GNU Linux-libre 6.8.7-gnu
[releases.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  * (C) Copyright 2020 Hewlett Packard Enterprise Development LP
7  * Copyright (c) 2004-2009 Silicon Graphics, Inc.  All Rights Reserved.
8  */
9
10 /*
11  * Cross Partition Communication (XPC) support - standard version.
12  *
13  *      XPC provides a message passing capability that crosses partition
14  *      boundaries. This module is made up of two parts:
15  *
16  *          partition   This part detects the presence/absence of other
17  *                      partitions. It provides a heartbeat and monitors
18  *                      the heartbeats of other partitions.
19  *
20  *          channel     This part manages the channels and sends/receives
21  *                      messages across them to/from other partitions.
22  *
23  *      There are a couple of additional functions residing in XP, which
24  *      provide an interface to XPC for its users.
25  *
26  *
27  *      Caveats:
28  *
29  *        . Currently on sn2, we have no way to determine which nasid an IRQ
30  *          came from. Thus, xpc_send_IRQ_sn2() does a remote amo write
31  *          followed by an IPI. The amo indicates where data is to be pulled
32  *          from, so after the IPI arrives, the remote partition checks the amo
33  *          word. The IPI can actually arrive before the amo however, so other
34  *          code must periodically check for this case. Also, remote amo
35  *          operations do not reliably time out. Thus we do a remote PIO read
36  *          solely to know whether the remote partition is down and whether we
37  *          should stop sending IPIs to it. This remote PIO read operation is
38  *          set up in a special nofault region so SAL knows to ignore (and
39  *          cleanup) any errors due to the remote amo write, PIO read, and/or
40  *          PIO write operations.
41  *
42  *          If/when new hardware solves this IPI problem, we should abandon
43  *          the current approach.
44  *
45  */
46
47 #include <linux/module.h>
48 #include <linux/slab.h>
49 #include <linux/sysctl.h>
50 #include <linux/device.h>
51 #include <linux/delay.h>
52 #include <linux/reboot.h>
53 #include <linux/kdebug.h>
54 #include <linux/kthread.h>
55 #include "xpc.h"
56
57 #ifdef CONFIG_X86_64
58 #include <asm/traps.h>
59 #endif
60
61 /* define two XPC debug device structures to be used with dev_dbg() et al */
62
63 static struct device_driver xpc_dbg_name = {
64         .name = "xpc"
65 };
66
67 static struct device xpc_part_dbg_subname = {
68         .init_name = "",        /* set to "part" at xpc_init() time */
69         .driver = &xpc_dbg_name
70 };
71
72 static struct device xpc_chan_dbg_subname = {
73         .init_name = "",        /* set to "chan" at xpc_init() time */
74         .driver = &xpc_dbg_name
75 };
76
77 struct device *xpc_part = &xpc_part_dbg_subname;
78 struct device *xpc_chan = &xpc_chan_dbg_subname;
79
80 static int xpc_kdebug_ignore;
81
82 /* systune related variables for /proc/sys directories */
83
84 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
85 static int xpc_hb_min_interval = 1;
86 static int xpc_hb_max_interval = 10;
87
88 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
89 static int xpc_hb_check_min_interval = 10;
90 static int xpc_hb_check_max_interval = 120;
91
92 int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
93 static int xpc_disengage_min_timelimit; /* = 0 */
94 static int xpc_disengage_max_timelimit = 120;
95
96 static struct ctl_table xpc_sys_xpc_hb[] = {
97         {
98          .procname = "hb_interval",
99          .data = &xpc_hb_interval,
100          .maxlen = sizeof(int),
101          .mode = 0644,
102          .proc_handler = proc_dointvec_minmax,
103          .extra1 = &xpc_hb_min_interval,
104          .extra2 = &xpc_hb_max_interval},
105         {
106          .procname = "hb_check_interval",
107          .data = &xpc_hb_check_interval,
108          .maxlen = sizeof(int),
109          .mode = 0644,
110          .proc_handler = proc_dointvec_minmax,
111          .extra1 = &xpc_hb_check_min_interval,
112          .extra2 = &xpc_hb_check_max_interval},
113 };
114 static struct ctl_table xpc_sys_xpc[] = {
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 struct ctl_table_header *xpc_sysctl;
126 static struct ctl_table_header *xpc_sysctl_hb;
127
128 /* non-zero if any remote partition disengage was timed out */
129 int xpc_disengage_timedout;
130
131 /* #of activate IRQs received and not yet processed */
132 int xpc_activate_IRQ_rcvd;
133 DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
134
135 /* IRQ handler notifies this wait queue on receipt of an IRQ */
136 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
137
138 static unsigned long xpc_hb_check_timeout;
139 static struct timer_list xpc_hb_timer;
140
141 /* notification that the xpc_hb_checker thread has exited */
142 static DECLARE_COMPLETION(xpc_hb_checker_exited);
143
144 /* notification that the xpc_discovery thread has exited */
145 static DECLARE_COMPLETION(xpc_discovery_exited);
146
147 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
148
149 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
150 static struct notifier_block xpc_reboot_notifier = {
151         .notifier_call = xpc_system_reboot,
152 };
153
154 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
155 static struct notifier_block xpc_die_notifier = {
156         .notifier_call = xpc_system_die,
157 };
158
159 struct xpc_arch_operations xpc_arch_ops;
160
161 /*
162  * Timer function to enforce the timelimit on the partition disengage.
