/* * This file is subject to the terms and conditions of the GNU General Public * License. See the file "COPYING" in the main directory of this archive * for more details. * * (C) Copyright 2020 Hewlett Packard Enterprise Development LP * Copyright (c) 2004-2009 Silicon Graphics, Inc. All Rights Reserved.
*/
/* * Cross Partition Communication (XPC) support - standard version. * * XPC provides a message passing capability that crosses partition * boundaries. This module is made up of two parts: * * partition This part detects the presence/absence of other * partitions. It provides a heartbeat and monitors * the heartbeats of other partitions. * * channel This part manages the channels and sends/receives * messages across them to/from other partitions. * * There are a couple of additional functions residing in XP, which * provide an interface to XPC for its users. * * * Caveats: * * . Currently on sn2, we have no way to determine which nasid an IRQ * came from. Thus, xpc_send_IRQ_sn2() does a remote amo write * followed by an IPI. The amo indicates where data is to be pulled * from, so after the IPI arrives, the remote partition checks the amo * word. The IPI can actually arrive before the amo however, so other * code must periodically check for this case. Also, remote amo * operations do not reliably time out. Thus we do a remote PIO read * solely to know whether the remote partition is down and whether we * should stop sending IPIs to it. This remote PIO read operation is * set up in a special nofault region so SAL knows to ignore (and * cleanup) any errors due to the remote amo write, PIO read, and/or * PIO write operations. * * If/when new hardware solves this IPI problem, we should abandon * the current approach. *
*/
/* * Timer function to enforce the timelimit on the partition disengage.
*/ staticvoid
xpc_timeout_partition_disengage(struct timer_list *t)
{ struct xpc_partition *part = timer_container_of(part, t,
disengage_timer);
/* * Timer to produce the heartbeat. The timer structures function is * already set when this is initially called. A tunable is used to * specify when the next timeout should occur.
*/ staticvoid
xpc_hb_beater(struct timer_list *unused)
{
xpc_arch_ops.increment_heartbeat();
if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
wake_up_interruptible(&xpc_activate_IRQ_wq);
/* * At periodic intervals, scan through all active partitions and ensure * their heartbeat is still active. If not, the partition is deactivated.
*/ staticvoid
xpc_check_remote_hb(void)
{ struct xpc_partition *part; short partid; enum xp_retval ret;
for (partid = 0; partid < xp_max_npartitions; partid++) {
ret = xpc_arch_ops.get_remote_heartbeat(part); if (ret != xpSuccess)
XPC_DEACTIVATE_PARTITION(part, ret);
}
}
/* * This thread is responsible for nearly all of the partition * activation/deactivation.
*/ staticint
xpc_hb_checker(void *ignore)
{ int force_IRQ = 0;
/* this thread was marked active by xpc_hb_init() */
/* wait for IRQ or timeout */
(void)wait_event_interruptible(xpc_activate_IRQ_wq,
(time_is_before_eq_jiffies(
xpc_hb_check_timeout) ||
xpc_activate_IRQ_rcvd > 0 ||
xpc_exiting));
}
xpc_stop_hb_beater();
dev_dbg(xpc_part, "heartbeat checker is exiting\n");
/* mark this thread as having exited */
complete(&xpc_hb_checker_exited); return 0;
}
/* * This thread will attempt to discover other partitions to activate * based on info provided by SAL. This new thread is short lived and * will exit once discovery is complete.
*/ staticint
xpc_initiate_discovery(void *ignore)
{
xpc_discovery();
dev_dbg(xpc_part, "discovery thread is exiting\n");
/* mark this thread as having exited */
complete(&xpc_discovery_exited); return 0;
}
/* * The first kthread assigned to a newly activated partition is the one * created by XPC HB with which it calls xpc_activating(). XPC hangs on to * that kthread until the partition is brought down, at which time that kthread * returns back to XPC HB. (The return of that kthread will signify to XPC HB * that XPC has dismantled all communication infrastructure for the associated * partition.) This kthread becomes the channel manager for that partition. * * Each active partition has a channel manager, who, besides connecting and * disconnecting channels, will ensure that each of the partition's connected * channels has the required number of assigned kthreads to get the work done.
*/ staticvoid
xpc_channel_mgr(struct xpc_partition *part)
{ while (part->act_state != XPC_P_AS_DEACTIVATING ||
atomic_read(&part->nchannels_active) > 0 ||
!xpc_partition_disengaged(part)) {
xpc_process_sent_chctl_flags(part);
/* * Wait until we've been requested to activate kthreads or * all of the channel's message queues have been torn down or * a signal is pending. * * The channel_mgr_requests is set to 1 after being awakened, * This is done to prevent the channel mgr from making one pass * through the loop for each request, since he will * be servicing all the requests in one pass. The reason it's * set to 1 instead of 0 is so that other kthreads will know * that the channel mgr is running and won't bother trying to * wake him up.
