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
6 * Copyright (c) 2004-2008 Silicon Graphics, Inc. All Rights Reserved.
10 * Cross Partition Communication (XPC) channel support.
12 * This is the part of XPC that manages the channels and
13 * sends/receives messages across them to/from other partitions.
17 #include <linux/kernel.h>
18 #include <linux/init.h>
19 #include <linux/sched.h>
20 #include <linux/cache.h>
21 #include <linux/interrupt.h>
22 #include <linux/mutex.h>
23 #include <linux/completion.h>
24 #include <asm/sn/bte.h>
25 #include <asm/sn/sn_sal.h>
29 * Guarantee that the kzalloc'd memory is cacheline aligned.
32 xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
34 /* see if kzalloc will give us cachline aligned memory by default */
35 *base = kzalloc(size, flags);
39 if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
44 /* nope, we'll have to do it ourselves */
45 *base = kzalloc(size + L1_CACHE_BYTES, flags);
49 return (void *)L1_CACHE_ALIGN((u64)*base);
53 * Set up the initial values for the XPartition Communication channels.
56 xpc_initialize_channels(struct xpc_partition *part, partid_t partid)
59 struct xpc_channel *ch;
61 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
62 ch = &part->channels[ch_number];
65 ch->number = ch_number;
66 ch->flags = XPC_C_DISCONNECTED;
68 ch->local_GP = &part->local_GPs[ch_number];
69 ch->local_openclose_args =
70 &part->local_openclose_args[ch_number];
72 atomic_set(&ch->kthreads_assigned, 0);
73 atomic_set(&ch->kthreads_idle, 0);
74 atomic_set(&ch->kthreads_active, 0);
76 atomic_set(&ch->references, 0);
77 atomic_set(&ch->n_to_notify, 0);
79 spin_lock_init(&ch->lock);
80 mutex_init(&ch->msg_to_pull_mutex);
81 init_completion(&ch->wdisconnect_wait);
83 atomic_set(&ch->n_on_msg_allocate_wq, 0);
84 init_waitqueue_head(&ch->msg_allocate_wq);
85 init_waitqueue_head(&ch->idle_wq);
90 * Setup the infrastructure necessary to support XPartition Communication
91 * between the specified remote partition and the local one.
94 xpc_setup_infrastructure(struct xpc_partition *part)
97 struct timer_list *timer;
98 partid_t partid = XPC_PARTID(part);
101 * Zero out MOST of the entry for this partition. Only the fields
102 * starting with `nchannels' will be zeroed. The preceding fields must
103 * remain `viable' across partition ups and downs, since they may be
104 * referenced during this memset() operation.
106 memset(&part->nchannels, 0, sizeof(struct xpc_partition) -
107 offsetof(struct xpc_partition, nchannels));
110 * Allocate all of the channel structures as a contiguous chunk of
113 part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_NCHANNELS,
115 if (part->channels == NULL) {
116 dev_err(xpc_chan, "can't get memory for channels\n");
120 part->nchannels = XPC_NCHANNELS;
122 /* allocate all the required GET/PUT values */
124 part->local_GPs = xpc_kzalloc_cacheline_aligned(XPC_GP_SIZE,
126 &part->local_GPs_base);
127 if (part->local_GPs == NULL) {
128 kfree(part->channels);
129 part->channels = NULL;
130 dev_err(xpc_chan, "can't get memory for local get/put "
135 part->remote_GPs = xpc_kzalloc_cacheline_aligned(XPC_GP_SIZE,
139 if (part->remote_GPs == NULL) {
140 dev_err(xpc_chan, "can't get memory for remote get/put "
142 kfree(part->local_GPs_base);
143 part->local_GPs = NULL;
144 kfree(part->channels);
145 part->channels = NULL;
149 /* allocate all the required open and close args */
151 part->local_openclose_args =
152 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
153 &part->local_openclose_args_base);
154 if (part->local_openclose_args == NULL) {
155 dev_err(xpc_chan, "can't get memory for local connect args\n");
156 kfree(part->remote_GPs_base);
157 part->remote_GPs = NULL;
158 kfree(part->local_GPs_base);
159 part->local_GPs = NULL;
160 kfree(part->channels);
161 part->channels = NULL;
165 part->remote_openclose_args =
166 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
167 &part->remote_openclose_args_base);
168 if (part->remote_openclose_args == NULL) {
169 dev_err(xpc_chan, "can't get memory for remote connect args\n");
170 kfree(part->local_openclose_args_base);
171 part->local_openclose_args = NULL;
172 kfree(part->remote_GPs_base);
173 part->remote_GPs = NULL;
174 kfree(part->local_GPs_base);
175 part->local_GPs = NULL;
176 kfree(part->channels);
177 part->channels = NULL;
181 xpc_initialize_channels(part, partid);
183 atomic_set(&part->nchannels_active, 0);
184 atomic_set(&part->nchannels_engaged, 0);
186 /* local_IPI_amo were set to 0 by an earlier memset() */
188 /* Initialize this partitions AMO_t structure */
189 part->local_IPI_amo_va = xpc_IPI_init(partid);
191 spin_lock_init(&part->IPI_lock);
193 atomic_set(&part->channel_mgr_requests, 1);
194 init_waitqueue_head(&part->channel_mgr_wq);
196 sprintf(part->IPI_owner, "xpc%02d", partid);
197 ret = request_irq(SGI_XPC_NOTIFY, xpc_notify_IRQ_handler, IRQF_SHARED,
198 part->IPI_owner, (void *)(u64)partid);
200 dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
202 kfree(part->remote_openclose_args_base);
203 part->remote_openclose_args = NULL;
204 kfree(part->local_openclose_args_base);
205 part->local_openclose_args = NULL;
206 kfree(part->remote_GPs_base);
207 part->remote_GPs = NULL;
208 kfree(part->local_GPs_base);
209 part->local_GPs = NULL;
210 kfree(part->channels);
211 part->channels = NULL;
212 return xpcLackOfResources;
215 /* Setup a timer to check for dropped IPIs */
216 timer = &part->dropped_IPI_timer;
218 timer->function = (void (*)(unsigned long))xpc_dropped_IPI_check;
219 timer->data = (unsigned long)part;
220 timer->expires = jiffies + XPC_P_DROPPED_IPI_WAIT;
224 * With the setting of the partition setup_state to XPC_P_SETUP, we're
225 * declaring that this partition is ready to go.
227 part->setup_state = XPC_P_SETUP;
230 * Setup the per partition specific variables required by the
231 * remote partition to establish channel connections with us.
233 * The setting of the magic # indicates that these per partition
234 * specific variables are ready to be used.
236 xpc_vars_part[partid].GPs_pa = __pa(part->local_GPs);
237 xpc_vars_part[partid].openclose_args_pa =
238 __pa(part->local_openclose_args);
239 xpc_vars_part[partid].IPI_amo_pa = __pa(part->local_IPI_amo_va);
240 cpuid = raw_smp_processor_id(); /* any CPU in this partition will do */
241 xpc_vars_part[partid].IPI_nasid = cpuid_to_nasid(cpuid);
242 xpc_vars_part[partid].IPI_phys_cpuid = cpu_physical_id(cpuid);
243 xpc_vars_part[partid].nchannels = part->nchannels;
244 xpc_vars_part[partid].magic = XPC_VP_MAGIC1;
250 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
251 * (or multiple cachelines) from a remote partition.
253 * src must be a cacheline aligned physical address on the remote partition.
254 * dst must be a cacheline aligned virtual address on this partition.