163  */
164 static void
165 xpc_timeout_partition_disengage(struct timer_list *t)
166 {
167         struct xpc_partition *part = from_timer(part, t, disengage_timer);
168
169         DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
170
171         xpc_partition_disengaged_from_timer(part);
172
173         DBUG_ON(part->disengage_timeout != 0);
174         DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
175 }
176
177 /*
178  * Timer to produce the heartbeat.  The timer structures function is
179  * already set when this is initially called.  A tunable is used to
180  * specify when the next timeout should occur.
181  */
182 static void
183 xpc_hb_beater(struct timer_list *unused)
184 {
185         xpc_arch_ops.increment_heartbeat();
186
187         if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
188                 wake_up_interruptible(&xpc_activate_IRQ_wq);
189
190         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
191         add_timer(&xpc_hb_timer);
192 }
193
194 static void
195 xpc_start_hb_beater(void)
196 {
197         xpc_arch_ops.heartbeat_init();
198         timer_setup(&xpc_hb_timer, xpc_hb_beater, 0);
199         xpc_hb_beater(NULL);
200 }
201
202 static void
203 xpc_stop_hb_beater(void)
204 {
205         del_timer_sync(&xpc_hb_timer);
206         xpc_arch_ops.heartbeat_exit();
207 }
208
209 /*
210  * At periodic intervals, scan through all active partitions and ensure
211  * their heartbeat is still active.  If not, the partition is deactivated.
212  */
213 static void
214 xpc_check_remote_hb(void)
215 {
216         struct xpc_partition *part;
217         short partid;
218         enum xp_retval ret;
219
220         for (partid = 0; partid < xp_max_npartitions; partid++) {
221
222                 if (xpc_exiting)
223                         break;
224
225                 if (partid == xp_partition_id)
226                         continue;
227
228                 part = &xpc_partitions[partid];
229
230                 if (part->act_state == XPC_P_AS_INACTIVE ||
231                     part->act_state == XPC_P_AS_DEACTIVATING) {
232                         continue;
233                 }
234
235                 ret = xpc_arch_ops.get_remote_heartbeat(part);
236                 if (ret != xpSuccess)
237                         XPC_DEACTIVATE_PARTITION(part, ret);
238         }
239 }
240
241 /*
242  * This thread is responsible for nearly all of the partition
243  * activation/deactivation.
244  */
245 static int
246 xpc_hb_checker(void *ignore)
247 {
248         int force_IRQ = 0;
249
250         /* this thread was marked active by xpc_hb_init() */
251
252         set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
253
254         /* set our heartbeating to other partitions into motion */
255         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
256         xpc_start_hb_beater();
257
258         while (!xpc_exiting) {
259
260                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
261                         "been received\n",
262                         (int)(xpc_hb_check_timeout - jiffies),
263                         xpc_activate_IRQ_rcvd);
264
265                 /* checking of remote heartbeats is skewed by IRQ handling */
266                 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
267                         xpc_hb_check_timeout = jiffies +
268                             (xpc_hb_check_interval * HZ);
269
270                         dev_dbg(xpc_part, "checking remote heartbeats\n");
271                         xpc_check_remote_hb();
272                 }
273
274                 /* check for outstanding IRQs */
275                 if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
276                         force_IRQ = 0;
277                         dev_dbg(xpc_part, "processing activate IRQs "
278                                 "received\n");
279                         xpc_arch_ops.process_activate_IRQ_rcvd();
280                 }
281
282                 /* wait for IRQ or timeout */
283                 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
284                                                (time_is_before_eq_jiffies(
285                                                 xpc_hb_check_timeout) ||
286                                                 xpc_activate_IRQ_rcvd > 0 ||
287                                                 xpc_exiting));
288         }
289
290         xpc_stop_hb_beater();
291
292         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
293
294         /* mark this thread as having exited */
295         complete(&xpc_hb_checker_exited);
296         return 0;
297 }
298
299 /*
300  * This thread will attempt to discover other partitions to activate
301  * based on info provided by SAL. This new thread is short lived and
302  * will exit once discovery is complete.
303  */
304 static int
305 xpc_initiate_discovery(void *ignore)
306 {
307         xpc_discovery();
308
309         dev_dbg(xpc_part, "discovery thread is exiting\n");
310
311         /* mark this thread as having exited */
312         complete(&xpc_discovery_exited);
313         return 0;
314 }
315
316 /*
317  * The first kthread assigned to a newly activated partition is the one
318  * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
319  * that kthread until the partition is brought down, at which time that kthread
320  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
321  * that XPC has dismantled all communication infrastructure for the associated
322  * partition.) This kthread becomes the channel manager for that partition.
323  *
324  * Each active partition has a channel manager, who, besides connecting and
325  * disconnecting channels, will ensure that each of the partition's connected
326  * channels has the required number of assigned kthreads to get the work done.
327  */
328 static void
329 xpc_channel_mgr(struct xpc_partition *part)
330 {
331         while (part->act_state != XPC_P_AS_DEACTIVATING ||
332                atomic_read(&part->nchannels_active) > 0 ||
333                !xpc_partition_disengaged(part)) {
334
335                 xpc_process_sent_chctl_flags(part);
336
337                 /*
338                  * Wait until we've been requested to activate kthreads or
339                  * all of the channel's message queues have been torn down or
340                  * a signal is pending.
