Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6
[linux-2.6] / arch / ia64 / sn / kernel / 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) 2004-2005 Silicon Graphics, Inc.  All Rights Reserved.
7  */
8
9
10 /*
11  * Cross Partition Communication (XPC) channel support.
12  *
13  *      This is the part of XPC that manages the channels and
14  *      sends/receives messages across them to/from other partitions.
15  *
16  */
17
18
19 #include <linux/kernel.h>
20 #include <linux/init.h>
21 #include <linux/sched.h>
22 #include <linux/cache.h>
23 #include <linux/interrupt.h>
24 #include <linux/slab.h>
25 #include <asm/sn/bte.h>
26 #include <asm/sn/sn_sal.h>
27 #include "xpc.h"
28
29
30 /*
31  * Set up the initial values for the XPartition Communication channels.
32  */
33 static void
34 xpc_initialize_channels(struct xpc_partition *part, partid_t partid)
35 {
36         int ch_number;
37         struct xpc_channel *ch;
38
39
40         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
41                 ch = &part->channels[ch_number];
42
43                 ch->partid = partid;
44                 ch->number = ch_number;
45                 ch->flags = XPC_C_DISCONNECTED;
46
47                 ch->local_GP = &part->local_GPs[ch_number];
48                 ch->local_openclose_args =
49                                         &part->local_openclose_args[ch_number];
50
51                 atomic_set(&ch->kthreads_assigned, 0);
52                 atomic_set(&ch->kthreads_idle, 0);
53                 atomic_set(&ch->kthreads_active, 0);
54
55                 atomic_set(&ch->references, 0);
56                 atomic_set(&ch->n_to_notify, 0);
57
58                 spin_lock_init(&ch->lock);
59                 sema_init(&ch->msg_to_pull_sema, 1);    /* mutex */
60                 sema_init(&ch->wdisconnect_sema, 0);    /* event wait */
61
62                 atomic_set(&ch->n_on_msg_allocate_wq, 0);
63                 init_waitqueue_head(&ch->msg_allocate_wq);
64                 init_waitqueue_head(&ch->idle_wq);
65         }
66 }
67
68
69 /*
70  * Setup the infrastructure necessary to support XPartition Communication
71  * between the specified remote partition and the local one.
72  */
73 enum xpc_retval
74 xpc_setup_infrastructure(struct xpc_partition *part)
75 {
76         int ret, cpuid;
77         struct timer_list *timer;
78         partid_t partid = XPC_PARTID(part);
79
80
81         /*
82          * Zero out MOST of the entry for this partition. Only the fields
83          * starting with `nchannels' will be zeroed. The preceding fields must
84          * remain `viable' across partition ups and downs, since they may be
85          * referenced during this memset() operation.
86          */
87         memset(&part->nchannels, 0, sizeof(struct xpc_partition) -
88                                 offsetof(struct xpc_partition, nchannels));
89
90         /*
91          * Allocate all of the channel structures as a contiguous chunk of
92          * memory.
93          */
94         part->channels = kmalloc(sizeof(struct xpc_channel) * XPC_NCHANNELS,
95                                                                 GFP_KERNEL);
96         if (part->channels == NULL) {
97                 dev_err(xpc_chan, "can't get memory for channels\n");
98                 return xpcNoMemory;
99         }
100         memset(part->channels, 0, sizeof(struct xpc_channel) * XPC_NCHANNELS);
101
102         part->nchannels = XPC_NCHANNELS;
103
104
105         /* allocate all the required GET/PUT values */
106
107         part->local_GPs = xpc_kmalloc_cacheline_aligned(XPC_GP_SIZE,
108                                         GFP_KERNEL, &part->local_GPs_base);
109         if (part->local_GPs == NULL) {
110                 kfree(part->channels);
111                 part->channels = NULL;
112                 dev_err(xpc_chan, "can't get memory for local get/put "
113                         "values\n");
114                 return xpcNoMemory;
115         }
116         memset(part->local_GPs, 0, XPC_GP_SIZE);
117
118         part->remote_GPs = xpc_kmalloc_cacheline_aligned(XPC_GP_SIZE,
119                                         GFP_KERNEL, &part->remote_GPs_base);
120         if (part->remote_GPs == NULL) {
121                 kfree(part->channels);
122                 part->channels = NULL;
123                 kfree(part->local_GPs_base);
124                 part->local_GPs = NULL;
125                 dev_err(xpc_chan, "can't get memory for remote get/put "
126                         "values\n");
127                 return xpcNoMemory;
128         }
129         memset(part->remote_GPs, 0, XPC_GP_SIZE);
130
131
132         /* allocate all the required open and close args */
133
134         part->local_openclose_args = xpc_kmalloc_cacheline_aligned(
135                                         XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
136                                         &part->local_openclose_args_base);
137         if (part->local_openclose_args == NULL) {
138                 kfree(part->channels);
139                 part->channels = NULL;
140                 kfree(part->local_GPs_base);
141                 part->local_GPs = NULL;
142                 kfree(part->remote_GPs_base);
143                 part->remote_GPs = NULL;
144                 dev_err(xpc_chan, "can't get memory for local connect args\n");
145                 return xpcNoMemory;
146         }
147         memset(part->local_openclose_args, 0, XPC_OPENCLOSE_ARGS_SIZE);
148
149         part->remote_openclose_args = xpc_kmalloc_cacheline_aligned(
150                                         XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
151                                         &part->remote_openclose_args_base);
152         if (part->remote_openclose_args == NULL) {
153                 kfree(part->channels);
154                 part->channels = NULL;
155                 kfree(part->local_GPs_base);
156                 part->local_GPs = NULL;
157                 kfree(part->remote_GPs_base);
158                 part->remote_GPs = NULL;
159                 kfree(part->local_openclose_args_base);
160                 part->local_openclose_args = NULL;
161                 dev_err(xpc_chan, "can't get memory for remote connect args\n");
162                 return xpcNoMemory;
163         }
164         memset(part->remote_openclose_args, 0, XPC_OPENCLOSE_ARGS_SIZE);
165
166
167         xpc_initialize_channels(part, partid);
168
169         atomic_set(&part->nchannels_active, 0);
170         atomic_set(&part->nchannels_engaged, 0);
171
172
173         /* local_IPI_amo were set to 0 by an earlier memset() */
174
175         /* Initialize this partitions AMO_t structure */
176         part->local_IPI_amo_va = xpc_IPI_init(partid);
177
178         spin_lock_init(&part->IPI_lock);
179
180         atomic_set(&part->channel_mgr_requests, 1);
181         init_waitqueue_head(&part->channel_mgr_wq);
182
183         sprintf(part->IPI_owner, "xpc%02d", partid);
184         ret = request_irq(SGI_XPC_NOTIFY, xpc_notify_IRQ_handler, SA_SHIRQ,
185                                 part->IPI_owner, (void *) (u64) partid);
186         if (ret != 0) {
187                 kfree(part->channels);
188                 part->channels = NULL;
189                 kfree(part->local_GPs_base);
190                 part->local_GPs = NULL;
191                 kfree(part->remote_GPs_base);
192                 part->remote_GPs = NULL;
193                 kfree(part->local_openclose_args_base);
194                 part->local_openclose_args = NULL;
195                 kfree(part->remote_openclose_args_base);
196                 part->remote_openclose_args = NULL;
197                 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
198                         "errno=%d\n", -ret);
199                 return xpcLackOfResources;
200         }
201
202         /* Setup a timer to check for dropped IPIs */
203         timer = &part->dropped_IPI_timer;
204         init_timer(timer);
205         timer->function = (void (*)(unsigned long)) xpc_dropped_IPI_check;
206         timer->data = (unsigned long) part;
207         timer->expires = jiffies + XPC_P_DROPPED_IPI_WAIT;
208         add_timer(timer);
209
210         /*
211          * With the setting of the partition setup_state to XPC_P_SETUP, we're
212          * declaring that this partition is ready to go.
213          */
214         part->setup_state = XPC_P_SETUP;
215
216
217         /*
218          * Setup the per partition specific variables required by the
219          * remote partition to establish channel connections with us.
220          *
221          * The setting of the magic # indicates that these per partition
222          * specific variables are ready to be used.
223          */
224         xpc_vars_part[partid].GPs_pa = __pa(part->local_GPs);
225         xpc_vars_part[partid].openclose_args_pa =
226                                         __pa(part->local_openclose_args);
227         xpc_vars_part[partid].IPI_amo_pa = __pa(part->local_IPI_amo_va);
228         cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
229         xpc_vars_part[partid].IPI_nasid = cpuid_to_nasid(cpuid);
230         xpc_vars_part[partid].IPI_phys_cpuid = cpu_physical_id(cpuid);
231         xpc_vars_part[partid].nchannels = part->nchannels;
232         xpc_vars_part[partid].magic = XPC_VP_MAGIC1;
233
234         return xpcSuccess;
235 }
236
237
238 /*
239  * Create a wrapper that hides the underlying mechanism for pulling a cacheline
240  * (or multiple cachelines) from a remote partition.
241  *
242  * src must be a cacheline aligned physical address on the remote partition.
243  * dst must be a cacheline aligned virtual address on this partition.
244  * cnt must be an cacheline sized
245  */
246 static enum xpc_retval
247 xpc_pull_remote_cachelines(struct xpc_partition *part, void *dst,
248                                 const void *src, size_t cnt)
249 {
250         bte_result_t bte_ret;
251
252
253         DBUG_ON((u64) src != L1_CACHE_ALIGN((u64) src));
254         DBUG_ON((u64) dst != L1_CACHE_ALIGN((u64) dst));
255         DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
256
257         if (part->act_state == XPC_P_DEACTIVATING) {
258                 return part->reason;
259         }
260
261         bte_ret = xp_bte_copy((u64) src, (u64) ia64_tpa((u64) dst),
262                                 (u64) cnt, (BTE_NORMAL | BTE_WACQUIRE), NULL);
263         if (bte_ret == BTE_SUCCESS) {
264                 return xpcSuccess;
265         }
266
267         dev_dbg(xpc_chan, "xp_bte_copy() from partition %d failed, ret=%d\n",
268                 XPC_PARTID(part), bte_ret);
269
270         return xpc_map_bte_errors(bte_ret);
271 }
272
273
274 /*
275  * Pull the remote per partititon specific variables from the specified
276  * partition.
