sgi-xp: support runtime selection of xp_max_npartitions
[linux-2.6.git] / drivers / misc / sgi-xp / xpc_channel.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) channel support.
11  *
12  *      This is the part of XPC that manages the channels and
13  *      sends/receives messages across them to/from other partitions.
14  *
15  */
16
17 #include <linux/kernel.h>
18 #include <linux/init.h>
19 #include <linux/sched.h>
20 #include <linux/cache.h>
21 #include <linux/interrupt.h>
22 #include <linux/mutex.h>
23 #include <linux/completion.h>
24 #include <asm/sn/bte.h>
25 #include <asm/sn/sn_sal.h>
26 #include "xpc.h"
27
28 /*
29  * Guarantee that the kzalloc'd memory is cacheline aligned.
30  */
31 static void *
32 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
33 {
34         /* see if kzalloc will give us cachline aligned memory by default */
35         *base = kzalloc(size, flags);
36         if (*base == NULL)
37                 return NULL;
38
39         if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
40                 return *base;
41
42         kfree(*base);
43
44         /* nope, we'll have to do it ourselves */
45         *base = kzalloc(size + L1_CACHE_BYTES, flags);
46         if (*base == NULL)
47                 return NULL;
48
49         return (void *)L1_CACHE_ALIGN((u64)*base);
50 }
51
52 /*
53  * Set up the initial values for the XPartition Communication channels.
54  */
55 static void
56 xpc_initialize_channels(struct xpc_partition *part, short partid)
57 {
58         int ch_number;
59         struct xpc_channel *ch;
60
61         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
62                 ch = &part->channels[ch_number];
63
64                 ch->partid = partid;
65                 ch->number = ch_number;
66                 ch->flags = XPC_C_DISCONNECTED;
67
68                 ch->local_GP = &part->local_GPs[ch_number];
69                 ch->local_openclose_args =
70                     &part->local_openclose_args[ch_number];
71
72                 atomic_set(&ch->kthreads_assigned, 0);
73                 atomic_set(&ch->kthreads_idle, 0);
74                 atomic_set(&ch->kthreads_active, 0);
75
76                 atomic_set(&ch->references, 0);
77                 atomic_set(&ch->n_to_notify, 0);
78
79                 spin_lock_init(&ch->lock);
80                 mutex_init(&ch->msg_to_pull_mutex);
81                 init_completion(&ch->wdisconnect_wait);
82
83                 atomic_set(&ch->n_on_msg_allocate_wq, 0);
84                 init_waitqueue_head(&ch->msg_allocate_wq);
85                 init_waitqueue_head(&ch->idle_wq);
86         }
87 }
88
89 /*
90  * Setup the infrastructure necessary to support XPartition Communication
91  * between the specified remote partition and the local one.
92  */
93 enum xp_retval
94 xpc_setup_infrastructure(struct xpc_partition *part)
95 {
96         int ret, cpuid;
97         struct timer_list *timer;
98         short partid = XPC_PARTID(part);
99
100         /*
101          * Zero out MOST of the entry for this partition. Only the fields
102          * starting with `nchannels' will be zeroed. The preceding fields must
103          * remain `viable' across partition ups and downs, since they may be
104          * referenced during this memset() operation.
105          */
106         memset(&part->nchannels, 0, sizeof(struct xpc_partition) -
107                offsetof(struct xpc_partition, nchannels));
108
109         /*
110          * Allocate all of the channel structures as a contiguous chunk of
111          * memory.
112          */
113         part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
114                                  GFP_KERNEL);
115         if (part->channels == NULL) {
116                 dev_err(xpc_chan, "can't get memory for channels\n");
117                 return xpNoMemory;
118         }
119
120         part->nchannels = XPC_MAX_NCHANNELS;
121
122         /* allocate all the required GET/PUT values */
123
124         part->local_GPs = xpc_kzalloc_cacheline_aligned(XPC_GP_SIZE,
125                                                         GFP_KERNEL,
126                                                         &part->local_GPs_base);
127         if (part->local_GPs == NULL) {
128                 kfree(part->channels);
129                 part->channels = NULL;
130                 dev_err(xpc_chan, "can't get memory for local get/put "
131                         "values\n");
132                 return xpNoMemory;
133         }
134
135         part->remote_GPs = xpc_kzalloc_cacheline_aligned(XPC_GP_SIZE,
136                                                          GFP_KERNEL,
137                                                          &part->
138                                                          remote_GPs_base);
139         if (part->remote_GPs == NULL) {
140                 dev_err(xpc_chan, "can't get memory for remote get/put "
141                         "values\n");
142                 kfree(part->local_GPs_base);
143                 part->local_GPs = NULL;
144                 kfree(part->channels);
145                 part->channels = NULL;
146                 return xpNoMemory;
147         }
148
149         /* allocate all the required open and close args */
150
151         part->local_openclose_args =
152             xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
153                                           &part->local_openclose_args_base);
154         if (part->local_openclose_args == NULL) {
155                 dev_err(xpc_chan, "can't get memory for local connect args\n");
156                 kfree(part->remote_GPs_base);
157                 part->remote_GPs = NULL;
158                 kfree(part->local_GPs_base);
159                 part->local_GPs = NULL;
160                 kfree(part->channels);
161                 part->channels = NULL;
162                 return xpNoMemory;
163         }
164
165         part->remote_openclose_args =
166             xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
167                                           &part->remote_openclose_args_base);
168         if (part->remote_openclose_args == NULL) {
169                 dev_err(xpc_chan, "can't get memory for remote connect args\n");
170                 kfree(part->local_openclose_args_base);
171                 part->local_openclose_args = NULL;
172                 kfree(part->remote_GPs_base);
173                 part->remote_GPs = NULL;
174                 kfree(part->local_GPs_base);
175                 part->local_GPs = NULL;
176                 kfree(part->channels);
177                 part->channels = NULL;
178                 return xpNoMemory;
179         }
180
181         xpc_initialize_channels(part, partid);
182
183         atomic_set(&part->nchannels_active, 0);
184         atomic_set(&part->nchannels_engaged, 0);
185
186         /* local_IPI_amo were set to 0 by an earlier memset() */
187
188         /* Initialize this partitions AMO_t structure */
189         part->local_IPI_amo_va = xpc_IPI_init(partid);
190
191         spin_lock_init(&part->IPI_lock);
192
193         atomic_set(&part->channel_mgr_requests, 1);
194         init_waitqueue_head(&part->channel_mgr_wq);
195
196         sprintf(part->IPI_owner, "xpc%02d", partid);
197         ret = request_irq(SGI_XPC_NOTIFY, xpc_notify_IRQ_handler, IRQF_SHARED,
198                           part->IPI_owner, (void *)(u64)partid);
199         if (ret != 0) {
200                 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
201                         "errno=%d\n", -ret);
202                 kfree(part->remote_openclose_args_base);
203                 part->remote_openclose_args = NULL;
204                 kfree(part->local_openclose_args_base);
205                 part->local_openclose_args = NULL;
206                 kfree(part->remote_GPs_base);
207                 part->remote_GPs = NULL;
208                 kfree(part->local_GPs_base);
209                 part->local_GPs = NULL;
210                 kfree(part->channels);
211                 part->channels = NULL;
212                 return xpLackOfResources;
213         }
214
215         /* Setup a timer to check for dropped IPIs */
216         timer = &part->dropped_IPI_timer;
217         init_timer(timer);
218         timer->function = (void (*)(unsigned long))xpc_dropped_IPI_check;
219         timer->data = (unsigned long)part;
220         timer->expires = jiffies + XPC_P_DROPPED_IPI_WAIT;
221         add_timer(timer);
222
223         /*
224          * With the setting of the partition setup_state to XPC_P_SETUP, we're
225          * declaring that this partition is ready to go.
226          */
227         part->setup_state = XPC_P_SETUP;
228
229         /*
230          * Setup the per partition specific variables required by the
231          * remote partition to establish channel connections with us.
232          *
233          * The setting of the magic # indicates that these per partition
234          * specific variables are ready to be used.
235          */
236         xpc_vars_part[partid].GPs_pa = __pa(part->local_GPs);
237         xpc_vars_part[partid].openclose_args_pa =
238             __pa(part->local_openclose_args);
239         xpc_vars_part[partid].IPI_amo_pa = __pa(part->local_IPI_amo_va);
240         cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
241         xpc_vars_part[partid].IPI_nasid = cpuid_to_nasid(cpuid);
242         xpc_vars_part[partid].IPI_phys_cpuid = cpu_physical_id(cpuid);
243         xpc_vars_part[partid].nchannels = part->nchannels;
244         xpc_vars_part[partid].magic = XPC_VP_MAGIC1;
245
246         return xpSuccess;
247 }
248
249 /*
250  * Create a wrapper that hides the underlying mechanism for pulling a cacheline
251  * (or multiple cachelines) from a remote partition.
252  *
253  * src must be a cacheline aligned physical address on the remote partition.
254  * dst must be a cacheline aligned virtual address on this partition.
255  * cnt must be an cacheline sized
256  */
257 static enum xp_retval
258 xpc_pull_remote_cachelines(struct xpc_partition *part, void *dst,
259                            const void *src, size_t cnt)
260 {
261         bte_result_t bte_ret;
262
263         DBUG_ON((u64)src != L1_CACHE_ALIGN((u64)src));
264         DBUG_ON((u64)dst != L1_CACHE_ALIGN((u64)dst));
265         DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
266
267         if (part->act_state == XPC_P_DEACTIVATING)
268                 return part->reason;
269
270         bte_ret = xp_bte_copy((u64)src, (u64)dst, (u64)cnt,
271                               (BTE_NORMAL | BTE_WACQUIRE), NULL);
272         if (bte_ret == BTE_SUCCESS)
273                 return xpSuccess;
274
275         dev_dbg(xpc_chan, "xp_bte_copy() from partition %d failed, ret=%d\n",
276                 XPC_PARTID(part), bte_ret);
277
278         return xpc_map_bte_errors(bte_ret);
279 }
280
281 /*
282  * Pull the remote per partition specific variables from the specified
283  * partition.
