195ac1b8e2620e06c06d9cc89f7e3c50a689c35f
[safe/jmp/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         /* wake the thread that is waiting for this channel to disconnect */
796         if (ch->flags & XPC_C_WDISCONNECT) {
797                 spin_unlock_irqrestore(&ch->lock, *irq_flags);
798                 up(&ch->wdisconnect_sema);
799                 spin_lock_irqsave(&ch->lock, *irq_flags);
800         }
801 }
802
803
804 /*
805  * Process a change in the channel's remote connection state.
806  */
807 static void
808 xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
809                                 u8 IPI_flags)
810 {
811         unsigned long irq_flags;
812         struct xpc_openclose_args *args =
813                                 &part->remote_openclose_args[ch_number];
814         struct xpc_channel *ch = &part->channels[ch_number];
815         enum xpc_retval reason;
816
817
818
819         spin_lock_irqsave(&ch->lock, irq_flags);
820
821
822         if (IPI_flags & XPC_IPI_CLOSEREQUEST) {
823
824                 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREQUEST (reason=%d) received "
825                         "from partid=%d, channel=%d\n", args->reason,
826                         ch->partid, ch->number);
827
828                 /*
829                  * If RCLOSEREQUEST is set, we're probably waiting for
830                  * RCLOSEREPLY. We should find it and a ROPENREQUEST packed
831                  * with this RCLOSEREQUEST in the IPI_flags.
832                  */
833
834                 if (ch->flags & XPC_C_RCLOSEREQUEST) {
835                         DBUG_ON(!(ch->flags & XPC_C_DISCONNECTING));
836                         DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
837                         DBUG_ON(!(ch->flags & XPC_C_CLOSEREPLY));
838                         DBUG_ON(ch->flags & XPC_C_RCLOSEREPLY);
839
840                         DBUG_ON(!(IPI_flags & XPC_IPI_CLOSEREPLY));
841                         IPI_flags &= ~XPC_IPI_CLOSEREPLY;
842                         ch->flags |= XPC_C_RCLOSEREPLY;
843
844                         /* both sides have finished disconnecting */
845                         xpc_process_disconnect(ch, &irq_flags);
846                 }
847
848                 if (ch->flags & XPC_C_DISCONNECTED) {
849                         // >>> explain this section
850
851                         if (!(IPI_flags & XPC_IPI_OPENREQUEST)) {
852                                 DBUG_ON(part->act_state !=
853                                                         XPC_P_DEACTIVATING);
854                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
855                                 return;
856                         }
857
858                         XPC_SET_REASON(ch, 0, 0);
859                         ch->flags &= ~XPC_C_DISCONNECTED;
860
861                         atomic_inc(&part->nchannels_active);
862                         ch->flags |= (XPC_C_CONNECTING | XPC_C_ROPENREQUEST);
863                 }
864
865                 IPI_flags &= ~(XPC_IPI_OPENREQUEST | XPC_IPI_OPENREPLY);
866
867                 /*
868                  * The meaningful CLOSEREQUEST connection state fields are:
869                  *      reason = reason connection is to be closed
870                  */
871
872                 ch->flags |= XPC_C_RCLOSEREQUEST;
873
874                 if (!(ch->flags & XPC_C_DISCONNECTING)) {
875                         reason = args->reason;
876                         if (reason <= xpcSuccess || reason > xpcUnknownReason) {
877                                 reason = xpcUnknownReason;
878                         } else if (reason == xpcUnregistering) {
879                                 reason = xpcOtherUnregistering;
880                         }
881
882                         XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
883                 } else {
884                         xpc_process_disconnect(ch, &irq_flags);
885                 }
886         }
887
888
889         if (IPI_flags & XPC_IPI_CLOSEREPLY) {
890
891                 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREPLY received from partid=%d,"
892                         " channel=%d\n", ch->partid, ch->number);
893
894                 if (ch->flags & XPC_C_DISCONNECTED) {
895                         DBUG_ON(part->act_state != XPC_P_DEACTIVATING);
896                         spin_unlock_irqrestore(&ch->lock, irq_flags);
897                         return;
898                 }
899
900                 DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
901                 DBUG_ON(!(ch->flags & XPC_C_RCLOSEREQUEST));
902
903                 ch->flags |= XPC_C_RCLOSEREPLY;
904
905                 if (ch->flags & XPC_C_CLOSEREPLY) {
906                         /* both sides have finished disconnecting */
907                         xpc_process_disconnect(ch, &irq_flags);
908                 }
909         }
910
911
912         if (IPI_flags & XPC_IPI_OPENREQUEST) {
913
914                 dev_dbg(xpc_chan, "XPC_IPI_OPENREQUEST (msg_size=%d, "
915                         "local_nentries=%d) received from partid=%d, "
916                         "channel=%d\n", args->msg_size, args->local_nentries,
917                         ch->partid, ch->number);
918
919                 if ((ch->flags & (XPC_C_DISCONNECTING | XPC_C_WDISCONNECT)) ||
920                                         part->act_state == XPC_P_DEACTIVATING) {
921                         spin_unlock_irqrestore(&ch->lock, irq_flags);
922                         return;
923                 }
924                 DBUG_ON(!(ch->flags & (XPC_C_DISCONNECTED |
925                                                         XPC_C_OPENREQUEST)));
926                 DBUG_ON(ch->flags & (XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
927                                         XPC_C_OPENREPLY | XPC_C_CONNECTED));
928
929                 /*
930                  * The meaningful OPENREQUEST connection state fields are:
931                  *      msg_size = size of channel's messages in bytes
932                  *      local_nentries = remote partition's local_nentries
933                  */
934                 DBUG_ON(args->msg_size == 0);
935                 DBUG_ON(args->local_nentries == 0);
936
937                 ch->flags |= (XPC_C_ROPENREQUEST | XPC_C_CONNECTING);
938                 ch->remote_nentries = args->local_nentries;
939
940
941                 if (ch->flags & XPC_C_OPENREQUEST) {
942                         if (args->msg_size != ch->msg_size) {
943                                 XPC_DISCONNECT_CHANNEL(ch, xpcUnequalMsgSizes,
944                                                                 &irq_flags);
945                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
946                                 return;
947                         }
948                 } else {
949                         ch->msg_size = args->msg_size;
950
951                         XPC_SET_REASON(ch, 0, 0);
952                         ch->flags &= ~XPC_C_DISCONNECTED;
953
