsigned long data)
172 {
173         struct xpc_partition *part = (struct xpc_partition *)data;
174
175         DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
176
177         (void)xpc_partition_disengaged(part);
178
179         DBUG_ON(part->disengage_timeout != 0);
180         DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
181 }
182
183 /*
184  * Timer to produce the heartbeat.  The timer structures function is
185  * already set when this is initially called.  A tunable is used to
186  * specify when the next timeout should occur.
187  */
188 static void
189 xpc_hb_beater(unsigned long dummy)
190 {
191         xpc_arch_ops.increment_heartbeat();
192
193         if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
194                 wake_up_interruptible(&xpc_activate_IRQ_wq);
195
196         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
197         add_timer(&xpc_hb_timer);
198 }
199
200 static void
201 xpc_start_hb_beater(void)
202 {
203         xpc_arch_ops.heartbeat_init();
204         init_timer(&xpc_hb_timer);
205         xpc_hb_timer.function = xpc_hb_beater;
206         xpc_hb_beater(0);
207 }
208
209 static void
210 xpc_stop_hb_beater(void)
211 {
212         del_timer_sync(&xpc_hb_timer);
213         xpc_arch_ops.heartbeat_exit();
214 }
215
216 /*
217  * At periodic intervals, scan through all active partitions and ensure
218  * their heartbeat is still active.  If not, the partition is deactivated.
219  */
220 static void
221 xpc_check_remote_hb(void)
222 {
223         struct xpc_partition *part;
224         short partid;
225         enum xp_retval ret;
226
227         for (partid = 0; partid < xp_max_npartitions; partid++) {
228
229                 if (xpc_exiting)
230                         break;
231
232                 if (partid == xp_partition_id)
233                         continue;
234
235                 part = &xpc_partitions[partid];
236
237                 if (part->act_state == XPC_P_AS_INACTIVE ||
238                     part->act_state == XPC_P_AS_DEACTIVATING) {
239                         continue;
240                 }
241
242                 ret = xpc_arch_ops.get_remote_heartbeat(part);
243                 if (ret != xpSuccess)
244                         XPC_DEACTIVATE_PARTITION(part, ret);
245         }
246 }
247
248 /*
249  * This thread is responsible for nearly all of the partition
250  * activation/deactivation.
251  */
252 static int
253 xpc_hb_checker(void *ignore)
254 {
255         int force_IRQ = 0;
256
257         /* this thread was marked active by xpc_hb_init() */
258
259         set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
260
261         /* set our heartbeating to other partitions into motion */
262         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
263         xpc_start_hb_beater();
264
265         while (!xpc_exiting) {
266
267                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
268                         "been received\n",
269                         (int)(xpc_hb_check_timeout - jiffies),
270                         xpc_activate_IRQ_rcvd);
271
272                 /* checking of remote heartbeats is skewed by IRQ handling */
273                 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
274                         xpc_hb_check_timeout = jiffies +
275                             (xpc_hb_check_interval * HZ);
276
277                         dev_dbg(xpc_part, "checking remote heartbeats\n");
278                         xpc_check_remote_hb();
279
280                         /*
281                          * On sn2 we need to periodically recheck to ensure no
282                          * IRQ/amo pairs have been missed.
283                          */
284                         if (is_shub())
285                                 force_IRQ = 1;
286                 }
287
288                 /* check for outstanding IRQs */
289                 if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
290                         force_IRQ = 0;
291                         dev_dbg(xpc_part, "processing activate IRQs "
292                                 "received\n");
293                         xpc_arch_ops.process_activate_IRQ_rcvd();
294                 }
295
296                 /* wait for IRQ or timeout */
297                 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
298                                                (time_is_before_eq_jiffies(
299                                                 xpc_hb_check_timeout) ||
300                                                 xpc_activate_IRQ_rcvd > 0 ||
301                                                 xpc_exiting));
302         }
303
304         xpc_stop_hb_beater();
305
306         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
307
308         /* mark this thread as having exited */
309         complete(&xpc_hb_checker_exited);
310         return 0;
311 }
312
313 /*
314  * This thread will attempt to discover other partitions to activate
315  * based on info provided by SAL. This new thread is short lived and
316  * will exit once discovery is complete.
317  */
318 static int
319 xpc_initiate_discovery(void *ignore)
320 {
321         xpc_discovery();
322
323         dev_dbg(xpc_part, "discovery thread is exiting\n");
324
325         /* mark this thread as having exited */
326         complete(&xpc_discovery_exited);
327         return 0;
328 }
329
330 /*
331  * The first kthread assigned to a newly activated partition is the one
332  * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
333  * that kthread until the partition is brought down, at which time that kthread
334  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
335  * that XPC has dismantled all communication infrastructure for the associated
336  * partition.) This kthread becomes the channel manager for that partition.