284  */
285 enum xp_retval
286 xpc_pull_remote_vars_part(struct xpc_partition *part)
287 {
288         u8 buffer[L1_CACHE_BYTES * 2];
289         struct xpc_vars_part *pulled_entry_cacheline =
290             (struct xpc_vars_part *)L1_CACHE_ALIGN((u64)buffer);
291         struct xpc_vars_part *pulled_entry;
292         u64 remote_entry_cacheline_pa, remote_entry_pa;
293         short partid = XPC_PARTID(part);
294         enum xp_retval ret;
295
296         /* pull the cacheline that contains the variables we're interested in */
297
298         DBUG_ON(part->remote_vars_part_pa !=
299                 L1_CACHE_ALIGN(part->remote_vars_part_pa));
300         DBUG_ON(sizeof(struct xpc_vars_part) != L1_CACHE_BYTES / 2);
301
302         remote_entry_pa = part->remote_vars_part_pa +
303             sn_partition_id * sizeof(struct xpc_vars_part);
304
305         remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
306
307         pulled_entry = (struct xpc_vars_part *)((u64)pulled_entry_cacheline +
308                                                 (remote_entry_pa &
309                                                  (L1_CACHE_BYTES - 1)));
310
311         ret = xpc_pull_remote_cachelines(part, pulled_entry_cacheline,
312                                          (void *)remote_entry_cacheline_pa,
313                                          L1_CACHE_BYTES);
314         if (ret != xpSuccess) {
315                 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
316                         "partition %d, ret=%d\n", partid, ret);
317                 return ret;
318         }
319
320         /* see if they've been set up yet */
321
322         if (pulled_entry->magic != XPC_VP_MAGIC1 &&
323             pulled_entry->magic != XPC_VP_MAGIC2) {
324
325                 if (pulled_entry->magic != 0) {
326                         dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
327                                 "partition %d has bad magic value (=0x%lx)\n",
328                                 partid, sn_partition_id, pulled_entry->magic);
329                         return xpBadMagic;
330                 }
331
332                 /* they've not been initialized yet */
333                 return xpRetry;
334         }
335
336         if (xpc_vars_part[partid].magic == XPC_VP_MAGIC1) {
337
338                 /* validate the variables */
339
340                 if (pulled_entry->GPs_pa == 0 ||
341                     pulled_entry->openclose_args_pa == 0 ||
342                     pulled_entry->IPI_amo_pa == 0) {
343
344                         dev_err(xpc_chan, "partition %d's XPC vars_part for "
345                                 "partition %d are not valid\n", partid,
346                                 sn_partition_id);
347                         return xpInvalidAddress;
348                 }
349
350                 /* the variables we imported look to be valid */
351
352                 part->remote_GPs_pa = pulled_entry->GPs_pa;
353                 part->remote_openclose_args_pa =
354                     pulled_entry->openclose_args_pa;
355                 part->remote_IPI_amo_va =
356                     (AMO_t *)__va(pulled_entry->IPI_amo_pa);
357                 part->remote_IPI_nasid = pulled_entry->IPI_nasid;
358                 part->remote_IPI_phys_cpuid = pulled_entry->IPI_phys_cpuid;
359
360                 if (part->nchannels > pulled_entry->nchannels)
361                         part->nchannels = pulled_entry->nchannels;
362
363                 /* let the other side know that we've pulled their variables */
364
365                 xpc_vars_part[partid].magic = XPC_VP_MAGIC2;
366         }
367
368         if (pulled_entry->magic == XPC_VP_MAGIC1)
369                 return xpRetry;
370
371         return xpSuccess;
372 }
373
374 /*
375  * Get the IPI flags and pull the openclose args and/or remote GPs as needed.
376  */
377 static u64
378 xpc_get_IPI_flags(struct xpc_partition *part)
379 {
380         unsigned long irq_flags;
381         u64 IPI_amo;
382         enum xp_retval ret;
383
384         /*
385          * See if there are any IPI flags to be handled.
386          */
387
388         spin_lock_irqsave(&part->IPI_lock, irq_flags);
389         IPI_amo = part->local_IPI_amo;
390         if (IPI_amo != 0)
391                 part->local_IPI_amo = 0;
392
393         spin_unlock_irqrestore(&part->IPI_lock, irq_flags);
394
395         if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_amo)) {
396                 ret = xpc_pull_remote_cachelines(part,
397                                                  part->remote_openclose_args,
398                                                  (void *)part->
399                                                  remote_openclose_args_pa,
400                                                  XPC_OPENCLOSE_ARGS_SIZE);
401                 if (ret != xpSuccess) {
402                         XPC_DEACTIVATE_PARTITION(part, ret);
403
404                         dev_dbg(xpc_chan, "failed to pull openclose args from "
405                                 "partition %d, ret=%d\n", XPC_PARTID(part),
406                                 ret);
407
408                         /* don't bother processing IPIs anymore */
409                         IPI_amo = 0;
410                 }
411         }
412
413         if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_amo)) {
414                 ret = xpc_pull_remote_cachelines(part, part->remote_GPs,
415                                                  (void *)part->remote_GPs_pa,
416                                                  XPC_GP_SIZE);
417                 if (ret != xpSuccess) {
418                         XPC_DEACTIVATE_PARTITION(part, ret);
419
420                         dev_dbg(xpc_chan, "failed to pull GPs from partition "
421                                 "%d, ret=%d\n", XPC_PARTID(part), ret);
422
423                         /* don't bother processing IPIs anymore */
424                         IPI_amo = 0;
425                 }
426         }
427
428         return IPI_amo;
429 }
430
431 /*
432  * Allocate the local message queue and the notify queue.
433  */
434 static enum xp_retval
435 xpc_allocate_local_msgqueue(struct xpc_channel *ch)
436 {
437         unsigned long irq_flags;
438         int nentries;
439         size_t nbytes;
440
441         for (nentries = ch->local_nentries; nentries > 0; nentries--) {
442
443                 nbytes = nentries * ch->msg_size;
444                 ch->local_msgqueue = xpc_kzalloc_cacheline_aligned(nbytes,
445                                                                    GFP_KERNEL,
446                                                       &ch->local_msgqueue_base);
447                 if (ch->local_msgqueue == NULL)
448                         continue;
449
450                 nbytes = nentries * sizeof(struct xpc_notify);
451                 ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);
452                 if (ch->notify_queue == NULL) {
453                         kfree(ch->local_msgqueue_base);
454                         ch->local_msgqueue = NULL;
455                         continue;
456                 }
457
458                 spin_lock_irqsave(&ch->lock, irq_flags);
459                 if (nentries < ch->local_nentries) {
460                         dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
461                                 "partid=%d, channel=%d\n", nentries,
462                                 ch->local_nentries, ch->partid, ch->number);
463
464                         ch->local_nentries = nentries;
465                 }
466                 spin_unlock_irqrestore(&ch->lock, irq_flags);
467                 return xpSuccess;
468         }
469
470         dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
471                 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
472         return xpNoMemory;
473 }
474
475 /*
476  * Allocate the cached remote message queue.
477  */
478 static enum xp_retval
479 xpc_allocate_remote_msgqueue(struct xpc_channel *ch)
480 {
481         unsigned long irq_flags;
482         int nentries;
483         size_t nbytes;
484
485         DBUG_ON(ch->remote_nentries <= 0);
486
487         for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
488
489                 nbytes = nentries * ch->msg_size;
490                 ch->remote_msgqueue = xpc_kzalloc_cacheline_aligned(nbytes,
491                                                                     GFP_KERNEL,
492                                                      &ch->remote_msgqueue_base);
493                 if (ch->remote_msgqueue == NULL)
494                         continue;
495
496                 spin_lock_irqsave(&ch->lock, irq_flags);
497                 if (nentries < ch->remote_nentries) {
498                         dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
499                                 "partid=%d, channel=%d\n", nentries,
500                                 ch->remote_nentries, ch->partid, ch->number);
501
502                         ch->remote_nentries = nentries;
503                 }
504                 spin_unlock_irqrestore(&ch->lock, irq_flags);
505                 return xpSuccess;
506         }
507
508         dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
509                 "partid=%d, channel=%d\n", ch->partid, ch->number);
510         return xpNoMemory;
511 }
512
513 /*
514  * Allocate message queues and other stuff associated with a channel.
515  *
516  * Note: Assumes all of the channel sizes are filled in.
517  */
518 static enum xp_retval
519 xpc_allocate_msgqueues(struct xpc_channel *ch)
520 {
521         unsigned long irq_flags;
522         enum xp_retval ret;
523
524         DBUG_ON(ch->flags & XPC_C_SETUP);
525
526         ret = xpc_allocate_local_msgqueue(ch);
527         if (ret != xpSuccess)
528                 return ret;
529
530         ret = xpc_allocate_remote_msgqueue(ch);
531         if (ret != xpSuccess) {
532                 kfree(ch->local_msgqueue_base);
533                 ch->local_msgqueue = NULL;
534                 kfree(ch->notify_queue);
535                 ch->notify_queue = NULL;
536                 return ret;
537         }
538
539         spin_lock_irqsave(&ch->lock, irq_flags);
540         ch->flags |= XPC_C_SETUP;
541         spin_unlock_irqrestore(&ch->lock, irq_flags);
542
543         return xpSuccess;
544 }
545
546 /*
547  * Process a connect message from a remote partition.
548  *
549  * Note: xpc_process_connect() is expecting to be called with the
550  * spin_lock_irqsave held and will leave it locked upon return.