954                         atomic_inc(&part->nchannels_active);
955                 }
956
957                 xpc_process_connect(ch, &irq_flags);
958         }
959
960
961         if (IPI_flags & XPC_IPI_OPENREPLY) {
962
963                 dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY (local_msgqueue_pa=0x%lx, "
964                         "local_nentries=%d, remote_nentries=%d) received from "
965                         "partid=%d, channel=%d\n", args->local_msgqueue_pa,
966                         args->local_nentries, args->remote_nentries,
967                         ch->partid, ch->number);
968
969                 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED)) {
970                         spin_unlock_irqrestore(&ch->lock, irq_flags);
971                         return;
972                 }
973                 DBUG_ON(!(ch->flags & XPC_C_OPENREQUEST));
974                 DBUG_ON(!(ch->flags & XPC_C_ROPENREQUEST));
975                 DBUG_ON(ch->flags & XPC_C_CONNECTED);
976
977                 /*
978                  * The meaningful OPENREPLY connection state fields are:
979                  *      local_msgqueue_pa = physical address of remote
980                  *                          partition's local_msgqueue
981                  *      local_nentries = remote partition's local_nentries
982                  *      remote_nentries = remote partition's remote_nentries
983                  */
984                 DBUG_ON(args->local_msgqueue_pa == 0);
985                 DBUG_ON(args->local_nentries == 0);
986                 DBUG_ON(args->remote_nentries == 0);
987
988                 ch->flags |= XPC_C_ROPENREPLY;
989                 ch->remote_msgqueue_pa = args->local_msgqueue_pa;
990
991                 if (args->local_nentries < ch->remote_nentries) {
992                         dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
993                                 "remote_nentries=%d, old remote_nentries=%d, "
994                                 "partid=%d, channel=%d\n",
995                                 args->local_nentries, ch->remote_nentries,
996                                 ch->partid, ch->number);
997
998                         ch->remote_nentries = args->local_nentries;
999                 }
1000                 if (args->remote_nentries < ch->local_nentries) {
1001                         dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1002                                 "local_nentries=%d, old local_nentries=%d, "
1003                                 "partid=%d, channel=%d\n",
1004                                 args->remote_nentries, ch->local_nentries,
1005                                 ch->partid, ch->number);
1006
1007                         ch->local_nentries = args->remote_nentries;
1008                 }
1009
1010                 xpc_process_connect(ch, &irq_flags);
1011         }
1012
1013         spin_unlock_irqrestore(&ch->lock, irq_flags);
1014 }
1015
1016
1017 /*
1018  * Attempt to establish a channel connection to a remote partition.
1019  */
1020 static enum xpc_retval
1021 xpc_connect_channel(struct xpc_channel *ch)
1022 {
1023         unsigned long irq_flags;
1024         struct xpc_registration *registration = &xpc_registrations[ch->number];
1025
1026
1027         if (down_interruptible(&registration->sema) != 0) {
1028                 return xpcInterrupted;
1029         }
1030
1031         if (!XPC_CHANNEL_REGISTERED(ch->number)) {
1032                 up(&registration->sema);
1033                 return xpcUnregistered;
1034         }
1035
1036         spin_lock_irqsave(&ch->lock, irq_flags);
1037
1038         DBUG_ON(ch->flags & XPC_C_CONNECTED);
1039         DBUG_ON(ch->flags & XPC_C_OPENREQUEST);
1040
1041         if (ch->flags & XPC_C_DISCONNECTING) {
1042                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1043                 up(&registration->sema);
1044                 return ch->reason;
1045         }
1046
1047
1048         /* add info from the channel connect registration to the channel */
1049
1050         ch->kthreads_assigned_limit = registration->assigned_limit;
1051         ch->kthreads_idle_limit = registration->idle_limit;
1052         DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
1053         DBUG_ON(atomic_read(&ch->kthreads_idle) != 0);
1054         DBUG_ON(atomic_read(&ch->kthreads_active) != 0);
1055
1056         ch->func = registration->func;
1057         DBUG_ON(registration->func == NULL);
1058         ch->key = registration->key;
1059
1060         ch->local_nentries = registration->nentries;
1061
1062         if (ch->flags & XPC_C_ROPENREQUEST) {
1063                 if (registration->msg_size != ch->msg_size) {
1064                         /* the local and remote sides aren't the same */
1065
1066                         /*
1067                          * Because XPC_DISCONNECT_CHANNEL() can block we're
1068                          * forced to up the registration sema before we unlock
1069                          * the channel lock. But that's okay here because we're
1070                          * done with the part that required the registration
1071                          * sema. XPC_DISCONNECT_CHANNEL() requires that the
1072                          * channel lock be locked and will unlock and relock
1073                          * the channel lock as needed.
1074                          */
1075                         up(&registration->sema);
1076                         XPC_DISCONNECT_CHANNEL(ch, xpcUnequalMsgSizes,
1077                                                                 &irq_flags);
1078                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1079                         return xpcUnequalMsgSizes;
1080                 }
1081         } else {
1082                 ch->msg_size = registration->msg_size;
1083
1084                 XPC_SET_REASON(ch, 0, 0);
1085                 ch->flags &= ~XPC_C_DISCONNECTED;
1086
1087                 atomic_inc(&xpc_partitions[ch->partid].nchannels_active);
1088         }
1089
1090         up(&registration->sema);
1091
1092
1093         /* initiate the connection */
1094
1095         ch->flags |= (XPC_C_OPENREQUEST | XPC_C_CONNECTING);
1096         xpc_IPI_send_openrequest(ch, &irq_flags);
1097
1098         xpc_process_connect(ch, &irq_flags);
1099
1100         spin_unlock_irqrestore(&ch->lock, irq_flags);
1101
1102         return xpcSuccess;
1103 }
1104
1105
1106 /*
1107  * Clear some of the msg flags in the local message queue.
1108  */
1109 static inline void
1110 xpc_clear_local_msgqueue_flags(struct xpc_channel *ch)
1111 {
1112         struct xpc_msg *msg;
1113         s64 get;
1114
1115
1116         get = ch->w_remote_GP.get;
1117         do {
1118                 msg = (struct xpc_msg *) ((u64) ch->local_msgqueue +
1119                                 (get % ch->local_nentries) * ch->msg_size);
1120                 msg->flags = 0;
1121         } while (++get < (volatile s64) ch->remote_GP.get);
1122 }
1123
1124
1125 /*
1126  * Clear some of the msg flags in the remote message queue.