337  *
338  * Each active partition has a channel manager, who, besides connecting and
339  * disconnecting channels, will ensure that each of the partition's connected
340  * channels has the required number of assigned kthreads to get the work done.
341  */
342 static void
343 xpc_channel_mgr(struct xpc_partition *part)
344 {
345         while (part->act_state != XPC_P_AS_DEACTIVATING ||
346                atomic_read(&part->nchannels_active) > 0 ||
347                !xpc_partition_disengaged(part)) {
348
349                 xpc_process_sent_chctl_flags(part);
350
351                 /*
352                  * Wait until we've been requested to activate kthreads or
353                  * all of the channel's message queues have been torn down or
354                  * a signal is pending.
355                  *
356                  * The channel_mgr_requests is set to 1 after being awakened,
357                  * This is done to prevent the channel mgr from making one pass
358                  * through the loop for each request, since he will
359                  * be servicing all the requests in one pass. The reason it's
360                  * set to 1 instead of 0 is so that other kthreads will know
361                  * that the channel mgr is running and won't bother trying to
362                  * wake him up.
363                  */
364                 atomic_dec(&part->channel_mgr_requests);
365                 (void)wait_event_interruptible(part->channel_mgr_wq,
366                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
367                                  part->chctl.all_flags != 0 ||
368                                  (part->act_state == XPC_P_AS_DEACTIVATING &&
369                                  atomic_read(&part->nchannels_active) == 0 &&
370                                  xpc_partition_disengaged(part))));
371                 atomic_set(&part->channel_mgr_requests, 1);
372         }
373 }
374
375 /*
376  * Guarantee that the kzalloc'd memory is cacheline aligned.
377  */
378 void *
379 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
380 {
381         /* see if kzalloc will give us cachline aligned memory by default */
382         *base = kzalloc(size, flags);
383         if (*base == NULL)
384                 return NULL;
385
386         if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
387                 return *base;
388
389         kfree(*base);
390
391         /* nope, we'll have to do it ourselves */
392         *base = kzalloc(size + L1_CACHE_BYTES, flags);
393         if (*base == NULL)
394                 return NULL;
395
396         return (void *)L1_CACHE_ALIGN((u64)*base);
397 }
398
399 /*
400  * Setup the channel structures necessary to support XPartition Communication
401  * between the specified remote partition and the local one.
402  */
403 static enum xp_retval
404 xpc_setup_ch_structures(struct xpc_partition *part)
405 {
406         enum xp_retval ret;
407         int ch_number;
408         struct xpc_channel *ch;
409         short partid = XPC_PARTID(part);
410
411         /*
412          * Allocate all of the channel structures as a contiguous chunk of
413          * memory.
414          */
415         DBUG_ON(part->channels != NULL);
416         part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
417                                  GFP_KERNEL);
418         if (part->channels == NULL) {
419                 dev_err(xpc_chan, "can't get memory for channels\n");
420                 return xpNoMemory;
421         }
422
423         /* allocate the remote open and close args */
424
425         part->remote_openclose_args =
426             xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
427                                           GFP_KERNEL, &part->
428                                           remote_openclose_args_base);
429         if (part->remote_openclose_args == NULL) {
430                 dev_err(xpc_chan, "can't get memory for remote connect args\n");
431                 ret = xpNoMemory;
432                 goto out_1;
433         }
434
435         part->chctl.all_flags = 0;
436         spin_lock_init(&part->chctl_lock);
437
438         atomic_set(&part->channel_mgr_requests, 1);
439         init_waitqueue_head(&part->channel_mgr_wq);
440
441         part->nchannels = XPC_MAX_NCHANNELS;
442
443         atomic_set(&part->nchannels_active, 0);
444         atomic_set(&part->nchannels_engaged, 0);
445
446         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
447                 ch = &part->channels[ch_number];
448
449                 ch->partid = partid;
450                 ch->number = ch_number;
451                 ch->flags = XPC_C_DISCONNECTED;
452
453                 atomic_set(&ch->kthreads_assigned, 0);
454                 atomic_set(&ch->kthreads_idle, 0);
455                 atomic_set(&ch->kthreads_active, 0);
456
457                 atomic_set(&ch->references, 0);
458                 atomic_set(&ch->n_to_notify, 0);
459
460                 spin_lock_init(&ch->lock);
461                 init_completion(&ch->wdisconnect_wait);
462
463                 atomic_set(&ch->n_on_msg_allocate_wq, 0);
464                 init_waitqueue_head(&ch->msg_allocate_wq);
465                 init_waitqueue_head(&ch->idle_wq);
466         }
467
468         ret = xpc_arch_ops.setup_ch_structures(part);
469         if (ret != xpSuccess)
470                 goto out_2;
471
472         /*
473          * With the setting of the partition setup_state to XPC_P_SS_SETUP,
474          * we're declaring that this partition is ready to go.