277  */
278 enum xpc_retval
279 xpc_pull_remote_vars_part(struct xpc_partition *part)
280 {
281         u8 buffer[L1_CACHE_BYTES * 2];
282         struct xpc_vars_part *pulled_entry_cacheline =
283                         (struct xpc_vars_part *) L1_CACHE_ALIGN((u64) buffer);
284         struct xpc_vars_part *pulled_entry;
285         u64 remote_entry_cacheline_pa, remote_entry_pa;
286         partid_t partid = XPC_PARTID(part);
287         enum xpc_retval ret;
288
289
290         /* pull the cacheline that contains the variables we're interested in */
291
292         DBUG_ON(part->remote_vars_part_pa !=
293                                 L1_CACHE_ALIGN(part->remote_vars_part_pa));
294         DBUG_ON(sizeof(struct xpc_vars_part) != L1_CACHE_BYTES / 2);
295
296         remote_entry_pa = part->remote_vars_part_pa +
297                         sn_partition_id * sizeof(struct xpc_vars_part);
298
299         remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
300
301         pulled_entry = (struct xpc_vars_part *) ((u64) pulled_entry_cacheline +
302                                 (remote_entry_pa & (L1_CACHE_BYTES - 1)));
303
304         ret = xpc_pull_remote_cachelines(part, pulled_entry_cacheline,
305                                         (void *) remote_entry_cacheline_pa,
306                                         L1_CACHE_BYTES);
307         if (ret != xpcSuccess) {
308                 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
309                         "partition %d, ret=%d\n", partid, ret);
310                 return ret;
311         }
312
313
314         /* see if they've been set up yet */
315
316         if (pulled_entry->magic != XPC_VP_MAGIC1 &&
317                                 pulled_entry->magic != XPC_VP_MAGIC2) {
318
319                 if (pulled_entry->magic != 0) {
320                         dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
321                                 "partition %d has bad magic value (=0x%lx)\n",
322                                 partid, sn_partition_id, pulled_entry->magic);
323                         return xpcBadMagic;
324                 }
325
326                 /* they've not been initialized yet */
327                 return xpcRetry;
328         }
329
330         if (xpc_vars_part[partid].magic == XPC_VP_MAGIC1) {
331
332                 /* validate the variables */
333
334                 if (pulled_entry->GPs_pa == 0 ||
335                                 pulled_entry->openclose_args_pa == 0 ||
336                                         pulled_entry->IPI_amo_pa == 0) {
337
338                         dev_err(xpc_chan, "partition %d's XPC vars_part for "
339                                 "partition %d are not valid\n", partid,
340                                 sn_partition_id);
341                         return xpcInvalidAddress;
342                 }
343
344                 /* the variables we imported look to be valid */
345
346                 part->remote_GPs_pa = pulled_entry->GPs_pa;
347                 part->remote_openclose_args_pa =
348                                         pulled_entry->openclose_args_pa;
349                 part->remote_IPI_amo_va =
350                                       (AMO_t *) __va(pulled_entry->IPI_amo_pa);
351                 part->remote_IPI_nasid = pulled_entry->IPI_nasid;
352                 part->remote_IPI_phys_cpuid = pulled_entry->IPI_phys_cpuid;
353
354                 if (part->nchannels > pulled_entry->nchannels) {
355                         part->nchannels = pulled_entry->nchannels;
356                 }
357
358                 /* let the other side know that we've pulled their variables */
359
360                 xpc_vars_part[partid].magic = XPC_VP_MAGIC2;
361         }
362
363         if (pulled_entry->magic == XPC_VP_MAGIC1) {
364                 return xpcRetry;
365         }
366
367         return xpcSuccess;
368 }
369
370
371 /*
372  * Get the IPI flags and pull the openclose args and/or remote GPs as needed.
373  */
374 static u64
375 xpc_get_IPI_flags(struct xpc_partition *part)
376 {
377         unsigned long irq_flags;
378         u64 IPI_amo;
379         enum xpc_retval ret;
380
381
382         /*
383          * See if there are any IPI flags to be handled.
384          */
385
386         spin_lock_irqsave(&part->IPI_lock, irq_flags);
387         if ((IPI_amo = part->local_IPI_amo) != 0) {
388                 part->local_IPI_amo = 0;
389         }
390         spin_unlock_irqrestore(&part->IPI_lock, irq_flags);
391
392
393         if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_amo)) {
394                 ret = xpc_pull_remote_cachelines(part,
395                                         part->remote_openclose_args,
396                                         (void *) part->remote_openclose_args_pa,
397                                         XPC_OPENCLOSE_ARGS_SIZE);
398                 if (ret != xpcSuccess) {
399                         XPC_DEACTIVATE_PARTITION(part, ret);
400
401                         dev_dbg(xpc_chan, "failed to pull openclose args from "
402                                 "partition %d, ret=%d\n", XPC_PARTID(part),
403                                 ret);
404
405                         /* don't bother processing IPIs anymore */
406                         IPI_amo = 0;
407                 }
408         }
409
410         if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_amo)) {
411                 ret = xpc_pull_remote_cachelines(part, part->remote_GPs,
412                                                 (void *) part->remote_GPs_pa,
413                                                 XPC_GP_SIZE);
414                 if (ret != xpcSuccess) {
415                         XPC_DEACTIVATE_PARTITION(part, ret);
416
417                         dev_dbg(xpc_chan, "failed to pull GPs from partition "
418                                 "%d, ret=%d\n", XPC_PARTID(part), ret);
419
420                         /* don't bother processing IPIs anymore */
421                         IPI_amo = 0;
422                 }
423         }
424
425         return IPI_amo;
426 }
427
428
429 /*
430  * Allocate the local message queue and the notify queue.
431  */
432 static enum xpc_retval
433 xpc_allocate_local_msgqueue(struct xpc_channel *ch)
434 {
435         unsigned long irq_flags;
436         int nentries;
437         size_t nbytes;
438
439
440         // >>> may want to check for ch->flags & XPC_C_DISCONNECTING between
441         // >>> iterations of the for-loop, bail if set?
442
443         // >>> should we impose a minumum #of entries? like 4 or 8?
444         for (nentries = ch->local_nentries; nentries > 0; nentries--) {
445
446                 nbytes = nentries * ch->msg_size;
447                 ch->local_msgqueue = xpc_kmalloc_cacheline_aligned(nbytes,
448                                                 (GFP_KERNEL | GFP_DMA),
449                                                 &ch->local_msgqueue_base);
450                 if (ch->local_msgqueue == NULL) {
451                         continue;
452                 }
453                 memset(ch->local_msgqueue, 0, nbytes);
454
455                 nbytes = nentries * sizeof(struct xpc_notify);
456                 ch->notify_queue = kmalloc(nbytes, (GFP_KERNEL | GFP_DMA));
457                 if (ch->notify_queue == NULL) {
458                         kfree(ch->local_msgqueue_base);
459                         ch->local_msgqueue = NULL;
460                         continue;
461                 }
462                 memset(ch->notify_queue, 0, nbytes);
463
464                 spin_lock_irqsave(&ch->lock, irq_flags);
465                 if (nentries < ch->local_nentries) {
466                         dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
467                                 "partid=%d, channel=%d\n", nentries,
468                                 ch->local_nentries, ch->partid, ch->number);
469
470                         ch->local_nentries = nentries;
471                 }
472                 spin_unlock_irqrestore(&ch->lock, irq_flags);
473                 return xpcSuccess;
474         }
475
476         dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
477                 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
478         return xpcNoMemory;
479 }
480
481
482 /*
483  * Allocate the cached remote message queue.
484  */
485 static enum xpc_retval
486 xpc_allocate_remote_msgqueue(struct xpc_channel *ch)
487 {
488         unsigned long irq_flags;
489         int nentries;
490         size_t nbytes;
491
492
493         DBUG_ON(ch->remote_nentries <= 0);
494
495         // >>> may want to check for ch->flags & XPC_C_DISCONNECTING between
496         // >>> iterations of the for-loop, bail if set?
497
498         // >>> should we impose a minumum #of entries? like 4 or 8?
499         for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
500
501                 nbytes = nentries * ch->msg_size;
502                 ch->remote_msgqueue = xpc_kmalloc_cacheline_aligned(nbytes,
503                                                 (GFP_KERNEL | GFP_DMA),
504                                                 &ch->remote_msgqueue_base);
505                 if (ch->remote_msgqueue == NULL) {
506                         continue;
507                 }
508                 memset(ch->remote_msgqueue, 0, nbytes);
509
510                 spin_lock_irqsave(&ch->lock, irq_flags);
511                 if (nentries < ch->remote_nentries) {
512                         dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
513                                 "partid=%d, channel=%d\n", nentries,
514                                 ch->remote_nentries, ch->partid, ch->number);
515
516                         ch->remote_nentries = nentries;
517                 }
518                 spin_unlock_irqrestore(&ch->lock, irq_flags);
519                 return xpcSuccess;
520         }
521
522         dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
523                 "partid=%d, channel=%d\n", ch->partid, ch->number);
524         return xpcNoMemory;
525 }
526
527
528 /*
529  * Allocate message queues and other stuff associated with a channel.
530  *
531  * Note: Assumes all of the channel sizes are filled in.
532  */
533 static enum xpc_retval
534 xpc_allocate_msgqueues(struct xpc_channel *ch)
535 {
536         unsigned long irq_flags;
537         int i;
538         enum xpc_retval ret;
539
540
541         DBUG_ON(ch->flags & XPC_C_SETUP);
542
543         if ((ret = xpc_allocate_local_msgqueue(ch)) != xpcSuccess) {
544                 return ret;
545         }
546
547         if ((ret = xpc_allocate_remote_msgqueue(ch)) != xpcSuccess) {
548                 kfree(ch->local_msgqueue_base);
549                 ch->local_msgqueue = NULL;
550                 kfree(ch->notify_queue);
551                 ch->notify_queue = NULL;
552                 return ret;
553         }
554
555         for (i = 0; i < ch->local_nentries; i++) {
556                 /* use a semaphore as an event wait queue */
557                 sema_init(&ch->notify_queue[i].sema, 0);
558         }
559
560         spin_lock_irqsave(&ch->lock, irq_flags);
561         ch->flags |= XPC_C_SETUP;
562         spin_unlock_irqrestore(&ch->lock, irq_flags);
563
564         return xpcSuccess;
565 }
566
567
568 /*
569  * Process a connect message from a remote partition.