1127  */
1128 static inline void
1129 xpc_clear_remote_msgqueue_flags(struct xpc_channel *ch)
1130 {
1131         struct xpc_msg *msg;
1132         s64 put;
1133
1134
1135         put = ch->w_remote_GP.put;
1136         do {
1137                 msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue +
1138                                 (put % ch->remote_nentries) * ch->msg_size);
1139                 msg->flags = 0;
1140         } while (++put < (volatile s64) ch->remote_GP.put);
1141 }
1142
1143
1144 static void
1145 xpc_process_msg_IPI(struct xpc_partition *part, int ch_number)
1146 {
1147         struct xpc_channel *ch = &part->channels[ch_number];
1148         int nmsgs_sent;
1149
1150
1151         ch->remote_GP = part->remote_GPs[ch_number];
1152
1153
1154         /* See what, if anything, has changed for each connected channel */
1155
1156         xpc_msgqueue_ref(ch);
1157
1158         if (ch->w_remote_GP.get == ch->remote_GP.get &&
1159                                 ch->w_remote_GP.put == ch->remote_GP.put) {
1160                 /* nothing changed since GPs were last pulled */
1161                 xpc_msgqueue_deref(ch);
1162                 return;
1163         }
1164
1165         if (!(ch->flags & XPC_C_CONNECTED)){
1166                 xpc_msgqueue_deref(ch);
1167                 return;
1168         }
1169
1170
1171         /*
1172          * First check to see if messages recently sent by us have been
1173          * received by the other side. (The remote GET value will have
1174          * changed since we last looked at it.)
1175          */
1176
1177         if (ch->w_remote_GP.get != ch->remote_GP.get) {
1178
1179                 /*
1180                  * We need to notify any senders that want to be notified
1181                  * that their sent messages have been received by their
1182                  * intended recipients. We need to do this before updating
1183                  * w_remote_GP.get so that we don't allocate the same message
1184                  * queue entries prematurely (see xpc_allocate_msg()).
1185                  */
1186                 if (atomic_read(&ch->n_to_notify) > 0) {
1187                         /*
1188                          * Notify senders that messages sent have been
1189                          * received and delivered by the other side.
1190                          */
1191                         xpc_notify_senders(ch, xpcMsgDelivered,
1192                                                         ch->remote_GP.get);
1193                 }
1194
1195                 /*
1196                  * Clear msg->flags in previously sent messages, so that
1197                  * they're ready for xpc_allocate_msg().
1198                  */
1199                 xpc_clear_local_msgqueue_flags(ch);
1200
1201                 ch->w_remote_GP.get = ch->remote_GP.get;
1202
1203                 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1204                         "channel=%d\n", ch->w_remote_GP.get, ch->partid,
1205                         ch->number);
1206
1207                 /*
1208                  * If anyone was waiting for message queue entries to become
1209                  * available, wake them up.
1210                  */
1211                 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0) {
1212                         wake_up(&ch->msg_allocate_wq);
1213                 }
1214         }
1215
1216
1217         /*
1218          * Now check for newly sent messages by the other side. (The remote
1219          * PUT value will have changed since we last looked at it.)
1220          */
1221
1222         if (ch->w_remote_GP.put != ch->remote_GP.put) {
1223                 /*
1224                  * Clear msg->flags in previously received messages, so that
1225                  * they're ready for xpc_get_deliverable_msg().
1226                  */
1227                 xpc_clear_remote_msgqueue_flags(ch);
1228
1229                 ch->w_remote_GP.put = ch->remote_GP.put;
1230
1231                 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1232                         "channel=%d\n", ch->w_remote_GP.put, ch->partid,
1233                         ch->number);
1234
1235                 nmsgs_sent = ch->w_remote_GP.put - ch->w_local_GP.get;
1236                 if (nmsgs_sent > 0) {
1237                         dev_dbg(xpc_chan, "msgs waiting to be copied and "
1238                                 "delivered=%d, partid=%d, channel=%d\n",
1239                                 nmsgs_sent, ch->partid, ch->number);
1240
1241                         if (ch->flags & XPC_C_CONNECTCALLOUT) {
1242                                 xpc_activate_kthreads(ch, nmsgs_sent);
1243                         }
1244                 }
1245         }
1246
1247         xpc_msgqueue_deref(ch);
1248 }
1249
1250
1251 void
1252 xpc_process_channel_activity(struct xpc_partition *part)
1253 {
1254         unsigned long irq_flags;
1255         u64 IPI_amo, IPI_flags;
1256         struct xpc_channel *ch;
1257         int ch_number;
1258         u32 ch_flags;
1259
1260
1261         IPI_amo = xpc_get_IPI_flags(part);
1262
1263         /*
1264          * Initiate channel connections for registered channels.
1265          *
1266          * For each connected channel that has pending messages activate idle
1267          * kthreads and/or create new kthreads as needed.
1268          */
1269
1270         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1271                 ch = &part->channels[ch_number];
1272
1273
1274                 /*
1275                  * Process any open or close related IPI flags, and then deal
1276                  * with connecting or disconnecting the channel as required.
1277                  */
1278
1279                 IPI_flags = XPC_GET_IPI_FLAGS(IPI_amo, ch_number);
1280
1281                 if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_flags)) {
1282                         xpc_process_openclose_IPI(part, ch_number, IPI_flags);
1283                 }
1284
1285                 ch_flags = ch->flags;   /* need an atomic snapshot of flags */
1286
1287                 if (ch_flags & XPC_C_DISCONNECTING) {
1288                         spin_lock_irqsave(&ch->lock, irq_flags);
1289                         xpc_process_disconnect(ch, &irq_flags);
1290                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1291                         continue;
1292                 }
1293
1294                 if (part->act_state == XPC_P_DEACTIVATING) {
1295                         continue;
1296                 }
1297
1298                 if (!(ch_flags & XPC_C_CONNECTED)) {
1299                         if (!(ch_flags & XPC_C_OPENREQUEST)) {
1300                                 DBUG_ON(ch_flags & XPC_C_SETUP);
1301                                 (void) xpc_connect_channel(ch);
1302                         } else {
1303                                 spin_lock_irqsave(&ch->lock, irq_flags);
1304                                 xpc_process_connect(ch, &irq_flags);
1305                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1306                         }
1307                         continue;
1308                 }
1309
1310
1311                 /*
1312                  * Process any message related IPI flags, this may involve the
1313                  * activation of kthreads to deliver any pending messages sent
1314                  * from the other partition.
1315                  */
1316
1317                 if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_flags)) {
1318                         xpc_process_msg_IPI(part, ch_number);
1319                 }
1320         }
1321 }
1322
1323
1324 /*
1325  * XPC's heartbeat code calls this function to inform XPC that a partition is
1326  * going down.  XPC responds by tearing down the XPartition Communication
1327  * infrastructure used for the just downed partition.
1328  *
1329  * XPC's heartbeat code will never call this function and xpc_partition_up()
1330  * at the same time. Nor will it ever make multiple calls to either function
1331  * at the same time.