475          */
476         part->setup_state = XPC_P_SS_SETUP;
477
478         return xpSuccess;
479
480         /* setup of ch structures failed */
481 out_2:
482         kfree(part->remote_openclose_args_base);
483         part->remote_openclose_args = NULL;
484 out_1:
485         kfree(part->channels);
486         part->channels = NULL;
487         return ret;
488 }
489
490 /*
491  * Teardown the channel structures necessary to support XPartition Communication
492  * between the specified remote partition and the local one.
493  */
494 static void
495 xpc_teardown_ch_structures(struct xpc_partition *part)
496 {
497         DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
498         DBUG_ON(atomic_read(&part->nchannels_active) != 0);
499
500         /*
501          * Make this partition inaccessible to local processes by marking it
502          * as no longer setup. Then wait before proceeding with the teardown
503          * until all existing references cease.
504          */
505         DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
506         part->setup_state = XPC_P_SS_WTEARDOWN;
507
508         wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
509
510         /* now we can begin tearing down the infrastructure */
511
512         xpc_arch_ops.teardown_ch_structures(part);
513
514         kfree(part->remote_openclose_args_base);
515         part->remote_openclose_args = NULL;
516         kfree(part->channels);
517         part->channels = NULL;
518
519         part->setup_state = XPC_P_SS_TORNDOWN;
520 }
521
522 /*
523  * When XPC HB determines that a partition has come up, it will create a new
524  * kthread and that kthread will call this function to attempt to set up the
525  * basic infrastructure used for Cross Partition Communication with the newly
526  * upped partition.
527  *
528  * The kthread that was created by XPC HB and which setup the XPC
529  * infrastructure will remain assigned to the partition becoming the channel
530  * manager for that partition until the partition is deactivating, at which
531  * time the kthread will teardown the XPC infrastructure and then exit.
532  */
533 static int
534 xpc_activating(void *__partid)
535 {
536         short partid = (u64)__partid;
537         struct xpc_partition *part = &xpc_partitions[partid];
538         unsigned long irq_flags;
539
540         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
541
542         spin_lock_irqsave(&part->act_lock, irq_flags);
543
544         if (part->act_state == XPC_P_AS_DEACTIVATING) {
545                 part->act_state = XPC_P_AS_INACTIVE;
546                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
547                 part->remote_rp_pa = 0;
548                 return 0;
549         }
550
551         /* indicate the thread is activating */
552         DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
553         part->act_state = XPC_P_AS_ACTIVATING;
554
555         XPC_SET_REASON(part, 0, 0);
556         spin_unlock_irqrestore(&part->act_lock, irq_flags);
557
558         dev_dbg(xpc_part, "activating partition %d\n", partid);
559
560         xpc_arch_ops.allow_hb(partid);
561
562         if (xpc_setup_ch_structures(part) == xpSuccess) {
563                 (void)xpc_part_ref(part);       /* this will always succeed */
564
565                 if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
566                         xpc_mark_partition_active(part);
567                         xpc_channel_mgr(part);
568                         /* won't return until partition is deactivating */
569                 }
570
571                 xpc_part_deref(part);
572                 xpc_teardown_ch_structures(part);
573         }
574
575         xpc_arch_ops.disallow_hb(partid);
576         xpc_mark_partition_inactive(part);
577
578         if (part->reason == xpReactivating) {
579                 /* interrupting ourselves results in activating partition */
580                 xpc_arch_ops.request_partition_reactivation(part);
581         }
582
583         return 0;
584 }
585
586 void
587 xpc_activate_partition(struct xpc_partition *part)
588 {
589         short partid = XPC_PARTID(part);
590         unsigned long irq_flags;
591         struct task_struct *kthread;
592
593         spin_lock_irqsave(&part->act_lock, irq_flags);
594
595         DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
596
597         part->act_state = XPC_P_AS_ACTIVATION_REQ;
598         XPC_SET_REASON(part, xpCloneKThread, __LINE__);
599
600         spin_unlock_irqrestore(&part->act_lock, irq_flags);
601
602         kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
603                               partid);
604         if (IS_ERR(kthread)) {
605                 spin_lock_irqsave(&part->act_lock, irq_flags);
606                 part->act_state = XPC_P_AS_INACTIVE;
607                 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
608                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
609         }
610 }
611
612 void
613 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
614 {
615         int idle = atomic_read(&ch->kthreads_idle);
616         int assigned = atomic_read(&ch->kthreads_assigned);
617         int wakeup;
618
619         DBUG_ON(needed <= 0);
620
621         if (idle > 0) {
622                 wakeup = (needed > idle) ? idle : needed;
623                 needed -= wakeup;
624
625                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
626                         "channel=%d\n", wakeup, ch->partid, ch->number);
627
628                 /* only wakeup the requested number of kthreads */
629                 wake_up_nr(&ch->idle_wq, wakeup);
630         }
631
632         if (needed <= 0)
633                 return;
634
635         if (needed + assigned > ch->kthreads_assigned_limit) {
636                 needed = ch->kthreads_assigned_limit - assigned;
637                 if (needed <= 0)
638                         return;
639         }
640
641         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
642                 needed, ch->partid, ch->number);
643
644         xpc_create_kthreads(ch, needed, 0);
645 }
646
647 /*
648  * This function is where XPC's kthreads wait for messages to deliver.
