void
xpc_partition_going_down(struct xpc_partition *part, enum xp_retval reason)
{
unsigned long irq_flags;
int ch_number;
struct xpc_channel *ch;
dev_dbg(xpc_chan, "deactivating partition %d, reason=%d\n",
XPC_PARTID(part), reason);
if (!xpc_part_ref(part)) {
return;
}
for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
ch = &part->channels[ch_number];
xpc_msgqueue_ref(ch);
spin_lock_irqsave(&ch->lock, irq_flags);
XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
spin_unlock_irqrestore(&ch->lock, irq_flags);
xpc_msgqueue_deref(ch);
}
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
/*
* XPC's heartbeat code calls this function to inform XPC that a partition is
* going down. XPC responds by tearing down the XPartition Communication
* infrastructure used for the just downed partition.
*
* XPC's heartbeat code will never call this function and xpc_partition_up()
* at the same time. Nor will it ever make multiple calls to either function
* at the same time.
*/
void
xpc_partition_going_down(struct xpc_partition *part, enum xp_retval reason)
{
unsigned long irq_flags;
int ch_number;
struct xpc_channel *ch;
dev_dbg(xpc_chan, "deactivating partition %d, reason=%d\n",
XPC_PARTID(part), reason);
if (!xpc_part_ref(part)) {
/* infrastructure for this partition isn't currently set up */
return;
}
/* disconnect channels associated with the partition going down */
for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
ch = &part->channels[ch_number];
xpc_msgqueue_ref(ch);
spin_lock_irqsave(&ch->lock, irq_flags);
XPC_DISCONNECT_CHANNEL(ch, reason, &irq_flags);
spin_unlock_irqrestore(&ch->lock, irq_flags);
xpc_msgqueue_deref(ch);
}
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
static irqreturn_t
xpc_handle_notify_IRQ_uv(int irq, void *dev_id)
{
struct xpc_notify_mq_msg_uv *msg;
short partid;
struct xpc_partition *part;
while ((msg = gru_get_next_message(xpc_notify_mq_uv->gru_mq_desc)) !=
NULL) {
partid = msg->hdr.partid;
if (partid < 0 || partid >= XP_MAX_NPARTITIONS_UV) {
dev_err(xpc_part, "xpc_handle_notify_IRQ_uv() received "
"invalid partid=0x%x in message\n", partid);
} else {
part = &xpc_partitions[partid];
if (xpc_part_ref(part)) {
xpc_handle_notify_mq_msg_uv(part, msg);
xpc_part_deref(part);
}
}
gru_free_message(xpc_notify_mq_uv->gru_mq_desc, msg);
}
return IRQ_HANDLED;
}
/*
* When XPC HB determines that a partition has come up, it will create a new
* kthread and that kthread will call this function to attempt to set up the
* basic infrastructure used for Cross Partition Communication with the newly
* upped partition.
*
* The kthread that was created by XPC HB and which setup the XPC
* infrastructure will remain assigned to the partition until the partition
* goes down. At which time the kthread will teardown the XPC infrastructure
* and then exit.
*
* XPC HB will put the remote partition's XPC per partition specific variables
* physical address into xpc_partitions[partid].remote_vars_part_pa prior to
* calling xpc_partition_up().
*/
static void
xpc_partition_up(struct xpc_partition *part)
{
DBUG_ON(part->channels != NULL);
dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part));
if (xpc_setup_infrastructure(part) != xpSuccess)
return;
/*
* The kthread that XPC HB called us with will become the
* channel manager for this partition. It will not return
* back to XPC HB until the partition's XPC infrastructure
* has been dismantled.
*/
(void)xpc_part_ref(part); /* this will always succeed */
if (xpc_make_first_contact(part) == xpSuccess)
xpc_channel_mgr(part);
xpc_part_deref(part);
xpc_teardown_infrastructure(part);
}
/*
* Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
* because the write to their associated IPI amo completed after the IRQ/IPI
* was received.
*/
void
xpc_dropped_IPI_check(struct xpc_partition *part)
{
if (xpc_part_ref(part)) {
xpc_check_for_channel_activity(part);
part->dropped_IPI_timer.expires = jiffies +
XPC_P_DROPPED_IPI_WAIT;
add_timer(&part->dropped_IPI_timer);
xpc_part_deref(part);
}
}
/*
* When XPC HB determines that a partition has come up, it will create a new
* kthread and that kthread will call this function to attempt to set up the
* basic infrastructure used for Cross Partition Communication with the newly
* upped partition.