1332  */
1333 void
1334 xpc_partition_going_down(struct xpc_partition *part, enum xpc_retval reason)
1335 {
1336         unsigned long irq_flags;
1337         int ch_number;
1338         struct xpc_channel *ch;
1339
1340
1341         dev_dbg(xpc_chan, "deactivating partition %d, reason=%d\n",
1342                 XPC_PARTID(part), reason);
1343
1344         if (!xpc_part_ref(part)) {
1345                 /* infrastructure for this partition isn't currently set up */
1346                 return;
1347         }
1348
1349
1350         /* disconnect channels associated with the partition going down */
1351
1352         for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1353                 ch = &part->channels[ch_number];
1354
1355                 xpc_msgqueue_ref(ch);
1356                 spin_lock_irqsave(&ch->lock, irq_flags);
1357
1358                 XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
1359
1360                 spin_unlock_irqrestore(&ch->lock, irq_flags);
1361                 xpc_msgqueue_deref(ch);
1362         }
1363
1364         xpc_wakeup_channel_mgr(part);
1365
1366         xpc_part_deref(part);
1367 }
1368
1369
1370 /*
1371  * Teardown the infrastructure necessary to support XPartition Communication
1372  * between the specified remote partition and the local one.
1373  */
1374 void
1375 xpc_teardown_infrastructure(struct xpc_partition *part)
1376 {
1377         partid_t partid = XPC_PARTID(part);
1378
1379
1380         /*
1381          * We start off by making this partition inaccessible to local
1382          * processes by marking it as no longer setup. Then we make it
1383          * inaccessible to remote processes by clearing the XPC per partition
1384          * specific variable's magic # (which indicates that these variables
1385          * are no longer valid) and by ignoring all XPC notify IPIs sent to
1386          * this partition.
1387          */
1388
1389         DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
1390         DBUG_ON(atomic_read(&part->nchannels_active) != 0);
1391         DBUG_ON(part->setup_state != XPC_P_SETUP);
1392         part->setup_state = XPC_P_WTEARDOWN;
1393
1394         xpc_vars_part[partid].magic = 0;
1395
1396
1397         free_irq(SGI_XPC_NOTIFY, (void *) (u64) partid);
1398
1399
1400         /*
1401          * Before proceding with the teardown we have to wait until all
1402          * existing references cease.
1403          */
1404         wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1405
1406
1407         /* now we can begin tearing down the infrastructure */
1408
1409         part->setup_state = XPC_P_TORNDOWN;
1410
1411         /* in case we've still got outstanding timers registered... */
1412         del_timer_sync(&part->dropped_IPI_timer);
1413
1414         kfree(part->remote_openclose_args_base);
1415         part->remote_openclose_args = NULL;
1416         kfree(part->local_openclose_args_base);
1417         part->local_openclose_args = NULL;
1418         kfree(part->remote_GPs_base);
1419         part->remote_GPs = NULL;
1420         kfree(part->local_GPs_base);
1421         part->local_GPs = NULL;
1422         kfree(part->channels);
1423         part->channels = NULL;
1424         part->local_IPI_amo_va = NULL;
1425 }
1426
1427
1428 /*
1429  * Called by XP at the time of channel connection registration to cause
1430  * XPC to establish connections to all currently active partitions.
1431  */
1432 void
1433 xpc_initiate_connect(int ch_number)
1434 {
1435         partid_t partid;
1436         struct xpc_partition *part;
1437         struct xpc_channel *ch;
1438
1439
1440         DBUG_ON(ch_number < 0 || ch_number >= XPC_NCHANNELS);
1441
1442         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1443                 part = &xpc_partitions[partid];
1444
1445                 if (xpc_part_ref(part)) {
1446                         ch = &part->channels[ch_number];
1447
1448                         if (!(ch->flags & XPC_C_DISCONNECTING)) {
1449                                 DBUG_ON(ch->flags & XPC_C_OPENREQUEST);
1450                                 DBUG_ON(ch->flags & XPC_C_CONNECTED);
1451                                 DBUG_ON(ch->flags & XPC_C_SETUP);
1452
1453                                 /*
1454                                  * Initiate the establishment of a connection
1455                                  * on the newly registered channel to the
1456                                  * remote partition.
1457                                  */
1458                                 xpc_wakeup_channel_mgr(part);
1459                         }
1460
1461                         xpc_part_deref(part);
1462                 }
1463         }
1464 }
1465
1466
1467 void
1468 xpc_connected_callout(struct xpc_channel *ch)
1469 {
1470         unsigned long irq_flags;
1471
1472
1473         /* let the registerer know that a connection has been established */
1474
1475         if (ch->func != NULL) {
1476                 dev_dbg(xpc_chan, "ch->func() called, reason=xpcConnected, "
1477                         "partid=%d, channel=%d\n", ch->partid, ch->number);
1478
1479                 ch->func(xpcConnected, ch->partid, ch->number,
1480                                 (void *) (u64) ch->local_nentries, ch->key);
1481
1482                 dev_dbg(xpc_chan, "ch->func() returned, reason=xpcConnected, "
1483                         "partid=%d, channel=%d\n", ch->partid, ch->number);
1484         }
1485
1486         spin_lock_irqsave(&ch->lock, irq_flags);
1487         ch->flags |= XPC_C_CONNECTCALLOUT;
1488         spin_unlock_irqrestore(&ch->lock, irq_flags);
1489 }
1490
1491
1492 /*
1493  * Called by XP at the time of channel connection unregistration to cause
1494  * XPC to teardown all current connections for the specified channel.
1495  *
1496  * Before returning xpc_initiate_disconnect() will wait until all connections
1497  * on the specified channel have been closed/torndown. So the caller can be
1498  * assured that they will not be receiving any more callouts from XPC to the
1499  * function they registered via xpc_connect().
1500  *
1501  * Arguments:
1502  *
1503  *      ch_number - channel # to unregister.
1504  */
1505 void
1506 xpc_initiate_disconnect(int ch_number)
1507 {
1508         unsigned long irq_flags;
1509         partid_t partid;
1510         struct xpc_partition *part;
1511         struct xpc_channel *ch;
1512
1513
1514         DBUG_ON(ch_number < 0 || ch_number >= XPC_NCHANNELS);
1515
1516         /* initiate the channel disconnect for every active partition */
1517         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1518                 part = &xpc_partitions[partid];
1519
1520                 if (xpc_part_ref(part)) {
1521                         ch = &part->channels[ch_number];
1522                         xpc_msgqueue_ref(ch);
1523
1524                         spin_lock_irqsave(&ch->lock, irq_flags);
1525
1526                         if (!(ch->flags & XPC_C_DISCONNECTED)) {
1527                                 ch->flags |= XPC_C_WDISCONNECT;
1528
1529                                 XPC_DISCONNECT_CHANNEL(ch, xpcUnregistering,
1530                                                                 &irq_flags);
1531                         }
1532
1533                         spin_unlock_irqrestore(&ch->lock, irq_flags);
1534
1535                         xpc_msgqueue_deref(ch);
1536                         xpc_part_deref(part);
1537                 }
1538         }
1539
1540         xpc_disconnect_wait(ch_number);
1541 }
1542
1543
1544 /*
1545  * To disconnect a channel, and reflect it back to all who may be waiting.