649  */
650 static void
651 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
652 {
653         int (*n_of_deliverable_payloads) (struct xpc_channel *) =
654                 xpc_arch_ops.n_of_deliverable_payloads;
655
656         do {
657                 /* deliver messages to their intended recipients */
658
659                 while (n_of_deliverable_payloads(ch) > 0 &&
660                        !(ch->flags & XPC_C_DISCONNECTING)) {
661                         xpc_deliver_payload(ch);
662                 }
663
664                 if (atomic_inc_return(&ch->kthreads_idle) >
665                     ch->kthreads_idle_limit) {
666                         /* too many idle kthreads on this channel */
667                         atomic_dec(&ch->kthreads_idle);
668                         break;
669                 }
670
671                 dev_dbg(xpc_chan, "idle kthread calling "
672                         "wait_event_interruptible_exclusive()\n");
673
674                 (void)wait_event_interruptible_exclusive(ch->idle_wq,
675                                 (n_of_deliverable_payloads(ch) > 0 ||
676                                  (ch->flags & XPC_C_DISCONNECTING)));
677
678                 atomic_dec(&ch->kthreads_idle);
679
680         } while (!(ch->flags & XPC_C_DISCONNECTING));
681 }
682
683 static int
684 xpc_kthread_start(void *args)
685 {
686         short partid = XPC_UNPACK_ARG1(args);
687         u16 ch_number = XPC_UNPACK_ARG2(args);
688         struct xpc_partition *part = &xpc_partitions[partid];
689         struct xpc_channel *ch;
690         int n_needed;
691         unsigned long irq_flags;
692         int (*n_of_deliverable_payloads) (struct xpc_channel *) =
693                 xpc_arch_ops.n_of_deliverable_payloads;
694
695         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
696                 partid, ch_number);
697
698         ch = &part->channels[ch_number];
699
700         if (!(ch->flags & XPC_C_DISCONNECTING)) {
701
702                 /* let registerer know that connection has been established */
703
704                 spin_lock_irqsave(&ch->lock, irq_flags);
705                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
706                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
707                         spin_unlock_irqrestore(&ch->lock, irq_flags);
708
709                         xpc_connected_callout(ch);
710
711                         spin_lock_irqsave(&ch->lock, irq_flags);
712                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
713                         spin_unlock_irqrestore(&ch->lock, irq_flags);
714
715                         /*
716                          * It is possible that while the callout was being
717                          * made that the remote partition sent some messages.
718                          * If that is the case, we may need to activate
719                          * additional kthreads to help deliver them. We only
720                          * need one less than total #of messages to deliver.
721                          */
722                         n_needed = n_of_deliverable_payloads(ch) - 1;
723                         if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
724                                 xpc_activate_kthreads(ch, n_needed);
725
726                 } else {
727                         spin_unlock_irqrestore(&ch->lock, irq_flags);
728                 }
729
730                 xpc_kthread_waitmsgs(part, ch);
731         }
732
733         /* let registerer know that connection is disconnecting */
734
735         spin_lock_irqsave(&ch->lock, irq_flags);
736         if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
737             !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
738                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
739                 spin_unlock_irqrestore(&ch->lock, irq_flags);
740
741                 xpc_disconnect_callout(ch, xpDisconnecting);
742
743                 spin_lock_irqsave(&ch->lock, irq_flags);
744                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
745         }
746         spin_unlock_irqrestore(&ch->lock, irq_flags);
747
748         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
749             atomic_dec_return(&part->nchannels_engaged) == 0) {
750                 xpc_arch_ops.indicate_partition_disengaged(part);
751         }
752
753         xpc_msgqueue_deref(ch);
754
755         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
756                 partid, ch_number);
757
758         xpc_part_deref(part);
759         return 0;
760 }
761
762 /*
763  * For each partition that XPC has established communications with, there is
764  * a minimum of one kernel thread assigned to perform any operation that
765  * may potentially sleep or block (basically the callouts to the asynchronous
766  * functions registered via xpc_connect()).