341                  *
342                  * The channel_mgr_requests is set to 1 after being awakened,
343                  * This is done to prevent the channel mgr from making one pass
344                  * through the loop for each request, since he will
345                  * be servicing all the requests in one pass. The reason it's
346                  * set to 1 instead of 0 is so that other kthreads will know
347                  * that the channel mgr is running and won't bother trying to
348                  * wake him up.
349                  */
350                 atomic_dec(&part->channel_mgr_requests);
351                 (void)wait_event_interruptible(part->channel_mgr_wq,
352                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
353                                  part->chctl.all_flags != 0 ||
354                                  (part->act_state == XPC_P_AS_DEACTIVATING &&
355                                  atomic_read(&part->nchannels_active) == 0 &&
356                                  xpc_partition_disengaged(part))));
357                 atomic_set(&part->channel_mgr_requests, 1);
358         }
359 }
360
361 /*
362  * Guarantee that the kzalloc'd memory is cacheline aligned.
363  */
364 void *
365 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
366 {
367         /* see if kzalloc will give us cachline aligned memory by default */
368         *base = kzalloc(size, flags);
369         if (*base == NULL)
370                 return NULL;
371
372         if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
373                 return *base;
374
375         kfree(*base);
376
377         /* nope, we'll have to do it ourselves */
378         *base = kzalloc(size + L1_CACHE_BYTES, flags);
379         if (*base == NULL)
380                 return NULL;
381
382         return (void *)L1_CACHE_ALIGN((u64)*base);
383 }
384
385 /*
386  * Setup the channel structures necessary to support XPartition Communication
387  * between the specified remote partition and the local one.
388  */
389 static enum xp_retval
390 xpc_setup_ch_structures(struct xpc_partition *part)
391 {
392         enum xp_retval ret;
393         int ch_number;
394         struct xpc_channel *ch;
395         short partid = XPC_PARTID(part);
396
397         /*
398          * Allocate all of the channel structures as a contiguous chunk of
399          * memory.
400          */
401         DBUG_ON(part->channels != NULL);
402         part->channels = kcalloc(XPC_MAX_NCHANNELS,
403                                  sizeof(struct xpc_channel),
404                                  GFP_KERNEL);
405         if (part->channels == NULL) {
406                 dev_err(xpc_chan, "can't get memory for channels\n");
407                 return xpNoMemory;
408         }
409
410         /* allocate the remote open and close args */
411
412         part->remote_openclose_args =
413             xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
414                                           GFP_KERNEL, &part->
415                                           remote_openclose_args_base);
416         if (part->remote_openclose_args == NULL) {
417                 dev_err(xpc_chan, "can't get memory for remote connect args\n");
418                 ret = xpNoMemory;
419                 goto out_1;
420         }
421
422         part->chctl.all_flags = 0;
423         spin_lock_init(&part->chctl_lock);
424
425         atomic_set(&part->channel_mgr_requests, 1);
426         init_waitqueue_head(&part->channel_mgr_wq);
427
428         part->nchannels = XPC_MAX_NCHANNELS;
429
430         atomic_set(&part->nchannels_active, 0);
431         atomic_set(&part->nchannels_engaged, 0);
432
433         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
434                 ch = &part->channels[ch_number];
435
436                 ch->partid = partid;
437                 ch->number = ch_number;
438                 ch->flags = XPC_C_DISCONNECTED;
439
440                 atomic_set(&ch->kthreads_assigned, 0);
441                 atomic_set(&ch->kthreads_idle, 0);
442                 atomic_set(&ch->kthreads_active, 0);
443
444                 atomic_set(&ch->references, 0);
445                 atomic_set(&ch->n_to_notify, 0);
446
447                 spin_lock_init(&ch->lock);
448                 init_completion(&ch->wdisconnect_wait);
449
450                 atomic_set(&ch->n_on_msg_allocate_wq, 0);
451                 init_waitqueue_head(&ch->msg_allocate_wq);
452                 init_waitqueue_head(&ch->idle_wq);
453         }
454
455         ret = xpc_arch_ops.setup_ch_structures(part);
456         if (ret != xpSuccess)
457                 goto out_2;
458
459         /*
460          * With the setting of the partition setup_state to XPC_P_SS_SETUP,
461          * we're declaring that this partition is ready to go.
462          */
463         part->setup_state = XPC_P_SS_SETUP;
464
465         return xpSuccess;
466
467         /* setup of ch structures failed */
468 out_2:
469         kfree(part->remote_openclose_args_base);
470         part->remote_openclose_args = NULL;
471 out_1:
472         kfree(part->channels);
473         part->channels = NULL;
474         return ret;
475 }
476
477 /*
478  * Teardown the channel structures necessary to support XPartition Communication
479  * between the specified remote partition and the local one.
480  */
481 static void
482 xpc_teardown_ch_structures(struct xpc_partition *part)
483 {
484         DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
485         DBUG_ON(atomic_read(&part->nchannels_active) != 0);
486
487         /*
488          * Make this partition inaccessible to local processes by marking it
489          * as no longer setup. Then wait before proceeding with the teardown
490          * until all existing references cease.