1546  *
1547  * An OPEN is not allowed until XPC_C_DISCONNECTING is cleared by
1548  * xpc_process_disconnect(), and if set, XPC_C_WDISCONNECT is cleared by
1549  * xpc_disconnect_wait().
1550  *
1551  * THE CHANNEL IS TO BE LOCKED BY THE CALLER AND WILL REMAIN LOCKED UPON RETURN.
1552  */
1553 void
1554 xpc_disconnect_channel(const int line, struct xpc_channel *ch,
1555                         enum xpc_retval reason, unsigned long *irq_flags)
1556 {
1557         u32 channel_was_connected = (ch->flags & XPC_C_CONNECTED);
1558
1559
1560         DBUG_ON(!spin_is_locked(&ch->lock));
1561
1562         if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED)) {
1563                 return;
1564         }
1565         DBUG_ON(!(ch->flags & (XPC_C_CONNECTING | XPC_C_CONNECTED)));
1566
1567         dev_dbg(xpc_chan, "reason=%d, line=%d, partid=%d, channel=%d\n",
1568                 reason, line, ch->partid, ch->number);
1569
1570         XPC_SET_REASON(ch, reason, line);
1571
1572         ch->flags |= (XPC_C_CLOSEREQUEST | XPC_C_DISCONNECTING);
1573         /* some of these may not have been set */
1574         ch->flags &= ~(XPC_C_OPENREQUEST | XPC_C_OPENREPLY |
1575                         XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
1576                         XPC_C_CONNECTING | XPC_C_CONNECTED);
1577
1578         xpc_IPI_send_closerequest(ch, irq_flags);
1579
1580         if (channel_was_connected) {
1581                 ch->flags |= XPC_C_WASCONNECTED;
1582         }
1583
1584         spin_unlock_irqrestore(&ch->lock, *irq_flags);
1585
1586         /* wake all idle kthreads so they can exit */
1587         if (atomic_read(&ch->kthreads_idle) > 0) {
1588                 wake_up_all(&ch->idle_wq);
1589         }
1590
1591         /* wake those waiting to allocate an entry from the local msg queue */
1592         if (atomic_read(&ch->n_on_msg_allocate_wq) > 0) {
1593                 wake_up(&ch->msg_allocate_wq);
1594         }
1595
1596         spin_lock_irqsave(&ch->lock, *irq_flags);
1597 }
1598
1599
1600 void
1601 xpc_disconnecting_callout(struct xpc_channel *ch)
1602 {
1603         /*
1604          * Let the channel's registerer know that the channel is being
1605          * disconnected. We don't want to do this if the registerer was never
1606          * informed of a connection being made.
1607          */
1608
1609         if (ch->func != NULL) {
1610                 dev_dbg(xpc_chan, "ch->func() called, reason=xpcDisconnecting,"
1611                         " partid=%d, channel=%d\n", ch->partid, ch->number);
1612
1613                 ch->func(xpcDisconnecting, ch->partid, ch->number, NULL,
1614                                                                 ch->key);
1615
1616                 dev_dbg(xpc_chan, "ch->func() returned, reason="
1617                         "xpcDisconnecting, partid=%d, channel=%d\n",
1618                         ch->partid, ch->number);
1619         }
1620 }
1621
1622
1623 /*
1624  * Wait for a message entry to become available for the specified channel,
1625  * but don't wait any longer than 1 jiffy.
1626  */
1627 static enum xpc_retval
1628 xpc_allocate_msg_wait(struct xpc_channel *ch)
1629 {
1630         enum xpc_retval ret;
1631
1632
1633         if (ch->flags & XPC_C_DISCONNECTING) {
1634                 DBUG_ON(ch->reason == xpcInterrupted);  // >>> Is this true?
1635                 return ch->reason;
1636         }
1637
1638         atomic_inc(&ch->n_on_msg_allocate_wq);
1639         ret = interruptible_sleep_on_timeout(&ch->msg_allocate_wq, 1);
1640         atomic_dec(&ch->n_on_msg_allocate_wq);
1641
1642         if (ch->flags & XPC_C_DISCONNECTING) {
1643                 ret = ch->reason;
1644                 DBUG_ON(ch->reason == xpcInterrupted);  // >>> Is this true?
1645         } else if (ret == 0) {
1646                 ret = xpcTimeout;
1647         } else {
1648                 ret = xpcInterrupted;
1649         }
1650
1651         return ret;
1652 }
1653
1654
1655 /*
1656  * Allocate an entry for a message from the message queue associated with the
1657  * specified channel.
1658  */
1659 static enum xpc_retval
1660 xpc_allocate_msg(struct xpc_channel *ch, u32 flags,
1661                         struct xpc_msg **address_of_msg)
1662 {
1663         struct xpc_msg *msg;
1664         enum xpc_retval ret;
1665         s64 put;
1666
1667
1668         /* this reference will be dropped in xpc_send_msg() */
1669         xpc_msgqueue_ref(ch);
1670
1671         if (ch->flags & XPC_C_DISCONNECTING) {
1672                 xpc_msgqueue_deref(ch);
1673                 return ch->reason;
1674         }
1675         if (!(ch->flags & XPC_C_CONNECTED)) {
1676                 xpc_msgqueue_deref(ch);
1677                 return xpcNotConnected;
1678         }
1679
1680
1681         /*
1682          * Get the next available message entry from the local message queue.
1683          * If none are available, we'll make sure that we grab the latest
1684          * GP values.
1685          */
1686         ret = xpcTimeout;
1687
1688         while (1) {
1689
1690                 put = (volatile s64) ch->w_local_GP.put;
1691                 if (put - (volatile s64) ch->w_remote_GP.get <
1692                                                         ch->local_nentries) {
1693
1694                         /* There are available message entries. We need to try
1695                          * to secure one for ourselves. We'll do this by trying
1696                          * to increment w_local_GP.put as long as someone else
1697                          * doesn't beat us to it. If they do, we'll have to
1698                          * try again.
1699                          */
1700                         if (cmpxchg(&ch->w_local_GP.put, put, put + 1) ==
1701                                                                         put) {
1702                                 /* we got the entry referenced by put */
1703                                 break;
1704                         }
1705                         continue;       /* try again */
1706                 }
1707
1708
1709                 /*
1710                  * There aren't any available msg entries at this time.
