563aaf4a2ff688ffa55326df794e71f906ae3203
[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) 2004-2008 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  *        . We currently have no way to determine which nasid an IPI came
29  *          from. Thus, >>> xpc_IPI_send() does a remote AMO write followed by
30  *          an IPI. The AMO indicates where data is to be pulled from, so
31  *          after the IPI arrives, the remote partition checks the AMO word.
32  *          The IPI can actually arrive before the AMO however, so other code
33  *          must periodically check for this case. Also, remote AMO operations
34  *          do not reliably time out. Thus we do a remote PIO read solely to
35  *          know whether the remote partition is down and whether we should
36  *          stop sending IPIs to it. This remote PIO read operation is set up
37  *          in a special nofault region so SAL knows to ignore (and cleanup)
38  *          any errors due to the remote AMO write, PIO read, and/or PIO
39  *          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/kernel.h>
47 #include <linux/module.h>
48 #include <linux/init.h>
49 #include <linux/cache.h>
50 #include <linux/interrupt.h>
51 #include <linux/delay.h>
52 #include <linux/reboot.h>
53 #include <linux/completion.h>
54 #include <linux/kdebug.h>
55 #include <linux/kthread.h>
56 #include <linux/uaccess.h>
57 #include <asm/sn/intr.h>
58 #include <asm/sn/sn_sal.h>
59 #include "xpc.h"
60
61 /* define two XPC debug device structures to be used with dev_dbg() et al */
62
63 struct device_driver xpc_dbg_name = {
64         .name = "xpc"
65 };
66
67 struct device xpc_part_dbg_subname = {
68         .bus_id = {0},          /* set to "part" at xpc_init() time */
69         .driver = &xpc_dbg_name
70 };
71
72 struct device xpc_chan_dbg_subname = {
73         .bus_id = {0},          /* 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 ctl_table xpc_sys_xpc_hb_dir[] = {
97         {
98          .ctl_name = CTL_UNNUMBERED,
99          .procname = "hb_interval",
100          .data = &xpc_hb_interval,
101          .maxlen = sizeof(int),
102          .mode = 0644,
103          .proc_handler = &proc_dointvec_minmax,
104          .strategy = &sysctl_intvec,
105          .extra1 = &xpc_hb_min_interval,
106          .extra2 = &xpc_hb_max_interval},
107         {
108          .ctl_name = CTL_UNNUMBERED,
109          .procname = "hb_check_interval",
110          .data = &xpc_hb_check_interval,
111          .maxlen = sizeof(int),
112          .mode = 0644,
113          .proc_handler = &proc_dointvec_minmax,
114          .strategy = &sysctl_intvec,
115          .extra1 = &xpc_hb_check_min_interval,
116          .extra2 = &xpc_hb_check_max_interval},
117         {}
118 };
119 static ctl_table xpc_sys_xpc_dir[] = {
120         {
121          .ctl_name = CTL_UNNUMBERED,
122          .procname = "hb",
123          .mode = 0555,
124          .child = xpc_sys_xpc_hb_dir},
125         {
126          .ctl_name = CTL_UNNUMBERED,
127          .procname = "disengage_timelimit",
128          .data = &xpc_disengage_timelimit,
129          .maxlen = sizeof(int),
130          .mode = 0644,
131          .proc_handler = &proc_dointvec_minmax,
132          .strategy = &sysctl_intvec,
133          .extra1 = &xpc_disengage_min_timelimit,
134          .extra2 = &xpc_disengage_max_timelimit},
135         {}
136 };
137 static ctl_table xpc_sys_dir[] = {
138         {
139          .ctl_name = CTL_UNNUMBERED,
140          .procname = "xpc",
141          .mode = 0555,
142          .child = xpc_sys_xpc_dir},
143         {}
144 };
145 static struct ctl_table_header *xpc_sysctl;
146
147 /* non-zero if any remote partition disengage was timed out */
148 int xpc_disengage_timedout;
149
150 /* #of activate IRQs received */
151 atomic_t xpc_activate_IRQ_rcvd = ATOMIC_INIT(0);
152
153 /* IRQ handler notifies this wait queue on receipt of an IRQ */
154 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
155
156 static unsigned long xpc_hb_check_timeout;
157 static struct timer_list xpc_hb_timer;
158 void *xpc_heartbeating_to_mask;
159
160 /* notification that the xpc_hb_checker thread has exited */
161 static DECLARE_COMPLETION(xpc_hb_checker_exited);
162
163 /* notification that the xpc_discovery thread has exited */
164 static DECLARE_COMPLETION(xpc_discovery_exited);
165
166 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
167
168 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
169 static struct notifier_block xpc_reboot_notifier = {
170         .notifier_call = xpc_system_reboot,
171 };
172
173 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
174 static struct notifier_block xpc_die_notifier = {
175         .notifier_call = xpc_system_die,
176 };
177
178 enum xp_retval (*xpc_rsvd_page_init) (struct xpc_rsvd_page *rp);
179 void (*xpc_heartbeat_init) (void);
180 void (*xpc_heartbeat_exit) (void);
181 void (*xpc_increment_heartbeat) (void);
182 void (*xpc_offline_heartbeat) (void);
183 void (*xpc_online_heartbeat) (void);