570  *
571  * Note: xpc_process_connect() is expecting to be called with the
572  * spin_lock_irqsave held and will leave it locked upon return.
573  */
574 static void
575 xpc_process_connect(struct xpc_channel *ch, unsigned long *irq_flags)
576 {
577         enum xpc_retval ret;
578
579
580         DBUG_ON(!spin_is_locked(&ch->lock));
581
582         if (!(ch->flags & XPC_C_OPENREQUEST) ||
583                                 !(ch->flags & XPC_C_ROPENREQUEST)) {
584                 /* nothing more to do for now */
585                 return;
586         }
587         DBUG_ON(!(ch->flags & XPC_C_CONNECTING));
588
589         if (!(ch->flags & XPC_C_SETUP)) {
590                 spin_unlock_irqrestore(&ch->lock, *irq_flags);
591                 ret = xpc_allocate_msgqueues(ch);
592                 spin_lock_irqsave(&ch->lock, *irq_flags);
593
594                 if (ret != xpcSuccess) {
595                         XPC_DISCONNECT_CHANNEL(ch, ret, irq_flags);
596                 }
597                 if (ch->flags & (XPC_C_CONNECTED | XPC_C_DISCONNECTING)) {
598                         return;
599                 }
600
601                 DBUG_ON(!(ch->flags & XPC_C_SETUP));
602                 DBUG_ON(ch->local_msgqueue == NULL);
603                 DBUG_ON(ch->remote_msgqueue == NULL);
604         }
605
606         if (!(ch->flags & XPC_C_OPENREPLY)) {
607                 ch->flags |= XPC_C_OPENREPLY;
608                 xpc_IPI_send_openreply(ch, irq_flags);
609         }
610
611         if (!(ch->flags & XPC_C_ROPENREPLY)) {
612                 return;
613         }
614
615         DBUG_ON(ch->remote_msgqueue_pa == 0);
616
617         ch->flags = (XPC_C_CONNECTED | XPC_C_SETUP);    /* clear all else */
618
619         dev_info(xpc_chan, "channel %d to partition %d connected\n",
620                 ch->number, ch->partid);
621
622         spin_unlock_irqrestore(&ch->lock, *irq_flags);
623         xpc_create_kthreads(ch, 1);
624         spin_lock_irqsave(&ch->lock, *irq_flags);
625 }
626
627
628 /*
629  * Notify those who wanted to be notified upon delivery of their message.
630  */
631 static void
632 xpc_notify_senders(struct xpc_channel *ch, enum xpc_retval reason, s64 put)
633 {
634         struct xpc_notify *notify;
635         u8 notify_type;
636         s64 get = ch->w_remote_GP.get - 1;
637
638
639         while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
640
641                 notify = &ch->notify_queue[get % ch->local_nentries];
642
643                 /*
644                  * See if the notify entry indicates it was associated with
645                  * a message who's sender wants to be notified. It is possible
646                  * that it is, but someone else is doing or has done the
647                  * notification.
648                  */
649                 notify_type = notify->type;
650                 if (notify_type == 0 ||
651                                 cmpxchg(&notify->type, notify_type, 0) !=
652                                                                 notify_type) {
653                         continue;
654                 }
655
656                 DBUG_ON(notify_type != XPC_N_CALL);
657
658                 atomic_dec(&ch->n_to_notify);
659
660                 if (notify->func != NULL) {
661                         dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
662                                 "msg_number=%ld, partid=%d, channel=%d\n",
663                                 (void *) notify, get, ch->partid, ch->number);
664
665                         notify->func(reason, ch->partid, ch->number,
666                                                                 notify->key);
667
668                         dev_dbg(xpc_chan, "notify->func() returned, "
669                                 "notify=0x%p, msg_number=%ld, partid=%d, "
670                                 "channel=%d\n", (void *) notify, get,
671                                 ch->partid, ch->number);
672                 }
673         }
674 }
675
676
677 /*
678  * Free up message queues and other stuff that were allocated for the specified
679  * channel.
680  *
681  * Note: ch->reason and ch->reason_line are left set for debugging purposes,
682  * they're cleared when XPC_C_DISCONNECTED is cleared.
683  */
684 static void
685 xpc_free_msgqueues(struct xpc_channel *ch)
686 {
687         DBUG_ON(!spin_is_locked(&ch->lock));
688         DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
689
690         ch->remote_msgqueue_pa = 0;
691         ch->func = NULL;
692         ch->key = NULL;
693         ch->msg_size = 0;
694         ch->local_nentries = 0;
695         ch->remote_nentries = 0;
696         ch->kthreads_assigned_limit = 0;
697         ch->kthreads_idle_limit = 0;
698
699         ch->local_GP->get = 0;
700         ch->local_GP->put = 0;
701         ch->remote_GP.get = 0;
702         ch->remote_GP.put = 0;
703         ch->w_local_GP.get = 0;
704         ch->w_local_GP.put = 0;
705         ch->w_remote_GP.get = 0;
706         ch->w_remote_GP.put = 0;
707         ch->next_msg_to_pull = 0;
708
709         if (ch->flags & XPC_C_SETUP) {
710                 ch->flags &= ~XPC_C_SETUP;
711
712                 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
713                         ch->flags, ch->partid, ch->number);
714
715                 kfree(ch->local_msgqueue_base);
716                 ch->local_msgqueue = NULL;
717                 kfree(ch->remote_msgqueue_base);
718                 ch->remote_msgqueue = NULL;
719                 kfree(ch->notify_queue);
720                 ch->notify_queue = NULL;
721         }
722 }
723
724
725 /*
726  * spin_lock_irqsave() is expected to be held on entry.
727  */
728 static void
729 xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
730 {
731         struct xpc_partition *part = &xpc_partitions[ch->partid];
732         u32 channel_was_connected = (ch->flags & XPC_C_WASCONNECTED);
733
734
735         DBUG_ON(!spin_is_locked(&ch->lock));
736
737         if (!(ch->flags & XPC_C_DISCONNECTING)) {
738                 return;
739         }
740
741         DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
742
743         /* make sure all activity has settled down first */
744
745         if (atomic_read(&ch->references) > 0) {
746                 return;
747         }
748         DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
749
750         if (part->act_state == XPC_P_DEACTIVATING) {
751                 /* can't proceed until the other side disengages from us */
752                 if (xpc_partition_engaged(1UL << ch->partid)) {
753                         return;
754                 }
755
756         } else {
757
758                 /* as long as the other side is up do the full protocol */
759
760                 if (!(ch->flags & XPC_C_RCLOSEREQUEST)) {
761                         return;
762                 }
763
764                 if (!(ch->flags & XPC_C_CLOSEREPLY)) {
765                         ch->flags |= XPC_C_CLOSEREPLY;
766                         xpc_IPI_send_closereply(ch, irq_flags);
767                 }
768
769                 if (!(ch->flags & XPC_C_RCLOSEREPLY)) {
770                         return;
771                 }
772         }
773
774         /* wake those waiting for notify completion */
775         if (atomic_read(&ch->n_to_notify) > 0) {
776                 /* >>> we do callout while holding ch->lock */
777                 xpc_notify_senders(ch, ch->reason, ch->w_local_GP.put);
778         }
779
780         /* both sides are disconnected now */
781
782         /* it's now safe to free the channel's message queues */
783         xpc_free_msgqueues(ch);
784
785         /* mark disconnected, clear all other flags except XPC_C_WDISCONNECT */
786         ch->flags = (XPC_C_DISCONNECTED | (ch->flags & XPC_C_WDISCONNECT));
787
788         atomic_dec(&part->nchannels_active);
789
790         if (channel_was_connected) {
791                 dev_info(xpc_chan, "channel %d to partition %d disconnected, "
792                         "reason=%d\n", ch->number, ch->partid, ch->reason);
793         }
794
795         if (ch->flags & XPC_C_WDISCONNECT) {
796                 spin_unlock_irqrestore(&ch->lock, *irq_flags);
797                 up(&ch->wdisconnect_sema);
798                 spin_lock_irqsave(&ch->lock, *irq_flags);
799
800         } else if (ch->delayed_IPI_flags) {
801                 if (part->act_state != XPC_P_DEACTIVATING) {
802                         /* time to take action on any delayed IPI flags */
803                         spin_lock(&part->IPI_lock);
804                         XPC_SET_IPI_FLAGS(part->local_IPI_amo, ch->number,
805                                                         ch->delayed_IPI_flags);
806                         spin_unlock(&part->IPI_lock);
807                 }
808                 ch->delayed_IPI_flags = 0;
809         }
810 }
811
812
813 /*
814  * Process a change in the channel's remote connection state.
815  */
816 static void
817 xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
818                                 u8 IPI_flags)
819 {
820         unsigned long irq_flags;
821         struct xpc_openclose_args *args =
822                                 &part->remote_openclose_args[ch_number];
823         struct xpc_channel *ch = &part->channels[ch_number];
824         enum xpc_retval reason;
825
826
827
828         spin_lock_irqsave(&ch->lock, irq_flags);
829
830 again:
831
832         if ((ch->flags & XPC_C_DISCONNECTED) &&
833                                         (ch->flags & XPC_C_WDISCONNECT)) {
834                 /*
835                  * Delay processing IPI flags until thread waiting disconnect
836                  * has had a chance to see that the channel is disconnected.
837                  */
838                 ch->delayed_IPI_flags |= IPI_flags;
839                 spin_unlock_irqrestore(&ch->lock, irq_flags);
840                 return;
841         }
842
843
844         if (IPI_flags & XPC_IPI_CLOSEREQUEST) {
845
846                 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREQUEST (reason=%d) received "
847                         "from partid=%d, channel=%d\n", args->reason,
848                         ch->partid, ch->number);
849
850                 /*
851                  * If RCLOSEREQUEST is set, we're probably waiting for
852                  * RCLOSEREPLY. We should find it and a ROPENREQUEST packed
853                  * with this RCLOSEREQUEST in the IPI_flags.