255 * cnt must be an cacheline sized
257 static enum xpc_retval
258 xpc_pull_remote_cachelines(struct xpc_partition *part, void *dst,
259 const void *src, size_t cnt)
261 bte_result_t bte_ret;
263 DBUG_ON((u64)src != L1_CACHE_ALIGN((u64)src));
264 DBUG_ON((u64)dst != L1_CACHE_ALIGN((u64)dst));
265 DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));
267 if (part->act_state == XPC_P_DEACTIVATING)
270 bte_ret = xp_bte_copy((u64)src, (u64)dst, (u64)cnt,
271 (BTE_NORMAL | BTE_WACQUIRE), NULL);
272 if (bte_ret == BTE_SUCCESS)
275 dev_dbg(xpc_chan, "xp_bte_copy() from partition %d failed, ret=%d\n",
276 XPC_PARTID(part), bte_ret);
278 return xpc_map_bte_errors(bte_ret);
282 * Pull the remote per partition specific variables from the specified
286 xpc_pull_remote_vars_part(struct xpc_partition *part)
288 u8 buffer[L1_CACHE_BYTES * 2];
289 struct xpc_vars_part *pulled_entry_cacheline =
290 (struct xpc_vars_part *)L1_CACHE_ALIGN((u64)buffer);
291 struct xpc_vars_part *pulled_entry;
292 u64 remote_entry_cacheline_pa, remote_entry_pa;
293 partid_t partid = XPC_PARTID(part);
296 /* pull the cacheline that contains the variables we're interested in */
298 DBUG_ON(part->remote_vars_part_pa !=
299 L1_CACHE_ALIGN(part->remote_vars_part_pa));
300 DBUG_ON(sizeof(struct xpc_vars_part) != L1_CACHE_BYTES / 2);
302 remote_entry_pa = part->remote_vars_part_pa +
303 sn_partition_id * sizeof(struct xpc_vars_part);
305 remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));
307 pulled_entry = (struct xpc_vars_part *)((u64)pulled_entry_cacheline +
309 (L1_CACHE_BYTES - 1)));
311 ret = xpc_pull_remote_cachelines(part, pulled_entry_cacheline,
312 (void *)remote_entry_cacheline_pa,
314 if (ret != xpcSuccess) {
315 dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
316 "partition %d, ret=%d\n", partid, ret);
320 /* see if they've been set up yet */
322 if (pulled_entry->magic != XPC_VP_MAGIC1 &&
323 pulled_entry->magic != XPC_VP_MAGIC2) {
325 if (pulled_entry->magic != 0) {
326 dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
327 "partition %d has bad magic value (=0x%lx)\n",
328 partid, sn_partition_id, pulled_entry->magic);
332 /* they've not been initialized yet */
336 if (xpc_vars_part[partid].magic == XPC_VP_MAGIC1) {
338 /* validate the variables */
340 if (pulled_entry->GPs_pa == 0 ||
341 pulled_entry->openclose_args_pa == 0 ||
342 pulled_entry->IPI_amo_pa == 0) {
344 dev_err(xpc_chan, "partition %d's XPC vars_part for "
345 "partition %d are not valid\n", partid,
347 return xpcInvalidAddress;
350 /* the variables we imported look to be valid */
352 part->remote_GPs_pa = pulled_entry->GPs_pa;
353 part->remote_openclose_args_pa =
354 pulled_entry->openclose_args_pa;
355 part->remote_IPI_amo_va =
356 (AMO_t *)__va(pulled_entry->IPI_amo_pa);
357 part->remote_IPI_nasid = pulled_entry->IPI_nasid;
358 part->remote_IPI_phys_cpuid = pulled_entry->IPI_phys_cpuid;
360 if (part->nchannels > pulled_entry->nchannels)
361 part->nchannels = pulled_entry->nchannels;
363 /* let the other side know that we've pulled their variables */
365 xpc_vars_part[partid].magic = XPC_VP_MAGIC2;
368 if (pulled_entry->magic == XPC_VP_MAGIC1)
375 * Get the IPI flags and pull the openclose args and/or remote GPs as needed.
378 xpc_get_IPI_flags(struct xpc_partition *part)
380 unsigned long irq_flags;
385 * See if there are any IPI flags to be handled.
388 spin_lock_irqsave(&part->IPI_lock, irq_flags);
389 IPI_amo = part->local_IPI_amo;
391 part->local_IPI_amo = 0;
393 spin_unlock_irqrestore(&part->IPI_lock, irq_flags);
395 if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_amo)) {
396 ret = xpc_pull_remote_cachelines(part,
397 part->remote_openclose_args,
399 remote_openclose_args_pa,
400 XPC_OPENCLOSE_ARGS_SIZE);
401 if (ret != xpcSuccess) {
402 XPC_DEACTIVATE_PARTITION(part, ret);
404 dev_dbg(xpc_chan, "failed to pull openclose args from "
405 "partition %d, ret=%d\n", XPC_PARTID(part),
408 /* don't bother processing IPIs anymore */
413 if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_amo)) {
414 ret = xpc_pull_remote_cachelines(part, part->remote_GPs,
415 (void *)part->remote_GPs_pa,
417 if (ret != xpcSuccess) {
418 XPC_DEACTIVATE_PARTITION(part, ret);
420 dev_dbg(xpc_chan, "failed to pull GPs from partition "
421 "%d, ret=%d\n", XPC_PARTID(part), ret);
423 /* don't bother processing IPIs anymore */
432 * Allocate the local message queue and the notify queue.
434 static enum xpc_retval
435 xpc_allocate_local_msgqueue(struct xpc_channel *ch)
437 unsigned long irq_flags;
441 for (nentries = ch->local_nentries; nentries > 0; nentries--) {
443 nbytes = nentries * ch->msg_size;
444 ch->local_msgqueue = xpc_kzalloc_cacheline_aligned(nbytes,
446 &ch->local_msgqueue_base);
447 if (ch->local_msgqueue == NULL)
450 nbytes = nentries * sizeof(struct xpc_notify);
451 ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);
452 if (ch->notify_queue == NULL) {
453 kfree(ch->local_msgqueue_base);
454 ch->local_msgqueue = NULL;
458 spin_lock_irqsave(&ch->lock, irq_flags);
459 if (nentries < ch->local_nentries) {
460 dev_dbg(xpc_chan, "nentries=%d local_nentries=%d, "
461 "partid=%d, channel=%d\n", nentries,
462 ch->local_nentries, ch->partid, ch->number);
464 ch->local_nentries = nentries;
466 spin_unlock_irqrestore(&ch->lock, irq_flags);
470 dev_dbg(xpc_chan, "can't get memory for local message queue and notify "
471 "queue, partid=%d, channel=%d\n", ch->partid, ch->number);
476 * Allocate the cached remote message queue.
478 static enum xpc_retval
479 xpc_allocate_remote_msgqueue(struct xpc_channel *ch)
481 unsigned long irq_flags;
485 DBUG_ON(ch->remote_nentries <= 0);
487 for (nentries = ch->remote_nentries; nentries > 0; nentries--) {
489 nbytes = nentries * ch->msg_size;
490 ch->remote_msgqueue = xpc_kzalloc_cacheline_aligned(nbytes,
492 &ch->remote_msgqueue_base);
493 if (ch->remote_msgqueue == NULL)
496 spin_lock_irqsave(&ch->lock, irq_flags);
497 if (nentries < ch->remote_nentries) {
498 dev_dbg(xpc_chan, "nentries=%d remote_nentries=%d, "
499 "partid=%d, channel=%d\n", nentries,
500 ch->remote_nentries, ch->partid, ch->number);
502 ch->remote_nentries = nentries;
504 spin_unlock_irqrestore(&ch->lock, irq_flags);
508 dev_dbg(xpc_chan, "can't get memory for cached remote message queue, "
509 "partid=%d, channel=%d\n", ch->partid, ch->number);
514 * Allocate message queues and other stuff associated with a channel.
516 * Note: Assumes all of the channel sizes are filled in.
518 static enum xpc_retval
519 xpc_allocate_msgqueues(struct xpc_channel *ch)
521 unsigned long irq_flags;
524 DBUG_ON(ch->flags & XPC_C_SETUP);
526 ret = xpc_allocate_local_msgqueue(ch);
527 if (ret != xpcSuccess)
530 ret = xpc_allocate_remote_msgqueue(ch);
531 if (ret != xpcSuccess) {
532 kfree(ch->local_msgqueue_base);
533 ch->local_msgqueue = NULL;
534 kfree(ch->notify_queue);
535 ch->notify_queue = NULL;
539 spin_lock_irqsave(&ch->lock, irq_flags);
540 ch->flags |= XPC_C_SETUP;
541 spin_unlock_irqrestore(&ch->lock, irq_flags);
547 * Process a connect message from a remote partition.
549 * Note: xpc_process_connect() is expecting to be called with the
550 * spin_lock_irqsave held and will leave it locked upon return.