184 void (*xpc_check_remote_hb) (void);
185
186 enum xp_retval (*xpc_make_first_contact) (struct xpc_partition *part);
187 void (*xpc_notify_senders_of_disconnect) (struct xpc_channel *ch);
188 u64 (*xpc_get_IPI_flags) (struct xpc_partition *part);
189 void (*xpc_process_msg_IPI) (struct xpc_partition *part, int ch_number);
190 int (*xpc_n_of_deliverable_msgs) (struct xpc_channel *ch);
191 struct xpc_msg *(*xpc_get_deliverable_msg) (struct xpc_channel *ch);
192
193 void (*xpc_request_partition_activation) (struct xpc_rsvd_page *remote_rp,
194                                           u64 remote_rp_pa, int nasid);
195 void (*xpc_request_partition_reactivation) (struct xpc_partition *part);
196 void (*xpc_request_partition_deactivation) (struct xpc_partition *part);
197 void (*xpc_cancel_partition_deactivation_request) (struct xpc_partition *part);
198
199 void (*xpc_process_activate_IRQ_rcvd) (int n_IRQs_expected);
200 enum xp_retval (*xpc_setup_infrastructure) (struct xpc_partition *part);
201 void (*xpc_teardown_infrastructure) (struct xpc_partition *part);
202
203 void (*xpc_indicate_partition_engaged) (struct xpc_partition *part);
204 int (*xpc_partition_engaged) (short partid);
205 int (*xpc_any_partition_engaged) (void);
206 void (*xpc_indicate_partition_disengaged) (struct xpc_partition *part);
207 void (*xpc_assume_partition_disengaged) (short partid);
208
209 void (*xpc_send_channel_closerequest) (struct xpc_channel *ch,
210                                        unsigned long *irq_flags);
211 void (*xpc_send_channel_closereply) (struct xpc_channel *ch,
212                                      unsigned long *irq_flags);
213 void (*xpc_send_channel_openrequest) (struct xpc_channel *ch,
214                                       unsigned long *irq_flags);
215 void (*xpc_send_channel_openreply) (struct xpc_channel *ch,
216                                     unsigned long *irq_flags);
217
218 enum xp_retval (*xpc_send_msg) (struct xpc_channel *ch, u32 flags,
219                                 void *payload, u16 payload_size, u8 notify_type,
220                                 xpc_notify_func func, void *key);
221 void (*xpc_received_msg) (struct xpc_channel *ch, struct xpc_msg *msg);
222
223 /*
224  * Timer function to enforce the timelimit on the partition disengage.
225  */
226 static void
227 xpc_timeout_partition_disengage(unsigned long data)
228 {
229         struct xpc_partition *part = (struct xpc_partition *)data;
230
231         DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
232
233         (void)xpc_partition_disengaged(part);
234
235         DBUG_ON(part->disengage_timeout != 0);
236         DBUG_ON(xpc_partition_engaged(XPC_PARTID(part)));
237 }
238
239 /*
240  * Timer to produce the heartbeat.  The timer structures function is
241  * already set when this is initially called.  A tunable is used to
242  * specify when the next timeout should occur.
243  */
244 static void
245 xpc_hb_beater(unsigned long dummy)
246 {
247         xpc_increment_heartbeat();
248
249         if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
250                 wake_up_interruptible(&xpc_activate_IRQ_wq);
251
252         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
253         add_timer(&xpc_hb_timer);
254 }
255
256 static void
257 xpc_start_hb_beater(void)
258 {
259         xpc_heartbeat_init();
260         init_timer(&xpc_hb_timer);
261         xpc_hb_timer.function = xpc_hb_beater;
262         xpc_hb_beater(0);
263 }
264
265 static void
266 xpc_stop_hb_beater(void)
267 {
268         del_timer_sync(&xpc_hb_timer);
269         xpc_heartbeat_exit();
270 }
271
272 /*
273  * This thread is responsible for nearly all of the partition
274  * activation/deactivation.
275  */
276 static int
277 xpc_hb_checker(void *ignore)
278 {
279         int last_IRQ_count = 0;
280         int new_IRQ_count;
281         int force_IRQ = 0;
282
283         /* this thread was marked active by xpc_hb_init() */
284
285         set_cpus_allowed_ptr(current, &cpumask_of_cpu(XPC_HB_CHECK_CPU));
286
287         /* set our heartbeating to other partitions into motion */
288         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
289         xpc_start_hb_beater();
290
291         while (!xpc_exiting) {
292
293                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
294                         "been received\n",
295                         (int)(xpc_hb_check_timeout - jiffies),
296                         atomic_read(&xpc_activate_IRQ_rcvd) - last_IRQ_count);
297
298                 /* checking of remote heartbeats is skewed by IRQ handling */
299                 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
300                         dev_dbg(xpc_part, "checking remote heartbeats\n");
301                         xpc_check_remote_hb();
302
303                         /*
304                          * We need to periodically recheck to ensure no
305                          * IPI/AMO pairs have been missed.  That check
306                          * must always reset xpc_hb_check_timeout.