854                  */
855
856                 if (ch->flags & XPC_C_RCLOSEREQUEST) {
857                         DBUG_ON(!(ch->flags & XPC_C_DISCONNECTING));
858                         DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
859                         DBUG_ON(!(ch->flags & XPC_C_CLOSEREPLY));
860                         DBUG_ON(ch->flags & XPC_C_RCLOSEREPLY);
861
862                         DBUG_ON(!(IPI_flags & XPC_IPI_CLOSEREPLY));
863                         IPI_flags &= ~XPC_IPI_CLOSEREPLY;
864                         ch->flags |= XPC_C_RCLOSEREPLY;
865
866                         /* both sides have finished disconnecting */
867                         xpc_process_disconnect(ch, &irq_flags);
868                         DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
869                         goto again;
870                 }
871
872                 if (ch->flags & XPC_C_DISCONNECTED) {
873                         if (!(IPI_flags & XPC_IPI_OPENREQUEST)) {
874                                 if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo,
875                                          ch_number) & XPC_IPI_OPENREQUEST)) {
876
877                                         DBUG_ON(ch->delayed_IPI_flags != 0);
878                                         spin_lock(&part->IPI_lock);
879                                         XPC_SET_IPI_FLAGS(part->local_IPI_amo,
880                                                         ch_number,
881                                                         XPC_IPI_CLOSEREQUEST);
882                                         spin_unlock(&part->IPI_lock);
883                                 }
884                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
885                                 return;
886                         }
887
888                         XPC_SET_REASON(ch, 0, 0);
889                         ch->flags &= ~XPC_C_DISCONNECTED;
890
891                         atomic_inc(&part->nchannels_active);
892                         ch->flags |= (XPC_C_CONNECTING | XPC_C_ROPENREQUEST);
893                 }
894
895                 IPI_flags &= ~(XPC_IPI_OPENREQUEST | XPC_IPI_OPENREPLY);
896
897                 /*
898                  * The meaningful CLOSEREQUEST connection state fields are:
899                  *      reason = reason connection is to be closed
900                  */
901
902                 ch->flags |= XPC_C_RCLOSEREQUEST;
903
904                 if (!(ch->flags & XPC_C_DISCONNECTING)) {
905                         reason = args->reason;
906                         if (reason <= xpcSuccess || reason > xpcUnknownReason) {
907                                 reason = xpcUnknownReason;
908                         } else if (reason == xpcUnregistering) {
909                                 reason = xpcOtherUnregistering;
910                         }
911
912                         XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
913
914                         DBUG_ON(IPI_flags & XPC_IPI_CLOSEREPLY);
915                         spin_unlock_irqrestore(&ch->lock, irq_flags);
916                         return;
917                 }
918
919                 xpc_process_disconnect(ch, &irq_flags);
920         }
921
922
923         if (IPI_flags & XPC_IPI_CLOSEREPLY) {
924
925                 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREPLY received from partid=%d,"
926                         " channel=%d\n", ch->partid, ch->number);
927
928                 if (ch->flags & XPC_C_DISCONNECTED) {
929                         DBUG_ON(part->act_state != XPC_P_DEACTIVATING);
930                         spin_unlock_irqrestore(&ch->lock, irq_flags);
931                         return;
932                 }
933
934                 DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
935
936                 if (!(ch->flags & XPC_C_RCLOSEREQUEST)) {
937                         if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo, ch_number)
938                                                 & XPC_IPI_CLOSEREQUEST)) {
939
940                                 DBUG_ON(ch->delayed_IPI_flags != 0);
941                                 spin_lock(&part->IPI_lock);
942                                 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
943                                                 ch_number, XPC_IPI_CLOSEREPLY);
944                                 spin_unlock(&part->IPI_lock);
945                         }
946                         spin_unlock_irqrestore(&ch->lock, irq_flags);
947                         return;
948                 }
949
950                 ch->flags |= XPC_C_RCLOSEREPLY;
951
952                 if (ch->flags & XPC_C_CLOSEREPLY) {
953                         /* both sides have finished disconnecting */
954                         xpc_process_disconnect(ch, &irq_flags);
955                 }
956         }
957
958
959         if (IPI_flags & XPC_IPI_OPENREQUEST) {
960
961                 dev_dbg(xpc_chan, "XPC_IPI_OPENREQUEST (msg_size=%d, "
962                         "local_nentries=%d) received from partid=%d, "
963                         "channel=%d\n", args->msg_size, args->local_nentries,
964                         ch->partid, ch->number);
965
966                 if (part->act_state == XPC_P_DEACTIVATING ||
967                                         (ch->flags & XPC_C_ROPENREQUEST)) {
968                         spin_unlock_irqrestore(&ch->lock, irq_flags);
969                         return;
970                 }
971
972                 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_WDISCONNECT)) {
973                         ch->delayed_IPI_flags |= XPC_IPI_OPENREQUEST;
974                         spin_unlock_irqrestore(&ch->lock, irq_flags);
975                         return;
976                 }
977                 DBUG_ON(!(ch->flags & (XPC_C_DISCONNECTED |
978                                                         XPC_C_OPENREQUEST)));
979                 DBUG_ON(ch->flags & (XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
980                                         XPC_C_OPENREPLY | XPC_C_CONNECTED));
981
982                 /*
983                  * The meaningful OPENREQUEST connection state fields are:
984                  *      msg_size = size of channel's messages in bytes
985                  *      local_nentries = remote partition's local_nentries
986                  */
987                 if (args->msg_size == 0 || args->local_nentries == 0) {
988                         /* assume OPENREQUEST was delayed by mistake */
989                         spin_unlock_irqrestore(&ch->lock, irq_flags);
990                         return;
991                 }
992
993                 ch->flags |= (XPC_C_ROPENREQUEST | XPC_C_CONNECTING);
994                 ch->remote_nentries = args->local_nentries;
995
996
997                 if (ch->flags & XPC_C_OPENREQUEST) {
998                         if (args->msg_size != ch->msg_size) {
999                                 XPC_DISCONNECT_CHANNEL(ch, xpcUnequalMsgSizes,
1000                                                                 &irq_flags);
1001                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1002                                 return;
1003                         }
1004                 } else {
1005                         ch->msg_size = args->msg_size;
1006
1007                         XPC_SET_REASON(ch, 0, 0);
1008                         ch->flags &= ~XPC_C_DISCONNECTED;
1009
1010                         atomic_inc(&part->nchannels_active);
1011                 }
1012
1013                 xpc_process_connect(ch, &irq_flags);
1014         }
1015
1016
1017         if (IPI_flags & XPC_IPI_OPENREPLY) {
1018
1019                 dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY (local_msgqueue_pa=0x%lx, "
1020                         "local_nentries=%d, remote_nentries=%d) received from "
1021                         "partid=%d, channel=%d\n", args->local_msgqueue_pa,
1022                         args->local_nentries, args->remote_nentries,
1023                         ch->partid, ch->number);
1024
1025                 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED)) {
1026                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1027                         return;
1028                 }
1029                 if (!(ch->flags & XPC_C_OPENREQUEST)) {
1030                         XPC_DISCONNECT_CHANNEL(ch, xpcOpenCloseError,
1031                                                                 &irq_flags);
1032                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1033                         return;
1034                 }
1035
1036                 DBUG_ON(!(ch->flags & XPC_C_ROPENREQUEST));
1037                 DBUG_ON(ch->flags & XPC_C_CONNECTED);
1038
1039                 /*
1040                  * The meaningful OPENREPLY connection state fields are:
1041                  *      local_msgqueue_pa = physical address of remote
1042                  *                          partition's local_msgqueue
1043                  *      local_nentries = remote partition's local_nentries
1044                  *      remote_nentries = remote partition's remote_nentries
1045                  */
1046                 DBUG_ON(args->local_msgqueue_pa == 0);
1047                 DBUG_ON(args->local_nentries == 0);
1048                 DBUG_ON(args->remote_nentries == 0);
1049
1050                 ch->flags |= XPC_C_ROPENREPLY;
1051                 ch->remote_msgqueue_pa = args->local_msgqueue_pa;
1052
1053                 if (args->local_nentries < ch->remote_nentries) {
1054                         dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1055                                 "remote_nentries=%d, old remote_nentries=%d, "
1056                                 "partid=%d, channel=%d\n",
1057                                 args->local_nentries, ch->remote_nentries,
1058                                 ch->partid, ch->number);
1059
1060                         ch->remote_nentries = args->local_nentries;
1061                 }
1062                 if (args->remote_nentries < ch->local_nentries) {
1063                         dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1064                                 "local_nentries=%d, old local_nentries=%d, "
1065                                 "partid=%d, channel=%d\n",
1066                                 args->remote_nentries, ch->local_nentries,
1067                                 ch->partid, ch->number);
1068
1069                         ch->local_nentries = args->remote_nentries;
1070                 }
1071
1072                 xpc_process_connect(ch, &irq_flags);
1073         }
1074
1075         spin_unlock_irqrestore(&ch->lock, irq_flags);
1076 }
1077
1078
1079 /*
1080  * Attempt to establish a channel connection to a remote partition.
1081  */
1082 static enum xpc_retval
1083 xpc_connect_channel(struct xpc_channel *ch)
1084 {
1085         unsigned long irq_flags;
1086         struct xpc_registration *registration = &xpc_registrations[ch->number];
1087
1088
1089         if (down_trylock(&registration->sema) != 0) {
1090                 return xpcRetry;
1091         }
1092
1093         if (!XPC_CHANNEL_REGISTERED(ch->number)) {
1094                 up(&registration->sema);
1095                 return xpcUnregistered;
1096         }
1097
1098         spin_lock_irqsave(&ch->lock, irq_flags);
1099
1100         DBUG_ON(ch->flags & XPC_C_CONNECTED);
1101         DBUG_ON(ch->flags & XPC_C_OPENREQUEST);
1102
1103         if (ch->flags & XPC_C_DISCONNECTING) {
1104                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1105                 up(&registration->sema);
1106                 return ch->reason;
1107         }
1108
1109
1110         /* add info from the channel connect registration to the channel */
1111
1112         ch->kthreads_assigned_limit = registration->assigned_limit;
1113         ch->kthreads_idle_limit = registration->idle_limit;
1114         DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
1115         DBUG_ON(atomic_read(&ch->kthreads_idle) != 0);
1116         DBUG_ON(atomic_read(&ch->kthreads_active) != 0);
1117
1118         ch->func = registration->func;
1119         DBUG_ON(registration->func == NULL);
1120         ch->key = registration->key;
1121
1122         ch->local_nentries = registration->nentries;
1123
1124         if (ch->flags & XPC_C_ROPENREQUEST) {
1125                 if (registration->msg_size != ch->msg_size) {
1126                         /* the local and remote sides aren't the same */
1127
1128                         /*
1129                          * Because XPC_DISCONNECT_CHANNEL() can block we're
1130                          * forced to up the registration sema before we unlock
1131                          * the channel lock. But that's okay here because we're
1132                          * done with the part that required the registration
1133                          * sema. XPC_DISCONNECT_CHANNEL() requires that the
1134                          * channel lock be locked and will unlock and relock
1135                          * the channel lock as needed.