551  */
552 static void
553 xpc_process_connect(struct xpc_channel *ch, unsigned long *irq_flags)
554 {
555         enum xp_retval ret;
556
557         DBUG_ON(!spin_is_locked(&ch->lock));
558
559         if (!(ch->flags & XPC_C_OPENREQUEST) ||
560             !(ch->flags & XPC_C_ROPENREQUEST)) {
561                 /* nothing more to do for now */
562                 return;
563         }
564         DBUG_ON(!(ch->flags & XPC_C_CONNECTING));
565
566         if (!(ch->flags & XPC_C_SETUP)) {
567                 spin_unlock_irqrestore(&ch->lock, *irq_flags);
568                 ret = xpc_allocate_msgqueues(ch);
569                 spin_lock_irqsave(&ch->lock, *irq_flags);
570
571                 if (ret != xpSuccess)
572                         XPC_DISCONNECT_CHANNEL(ch, ret, irq_flags);
573
574                 if (ch->flags & (XPC_C_CONNECTED | XPC_C_DISCONNECTING))
575                         return;
576
577                 DBUG_ON(!(ch->flags & XPC_C_SETUP));
578                 DBUG_ON(ch->local_msgqueue == NULL);
579                 DBUG_ON(ch->remote_msgqueue == NULL);
580         }
581
582         if (!(ch->flags & XPC_C_OPENREPLY)) {
583                 ch->flags |= XPC_C_OPENREPLY;
584                 xpc_IPI_send_openreply(ch, irq_flags);
585         }
586
587         if (!(ch->flags & XPC_C_ROPENREPLY))
588                 return;
589
590         DBUG_ON(ch->remote_msgqueue_pa == 0);
591
592         ch->flags = (XPC_C_CONNECTED | XPC_C_SETUP);    /* clear all else */
593
594         dev_info(xpc_chan, "channel %d to partition %d connected\n",
595                  ch->number, ch->partid);
596
597         spin_unlock_irqrestore(&ch->lock, *irq_flags);
598         xpc_create_kthreads(ch, 1, 0);
599         spin_lock_irqsave(&ch->lock, *irq_flags);
600 }
601
602 /*
603  * Notify those who wanted to be notified upon delivery of their message.
604  */
605 static void
606 xpc_notify_senders(struct xpc_channel *ch, enum xp_retval reason, s64 put)
607 {
608         struct xpc_notify *notify;
609         u8 notify_type;
610         s64 get = ch->w_remote_GP.get - 1;
611
612         while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
613
614                 notify = &ch->notify_queue[get % ch->local_nentries];
615
616                 /*
617                  * See if the notify entry indicates it was associated with
618                  * a message who's sender wants to be notified. It is possible
619                  * that it is, but someone else is doing or has done the
620                  * notification.
621                  */
622                 notify_type = notify->type;
623                 if (notify_type == 0 ||
624                     cmpxchg(&notify->type, notify_type, 0) != notify_type) {
625                         continue;
626                 }
627
628                 DBUG_ON(notify_type != XPC_N_CALL);
629
630                 atomic_dec(&ch->n_to_notify);
631
632                 if (notify->func != NULL) {
633                         dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
634                                 "msg_number=%ld, partid=%d, channel=%d\n",
635                                 (void *)notify, get, ch->partid, ch->number);
636
637                         notify->func(reason, ch->partid, ch->number,
638                                      notify->key);
639
640                         dev_dbg(xpc_chan, "notify->func() returned, "
641                                 "notify=0x%p, msg_number=%ld, partid=%d, "
642                                 "channel=%d\n", (void *)notify, get,
643                                 ch->partid, ch->number);
644                 }
645         }
646 }
647
648 /*
649  * Free up message queues and other stuff that were allocated for the specified
650  * channel.
651  *
652  * Note: ch->reason and ch->reason_line are left set for debugging purposes,
653  * they're cleared when XPC_C_DISCONNECTED is cleared.
654  */
655 static void
656 xpc_free_msgqueues(struct xpc_channel *ch)
657 {
658         DBUG_ON(!spin_is_locked(&ch->lock));
659         DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
660
661         ch->remote_msgqueue_pa = 0;
662         ch->func = NULL;
663         ch->key = NULL;
664         ch->msg_size = 0;
665         ch->local_nentries = 0;
666         ch->remote_nentries = 0;
667         ch->kthreads_assigned_limit = 0;
668         ch->kthreads_idle_limit = 0;
669
670         ch->local_GP->get = 0;
671         ch->local_GP->put = 0;
672         ch->remote_GP.get = 0;
673         ch->remote_GP.put = 0;
674         ch->w_local_GP.get = 0;
675         ch->w_local_GP.put = 0;
676         ch->w_remote_GP.get = 0;
677         ch->w_remote_GP.put = 0;
678         ch->next_msg_to_pull = 0;
679
680         if (ch->flags & XPC_C_SETUP) {
681                 ch->flags &= ~XPC_C_SETUP;
682
683                 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
684                         ch->flags, ch->partid, ch->number);
685
686                 kfree(ch->local_msgqueue_base);
687                 ch->local_msgqueue = NULL;
688                 kfree(ch->remote_msgqueue_base);
689                 ch->remote_msgqueue = NULL;
690                 kfree(ch->notify_queue);
691                 ch->notify_queue = NULL;
692         }
693 }
694
695 /*
696  * spin_lock_irqsave() is expected to be held on entry.
697  */
698 static void
699 xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
700 {
701         struct xpc_partition *part = &xpc_partitions[ch->partid];
702         u32 channel_was_connected = (ch->flags & XPC_C_WASCONNECTED);
703
704         DBUG_ON(!spin_is_locked(&ch->lock));
705
706         if (!(ch->flags & XPC_C_DISCONNECTING))
707                 return;
708
709         DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
710
711         /* make sure all activity has settled down first */
712
713         if (atomic_read(&ch->kthreads_assigned) > 0 ||
714             atomic_read(&ch->references) > 0) {
715                 return;
716         }
717         DBUG_ON((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
718                 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT_MADE));
719
720         if (part->act_state == XPC_P_DEACTIVATING) {
721                 /* can't proceed until the other side disengages from us */
722                 if (xpc_partition_engaged(1UL << ch->partid))
723                         return;
724
725         } else {
726
727                 /* as long as the other side is up do the full protocol */
728
729                 if (!(ch->flags & XPC_C_RCLOSEREQUEST))
730                         return;
731
732                 if (!(ch->flags & XPC_C_CLOSEREPLY)) {
733                         ch->flags |= XPC_C_CLOSEREPLY;
734                         xpc_IPI_send_closereply(ch, irq_flags);
735                 }
736
737                 if (!(ch->flags & XPC_C_RCLOSEREPLY))
738                         return;
739         }
740
741         /* wake those waiting for notify completion */
742         if (atomic_read(&ch->n_to_notify) > 0) {
743                 /* >>> we do callout while holding ch->lock */
744                 xpc_notify_senders(ch, ch->reason, ch->w_local_GP.put);
745         }
746
747         /* both sides are disconnected now */
748
749         if (ch->flags & XPC_C_DISCONNECTINGCALLOUT_MADE) {
750                 spin_unlock_irqrestore(&ch->lock, *irq_flags);
751                 xpc_disconnect_callout(ch, xpDisconnected);
752                 spin_lock_irqsave(&ch->lock, *irq_flags);
753         }
754
755         /* it's now safe to free the channel's message queues */
756         xpc_free_msgqueues(ch);
757
758         /* mark disconnected, clear all other flags except XPC_C_WDISCONNECT */
759         ch->flags = (XPC_C_DISCONNECTED | (ch->flags & XPC_C_WDISCONNECT));
760
761         atomic_dec(&part->nchannels_active);
762
763         if (channel_was_connected) {
764                 dev_info(xpc_chan, "channel %d to partition %d disconnected, "
765                          "reason=%d\n", ch->number, ch->partid, ch->reason);
766         }
767
768         if (ch->flags & XPC_C_WDISCONNECT) {
769                 /* we won't lose the CPU since we're holding ch->lock */
770                 complete(&ch->wdisconnect_wait);
771         } else if (ch->delayed_IPI_flags) {
772                 if (part->act_state != XPC_P_DEACTIVATING) {
773                         /* time to take action on any delayed IPI flags */
774                         spin_lock(&part->IPI_lock);
775                         XPC_SET_IPI_FLAGS(part->local_IPI_amo, ch->number,
776                                           ch->delayed_IPI_flags);
777                         spin_unlock(&part->IPI_lock);
778                 }
779                 ch->delayed_IPI_flags = 0;
780         }
781 }
782
783 /*
784  * Process a change in the channel's remote connection state.
785  */
786 static void
787 xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
788                           u8 IPI_flags)
789 {
790         unsigned long irq_flags;
791         struct xpc_openclose_args *args =
792             &part->remote_openclose_args[ch_number];
793         struct xpc_channel *ch = &part->channels[ch_number];
794         enum xp_retval reason;
795
796         spin_lock_irqsave(&ch->lock, irq_flags);
797
798 again:
799
800         if ((ch->flags & XPC_C_DISCONNECTED) &&
801             (ch->flags & XPC_C_WDISCONNECT)) {
802                 /*
803                  * Delay processing IPI flags until thread waiting disconnect
804                  * has had a chance to see that the channel is disconnected.
805                  */
806                 ch->delayed_IPI_flags |= IPI_flags;
807                 spin_unlock_irqrestore(&ch->lock, irq_flags);
808                 return;
809         }
810
811         if (IPI_flags & XPC_IPI_CLOSEREQUEST) {
812
813                 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREQUEST (reason=%d) received "
814                         "from partid=%d, channel=%d\n", args->reason,
815                         ch->partid, ch->number);
816
817                 /*
818                  * If RCLOSEREQUEST is set, we're probably waiting for
819                  * RCLOSEREPLY. We should find it and a ROPENREQUEST packed
820                  * with this RCLOSEREQUEST in the IPI_flags.