307                          */
308                         force_IRQ = 1;
309                 }
310
311                 /* check for outstanding IRQs */
312                 new_IRQ_count = atomic_read(&xpc_activate_IRQ_rcvd);
313                 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
314                         force_IRQ = 0;
315
316                         dev_dbg(xpc_part, "found an IRQ to process; will be "
317                                 "resetting xpc_hb_check_timeout\n");
318
319                         xpc_process_activate_IRQ_rcvd(new_IRQ_count -
320                                                       last_IRQ_count);
321                         last_IRQ_count = new_IRQ_count;
322
323                         xpc_hb_check_timeout = jiffies +
324                             (xpc_hb_check_interval * HZ);
325                 }
326
327                 /* wait for IRQ or timeout */
328                 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
329                                                (last_IRQ_count < atomic_read(
330                                                 &xpc_activate_IRQ_rcvd)
331                                                 || time_is_before_eq_jiffies(
332                                                 xpc_hb_check_timeout) ||
333                                                 xpc_exiting));
334         }
335
336         xpc_stop_hb_beater();
337
338         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
339
340         /* mark this thread as having exited */
341         complete(&xpc_hb_checker_exited);
342         return 0;
343 }
344
345 /*
346  * This thread will attempt to discover other partitions to activate
347  * based on info provided by SAL. This new thread is short lived and
348  * will exit once discovery is complete.
349  */
350 static int
351 xpc_initiate_discovery(void *ignore)
352 {
353         xpc_discovery();
354
355         dev_dbg(xpc_part, "discovery thread is exiting\n");
356
357         /* mark this thread as having exited */
358         complete(&xpc_discovery_exited);
359         return 0;
360 }
361
362 /*
363  * The first kthread assigned to a newly activated partition is the one
364  * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
365  * that kthread until the partition is brought down, at which time that kthread
366  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
367  * that XPC has dismantled all communication infrastructure for the associated
368  * partition.) This kthread becomes the channel manager for that partition.
369  *
370  * Each active partition has a channel manager, who, besides connecting and
371  * disconnecting channels, will ensure that each of the partition's connected
372  * channels has the required number of assigned kthreads to get the work done.
373  */
374 static void
375 xpc_channel_mgr(struct xpc_partition *part)
376 {
377         while (part->act_state != XPC_P_DEACTIVATING ||
378                atomic_read(&part->nchannels_active) > 0 ||
379                !xpc_partition_disengaged(part)) {
380
381                 xpc_process_channel_activity(part);
382
383                 /*
384                  * Wait until we've been requested to activate kthreads or
385                  * all of the channel's message queues have been torn down or
386                  * a signal is pending.
387                  *
388                  * The channel_mgr_requests is set to 1 after being awakened,
389                  * This is done to prevent the channel mgr from making one pass
390                  * through the loop for each request, since he will
391                  * be servicing all the requests in one pass. The reason it's
392                  * set to 1 instead of 0 is so that other kthreads will know
393                  * that the channel mgr is running and won't bother trying to
394                  * wake him up.
395                  */
396                 atomic_dec(&part->channel_mgr_requests);
397                 (void)wait_event_interruptible(part->channel_mgr_wq,
398                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
399                                  part->local_IPI_amo != 0 ||
400                                  (part->act_state == XPC_P_DEACTIVATING &&
401                                  atomic_read(&part->nchannels_active) == 0 &&
402                                  xpc_partition_disengaged(part))));
403                 atomic_set(&part->channel_mgr_requests, 1);
404         }
405 }
406
407 /*
408  * When XPC HB determines that a partition has come up, it will create a new
409  * kthread and that kthread will call this function to attempt to set up the
410  * basic infrastructure used for Cross Partition Communication with the newly
411  * upped partition.
412  *
413  * The kthread that was created by XPC HB and which setup the XPC
414  * infrastructure will remain assigned to the partition becoming the channel
415  * manager for that partition until the partition is deactivating, at which
416  * time the kthread will teardown the XPC infrastructure and then exit.