1136                          */
1137                         up(&registration->sema);
1138                         XPC_DISCONNECT_CHANNEL(ch, xpcUnequalMsgSizes,
1139                                                                 &irq_flags);
1140                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1141                         return xpcUnequalMsgSizes;
1142                 }
1143         } else {
1144                 ch->msg_size = registration->msg_size;
1145
1146                 XPC_SET_REASON(ch, 0, 0);
1147                 ch->flags &= ~XPC_C_DISCONNECTED;
1148
1149                 atomic_inc(&xpc_partitions[ch->partid].nchannels_active);
1150         }
1151
1152         up(&registration->sema);
1153
1154
1155         /* initiate the connection */
1156
1157         ch->flags |= (XPC_C_OPENREQUEST | XPC_C_CONNECTING);
1158         xpc_IPI_send_openrequest(ch, &irq_flags);
1159
1160         xpc_process_connect(ch, &irq_flags);
1161
1162         spin_unlock_irqrestore(&ch->lock, irq_flags);
1163
1164         return xpcSuccess;
1165 }
1166
1167
1168 /*
1169  * Clear some of the msg flags in the local message queue.
1170  */
1171 static inline void
1172 xpc_clear_local_msgqueue_flags(struct xpc_channel *ch)
1173 {
1174         struct xpc_msg *msg;
1175         s64 get;
1176
1177
1178         get = ch->w_remote_GP.get;
1179         do {
1180                 msg = (struct xpc_msg *) ((u64) ch->local_msgqueue +
1181                                 (get % ch->local_nentries) * ch->msg_size);
1182                 msg->flags = 0;
1183         } while (++get < (volatile s64) ch->remote_GP.get);
1184 }
1185
1186
1187 /*
1188  * Clear some of the msg flags in the remote message queue.
1189  */
1190 static inline void
1191 xpc_clear_remote_msgqueue_flags(struct xpc_channel *ch)
1192 {
1193         struct xpc_msg *msg;
1194         s64 put;
1195
1196
1197         put = ch->w_remote_GP.put;
1198         do {
1199                 msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue +
1200                                 (put % ch->remote_nentries) * ch->msg_size);
1201                 msg->flags = 0;
1202         } while (++put < (volatile s64) ch->remote_GP.put);
1203 }
1204
1205
1206 static void
1207 xpc_process_msg_IPI(struct xpc_partition *part, int ch_number)
1208 {
1209         struct xpc_channel *ch = &part->channels[ch_number];
1210         int nmsgs_sent;
1211
1212
1213         ch->remote_GP = part->remote_GPs[ch_number];
1214
1215
1216         /* See what, if anything, has changed for each connected channel */
1217
1218         xpc_msgqueue_ref(ch);
1219
1220         if (ch->w_remote_GP.get == ch->remote_GP.get &&
1221                                 ch->w_remote_GP.put == ch->remote_GP.put) {
1222                 /* nothing changed since GPs were last pulled */
1223                 xpc_msgqueue_deref(ch);
1224                 return;
1225         }
1226
1227         if (!(ch->flags & XPC_C_CONNECTED)){
1228                 xpc_msgqueue_deref(ch);
1229                 return;
1230         }
1231
1232
1233         /*
1234          * First check to see if messages recently sent by us have been
1235          * received by the other side. (The remote GET value will have
1236          * changed since we last looked at it.)
1237          */
1238
1239         if (ch->w_remote_GP.get != ch->remote_GP.get) {
1240
1241                 /*
1242                  * We need to notify any senders that want to be notified
1243                  * that their sent messages have been received by their
1244                  * intended recipients. We need to do this before updating
1245                  * w_remote_GP.get so that we don't allocate the same message
1246                  * queue entries prematurely (see xpc_allocate_msg()).
1247                  */
1248                 if (atomic_read(&ch->n_to_notify) > 0) {
1249                         /*
1250                          * Notify senders that messages sent have been
1251                          * received and delivered by the other side.
1252                          */
1253                         xpc_notify_senders(ch, xpcMsgDelivered,
1254                                                         ch->remote_GP.get);
1255                 }
1256
1257                 /*
1258                  * Clear msg->flags in previously sent messages, so that
1259                  * they're ready for xpc_allocate_msg().
1260                  */
1261                 xpc_clear_local_msgqueue_flags(ch);
1262
1263                 ch->w_remote_GP.get = ch->remote_GP.get;
1264
1265                 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1266                         "channel=%d\n", ch->w_remote_GP.get, ch->partid,
1267                         ch->number);
1268
1269                 /*
1270                  * If anyone was waiting for message queue entries to become
1271                  * available, wake them up.
1272                  */
1273                 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0) {
1274                         wake_up(&ch->msg_allocate_wq);
1275                 }
1276         }
1277
1278
1279         /*
1280          * Now check for newly sent messages by the other side. (The remote
1281          * PUT value will have changed since we last looked at it.)
1282          */
1283
1284         if (ch->w_remote_GP.put != ch->remote_GP.put) {
1285                 /*
1286                  * Clear msg->flags in previously received messages, so that
1287                  * they're ready for xpc_get_deliverable_msg().
1288                  */
1289                 xpc_clear_remote_msgqueue_flags(ch);
1290
1291                 ch->w_remote_GP.put = ch->remote_GP.put;
1292
1293                 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1294                         "channel=%d\n", ch->w_remote_GP.put, ch->partid,
1295                         ch->number);
1296
1297                 nmsgs_sent = ch->w_remote_GP.put - ch->w_local_GP.get;
1298                 if (nmsgs_sent > 0) {
1299                         dev_dbg(xpc_chan, "msgs waiting to be copied and "
1300                                 "delivered=%d, partid=%d, channel=%d\n",
1301                                 nmsgs_sent, ch->partid, ch->number);
1302
1303                         if (ch->flags & XPC_C_CONNECTCALLOUT) {
1304                                 xpc_activate_kthreads(ch, nmsgs_sent);
1305                         }
1306                 }
1307         }
1308
1309         xpc_msgqueue_deref(ch);
1310 }
1311
1312
1313 void
1314 xpc_process_channel_activity(struct xpc_partition *part)
1315 {
1316         unsigned long irq_flags;
1317         u64 IPI_amo, IPI_flags;
1318         struct xpc_channel *ch;
1319         int ch_number;
1320         u32 ch_flags;
1321
1322
1323         IPI_amo = xpc_get_IPI_flags(part);
1324
1325         /*
1326          * Initiate channel connections for registered channels.
1327          *
1328          * For each connected channel that has pending messages activate idle
1329          * kthreads and/or create new kthreads as needed.
1330          */
1331
1332         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1333                 ch = &part->channels[ch_number];
1334
1335
1336                 /*
1337                  * Process any open or close related IPI flags, and then deal
1338                  * with connecting or disconnecting the channel as required.
1339                  */
1340
1341                 IPI_flags = XPC_GET_IPI_FLAGS(IPI_amo, ch_number);
1342
1343                 if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_flags)) {
1344                         xpc_process_openclose_IPI(part, ch_number, IPI_flags);
1345                 }
1346
1347                 ch_flags = ch->flags;   /* need an atomic snapshot of flags */
1348
1349                 if (ch_flags & XPC_C_DISCONNECTING) {
1350                         spin_lock_irqsave(&ch->lock, irq_flags);
1351                         xpc_process_disconnect(ch, &irq_flags);
1352                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1353                         continue;
1354                 }
1355
1356                 if (part->act_state == XPC_P_DEACTIVATING) {
1357                         continue;
1358                 }
1359
1360                 if (!(ch_flags & XPC_C_CONNECTED)) {
1361                         if (!(ch_flags & XPC_C_OPENREQUEST)) {
1362                                 DBUG_ON(ch_flags & XPC_C_SETUP);
1363                                 (void) xpc_connect_channel(ch);
1364                         } else {
1365                                 spin_lock_irqsave(&ch->lock, irq_flags);
1366                                 xpc_process_connect(ch, &irq_flags);
1367                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1368                         }
1369                         continue;
1370                 }
1371
1372
1373                 /*
1374                  * Process any message related IPI flags, this may involve the
1375                  * activation of kthreads to deliver any pending messages sent
1376                  * from the other partition.
1377                  */
1378
1379                 if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_flags)) {
1380                         xpc_process_msg_IPI(part, ch_number);
1381                 }
1382         }
1383 }
1384
1385
1386 /*
1387  * XPC's heartbeat code calls this function to inform XPC that a partition is
1388  * going down.  XPC responds by tearing down the XPartition Communication
1389  * infrastructure used for the just downed partition.
1390  *
1391  * XPC's heartbeat code will never call this function and xpc_partition_up()
1392  * at the same time. Nor will it ever make multiple calls to either function
1393  * at the same time.
1394  */
1395 void
1396 xpc_partition_going_down(struct xpc_partition *part, enum xpc_retval reason)
1397 {
1398         unsigned long irq_flags;
1399         int ch_number;
1400         struct xpc_channel *ch;
1401
1402
1403         dev_dbg(xpc_chan, "deactivating partition %d, reason=%d\n",
1404                 XPC_PARTID(part), reason);
1405
1406         if (!xpc_part_ref(part)) {
1407                 /* infrastructure for this partition isn't currently set up */
1408                 return;
1409         }
1410
1411
1412         /* disconnect channels associated with the partition going down */
1413
1414         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1415                 ch = &part->channels[ch_number];
1416
1417                 xpc_msgqueue_ref(ch);
1418                 spin_lock_irqsave(&ch->lock, irq_flags);
1419
1420                 XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
1421
1422                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1423                 xpc_msgqueue_deref(ch);
1424         }
1425
1426         xpc_wakeup_channel_mgr(part);
1427
1428         xpc_part_deref(part);
1429 }
1430
1431
1432 /*
1433  * Teardown the infrastructure necessary to support XPartition Communication
1434  * between the specified remote partition and the local one.