821                  */
822
823                 if (ch->flags & XPC_C_RCLOSEREQUEST) {
824                         DBUG_ON(!(ch->flags & XPC_C_DISCONNECTING));
825                         DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
826                         DBUG_ON(!(ch->flags & XPC_C_CLOSEREPLY));
827                         DBUG_ON(ch->flags & XPC_C_RCLOSEREPLY);
828
829                         DBUG_ON(!(IPI_flags & XPC_IPI_CLOSEREPLY));
830                         IPI_flags &= ~XPC_IPI_CLOSEREPLY;
831                         ch->flags |= XPC_C_RCLOSEREPLY;
832
833                         /* both sides have finished disconnecting */
834                         xpc_process_disconnect(ch, &irq_flags);
835                         DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
836                         goto again;
837                 }
838
839                 if (ch->flags & XPC_C_DISCONNECTED) {
840                         if (!(IPI_flags & XPC_IPI_OPENREQUEST)) {
841                                 if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo,
842                                                        ch_number) &
843                                      XPC_IPI_OPENREQUEST)) {
844
845                                         DBUG_ON(ch->delayed_IPI_flags != 0);
846                                         spin_lock(&part->IPI_lock);
847                                         XPC_SET_IPI_FLAGS(part->local_IPI_amo,
848                                                           ch_number,
849                                                           XPC_IPI_CLOSEREQUEST);
850                                         spin_unlock(&part->IPI_lock);
851                                 }
852                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
853                                 return;
854                         }
855
856                         XPC_SET_REASON(ch, 0, 0);
857                         ch->flags &= ~XPC_C_DISCONNECTED;
858
859                         atomic_inc(&part->nchannels_active);
860                         ch->flags |= (XPC_C_CONNECTING | XPC_C_ROPENREQUEST);
861                 }
862
863                 IPI_flags &= ~(XPC_IPI_OPENREQUEST | XPC_IPI_OPENREPLY);
864
865                 /*
866                  * The meaningful CLOSEREQUEST connection state fields are:
867                  *      reason = reason connection is to be closed
868                  */
869
870                 ch->flags |= XPC_C_RCLOSEREQUEST;
871
872                 if (!(ch->flags & XPC_C_DISCONNECTING)) {
873                         reason = args->reason;
874                         if (reason <= xpSuccess || reason > xpUnknownReason)
875                                 reason = xpUnknownReason;
876                         else if (reason == xpUnregistering)
877                                 reason = xpOtherUnregistering;
878
879                         XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
880
881                         DBUG_ON(IPI_flags & XPC_IPI_CLOSEREPLY);
882                         spin_unlock_irqrestore(&ch->lock, irq_flags);
883                         return;
884                 }
885
886                 xpc_process_disconnect(ch, &irq_flags);
887         }
888
889         if (IPI_flags & XPC_IPI_CLOSEREPLY) {
890
891                 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREPLY received from partid=%d,"
892                         " channel=%d\n", ch->partid, ch->number);
893
894                 if (ch->flags & XPC_C_DISCONNECTED) {
895                         DBUG_ON(part->act_state != XPC_P_DEACTIVATING);
896                         spin_unlock_irqrestore(&ch->lock, irq_flags);
897                         return;
898                 }
899
900                 DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
901
902                 if (!(ch->flags & XPC_C_RCLOSEREQUEST)) {
903                         if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo, ch_number)
904                              & XPC_IPI_CLOSEREQUEST)) {
905
906                                 DBUG_ON(ch->delayed_IPI_flags != 0);
907                                 spin_lock(&part->IPI_lock);
908                                 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
909                                                   ch_number,
910                                                   XPC_IPI_CLOSEREPLY);
911                                 spin_unlock(&part->IPI_lock);
912                         }
913                         spin_unlock_irqrestore(&ch->lock, irq_flags);
914                         return;
915                 }
916
917                 ch->flags |= XPC_C_RCLOSEREPLY;
918
919                 if (ch->flags & XPC_C_CLOSEREPLY) {
920                         /* both sides have finished disconnecting */
921                         xpc_process_disconnect(ch, &irq_flags);
922                 }
923         }
924
925         if (IPI_flags & XPC_IPI_OPENREQUEST) {
926
927                 dev_dbg(xpc_chan, "XPC_IPI_OPENREQUEST (msg_size=%d, "
928                         "local_nentries=%d) received from partid=%d, "
929                         "channel=%d\n", args->msg_size, args->local_nentries,
930                         ch->partid, ch->number);
931
932                 if (part->act_state == XPC_P_DEACTIVATING ||
933                     (ch->flags & XPC_C_ROPENREQUEST)) {
934                         spin_unlock_irqrestore(&ch->lock, irq_flags);
935                         return;
936                 }
937
938                 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_WDISCONNECT)) {
939                         ch->delayed_IPI_flags |= XPC_IPI_OPENREQUEST;
940                         spin_unlock_irqrestore(&ch->lock, irq_flags);
941                         return;
942                 }
943                 DBUG_ON(!(ch->flags & (XPC_C_DISCONNECTED |
944                                        XPC_C_OPENREQUEST)));
945                 DBUG_ON(ch->flags & (XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
946                                      XPC_C_OPENREPLY | XPC_C_CONNECTED));
947
948                 /*
949                  * The meaningful OPENREQUEST connection state fields are:
950                  *      msg_size = size of channel's messages in bytes
951                  *      local_nentries = remote partition's local_nentries
952                  */
953                 if (args->msg_size == 0 || args->local_nentries == 0) {
954                         /* assume OPENREQUEST was delayed by mistake */
955                         spin_unlock_irqrestore(&ch->lock, irq_flags);
956                         return;
957                 }
958
959                 ch->flags |= (XPC_C_ROPENREQUEST | XPC_C_CONNECTING);
960                 ch->remote_nentries = args->local_nentries;
961
962                 if (ch->flags & XPC_C_OPENREQUEST) {
963                         if (args->msg_size != ch->msg_size) {
964                                 XPC_DISCONNECT_CHANNEL(ch, xpUnequalMsgSizes,
965                                                        &irq_flags);
966                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
967                                 return;
968                         }
969                 } else {
970                         ch->msg_size = args->msg_size;
971
972                         XPC_SET_REASON(ch, 0, 0);
973                         ch->flags &= ~XPC_C_DISCONNECTED;
974
975                         atomic_inc(&part->nchannels_active);
976                 }
977
978                 xpc_process_connect(ch, &irq_flags);
979         }
980
981         if (IPI_flags & XPC_IPI_OPENREPLY) {
982
983                 dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY (local_msgqueue_pa=0x%lx, "
984                         "local_nentries=%d, remote_nentries=%d) received from "
985                         "partid=%d, channel=%d\n", args->local_msgqueue_pa,
986                         args->local_nentries, args->remote_nentries,
987                         ch->partid, ch->number);
988
989                 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED)) {
990                         spin_unlock_irqrestore(&ch->lock, irq_flags);
991                         return;
992                 }
993                 if (!(ch->flags & XPC_C_OPENREQUEST)) {
994                         XPC_DISCONNECT_CHANNEL(ch, xpOpenCloseError,
995                                                &irq_flags);
996                         spin_unlock_irqrestore(&ch->lock, irq_flags);
997                         return;
998                 }
999
1000                 DBUG_ON(!(ch->flags & XPC_C_ROPENREQUEST));
1001                 DBUG_ON(ch->flags & XPC_C_CONNECTED);
1002
1003                 /*
1004                  * The meaningful OPENREPLY connection state fields are:
1005                  *      local_msgqueue_pa = physical address of remote
1006                  *                          partition's local_msgqueue
1007                  *      local_nentries = remote partition's local_nentries
1008                  *      remote_nentries = remote partition's remote_nentries
1009                  */
1010                 DBUG_ON(args->local_msgqueue_pa == 0);
1011                 DBUG_ON(args->local_nentries == 0);
1012                 DBUG_ON(args->remote_nentries == 0);
1013
1014                 ch->flags |= XPC_C_ROPENREPLY;
1015                 ch->remote_msgqueue_pa = args->local_msgqueue_pa;
1016
1017                 if (args->local_nentries < ch->remote_nentries) {
1018                         dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1019                                 "remote_nentries=%d, old remote_nentries=%d, "
1020                                 "partid=%d, channel=%d\n",
1021                                 args->local_nentries, ch->remote_nentries,
1022                                 ch->partid, ch->number);
1023
1024                         ch->remote_nentries = args->local_nentries;
1025                 }
1026                 if (args->remote_nentries < ch->local_nentries) {
1027                         dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1028                                 "local_nentries=%d, old local_nentries=%d, "
1029                                 "partid=%d, channel=%d\n",
1030                                 args->remote_nentries, ch->local_nentries,
1031                                 ch->partid, ch->number);
1032
1033                         ch->local_nentries = args->remote_nentries;
1034                 }
1035
1036                 xpc_process_connect(ch, &irq_flags);
1037         }
1038
1039         spin_unlock_irqrestore(&ch->lock, irq_flags);
1040 }
1041
1042 /*
1043  * Attempt to establish a channel connection to a remote partition.
1044  */
1045 static enum xp_retval
1046 xpc_connect_channel(struct xpc_channel *ch)
1047 {
1048         unsigned long irq_flags;
1049         struct xpc_registration *registration = &xpc_registrations[ch->number];
1050
1051         if (mutex_trylock(&registration->mutex) == 0)
1052                 return xpRetry;
1053
1054         if (!XPC_CHANNEL_REGISTERED(ch->number)) {
1055                 mutex_unlock(&registration->mutex);
1056                 return xpUnregistered;
1057         }
1058
1059         spin_lock_irqsave(&ch->lock, irq_flags);
1060
1061         DBUG_ON(ch->flags & XPC_C_CONNECTED);
1062         DBUG_ON(ch->flags & XPC_C_OPENREQUEST);
1063
1064         if (ch->flags & XPC_C_DISCONNECTING) {
1065                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1066                 mutex_unlock(&registration->mutex);
1067                 return ch->reason;
1068         }
1069
1070         /* add info from the channel connect registration to the channel */
1071
1072         ch->kthreads_assigned_limit = registration->assigned_limit;
1073         ch->kthreads_idle_limit = registration->idle_limit;
1074         DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
1075         DBUG_ON(atomic_read(&ch->kthreads_idle) != 0);
1076         DBUG_ON(atomic_read(&ch->kthreads_active) != 0);
1077
1078         ch->func = registration->func;
1079         DBUG_ON(registration->func == NULL);
1080         ch->key = registration->key;
1081
1082         ch->local_nentries = registration->nentries;
1083
1084         if (ch->flags & XPC_C_ROPENREQUEST) {
1085                 if (registration->msg_size != ch->msg_size) {
1086                         /* the local and remote sides aren't the same */
1087
1088                         /*
1089                          * Because XPC_DISCONNECT_CHANNEL() can block we're
1090                          * forced to up the registration sema before we unlock
1091                          * the channel lock. But that's okay here because we're
1092                          * done with the part that required the registration
1093                          * sema. XPC_DISCONNECT_CHANNEL() requires that the
1094                          * channel lock be locked and will unlock and relock
1095                          * the channel lock as needed.