417  */
418 static int
419 xpc_activating(void *__partid)
420 {
421         short partid = (u64)__partid;
422         struct xpc_partition *part = &xpc_partitions[partid];
423         unsigned long irq_flags;
424
425         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
426
427         spin_lock_irqsave(&part->act_lock, irq_flags);
428
429         if (part->act_state == XPC_P_DEACTIVATING) {
430                 part->act_state = XPC_P_INACTIVE;
431                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
432                 part->remote_rp_pa = 0;
433                 return 0;
434         }
435
436         /* indicate the thread is activating */
437         DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
438         part->act_state = XPC_P_ACTIVATING;
439
440         XPC_SET_REASON(part, 0, 0);
441         spin_unlock_irqrestore(&part->act_lock, irq_flags);
442
443         dev_dbg(xpc_part, "activating partition %d\n", partid);
444
445         xpc_allow_hb(partid);
446
447         if (xpc_setup_infrastructure(part) == xpSuccess) {
448                 (void)xpc_part_ref(part);       /* this will always succeed */
449
450                 if (xpc_make_first_contact(part) == xpSuccess) {
451                         xpc_mark_partition_active(part);
452                         xpc_channel_mgr(part);
453                         /* won't return until partition is deactivating */
454                 }
455
456                 xpc_part_deref(part);
457                 xpc_teardown_infrastructure(part);
458         }
459
460         xpc_disallow_hb(partid);
461         xpc_mark_partition_inactive(part);
462
463         if (part->reason == xpReactivating) {
464                 /* interrupting ourselves results in activating partition */
465                 xpc_request_partition_reactivation(part);
466         }
467
468         return 0;
469 }
470
471 void
472 xpc_activate_partition(struct xpc_partition *part)
473 {
474         short partid = XPC_PARTID(part);
475         unsigned long irq_flags;
476         struct task_struct *kthread;
477
478         spin_lock_irqsave(&part->act_lock, irq_flags);
479
480         DBUG_ON(part->act_state != XPC_P_INACTIVE);
481
482         part->act_state = XPC_P_ACTIVATION_REQ;
483         XPC_SET_REASON(part, xpCloneKThread, __LINE__);
484
485         spin_unlock_irqrestore(&part->act_lock, irq_flags);
486
487         kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
488                               partid);
489         if (IS_ERR(kthread)) {
490                 spin_lock_irqsave(&part->act_lock, irq_flags);
491                 part->act_state = XPC_P_INACTIVE;
492                 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
493                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
494         }
495 }
496
497 void
498 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
499 {
500         int idle = atomic_read(&ch->kthreads_idle);
501         int assigned = atomic_read(&ch->kthreads_assigned);
502         int wakeup;
503
504         DBUG_ON(needed <= 0);
505
506         if (idle > 0) {
507                 wakeup = (needed > idle) ? idle : needed;
508                 needed -= wakeup;
509
510                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
511                         "channel=%d\n", wakeup, ch->partid, ch->number);
512
513                 /* only wakeup the requested number of kthreads */
514                 wake_up_nr(&ch->idle_wq, wakeup);
515         }
516
517         if (needed <= 0)
518                 return;
519
520         if (needed + assigned > ch->kthreads_assigned_limit) {
521                 needed = ch->kthreads_assigned_limit - assigned;
522                 if (needed <= 0)
523                         return;
524         }
525
526         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
527                 needed, ch->partid, ch->number);
528
529         xpc_create_kthreads(ch, needed, 0);
530 }
531
532 /*
533  * This function is where XPC's kthreads wait for messages to deliver.
534  */
535 static void
536 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
537 {
538         do {
539                 /* deliver messages to their intended recipients */
540
541                 while (xpc_n_of_deliverable_msgs(ch) > 0 &&
542                        !(ch->flags & XPC_C_DISCONNECTING)) {
543                         xpc_deliver_msg(ch);
544                 }
545
546                 if (atomic_inc_return(&ch->kthreads_idle) >
547                     ch->kthreads_idle_limit) {
548                         /* too many idle kthreads on this channel */
549                         atomic_dec(&ch->kthreads_idle);
550                         break;
551                 }
552
553                 dev_dbg(xpc_chan, "idle kthread calling "
554                         "wait_event_interruptible_exclusive()\n");
555
556                 (void)wait_event_interruptible_exclusive(ch->idle_wq,
557                                 (xpc_n_of_deliverable_msgs(ch) > 0 ||
558                                  (ch->flags & XPC_C_DISCONNECTING)));
559
560                 atomic_dec(&ch->kthreads_idle);
561
562         } while (!(ch->flags & XPC_C_DISCONNECTING));
563 }
564
565 static int
566 xpc_kthread_start(void *args)
567 {
568         short partid = XPC_UNPACK_ARG1(args);
569         u16 ch_number = XPC_UNPACK_ARG2(args);
570         struct xpc_partition *part = &xpc_partitions[partid];
571         struct xpc_channel *ch;
572         int n_needed;
573         unsigned long irq_flags;
574
575         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
576                 partid, ch_number);
577
578         ch = &part->channels[ch_number];
579
580         if (!(ch->flags & XPC_C_DISCONNECTING)) {
581
582                 /* let registerer know that connection has been established */
583
584                 spin_lock_irqsave(&ch->lock, irq_flags);
585                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
586                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
587                         spin_unlock_irqrestore(&ch->lock, irq_flags);
588
589                         xpc_connected_callout(ch);
590
591                         spin_lock_irqsave(&ch->lock, irq_flags);
592                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
593                         spin_unlock_irqrestore(&ch->lock, irq_flags);
594
595                         /*
596                          * It is possible that while the callout was being
597                          * made that the remote partition sent some messages.
598                          * If that is the case, we may need to activate
599                          * additional kthreads to help deliver them. We only
600                          * need one less than total #of messages to deliver.