1435  */
1436 void
1437 xpc_teardown_infrastructure(struct xpc_partition *part)
1438 {
1439         partid_t partid = XPC_PARTID(part);
1440
1441
1442         /*
1443          * We start off by making this partition inaccessible to local
1444          * processes by marking it as no longer setup. Then we make it
1445          * inaccessible to remote processes by clearing the XPC per partition
1446          * specific variable's magic # (which indicates that these variables
1447          * are no longer valid) and by ignoring all XPC notify IPIs sent to
1448          * this partition.
1449          */
1450
1451         DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1452         DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1453         DBUG_ON(part->setup_state != XPC_P_SETUP);
1454         part->setup_state = XPC_P_WTEARDOWN;
1455
1456         xpc_vars_part[partid].magic = 0;
1457
1458
1459         free_irq(SGI_XPC_NOTIFY, (void *) (u64) partid);
1460
1461
1462         /*
1463          * Before proceding with the teardown we have to wait until all
1464          * existing references cease.
1465          */
1466         wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1467
1468
1469         /* now we can begin tearing down the infrastructure */
1470
1471         part->setup_state = XPC_P_TORNDOWN;
1472
1473         /* in case we've still got outstanding timers registered... */
1474         del_timer_sync(&part->dropped_IPI_timer);
1475
1476         kfree(part->remote_openclose_args_base);
1477         part->remote_openclose_args = NULL;
1478         kfree(part->local_openclose_args_base);
1479         part->local_openclose_args = NULL;
1480         kfree(part->remote_GPs_base);
1481         part->remote_GPs = NULL;
1482         kfree(part->local_GPs_base);
1483         part->local_GPs = NULL;
1484         kfree(part->channels);
1485         part->channels = NULL;
1486         part->local_IPI_amo_va = NULL;
1487 }
1488
1489
1490 /*
1491  * Called by XP at the time of channel connection registration to cause
1492  * XPC to establish connections to all currently active partitions.
1493  */
1494 void
1495 xpc_initiate_connect(int ch_number)
1496 {
1497         partid_t partid;
1498         struct xpc_partition *part;
1499         struct xpc_channel *ch;
1500
1501
1502         DBUG_ON(ch_number < 0 || ch_number >= XPC_NCHANNELS);
1503
1504         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1505                 part = &xpc_partitions[partid];
1506
1507                 if (xpc_part_ref(part)) {
1508                         ch = &part->channels[ch_number];
1509
1510                         /*
1511                          * Initiate the establishment of a connection on the
1512                          * newly registered channel to the remote partition.
1513                          */
1514                         xpc_wakeup_channel_mgr(part);
1515                         xpc_part_deref(part);
1516                 }
1517         }
1518 }
1519
1520
1521 void
1522 xpc_connected_callout(struct xpc_channel *ch)
1523 {
1524         /* let the registerer know that a connection has been established */
1525
1526         if (ch->func != NULL) {
1527                 dev_dbg(xpc_chan, "ch->func() called, reason=xpcConnected, "
1528                         "partid=%d, channel=%d\n", ch->partid, ch->number);
1529
1530                 ch->func(xpcConnected, ch->partid, ch->number,
1531                                 (void *) (u64) ch->local_nentries, ch->key);
1532
1533                 dev_dbg(xpc_chan, "ch->func() returned, reason=xpcConnected, "
1534                         "partid=%d, channel=%d\n", ch->partid, ch->number);
1535         }
1536 }
1537
1538
1539 /*
1540  * Called by XP at the time of channel connection unregistration to cause
1541  * XPC to teardown all current connections for the specified channel.
1542  *
1543  * Before returning xpc_initiate_disconnect() will wait until all connections
1544  * on the specified channel have been closed/torndown. So the caller can be
1545  * assured that they will not be receiving any more callouts from XPC to the
1546  * function they registered via xpc_connect().
1547  *
1548  * Arguments:
1549  *
1550  *      ch_number - channel # to unregister.
1551  */
1552 void
1553 xpc_initiate_disconnect(int ch_number)
1554 {
1555         unsigned long irq_flags;
1556         partid_t partid;
1557         struct xpc_partition *part;
1558         struct xpc_channel *ch;
1559
1560
1561         DBUG_ON(ch_number < 0 || ch_number >= XPC_NCHANNELS);
1562
1563         /* initiate the channel disconnect for every active partition */
1564         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1565                 part = &xpc_partitions[partid];
1566
1567                 if (xpc_part_ref(part)) {
1568                         ch = &part->channels[ch_number];
1569                         xpc_msgqueue_ref(ch);
1570
1571                         spin_lock_irqsave(&ch->lock, irq_flags);
1572
1573                         if (!(ch->flags & XPC_C_DISCONNECTED)) {
1574                                 ch->flags |= XPC_C_WDISCONNECT;
1575
1576                                 XPC_DISCONNECT_CHANNEL(ch, xpcUnregistering,
1577                                                                 &irq_flags);
1578                         }
1579
1580                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1581
1582                         xpc_msgqueue_deref(ch);
1583                         xpc_part_deref(part);
1584                 }
1585         }
1586
1587         xpc_disconnect_wait(ch_number);
1588 }
1589
1590
1591 /*
1592  * To disconnect a channel, and reflect it back to all who may be waiting.
1593  *
1594  * An OPEN is not allowed until XPC_C_DISCONNECTING is cleared by
1595  * xpc_process_disconnect(), and if set, XPC_C_WDISCONNECT is cleared by
1596  * xpc_disconnect_wait().
1597  *
1598  * THE CHANNEL IS TO BE LOCKED BY THE CALLER AND WILL REMAIN LOCKED UPON RETURN.
1599  */
1600 void
1601 xpc_disconnect_channel(const int line, struct xpc_channel *ch,
1602                         enum xpc_retval reason, unsigned long *irq_flags)
1603 {
1604         u32 channel_was_connected = (ch->flags & XPC_C_CONNECTED);
1605
1606
1607         DBUG_ON(!spin_is_locked(&ch->lock));
1608
1609         if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED)) {
1610                 return;
1611         }
1612         DBUG_ON(!(ch->flags & (XPC_C_CONNECTING | XPC_C_CONNECTED)));
1613
1614         dev_dbg(xpc_chan, "reason=%d, line=%d, partid=%d, channel=%d\n",
1615                 reason, line, ch->partid, ch->number);
1616
1617         XPC_SET_REASON(ch, reason, line);
1618
1619         ch->flags |= (XPC_C_CLOSEREQUEST | XPC_C_DISCONNECTING);
1620         /* some of these may not have been set */
1621         ch->flags &= ~(XPC_C_OPENREQUEST | XPC_C_OPENREPLY |
1622                         XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
1623                         XPC_C_CONNECTING | XPC_C_CONNECTED);
1624
1625         xpc_IPI_send_closerequest(ch, irq_flags);
1626
1627         if (channel_was_connected) {
1628                 ch->flags |= XPC_C_WASCONNECTED;
1629         }
1630
1631         spin_unlock_irqrestore(&ch->lock, *irq_flags);
1632
1633         /* wake all idle kthreads so they can exit */
1634         if (atomic_read(&ch->kthreads_idle) > 0) {
1635                 wake_up_all(&ch->idle_wq);
1636         }
1637
1638         /* wake those waiting to allocate an entry from the local msg queue */
1639         if (atomic_read(&ch->n_on_msg_allocate_wq) > 0) {
1640                 wake_up(&ch->msg_allocate_wq);
1641         }
1642
1643         spin_lock_irqsave(&ch->lock, *irq_flags);
1644 }
1645
1646
1647 void
1648 xpc_disconnecting_callout(struct xpc_channel *ch)
1649 {
1650         /*
1651          * Let the channel's registerer know that the channel is being
1652          * disconnected. We don't want to do this if the registerer was never
1653          * informed of a connection being made.
1654          */
1655
1656         if (ch->func != NULL) {
1657                 dev_dbg(xpc_chan, "ch->func() called, reason=xpcDisconnecting,"
1658                         " partid=%d, channel=%d\n", ch->partid, ch->number);
1659
1660                 ch->func(xpcDisconnecting, ch->partid, ch->number, NULL,
1661                                                                 ch->key);
1662
1663                 dev_dbg(xpc_chan, "ch->func() returned, reason="
1664                         "xpcDisconnecting, partid=%d, channel=%d\n",
1665                         ch->partid, ch->number);
1666         }
1667 }
1668
1669
1670 /*
1671  * Wait for a message entry to become available for the specified channel,
1672  * but don't wait any longer than 1 jiffy.
1673  */
1674 static enum xpc_retval
1675 xpc_allocate_msg_wait(struct xpc_channel *ch)
1676 {
1677         enum xpc_retval ret;
1678
1679
1680         if (ch->flags & XPC_C_DISCONNECTING) {
1681                 DBUG_ON(ch->reason == xpcInterrupted);  // >>> Is this true?
1682                 return ch->reason;
1683         }
1684
1685         atomic_inc(&ch->n_on_msg_allocate_wq);
1686         ret = interruptible_sleep_on_timeout(&ch->msg_allocate_wq, 1);
1687         atomic_dec(&ch->n_on_msg_allocate_wq);
1688
1689         if (ch->flags & XPC_C_DISCONNECTING) {
1690                 ret = ch->reason;
1691                 DBUG_ON(ch->reason == xpcInterrupted);  // >>> Is this true?
1692         } else if (ret == 0) {
1693                 ret = xpcTimeout;
1694         } else {
1695                 ret = xpcInterrupted;
1696         }
1697
1698         return ret;
1699 }
1700
1701
1702 /*
1703  * Allocate an entry for a message from the message queue associated with the
1704  * specified channel.
1705  */
1706 static enum xpc_retval
1707 xpc_allocate_msg(struct xpc_channel *ch, u32 flags,
1708                         struct xpc_msg **address_of_msg)
1709 {
1710         struct xpc_msg *msg;
1711         enum xpc_retval ret;
1712         s64 put;
1713
1714
1715         /* this reference will be dropped in xpc_send_msg() */
1716         xpc_msgqueue_ref(ch);
1717
1718         if (ch->flags & XPC_C_DISCONNECTING) {
1719                 xpc_msgqueue_deref(ch);
1720                 return ch->reason;
1721         }
1722         if (!(ch->flags & XPC_C_CONNECTED)) {
1723                 xpc_msgqueue_deref(ch);
1724                 return xpcNotConnected;
1725         }
1726
1727
1728         /*
1729          * Get the next available message entry from the local message queue.
1730          * If none are available, we'll make sure that we grab the latest
1731          * GP values.