767  *
768  * Additional kthreads are created and destroyed by XPC as the workload
769  * demands.
770  *
771  * A kthread is assigned to one of the active channels that exists for a given
772  * partition.
773  */
774 void
775 xpc_create_kthreads(struct xpc_channel *ch, int needed,
776                     int ignore_disconnecting)
777 {
778         unsigned long irq_flags;
779         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
780         struct xpc_partition *part = &xpc_partitions[ch->partid];
781         struct task_struct *kthread;
782         void (*indicate_partition_disengaged) (struct xpc_partition *) =
783                 xpc_arch_ops.indicate_partition_disengaged;
784
785         while (needed-- > 0) {
786
787                 /*
788                  * The following is done on behalf of the newly created
789                  * kthread. That kthread is responsible for doing the
790                  * counterpart to the following before it exits.
791                  */
792                 if (ignore_disconnecting) {
793                         if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
794                                 /* kthreads assigned had gone to zero */
795                                 BUG_ON(!(ch->flags &
796                                          XPC_C_DISCONNECTINGCALLOUT_MADE));
797                                 break;
798                         }
799
800                 } else if (ch->flags & XPC_C_DISCONNECTING) {
801                         break;
802
803                 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
804                            atomic_inc_return(&part->nchannels_engaged) == 1) {
805                         xpc_arch_ops.indicate_partition_engaged(part);
806                 }
807                 (void)xpc_part_ref(part);
808                 xpc_msgqueue_ref(ch);
809
810                 kthread = kthread_run(xpc_kthread_start, (void *)args,
811                                       "xpc%02dc%d", ch->partid, ch->number);
812                 if (IS_ERR(kthread)) {
813                         /* the fork failed */
814
815                         /*
816                          * NOTE: if (ignore_disconnecting &&
817                          * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
818                          * then we'll deadlock if all other kthreads assigned
819                          * to this channel are blocked in the channel's
820                          * registerer, because the only thing that will unblock
821                          * them is the xpDisconnecting callout that this
822                          * failed kthread_run() would have made.
823                          */
824
825                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
826                             atomic_dec_return(&part->nchannels_engaged) == 0) {
827                                 indicate_partition_disengaged(part);
828                         }
829                         xpc_msgqueue_deref(ch);
830                         xpc_part_deref(part);
831
832                         if (atomic_read(&ch->kthreads_assigned) <
833                             ch->kthreads_idle_limit) {
834                                 /*
835                                  * Flag this as an error only if we have an
836                                  * insufficient #of kthreads for the channel
837                                  * to function.
838                                  */
839                                 spin_lock_irqsave(&ch->lock, irq_flags);
840                                 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
841                                                        &irq_flags);
842                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
843                         }
844                         break;
845                 }
846         }
847 }
848
849 void
850 xpc_disconnect_wait(int ch_number)
851 {
852         unsigned long irq_flags;
853         short partid;
854         struct xpc_partition *part;
855         struct xpc_channel *ch;
856         int wakeup_channel_mgr;
857
858         /* now wait for all callouts to the caller's function to cease */
859         for (partid = 0; partid < xp_max_npartitions; partid++) {
860                 part = &xpc_partitions[partid];
861
862                 if (!xpc_part_ref(part))
863                         continue;
864
865                 ch = &part->channels[ch_number];
866
867                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
868                         xpc_part_deref(part);
869                         continue;
870                 }
871
872                 wait_for_completion(&ch->wdisconnect_wait);
873
874                 spin_lock_irqsave(&ch->lock, irq_flags);
875                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
876                 wakeup_channel_mgr = 0;
877
878                 if (ch->delayed_chctl_flags) {
879                         if (part->act_state != XPC_P_AS_DEACTIVATING) {
880                                 spin_lock(&part->chctl_lock);
881                                 part->chctl.flags[ch->number] |=
882                                     ch->delayed_chctl_flags;
883                                 spin_unlock(&part->chctl_lock);
884                                 wakeup_channel_mgr = 1;
885                         }
886                         ch->delayed_chctl_flags = 0;
887                 }
888
889                 ch->flags &= ~XPC_C_WDISCONNECT;
890                 spin_unlock_irqrestore(&ch->lock, irq_flags);
891
892                 if (wakeup_channel_mgr)
893                         xpc_wakeup_channel_mgr(part);
894
895                 xpc_part_deref(part);
896         }
897 }
898
899 static int
900 xpc_setup_partitions(void)
901 {
902         short partid;
903         struct xpc_partition *part;
904
905         xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
906                                  xp_max_npartitions, GFP_KERNEL);
907         if (xpc_partitions == NULL) {
908                 dev_err(xpc_part, "can't get memory for partition structure\n");
909                 return -ENOMEM;
910         }
911
912         /*
913          * The first few fields of each entry of xpc_partitions[] need to
914          * be initialized now so that calls to xpc_connect() and
915          * xpc_disconnect() can be made prior to the activation of any remote
916          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
917          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
918          * PARTITION HAS BEEN ACTIVATED.