1096                          */
1097                         mutex_unlock(&registration->mutex);
1098                         XPC_DISCONNECT_CHANNEL(ch, xpUnequalMsgSizes,
1099                                                &irq_flags);
1100                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1101                         return xpUnequalMsgSizes;
1102                 }
1103         } else {
1104                 ch->msg_size = registration->msg_size;
1105
1106                 XPC_SET_REASON(ch, 0, 0);
1107                 ch->flags &= ~XPC_C_DISCONNECTED;
1108
1109                 atomic_inc(&xpc_partitions[ch->partid].nchannels_active);
1110         }
1111
1112         mutex_unlock(&registration->mutex);
1113
1114         /* initiate the connection */
1115
1116         ch->flags |= (XPC_C_OPENREQUEST | XPC_C_CONNECTING);
1117         xpc_IPI_send_openrequest(ch, &irq_flags);
1118
1119         xpc_process_connect(ch, &irq_flags);
1120
1121         spin_unlock_irqrestore(&ch->lock, irq_flags);
1122
1123         return xpSuccess;
1124 }
1125
1126 /*
1127  * Clear some of the msg flags in the local message queue.
1128  */
1129 static inline void
1130 xpc_clear_local_msgqueue_flags(struct xpc_channel *ch)
1131 {
1132         struct xpc_msg *msg;
1133         s64 get;
1134
1135         get = ch->w_remote_GP.get;
1136         do {
1137                 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1138                                          (get % ch->local_nentries) *
1139                                          ch->msg_size);
1140                 msg->flags = 0;
1141         } while (++get < ch->remote_GP.get);
1142 }
1143
1144 /*
1145  * Clear some of the msg flags in the remote message queue.
1146  */
1147 static inline void
1148 xpc_clear_remote_msgqueue_flags(struct xpc_channel *ch)
1149 {
1150         struct xpc_msg *msg;
1151         s64 put;
1152
1153         put = ch->w_remote_GP.put;
1154         do {
1155                 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
1156                                          (put % ch->remote_nentries) *
1157                                          ch->msg_size);
1158                 msg->flags = 0;
1159         } while (++put < ch->remote_GP.put);
1160 }
1161
1162 static void
1163 xpc_process_msg_IPI(struct xpc_partition *part, int ch_number)
1164 {
1165         struct xpc_channel *ch = &part->channels[ch_number];
1166         int nmsgs_sent;
1167
1168         ch->remote_GP = part->remote_GPs[ch_number];
1169
1170         /* See what, if anything, has changed for each connected channel */
1171
1172         xpc_msgqueue_ref(ch);
1173
1174         if (ch->w_remote_GP.get == ch->remote_GP.get &&
1175             ch->w_remote_GP.put == ch->remote_GP.put) {
1176                 /* nothing changed since GPs were last pulled */
1177                 xpc_msgqueue_deref(ch);
1178                 return;
1179         }
1180
1181         if (!(ch->flags & XPC_C_CONNECTED)) {
1182                 xpc_msgqueue_deref(ch);
1183                 return;
1184         }
1185
1186         /*
1187          * First check to see if messages recently sent by us have been
1188          * received by the other side. (The remote GET value will have
1189          * changed since we last looked at it.)
1190          */
1191
1192         if (ch->w_remote_GP.get != ch->remote_GP.get) {
1193
1194                 /*
1195                  * We need to notify any senders that want to be notified
1196                  * that their sent messages have been received by their
1197                  * intended recipients. We need to do this before updating
1198                  * w_remote_GP.get so that we don't allocate the same message
1199                  * queue entries prematurely (see xpc_allocate_msg()).
1200                  */
1201                 if (atomic_read(&ch->n_to_notify) > 0) {
1202                         /*
1203                          * Notify senders that messages sent have been
1204                          * received and delivered by the other side.
1205                          */
1206                         xpc_notify_senders(ch, xpMsgDelivered,
1207                                            ch->remote_GP.get);
1208                 }
1209
1210                 /*
1211                  * Clear msg->flags in previously sent messages, so that
1212                  * they're ready for xpc_allocate_msg().
1213                  */
1214                 xpc_clear_local_msgqueue_flags(ch);
1215
1216                 ch->w_remote_GP.get = ch->remote_GP.get;
1217
1218                 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1219                         "channel=%d\n", ch->w_remote_GP.get, ch->partid,
1220                         ch->number);
1221
1222                 /*
1223                  * If anyone was waiting for message queue entries to become
1224                  * available, wake them up.
1225                  */
1226                 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1227                         wake_up(&ch->msg_allocate_wq);
1228         }
1229
1230         /*
1231          * Now check for newly sent messages by the other side. (The remote
1232          * PUT value will have changed since we last looked at it.)
1233          */
1234
1235         if (ch->w_remote_GP.put != ch->remote_GP.put) {
1236                 /*
1237                  * Clear msg->flags in previously received messages, so that
1238                  * they're ready for xpc_get_deliverable_msg().
1239                  */
1240                 xpc_clear_remote_msgqueue_flags(ch);
1241
1242                 ch->w_remote_GP.put = ch->remote_GP.put;
1243
1244                 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1245                         "channel=%d\n", ch->w_remote_GP.put, ch->partid,
1246                         ch->number);
1247
1248                 nmsgs_sent = ch->w_remote_GP.put - ch->w_local_GP.get;
1249                 if (nmsgs_sent > 0) {
1250                         dev_dbg(xpc_chan, "msgs waiting to be copied and "
1251                                 "delivered=%d, partid=%d, channel=%d\n",
1252                                 nmsgs_sent, ch->partid, ch->number);
1253
1254                         if (ch->flags & XPC_C_CONNECTEDCALLOUT_MADE)
1255                                 xpc_activate_kthreads(ch, nmsgs_sent);
1256                 }
1257         }
1258
1259         xpc_msgqueue_deref(ch);
1260 }
1261
1262 void
1263 xpc_process_channel_activity(struct xpc_partition *part)
1264 {
1265         unsigned long irq_flags;
1266         u64 IPI_amo, IPI_flags;
1267         struct xpc_channel *ch;
1268         int ch_number;
1269         u32 ch_flags;
1270
1271         IPI_amo = xpc_get_IPI_flags(part);
1272
1273         /*
1274          * Initiate channel connections for registered channels.
1275          *
1276          * For each connected channel that has pending messages activate idle
1277          * kthreads and/or create new kthreads as needed.
1278          */
1279
1280         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1281                 ch = &part->channels[ch_number];
1282
1283                 /*
1284                  * Process any open or close related IPI flags, and then deal
1285                  * with connecting or disconnecting the channel as required.
1286                  */
1287
1288                 IPI_flags = XPC_GET_IPI_FLAGS(IPI_amo, ch_number);
1289
1290                 if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_flags))
1291                         xpc_process_openclose_IPI(part, ch_number, IPI_flags);
1292
1293                 ch_flags = ch->flags;   /* need an atomic snapshot of flags */
1294
1295                 if (ch_flags & XPC_C_DISCONNECTING) {
1296                         spin_lock_irqsave(&ch->lock, irq_flags);
1297                         xpc_process_disconnect(ch, &irq_flags);
1298                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1299                         continue;
1300                 }
1301
1302                 if (part->act_state == XPC_P_DEACTIVATING)
1303                         continue;
1304
1305                 if (!(ch_flags & XPC_C_CONNECTED)) {
1306                         if (!(ch_flags & XPC_C_OPENREQUEST)) {
1307                                 DBUG_ON(ch_flags & XPC_C_SETUP);
1308                                 (void)xpc_connect_channel(ch);
1309                         } else {
1310                                 spin_lock_irqsave(&ch->lock, irq_flags);
1311                                 xpc_process_connect(ch, &irq_flags);
1312                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1313                         }
1314                         continue;
1315                 }
1316
1317                 /*
1318                  * Process any message related IPI flags, this may involve the
1319                  * activation of kthreads to deliver any pending messages sent
1320                  * from the other partition.
1321                  */
1322
1323                 if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_flags))
1324                         xpc_process_msg_IPI(part, ch_number);
1325         }
1326 }
1327
1328 /*
1329  * XPC's heartbeat code calls this function to inform XPC that a partition is
1330  * going down.  XPC responds by tearing down the XPartition Communication
1331  * infrastructure used for the just downed partition.
1332  *
1333  * XPC's heartbeat code will never call this function and xpc_partition_up()
1334  * at the same time. Nor will it ever make multiple calls to either function
1335  * at the same time.
1336  */
1337 void
1338 xpc_partition_going_down(struct xpc_partition *part, enum xp_retval reason)
1339 {
1340         unsigned long irq_flags;
1341         int ch_number;
1342         struct xpc_channel *ch;
1343
1344         dev_dbg(xpc_chan, "deactivating partition %d, reason=%d\n",
1345                 XPC_PARTID(part), reason);
1346
1347         if (!xpc_part_ref(part)) {
1348                 /* infrastructure for this partition isn't currently set up */
1349                 return;
1350         }
1351
1352         /* disconnect channels associated with the partition going down */
1353
1354         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1355                 ch = &part->channels[ch_number];
1356
1357                 xpc_msgqueue_ref(ch);
1358                 spin_lock_irqsave(&ch->lock, irq_flags);
1359
1360                 XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
1361
1362                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1363                 xpc_msgqueue_deref(ch);
1364         }
1365
1366         xpc_wakeup_channel_mgr(part);
1367
1368         xpc_part_deref(part);
1369 }
1370
1371 /*
1372  * Teardown the infrastructure necessary to support XPartition Communication
1373  * between the specified remote partition and the local one.