1732          */
1733         ret = xpcTimeout;
1734
1735         while (1) {
1736
1737                 put = (volatile s64) ch->w_local_GP.put;
1738                 if (put - (volatile s64) ch->w_remote_GP.get <
1739                                                         ch->local_nentries) {
1740
1741                         /* There are available message entries. We need to try
1742                          * to secure one for ourselves. We'll do this by trying
1743                          * to increment w_local_GP.put as long as someone else
1744                          * doesn't beat us to it. If they do, we'll have to
1745                          * try again.
1746                          */
1747                         if (cmpxchg(&ch->w_local_GP.put, put, put + 1) ==
1748                                                                         put) {
1749                                 /* we got the entry referenced by put */
1750                                 break;
1751                         }
1752                         continue;       /* try again */
1753                 }
1754
1755
1756                 /*
1757                  * There aren't any available msg entries at this time.
1758                  *
1759                  * In waiting for a message entry to become available,
1760                  * we set a timeout in case the other side is not
1761                  * sending completion IPIs. This lets us fake an IPI
1762                  * that will cause the IPI handler to fetch the latest
1763                  * GP values as if an IPI was sent by the other side.
1764                  */
1765                 if (ret == xpcTimeout) {
1766                         xpc_IPI_send_local_msgrequest(ch);
1767                 }
1768
1769                 if (flags & XPC_NOWAIT) {
1770                         xpc_msgqueue_deref(ch);
1771                         return xpcNoWait;
1772                 }
1773
1774                 ret = xpc_allocate_msg_wait(ch);
1775                 if (ret != xpcInterrupted && ret != xpcTimeout) {
1776                         xpc_msgqueue_deref(ch);
1777                         return ret;
1778                 }
1779         }
1780
1781
1782         /* get the message's address and initialize it */
1783         msg = (struct xpc_msg *) ((u64) ch->local_msgqueue +
1784                                 (put % ch->local_nentries) * ch->msg_size);
1785
1786
1787         DBUG_ON(msg->flags != 0);
1788         msg->number = put;
1789
1790         dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
1791                 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
1792                 (void *) msg, msg->number, ch->partid, ch->number);
1793
1794         *address_of_msg = msg;
1795
1796         return xpcSuccess;
1797 }
1798
1799
1800 /*
1801  * Allocate an entry for a message from the message queue associated with the
1802  * specified channel. NOTE that this routine can sleep waiting for a message
1803  * entry to become available. To not sleep, pass in the XPC_NOWAIT flag.
1804  *
1805  * Arguments:
1806  *
1807  *      partid - ID of partition to which the channel is connected.
1808  *      ch_number - channel #.
1809  *      flags - see xpc.h for valid flags.
1810  *      payload - address of the allocated payload area pointer (filled in on
1811  *                return) in which the user-defined message is constructed.
1812  */
1813 enum xpc_retval
1814 xpc_initiate_allocate(partid_t partid, int ch_number, u32 flags, void **payload)
1815 {
1816         struct xpc_partition *part = &xpc_partitions[partid];
1817         enum xpc_retval ret = xpcUnknownReason;
1818         struct xpc_msg *msg;
1819
1820
1821         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
1822         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1823
1824         *payload = NULL;
1825
1826         if (xpc_part_ref(part)) {
1827                 ret = xpc_allocate_msg(&part->channels[ch_number], flags, &msg);
1828                 xpc_part_deref(part);
1829
1830                 if (msg != NULL) {
1831                         *payload = &msg->payload;
1832                 }
1833         }
1834
1835         return ret;
1836 }
1837
1838
1839 /*
1840  * Now we actually send the messages that are ready to be sent by advancing
1841  * the local message queue's Put value and then send an IPI to the recipient
1842  * partition.
1843  */
1844 static void
1845 xpc_send_msgs(struct xpc_channel *ch, s64 initial_put)
1846 {
1847         struct xpc_msg *msg;
1848         s64 put = initial_put + 1;
1849         int send_IPI = 0;
1850
1851
1852         while (1) {
1853
1854                 while (1) {
1855                         if (put == (volatile s64) ch->w_local_GP.put) {
1856                                 break;
1857                         }
1858
1859                         msg = (struct xpc_msg *) ((u64) ch->local_msgqueue +
1860                                (put % ch->local_nentries) * ch->msg_size);
1861
1862                         if (!(msg->flags & XPC_M_READY)) {
1863                                 break;
1864                         }
1865
1866                         put++;
1867                 }
1868
1869                 if (put == initial_put) {
1870                         /* nothing's changed */
1871                         break;
1872                 }
1873
1874                 if (cmpxchg_rel(&ch->local_GP->put, initial_put, put) !=
1875                                                                 initial_put) {
1876                         /* someone else beat us to it */
1877                         DBUG_ON((volatile s64) ch->local_GP->put < initial_put);
1878                         break;
1879                 }
1880
1881                 /* we just set the new value of local_GP->put */
1882
1883                 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
1884                         "channel=%d\n", put, ch->partid, ch->number);
1885
1886                 send_IPI = 1;
1887
1888                 /*
1889                  * We need to ensure that the message referenced by
1890                  * local_GP->put is not XPC_M_READY or that local_GP->put
1891                  * equals w_local_GP.put, so we'll go have a look.
1892                  */
1893                 initial_put = put;
1894         }
1895
1896         if (send_IPI) {
1897                 xpc_IPI_send_msgrequest(ch);
1898         }
1899 }
1900
1901
1902 /*
1903  * Common code that does the actual sending of the message by advancing the
1904  * local message queue's Put value and sends an IPI to the partition the
1905  * message is being sent to.
1906  */
1907 static enum xpc_retval
1908 xpc_send_msg(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type,
1909                         xpc_notify_func func, void *key)
1910 {
1911         enum xpc_retval ret = xpcSuccess;
1912         struct xpc_notify *notify = notify;
1913         s64 put, msg_number = msg->number;
1914
1915
1916         DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
1917         DBUG_ON((((u64) msg - (u64) ch->local_msgqueue) / ch->msg_size) !=
1918                                         msg_number % ch->local_nentries);
1919         DBUG_ON(msg->flags & XPC_M_READY);
1920
1921         if (ch->flags & XPC_C_DISCONNECTING) {
1922                 /* drop the reference grabbed in xpc_allocate_msg() */
1923                 xpc_msgqueue_deref(ch);
1924                 return ch->reason;
1925         }
1926
1927         if (notify_type != 0) {
1928                 /*
1929                  * Tell the remote side to send an ACK interrupt when the
1930                  * message has been delivered.
1931                  */
1932                 msg->flags |= XPC_M_INTERRUPT;
1933
1934                 atomic_inc(&ch->n_to_notify);
1935
1936                 notify = &ch->notify_queue[msg_number % ch->local_nentries];
1937                 notify->func = func;
1938                 notify->key = key;
1939                 notify->type = notify_type;
1940
1941                 // >>> is a mb() needed here?
1942
1943                 if (ch->flags & XPC_C_DISCONNECTING) {
1944                         /*
1945                          * An error occurred between our last error check and
1946                          * this one. We will try to clear the type field from
1947                          * the notify entry. If we succeed then
1948                          * xpc_disconnect_channel() didn't already process
1949                          * the notify entry.
1950                          */
1951                         if (cmpxchg(&notify->type, notify_type, 0) ==
1952                                                                 notify_type) {
1953                                 atomic_dec(&ch->n_to_notify);
1954                                 ret = ch->reason;
1955                         }
1956
1957                         /* drop the reference grabbed in xpc_allocate_msg() */
1958                         xpc_msgqueue_deref(ch);
1959                         return ret;
1960                 }
1961         }
1962
1963         msg->flags |= XPC_M_READY;
1964
1965         /*
1966          * The preceding store of msg->flags must occur before the following
1967          * load of ch->local_GP->put.
1968          */
1969         mb();
1970
1971         /* see if the message is next in line to be sent, if so send it */
1972
1973         put = ch->local_GP->put;
1974         if (put == msg_number) {
1975                 xpc_send_msgs(ch, put);
1976         }
1977
1978         /* drop the reference grabbed in xpc_allocate_msg() */
1979         xpc_msgqueue_deref(ch);
1980         return ret;
1981 }
1982
1983
1984 /*
1985  * Send a message previously allocated using xpc_initiate_allocate() on the
1986  * specified channel connected to the specified partition.
1987  *
1988  * This routine will not wait for the message to be received, nor will
1989  * notification be given when it does happen. Once this routine has returned
1990  * the message entry allocated via xpc_initiate_allocate() is no longer
1991  * accessable to the caller.
1992  *
1993  * This routine, although called by users, does not call xpc_part_ref() to
1994  * ensure that the partition infrastructure is in place. It relies on the
1995  * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
1996  *
1997  * Arguments:
1998  *
1999  *      partid - ID of partition to which the channel is connected.
2000  *      ch_number - channel # to send message on.
2001  *      payload - pointer to the payload area allocated via
2002  *                      xpc_initiate_allocate().
2003  */
2004 enum xpc_retval
2005 xpc_initiate_send(partid_t partid, int ch_number, void *payload)
2006 {
2007         struct xpc_partition *part = &xpc_partitions[partid];
2008         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2009         enum xpc_retval ret;
2010
2011
2012         dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *) msg,
2013                 partid, ch_number);
2014
2015         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
2016         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2017         DBUG_ON(msg == NULL);
2018
2019         ret = xpc_send_msg(&part->channels[ch_number], msg, 0, NULL, NULL);
2020
2021         return ret;
2022 }
2023
2024
2025 /*
2026  * Send a message previously allocated using xpc_initiate_allocate on the
2027  * specified channel connected to the specified partition.
2028  *
2029  * This routine will not wait for the message to be sent. Once this routine
2030  * has returned the message entry allocated via xpc_initiate_allocate() is no
2031  * longer accessable to the caller.
2032  *
2033  * Once the remote end of the channel has received the message, the function
2034  * passed as an argument to xpc_initiate_send_notify() will be called. This
2035  * allows the sender to free up or re-use any buffers referenced by the
2036  * message, but does NOT mean the message has been processed at the remote
2037  * end by a receiver.
2038  *
2039  * If this routine returns an error, the caller's function will NOT be called.
2040  *
2041  * This routine, although called by users, does not call xpc_part_ref() to
2042  * ensure that the partition infrastructure is in place. It relies on the
2043  * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
2044  *
2045  * Arguments:
2046  *
2047  *      partid - ID of partition to which the channel is connected.