553 xpc_process_connect(struct xpc_channel *ch, unsigned long *irq_flags)
557 DBUG_ON(!spin_is_locked(&ch->lock));
559 if (!(ch->flags & XPC_C_OPENREQUEST) ||
560 !(ch->flags & XPC_C_ROPENREQUEST)) {
561 /* nothing more to do for now */
564 DBUG_ON(!(ch->flags & XPC_C_CONNECTING));
566 if (!(ch->flags & XPC_C_SETUP)) {
567 spin_unlock_irqrestore(&ch->lock, *irq_flags);
568 ret = xpc_allocate_msgqueues(ch);
569 spin_lock_irqsave(&ch->lock, *irq_flags);
571 if (ret != xpcSuccess)
572 XPC_DISCONNECT_CHANNEL(ch, ret, irq_flags);
574 if (ch->flags & (XPC_C_CONNECTED | XPC_C_DISCONNECTING))
577 DBUG_ON(!(ch->flags & XPC_C_SETUP));
578 DBUG_ON(ch->local_msgqueue == NULL);
579 DBUG_ON(ch->remote_msgqueue == NULL);
582 if (!(ch->flags & XPC_C_OPENREPLY)) {
583 ch->flags |= XPC_C_OPENREPLY;
584 xpc_IPI_send_openreply(ch, irq_flags);
587 if (!(ch->flags & XPC_C_ROPENREPLY))
590 DBUG_ON(ch->remote_msgqueue_pa == 0);
592 ch->flags = (XPC_C_CONNECTED | XPC_C_SETUP); /* clear all else */
594 dev_info(xpc_chan, "channel %d to partition %d connected\n",
595 ch->number, ch->partid);
597 spin_unlock_irqrestore(&ch->lock, *irq_flags);
598 xpc_create_kthreads(ch, 1, 0);
599 spin_lock_irqsave(&ch->lock, *irq_flags);
603 * Notify those who wanted to be notified upon delivery of their message.
606 xpc_notify_senders(struct xpc_channel *ch, enum xpc_retval reason, s64 put)
608 struct xpc_notify *notify;
610 s64 get = ch->w_remote_GP.get - 1;
612 while (++get < put && atomic_read(&ch->n_to_notify) > 0) {
614 notify = &ch->notify_queue[get % ch->local_nentries];
617 * See if the notify entry indicates it was associated with
618 * a message who's sender wants to be notified. It is possible
619 * that it is, but someone else is doing or has done the
622 notify_type = notify->type;
623 if (notify_type == 0 ||
624 cmpxchg(¬ify->type, notify_type, 0) != notify_type) {
628 DBUG_ON(notify_type != XPC_N_CALL);
630 atomic_dec(&ch->n_to_notify);
632 if (notify->func != NULL) {
633 dev_dbg(xpc_chan, "notify->func() called, notify=0x%p, "
634 "msg_number=%ld, partid=%d, channel=%d\n",
635 (void *)notify, get, ch->partid, ch->number);
637 notify->func(reason, ch->partid, ch->number,
640 dev_dbg(xpc_chan, "notify->func() returned, "
641 "notify=0x%p, msg_number=%ld, partid=%d, "
642 "channel=%d\n", (void *)notify, get,
643 ch->partid, ch->number);
649 * Free up message queues and other stuff that were allocated for the specified
652 * Note: ch->reason and ch->reason_line are left set for debugging purposes,
653 * they're cleared when XPC_C_DISCONNECTED is cleared.
656 xpc_free_msgqueues(struct xpc_channel *ch)
658 DBUG_ON(!spin_is_locked(&ch->lock));
659 DBUG_ON(atomic_read(&ch->n_to_notify) != 0);
661 ch->remote_msgqueue_pa = 0;
665 ch->local_nentries = 0;
666 ch->remote_nentries = 0;
667 ch->kthreads_assigned_limit = 0;
668 ch->kthreads_idle_limit = 0;
670 ch->local_GP->get = 0;
671 ch->local_GP->put = 0;
672 ch->remote_GP.get = 0;
673 ch->remote_GP.put = 0;
674 ch->w_local_GP.get = 0;
675 ch->w_local_GP.put = 0;
676 ch->w_remote_GP.get = 0;
677 ch->w_remote_GP.put = 0;
678 ch->next_msg_to_pull = 0;
680 if (ch->flags & XPC_C_SETUP) {
681 ch->flags &= ~XPC_C_SETUP;
683 dev_dbg(xpc_chan, "ch->flags=0x%x, partid=%d, channel=%d\n",
684 ch->flags, ch->partid, ch->number);
686 kfree(ch->local_msgqueue_base);
687 ch->local_msgqueue = NULL;
688 kfree(ch->remote_msgqueue_base);
689 ch->remote_msgqueue = NULL;
690 kfree(ch->notify_queue);
691 ch->notify_queue = NULL;
696 * spin_lock_irqsave() is expected to be held on entry.
699 xpc_process_disconnect(struct xpc_channel *ch, unsigned long *irq_flags)
701 struct xpc_partition *part = &xpc_partitions[ch->partid];
702 u32 channel_was_connected = (ch->flags & XPC_C_WASCONNECTED);
704 DBUG_ON(!spin_is_locked(&ch->lock));
706 if (!(ch->flags & XPC_C_DISCONNECTING))
709 DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
711 /* make sure all activity has settled down first */
713 if (atomic_read(&ch->kthreads_assigned) > 0 ||
714 atomic_read(&ch->references) > 0) {
717 DBUG_ON((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
718 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT_MADE));
720 if (part->act_state == XPC_P_DEACTIVATING) {
721 /* can't proceed until the other side disengages from us */
722 if (xpc_partition_engaged(1UL << ch->partid))
727 /* as long as the other side is up do the full protocol */
729 if (!(ch->flags & XPC_C_RCLOSEREQUEST))
732 if (!(ch->flags & XPC_C_CLOSEREPLY)) {
733 ch->flags |= XPC_C_CLOSEREPLY;
734 xpc_IPI_send_closereply(ch, irq_flags);
737 if (!(ch->flags & XPC_C_RCLOSEREPLY))
741 /* wake those waiting for notify completion */
742 if (atomic_read(&ch->n_to_notify) > 0) {
743 /* >>> we do callout while holding ch->lock */
744 xpc_notify_senders(ch, ch->reason, ch->w_local_GP.put);
747 /* both sides are disconnected now */
749 if (ch->flags & XPC_C_DISCONNECTINGCALLOUT_MADE) {
750 spin_unlock_irqrestore(&ch->lock, *irq_flags);
751 xpc_disconnect_callout(ch, xpcDisconnected);
752 spin_lock_irqsave(&ch->lock, *irq_flags);
755 /* it's now safe to free the channel's message queues */
756 xpc_free_msgqueues(ch);
758 /* mark disconnected, clear all other flags except XPC_C_WDISCONNECT */
759 ch->flags = (XPC_C_DISCONNECTED | (ch->flags & XPC_C_WDISCONNECT));
761 atomic_dec(&part->nchannels_active);
763 if (channel_was_connected) {
764 dev_info(xpc_chan, "channel %d to partition %d disconnected, "
765 "reason=%d\n", ch->number, ch->partid, ch->reason);
768 if (ch->flags & XPC_C_WDISCONNECT) {
769 /* we won't lose the CPU since we're holding ch->lock */
770 complete(&ch->wdisconnect_wait);
771 } else if (ch->delayed_IPI_flags) {
772 if (part->act_state != XPC_P_DEACTIVATING) {
773 /* time to take action on any delayed IPI flags */
774 spin_lock(&part->IPI_lock);
775 XPC_SET_IPI_FLAGS(part->local_IPI_amo, ch->number,
776 ch->delayed_IPI_flags);
777 spin_unlock(&part->IPI_lock);
779 ch->delayed_IPI_flags = 0;
784 * Process a change in the channel's remote connection state.
787 xpc_process_openclose_IPI(struct xpc_partition *part, int ch_number,
790 unsigned long irq_flags;
791 struct xpc_openclose_args *args =
792 &part->remote_openclose_args[ch_number];
793 struct xpc_channel *ch = &part->channels[ch_number];
794 enum xpc_retval reason;
796 spin_lock_irqsave(&ch->lock, irq_flags);
800 if ((ch->flags & XPC_C_DISCONNECTED) &&
801 (ch->flags & XPC_C_WDISCONNECT)) {
803 * Delay processing IPI flags until thread waiting disconnect
804 * has had a chance to see that the channel is disconnected.
806 ch->delayed_IPI_flags |= IPI_flags;
807 spin_unlock_irqrestore(&ch->lock, irq_flags);
811 if (IPI_flags & XPC_IPI_CLOSEREQUEST) {
813 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREQUEST (reason=%d) received "
814 "from partid=%d, channel=%d\n", args->reason,
815 ch->partid, ch->number);
818 * If RCLOSEREQUEST is set, we're probably waiting for
819 * RCLOSEREPLY. We should find it and a ROPENREQUEST packed
820 * with this RCLOSEREQUEST in the IPI_flags.