1374  */
1375 void
1376 xpc_teardown_infrastructure(struct xpc_partition *part)
1377 {
1378         short partid = XPC_PARTID(part);
1379
1380         /*
1381          * We start off by making this partition inaccessible to local
1382          * processes by marking it as no longer setup. Then we make it
1383          * inaccessible to remote processes by clearing the XPC per partition
1384          * specific variable's magic # (which indicates that these variables
1385          * are no longer valid) and by ignoring all XPC notify IPIs sent to
1386          * this partition.
1387          */
1388
1389         DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1390         DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1391         DBUG_ON(part->setup_state != XPC_P_SETUP);
1392         part->setup_state = XPC_P_WTEARDOWN;
1393
1394         xpc_vars_part[partid].magic = 0;
1395
1396         free_irq(SGI_XPC_NOTIFY, (void *)(u64)partid);
1397
1398         /*
1399          * Before proceeding with the teardown we have to wait until all
1400          * existing references cease.
1401          */
1402         wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1403
1404         /* now we can begin tearing down the infrastructure */
1405
1406         part->setup_state = XPC_P_TORNDOWN;
1407
1408         /* in case we've still got outstanding timers registered... */
1409         del_timer_sync(&part->dropped_IPI_timer);
1410
1411         kfree(part->remote_openclose_args_base);
1412         part->remote_openclose_args = NULL;
1413         kfree(part->local_openclose_args_base);
1414         part->local_openclose_args = NULL;
1415         kfree(part->remote_GPs_base);
1416         part->remote_GPs = NULL;
1417         kfree(part->local_GPs_base);
1418         part->local_GPs = NULL;
1419         kfree(part->channels);
1420         part->channels = NULL;
1421         part->local_IPI_amo_va = NULL;
1422 }
1423
1424 /*
1425  * Called by XP at the time of channel connection registration to cause
1426  * XPC to establish connections to all currently active partitions.
1427  */
1428 void
1429 xpc_initiate_connect(int ch_number)
1430 {
1431         short partid;
1432         struct xpc_partition *part;
1433         struct xpc_channel *ch;
1434
1435         DBUG_ON(ch_number < 0 || ch_number >= XPC_MAX_NCHANNELS);
1436
1437         for (partid = 0; partid < xp_max_npartitions; partid++) {
1438                 part = &xpc_partitions[partid];
1439
1440                 if (xpc_part_ref(part)) {
1441                         ch = &part->channels[ch_number];
1442
1443                         /*
1444                          * Initiate the establishment of a connection on the
1445                          * newly registered channel to the remote partition.
1446                          */
1447                         xpc_wakeup_channel_mgr(part);
1448                         xpc_part_deref(part);
1449                 }
1450         }
1451 }
1452
1453 void
1454 xpc_connected_callout(struct xpc_channel *ch)
1455 {
1456         /* let the registerer know that a connection has been established */
1457
1458         if (ch->func != NULL) {
1459                 dev_dbg(xpc_chan, "ch->func() called, reason=xpConnected, "
1460                         "partid=%d, channel=%d\n", ch->partid, ch->number);
1461
1462                 ch->func(xpConnected, ch->partid, ch->number,
1463                          (void *)(u64)ch->local_nentries, ch->key);
1464
1465                 dev_dbg(xpc_chan, "ch->func() returned, reason=xpConnected, "
1466                         "partid=%d, channel=%d\n", ch->partid, ch->number);
1467         }
1468 }
1469
1470 /*
1471  * Called by XP at the time of channel connection unregistration to cause
1472  * XPC to teardown all current connections for the specified channel.
1473  *
1474  * Before returning xpc_initiate_disconnect() will wait until all connections
1475  * on the specified channel have been closed/torndown. So the caller can be
1476  * assured that they will not be receiving any more callouts from XPC to the
1477  * function they registered via xpc_connect().
1478  *
1479  * Arguments:
1480  *
1481  *      ch_number - channel # to unregister.
1482  */
1483 void
1484 xpc_initiate_disconnect(int ch_number)
1485 {
1486         unsigned long irq_flags;
1487         short partid;
1488         struct xpc_partition *part;
1489         struct xpc_channel *ch;
1490
1491         DBUG_ON(ch_number < 0 || ch_number >= XPC_MAX_NCHANNELS);
1492
1493         /* initiate the channel disconnect for every active partition */
1494         for (partid = 0; partid < xp_max_npartitions; partid++) {
1495                 part = &xpc_partitions[partid];
1496
1497                 if (xpc_part_ref(part)) {
1498                         ch = &part->channels[ch_number];
1499                         xpc_msgqueue_ref(ch);
1500
1501                         spin_lock_irqsave(&ch->lock, irq_flags);
1502
1503                         if (!(ch->flags & XPC_C_DISCONNECTED)) {
1504                                 ch->flags |= XPC_C_WDISCONNECT;
1505
1506                                 XPC_DISCONNECT_CHANNEL(ch, xpUnregistering,
1507                                                        &irq_flags);
1508                         }
1509
1510                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1511
1512                         xpc_msgqueue_deref(ch);
1513                         xpc_part_deref(part);
1514                 }
1515         }
1516
1517         xpc_disconnect_wait(ch_number);
1518 }
1519
1520 /*
1521  * To disconnect a channel, and reflect it back to all who may be waiting.
1522  *
1523  * An OPEN is not allowed until XPC_C_DISCONNECTING is cleared by
1524  * xpc_process_disconnect(), and if set, XPC_C_WDISCONNECT is cleared by
1525  * xpc_disconnect_wait().
1526  *
1527  * THE CHANNEL IS TO BE LOCKED BY THE CALLER AND WILL REMAIN LOCKED UPON RETURN.
1528  */
1529 void
1530 xpc_disconnect_channel(const int line, struct xpc_channel *ch,
1531                        enum xp_retval reason, unsigned long *irq_flags)
1532 {
1533         u32 channel_was_connected = (ch->flags & XPC_C_CONNECTED);
1534
1535         DBUG_ON(!spin_is_locked(&ch->lock));
1536
1537         if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED))
1538                 return;
1539
1540         DBUG_ON(!(ch->flags & (XPC_C_CONNECTING | XPC_C_CONNECTED)));
1541
1542         dev_dbg(xpc_chan, "reason=%d, line=%d, partid=%d, channel=%d\n",
1543                 reason, line, ch->partid, ch->number);
1544
1545         XPC_SET_REASON(ch, reason, line);
1546
1547         ch->flags |= (XPC_C_CLOSEREQUEST | XPC_C_DISCONNECTING);
1548         /* some of these may not have been set */
1549         ch->flags &= ~(XPC_C_OPENREQUEST | XPC_C_OPENREPLY |
1550                        XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
1551                        XPC_C_CONNECTING | XPC_C_CONNECTED);
1552
1553         xpc_IPI_send_closerequest(ch, irq_flags);
1554
1555         if (channel_was_connected)
1556                 ch->flags |= XPC_C_WASCONNECTED;
1557
1558         spin_unlock_irqrestore(&ch->lock, *irq_flags);
1559
1560         /* wake all idle kthreads so they can exit */
1561         if (atomic_read(&ch->kthreads_idle) > 0) {
1562                 wake_up_all(&ch->idle_wq);
1563
1564         } else if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
1565                    !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
1566                 /* start a kthread that will do the xpDisconnecting callout */
1567                 xpc_create_kthreads(ch, 1, 1);
1568         }
1569
1570         /* wake those waiting to allocate an entry from the local msg queue */
1571         if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1572                 wake_up(&ch->msg_allocate_wq);
1573
1574         spin_lock_irqsave(&ch->lock, *irq_flags);
1575 }
1576
1577 void
1578 xpc_disconnect_callout(struct xpc_channel *ch, enum xp_retval reason)
1579 {
1580         /*
1581          * Let the channel's registerer know that the channel is being
1582          * disconnected. We don't want to do this if the registerer was never
1583          * informed of a connection being made.
1584          */
1585
1586         if (ch->func != NULL) {
1587                 dev_dbg(xpc_chan, "ch->func() called, reason=%d, partid=%d, "
1588                         "channel=%d\n", reason, ch->partid, ch->number);
1589
1590                 ch->func(reason, ch->partid, ch->number, NULL, ch->key);
1591
1592                 dev_dbg(xpc_chan, "ch->func() returned, reason=%d, partid=%d, "
1593                         "channel=%d\n", reason, ch->partid, ch->number);
1594         }
1595 }
1596
1597 /*
1598  * Wait for a message entry to become available for the specified channel,
1599  * but don't wait any longer than 1 jiffy.
1600  */
1601 static enum xp_retval
1602 xpc_allocate_msg_wait(struct xpc_channel *ch)
1603 {
1604         enum xp_retval ret;
1605
1606         if (ch->flags & XPC_C_DISCONNECTING) {
1607                 DBUG_ON(ch->reason == xpInterrupted);
1608                 return ch->reason;
1609         }
1610
1611         atomic_inc(&ch->n_on_msg_allocate_wq);
1612         ret = interruptible_sleep_on_timeout(&ch->msg_allocate_wq, 1);
1613         atomic_dec(&ch->n_on_msg_allocate_wq);
1614
1615         if (ch->flags & XPC_C_DISCONNECTING) {
1616                 ret = ch->reason;
1617                 DBUG_ON(ch->reason == xpInterrupted);
1618         } else if (ret == 0) {
1619                 ret = xpTimeout;
1620         } else {
1621                 ret = xpInterrupted;
1622         }
1623
1624         return ret;
1625 }
1626
1627 /*
1628  * Allocate an entry for a message from the message queue associated with the
1629  * specified channel.
1630  */
1631 static enum xp_retval
1632 xpc_allocate_msg(struct xpc_channel *ch, u32 flags,
1633                  struct xpc_msg **address_of_msg)
1634 {
1635         struct xpc_msg *msg;
1636         enum xp_retval ret;
1637         s64 put;
1638
1639         /* this reference will be dropped in xpc_send_msg() */
1640         xpc_msgqueue_ref(ch);
1641
1642         if (ch->flags & XPC_C_DISCONNECTING) {
1643                 xpc_msgqueue_deref(ch);
1644                 return ch->reason;
1645         }
1646         if (!(ch->flags & XPC_C_CONNECTED)) {
1647                 xpc_msgqueue_deref(ch);
1648                 return xpNotConnected;
1649         }
1650
1651         /*
1652          * Get the next available message entry from the local message queue.