601                          */
602                         n_needed = xpc_n_of_deliverable_msgs(ch) - 1;
603                         if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
604                                 xpc_activate_kthreads(ch, n_needed);
605
606                 } else {
607                         spin_unlock_irqrestore(&ch->lock, irq_flags);
608                 }
609
610                 xpc_kthread_waitmsgs(part, ch);
611         }
612
613         /* let registerer know that connection is disconnecting */
614
615         spin_lock_irqsave(&ch->lock, irq_flags);
616         if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
617             !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
618                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
619                 spin_unlock_irqrestore(&ch->lock, irq_flags);
620
621                 xpc_disconnect_callout(ch, xpDisconnecting);
622
623                 spin_lock_irqsave(&ch->lock, irq_flags);
624                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
625         }
626         spin_unlock_irqrestore(&ch->lock, irq_flags);
627
628         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
629             atomic_dec_return(&part->nchannels_engaged) == 0) {
630                 xpc_indicate_partition_disengaged(part);
631         }
632
633         xpc_msgqueue_deref(ch);
634
635         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
636                 partid, ch_number);
637
638         xpc_part_deref(part);
639         return 0;
640 }
641
642 /*
643  * For each partition that XPC has established communications with, there is
644  * a minimum of one kernel thread assigned to perform any operation that
645  * may potentially sleep or block (basically the callouts to the asynchronous
646  * functions registered via xpc_connect()).
647  *
648  * Additional kthreads are created and destroyed by XPC as the workload
649  * demands.
650  *
651  * A kthread is assigned to one of the active channels that exists for a given
652  * partition.
653  */
654 void
655 xpc_create_kthreads(struct xpc_channel *ch, int needed,
656                     int ignore_disconnecting)
657 {
658         unsigned long irq_flags;
659         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
660         struct xpc_partition *part = &xpc_partitions[ch->partid];
661         struct task_struct *kthread;
662
663         while (needed-- > 0) {
664
665                 /*
666                  * The following is done on behalf of the newly created
667                  * kthread. That kthread is responsible for doing the
668                  * counterpart to the following before it exits.
669                  */
670                 if (ignore_disconnecting) {
671                         if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
672                                 /* kthreads assigned had gone to zero */
673                                 BUG_ON(!(ch->flags &
674                                          XPC_C_DISCONNECTINGCALLOUT_MADE));
675                                 break;
676                         }
677
678                 } else if (ch->flags & XPC_C_DISCONNECTING) {
679                         break;
680
681                 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
682                            atomic_inc_return(&part->nchannels_engaged) == 1) {
683                                 xpc_indicate_partition_engaged(part);
684                 }
685                 (void)xpc_part_ref(part);
686                 xpc_msgqueue_ref(ch);
687
688                 kthread = kthread_run(xpc_kthread_start, (void *)args,
689                                       "xpc%02dc%d", ch->partid, ch->number);
690                 if (IS_ERR(kthread)) {
691                         /* the fork failed */
692
693                         /*
694                          * NOTE: if (ignore_disconnecting &&
695                          * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
696                          * then we'll deadlock if all other kthreads assigned
697                          * to this channel are blocked in the channel's
698                          * registerer, because the only thing that will unblock
699                          * them is the xpDisconnecting callout that this
700                          * failed kthread_run() would have made.
701                          */
702
703                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
704                             atomic_dec_return(&part->nchannels_engaged) == 0) {
705                                 xpc_indicate_partition_disengaged(part);
706                         }
707                         xpc_msgqueue_deref(ch);
708                         xpc_part_deref(part);
709
710                         if (atomic_read(&ch->kthreads_assigned) <
711                             ch->kthreads_idle_limit) {
712                                 /*
713                                  * Flag this as an error only if we have an
714                                  * insufficient #of kthreads for the channel
715                                  * to function.
716                                  */
717                                 spin_lock_irqsave(&ch->lock, irq_flags);
718                                 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
719                                                        &irq_flags);
720                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
721                         }
722                         break;
723                 }
724         }
725 }
726
727 void
728 xpc_disconnect_wait(int ch_number)
729 {
730         unsigned long irq_flags;
731         short partid;
732         struct xpc_partition *part;
733         struct xpc_channel *ch;
734         int wakeup_channel_mgr;
735
736         /* now wait for all callouts to the caller's function to cease */
737         for (partid = 0; partid < xp_max_npartitions; partid++) {
738                 part = &xpc_partitions[partid];
739
740                 if (!xpc_part_ref(part))
741                         continue;
742
743                 ch = &part->channels[ch_number];
744
745                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
746                         xpc_part_deref(part);
747                         continue;
748                 }
749
750                 wait_for_completion(&ch->wdisconnect_wait);
751
752                 spin_lock_irqsave(&ch->lock, irq_flags);
753                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
754                 wakeup_channel_mgr = 0;
755
756                 if (ch->delayed_IPI_flags) {
757                         if (part->act_state != XPC_P_DEACTIVATING) {
758                                 spin_lock(&part->IPI_lock);
759                                 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
760                                                   ch->number,
761                                                   ch->delayed_IPI_flags);
762                                 spin_unlock(&part->IPI_lock);
763                                 wakeup_channel_mgr = 1;
764                         }
765                         ch->delayed_IPI_flags = 0;
766                 }
767
768                 ch->flags &= ~XPC_C_WDISCONNECT;
769                 spin_unlock_irqrestore(&ch->lock, irq_flags);
770
771                 if (wakeup_channel_mgr)
772                         xpc_wakeup_channel_mgr(part);
773
774                 xpc_part_deref(part);
775         }
776 }
777
778 static void
779 xpc_do_exit(enum xp_retval reason)
780 {
781         short partid;
782         int active_part_count, printed_waiting_msg = 0;
783         struct xpc_partition *part;
784         unsigned long printmsg_time, disengage_timeout = 0;
785
786         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
787         DBUG_ON(xpc_exiting == 1);
788
789         /*
790          * Let the heartbeat checker thread and the discovery thread
791          * (if one is running) know that they should exit. Also wake up
792          * the heartbeat checker thread in case it's sleeping.