919          */
920         for (partid = 0; partid < xp_max_npartitions; partid++) {
921                 part = &xpc_partitions[partid];
922
923                 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
924
925                 part->activate_IRQ_rcvd = 0;
926                 spin_lock_init(&part->act_lock);
927                 part->act_state = XPC_P_AS_INACTIVE;
928                 XPC_SET_REASON(part, 0, 0);
929
930                 init_timer(&part->disengage_timer);
931                 part->disengage_timer.function =
932                     xpc_timeout_partition_disengage;
933                 part->disengage_timer.data = (unsigned long)part;
934
935                 part->setup_state = XPC_P_SS_UNSET;
936                 init_waitqueue_head(&part->teardown_wq);
937                 atomic_set(&part->references, 0);
938         }
939
940         return xpc_arch_ops.setup_partitions();
941 }
942
943 static void
944 xpc_teardown_partitions(void)
945 {
946         xpc_arch_ops.teardown_partitions();
947         kfree(xpc_partitions);
948 }
949
950 static void
951 xpc_do_exit(enum xp_retval reason)
952 {
953         short partid;
954         int active_part_count, printed_waiting_msg = 0;
955         struct xpc_partition *part;
956         unsigned long printmsg_time, disengage_timeout = 0;
957
958         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
959         DBUG_ON(xpc_exiting == 1);
960
961         /*
962          * Let the heartbeat checker thread and the discovery thread
963          * (if one is running) know that they should exit. Also wake up
964          * the heartbeat checker thread in case it's sleeping.
965          */
966         xpc_exiting = 1;
967         wake_up_interruptible(&xpc_activate_IRQ_wq);
968
969         /* wait for the discovery thread to exit */
970         wait_for_completion(&xpc_discovery_exited);
971
972         /* wait for the heartbeat checker thread to exit */
973         wait_for_completion(&xpc_hb_checker_exited);
974
975         /* sleep for a 1/3 of a second or so */
976         (void)msleep_interruptible(300);
977
978         /* wait for all partitions to become inactive */
979
980         printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
981         xpc_disengage_timedout = 0;
982
983         do {
984                 active_part_count = 0;
985
986                 for (partid = 0; partid < xp_max_npartitions; partid++) {
987                         part = &xpc_partitions[partid];
988
989                         if (xpc_partition_disengaged(part) &&
990                             part->act_state == XPC_P_AS_INACTIVE) {
991                                 continue;
992                         }
993
994                         active_part_count++;
995
996                         XPC_DEACTIVATE_PARTITION(part, reason);
997
998                         if (part->disengage_timeout > disengage_timeout)
999                                 disengage_timeout = part->disengage_timeout;
1000                 }
1001
1002                 if (xpc_arch_ops.any_partition_engaged()) {
1003                         if (time_is_before_jiffies(printmsg_time)) {
1004                                 dev_info(xpc_part, "waiting for remote "
1005                                          "partitions to deactivate, timeout in "
1006                                          "%ld seconds\n", (disengage_timeout -
1007                                          jiffies) / HZ);
1008                                 printmsg_time = jiffies +
1009                                     (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
1010                                 printed_waiting_msg = 1;
1011                         }
1012
1013                 } else if (active_part_count > 0) {
1014                         if (printed_waiting_msg) {
1015                                 dev_info(xpc_part, "waiting for local partition"
1016                                          " to deactivate\n");
1017                                 printed_waiting_msg = 0;
1018                         }
1019
1020                 } else {
1021                         if (!xpc_disengage_timedout) {
1022                                 dev_info(xpc_part, "all partitions have "
1023                                          "deactivated\n");
1024                         }
1025                         break;
1026                 }
1027
1028                 /* sleep for a 1/3 of a second or so */
1029                 (void)msleep_interruptible(300);
1030
1031         } while (1);
1032
1033         DBUG_ON(xpc_arch_ops.any_partition_engaged());
1034
1035         xpc_teardown_rsvd_page();
1036
1037         if (reason == xpUnloading) {
1038                 (void)unregister_die_notifier(&xpc_die_notifier);
1039                 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1040         }
1041
1042         /* clear the interface to XPC's functions */
1043         xpc_clear_interface();
1044
1045         if (xpc_sysctl)
1046                 unregister_sysctl_table(xpc_sysctl);
1047
1048         xpc_teardown_partitions();
1049
1050         if (is_shub())
1051                 xpc_exit_sn2();
1052         else if (is_uv())
1053                 xpc_exit_uv();
1054 }
1055
1056 /*
1057  * This function is called when the system is being rebooted.