1653          * If none are available, we'll make sure that we grab the latest
1654          * GP values.
1655          */
1656         ret = xpTimeout;
1657
1658         while (1) {
1659
1660                 put = ch->w_local_GP.put;
1661                 rmb();  /* guarantee that .put loads before .get */
1662                 if (put - ch->w_remote_GP.get < ch->local_nentries) {
1663
1664                         /* There are available message entries. We need to try
1665                          * to secure one for ourselves. We'll do this by trying
1666                          * to increment w_local_GP.put as long as someone else
1667                          * doesn't beat us to it. If they do, we'll have to
1668                          * try again.
1669                          */
1670                         if (cmpxchg(&ch->w_local_GP.put, put, put + 1) == put) {
1671                                 /* we got the entry referenced by put */
1672                                 break;
1673                         }
1674                         continue;       /* try again */
1675                 }
1676
1677                 /*
1678                  * There aren't any available msg entries at this time.
1679                  *
1680                  * In waiting for a message entry to become available,
1681                  * we set a timeout in case the other side is not
1682                  * sending completion IPIs. This lets us fake an IPI
1683                  * that will cause the IPI handler to fetch the latest
1684                  * GP values as if an IPI was sent by the other side.
1685                  */
1686                 if (ret == xpTimeout)
1687                         xpc_IPI_send_local_msgrequest(ch);
1688
1689                 if (flags & XPC_NOWAIT) {
1690                         xpc_msgqueue_deref(ch);
1691                         return xpNoWait;
1692                 }
1693
1694                 ret = xpc_allocate_msg_wait(ch);
1695                 if (ret != xpInterrupted && ret != xpTimeout) {
1696                         xpc_msgqueue_deref(ch);
1697                         return ret;
1698                 }
1699         }
1700
1701         /* get the message's address and initialize it */
1702         msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1703                                  (put % ch->local_nentries) * ch->msg_size);
1704
1705         DBUG_ON(msg->flags != 0);
1706         msg->number = put;
1707
1708         dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
1709                 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
1710                 (void *)msg, msg->number, ch->partid, ch->number);
1711
1712         *address_of_msg = msg;
1713
1714         return xpSuccess;
1715 }
1716
1717 /*
1718  * Allocate an entry for a message from the message queue associated with the
1719  * specified channel. NOTE that this routine can sleep waiting for a message
1720  * entry to become available. To not sleep, pass in the XPC_NOWAIT flag.
1721  *
1722  * Arguments:
1723  *
1724  *      partid - ID of partition to which the channel is connected.
1725  *      ch_number - channel #.
1726  *      flags - see xpc.h for valid flags.
1727  *      payload - address of the allocated payload area pointer (filled in on
1728  *                return) in which the user-defined message is constructed.
1729  */
1730 enum xp_retval
1731 xpc_initiate_allocate(short partid, int ch_number, u32 flags, void **payload)
1732 {
1733         struct xpc_partition *part = &xpc_partitions[partid];
1734         enum xp_retval ret = xpUnknownReason;
1735         struct xpc_msg *msg = NULL;
1736
1737         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
1738         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1739
1740         *payload = NULL;
1741
1742         if (xpc_part_ref(part)) {
1743                 ret = xpc_allocate_msg(&part->channels[ch_number], flags, &msg);
1744                 xpc_part_deref(part);
1745
1746                 if (msg != NULL)
1747                         *payload = &msg->payload;
1748         }
1749
1750         return ret;
1751 }
1752
1753 /*
1754  * Now we actually send the messages that are ready to be sent by advancing
1755  * the local message queue's Put value and then send an IPI to the recipient
1756  * partition.
1757  */
1758 static void
1759 xpc_send_msgs(struct xpc_channel *ch, s64 initial_put)
1760 {
1761         struct xpc_msg *msg;
1762         s64 put = initial_put + 1;
1763         int send_IPI = 0;
1764
1765         while (1) {
1766
1767                 while (1) {
1768                         if (put == ch->w_local_GP.put)
1769                                 break;
1770
1771                         msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1772                                                  (put % ch->local_nentries) *
1773                                                  ch->msg_size);
1774
1775                         if (!(msg->flags & XPC_M_READY))
1776                                 break;
1777
1778                         put++;
1779                 }
1780
1781                 if (put == initial_put) {
1782                         /* nothing's changed */
1783                         break;
1784                 }
1785
1786                 if (cmpxchg_rel(&ch->local_GP->put, initial_put, put) !=
1787                     initial_put) {
1788                         /* someone else beat us to it */
1789                         DBUG_ON(ch->local_GP->put < initial_put);
1790                         break;
1791                 }
1792
1793                 /* we just set the new value of local_GP->put */
1794
1795                 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
1796                         "channel=%d\n", put, ch->partid, ch->number);
1797
1798                 send_IPI = 1;
1799
1800                 /*
1801                  * We need to ensure that the message referenced by
1802                  * local_GP->put is not XPC_M_READY or that local_GP->put
1803                  * equals w_local_GP.put, so we'll go have a look.
1804                  */
1805                 initial_put = put;
1806         }
1807
1808         if (send_IPI)
1809                 xpc_IPI_send_msgrequest(ch);
1810 }
1811
1812 /*
1813  * Common code that does the actual sending of the message by advancing the
1814  * local message queue's Put value and sends an IPI to the partition the
1815  * message is being sent to.
1816  */
1817 static enum xp_retval
1818 xpc_send_msg(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type,
1819              xpc_notify_func func, void *key)
1820 {
1821         enum xp_retval ret = xpSuccess;
1822         struct xpc_notify *notify = notify;
1823         s64 put, msg_number = msg->number;
1824
1825         DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
1826         DBUG_ON((((u64)msg - (u64)ch->local_msgqueue) / ch->msg_size) !=
1827                 msg_number % ch->local_nentries);
1828         DBUG_ON(msg->flags & XPC_M_READY);
1829
1830         if (ch->flags & XPC_C_DISCONNECTING) {
1831                 /* drop the reference grabbed in xpc_allocate_msg() */
1832                 xpc_msgqueue_deref(ch);
1833                 return ch->reason;
1834         }
1835
1836         if (notify_type != 0) {
1837                 /*
1838                  * Tell the remote side to send an ACK interrupt when the
1839                  * message has been delivered.
1840                  */
1841                 msg->flags |= XPC_M_INTERRUPT;
1842
1843                 atomic_inc(&ch->n_to_notify);
1844
1845                 notify = &ch->notify_queue[msg_number % ch->local_nentries];
1846                 notify->func = func;
1847                 notify->key = key;
1848                 notify->type = notify_type;
1849
1850                 /* >>> is a mb() needed here? */
1851
1852                 if (ch->flags & XPC_C_DISCONNECTING) {
1853                         /*
1854                          * An error occurred between our last error check and
1855                          * this one. We will try to clear the type field from
1856                          * the notify entry. If we succeed then
1857                          * xpc_disconnect_channel() didn't already process
1858                          * the notify entry.
1859                          */
1860                         if (cmpxchg(&notify->type, notify_type, 0) ==
1861                             notify_type) {
1862                                 atomic_dec(&ch->n_to_notify);
1863                                 ret = ch->reason;
1864                         }
1865
1866                         /* drop the reference grabbed in xpc_allocate_msg() */
1867                         xpc_msgqueue_deref(ch);
1868                         return ret;
1869                 }
1870         }
1871
1872         msg->flags |= XPC_M_READY;
1873
1874         /*
1875          * The preceding store of msg->flags must occur before the following
1876          * load of ch->local_GP->put.
1877          */
1878         mb();
1879
1880         /* see if the message is next in line to be sent, if so send it */
1881
1882         put = ch->local_GP->put;
1883         if (put == msg_number)
1884                 xpc_send_msgs(ch, put);
1885
1886         /* drop the reference grabbed in xpc_allocate_msg() */
1887         xpc_msgqueue_deref(ch);
1888         return ret;
1889 }
1890
1891 /*
1892  * Send a message previously allocated using xpc_initiate_allocate() on the
1893  * specified channel connected to the specified partition.
1894  *
1895  * This routine will not wait for the message to be received, nor will
1896  * notification be given when it does happen. Once this routine has returned
1897  * the message entry allocated via xpc_initiate_allocate() is no longer
1898  * accessable to the caller.
1899  *
1900  * This routine, although called by users, does not call xpc_part_ref() to
1901  * ensure that the partition infrastructure is in place. It relies on the
1902  * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
1903  *
1904  * Arguments:
1905  *
1906  *      partid - ID of partition to which the channel is connected.
1907  *      ch_number - channel # to send message on.
1908  *      payload - pointer to the payload area allocated via
1909  *                      xpc_initiate_allocate().
1910  */
1911 enum xp_retval
1912 xpc_initiate_send(short partid, int ch_number, void *payload)
1913 {
1914         struct xpc_partition *part = &xpc_partitions[partid];
1915         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
1916         enum xp_retval ret;
1917
1918         dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *)msg,
1919                 partid, ch_number);
1920
1921         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
1922         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1923         DBUG_ON(msg == NULL);
1924
1925         ret = xpc_send_msg(&part->channels[ch_number], msg, 0, NULL, NULL);
1926
1927         return ret;
1928 }
1929
1930 /*
1931  * Send a message previously allocated using xpc_initiate_allocate on the
1932  * specified channel connected to the specified partition.
1933  *
1934  * This routine will not wait for the message to be sent. Once this routine
1935  * has returned the message entry allocated via xpc_initiate_allocate() is no
1936  * longer accessable to the caller.
1937  *
1938  * Once the remote end of the channel has received the message, the function
1939  * passed as an argument to xpc_initiate_send_notify() will be called. This
1940  * allows the sender to free up or re-use any buffers referenced by the
1941  * message, but does NOT mean the message has been processed at the remote
1942  * end by a receiver.
1943  *
1944  * If this routine returns an error, the caller's function will NOT be called.