*/
atomic_dec(&part->channel_mgr_requests);
(void)wait_event_interruptible(part->channel_mgr_wq,
(atomic_read(&part->channel_mgr_requests) > 0 ||
part->chctl.all_flags != 0 ||
(part->act_state == XPC_P_AS_DEACTIVATING &&
atomic_read(&part->nchannels_active) == 0 &&
xpc_partition_disengaged(part))));
atomic_set(&part->channel_mgr_requests, 1);
}
}
/* * Guarantee that the kzalloc'd memory is cacheline aligned.
*/ void *
xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
{ /* see if kzalloc will give us cachline aligned memory by default */
*base = kzalloc(size, flags); if (*base == NULL) return NULL;
if ((u64)*base == L1_CACHE_ALIGN((u64)*base)) return *base;
kfree(*base);
/* nope, we'll have to do it ourselves */
*base = kzalloc(size + L1_CACHE_BYTES, flags); if (*base == NULL) return NULL;
return (void *)L1_CACHE_ALIGN((u64)*base);
}
/* * Setup the channel structures necessary to support XPartition Communication * between the specified remote partition and the local one.
*/ staticenum xp_retval
xpc_setup_ch_structures(struct xpc_partition *part)
{ enum xp_retval ret; int ch_number; struct xpc_channel *ch; short partid = XPC_PARTID(part);
/* * Allocate all of the channel structures as a contiguous chunk of * memory.
*/
DBUG_ON(part->channels != NULL);
part->channels = kcalloc(XPC_MAX_NCHANNELS, sizeof(struct xpc_channel),
GFP_KERNEL); if (part->channels == NULL) {
dev_err(xpc_chan, "can't get memory for channels\n"); return xpNoMemory;
}
/* allocate the remote open and close args */
part->remote_openclose_args =
xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
GFP_KERNEL, &part->
remote_openclose_args_base); if (part->remote_openclose_args == NULL) {
dev_err(xpc_chan, "can't get memory for remote connect args\n");
ret = xpNoMemory; goto out_1;
}
ret = xpc_arch_ops.setup_ch_structures(part); if (ret != xpSuccess) goto out_2;
/* * With the setting of the partition setup_state to XPC_P_SS_SETUP, * we're declaring that this partition is ready to go.
*/
part->setup_state = XPC_P_SS_SETUP;
/* * Teardown the channel structures necessary to support XPartition Communication * between the specified remote partition and the local one.
*/ staticvoid
xpc_teardown_ch_structures(struct xpc_partition *part)
{
DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
DBUG_ON(atomic_read(&part->nchannels_active) != 0);
/* * Make this partition inaccessible to local processes by marking it * as no longer setup. Then wait before proceeding with the teardown * until all existing references cease.
*/
DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
part->setup_state = XPC_P_SS_WTEARDOWN;
/* * 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.
*/ staticint
xpc_activating(void *__partid)
{ short partid = (u64)__partid; struct xpc_partition *part = &xpc_partitions[partid]; unsignedlong irq_flags;
void
xpc_activate_kthreads(struct xpc_channel *ch, int needed)
{ int idle = atomic_read(&ch->kthreads_idle); int assigned = atomic_read(&ch->kthreads_assigned); int wakeup;
/* only wakeup the requested number of kthreads */
wake_up_nr(&ch->idle_wq, wakeup);
}
if (needed <= 0) return;
if (needed + assigned > ch->kthreads_assigned_limit) {
needed = ch->kthreads_assigned_limit - assigned; if (needed <= 0) return;
}
dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
needed, ch->partid, ch->number);
xpc_create_kthreads(ch, needed, 0);
}
/* * This function is where XPC's kthreads wait for messages to deliver.
*/ staticvoid
xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
{ int (*n_of_deliverable_payloads) (struct xpc_channel *) =
xpc_arch_ops.n_of_deliverable_payloads;
do { /* deliver messages to their intended recipients */
if (atomic_inc_return(&ch->kthreads_idle) >
ch->kthreads_idle_limit) { /* too many idle kthreads on this channel */
atomic_dec(&ch->kthreads_idle); break;
}
/* * It is possible that while the callout was being * made that the remote partition sent some messages. * If that is the case, we may need to activate * additional kthreads to help deliver them. We only * need one less than total #of messages to deliver.