823 if (ch->flags & XPC_C_RCLOSEREQUEST) {
824 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTING));
825 DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
826 DBUG_ON(!(ch->flags & XPC_C_CLOSEREPLY));
827 DBUG_ON(ch->flags & XPC_C_RCLOSEREPLY);
829 DBUG_ON(!(IPI_flags & XPC_IPI_CLOSEREPLY));
830 IPI_flags &= ~XPC_IPI_CLOSEREPLY;
831 ch->flags |= XPC_C_RCLOSEREPLY;
833 /* both sides have finished disconnecting */
834 xpc_process_disconnect(ch, &irq_flags);
835 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
839 if (ch->flags & XPC_C_DISCONNECTED) {
840 if (!(IPI_flags & XPC_IPI_OPENREQUEST)) {
841 if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo,
843 XPC_IPI_OPENREQUEST)) {
845 DBUG_ON(ch->delayed_IPI_flags != 0);
846 spin_lock(&part->IPI_lock);
847 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
849 XPC_IPI_CLOSEREQUEST);
850 spin_unlock(&part->IPI_lock);
852 spin_unlock_irqrestore(&ch->lock, irq_flags);
856 XPC_SET_REASON(ch, 0, 0);
857 ch->flags &= ~XPC_C_DISCONNECTED;
859 atomic_inc(&part->nchannels_active);
860 ch->flags |= (XPC_C_CONNECTING | XPC_C_ROPENREQUEST);
863 IPI_flags &= ~(XPC_IPI_OPENREQUEST | XPC_IPI_OPENREPLY);
866 * The meaningful CLOSEREQUEST connection state fields are:
867 * reason = reason connection is to be closed
870 ch->flags |= XPC_C_RCLOSEREQUEST;
872 if (!(ch->flags & XPC_C_DISCONNECTING)) {
873 reason = args->reason;
874 if (reason <= xpcSuccess || reason > xpcUnknownReason)
875 reason = xpcUnknownReason;
876 else if (reason == xpcUnregistering)
877 reason = xpcOtherUnregistering;
879 XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
881 DBUG_ON(IPI_flags & XPC_IPI_CLOSEREPLY);
882 spin_unlock_irqrestore(&ch->lock, irq_flags);
886 xpc_process_disconnect(ch, &irq_flags);
889 if (IPI_flags & XPC_IPI_CLOSEREPLY) {
891 dev_dbg(xpc_chan, "XPC_IPI_CLOSEREPLY received from partid=%d,"
892 " channel=%d\n", ch->partid, ch->number);
894 if (ch->flags & XPC_C_DISCONNECTED) {
895 DBUG_ON(part->act_state != XPC_P_DEACTIVATING);
896 spin_unlock_irqrestore(&ch->lock, irq_flags);
900 DBUG_ON(!(ch->flags & XPC_C_CLOSEREQUEST));
902 if (!(ch->flags & XPC_C_RCLOSEREQUEST)) {
903 if ((XPC_GET_IPI_FLAGS(part->local_IPI_amo, ch_number)
904 & XPC_IPI_CLOSEREQUEST)) {
906 DBUG_ON(ch->delayed_IPI_flags != 0);
907 spin_lock(&part->IPI_lock);
908 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
911 spin_unlock(&part->IPI_lock);
913 spin_unlock_irqrestore(&ch->lock, irq_flags);
917 ch->flags |= XPC_C_RCLOSEREPLY;
919 if (ch->flags & XPC_C_CLOSEREPLY) {
920 /* both sides have finished disconnecting */
921 xpc_process_disconnect(ch, &irq_flags);
925 if (IPI_flags & XPC_IPI_OPENREQUEST) {
927 dev_dbg(xpc_chan, "XPC_IPI_OPENREQUEST (msg_size=%d, "
928 "local_nentries=%d) received from partid=%d, "
929 "channel=%d\n", args->msg_size, args->local_nentries,
930 ch->partid, ch->number);
932 if (part->act_state == XPC_P_DEACTIVATING ||
933 (ch->flags & XPC_C_ROPENREQUEST)) {
934 spin_unlock_irqrestore(&ch->lock, irq_flags);
938 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_WDISCONNECT)) {
939 ch->delayed_IPI_flags |= XPC_IPI_OPENREQUEST;
940 spin_unlock_irqrestore(&ch->lock, irq_flags);
943 DBUG_ON(!(ch->flags & (XPC_C_DISCONNECTED |
944 XPC_C_OPENREQUEST)));
945 DBUG_ON(ch->flags & (XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
946 XPC_C_OPENREPLY | XPC_C_CONNECTED));
949 * The meaningful OPENREQUEST connection state fields are:
950 * msg_size = size of channel's messages in bytes
951 * local_nentries = remote partition's local_nentries
953 if (args->msg_size == 0 || args->local_nentries == 0) {
954 /* assume OPENREQUEST was delayed by mistake */
955 spin_unlock_irqrestore(&ch->lock, irq_flags);
959 ch->flags |= (XPC_C_ROPENREQUEST | XPC_C_CONNECTING);
960 ch->remote_nentries = args->local_nentries;
962 if (ch->flags & XPC_C_OPENREQUEST) {
963 if (args->msg_size != ch->msg_size) {
964 XPC_DISCONNECT_CHANNEL(ch, xpcUnequalMsgSizes,
966 spin_unlock_irqrestore(&ch->lock, irq_flags);
970 ch->msg_size = args->msg_size;
972 XPC_SET_REASON(ch, 0, 0);
973 ch->flags &= ~XPC_C_DISCONNECTED;
975 atomic_inc(&part->nchannels_active);
978 xpc_process_connect(ch, &irq_flags);
981 if (IPI_flags & XPC_IPI_OPENREPLY) {
983 dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY (local_msgqueue_pa=0x%lx, "
984 "local_nentries=%d, remote_nentries=%d) received from "
985 "partid=%d, channel=%d\n", args->local_msgqueue_pa,
986 args->local_nentries, args->remote_nentries,
987 ch->partid, ch->number);
989 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED)) {
990 spin_unlock_irqrestore(&ch->lock, irq_flags);
993 if (!(ch->flags & XPC_C_OPENREQUEST)) {
994 XPC_DISCONNECT_CHANNEL(ch, xpcOpenCloseError,
996 spin_unlock_irqrestore(&ch->lock, irq_flags);
1000 DBUG_ON(!(ch->flags & XPC_C_ROPENREQUEST));
1001 DBUG_ON(ch->flags & XPC_C_CONNECTED);
1004 * The meaningful OPENREPLY connection state fields are:
1005 * local_msgqueue_pa = physical address of remote
1006 * partition's local_msgqueue
1007 * local_nentries = remote partition's local_nentries
1008 * remote_nentries = remote partition's remote_nentries
1010 DBUG_ON(args->local_msgqueue_pa == 0);
1011 DBUG_ON(args->local_nentries == 0);
1012 DBUG_ON(args->remote_nentries == 0);
1014 ch->flags |= XPC_C_ROPENREPLY;
1015 ch->remote_msgqueue_pa = args->local_msgqueue_pa;
1017 if (args->local_nentries < ch->remote_nentries) {
1018 dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1019 "remote_nentries=%d, old remote_nentries=%d, "
1020 "partid=%d, channel=%d\n",
1021 args->local_nentries, ch->remote_nentries,
1022 ch->partid, ch->number);
1024 ch->remote_nentries = args->local_nentries;
1026 if (args->remote_nentries < ch->local_nentries) {
1027 dev_dbg(xpc_chan, "XPC_IPI_OPENREPLY: new "
1028 "local_nentries=%d, old local_nentries=%d, "
1029 "partid=%d, channel=%d\n",
1030 args->remote_nentries, ch->local_nentries,
1031 ch->partid, ch->number);
1033 ch->local_nentries = args->remote_nentries;
1036 xpc_process_connect(ch, &irq_flags);
1039 spin_unlock_irqrestore(&ch->lock, irq_flags);
1043 * Attempt to establish a channel connection to a remote partition.
1045 static enum xpc_retval
1046 xpc_connect_channel(struct xpc_channel *ch)
1048 unsigned long irq_flags;
1049 struct xpc_registration *registration = &xpc_registrations[ch->number];
1051 if (mutex_trylock(®istration->mutex) == 0)
1054 if (!XPC_CHANNEL_REGISTERED(ch->number)) {
1055 mutex_unlock(®istration->mutex);
1056 return xpcUnregistered;
1059 spin_lock_irqsave(&ch->lock, irq_flags);
1061 DBUG_ON(ch->flags & XPC_C_CONNECTED);
1062 DBUG_ON(ch->flags & XPC_C_OPENREQUEST);
1064 if (ch->flags & XPC_C_DISCONNECTING) {
1065 spin_unlock_irqrestore(&ch->lock, irq_flags);
1066 mutex_unlock(®istration->mutex);
1070 /* add info from the channel connect registration to the channel */
1072 ch->kthreads_assigned_limit = registration->assigned_limit;
1073 ch->kthreads_idle_limit = registration->idle_limit;
1074 DBUG_ON(atomic_read(&ch->kthreads_assigned) != 0);
1075 DBUG_ON(atomic_read(&ch->kthreads_idle) != 0);
1076 DBUG_ON(atomic_read(&ch->kthreads_active) != 0);
1078 ch->func = registration->func;
1079 DBUG_ON(registration->func == NULL);
1080 ch->key = registration->key;
1082 ch->local_nentries = registration->nentries;
1084 if (ch->flags & XPC_C_ROPENREQUEST) {
1085 if (registration->msg_size != ch->msg_size) {
1086 /* the local and remote sides aren't the same */
1089 * Because XPC_DISCONNECT_CHANNEL() can block we're
1090 * forced to up the registration sema before we unlock
1091 * the channel lock. But that's okay here because we're
1092 * done with the part that required the registration
1093 * sema. XPC_DISCONNECT_CHANNEL() requires that the
1094 * channel lock be locked and will unlock and relock
1095 * the channel lock as needed.