491          */
492         DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
493         part->setup_state = XPC_P_SS_WTEARDOWN;
494
495         wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
496
497         /* now we can begin tearing down the infrastructure */
498
499         xpc_arch_ops.teardown_ch_structures(part);
500
501         kfree(part->remote_openclose_args_base);
502         part->remote_openclose_args = NULL;
503         kfree(part->channels);
504         part->channels = NULL;
505
506         part->setup_state = XPC_P_SS_TORNDOWN;
507 }
508
509 /*
510  * When XPC HB determines that a partition has come up, it will create a new
511  * kthread and that kthread will call this function to attempt to set up the
512  * basic infrastructure used for Cross Partition Communication with the newly
513  * upped partition.
514  *
515  * The kthread that was created by XPC HB and which setup the XPC
516  * infrastructure will remain assigned to the partition becoming the channel
517  * manager for that partition until the partition is deactivating, at which
518  * time the kthread will teardown the XPC infrastructure and then exit.
519  */
520 static int
521 xpc_activating(void *__partid)
522 {
523         short partid = (u64)__partid;
524         struct xpc_partition *part = &xpc_partitions[partid];
525         unsigned long irq_flags;
526
527         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
528
529         spin_lock_irqsave(&part->act_lock, irq_flags);
530
531         if (part->act_state == XPC_P_AS_DEACTIVATING) {
532                 part->act_state = XPC_P_AS_INACTIVE;
533                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
534                 part->remote_rp_pa = 0;
535                 return 0;
536         }
537
538         /* indicate the thread is activating */
539         DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
540         part->act_state = XPC_P_AS_ACTIVATING;
541
542         XPC_SET_REASON(part, 0, 0);
543         spin_unlock_irqrestore(&part->act_lock, irq_flags);
544
545         dev_dbg(xpc_part, "activating partition %d\n", partid);
546
547         xpc_arch_ops.allow_hb(partid);
548
549         if (xpc_setup_ch_structures(part) == xpSuccess) {
550                 (void)xpc_part_ref(part);       /* this will always succeed */
551
552                 if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
553                         xpc_mark_partition_active(part);
554                         xpc_channel_mgr(part);
555                         /* won't return until partition is deactivating */
556                 }
557
558                 xpc_part_deref(part);
559                 xpc_teardown_ch_structures(part);
560         }
561
562         xpc_arch_ops.disallow_hb(partid);
563         xpc_mark_partition_inactive(part);
564
565         if (part->reason == xpReactivating) {
566                 /* interrupting ourselves results in activating partition */
567                 xpc_arch_ops.request_partition_reactivation(part);
568         }
569
570         return 0;
571 }
572
573 void
574 xpc_activate_partition(struct xpc_partition *part)
575 {
576         short partid = XPC_PARTID(part);
577         unsigned long irq_flags;
578         struct task_struct *kthread;
579
580         spin_lock_irqsave(&part->act_lock, irq_flags);
581
582         DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
583
584         part->act_state = XPC_P_AS_ACTIVATION_REQ;
585         XPC_SET_REASON(part, xpCloneKThread, __LINE__);
586
587         spin_unlock_irqrestore(&part->act_lock, irq_flags);
588
589         kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
590                               partid);
591         if (IS_ERR(kthread)) {
592                 spin_lock_irqsave(&part->act_lock, irq_flags);
593                 part->act_state = XPC_P_AS_INACTIVE;
594                 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
595                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
596         }
597 }
598
599 void
600 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
601 {
602         int idle = atomic_read(&ch->kthreads_idle);
603         int assigned = atomic_read(&ch->kthreads_assigned);
604         int wakeup;
605
606         DBUG_ON(needed <= 0);
607
608         if (idle > 0) {
609                 wakeup = (needed > idle) ? idle : needed;
610                 needed -= wakeup;
611
612                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
613                         "channel=%d\n", wakeup, ch->partid, ch->number);
614
615                 /* only wakeup the requested number of kthreads */
616                 wake_up_nr(&ch->idle_wq, wakeup);
617         }
618
619         if (needed <= 0)
620                 return;
621
622         if (needed + assigned > ch->kthreads_assigned_limit) {
623                 needed = ch->kthreads_assigned_limit - assigned;
624                 if (needed <= 0)
625                         return;
626         }
627
628         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
629                 needed, ch->partid, ch->number);
630
631         xpc_create_kthreads(ch, needed, 0);
632 }
633
634 /*
635  * This function is where XPC's kthreads wait for messages to deliver.