*/
n_needed = n_of_deliverable_payloads(ch) - 1; if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
xpc_activate_kthreads(ch, n_needed);
/* * 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)
{ unsignedlong 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) {
/* * 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;
}
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_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_disconnect_wait(int ch_number)
{ unsignedlong 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 = 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;
}
if (wakeup_channel_mgr)
xpc_wakeup_channel_mgr(part);
xpc_part_deref(part);
}
}
staticint
xpc_setup_partitions(void)
{ short partid; struct xpc_partition *part;
xpc_partitions = kcalloc(xp_max_npartitions, sizeof(struct xpc_partition),
GFP_KERNEL); if (xpc_partitions == NULL) {
dev_err(xpc_part, "can't get memory for partition structure\n"); return -ENOMEM;
}
/* * The first few fields of each entry of xpc_partitions[] need to * be initialized now so that calls to xpc_connect() and * xpc_disconnect() can be made prior to the activation of any remote * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING * PARTITION HAS BEEN ACTIVATED.
*/ for (partid = 0; partid < xp_max_npartitions; partid++) {
part = &xpc_partitions[partid];
staticvoid
xpc_do_exit(enum xp_retval reason)
{ short partid; int active_part_count, printed_waiting_msg = 0; struct xpc_partition *part; unsignedlong printmsg_time, disengage_timeout = 0;
/* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
DBUG_ON(xpc_exiting == 1);
/* * Let the heartbeat checker thread and the discovery thread * (if one is running) know that they should exit. Also wake up * the heartbeat checker thread in case it's sleeping.
*/
xpc_exiting = 1;
wake_up_interruptible(&xpc_activate_IRQ_wq);
/* wait for the discovery thread to exit */
wait_for_completion(&xpc_discovery_exited);
/* wait for the heartbeat checker thread to exit */
wait_for_completion(&xpc_hb_checker_exited);
/* sleep for a 1/3 of a second or so */
(void)msleep_interruptible(300);
for (partid = 0; partid < xp_max_npartitions; partid++) {
part = &xpc_partitions[partid];
if (xpc_partition_disengaged(part) &&
part->act_state == XPC_P_AS_INACTIVE) { continue;
}
active_part_count++;
XPC_DEACTIVATE_PARTITION(part, reason);
if (part->disengage_timeout > disengage_timeout)
disengage_timeout = part->disengage_timeout;
}
if (xpc_arch_ops.any_partition_engaged()) { if (time_is_before_jiffies(printmsg_time)) {
dev_info(xpc_part, "waiting for remote " "partitions to deactivate, timeout in " "%ld seconds\n", (disengage_timeout -
jiffies) / HZ);
printmsg_time = jiffies +
(XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
printed_waiting_msg = 1;
}
} elseif (active_part_count > 0) { if (printed_waiting_msg) {
dev_info(xpc_part, "waiting for local partition" " to deactivate\n");
printed_waiting_msg = 0;
}
} else { if (!xpc_disengage_timedout) {
dev_info(xpc_part, "all partitions have " "deactivated\n");
} break;
}
/* sleep for a 1/3 of a second or so */
(void)msleep_interruptible(300);
} while (1);
DBUG_ON(xpc_arch_ops.any_partition_engaged());
xpc_teardown_rsvd_page();
if (reason == xpUnloading) {
(void)unregister_die_notifier(&xpc_die_notifier);
(void)unregister_reboot_notifier(&xpc_reboot_notifier);
}
/* clear the interface to XPC's functions */
xpc_clear_interface();
if (xpc_sysctl)
unregister_sysctl_table(xpc_sysctl); if (xpc_sysctl_hb)
unregister_sysctl_table(xpc_sysctl_hb);
xpc_teardown_partitions();
if (is_uv_system())
xpc_exit_uv();
}
/* * This function is called when the system is being rebooted.
*/ staticint
xpc_system_reboot(struct notifier_block *nb, unsignedlong event, void *unused)
{ enum xp_retval reason;
switch (event) { case SYS_RESTART:
reason = xpSystemReboot; break; case SYS_HALT:
reason = xpSystemHalt; break; case SYS_POWER_OFF:
reason = xpSystemPoweroff; break; default:
reason = xpSystemGoingDown;
}
xpc_do_exit(reason); return NOTIFY_DONE;
}
/* Used to only allow one cpu to complete disconnect */ staticunsignedint xpc_die_disconnecting;
/* * Notify other partitions to deactivate from us by first disengaging from all * references to our memory.