1097 mutex_unlock(®istration->mutex);
1098 XPC_DISCONNECT_CHANNEL(ch, xpcUnequalMsgSizes,
1100 spin_unlock_irqrestore(&ch->lock, irq_flags);
1101 return xpcUnequalMsgSizes;
1104 ch->msg_size = registration->msg_size;
1106 XPC_SET_REASON(ch, 0, 0);
1107 ch->flags &= ~XPC_C_DISCONNECTED;
1109 atomic_inc(&xpc_partitions[ch->partid].nchannels_active);
1112 mutex_unlock(®istration->mutex);
1114 /* initiate the connection */
1116 ch->flags |= (XPC_C_OPENREQUEST | XPC_C_CONNECTING);
1117 xpc_IPI_send_openrequest(ch, &irq_flags);
1119 xpc_process_connect(ch, &irq_flags);
1121 spin_unlock_irqrestore(&ch->lock, irq_flags);
1127 * Clear some of the msg flags in the local message queue.
1130 xpc_clear_local_msgqueue_flags(struct xpc_channel *ch)
1132 struct xpc_msg *msg;
1135 get = ch->w_remote_GP.get;
1137 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1138 (get % ch->local_nentries) *
1141 } while (++get < ch->remote_GP.get);
1145 * Clear some of the msg flags in the remote message queue.
1148 xpc_clear_remote_msgqueue_flags(struct xpc_channel *ch)
1150 struct xpc_msg *msg;
1153 put = ch->w_remote_GP.put;
1155 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
1156 (put % ch->remote_nentries) *
1159 } while (++put < ch->remote_GP.put);
1163 xpc_process_msg_IPI(struct xpc_partition *part, int ch_number)
1165 struct xpc_channel *ch = &part->channels[ch_number];
1168 ch->remote_GP = part->remote_GPs[ch_number];
1170 /* See what, if anything, has changed for each connected channel */
1172 xpc_msgqueue_ref(ch);
1174 if (ch->w_remote_GP.get == ch->remote_GP.get &&
1175 ch->w_remote_GP.put == ch->remote_GP.put) {
1176 /* nothing changed since GPs were last pulled */
1177 xpc_msgqueue_deref(ch);
1181 if (!(ch->flags & XPC_C_CONNECTED)) {
1182 xpc_msgqueue_deref(ch);
1187 * First check to see if messages recently sent by us have been
1188 * received by the other side. (The remote GET value will have
1189 * changed since we last looked at it.)
1192 if (ch->w_remote_GP.get != ch->remote_GP.get) {
1195 * We need to notify any senders that want to be notified
1196 * that their sent messages have been received by their
1197 * intended recipients. We need to do this before updating
1198 * w_remote_GP.get so that we don't allocate the same message
1199 * queue entries prematurely (see xpc_allocate_msg()).
1201 if (atomic_read(&ch->n_to_notify) > 0) {
1203 * Notify senders that messages sent have been
1204 * received and delivered by the other side.
1206 xpc_notify_senders(ch, xpcMsgDelivered,
1211 * Clear msg->flags in previously sent messages, so that
1212 * they're ready for xpc_allocate_msg().
1214 xpc_clear_local_msgqueue_flags(ch);
1216 ch->w_remote_GP.get = ch->remote_GP.get;
1218 dev_dbg(xpc_chan, "w_remote_GP.get changed to %ld, partid=%d, "
1219 "channel=%d\n", ch->w_remote_GP.get, ch->partid,
1223 * If anyone was waiting for message queue entries to become
1224 * available, wake them up.
1226 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1227 wake_up(&ch->msg_allocate_wq);
1231 * Now check for newly sent messages by the other side. (The remote
1232 * PUT value will have changed since we last looked at it.)
1235 if (ch->w_remote_GP.put != ch->remote_GP.put) {
1237 * Clear msg->flags in previously received messages, so that
1238 * they're ready for xpc_get_deliverable_msg().
1240 xpc_clear_remote_msgqueue_flags(ch);
1242 ch->w_remote_GP.put = ch->remote_GP.put;
1244 dev_dbg(xpc_chan, "w_remote_GP.put changed to %ld, partid=%d, "
1245 "channel=%d\n", ch->w_remote_GP.put, ch->partid,
1248 nmsgs_sent = ch->w_remote_GP.put - ch->w_local_GP.get;
1249 if (nmsgs_sent > 0) {
1250 dev_dbg(xpc_chan, "msgs waiting to be copied and "
1251 "delivered=%d, partid=%d, channel=%d\n",
1252 nmsgs_sent, ch->partid, ch->number);
1254 if (ch->flags & XPC_C_CONNECTEDCALLOUT_MADE)
1255 xpc_activate_kthreads(ch, nmsgs_sent);
1259 xpc_msgqueue_deref(ch);
1263 xpc_process_channel_activity(struct xpc_partition *part)
1265 unsigned long irq_flags;
1266 u64 IPI_amo, IPI_flags;
1267 struct xpc_channel *ch;
1271 IPI_amo = xpc_get_IPI_flags(part);
1274 * Initiate channel connections for registered channels.
1276 * For each connected channel that has pending messages activate idle
1277 * kthreads and/or create new kthreads as needed.
1280 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1281 ch = &part->channels[ch_number];
1284 * Process any open or close related IPI flags, and then deal
1285 * with connecting or disconnecting the channel as required.
1288 IPI_flags = XPC_GET_IPI_FLAGS(IPI_amo, ch_number);
1290 if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_flags))
1291 xpc_process_openclose_IPI(part, ch_number, IPI_flags);
1293 ch_flags = ch->flags; /* need an atomic snapshot of flags */
1295 if (ch_flags & XPC_C_DISCONNECTING) {
1296 spin_lock_irqsave(&ch->lock, irq_flags);
1297 xpc_process_disconnect(ch, &irq_flags);
1298 spin_unlock_irqrestore(&ch->lock, irq_flags);
1302 if (part->act_state == XPC_P_DEACTIVATING)
1305 if (!(ch_flags & XPC_C_CONNECTED)) {
1306 if (!(ch_flags & XPC_C_OPENREQUEST)) {
1307 DBUG_ON(ch_flags & XPC_C_SETUP);
1308 (void)xpc_connect_channel(ch);
1310 spin_lock_irqsave(&ch->lock, irq_flags);
1311 xpc_process_connect(ch, &irq_flags);
1312 spin_unlock_irqrestore(&ch->lock, irq_flags);
1318 * Process any message related IPI flags, this may involve the
1319 * activation of kthreads to deliver any pending messages sent
1320 * from the other partition.
1323 if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_flags))
1324 xpc_process_msg_IPI(part, ch_number);
1329 * XPC's heartbeat code calls this function to inform XPC that a partition is
1330 * going down. XPC responds by tearing down the XPartition Communication
1331 * infrastructure used for the just downed partition.
1333 * XPC's heartbeat code will never call this function and xpc_partition_up()
1334 * at the same time. Nor will it ever make multiple calls to either function
1338 xpc_partition_going_down(struct xpc_partition *part, enum xpc_retval reason)
1340 unsigned long irq_flags;
1342 struct xpc_channel *ch;
1344 dev_dbg(xpc_chan, "deactivating partition %d, reason=%d\n",
1345 XPC_PARTID(part), reason);
1347 if (!xpc_part_ref(part)) {
1348 /* infrastructure for this partition isn't currently set up */
1352 /* disconnect channels associated with the partition going down */
1354 for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
1355 ch = &part->channels[ch_number];
1357 xpc_msgqueue_ref(ch);
1358 spin_lock_irqsave(&ch->lock, irq_flags);
1360 XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
1362 spin_unlock_irqrestore(&ch->lock, irq_flags);
1363 xpc_msgqueue_deref(ch);
1366 xpc_wakeup_channel_mgr(part);
1368 xpc_part_deref(part);
1372 * Teardown the infrastructure necessary to support XPartition Communication
1373 * between the specified remote partition and the local one.