636  */
637 static void
638 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
639 {
640         int (*n_of_deliverable_payloads) (struct xpc_channel *) =
641                 xpc_arch_ops.n_of_deliverable_payloads;
642
643         do {
644                 /* deliver messages to their intended recipients */
645
646                 while (n_of_deliverable_payloads(ch) > 0 &&
647                        !(ch->flags & XPC_C_DISCONNECTING)) {
648                         xpc_deliver_payload(ch);
649                 }
650
651                 if (atomic_inc_return(&ch->kthreads_idle) >
652                     ch->kthreads_idle_limit) {
653                         /* too many idle kthreads on this channel */
654                         atomic_dec(&ch->kthreads_idle);
655                         break;
656                 }
657
658                 dev_dbg(xpc_chan, "idle kthread calling "
659                         "wait_event_interruptible_exclusive()\n");
660
661                 (void)wait_event_interruptible_exclusive(ch->idle_wq,
662                                 (n_of_deliverable_payloads(ch) > 0 ||
663                                  (ch->flags & XPC_C_DISCONNECTING)));
664
665                 atomic_dec(&ch->kthreads_idle);
666
667         } while (!(ch->flags & XPC_C_DISCONNECTING));
668 }
669
670 static int
671 xpc_kthread_start(void *args)
672 {
673         short partid = XPC_UNPACK_ARG1(args);
674         u16 ch_number = XPC_UNPACK_ARG2(args);
675         struct xpc_partition *part = &xpc_partitions[partid];
676         struct xpc_channel *ch;
677         int n_needed;
678         unsigned long irq_flags;
679         int (*n_of_deliverable_payloads) (struct xpc_channel *) =
680                 xpc_arch_ops.n_of_deliverable_payloads;
681
682         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
683                 partid, ch_number);
684
685         ch = &part->channels[ch_number];
686
687         if (!(ch->flags & XPC_C_DISCONNECTING)) {
688
689                 /* let registerer know that connection has been established */
690
691                 spin_lock_irqsave(&ch->lock, irq_flags);
692                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
693                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
694                         spin_unlock_irqrestore(&ch->lock, irq_flags);
695
696                         xpc_connected_callout(ch);
697
698                         spin_lock_irqsave(&ch->lock, irq_flags);
699                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
700                         spin_unlock_irqrestore(&ch->lock, irq_flags);
701
702                         /*
703                          * It is possible that while the callout was being
704                          * made that the remote partition sent some messages.
705                          * If that is the case, we may need to activate
706                          * additional kthreads to help deliver them. We only
707                          * need one less than total #of messages to deliver.
708                          */
709                         n_needed = n_of_deliverable_payloads(ch) - 1;
710                         if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
711                                 xpc_activate_kthreads(ch, n_needed);
712
713                 } else {
714                         spin_unlock_irqrestore(&ch->lock, irq_flags);
715                 }
716
717                 xpc_kthread_waitmsgs(part, ch);
718         }
719
720         /* let registerer know that connection is disconnecting */
721
722         spin_lock_irqsave(&ch->lock, irq_flags);
723         if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
724             !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
725                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
726                 spin_unlock_irqrestore(&ch->lock, irq_flags);
727
728                 xpc_disconnect_callout(ch, xpDisconnecting);
729
730                 spin_lock_irqsave(&ch->lock, irq_flags);
731                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
732         }
733         spin_unlock_irqrestore(&ch->lock, irq_flags);
734
735         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
736             atomic_dec_return(&part->nchannels_engaged) == 0) {
737                 xpc_arch_ops.indicate_partition_disengaged(part);
738         }
739
740         xpc_msgqueue_deref(ch);
741
742         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
743                 partid, ch_number);
744
745         xpc_part_deref(part);
746         return 0;
747 }
748
749 /*
750  * For each partition that XPC has established communications with, there is
751  * a minimum of one kernel thread assigned to perform any operation that
752  * may potentially sleep or block (basically the callouts to the asynchronous
753  * functions registered via xpc_connect()).
754  *
755  * Additional kthreads are created and destroyed by XPC as the workload
756  * demands.
757  *
758  * A kthread is assigned to one of the active channels that exists for a given
759  * partition.
760  */
761 void
762 xpc_create_kthreads(struct xpc_channel *ch, int needed,
763                     int ignore_disconnecting)
764 {
765         unsigned long irq_flags;
766         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
767         struct xpc_partition *part = &xpc_partitions[ch->partid];
768         struct task_struct *kthread;
769         void (*indicate_partition_disengaged) (struct xpc_partition *) =
770                 xpc_arch_ops.indicate_partition_disengaged;
771
772         while (needed-- > 0) {
773
774                 /*
775                  * The following is done on behalf of the newly created
776                  * kthread. That kthread is responsible for doing the
777                  * counterpart to the following before it exits.
778                  */
779                 if (ignore_disconnecting) {
780                         if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
781                                 /* kthreads assigned had gone to zero */
782                                 BUG_ON(!(ch->flags &
783                                          XPC_C_DISCONNECTINGCALLOUT_MADE));
784                                 break;
785                         }
786
787                 } else if (ch->flags & XPC_C_DISCONNECTING) {
788                         break;
789
790                 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
791                            atomic_inc_return(&part->nchannels_engaged) == 1) {
792                         xpc_arch_ops.indicate_partition_engaged(part);
793                 }
794                 (void)xpc_part_ref(part);
795                 xpc_msgqueue_ref(ch);
796
797                 kthread = kthread_run(xpc_kthread_start, (void *)args,
798                                       "xpc%02dc%d", ch->partid, ch->number);
799                 if (IS_ERR(kthread)) {
800                         /* the fork failed */
801
802                         /*
803                          * NOTE: if (ignore_disconnecting &&
804                          * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
805                          * then we'll deadlock if all other kthreads assigned
806                          * to this channel are blocked in the channel's
807                          * registerer, because the only thing that will unblock
808                          * them is the xpDisconnecting callout that this
809                          * failed kthread_run() would have made.
810                          */
811
812                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
813                             atomic_dec_return(&part->nchannels_engaged) == 0) {
814                                 indicate_partition_disengaged(part);
815                         }
816                         xpc_msgqueue_deref(ch);
817                         xpc_part_deref(part);
818
819                         if (atomic_read(&ch->kthreads_assigned) <
820                             ch->kthreads_idle_limit) {
821                                 /*
822                                  * Flag this as an error only if we have an
823                                  * insufficient #of kthreads for the channel
824                                  * to function.