1711                  *
1712                  * In waiting for a message entry to become available,
1713                  * we set a timeout in case the other side is not
1714                  * sending completion IPIs. This lets us fake an IPI
1715                  * that will cause the IPI handler to fetch the latest
1716                  * GP values as if an IPI was sent by the other side.
1717                  */
1718                 if (ret == xpcTimeout) {
1719                         xpc_IPI_send_local_msgrequest(ch);
1720                 }
1721
1722                 if (flags & XPC_NOWAIT) {
1723                         xpc_msgqueue_deref(ch);
1724                         return xpcNoWait;
1725                 }
1726
1727                 ret = xpc_allocate_msg_wait(ch);
1728                 if (ret != xpcInterrupted && ret != xpcTimeout) {
1729                         xpc_msgqueue_deref(ch);
1730                         return ret;
1731                 }
1732         }
1733
1734
1735         /* get the message's address and initialize it */
1736         msg = (struct xpc_msg *) ((u64) ch->local_msgqueue +
1737                                 (put % ch->local_nentries) * ch->msg_size);
1738
1739
1740         DBUG_ON(msg->flags != 0);
1741         msg->number = put;
1742
1743         dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
1744                 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
1745                 (void *) msg, msg->number, ch->partid, ch->number);
1746
1747         *address_of_msg = msg;
1748
1749         return xpcSuccess;
1750 }
1751
1752
1753 /*
1754  * Allocate an entry for a message from the message queue associated with the
1755  * specified channel. NOTE that this routine can sleep waiting for a message
1756  * entry to become available. To not sleep, pass in the XPC_NOWAIT flag.
1757  *
1758  * Arguments:
1759  *
1760  *      partid - ID of partition to which the channel is connected.
1761  *      ch_number - channel #.
1762  *      flags - see xpc.h for valid flags.
1763  *      payload - address of the allocated payload area pointer (filled in on
1764  *                return) in which the user-defined message is constructed.
1765  */
1766 enum xpc_retval
1767 xpc_initiate_allocate(partid_t partid, int ch_number, u32 flags, void **payload)
1768 {
1769         struct xpc_partition *part = &xpc_partitions[partid];
1770         enum xpc_retval ret = xpcUnknownReason;
1771         struct xpc_msg *msg;
1772
1773
1774         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
1775         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1776
1777         *payload = NULL;
1778
1779         if (xpc_part_ref(part)) {
1780                 ret = xpc_allocate_msg(&part->channels[ch_number], flags, &msg);
1781                 xpc_part_deref(part);
1782
1783                 if (msg != NULL) {
1784                         *payload = &msg->payload;
1785                 }
1786         }
1787
1788         return ret;
1789 }
1790
1791
1792 /*
1793  * Now we actually send the messages that are ready to be sent by advancing
1794  * the local message queue's Put value and then send an IPI to the recipient
1795  * partition.
1796  */
1797 static void
1798 xpc_send_msgs(struct xpc_channel *ch, s64 initial_put)
1799 {
1800         struct xpc_msg *msg;
1801         s64 put = initial_put + 1;
1802         int send_IPI = 0;
1803
1804
1805         while (1) {
1806
1807                 while (1) {
1808                         if (put == (volatile s64) ch->w_local_GP.put) {
1809                                 break;
1810                         }
1811
1812                         msg = (struct xpc_msg *) ((u64) ch->local_msgqueue +
1813                                (put % ch->local_nentries) * ch->msg_size);
1814
1815                         if (!(msg->flags & XPC_M_READY)) {
1816                                 break;
1817                         }
1818
1819                         put++;
1820                 }
1821
1822                 if (put == initial_put) {
1823                         /* nothing's changed */
1824                         break;
1825                 }
1826
1827                 if (cmpxchg_rel(&ch->local_GP->put, initial_put, put) !=
1828                                                                 initial_put) {
1829                         /* someone else beat us to it */
1830                         DBUG_ON((volatile s64) ch->local_GP->put < initial_put);
1831                         break;
1832                 }
1833
1834                 /* we just set the new value of local_GP->put */
1835
1836                 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
1837                         "channel=%d\n", put, ch->partid, ch->number);
1838
1839                 send_IPI = 1;
1840
1841                 /*
1842                  * We need to ensure that the message referenced by
1843                  * local_GP->put is not XPC_M_READY or that local_GP->put
1844                  * equals w_local_GP.put, so we'll go have a look.
1845                  */
1846                 initial_put = put;
1847         }
1848
1849         if (send_IPI) {
1850                 xpc_IPI_send_msgrequest(ch);
1851         }
1852 }
1853
1854
1855 /*
1856  * Common code that does the actual sending of the message by advancing the
1857  * local message queue's Put value and sends an IPI to the partition the
1858  * message is being sent to.
1859  */
1860 static enum xpc_retval
1861 xpc_send_msg(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type,
1862                         xpc_notify_func func, void *key)
1863 {
1864         enum xpc_retval ret = xpcSuccess;
1865         struct xpc_notify *notify = notify;
1866         s64 put, msg_number = msg->number;
1867
1868
1869         DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
1870         DBUG_ON((((u64) msg - (u64) ch->local_msgqueue) / ch->msg_size) !=
1871                                         msg_number % ch->local_nentries);
1872         DBUG_ON(msg->flags & XPC_M_READY);
1873
1874         if (ch->flags & XPC_C_DISCONNECTING) {
1875                 /* drop the reference grabbed in xpc_allocate_msg() */
1876                 xpc_msgqueue_deref(ch);
1877                 return ch->reason;
1878         }
1879
1880         if (notify_type != 0) {
1881                 /*
1882                  * Tell the remote side to send an ACK interrupt when the
1883                  * message has been delivered.
1884                  */
1885                 msg->flags |= XPC_M_INTERRUPT;
1886
1887                 atomic_inc(&ch->n_to_notify);
1888
1889                 notify = &ch->notify_queue[msg_number % ch->local_nentries];
1890                 notify->func = func;
1891                 notify->key = key;
1892                 notify->type = notify_type;
1893
1894                 // >>> is a mb() needed here?
1895
1896                 if (ch->flags & XPC_C_DISCONNECTING) {
1897                         /*
1898                          * An error occurred between our last error check and
1899                          * this one. We will try to clear the type field from
1900                          * the notify entry. If we succeed then
1901                          * xpc_disconnect_channel() didn't already process
1902                          * the notify entry.
1903                          */
1904                         if (cmpxchg(&notify->type, notify_type, 0) ==
1905                                                                 notify_type) {
1906                                 atomic_dec(&ch->n_to_notify);
1907                                 ret = ch->reason;
1908                         }
1909
1910                         /* drop the reference grabbed in xpc_allocate_msg() */
1911                         xpc_msgqueue_deref(ch);
1912                         return ret;
1913                 }
1914         }
1915
1916         msg->flags |= XPC_M_READY;
1917
1918         /*
1919          * The preceding store of msg->flags must occur before the following
1920          * load of ch->local_GP->put.