1376 xpc_teardown_infrastructure(struct xpc_partition *part)
1378 partid_t partid = XPC_PARTID(part);
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
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;
1394 xpc_vars_part[partid].magic = 0;
1396 free_irq(SGI_XPC_NOTIFY, (void *)(u64)partid);
1399 * Before proceeding with the teardown we have to wait until all
1400 * existing references cease.
1402 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
1404 /* now we can begin tearing down the infrastructure */
1406 part->setup_state = XPC_P_TORNDOWN;
1408 /* in case we've still got outstanding timers registered... */
1409 del_timer_sync(&part->dropped_IPI_timer);
1411 kfree(part->remote_openclose_args_base);
1412 part->remote_openclose_args = NULL;
1413 kfree(part->local_openclose_args_base);
1414 part->local_openclose_args = NULL;
1415 kfree(part->remote_GPs_base);
1416 part->remote_GPs = NULL;
1417 kfree(part->local_GPs_base);
1418 part->local_GPs = NULL;
1419 kfree(part->channels);
1420 part->channels = NULL;
1421 part->local_IPI_amo_va = NULL;
1425 * Called by XP at the time of channel connection registration to cause
1426 * XPC to establish connections to all currently active partitions.
1429 xpc_initiate_connect(int ch_number)
1432 struct xpc_partition *part;
1433 struct xpc_channel *ch;
1435 DBUG_ON(ch_number < 0 || ch_number >= XPC_NCHANNELS);
1437 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1438 part = &xpc_partitions[partid];
1440 if (xpc_part_ref(part)) {
1441 ch = &part->channels[ch_number];
1444 * Initiate the establishment of a connection on the
1445 * newly registered channel to the remote partition.
1447 xpc_wakeup_channel_mgr(part);
1448 xpc_part_deref(part);
1454 xpc_connected_callout(struct xpc_channel *ch)
1456 /* let the registerer know that a connection has been established */
1458 if (ch->func != NULL) {
1459 dev_dbg(xpc_chan, "ch->func() called, reason=xpcConnected, "
1460 "partid=%d, channel=%d\n", ch->partid, ch->number);
1462 ch->func(xpcConnected, ch->partid, ch->number,
1463 (void *)(u64)ch->local_nentries, ch->key);
1465 dev_dbg(xpc_chan, "ch->func() returned, reason=xpcConnected, "
1466 "partid=%d, channel=%d\n", ch->partid, ch->number);
1471 * Called by XP at the time of channel connection unregistration to cause
1472 * XPC to teardown all current connections for the specified channel.
1474 * Before returning xpc_initiate_disconnect() will wait until all connections
1475 * on the specified channel have been closed/torndown. So the caller can be
1476 * assured that they will not be receiving any more callouts from XPC to the
1477 * function they registered via xpc_connect().
1481 * ch_number - channel # to unregister.
1484 xpc_initiate_disconnect(int ch_number)
1486 unsigned long irq_flags;
1488 struct xpc_partition *part;
1489 struct xpc_channel *ch;
1491 DBUG_ON(ch_number < 0 || ch_number >= XPC_NCHANNELS);
1493 /* initiate the channel disconnect for every active partition */
1494 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1495 part = &xpc_partitions[partid];
1497 if (xpc_part_ref(part)) {
1498 ch = &part->channels[ch_number];
1499 xpc_msgqueue_ref(ch);
1501 spin_lock_irqsave(&ch->lock, irq_flags);
1503 if (!(ch->flags & XPC_C_DISCONNECTED)) {
1504 ch->flags |= XPC_C_WDISCONNECT;
1506 XPC_DISCONNECT_CHANNEL(ch, xpcUnregistering,
1510 spin_unlock_irqrestore(&ch->lock, irq_flags);
1512 xpc_msgqueue_deref(ch);
1513 xpc_part_deref(part);
1517 xpc_disconnect_wait(ch_number);
1521 * To disconnect a channel, and reflect it back to all who may be waiting.
1523 * An OPEN is not allowed until XPC_C_DISCONNECTING is cleared by
1524 * xpc_process_disconnect(), and if set, XPC_C_WDISCONNECT is cleared by
1525 * xpc_disconnect_wait().
1527 * THE CHANNEL IS TO BE LOCKED BY THE CALLER AND WILL REMAIN LOCKED UPON RETURN.
1530 xpc_disconnect_channel(const int line, struct xpc_channel *ch,
1531 enum xpc_retval reason, unsigned long *irq_flags)
1533 u32 channel_was_connected = (ch->flags & XPC_C_CONNECTED);
1535 DBUG_ON(!spin_is_locked(&ch->lock));
1537 if (ch->flags & (XPC_C_DISCONNECTING | XPC_C_DISCONNECTED))
1540 DBUG_ON(!(ch->flags & (XPC_C_CONNECTING | XPC_C_CONNECTED)));
1542 dev_dbg(xpc_chan, "reason=%d, line=%d, partid=%d, channel=%d\n",
1543 reason, line, ch->partid, ch->number);
1545 XPC_SET_REASON(ch, reason, line);
1547 ch->flags |= (XPC_C_CLOSEREQUEST | XPC_C_DISCONNECTING);
1548 /* some of these may not have been set */
1549 ch->flags &= ~(XPC_C_OPENREQUEST | XPC_C_OPENREPLY |
1550 XPC_C_ROPENREQUEST | XPC_C_ROPENREPLY |
1551 XPC_C_CONNECTING | XPC_C_CONNECTED);
1553 xpc_IPI_send_closerequest(ch, irq_flags);
1555 if (channel_was_connected)
1556 ch->flags |= XPC_C_WASCONNECTED;
1558 spin_unlock_irqrestore(&ch->lock, *irq_flags);
1560 /* wake all idle kthreads so they can exit */
1561 if (atomic_read(&ch->kthreads_idle) > 0) {
1562 wake_up_all(&ch->idle_wq);
1564 } else if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
1565 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
1566 /* start a kthread that will do the xpcDisconnecting callout */
1567 xpc_create_kthreads(ch, 1, 1);
1570 /* wake those waiting to allocate an entry from the local msg queue */
1571 if (atomic_read(&ch->n_on_msg_allocate_wq) > 0)
1572 wake_up(&ch->msg_allocate_wq);
1574 spin_lock_irqsave(&ch->lock, *irq_flags);
1578 xpc_disconnect_callout(struct xpc_channel *ch, enum xpc_retval reason)
1581 * Let the channel's registerer know that the channel is being
1582 * disconnected. We don't want to do this if the registerer was never
1583 * informed of a connection being made.
1586 if (ch->func != NULL) {
1587 dev_dbg(xpc_chan, "ch->func() called, reason=%d, partid=%d, "
1588 "channel=%d\n", reason, ch->partid, ch->number);
1590 ch->func(reason, ch->partid, ch->number, NULL, ch->key);
1592 dev_dbg(xpc_chan, "ch->func() returned, reason=%d, partid=%d, "
1593 "channel=%d\n", reason, ch->partid, ch->number);
1598 * Wait for a message entry to become available for the specified channel,
1599 * but don't wait any longer than 1 jiffy.
1601 static enum xpc_retval
1602 xpc_allocate_msg_wait(struct xpc_channel *ch)
1604 enum xpc_retval ret;
1606 if (ch->flags & XPC_C_DISCONNECTING) {
1607 DBUG_ON(ch->reason == xpcInterrupted);
1611 atomic_inc(&ch->n_on_msg_allocate_wq);
1612 ret = interruptible_sleep_on_timeout(&ch->msg_allocate_wq, 1);
1613 atomic_dec(&ch->n_on_msg_allocate_wq);
1615 if (ch->flags & XPC_C_DISCONNECTING) {
1617 DBUG_ON(ch->reason == xpcInterrupted);
1618 } else if (ret == 0) {
1621 ret = xpcInterrupted;
1628 * Allocate an entry for a message from the message queue associated with the
1629 * specified channel.
1631 static enum xpc_retval
1632 xpc_allocate_msg(struct xpc_channel *ch, u32 flags,
1633 struct xpc_msg **address_of_msg)
1635 struct xpc_msg *msg;
1636 enum xpc_retval ret;
1639 /* this reference will be dropped in xpc_send_msg() */
1640 xpc_msgqueue_ref(ch);
1642 if (ch->flags & XPC_C_DISCONNECTING) {
1643 xpc_msgqueue_deref(ch);
1646 if (!(ch->flags & XPC_C_CONNECTED)) {
1647 xpc_msgqueue_deref(ch);
1648 return xpcNotConnected;
1652 * Get the next available message entry from the local message queue.