*
* The kthread that was created by XPC HB and which setup the XPC
* infrastructure will remain assigned to the partition becoming the channel
* manager for that partition until the partition is deactivating, at which
* time the kthread will teardown the XPC infrastructure and then exit.
*/
static int
xpc_activating(void *__partid)
{
short partid = (u64)__partid;
struct xpc_partition *part = &xpc_partitions[partid];
unsigned long irq_flags;
DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
spin_lock_irqsave(&part->act_lock, irq_flags);
if (part->act_state == XPC_P_AS_DEACTIVATING) {
part->act_state = XPC_P_AS_INACTIVE;
spin_unlock_irqrestore(&part->act_lock, irq_flags);
part->remote_rp_pa = 0;
return 0;
}
/* indicate the thread is activating */
DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
part->act_state = XPC_P_AS_ACTIVATING;
XPC_SET_REASON(part, 0, 0);
spin_unlock_irqrestore(&part->act_lock, irq_flags);
dev_dbg(xpc_part, "activating partition %d\n", partid);
xpc_allow_hb(partid);
if (xpc_setup_ch_structures(part) == xpSuccess) {
(void)xpc_part_ref(part); /* this will always succeed */
if (xpc_make_first_contact(part) == xpSuccess) {
xpc_mark_partition_active(part);
xpc_channel_mgr(part);
/* won't return until partition is deactivating */
}
xpc_part_deref(part);
xpc_teardown_ch_structures(part);
}
xpc_disallow_hb(partid);
xpc_mark_partition_inactive(part);
if (part->reason == xpReactivating) {
/* interrupting ourselves results in activating partition */
xpc_request_partition_reactivation(part);
}
return 0;
}
void
xpc_disconnect_wait(int ch_number)
{
unsigned long irq_flags;
short partid;
struct xpc_partition *part;
struct xpc_channel *ch;
int wakeup_channel_mgr;
/* now wait for all callouts to the caller's function to cease */
for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
part = &xpc_partitions[partid];
if (!xpc_part_ref(part))
continue;
ch = &part->channels[ch_number];
if (!(ch->flags & XPC_C_WDISCONNECT)) {
xpc_part_deref(part);
continue;
}
wait_for_completion(&ch->wdisconnect_wait);
spin_lock_irqsave(&ch->lock, irq_flags);
DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
wakeup_channel_mgr = 0;
if (ch->delayed_IPI_flags) {
if (part->act_state != XPC_P_DEACTIVATING) {
spin_lock(&part->IPI_lock);
XPC_SET_IPI_FLAGS(part->local_IPI_amo,
ch->number,
ch->delayed_IPI_flags);
spin_unlock(&part->IPI_lock);
wakeup_channel_mgr = 1;
}
ch->delayed_IPI_flags = 0;
}
ch->flags &= ~XPC_C_WDISCONNECT;
spin_unlock_irqrestore(&ch->lock, irq_flags);
if (wakeup_channel_mgr)
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
}
void
xpc_disconnect_wait(int ch_number)
{
unsigned long irq_flags;
short partid;
struct xpc_partition *part;
struct xpc_channel *ch;
int wakeup_channel_mgr;
/* */
for (partid = 0; partid < xp_max_npartitions; partid++) {
part = &xpc_partitions[partid];
if (!xpc_part_ref(part))
continue;
ch = &part->channels[ch_number];
if (!(ch->flags & XPC_C_WDISCONNECT)) {
xpc_part_deref(part);
continue;
}
wait_for_completion(&ch->wdisconnect_wait);
spin_lock_irqsave(&ch->lock, irq_flags);
DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
wakeup_channel_mgr = 0;
if (ch->delayed_chctl_flags) {
if (part->act_state != XPC_P_AS_DEACTIVATING) {
spin_lock(&part->chctl_lock);
part->chctl.flags[ch->number] |=
ch->delayed_chctl_flags;
spin_unlock(&part->chctl_lock);
wakeup_channel_mgr = 1;
}
ch->delayed_chctl_flags = 0;
}
ch->flags &= ~XPC_C_WDISCONNECT;
spin_unlock_irqrestore(&ch->lock, irq_flags);
if (wakeup_channel_mgr)
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
}
/*
* Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
* partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
* than one partition, we use an AMO_t structure per partition to indicate
* whether a partition has sent an IPI or not. If it has, then wake up the
* associated kthread to handle it.
*
* All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
* running on other partitions.
*
* Noteworthy Arguments:
*
* irq - Interrupt ReQuest number. NOT USED.
*
* dev_id - partid of IPI's potential sender.