825                                  */
826                                 spin_lock_irqsave(&ch->lock, irq_flags);
827                                 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
828                                                        &irq_flags);
829                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
830                         }
831                         break;
832                 }
833         }
834 }
835
836 void
837 xpc_disconnect_wait(int ch_number)
838 {
839         unsigned long irq_flags;
840         short partid;
841         struct xpc_partition *part;
842         struct xpc_channel *ch;
843         int wakeup_channel_mgr;
844
845         /* now wait for all callouts to the caller's function to cease */
846         for (partid = 0; partid < xp_max_npartitions; partid++) {
847                 part = &xpc_partitions[partid];
848
849                 if (!xpc_part_ref(part))
850                         continue;
851
852                 ch = &part->channels[ch_number];
853
854                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
855                         xpc_part_deref(part);
856                         continue;
857                 }
858
859                 wait_for_completion(&ch->wdisconnect_wait);
860
861                 spin_lock_irqsave(&ch->lock, irq_flags);
862                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
863                 wakeup_channel_mgr = 0;
864
865                 if (ch->delayed_chctl_flags) {
866                         if (part->act_state != XPC_P_AS_DEACTIVATING) {
867                                 spin_lock(&part->chctl_lock);
868                                 part->chctl.flags[ch->number] |=
869                                     ch->delayed_chctl_flags;
870                                 spin_unlock(&part->chctl_lock);
871                                 wakeup_channel_mgr = 1;
872                         }
873                         ch->delayed_chctl_flags = 0;
874                 }
875
876                 ch->flags &= ~XPC_C_WDISCONNECT;
877                 spin_unlock_irqrestore(&ch->lock, irq_flags);
878
879                 if (wakeup_channel_mgr)
880                         xpc_wakeup_channel_mgr(part);
881
882                 xpc_part_deref(part);
883         }
884 }
885
886 static int
887 xpc_setup_partitions(void)
888 {
889         short partid;
890         struct xpc_partition *part;
891
892         xpc_partitions = kcalloc(xp_max_npartitions,
893                                  sizeof(struct xpc_partition),
894                                  GFP_KERNEL);
895         if (xpc_partitions == NULL) {
896                 dev_err(xpc_part, "can't get memory for partition structure\n");
897                 return -ENOMEM;
898         }
899
900         /*
901          * The first few fields of each entry of xpc_partitions[] need to
902          * be initialized now so that calls to xpc_connect() and
903          * xpc_disconnect() can be made prior to the activation of any remote
904          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
905          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
906          * PARTITION HAS BEEN ACTIVATED.
907          */
908         for (partid = 0; partid < xp_max_npartitions; partid++) {
909                 part = &xpc_partitions[partid];
910
911                 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
912
913                 part->activate_IRQ_rcvd = 0;
914                 spin_lock_init(&part->act_lock);
915                 part->act_state = XPC_P_AS_INACTIVE;
916                 XPC_SET_REASON(part, 0, 0);
917
918                 timer_setup(&part->disengage_timer,
919                             xpc_timeout_partition_disengage, 0);
920
921                 part->setup_state = XPC_P_SS_UNSET;
922                 init_waitqueue_head(&part->teardown_wq);
923                 atomic_set(&part->references, 0);
924         }
925
926         return xpc_arch_ops.setup_partitions();
927 }
928
929 static void
930 xpc_teardown_partitions(void)
931 {
932         xpc_arch_ops.teardown_partitions();
933         kfree(xpc_partitions);
934 }
935
936 static void
937 xpc_do_exit(enum xp_retval reason)
938 {
939         short partid;
940         int active_part_count, printed_waiting_msg = 0;
941         struct xpc_partition *part;
942         unsigned long printmsg_time, disengage_timeout = 0;
943
944         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
945         DBUG_ON(xpc_exiting == 1);
946
947         /*
948          * Let the heartbeat checker thread and the discovery thread
949          * (if one is running) know that they should exit. Also wake up
950          * the heartbeat checker thread in case it's sleeping.