1945  *
1946  * This routine, although called by users, does not call xpc_part_ref() to
1947  * ensure that the partition infrastructure is in place. It relies on the
1948  * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
1949  *
1950  * Arguments:
1951  *
1952  *      partid - ID of partition to which the channel is connected.
1953  *      ch_number - channel # to send message on.
1954  *      payload - pointer to the payload area allocated via
1955  *                      xpc_initiate_allocate().
1956  *      func - function to call with asynchronous notification of message
1957  *                receipt. THIS FUNCTION MUST BE NON-BLOCKING.
1958  *      key - user-defined key to be passed to the function when it's called.
1959  */
1960 enum xp_retval
1961 xpc_initiate_send_notify(short partid, int ch_number, void *payload,
1962                          xpc_notify_func func, void *key)
1963 {
1964         struct xpc_partition *part = &xpc_partitions[partid];
1965         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
1966         enum xp_retval ret;
1967
1968         dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *)msg,
1969                 partid, ch_number);
1970
1971         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
1972         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1973         DBUG_ON(msg == NULL);
1974         DBUG_ON(func == NULL);
1975
1976         ret = xpc_send_msg(&part->channels[ch_number], msg, XPC_N_CALL,
1977                            func, key);
1978         return ret;
1979 }
1980
1981 static struct xpc_msg *
1982 xpc_pull_remote_msg(struct xpc_channel *ch, s64 get)
1983 {
1984         struct xpc_partition *part = &xpc_partitions[ch->partid];
1985         struct xpc_msg *remote_msg, *msg;
1986         u32 msg_index, nmsgs;
1987         u64 msg_offset;
1988         enum xp_retval ret;
1989
1990         if (mutex_lock_interruptible(&ch->msg_to_pull_mutex) != 0) {
1991                 /* we were interrupted by a signal */
1992                 return NULL;
1993         }
1994
1995         while (get >= ch->next_msg_to_pull) {
1996
1997                 /* pull as many messages as are ready and able to be pulled */
1998
1999                 msg_index = ch->next_msg_to_pull % ch->remote_nentries;
2000
2001                 DBUG_ON(ch->next_msg_to_pull >= ch->w_remote_GP.put);
2002                 nmsgs = ch->w_remote_GP.put - ch->next_msg_to_pull;
2003                 if (msg_index + nmsgs > ch->remote_nentries) {
2004                         /* ignore the ones that wrap the msg queue for now */
2005                         nmsgs = ch->remote_nentries - msg_index;
2006                 }
2007
2008                 msg_offset = msg_index * ch->msg_size;
2009                 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2010                 remote_msg = (struct xpc_msg *)(ch->remote_msgqueue_pa +
2011                                                 msg_offset);
2012
2013                 ret = xpc_pull_remote_cachelines(part, msg, remote_msg,
2014                                                  nmsgs * ch->msg_size);
2015                 if (ret != xpSuccess) {
2016
2017                         dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
2018                                 " msg %ld from partition %d, channel=%d, "
2019                                 "ret=%d\n", nmsgs, ch->next_msg_to_pull,
2020                                 ch->partid, ch->number, ret);
2021
2022                         XPC_DEACTIVATE_PARTITION(part, ret);
2023
2024                         mutex_unlock(&ch->msg_to_pull_mutex);
2025                         return NULL;
2026                 }
2027
2028                 ch->next_msg_to_pull += nmsgs;
2029         }
2030
2031         mutex_unlock(&ch->msg_to_pull_mutex);
2032
2033         /* return the message we were looking for */
2034         msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2035         msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2036
2037         return msg;
2038 }
2039
2040 /*
2041  * Get a message to be delivered.
2042  */
2043 static struct xpc_msg *
2044 xpc_get_deliverable_msg(struct xpc_channel *ch)
2045 {
2046         struct xpc_msg *msg = NULL;
2047         s64 get;
2048
2049         do {
2050                 if (ch->flags & XPC_C_DISCONNECTING)
2051                         break;
2052
2053                 get = ch->w_local_GP.get;
2054                 rmb();  /* guarantee that .get loads before .put */
2055                 if (get == ch->w_remote_GP.put)
2056                         break;
2057
2058                 /* There are messages waiting to be pulled and delivered.
2059                  * We need to try to secure one for ourselves. We'll do this
2060                  * by trying to increment w_local_GP.get and hope that no one
2061                  * else beats us to it. If they do, we'll we'll simply have
2062                  * to try again for the next one.
2063                  */
2064
2065                 if (cmpxchg(&ch->w_local_GP.get, get, get + 1) == get) {
2066                         /* we got the entry referenced by get */
2067
2068                         dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2069                                 "partid=%d, channel=%d\n", get + 1,
2070                                 ch->partid, ch->number);
2071
2072                         /* pull the message from the remote partition */
2073
2074                         msg = xpc_pull_remote_msg(ch, get);
2075
2076                         DBUG_ON(msg != NULL && msg->number != get);
2077                         DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2078                         DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2079
2080                         break;
2081                 }
2082
2083         } while (1);
2084
2085         return msg;
2086 }
2087
2088 /*
2089  * Deliver a message to its intended recipient.
2090  */
2091 void
2092 xpc_deliver_msg(struct xpc_channel *ch)
2093 {
2094         struct xpc_msg *msg;
2095
2096         msg = xpc_get_deliverable_msg(ch);
2097         if (msg != NULL) {
2098
2099                 /*
2100                  * This ref is taken to protect the payload itself from being
2101                  * freed before the user is finished with it, which the user
2102                  * indicates by calling xpc_initiate_received().
2103                  */
2104                 xpc_msgqueue_ref(ch);
2105
2106                 atomic_inc(&ch->kthreads_active);
2107
2108                 if (ch->func != NULL) {
2109                         dev_dbg(xpc_chan, "ch->func() called, msg=0x%p, "
2110                                 "msg_number=%ld, partid=%d, channel=%d\n",
2111                                 (void *)msg, msg->number, ch->partid,
2112                                 ch->number);
2113
2114                         /* deliver the message to its intended recipient */
2115                         ch->func(xpMsgReceived, ch->partid, ch->number,
2116                                  &msg->payload, ch->key);
2117
2118                         dev_dbg(xpc_chan, "ch->func() returned, msg=0x%p, "
2119                                 "msg_number=%ld, partid=%d, channel=%d\n",
2120                                 (void *)msg, msg->number, ch->partid,
2121                                 ch->number);
2122                 }
2123
2124                 atomic_dec(&ch->kthreads_active);
2125         }
2126 }
2127
2128 /*
2129  * Now we actually acknowledge the messages that have been delivered and ack'd
2130  * by advancing the cached remote message queue's Get value and if requested
2131  * send an IPI to the message sender's partition.
2132  */
2133 static void
2134 xpc_acknowledge_msgs(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2135 {
2136         struct xpc_msg *msg;
2137         s64 get = initial_get + 1;
2138         int send_IPI = 0;
2139
2140         while (1) {
2141
2142                 while (1) {
2143                         if (get == ch->w_local_GP.get)
2144                                 break;
2145
2146                         msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
2147                                                  (get % ch->remote_nentries) *
2148                                                  ch->msg_size);
2149
2150                         if (!(msg->flags & XPC_M_DONE))
2151                                 break;
2152
2153                         msg_flags |= msg->flags;
2154                         get++;
2155                 }
2156
2157                 if (get == initial_get) {
2158                         /* nothing's changed */
2159                         break;
2160                 }
2161
2162                 if (cmpxchg_rel(&ch->local_GP->get, initial_get, get) !=
2163                     initial_get) {
2164                         /* someone else beat us to it */
2165                         DBUG_ON(ch->local_GP->get <= initial_get);
2166                         break;
2167                 }
2168
2169                 /* we just set the new value of local_GP->get */
2170
2171                 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2172                         "channel=%d\n", get, ch->partid, ch->number);
2173
2174                 send_IPI = (msg_flags & XPC_M_INTERRUPT);
2175
2176                 /*
2177                  * We need to ensure that the message referenced by
2178                  * local_GP->get is not XPC_M_DONE or that local_GP->get
2179                  * equals w_local_GP.get, so we'll go have a look.
2180                  */
2181                 initial_get = get;
2182         }
2183
2184         if (send_IPI)
2185                 xpc_IPI_send_msgrequest(ch);
2186 }
2187
2188 /*
2189  * Acknowledge receipt of a delivered message.
2190  *
2191  * If a message has XPC_M_INTERRUPT set, send an interrupt to the partition
2192  * that sent the message.
2193  *
2194  * This function, although called by users, does not call xpc_part_ref() to
2195  * ensure that the partition infrastructure is in place. It relies on the
2196  * fact that we called xpc_msgqueue_ref() in xpc_deliver_msg().
2197  *
2198  * Arguments:
2199  *
2200  *      partid - ID of partition to which the channel is connected.
2201  *      ch_number - channel # message received on.
2202  *      payload - pointer to the payload area allocated via
2203  *                      xpc_initiate_allocate().
2204  */
2205 void
2206 xpc_initiate_received(short partid, int ch_number, void *payload)
2207 {
2208         struct xpc_partition *part = &xpc_partitions[partid];
2209         struct xpc_channel *ch;
2210         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2211         s64 get, msg_number = msg->number;
2212
2213         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
2214         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2215
2216         ch = &part->channels[ch_number];
2217
2218         dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2219                 (void *)msg, msg_number, ch->partid, ch->number);
2220
2221         DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) !=
2222                 msg_number % ch->remote_nentries);
2223         DBUG_ON(msg->flags & XPC_M_DONE);
2224
2225         msg->flags |= XPC_M_DONE;
2226
2227         /*
2228          * The preceding store of msg->flags must occur before the following
2229          * load of ch->local_GP->get.
2230          */
2231         mb();
2232
2233         /*
2234          * See if this message is next in line to be acknowledged as having
2235          * been delivered.
2236          */
2237         get = ch->local_GP->get;
2238         if (get == msg_number)
2239                 xpc_acknowledge_msgs(ch, get, msg->flags);
2240
2241         /* the call to xpc_msgqueue_ref() was done by xpc_deliver_msg()  */
2242         xpc_msgqueue_deref(ch);
2243 }