*/ staticvoid
xpc_die_deactivate(void)
{ struct xpc_partition *part; short partid; int any_engaged; long keep_waiting; long wait_to_print;
if (cmpxchg(&xpc_die_disconnecting, 0, 1)) return;
/* keep xpc_hb_checker thread from doing anything (just in case) */
xpc_exiting = 1;
/* * Though we requested that all other partitions deactivate from us, * we only wait until they've all disengaged or we've reached the * defined timelimit. * * Given that one iteration through the following while-loop takes * approximately 200 microseconds, calculate the #of loops to take * before bailing and the #of loops before printing a waiting message.
*/
keep_waiting = xpc_disengage_timelimit * 1000 * 5;
wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
while (1) {
any_engaged = xpc_arch_ops.any_partition_engaged(); if (!any_engaged) {
dev_info(xpc_part, "all partitions have deactivated\n"); break;
}
if (!keep_waiting--) { for (partid = 0; partid < xp_max_npartitions;
partid++) { if (xpc_arch_ops.partition_engaged(partid)) {
dev_info(xpc_part, "deactivate from " "remote partition %d timed " "out\n", partid);
}
} break;
}
if (!wait_to_print--) {
dev_info(xpc_part, "waiting for remote partitions to " "deactivate, timeout in %ld seconds\n",
keep_waiting / (1000 * 5));
wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1000 * 5;
}
udelay(200);
}
}
/* * This function is called when the system is being restarted or halted due * to some sort of system failure. If this is the case we need to notify the * other partitions to disengage from all references to our memory. * This function can also be called when our heartbeater could be offlined * for a time. In this case we need to notify other partitions to not worry * about the lack of a heartbeat.
*/ staticint
xpc_system_die(struct notifier_block *nb, unsignedlong event, void *_die_args)
{ struct die_args *die_args = _die_args;
switch (event) { case DIE_TRAP: if (die_args->trapnr == X86_TRAP_DF)
xpc_die_deactivate();
/* * Fill the partition reserved page with the information needed by * other partitions to discover we are alive and establish initial * communications.
*/
ret = xpc_setup_rsvd_page(); if (ret != 0) {
dev_err(xpc_part, "can't setup our reserved page\n"); goto out_2;
}
/* add ourselves to the reboot_notifier_list */
ret = register_reboot_notifier(&xpc_reboot_notifier); if (ret != 0)
dev_warn(xpc_part, "can't register reboot notifier\n");
/* add ourselves to the die_notifier list */
ret = register_die_notifier(&xpc_die_notifier); if (ret != 0)
dev_warn(xpc_part, "can't register die notifier\n");
/* * The real work-horse behind xpc. This processes incoming * interrupts and monitors remote heartbeats.
*/
kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME); if (IS_ERR(kthread)) {
dev_err(xpc_part, "failed while forking hb check thread\n");
ret = -EBUSY; goto out_3;
}
/* * Startup a thread that will attempt to discover other partitions to * activate based on info provided by SAL. This new thread is short * lived and will exit once discovery is complete.
*/
kthread = kthread_run(xpc_initiate_discovery, NULL,
XPC_DISCOVERY_THREAD_NAME); if (IS_ERR(kthread)) {
dev_err(xpc_part, "failed while forking discovery thread\n");
/* mark this new thread as a non-starter */
complete(&xpc_discovery_exited);
xpc_do_exit(xpUnloading); return -EBUSY;
}
/* set the interface to point at XPC's functions */
xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
xpc_initiate_send, xpc_initiate_send_notify,
xpc_initiate_received, xpc_initiate_partid_to_nasids);
return 0;
/* initialization was not successful */
out_3:
xpc_teardown_rsvd_page();
(void)unregister_die_notifier(&xpc_die_notifier);
(void)unregister_reboot_notifier(&xpc_reboot_notifier);
out_2: if (xpc_sysctl_hb)
unregister_sysctl_table(xpc_sysctl_hb); if (xpc_sysctl)
unregister_sysctl_table(xpc_sysctl);
xpc_teardown_partitions();
out_1: if (is_uv_system())
xpc_exit_uv(); return ret;
}
MODULE_AUTHOR("Silicon Graphics, Inc.");
MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
MODULE_LICENSE("GPL");
module_param(xpc_hb_interval, int, 0);
MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between " "heartbeat increments.");
module_param(xpc_hb_check_interval, int, 0);
MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between " "heartbeat checks.");
module_param(xpc_disengage_timelimit, int, 0);
MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait " "for disengage to complete.");
module_param(xpc_kdebug_ignore, int, 0);
MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by " "other partitions when dropping into kdebug.");
Messung V0.5
¤ Dauer der Verarbeitung: 0.4 Sekunden
(vorverarbeitet)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.