1921          */
1922         mb();
1923
1924         /* see if the message is next in line to be sent, if so send it */
1925
1926         put = ch->local_GP->put;
1927         if (put == msg_number) {
1928                 xpc_send_msgs(ch, put);
1929         }
1930
1931         /* drop the reference grabbed in xpc_allocate_msg() */
1932         xpc_msgqueue_deref(ch);
1933         return ret;
1934 }
1935
1936
1937 /*
1938  * Send a message previously allocated using xpc_initiate_allocate() on the
1939  * specified channel connected to the specified partition.
1940  *
1941  * This routine will not wait for the message to be received, nor will
1942  * notification be given when it does happen. Once this routine has returned
1943  * the message entry allocated via xpc_initiate_allocate() is no longer
1944  * accessable to the caller.
1945  *
1946  * This routine, although called by users, does not call xpc_part_ref() to
1947  * ensure that the partition infrastructure is in place. It relies on the
1948  * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
1949  *
1950  * Arguments:
1951  *
1952  *      partid - ID of partition to which the channel is connected.
1953  *      ch_number - channel # to send message on.
1954  *      payload - pointer to the payload area allocated via
1955  *                      xpc_initiate_allocate().
1956  */
1957 enum xpc_retval
1958 xpc_initiate_send(partid_t partid, int ch_number, void *payload)
1959 {
1960         struct xpc_partition *part = &xpc_partitions[partid];
1961         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
1962         enum xpc_retval ret;
1963
1964
1965         dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *) msg,
1966                 partid, ch_number);
1967
1968         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
1969         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1970         DBUG_ON(msg == NULL);
1971
1972         ret = xpc_send_msg(&part->channels[ch_number], msg, 0, NULL, NULL);
1973
1974         return ret;
1975 }
1976
1977
1978 /*
1979  * Send a message previously allocated using xpc_initiate_allocate on the
1980  * specified channel connected to the specified partition.
1981  *
1982  * This routine will not wait for the message to be sent. Once this routine
1983  * has returned the message entry allocated via xpc_initiate_allocate() is no
1984  * longer accessable to the caller.
1985  *
1986  * Once the remote end of the channel has received the message, the function
1987  * passed as an argument to xpc_initiate_send_notify() will be called. This
1988  * allows the sender to free up or re-use any buffers referenced by the
1989  * message, but does NOT mean the message has been processed at the remote
1990  * end by a receiver.
1991  *
1992  * If this routine returns an error, the caller's function will NOT be called.
1993  *
1994  * This routine, although called by users, does not call xpc_part_ref() to
1995  * ensure that the partition infrastructure is in place. It relies on the
1996  * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
1997  *
1998  * Arguments:
1999  *
2000  *      partid - ID of partition to which the channel is connected.
2001  *      ch_number - channel # to send message on.
2002  *      payload - pointer to the payload area allocated via
2003  *                      xpc_initiate_allocate().
2004  *      func - function to call with asynchronous notification of message
2005  *                receipt. THIS FUNCTION MUST BE NON-BLOCKING.
2006  *      key - user-defined key to be passed to the function when it's called.
2007  */
2008 enum xpc_retval
2009 xpc_initiate_send_notify(partid_t partid, int ch_number, void *payload,
2010                                 xpc_notify_func func, void *key)
2011 {
2012         struct xpc_partition *part = &xpc_partitions[partid];
2013         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2014         enum xpc_retval ret;
2015
2016
2017         dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *) msg,
2018                 partid, ch_number);
2019
2020         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
2021         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2022         DBUG_ON(msg == NULL);
2023         DBUG_ON(func == NULL);
2024
2025         ret = xpc_send_msg(&part->channels[ch_number], msg, XPC_N_CALL,
2026                                                                 func, key);
2027         return ret;
2028 }
2029
2030
2031 static struct xpc_msg *
2032 xpc_pull_remote_msg(struct xpc_channel *ch, s64 get)
2033 {
2034         struct xpc_partition *part = &xpc_partitions[ch->partid];
2035         struct xpc_msg *remote_msg, *msg;
2036         u32 msg_index, nmsgs;
2037         u64 msg_offset;
2038         enum xpc_retval ret;
2039
2040
2041         if (down_interruptible(&ch->msg_to_pull_sema) != 0) {
2042                 /* we were interrupted by a signal */
2043                 return NULL;
2044         }
2045
2046         while (get >= ch->next_msg_to_pull) {
2047
2048                 /* pull as many messages as are ready and able to be pulled */
2049
2050                 msg_index = ch->next_msg_to_pull % ch->remote_nentries;
2051
2052                 DBUG_ON(ch->next_msg_to_pull >=
2053                                         (volatile s64) ch->w_remote_GP.put);
2054                 nmsgs =  (volatile s64) ch->w_remote_GP.put -
2055                                                 ch->next_msg_to_pull;
2056                 if (msg_index + nmsgs > ch->remote_nentries) {
2057                         /* ignore the ones that wrap the msg queue for now */
2058                         nmsgs = ch->remote_nentries - msg_index;
2059                 }
2060
2061                 msg_offset = msg_index * ch->msg_size;
2062                 msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue +
2063                                                                 msg_offset);
2064                 remote_msg = (struct xpc_msg *) (ch->remote_msgqueue_pa +
2065                                                                 msg_offset);
2066
2067                 if ((ret = xpc_pull_remote_cachelines(part, msg, remote_msg,
2068                                 nmsgs * ch->msg_size)) != xpcSuccess) {
2069
2070                         dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
2071                                 " msg %ld from partition %d, channel=%d, "
2072                                 "ret=%d\n", nmsgs, ch->next_msg_to_pull,
2073                                 ch->partid, ch->number, ret);
2074
2075                         XPC_DEACTIVATE_PARTITION(part, ret);
2076
2077                         up(&ch->msg_to_pull_sema);
2078                         return NULL;
2079                 }
2080
2081                 mb();   /* >>> this may not be needed, we're not sure */
2082
2083                 ch->next_msg_to_pull += nmsgs;
2084         }
2085
2086         up(&ch->msg_to_pull_sema);
2087
2088         /* return the message we were looking for */
2089         msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2090         msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue + msg_offset);
2091
2092         return msg;
2093 }
2094
2095
2096 /*
2097  * Get a message to be delivered.