1058  */
1059 static int
1060 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1061 {
1062         enum xp_retval reason;
1063
1064         switch (event) {
1065         case SYS_RESTART:
1066                 reason = xpSystemReboot;
1067                 break;
1068         case SYS_HALT:
1069                 reason = xpSystemHalt;
1070                 break;
1071         case SYS_POWER_OFF:
1072                 reason = xpSystemPoweroff;
1073                 break;
1074         default:
1075                 reason = xpSystemGoingDown;
1076         }
1077
1078         xpc_do_exit(reason);
1079         return NOTIFY_DONE;
1080 }
1081
1082 /*
1083  * Notify other partitions to deactivate from us by first disengaging from all
1084  * references to our memory.
1085  */
1086 static void
1087 xpc_die_deactivate(void)
1088 {
1089         struct xpc_partition *part;
1090         short partid;
1091         int any_engaged;
1092         long keep_waiting;
1093         long wait_to_print;
1094
1095         /* keep xpc_hb_checker thread from doing anything (just in case) */
1096         xpc_exiting = 1;
1097
1098         xpc_arch_ops.disallow_all_hbs();   /*indicate we're deactivated */
1099
1100         for (partid = 0; partid < xp_max_npartitions; partid++) {
1101                 part = &xpc_partitions[partid];
1102
1103                 if (xpc_arch_ops.partition_engaged(partid) ||
1104                     part->act_state != XPC_P_AS_INACTIVE) {
1105                         xpc_arch_ops.request_partition_deactivation(part);
1106                         xpc_arch_ops.indicate_partition_disengaged(part);
1107                 }
1108         }
1109
1110         /*
1111          * Though we requested that all other partitions deactivate from us,
1112          * we only wait until they've all disengaged or we've reached the
1113          * defined timelimit.
1114          *
1115          * Given that one iteration through the following while-loop takes
1116          * approximately 200 microseconds, calculate the #of loops to take
1117          * before bailing and the #of loops before printing a waiting message.
1118          */
1119         keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1120         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1121
1122         while (1) {
1123                 any_engaged = xpc_arch_ops.any_partition_engaged();
1124                 if (!any_engaged) {
1125                         dev_info(xpc_part, "all partitions have deactivated\n");
1126                         break;
1127                 }
1128
1129                 if (!keep_waiting--) {
1130                         for (partid = 0; partid < xp_max_npartitions;
1131                              partid++) {
1132                                 if (xpc_arch_ops.partition_engaged(partid)) {
1133                                         dev_info(xpc_part, "deactivate from "
1134                                                  "remote partition %d timed "
1135                                                  "out\n", partid);
1136                                 }
1137                         }
1138                         break;
1139                 }
1140
1141                 if (!wait_to_print--) {
1142                         dev_info(xpc_part, "waiting for remote partitions to "
1143                                  "deactivate, timeout in %ld seconds\n",
1144                                  keep_waiting / (1000 * 5));
1145                         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1146                             1000 * 5;
1147                 }
1148
1149                 udelay(200);
1150         }
1151 }
1152
1153 /*
1154  * This function is called when the system is being restarted or halted due
1155  * to some sort of system failure. If this is the case we need to notify the
1156  * other partitions to disengage from all references to our memory.
1157  * This function can also be called when our heartbeater could be offlined
1158  * for a time. In this case we need to notify other partitions to not worry
1159  * about the lack of a heartbeat.
1160  */
1161 static int
1162 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1163 {
1164 #ifdef CONFIG_IA64              /* !!! temporary kludge */
1165         switch (event) {
1166         case DIE_MACHINE_RESTART:
1167         case DIE_MACHINE_HALT:
1168                 xpc_die_deactivate();
1169                 break;
1170
1171         case DIE_KDEBUG_ENTER:
1172                 /* Should lack of heartbeat be ignored by other partitions? */
1173                 if (!xpc_kdebug_ignore)
1174                         break;
1175
1176                 /* fall through */
1177         case DIE_MCA_MONARCH_ENTER:
1178         case DIE_INIT_MONARCH_ENTER:
1179                 xpc_arch_ops.offline_heartbeat();
1180                 break;
1181
1182         case DIE_KDEBUG_LEAVE:
1183                 /* Is lack of heartbeat being ignored by other partitions? */
1184                 if (!xpc_kdebug_ignore)
1185                         break;
1186
1187                 /* fall through */
1188         case DIE_MCA_MONARCH_LEAVE:
1189         case DIE_INIT_MONARCH_LEAVE:
1190                 xpc_arch_ops.online_heartbeat();
1191                 break;
1192         }
1193 #else
1194         xpc_die_deactivate();
1195 #endif
1196
1197         return NOTIFY_DONE;
1198 }
1199
1200 int __init
1201 xpc_init(void)
1202 {
1203         int ret;
1204         struct task_struct *kthread;
1205
1206         dev_set_name(xpc_part, "part");
1207         dev_set_name(xpc_chan, "chan");
1208
1209         if (is_shub()) {
1210                 /*
1211                  * The ia64-sn2 architecture supports at most 64 partitions.