*/
irqreturn_t
xpc_notify_IRQ_handler(int irq, void *dev_id)
{
short partid = (short)(u64)dev_id;
struct xpc_partition *part = &xpc_partitions[partid];
DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
if (xpc_part_ref(part)) {
xpc_check_for_channel_activity(part);
xpc_part_deref(part);
}
return IRQ_HANDLED;
}
void
xpc_initiate_connect(int ch_number)
{
short partid;
struct xpc_partition *part;
struct xpc_channel *ch;
DBUG_ON(ch_number < 0 || ch_number >= XPC_MAX_NCHANNELS);
for (partid = 0; partid < xp_max_npartitions; partid++) {
part = &xpc_partitions[partid];
if (xpc_part_ref(part)) {
ch = &part->channels[ch_number];
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
}
}
/*
* Send a message that contains the user's payload on the specified channel
* connected to the specified partition.
*
* NOTE that this routine can sleep waiting for a message entry to become
* available. To not sleep, pass in the XPC_NOWAIT flag.
*
* Once sent, this routine will not wait for the message to be received, nor
* will notification be given when it does happen.
*
* Arguments:
*
* partid - ID of partition to which the channel is connected.
* ch_number - channel # to send message on.
* flags - see xp.h for valid flags.
* payload - pointer to the payload which is to be sent.
* payload_size - size of the payload in bytes.
*/
enum xp_retval
xpc_initiate_send(short partid, int ch_number, u32 flags, void *payload,
u16 payload_size)
{
struct xpc_partition *part = &xpc_partitions[partid];
enum xp_retval ret = xpUnknownReason;
dev_dbg(xpc_chan, "payload=0x%p, partid=%d, channel=%d\n", payload,
partid, ch_number);
DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
DBUG_ON(ch_number < 0 || ch_number >= part->nchannels);
DBUG_ON(payload == NULL);
if (xpc_part_ref(part)) {
ret = xpc_send_payload(&part->channels[ch_number], flags,
payload, payload_size, 0, NULL, NULL);
xpc_part_deref(part);
}
return ret;
}
static irqreturn_t
xpc_handle_activate_IRQ_uv(int irq, void *dev_id)
{
struct xpc_activate_mq_msghdr_uv *msg_hdr;
short partid;
struct xpc_partition *part;
int wakeup_hb_checker = 0;
int part_referenced;
while (1) {
msg_hdr = gru_get_next_message(xpc_activate_mq_uv->gru_mq_desc);
if (msg_hdr == NULL)
break;
partid = msg_hdr->partid;
if (partid < 0 || partid >= XP_MAX_NPARTITIONS_UV) {
dev_err(xpc_part, "xpc_handle_activate_IRQ_uv() "
"received invalid partid=0x%x in message\n",
partid);
} else {
part = &xpc_partitions[partid];
part_referenced = xpc_part_ref(part);
xpc_handle_activate_mq_msg_uv(part, msg_hdr,
part_referenced,
&wakeup_hb_checker);
if (part_referenced)
xpc_part_deref(part);
}
gru_free_message(xpc_activate_mq_uv->gru_mq_desc, msg_hdr);
}
if (wakeup_hb_checker)
wake_up_interruptible(&xpc_activate_IRQ_wq);
return IRQ_HANDLED;
}
/*
* Called by XP at the time of channel connection unregistration to cause
* XPC to teardown all current connections for the specified channel.
*
* Before returning xpc_initiate_disconnect() will wait until all connections
* on the specified channel have been closed/torndown. So the caller can be
* assured that they will not be receiving any more callouts from XPC to the
* function they registered via xpc_connect().
*
* Arguments:
*
* ch_number - channel # to unregister.
*/
void
xpc_initiate_disconnect(int ch_number)
{
unsigned long irq_flags;
short partid;
struct xpc_partition *part;
struct xpc_channel *ch;
DBUG_ON(ch_number < 0 || ch_number >= XPC_MAX_NCHANNELS);
/* initiate the channel disconnect for every active partition */
for (partid = 0; partid < xp_max_npartitions; partid++) {
part = &xpc_partitions[partid];
if (xpc_part_ref(part)) {
ch = &part->channels[ch_number];
xpc_msgqueue_ref(ch);
spin_lock_irqsave(&ch->lock, irq_flags);
if (!(ch->flags & XPC_C_DISCONNECTED)) {
ch->flags |= XPC_C_WDISCONNECT;
XPC_DISCONNECT_CHANNEL(ch, xpUnregistering,
&irq_flags);
}
spin_unlock_irqrestore(&ch->lock, irq_flags);
xpc_msgqueue_deref(ch);
xpc_part_deref(part);
}
}
xpc_disconnect_wait(ch_number);
}
/*
* Called by XP at the time of channel connection registration to cause
* XPC to establish connections to all currently active partitions.