793          */
794         xpc_exiting = 1;
795         wake_up_interruptible(&xpc_activate_IRQ_wq);
796
797         /* wait for the discovery thread to exit */
798         wait_for_completion(&xpc_discovery_exited);
799
800         /* wait for the heartbeat checker thread to exit */
801         wait_for_completion(&xpc_hb_checker_exited);
802
803         /* sleep for a 1/3 of a second or so */
804         (void)msleep_interruptible(300);
805
806         /* wait for all partitions to become inactive */
807
808         printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
809         xpc_disengage_timedout = 0;
810
811         do {
812                 active_part_count = 0;
813
814                 for (partid = 0; partid < xp_max_npartitions; partid++) {
815                         part = &xpc_partitions[partid];
816
817                         if (xpc_partition_disengaged(part) &&
818                             part->act_state == XPC_P_INACTIVE) {
819                                 continue;
820                         }
821
822                         active_part_count++;
823
824                         XPC_DEACTIVATE_PARTITION(part, reason);
825
826                         if (part->disengage_timeout > disengage_timeout)
827                                 disengage_timeout = part->disengage_timeout;
828                 }
829
830                 if (xpc_any_partition_engaged()) {
831                         if (time_is_before_jiffies(printmsg_time)) {
832                                 dev_info(xpc_part, "waiting for remote "
833                                          "partitions to deactivate, timeout in "
834                                          "%ld seconds\n", (disengage_timeout -
835                                          jiffies) / HZ);
836                                 printmsg_time = jiffies +
837                                     (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
838                                 printed_waiting_msg = 1;
839                         }
840
841                 } else if (active_part_count > 0) {
842                         if (printed_waiting_msg) {
843                                 dev_info(xpc_part, "waiting for local partition"
844                                          " to deactivate\n");
845                                 printed_waiting_msg = 0;
846                         }
847
848                 } else {
849                         if (!xpc_disengage_timedout) {
850                                 dev_info(xpc_part, "all partitions have "
851                                          "deactivated\n");
852                         }
853                         break;
854                 }
855
856                 /* sleep for a 1/3 of a second or so */
857                 (void)msleep_interruptible(300);
858
859         } while (1);
860
861         DBUG_ON(xpc_any_partition_engaged());
862         DBUG_ON(xpc_any_hbs_allowed() != 0);
863
864         /* indicate to others that our reserved page is uninitialized */
865         xpc_rsvd_page->stamp = 0;
866
867         if (reason == xpUnloading) {
868                 (void)unregister_die_notifier(&xpc_die_notifier);
869                 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
870         }
871
872         /* clear the interface to XPC's functions */
873         xpc_clear_interface();
874
875         if (xpc_sysctl)
876                 unregister_sysctl_table(xpc_sysctl);
877
878         kfree(xpc_partitions);
879         kfree(xpc_remote_copy_buffer_base);
880
881         if (is_shub())
882                 xpc_exit_sn2();
883         else
884                 xpc_exit_uv();
885 }
886
887 /*
888  * This function is called when the system is being rebooted.
889  */
890 static int
891 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
892 {
893         enum xp_retval reason;
894
895         switch (event) {
896         case SYS_RESTART:
897                 reason = xpSystemReboot;
898                 break;
899         case SYS_HALT:
900                 reason = xpSystemHalt;
901                 break;
902         case SYS_POWER_OFF:
903                 reason = xpSystemPoweroff;
904                 break;
905         default:
906                 reason = xpSystemGoingDown;
907         }
908
909         xpc_do_exit(reason);
910         return NOTIFY_DONE;
911 }
912
913 /*
914  * Notify other partitions to deactivate from us by first disengaging from all
915  * references to our memory.