951          */
952         xpc_exiting = 1;
953         wake_up_interruptible(&xpc_activate_IRQ_wq);
954
955         /* wait for the discovery thread to exit */
956         wait_for_completion(&xpc_discovery_exited);
957
958         /* wait for the heartbeat checker thread to exit */
959         wait_for_completion(&xpc_hb_checker_exited);
960
961         /* sleep for a 1/3 of a second or so */
962         (void)msleep_interruptible(300);
963
964         /* wait for all partitions to become inactive */
965
966         printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
967         xpc_disengage_timedout = 0;
968
969         do {
970                 active_part_count = 0;
971
972                 for (partid = 0; partid < xp_max_npartitions; partid++) {
973                         part = &xpc_partitions[partid];
974
975                         if (xpc_partition_disengaged(part) &&
976                             part->act_state == XPC_P_AS_INACTIVE) {
977                                 continue;
978                         }
979
980                         active_part_count++;
981
982                         XPC_DEACTIVATE_PARTITION(part, reason);
983
984                         if (part->disengage_timeout > disengage_timeout)
985                                 disengage_timeout = part->disengage_timeout;
986                 }
987
988                 if (xpc_arch_ops.any_partition_engaged()) {
989                         if (time_is_before_jiffies(printmsg_time)) {
990                                 dev_info(xpc_part, "waiting for remote "
991                                          "partitions to deactivate, timeout in "
992                                          "%ld seconds\n", (disengage_timeout -
993                                          jiffies) / HZ);
994                                 printmsg_time = jiffies +
995                                     (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
996                                 printed_waiting_msg = 1;
997                         }
998
999                 } else if (active_part_count > 0) {
1000                         if (printed_waiting_msg) {
1001                                 dev_info(xpc_part, "waiting for local partition"
1002                                          " to deactivate\n");
1003                                 printed_waiting_msg = 0;
1004                         }
1005
1006                 } else {
1007                         if (!xpc_disengage_timedout) {
1008                                 dev_info(xpc_part, "all partitions have "
1009                                          "deactivated\n");
1010                         }
1011                         break;
1012                 }
1013
1014                 /* sleep for a 1/3 of a second or so */
1015                 (void)msleep_interruptible(300);
1016
1017         } while (1);
1018
1019         DBUG_ON(xpc_arch_ops.any_partition_engaged());
1020
1021         xpc_teardown_rsvd_page();
1022
1023         if (reason == xpUnloading) {
1024                 (void)unregister_die_notifier(&xpc_die_notifier);
1025                 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1026         }
1027
1028         /* clear the interface to XPC's functions */
1029         xpc_clear_interface();
1030
1031         if (xpc_sysctl)
1032                 unregister_sysctl_table(xpc_sysctl);
1033         if (xpc_sysctl_hb)
1034                 unregister_sysctl_table(xpc_sysctl_hb);
1035
1036         xpc_teardown_partitions();
1037
1038         if (is_uv_system())
1039                 xpc_exit_uv();
1040 }
1041
1042 /*
1043  * This function is called when the system is being rebooted.
1044  */
1045 static int
1046 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1047 {
1048         enum xp_retval reason;
1049
1050         switch (event) {
1051         case SYS_RESTART:
1052                 reason = xpSystemReboot;
1053                 break;
1054         case SYS_HALT:
1055                 reason = xpSystemHalt;
1056                 break;
1057         case SYS_POWER_OFF:
1058                 reason = xpSystemPoweroff;
1059                 break;
1060         default:
1061                 reason = xpSystemGoingDown;
1062         }
1063
1064         xpc_do_exit(reason);
1065         return NOTIFY_DONE;
1066 }
1067
1068 /* Used to only allow one cpu to complete disconnect */
1069 static unsigned int xpc_die_disconnecting;
1070
1071 /*
1072  * Notify other partitions to deactivate from us by first disengaging from all
1073  * references to our memory.
1074  */
1075 static void
1076 xpc_die_deactivate(void)
1077 {
1078         struct xpc_partition *part;
1079         short partid;
1080         int any_engaged;
1081         long keep_waiting;
1082         long wait_to_print;
1083
1084         if (cmpxchg(&xpc_die_disconnecting, 0, 1))
1085                 return;
1086
1087         /* keep xpc_hb_checker thread from doing anything (just in case) */
1088         xpc_exiting = 1;
1089
1090         xpc_arch_ops.disallow_all_hbs();   /*indicate we're deactivated */
1091
1092         for (partid = 0; partid < xp_max_npartitions; partid++) {
1093                 part = &xpc_partitions[partid];
1094
1095                 if (xpc_arch_ops.partition_engaged(partid) ||
1096                     part->act_state != XPC_P_AS_INACTIVE) {
1097                         xpc_arch_ops.request_partition_deactivation(part);
1098                         xpc_arch_ops.indicate_partition_disengaged(part);
1099                 }
1100         }
1101
1102         /*
1103          * Though we requested that all other partitions deactivate from us,
1104          * we only wait until they've all disengaged or we've reached the
1105          * defined timelimit.
1106          *
1107          * Given that one iteration through the following while-loop takes
1108          * approximately 200 microseconds, calculate the #of loops to take
1109          * before bailing and the #of loops before printing a waiting message.
1110          */
1111         keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1112         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1113
1114         while (1) {
1115                 any_engaged = xpc_arch_ops.any_partition_engaged();
1116                 if (!any_engaged) {
1117                         dev_info(xpc_part, "all partitions have deactivated\n");
1118                         break;
1119                 }
1120
1121                 if (!keep_waiting--) {
1122                         for (partid = 0; partid < xp_max_npartitions;
1123                              partid++) {
1124                                 if (xpc_arch_ops.partition_engaged(partid)) {
1125                                         dev_info(xpc_part, "deactivate from "
1126                                                  "remote partition %d timed "
1127                                                  "out\n", partid);
1128                                 }
1129                         }
1130                         break;
1131                 }
1132
1133                 if (!wait_to_print--) {
1134                         dev_info(xpc_part, "waiting for remote partitions to "
1135                                  "deactivate, timeout in %ld seconds\n",
1136                                  keep_waiting / (1000 * 5));
1137                         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1138                             1000 * 5;
1139                 }
1140
1141                 udelay(200);
1142         }
1143 }
1144
1145 /*
1146  * This function is called when the system is being restarted or halted due
1147  * to some sort of system failure. If this is the case we need to notify the
1148  * other partitions to disengage from all references to our memory.
1149  * This function can also be called when our heartbeater could be offlined
1150  * for a time. In this case we need to notify other partitions to not worry
1151  * about the lack of a heartbeat.