1212                  * And the inability to unregister remote amos restricts us
1213                  * further to only support exactly 64 partitions on this
1214                  * architecture, no less.
1215                  */
1216                 if (xp_max_npartitions != 64) {
1217                         dev_err(xpc_part, "max #of partitions not set to 64\n");
1218                         ret = -EINVAL;
1219                 } else {
1220                         ret = xpc_init_sn2();
1221                 }
1222
1223         } else if (is_uv()) {
1224                 ret = xpc_init_uv();
1225
1226         } else {
1227                 ret = -ENODEV;
1228         }
1229
1230         if (ret != 0)
1231                 return ret;
1232
1233         ret = xpc_setup_partitions();
1234         if (ret != 0) {
1235                 dev_err(xpc_part, "can't get memory for partition structure\n");
1236                 goto out_1;
1237         }
1238
1239         xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1240
1241         /*
1242          * Fill the partition reserved page with the information needed by
1243          * other partitions to discover we are alive and establish initial
1244          * communications.
1245          */
1246         ret = xpc_setup_rsvd_page();
1247         if (ret != 0) {
1248                 dev_err(xpc_part, "can't setup our reserved page\n");
1249                 goto out_2;
1250         }
1251
1252         /* add ourselves to the reboot_notifier_list */
1253         ret = register_reboot_notifier(&xpc_reboot_notifier);
1254         if (ret != 0)
1255                 dev_warn(xpc_part, "can't register reboot notifier\n");
1256
1257         /* add ourselves to the die_notifier list */
1258         ret = register_die_notifier(&xpc_die_notifier);
1259         if (ret != 0)
1260                 dev_warn(xpc_part, "can't register die notifier\n");
1261
1262         /*
1263          * The real work-horse behind xpc.  This processes incoming
1264          * interrupts and monitors remote heartbeats.
1265          */
1266         kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1267         if (IS_ERR(kthread)) {
1268                 dev_err(xpc_part, "failed while forking hb check thread\n");
1269                 ret = -EBUSY;
1270                 goto out_3;
1271         }
1272
1273         /*
1274          * Startup a thread that will attempt to discover other partitions to
1275          * activate based on info provided by SAL. This new thread is short
1276          * lived and will exit once discovery is complete.
1277          */
1278         kthread = kthread_run(xpc_initiate_discovery, NULL,
1279                               XPC_DISCOVERY_THREAD_NAME);
1280         if (IS_ERR(kthread)) {
1281                 dev_err(xpc_part, "failed while forking discovery thread\n");
1282
1283                 /* mark this new thread as a non-starter */
1284                 complete(&xpc_discovery_exited);
1285
1286                 xpc_do_exit(xpUnloading);
1287                 return -EBUSY;
1288         }
1289
1290         /* set the interface to point at XPC's functions */
1291         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1292                           xpc_initiate_send, xpc_initiate_send_notify,
1293                           xpc_initiate_received, xpc_initiate_partid_to_nasids);
1294
1295         return 0;
1296
1297         /* initialization was not successful */
1298 out_3:
1299         xpc_teardown_rsvd_page();
1300
1301         (void)unregister_die_notifier(&xpc_die_notifier);
1302         (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1303 out_2:
1304         if (xpc_sysctl)
1305                 unregister_sysctl_table(xpc_sysctl);
1306
1307         xpc_teardown_partitions();
1308 out_1:
1309         if (is_shub())
1310                 xpc_exit_sn2();
1311         else if (is_uv())
1312                 xpc_exit_uv();
1313         return ret;
1314 }
1315
1316 module_init(xpc_init);
1317
1318 void __exit
1319 xpc_exit(void)
1320 {
1321         xpc_do_exit(xpUnloading);
1322 }
1323
1324 module_exit(xpc_exit);
1325
1326 MODULE_AUTHOR("Silicon Graphics, Inc.");
1327 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1328 MODULE_LICENSE("GPL");
1329
1330 module_param(xpc_hb_interval, int, 0);
1331 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1332                  "heartbeat increments.");
1333
1334 module_param(xpc_hb_check_interval, int, 0);
1335 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1336                  "heartbeat checks.");
1337
1338 module_param(xpc_disengage_timelimit, int, 0);
1339 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1340                  "for disengage to complete.");
1341
1342 module_param(xpc_kdebug_ignore, int, 0);
1343 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1344                  "other partitions when dropping into kdebug.");