916  */
917 static void
918 xpc_die_deactivate(void)
919 {
920         struct xpc_partition *part;
921         short partid;
922         int any_engaged;
923         long time, printmsg_time, disengage_timeout;
924
925         /* keep xpc_hb_checker thread from doing anything (just in case) */
926         xpc_exiting = 1;
927
928         xpc_disallow_all_hbs(); /*indicate we're deactivated */
929
930         for (partid = 0; partid < xp_max_npartitions; partid++) {
931                 part = &xpc_partitions[partid];
932
933                 if (xpc_partition_engaged(partid) ||
934                     part->act_state != XPC_P_INACTIVE) {
935                         xpc_request_partition_deactivation(part);
936                         xpc_indicate_partition_disengaged(part);
937                 }
938         }
939
940         time = rtc_time();
941         printmsg_time = time +
942             (XPC_DEACTIVATE_PRINTMSG_INTERVAL * sn_rtc_cycles_per_second);
943         disengage_timeout = time +
944             (xpc_disengage_timelimit * sn_rtc_cycles_per_second);
945
946         /*
947          * Though we requested that all other partitions deactivate from us,
948          * we only wait until they've all disengaged.
949          */
950
951         while (1) {
952                 any_engaged = xpc_any_partition_engaged();
953                 if (!any_engaged) {
954                         dev_info(xpc_part, "all partitions have deactivated\n");
955                         break;
956                 }
957
958                 time = rtc_time();
959                 if (time >= disengage_timeout) {
960                         for (partid = 0; partid < xp_max_npartitions;
961                              partid++) {
962                                 if (xpc_partition_engaged(partid)) {
963                                         dev_info(xpc_part, "deactivate from "
964                                                  "remote partition %d timed "
965                                                  "out\n", partid);
966                                 }
967                         }
968                         break;
969                 }
970
971                 if (time >= printmsg_time) {
972                         dev_info(xpc_part, "waiting for remote partitions to "
973                                  "deactivate, timeout in %ld seconds\n",
974                                  (disengage_timeout - time) /
975                                  sn_rtc_cycles_per_second);
976                         printmsg_time = time +
977                             (XPC_DEACTIVATE_PRINTMSG_INTERVAL *
978                              sn_rtc_cycles_per_second);
979                 }
980         }
981 }
982
983 /*
984  * This function is called when the system is being restarted or halted due
985  * to some sort of system failure. If this is the case we need to notify the
986  * other partitions to disengage from all references to our memory.
987  * This function can also be called when our heartbeater could be offlined
988  * for a time. In this case we need to notify other partitions to not worry
989  * about the lack of a heartbeat.
990  */
991 static int
992 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
993 {
994         switch (event) {
995         case DIE_MACHINE_RESTART:
996         case DIE_MACHINE_HALT:
997                 xpc_die_deactivate();
998                 break;
999
1000         case DIE_KDEBUG_ENTER:
1001                 /* Should lack of heartbeat be ignored by other partitions? */
1002                 if (!xpc_kdebug_ignore)
1003                         break;
1004
1005                 /* fall through */
1006         case DIE_MCA_MONARCH_ENTER:
1007         case DIE_INIT_MONARCH_ENTER:
1008                 xpc_offline_heartbeat();
1009                 break;
1010
1011         case DIE_KDEBUG_LEAVE:
1012                 /* Is lack of heartbeat being ignored by other partitions? */
1013                 if (!xpc_kdebug_ignore)
1014                         break;
1015
1016                 /* fall through */
1017         case DIE_MCA_MONARCH_LEAVE:
1018         case DIE_INIT_MONARCH_LEAVE:
1019                 xpc_online_heartbeat();
1020                 break;
1021         }
1022
1023         return NOTIFY_DONE;
1024 }
1025
1026 int __init
1027 xpc_init(void)
1028 {
1029         int ret;
1030         short partid;
1031         struct xpc_partition *part;
1032         struct task_struct *kthread;
1033         size_t buf_size;
1034
1035         if (is_shub()) {
1036                 /*
1037                  * The ia64-sn2 architecture supports at most 64 partitions.
1038                  * And the inability to unregister remote AMOs restricts us
1039                  * further to only support exactly 64 partitions on this
1040                  * architecture, no less.
1041                  */
1042                 if (xp_max_npartitions != 64)
1043                         return -EINVAL;
1044
1045                 ret = xpc_init_sn2();
1046                 if (ret != 0)
1047                         return ret;
1048
1049         } else if (is_uv()) {
1050                 xpc_init_uv();
1051
1052         } else {
1053                 return -ENODEV;
1054         }
1055
1056         snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1057         snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1058
1059         buf_size = max(XPC_RP_VARS_SIZE,
1060                        XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES);
1061         xpc_remote_copy_buffer = xpc_kmalloc_cacheline_aligned(buf_size,
1062                                                                GFP_KERNEL,
1063                                                   &xpc_remote_copy_buffer_base);
1064         if (xpc_remote_copy_buffer == NULL) {
1065                 dev_err(xpc_part, "can't get memory for remote copy buffer\n");
1066                 ret = -ENOMEM;
1067                 goto out_1;
1068         }
1069
1070         xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
1071                                  xp_max_npartitions, GFP_KERNEL);
1072         if (xpc_partitions == NULL) {
1073                 dev_err(xpc_part, "can't get memory for partition structure\n");
1074                 ret = -ENOMEM;
1075                 goto out_2;
1076         }
1077
1078         /*
1079          * The first few fields of each entry of xpc_partitions[] need to
1080          * be initialized now so that calls to xpc_connect() and
1081          * xpc_disconnect() can be made prior to the activation of any remote
1082          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1083          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1084          * PARTITION HAS BEEN ACTIVATED.