1653 * If none are available, we'll make sure that we grab the latest
1660 put = ch->w_local_GP.put;
1661 rmb(); /* guarantee that .put loads before .get */
1662 if (put - ch->w_remote_GP.get < ch->local_nentries) {
1664 /* There are available message entries. We need to try
1665 * to secure one for ourselves. We'll do this by trying
1666 * to increment w_local_GP.put as long as someone else
1667 * doesn't beat us to it. If they do, we'll have to
1670 if (cmpxchg(&ch->w_local_GP.put, put, put + 1) == put) {
1671 /* we got the entry referenced by put */
1674 continue; /* try again */
1678 * There aren't any available msg entries at this time.
1680 * In waiting for a message entry to become available,
1681 * we set a timeout in case the other side is not
1682 * sending completion IPIs. This lets us fake an IPI
1683 * that will cause the IPI handler to fetch the latest
1684 * GP values as if an IPI was sent by the other side.
1686 if (ret == xpcTimeout)
1687 xpc_IPI_send_local_msgrequest(ch);
1689 if (flags & XPC_NOWAIT) {
1690 xpc_msgqueue_deref(ch);
1694 ret = xpc_allocate_msg_wait(ch);
1695 if (ret != xpcInterrupted && ret != xpcTimeout) {
1696 xpc_msgqueue_deref(ch);
1701 /* get the message's address and initialize it */
1702 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1703 (put % ch->local_nentries) * ch->msg_size);
1705 DBUG_ON(msg->flags != 0);
1708 dev_dbg(xpc_chan, "w_local_GP.put changed to %ld; msg=0x%p, "
1709 "msg_number=%ld, partid=%d, channel=%d\n", put + 1,
1710 (void *)msg, msg->number, ch->partid, ch->number);
1712 *address_of_msg = msg;
1718 * Allocate an entry for a message from the message queue associated with the
1719 * specified channel. NOTE that this routine can sleep waiting for a message
1720 * entry to become available. To not sleep, pass in the XPC_NOWAIT flag.
1724 * partid - ID of partition to which the channel is connected.
1725 * ch_number - channel #.
1726 * flags - see xpc.h for valid flags.
1727 * payload - address of the allocated payload area pointer (filled in on
1728 * return) in which the user-defined message is constructed.
1731 xpc_initiate_allocate(partid_t partid, int ch_number, u32 flags, void **payload)
1733 struct xpc_partition *part = &xpc_partitions[partid];
1734 enum xpc_retval ret = xpcUnknownReason;
1735 struct xpc_msg *msg = NULL;
1737 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
1738 DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1742 if (xpc_part_ref(part)) {
1743 ret = xpc_allocate_msg(&part->channels[ch_number], flags, &msg);
1744 xpc_part_deref(part);
1747 *payload = &msg->payload;
1754 * Now we actually send the messages that are ready to be sent by advancing
1755 * the local message queue's Put value and then send an IPI to the recipient
1759 xpc_send_msgs(struct xpc_channel *ch, s64 initial_put)
1761 struct xpc_msg *msg;
1762 s64 put = initial_put + 1;
1768 if (put == ch->w_local_GP.put)
1771 msg = (struct xpc_msg *)((u64)ch->local_msgqueue +
1772 (put % ch->local_nentries) *
1775 if (!(msg->flags & XPC_M_READY))
1781 if (put == initial_put) {
1782 /* nothing's changed */
1786 if (cmpxchg_rel(&ch->local_GP->put, initial_put, put) !=
1788 /* someone else beat us to it */
1789 DBUG_ON(ch->local_GP->put < initial_put);
1793 /* we just set the new value of local_GP->put */
1795 dev_dbg(xpc_chan, "local_GP->put changed to %ld, partid=%d, "
1796 "channel=%d\n", put, ch->partid, ch->number);
1801 * We need to ensure that the message referenced by
1802 * local_GP->put is not XPC_M_READY or that local_GP->put
1803 * equals w_local_GP.put, so we'll go have a look.
1809 xpc_IPI_send_msgrequest(ch);
1813 * Common code that does the actual sending of the message by advancing the
1814 * local message queue's Put value and sends an IPI to the partition the
1815 * message is being sent to.
1817 static enum xpc_retval
1818 xpc_send_msg(struct xpc_channel *ch, struct xpc_msg *msg, u8 notify_type,
1819 xpc_notify_func func, void *key)
1821 enum xpc_retval ret = xpcSuccess;
1822 struct xpc_notify *notify = notify;
1823 s64 put, msg_number = msg->number;
1825 DBUG_ON(notify_type == XPC_N_CALL && func == NULL);
1826 DBUG_ON((((u64)msg - (u64)ch->local_msgqueue) / ch->msg_size) !=
1827 msg_number % ch->local_nentries);
1828 DBUG_ON(msg->flags & XPC_M_READY);
1830 if (ch->flags & XPC_C_DISCONNECTING) {
1831 /* drop the reference grabbed in xpc_allocate_msg() */
1832 xpc_msgqueue_deref(ch);
1836 if (notify_type != 0) {
1838 * Tell the remote side to send an ACK interrupt when the
1839 * message has been delivered.
1841 msg->flags |= XPC_M_INTERRUPT;
1843 atomic_inc(&ch->n_to_notify);
1845 notify = &ch->notify_queue[msg_number % ch->local_nentries];
1846 notify->func = func;
1848 notify->type = notify_type;
1850 /* >>> is a mb() needed here? */
1852 if (ch->flags & XPC_C_DISCONNECTING) {
1854 * An error occurred between our last error check and
1855 * this one. We will try to clear the type field from
1856 * the notify entry. If we succeed then
1857 * xpc_disconnect_channel() didn't already process
1860 if (cmpxchg(¬ify->type, notify_type, 0) ==
1862 atomic_dec(&ch->n_to_notify);
1866 /* drop the reference grabbed in xpc_allocate_msg() */
1867 xpc_msgqueue_deref(ch);
1872 msg->flags |= XPC_M_READY;
1875 * The preceding store of msg->flags must occur before the following
1876 * load of ch->local_GP->put.
1880 /* see if the message is next in line to be sent, if so send it */
1882 put = ch->local_GP->put;
1883 if (put == msg_number)
1884 xpc_send_msgs(ch, put);
1886 /* drop the reference grabbed in xpc_allocate_msg() */
1887 xpc_msgqueue_deref(ch);
1892 * Send a message previously allocated using xpc_initiate_allocate() on the
1893 * specified channel connected to the specified partition.
1895 * This routine will not wait for the message to be received, nor will
1896 * notification be given when it does happen. Once this routine has returned
1897 * the message entry allocated via xpc_initiate_allocate() is no longer
1898 * accessable to the caller.
1900 * This routine, although called by users, does not call xpc_part_ref() to
1901 * ensure that the partition infrastructure is in place. It relies on the
1902 * fact that we called xpc_msgqueue_ref() in xpc_allocate_msg().
1906 * partid - ID of partition to which the channel is connected.
1907 * ch_number - channel # to send message on.
1908 * payload - pointer to the payload area allocated via
1909 * xpc_initiate_allocate().
1912 xpc_initiate_send(partid_t partid, int ch_number, void *payload)
1914 struct xpc_partition *part = &xpc_partitions[partid];
1915 struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
1916 enum xpc_retval ret;
1918 dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *)msg,
1921 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
1922 DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1923 DBUG_ON(msg == NULL);
1925 ret = xpc_send_msg(&part->channels[ch_number], msg, 0, NULL, NULL);
1931 * Send a message previously allocated using xpc_initiate_allocate on the
1932 * specified channel connected to the specified partition.
1934 * This routine will not wait for the message to be sent. Once this routine
1935 * has returned the message entry allocated via xpc_initiate_allocate() is no
1936 * longer accessable to the caller.
1938 * Once the remote end of the channel has received the message, the function
1939 * passed as an argument to xpc_initiate_send_notify() will be called. This
1940 * allows the sender to free up or re-use any buffers referenced by the
1941 * message, but does NOT mean the message has been processed at the remote
1942 * end by a receiver.
1944 * If this routine returns an error, the caller's function will NOT be called.
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().
1952 * partid - ID of partition to which the channel is connected.
1953 * ch_number - channel # to send message on.
1954 * payload - pointer to the payload area allocated via
1955 * xpc_initiate_allocate().