2048  *      ch_number - channel # to send message on.
2049  *      payload - pointer to the payload area allocated via
2050  *                      xpc_initiate_allocate().
2051  *      func - function to call with asynchronous notification of message
2052  *                receipt. THIS FUNCTION MUST BE NON-BLOCKING.
2053  *      key - user-defined key to be passed to the function when it's called.
2054  */
2055 enum xpc_retval
2056 xpc_initiate_send_notify(partid_t partid, int ch_number, void *payload,
2057                                 xpc_notify_func func, void *key)
2058 {
2059         struct xpc_partition *part = &xpc_partitions[partid];
2060         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2061         enum xpc_retval ret;
2062
2063
2064         dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *) msg,
2065                 partid, ch_number);
2066
2067         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
2068         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2069         DBUG_ON(msg == NULL);
2070         DBUG_ON(func == NULL);
2071
2072         ret = xpc_send_msg(&part->channels[ch_number], msg, XPC_N_CALL,
2073                                                                 func, key);
2074         return ret;
2075 }
2076
2077
2078 static struct xpc_msg *
2079 xpc_pull_remote_msg(struct xpc_channel *ch, s64 get)
2080 {
2081         struct xpc_partition *part = &xpc_partitions[ch->partid];
2082         struct xpc_msg *remote_msg, *msg;
2083         u32 msg_index, nmsgs;
2084         u64 msg_offset;
2085         enum xpc_retval ret;
2086
2087
2088         if (down_interruptible(&ch->msg_to_pull_sema) != 0) {
2089                 /* we were interrupted by a signal */
2090                 return NULL;
2091         }
2092
2093         while (get >= ch->next_msg_to_pull) {
2094
2095                 /* pull as many messages as are ready and able to be pulled */
2096
2097                 msg_index = ch->next_msg_to_pull % ch->remote_nentries;
2098
2099                 DBUG_ON(ch->next_msg_to_pull >=
2100                                         (volatile s64) ch->w_remote_GP.put);
2101                 nmsgs =  (volatile s64) ch->w_remote_GP.put -
2102                                                 ch->next_msg_to_pull;
2103                 if (msg_index + nmsgs > ch->remote_nentries) {
2104                         /* ignore the ones that wrap the msg queue for now */
2105                         nmsgs = ch->remote_nentries - msg_index;
2106                 }
2107
2108                 msg_offset = msg_index * ch->msg_size;
2109                 msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue +
2110                                                                 msg_offset);
2111                 remote_msg = (struct xpc_msg *) (ch->remote_msgqueue_pa +
2112                                                                 msg_offset);
2113
2114                 if ((ret = xpc_pull_remote_cachelines(part, msg, remote_msg,
2115                                 nmsgs * ch->msg_size)) != xpcSuccess) {
2116
2117                         dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
2118                                 " msg %ld from partition %d, channel=%d, "
2119                                 "ret=%d\n", nmsgs, ch->next_msg_to_pull,
2120                                 ch->partid, ch->number, ret);
2121
2122                         XPC_DEACTIVATE_PARTITION(part, ret);
2123
2124                         up(&ch->msg_to_pull_sema);
2125                         return NULL;
2126                 }
2127
2128                 mb();   /* >>> this may not be needed, we're not sure */
2129
2130                 ch->next_msg_to_pull += nmsgs;
2131         }
2132
2133         up(&ch->msg_to_pull_sema);
2134
2135         /* return the message we were looking for */
2136         msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2137         msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue + msg_offset);
2138
2139         return msg;
2140 }
2141
2142
2143 /*
2144  * Get a message to be delivered.
2145  */
2146 static struct xpc_msg *
2147 xpc_get_deliverable_msg(struct xpc_channel *ch)
2148 {
2149         struct xpc_msg *msg = NULL;
2150         s64 get;
2151
2152
2153         do {
2154                 if ((volatile u32) ch->flags & XPC_C_DISCONNECTING) {
2155                         break;
2156                 }
2157
2158                 get = (volatile s64) ch->w_local_GP.get;
2159                 if (get == (volatile s64) ch->w_remote_GP.put) {
2160                         break;
2161                 }
2162
2163                 /* There are messages waiting to be pulled and delivered.
2164                  * We need to try to secure one for ourselves. We'll do this
2165                  * by trying to increment w_local_GP.get and hope that no one
2166                  * else beats us to it. If they do, we'll we'll simply have
2167                  * to try again for the next one.
2168                  */
2169
2170                 if (cmpxchg(&ch->w_local_GP.get, get, get + 1) == get) {
2171                         /* we got the entry referenced by get */
2172
2173                         dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2174                                 "partid=%d, channel=%d\n", get + 1,
2175                                 ch->partid, ch->number);
2176
2177                         /* pull the message from the remote partition */
2178
2179                         msg = xpc_pull_remote_msg(ch, get);
2180
2181                         DBUG_ON(msg != NULL && msg->number != get);
2182                         DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2183                         DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2184
2185                         break;
2186                 }
2187
2188         } while (1);
2189
2190         return msg;
2191 }
2192
2193
2194 /*
2195  * Deliver a message to its intended recipient.
2196  */
2197 void
2198 xpc_deliver_msg(struct xpc_channel *ch)
2199 {
2200         struct xpc_msg *msg;
2201
2202
2203         if ((msg = xpc_get_deliverable_msg(ch)) != NULL) {
2204
2205                 /*
2206                  * This ref is taken to protect the payload itself from being
2207                  * freed before the user is finished with it, which the user
2208                  * indicates by calling xpc_initiate_received().
2209                  */
2210                 xpc_msgqueue_ref(ch);
2211
2212                 atomic_inc(&ch->kthreads_active);
2213
2214                 if (ch->func != NULL) {
2215                         dev_dbg(xpc_chan, "ch->func() called, msg=0x%p, "
2216                                 "msg_number=%ld, partid=%d, channel=%d\n",
2217                                 (void *) msg, msg->number, ch->partid,
2218                                 ch->number);
2219
2220                         /* deliver the message to its intended recipient */
2221                         ch->func(xpcMsgReceived, ch->partid, ch->number,
2222                                         &msg->payload, ch->key);
2223
2224                         dev_dbg(xpc_chan, "ch->func() returned, msg=0x%p, "
2225                                 "msg_number=%ld, partid=%d, channel=%d\n",
2226                                 (void *) msg, msg->number, ch->partid,
2227                                 ch->number);
2228                 }
2229
2230                 atomic_dec(&ch->kthreads_active);
2231         }
2232 }
2233
2234
2235 /*
2236  * Now we actually acknowledge the messages that have been delivered and ack'd
2237  * by advancing the cached remote message queue's Get value and if requested
2238  * send an IPI to the message sender's partition.
2239  */
2240 static void
2241 xpc_acknowledge_msgs(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2242 {
2243         struct xpc_msg *msg;
2244         s64 get = initial_get + 1;
2245         int send_IPI = 0;
2246
2247
2248         while (1) {
2249
2250                 while (1) {
2251                         if (get == (volatile s64) ch->w_local_GP.get) {
2252                                 break;
2253                         }
2254
2255                         msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue +
2256                                (get % ch->remote_nentries) * ch->msg_size);
2257
2258                         if (!(msg->flags & XPC_M_DONE)) {
2259                                 break;
2260                         }
2261
2262                         msg_flags |= msg->flags;
2263                         get++;
2264                 }
2265
2266                 if (get == initial_get) {
2267                         /* nothing's changed */
2268                         break;
2269                 }
2270
2271                 if (cmpxchg_rel(&ch->local_GP->get, initial_get, get) !=
2272                                                                 initial_get) {
2273                         /* someone else beat us to it */
2274                         DBUG_ON((volatile s64) ch->local_GP->get <=
2275                                                                 initial_get);
2276                         break;
2277                 }
2278
2279                 /* we just set the new value of local_GP->get */
2280
2281                 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2282                         "channel=%d\n", get, ch->partid, ch->number);
2283
2284                 send_IPI = (msg_flags & XPC_M_INTERRUPT);
2285
2286                 /*
2287                  * We need to ensure that the message referenced by
2288                  * local_GP->get is not XPC_M_DONE or that local_GP->get
2289                  * equals w_local_GP.get, so we'll go have a look.
2290                  */
2291                 initial_get = get;
2292         }
2293
2294         if (send_IPI) {
2295                 xpc_IPI_send_msgrequest(ch);
2296         }
2297 }
2298
2299
2300 /*
2301  * Acknowledge receipt of a delivered message.
2302  *
2303  * If a message has XPC_M_INTERRUPT set, send an interrupt to the partition
2304  * that sent the message.
2305  *
2306  * This function, although called by users, does not call xpc_part_ref() to
2307  * ensure that the partition infrastructure is in place. It relies on the
2308  * fact that we called xpc_msgqueue_ref() in xpc_deliver_msg().
2309  *
2310  * Arguments:
2311  *
2312  *      partid - ID of partition to which the channel is connected.
2313  *      ch_number - channel # message received on.
2314  *      payload - pointer to the payload area allocated via
2315  *                      xpc_initiate_allocate().
2316  */
2317 void
2318 xpc_initiate_received(partid_t partid, int ch_number, void *payload)
2319 {
2320         struct xpc_partition *part = &xpc_partitions[partid];
2321         struct xpc_channel *ch;
2322         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2323         s64 get, msg_number = msg->number;
2324
2325
2326         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
2327         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2328
2329         ch = &part->channels[ch_number];
2330
2331         dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2332                 (void *) msg, msg_number, ch->partid, ch->number);
2333
2334         DBUG_ON((((u64) msg - (u64) ch->remote_msgqueue) / ch->msg_size) !=
2335                                         msg_number % ch->remote_nentries);
2336         DBUG_ON(msg->flags & XPC_M_DONE);
2337
2338         msg->flags |= XPC_M_DONE;
2339
2340         /*
2341          * The preceding store of msg->flags must occur before the following
2342          * load of ch->local_GP->get.
2343          */
2344         mb();
2345
2346         /*
2347          * See if this message is next in line to be acknowledged as having
2348          * been delivered.
2349          */
2350         get = ch->local_GP->get;
2351         if (get == msg_number) {
2352                 xpc_acknowledge_msgs(ch, get, msg->flags);
2353         }
2354
2355         /* the call to xpc_msgqueue_ref() was done by xpc_deliver_msg()  */
2356         xpc_msgqueue_deref(ch);
2357 }
2358