2098  */
2099 static struct xpc_msg *
2100 xpc_get_deliverable_msg(struct xpc_channel *ch)
2101 {
2102         struct xpc_msg *msg = NULL;
2103         s64 get;
2104
2105
2106         do {
2107                 if ((volatile u32) ch->flags & XPC_C_DISCONNECTING) {
2108                         break;
2109                 }
2110
2111                 get = (volatile s64) ch->w_local_GP.get;
2112                 if (get == (volatile s64) ch->w_remote_GP.put) {
2113                         break;
2114                 }
2115
2116                 /* There are messages waiting to be pulled and delivered.
2117                  * We need to try to secure one for ourselves. We'll do this
2118                  * by trying to increment w_local_GP.get and hope that no one
2119                  * else beats us to it. If they do, we'll we'll simply have
2120                  * to try again for the next one.
2121                  */
2122
2123                 if (cmpxchg(&ch->w_local_GP.get, get, get + 1) == get) {
2124                         /* we got the entry referenced by get */
2125
2126                         dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2127                                 "partid=%d, channel=%d\n", get + 1,
2128                                 ch->partid, ch->number);
2129
2130                         /* pull the message from the remote partition */
2131
2132                         msg = xpc_pull_remote_msg(ch, get);
2133
2134                         DBUG_ON(msg != NULL && msg->number != get);
2135                         DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2136                         DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2137
2138                         break;
2139                 }
2140
2141         } while (1);
2142
2143         return msg;
2144 }
2145
2146
2147 /*
2148  * Deliver a message to its intended recipient.
2149  */
2150 void
2151 xpc_deliver_msg(struct xpc_channel *ch)
2152 {
2153         struct xpc_msg *msg;
2154
2155
2156         if ((msg = xpc_get_deliverable_msg(ch)) != NULL) {
2157
2158                 /*
2159                  * This ref is taken to protect the payload itself from being
2160                  * freed before the user is finished with it, which the user
2161                  * indicates by calling xpc_initiate_received().
2162                  */
2163                 xpc_msgqueue_ref(ch);
2164
2165                 atomic_inc(&ch->kthreads_active);
2166
2167                 if (ch->func != NULL) {
2168                         dev_dbg(xpc_chan, "ch->func() called, msg=0x%p, "
2169                                 "msg_number=%ld, partid=%d, channel=%d\n",
2170                                 (void *) msg, msg->number, ch->partid,
2171                                 ch->number);
2172
2173                         /* deliver the message to its intended recipient */
2174                         ch->func(xpcMsgReceived, ch->partid, ch->number,
2175                                         &msg->payload, ch->key);
2176
2177                         dev_dbg(xpc_chan, "ch->func() returned, msg=0x%p, "
2178                                 "msg_number=%ld, partid=%d, channel=%d\n",
2179                                 (void *) msg, msg->number, ch->partid,
2180                                 ch->number);
2181                 }
2182
2183                 atomic_dec(&ch->kthreads_active);
2184         }
2185 }
2186
2187
2188 /*
2189  * Now we actually acknowledge the messages that have been delivered and ack'd
2190  * by advancing the cached remote message queue's Get value and if requested
2191  * send an IPI to the message sender's partition.
2192  */
2193 static void
2194 xpc_acknowledge_msgs(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2195 {
2196         struct xpc_msg *msg;
2197         s64 get = initial_get + 1;
2198         int send_IPI = 0;
2199
2200
2201         while (1) {
2202
2203                 while (1) {
2204                         if (get == (volatile s64) ch->w_local_GP.get) {
2205                                 break;
2206                         }
2207
2208                         msg = (struct xpc_msg *) ((u64) ch->remote_msgqueue +
2209                                (get % ch->remote_nentries) * ch->msg_size);
2210
2211                         if (!(msg->flags & XPC_M_DONE)) {
2212                                 break;
2213                         }
2214
2215                         msg_flags |= msg->flags;
2216                         get++;
2217                 }
2218
2219                 if (get == initial_get) {
2220                         /* nothing's changed */
2221                         break;
2222                 }
2223
2224                 if (cmpxchg_rel(&ch->local_GP->get, initial_get, get) !=
2225                                                                 initial_get) {
2226                         /* someone else beat us to it */
2227                         DBUG_ON((volatile s64) ch->local_GP->get <=
2228                                                                 initial_get);
2229                         break;
2230                 }
2231
2232                 /* we just set the new value of local_GP->get */
2233
2234                 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2235                         "channel=%d\n", get, ch->partid, ch->number);
2236
2237                 send_IPI = (msg_flags & XPC_M_INTERRUPT);
2238
2239                 /*
2240                  * We need to ensure that the message referenced by
2241                  * local_GP->get is not XPC_M_DONE or that local_GP->get
2242                  * equals w_local_GP.get, so we'll go have a look.
2243                  */
2244                 initial_get = get;
2245         }
2246
2247         if (send_IPI) {
2248                 xpc_IPI_send_msgrequest(ch);
2249         }
2250 }
2251
2252
2253 /*
2254  * Acknowledge receipt of a delivered message.
2255  *
2256  * If a message has XPC_M_INTERRUPT set, send an interrupt to the partition
2257  * that sent the message.
2258  *
2259  * This function, although called by users, does not call xpc_part_ref() to
2260  * ensure that the partition infrastructure is in place. It relies on the
2261  * fact that we called xpc_msgqueue_ref() in xpc_deliver_msg().
2262  *
2263  * Arguments:
2264  *
2265  *      partid - ID of partition to which the channel is connected.
2266  *      ch_number - channel # message received on.
2267  *      payload - pointer to the payload area allocated via
2268  *                      xpc_initiate_allocate().
2269  */
2270 void
2271 xpc_initiate_received(partid_t partid, int ch_number, void *payload)
2272 {
2273         struct xpc_partition *part = &xpc_partitions[partid];
2274         struct xpc_channel *ch;
2275         struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2276         s64 get, msg_number = msg->number;
2277
2278
2279         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
2280         DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2281
2282         ch = &part->channels[ch_number];
2283
2284         dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2285                 (void *) msg, msg_number, ch->partid, ch->number);
2286
2287         DBUG_ON((((u64) msg - (u64) ch->remote_msgqueue) / ch->msg_size) !=
2288                                         msg_number % ch->remote_nentries);
2289         DBUG_ON(msg->flags & XPC_M_DONE);
2290
2291         msg->flags |= XPC_M_DONE;
2292
2293         /*
2294          * The preceding store of msg->flags must occur before the following
2295          * load of ch->local_GP->get.
2296          */
2297         mb();
2298
2299         /*
2300          * See if this message is next in line to be acknowledged as having
2301          * been delivered.
2302          */
2303         get = ch->local_GP->get;
2304         if (get == msg_number) {
2305                 xpc_acknowledge_msgs(ch, get, msg->flags);
2306         }
2307
2308         /* the call to xpc_msgqueue_ref() was done by xpc_deliver_msg()  */
2309         xpc_msgqueue_deref(ch);
2310 }
2311