1152  */
1153 static int
1154 xpc_system_die(struct notifier_block *nb, unsigned long event, void *_die_args)
1155 {
1156         struct die_args *die_args = _die_args;
1157
1158         switch (event) {
1159         case DIE_TRAP:
1160                 if (die_args->trapnr == X86_TRAP_DF)
1161                         xpc_die_deactivate();
1162
1163                 if (((die_args->trapnr == X86_TRAP_MF) ||
1164                      (die_args->trapnr == X86_TRAP_XF)) &&
1165                     !user_mode(die_args->regs))
1166                         xpc_die_deactivate();
1167
1168                 break;
1169         case DIE_INT3:
1170         case DIE_DEBUG:
1171                 break;
1172         case DIE_OOPS:
1173         case DIE_GPF:
1174         default:
1175                 xpc_die_deactivate();
1176         }
1177
1178         return NOTIFY_DONE;
1179 }
1180
1181 static int __init
1182 xpc_init(void)
1183 {
1184         int ret;
1185         struct task_struct *kthread;
1186
1187         dev_set_name(xpc_part, "part");
1188         dev_set_name(xpc_chan, "chan");
1189
1190         if (is_uv_system()) {
1191                 ret = xpc_init_uv();
1192
1193         } else {
1194                 ret = -ENODEV;
1195         }
1196
1197         if (ret != 0)
1198                 return ret;
1199
1200         ret = xpc_setup_partitions();
1201         if (ret != 0) {
1202                 dev_err(xpc_part, "can't get memory for partition structure\n");
1203                 goto out_1;
1204         }
1205
1206         xpc_sysctl = register_sysctl("xpc", xpc_sys_xpc);
1207         xpc_sysctl_hb = register_sysctl("xpc/hb", xpc_sys_xpc_hb);
1208
1209         /*
1210          * Fill the partition reserved page with the information needed by
1211          * other partitions to discover we are alive and establish initial
1212          * communications.
1213          */
1214         ret = xpc_setup_rsvd_page();
1215         if (ret != 0) {
1216                 dev_err(xpc_part, "can't setup our reserved page\n");
1217                 goto out_2;
1218         }
1219
1220         /* add ourselves to the reboot_notifier_list */
1221         ret = register_reboot_notifier(&xpc_reboot_notifier);
1222         if (ret != 0)
1223                 dev_warn(xpc_part, "can't register reboot notifier\n");
1224
1225         /* add ourselves to the die_notifier list */
1226         ret = register_die_notifier(&xpc_die_notifier);
1227         if (ret != 0)
1228                 dev_warn(xpc_part, "can't register die notifier\n");
1229
1230         /*
1231          * The real work-horse behind xpc.  This processes incoming
1232          * interrupts and monitors remote heartbeats.
1233          */
1234         kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1235         if (IS_ERR(kthread)) {
1236                 dev_err(xpc_part, "failed while forking hb check thread\n");
1237                 ret = -EBUSY;
1238                 goto out_3;
1239         }
1240
1241         /*
1242          * Startup a thread that will attempt to discover other partitions to
1243          * activate based on info provided by SAL. This new thread is short
1244          * lived and will exit once discovery is complete.
1245          */
1246         kthread = kthread_run(xpc_initiate_discovery, NULL,
1247                               XPC_DISCOVERY_THREAD_NAME);
1248         if (IS_ERR(kthread)) {
1249                 dev_err(xpc_part, "failed while forking discovery thread\n");
1250
1251                 /* mark this new thread as a non-starter */
1252                 complete(&xpc_discovery_exited);
1253
1254                 xpc_do_exit(xpUnloading);
1255                 return -EBUSY;
1256         }
1257
1258         /* set the interface to point at XPC's functions */
1259         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1260                           xpc_initiate_send, xpc_initiate_send_notify,
1261                           xpc_initiate_received, xpc_initiate_partid_to_nasids);
1262
1263         return 0;
1264
1265         /* initialization was not successful */
1266 out_3:
1267         xpc_teardown_rsvd_page();
1268
1269         (void)unregister_die_notifier(&xpc_die_notifier);
1270         (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1271 out_2:
1272         if (xpc_sysctl_hb)
1273                 unregister_sysctl_table(xpc_sysctl_hb);
1274         if (xpc_sysctl)
1275                 unregister_sysctl_table(xpc_sysctl);
1276
1277         xpc_teardown_partitions();
1278 out_1:
1279         if (is_uv_system())
1280                 xpc_exit_uv();
1281         return ret;
1282 }
1283
1284 module_init(xpc_init);
1285
1286 static void __exit
1287 xpc_exit(void)
1288 {
1289         xpc_do_exit(xpUnloading);
1290 }
1291
1292 module_exit(xpc_exit);
1293
1294 MODULE_AUTHOR("Silicon Graphics, Inc.");
1295 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1296 MODULE_LICENSE("GPL");
1297
1298 module_param(xpc_hb_interval, int, 0);
1299 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1300                  "heartbeat increments.");
1301
1302 module_param(xpc_hb_check_interval, int, 0);
1303 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1304                  "heartbeat checks.");
1305
1306 module_param(xpc_disengage_timelimit, int, 0);
1307 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1308                  "for disengage to complete.");
1309
1310 module_param(xpc_kdebug_ignore, int, 0);
1311 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1312                  "other partitions when dropping into kdebug.");