*/
void
xpc_initiate_connect(int ch_number)
{
short partid;
struct xpc_partition *part;
struct xpc_channel *ch;
DBUG_ON(ch_number < 0 || ch_number >= XPC_MAX_NCHANNELS);
for (partid = 0; partid < xp_max_npartitions; partid++) {
part = &xpc_partitions[partid];
if (xpc_part_ref(part)) {
ch = &part->channels[ch_number];
/*
* Initiate the establishment of a connection on the
* newly registered channel to the remote partition.
*/
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
}
}
/*
* For each partition that XPC has established communications with, there is
* a minimum of one kernel thread assigned to perform any operation that
* may potentially sleep or block (basically the callouts to the asynchronous
* functions registered via xpc_connect()).
*
* Additional kthreads are created and destroyed by XPC as the workload
* demands.
*
* A kthread is assigned to one of the active channels that exists for a given
* partition.
*/
void
xpc_create_kthreads(struct xpc_channel *ch, int needed,
int ignore_disconnecting)
{
unsigned long irq_flags;
u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
struct xpc_partition *part = &xpc_partitions[ch->partid];
struct task_struct *kthread;
while (needed-- > 0) {
/*
* The following is done on behalf of the newly created
* kthread. That kthread is responsible for doing the
* counterpart to the following before it exits.
*/
if (ignore_disconnecting) {
if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
/* kthreads assigned had gone to zero */
BUG_ON(!(ch->flags &
XPC_C_DISCONNECTINGCALLOUT_MADE));
break;
}
} else if (ch->flags & XPC_C_DISCONNECTING) {
break;
} else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
atomic_inc_return(&part->nchannels_engaged) == 1) {
xpc_indicate_partition_engaged(part);
}
(void)xpc_part_ref(part);
xpc_msgqueue_ref(ch);
kthread = kthread_run(xpc_kthread_start, (void *)args,
"xpc%02dc%d", ch->partid, ch->number);
if (IS_ERR(kthread)) {
/* the fork failed */
/*
* NOTE: if (ignore_disconnecting &&
* !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
* then we'll deadlock if all other kthreads assigned
* to this channel are blocked in the channel's
* registerer, because the only thing that will unblock
* them is the xpDisconnecting callout that this
* failed kthread_run() would have made.
*/
if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
atomic_dec_return(&part->nchannels_engaged) == 0) {
xpc_indicate_partition_disengaged(part);
}
xpc_msgqueue_deref(ch);
xpc_part_deref(part);
if (atomic_read(&ch->kthreads_assigned) <
ch->kthreads_idle_limit) {
/*
* Flag this as an error only if we have an
* insufficient #of kthreads for the channel
* to function.
*/
spin_lock_irqsave(&ch->lock, irq_flags);
XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
&irq_flags);
spin_unlock_irqrestore(&ch->lock, irq_flags);
}
break;
}
}
}
void
xpc_create_kthreads(struct xpc_channel *ch, int needed,
int ignore_disconnecting)
{
unsigned long irq_flags;
u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
struct xpc_partition *part = &xpc_partitions[ch->partid];
struct task_struct *kthread;
void (*indicate_partition_disengaged) (struct xpc_partition *) =
xpc_arch_ops.indicate_partition_disengaged;
while (needed-- > 0) {
/*
*/
if (ignore_disconnecting) {
if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
/* */
BUG_ON(!(ch->flags &
XPC_C_DISCONNECTINGCALLOUT_MADE));
break;
}
} else if (ch->flags & XPC_C_DISCONNECTING) {
break;
} else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
atomic_inc_return(&part->nchannels_engaged) == 1) {
xpc_arch_ops.indicate_partition_engaged(part);
}
(void)xpc_part_ref(part);
xpc_msgqueue_ref(ch);
kthread = kthread_run(xpc_kthread_start, (void *)args,
"xpc%02dc%d", ch->partid, ch->number);
if (IS_ERR(kthread)) {
/* */
/*
*/
if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
atomic_dec_return(&part->nchannels_engaged) == 0) {
indicate_partition_disengaged(part);
}
xpc_msgqueue_deref(ch);
xpc_part_deref(part);
if (atomic_read(&ch->kthreads_assigned) <
ch->kthreads_idle_limit) {
/*
*/
spin_lock_irqsave(&ch->lock, irq_flags);
XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
&irq_flags);
spin_unlock_irqrestore(&ch->lock, irq_flags);
}
break;
}
}
}