1085          */
1086         for (partid = 0; partid < xp_max_npartitions; partid++) {
1087                 part = &xpc_partitions[partid];
1088
1089                 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
1090
1091                 part->activate_IRQ_rcvd = 0;
1092                 spin_lock_init(&part->act_lock);
1093                 part->act_state = XPC_P_INACTIVE;
1094                 XPC_SET_REASON(part, 0, 0);
1095
1096                 init_timer(&part->disengage_timer);
1097                 part->disengage_timer.function =
1098                     xpc_timeout_partition_disengage;
1099                 part->disengage_timer.data = (unsigned long)part;
1100
1101                 part->setup_state = XPC_P_UNSET;
1102                 init_waitqueue_head(&part->teardown_wq);
1103                 atomic_set(&part->references, 0);
1104         }
1105
1106         xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1107
1108         /*
1109          * Fill the partition reserved page with the information needed by
1110          * other partitions to discover we are alive and establish initial
1111          * communications.
1112          */
1113         xpc_rsvd_page = xpc_setup_rsvd_page();
1114         if (xpc_rsvd_page == NULL) {
1115                 dev_err(xpc_part, "can't setup our reserved page\n");
1116                 ret = -EBUSY;
1117                 goto out_3;
1118         }
1119
1120         /* add ourselves to the reboot_notifier_list */
1121         ret = register_reboot_notifier(&xpc_reboot_notifier);
1122         if (ret != 0)
1123                 dev_warn(xpc_part, "can't register reboot notifier\n");
1124
1125         /* add ourselves to the die_notifier list */
1126         ret = register_die_notifier(&xpc_die_notifier);
1127         if (ret != 0)
1128                 dev_warn(xpc_part, "can't register die notifier\n");
1129
1130         /*
1131          * The real work-horse behind xpc.  This processes incoming
1132          * interrupts and monitors remote heartbeats.
1133          */
1134         kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1135         if (IS_ERR(kthread)) {
1136                 dev_err(xpc_part, "failed while forking hb check thread\n");
1137                 ret = -EBUSY;
1138                 goto out_4;
1139         }
1140
1141         /*
1142          * Startup a thread that will attempt to discover other partitions to
1143          * activate based on info provided by SAL. This new thread is short
1144          * lived and will exit once discovery is complete.
1145          */
1146         kthread = kthread_run(xpc_initiate_discovery, NULL,
1147                               XPC_DISCOVERY_THREAD_NAME);
1148         if (IS_ERR(kthread)) {
1149                 dev_err(xpc_part, "failed while forking discovery thread\n");
1150
1151                 /* mark this new thread as a non-starter */
1152                 complete(&xpc_discovery_exited);
1153
1154                 xpc_do_exit(xpUnloading);
1155                 return -EBUSY;
1156         }
1157
1158         /* set the interface to point at XPC's functions */
1159         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1160                           xpc_initiate_send, xpc_initiate_send_notify,
1161                           xpc_initiate_received, xpc_initiate_partid_to_nasids);
1162
1163         return 0;
1164
1165         /* initialization was not successful */
1166 out_4:
1167         /* indicate to others that our reserved page is uninitialized */
1168         xpc_rsvd_page->stamp = 0;
1169
1170         (void)unregister_die_notifier(&xpc_die_notifier);
1171         (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1172 out_3:
1173         if (xpc_sysctl)
1174                 unregister_sysctl_table(xpc_sysctl);
1175         kfree(xpc_partitions);
1176 out_2:
1177         kfree(xpc_remote_copy_buffer_base);
1178 out_1:
1179         if (is_shub())
1180                 xpc_exit_sn2();
1181         else
1182                 xpc_exit_uv();
1183         return ret;
1184 }
1185
1186 module_init(xpc_init);
1187
1188 void __exit
1189 xpc_exit(void)
1190 {
1191         xpc_do_exit(xpUnloading);
1192 }
1193
1194 module_exit(xpc_exit);
1195
1196 MODULE_AUTHOR("Silicon Graphics, Inc.");
1197 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1198 MODULE_LICENSE("GPL");
1199
1200 module_param(xpc_hb_interval, int, 0);
1201 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1202                  "heartbeat increments.");
1203
1204 module_param(xpc_hb_check_interval, int, 0);
1205 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1206                  "heartbeat checks.");
1207
1208 module_param(xpc_disengage_timelimit, int, 0);
1209 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1210                  "for disengage to complete.");
1211
1212 module_param(xpc_kdebug_ignore, int, 0);
1213 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1214                  "other partitions when dropping into kdebug.");