1956 * func - function to call with asynchronous notification of message
1957 * receipt. THIS FUNCTION MUST BE NON-BLOCKING.
1958 * key - user-defined key to be passed to the function when it's called.
1961 xpc_initiate_send_notify(partid_t partid, int ch_number, void *payload,
1962 xpc_notify_func func, void *key)
1964 struct xpc_partition *part = &xpc_partitions[partid];
1965 struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
1966 enum xpc_retval ret;
1968 dev_dbg(xpc_chan, "msg=0x%p, partid=%d, channel=%d\n", (void *)msg,
1971 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
1972 DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
1973 DBUG_ON(msg == NULL);
1974 DBUG_ON(func == NULL);
1976 ret = xpc_send_msg(&part->channels[ch_number], msg, XPC_N_CALL,
1981 static struct xpc_msg *
1982 xpc_pull_remote_msg(struct xpc_channel *ch, s64 get)
1984 struct xpc_partition *part = &xpc_partitions[ch->partid];
1985 struct xpc_msg *remote_msg, *msg;
1986 u32 msg_index, nmsgs;
1988 enum xpc_retval ret;
1990 if (mutex_lock_interruptible(&ch->msg_to_pull_mutex) != 0) {
1991 /* we were interrupted by a signal */
1995 while (get >= ch->next_msg_to_pull) {
1997 /* pull as many messages as are ready and able to be pulled */
1999 msg_index = ch->next_msg_to_pull % ch->remote_nentries;
2001 DBUG_ON(ch->next_msg_to_pull >= ch->w_remote_GP.put);
2002 nmsgs = ch->w_remote_GP.put - ch->next_msg_to_pull;
2003 if (msg_index + nmsgs > ch->remote_nentries) {
2004 /* ignore the ones that wrap the msg queue for now */
2005 nmsgs = ch->remote_nentries - msg_index;
2008 msg_offset = msg_index * ch->msg_size;
2009 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2010 remote_msg = (struct xpc_msg *)(ch->remote_msgqueue_pa +
2013 ret = xpc_pull_remote_cachelines(part, msg, remote_msg,
2014 nmsgs * ch->msg_size);
2015 if (ret != xpcSuccess) {
2017 dev_dbg(xpc_chan, "failed to pull %d msgs starting with"
2018 " msg %ld from partition %d, channel=%d, "
2019 "ret=%d\n", nmsgs, ch->next_msg_to_pull,
2020 ch->partid, ch->number, ret);
2022 XPC_DEACTIVATE_PARTITION(part, ret);
2024 mutex_unlock(&ch->msg_to_pull_mutex);
2028 ch->next_msg_to_pull += nmsgs;
2031 mutex_unlock(&ch->msg_to_pull_mutex);
2033 /* return the message we were looking for */
2034 msg_offset = (get % ch->remote_nentries) * ch->msg_size;
2035 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue + msg_offset);
2041 * Get a message to be delivered.
2043 static struct xpc_msg *
2044 xpc_get_deliverable_msg(struct xpc_channel *ch)
2046 struct xpc_msg *msg = NULL;
2050 if (ch->flags & XPC_C_DISCONNECTING)
2053 get = ch->w_local_GP.get;
2054 rmb(); /* guarantee that .get loads before .put */
2055 if (get == ch->w_remote_GP.put)
2058 /* There are messages waiting to be pulled and delivered.
2059 * We need to try to secure one for ourselves. We'll do this
2060 * by trying to increment w_local_GP.get and hope that no one
2061 * else beats us to it. If they do, we'll we'll simply have
2062 * to try again for the next one.
2065 if (cmpxchg(&ch->w_local_GP.get, get, get + 1) == get) {
2066 /* we got the entry referenced by get */
2068 dev_dbg(xpc_chan, "w_local_GP.get changed to %ld, "
2069 "partid=%d, channel=%d\n", get + 1,
2070 ch->partid, ch->number);
2072 /* pull the message from the remote partition */
2074 msg = xpc_pull_remote_msg(ch, get);
2076 DBUG_ON(msg != NULL && msg->number != get);
2077 DBUG_ON(msg != NULL && (msg->flags & XPC_M_DONE));
2078 DBUG_ON(msg != NULL && !(msg->flags & XPC_M_READY));
2089 * Deliver a message to its intended recipient.
2092 xpc_deliver_msg(struct xpc_channel *ch)
2094 struct xpc_msg *msg;
2096 msg = xpc_get_deliverable_msg(ch);
2100 * This ref is taken to protect the payload itself from being
2101 * freed before the user is finished with it, which the user
2102 * indicates by calling xpc_initiate_received().
2104 xpc_msgqueue_ref(ch);
2106 atomic_inc(&ch->kthreads_active);
2108 if (ch->func != NULL) {
2109 dev_dbg(xpc_chan, "ch->func() called, msg=0x%p, "
2110 "msg_number=%ld, partid=%d, channel=%d\n",
2111 (void *)msg, msg->number, ch->partid,
2114 /* deliver the message to its intended recipient */
2115 ch->func(xpcMsgReceived, ch->partid, ch->number,
2116 &msg->payload, ch->key);
2118 dev_dbg(xpc_chan, "ch->func() returned, msg=0x%p, "
2119 "msg_number=%ld, partid=%d, channel=%d\n",
2120 (void *)msg, msg->number, ch->partid,
2124 atomic_dec(&ch->kthreads_active);
2129 * Now we actually acknowledge the messages that have been delivered and ack'd
2130 * by advancing the cached remote message queue's Get value and if requested
2131 * send an IPI to the message sender's partition.
2134 xpc_acknowledge_msgs(struct xpc_channel *ch, s64 initial_get, u8 msg_flags)
2136 struct xpc_msg *msg;
2137 s64 get = initial_get + 1;
2143 if (get == ch->w_local_GP.get)
2146 msg = (struct xpc_msg *)((u64)ch->remote_msgqueue +
2147 (get % ch->remote_nentries) *
2150 if (!(msg->flags & XPC_M_DONE))
2153 msg_flags |= msg->flags;
2157 if (get == initial_get) {
2158 /* nothing's changed */
2162 if (cmpxchg_rel(&ch->local_GP->get, initial_get, get) !=
2164 /* someone else beat us to it */
2165 DBUG_ON(ch->local_GP->get <= initial_get);
2169 /* we just set the new value of local_GP->get */
2171 dev_dbg(xpc_chan, "local_GP->get changed to %ld, partid=%d, "
2172 "channel=%d\n", get, ch->partid, ch->number);
2174 send_IPI = (msg_flags & XPC_M_INTERRUPT);
2177 * We need to ensure that the message referenced by
2178 * local_GP->get is not XPC_M_DONE or that local_GP->get
2179 * equals w_local_GP.get, so we'll go have a look.
2185 xpc_IPI_send_msgrequest(ch);
2189 * Acknowledge receipt of a delivered message.
2191 * If a message has XPC_M_INTERRUPT set, send an interrupt to the partition
2192 * that sent the message.
2194 * This function, although called by users, does not call xpc_part_ref() to
2195 * ensure that the partition infrastructure is in place. It relies on the
2196 * fact that we called xpc_msgqueue_ref() in xpc_deliver_msg().
2200 * partid - ID of partition to which the channel is connected.
2201 * ch_number - channel # message received on.
2202 * payload - pointer to the payload area allocated via
2203 * xpc_initiate_allocate().
2206 xpc_initiate_received(partid_t partid, int ch_number, void *payload)
2208 struct xpc_partition *part = &xpc_partitions[partid];
2209 struct xpc_channel *ch;
2210 struct xpc_msg *msg = XPC_MSG_ADDRESS(payload);
2211 s64 get, msg_number = msg->number;
2213 DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
2214 DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
2216 ch = &part->channels[ch_number];
2218 dev_dbg(xpc_chan, "msg=0x%p, msg_number=%ld, partid=%d, channel=%d\n",
2219 (void *)msg, msg_number, ch->partid, ch->number);
2221 DBUG_ON((((u64)msg - (u64)ch->remote_msgqueue) / ch->msg_size) !=
2222 msg_number % ch->remote_nentries);
2223 DBUG_ON(msg->flags & XPC_M_DONE);
2225 msg->flags |= XPC_M_DONE;
2228 * The preceding store of msg->flags must occur before the following
2229 * load of ch->local_GP->get.
2234 * See if this message is next in line to be acknowledged as having
2237 get = ch->local_GP->get;
2238 if (get == msg_number)
2239 xpc_acknowledge_msgs(ch, get, msg->flags);
2241 /* the call to xpc_msgqueue_ref() was done by xpc_deliver_msg() */
2242 xpc_msgqueue_deref(ch);