xpc_main.c

linux 内核源代码

/*
 * 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.
 *
 * Copyright (c) 2004-2007 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:
 *
 *	  . We currently have no way to determine which nasid an IPI came
 *	    from. Thus, xpc_IPI_send() 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.
 *
 */


#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/init.h>
#include <linux/sched.h>
#include <linux/syscalls.h>
#include <linux/cache.h>
#include <linux/interrupt.h>
#include <linux/delay.h>
#include <linux/reboot.h>
#include <linux/completion.h>
#include <linux/kdebug.h>
#include <asm/sn/intr.h>
#include <asm/sn/sn_sal.h>
#include <asm/uaccess.h>
#include <asm/sn/xpc.h>


/* define two XPC debug device structures to be used with dev_dbg() et al */

struct device_driver xpc_dbg_name = {
	.name = "xpc"
};

struct device xpc_part_dbg_subname = {
	.bus_id = {0},		/* set to "part" at xpc_init() time */
	.driver = &xpc_dbg_name
};

struct device xpc_chan_dbg_subname = {
	.bus_id = {0},		/* set to "chan" at xpc_init() time */
	.driver = &xpc_dbg_name
};

struct device *xpc_part = &xpc_part_dbg_subname;
struct device *xpc_chan = &xpc_chan_dbg_subname;


static int xpc_kdebug_ignore;


/* systune related variables for /proc/sys directories */

static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
static int xpc_hb_min_interval = 1;
static int xpc_hb_max_interval = 10;

static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
static int xpc_hb_check_min_interval = 10;
static int xpc_hb_check_max_interval = 120;

int xpc_disengage_request_timelimit = XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT;
static int xpc_disengage_request_min_timelimit = 0;
static int xpc_disengage_request_max_timelimit = 120;

static ctl_table xpc_sys_xpc_hb_dir[] = {
	{
		.ctl_name 	= CTL_UNNUMBERED,
		.procname	= "hb_interval",
		.data		= &xpc_hb_interval,
		.maxlen		= sizeof(int),
		.mode		= 0644,
		.proc_handler	= &proc_dointvec_minmax,
		.strategy	= &sysctl_intvec,
		.extra1		= &xpc_hb_min_interval,
		.extra2		= &xpc_hb_max_interval
	},
	{
		.ctl_name	= CTL_UNNUMBERED,
		.procname	= "hb_check_interval",
		.data		= &xpc_hb_check_interval,
		.maxlen		= sizeof(int),
		.mode		= 0644,
		.proc_handler	= &proc_dointvec_minmax,
		.strategy	= &sysctl_intvec,
		.extra1		= &xpc_hb_check_min_interval,
		.extra2		= &xpc_hb_check_max_interval
	},
	{}
};
static ctl_table xpc_sys_xpc_dir[] = {
	{
		.ctl_name	= CTL_UNNUMBERED,
		.procname	= "hb",
		.mode		= 0555,
		.child		= xpc_sys_xpc_hb_dir
	},
	{
		.ctl_name	= CTL_UNNUMBERED,
		.procname	= "disengage_request_timelimit",
		.data		= &xpc_disengage_request_timelimit,
		.maxlen		= sizeof(int),
		.mode		= 0644,
		.proc_handler	= &proc_dointvec_minmax,
		.strategy	= &sysctl_intvec,
		.extra1		= &xpc_disengage_request_min_timelimit,
		.extra2		= &xpc_disengage_request_max_timelimit
	},
	{}
};
static ctl_table xpc_sys_dir[] = {
	{
		.ctl_name	= CTL_UNNUMBERED,
		.procname	= "xpc",
		.mode		= 0555,
		.child		= xpc_sys_xpc_dir
	},
	{}
};
static struct ctl_table_header *xpc_sysctl;

/* non-zero if any remote partition disengage request was timed out */
int xpc_disengage_request_timedout;

/* #of IRQs received */
static atomic_t xpc_act_IRQ_rcvd;

/* IRQ handler notifies this wait queue on receipt of an IRQ */
static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq);

static unsigned long xpc_hb_check_timeout;

/* notification that the xpc_hb_checker thread has exited */
static DECLARE_COMPLETION(xpc_hb_checker_exited);

/* notification that the xpc_discovery thread has exited */
static DECLARE_COMPLETION(xpc_discovery_exited);


static struct timer_list xpc_hb_timer;


static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);


static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
static struct notifier_block xpc_reboot_notifier = {
	.notifier_call = xpc_system_reboot,
};

static int xpc_system_die(struct notifier_block *, unsigned long, void *);
static struct notifier_block xpc_die_notifier = {
	.notifier_call = xpc_system_die,
};


/*
 * Timer function to enforce the timelimit on the partition disengage request.
 */
static void
xpc_timeout_partition_disengage_request(unsigned long data)
{
	struct xpc_partition *part = (struct xpc_partition *) data;


	DBUG_ON(jiffies < part->disengage_request_timeout);

	(void) xpc_partition_disengaged(part);

	DBUG_ON(part->disengage_request_timeout != 0);
	DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part)) != 0);
}


/*
 * Notify the heartbeat check thread that an IRQ has been received.
 */
static irqreturn_t
xpc_act_IRQ_handler(int irq, void *dev_id)
{
	atomic_inc(&xpc_act_IRQ_rcvd);
	wake_up_interruptible(&xpc_act_IRQ_wq);
	return IRQ_HANDLED;
}


/*
 * 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.
 */
static void
xpc_hb_beater(unsigned long dummy)
{
	xpc_vars->heartbeat++;

	if (jiffies >= xpc_hb_check_timeout) {
		wake_up_interruptible(&xpc_act_IRQ_wq);
	}

	xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
	add_timer(&xpc_hb_timer);
}


/*
 * This thread is responsible for nearly all of the partition
 * activation/deactivation.
 */
static int
xpc_hb_checker(void *ignore)
{
	int last_IRQ_count = 0;
	int new_IRQ_count;
	int force_IRQ=0;


	/* this thread was marked active by xpc_hb_init() */

	daemonize(XPC_HB_CHECK_THREAD_NAME);

	set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU));

	/* set our heartbeating to other partitions into motion */
	xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
	xpc_hb_beater(0);

	while (!(volatile int) xpc_exiting) {

		dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
			"been received\n",
			(int) (xpc_hb_check_timeout - jiffies),
			atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count);


		/* checking of remote heartbeats is skewed by IRQ handling */
		if (jiffies >= xpc_hb_check_timeout) {
			dev_dbg(xpc_part, "checking remote heartbeats\n");
			xpc_check_remote_hb();

			/*
			 * We need to periodically recheck to ensure no
			 * IPI/AMO pairs have been missed.  That check
			 * must always reset xpc_hb_check_timeout.
			 */
			force_IRQ = 1;
		}


		/* check for outstanding IRQs */
		new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd);
		if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
			force_IRQ = 0;

			dev_dbg(xpc_part, "found an IRQ to process; will be "
				"resetting xpc_hb_check_timeout\n");

			last_IRQ_count += xpc_identify_act_IRQ_sender();
			if (last_IRQ_count < new_IRQ_count) {
				/* retry once to help avoid missing AMO */
				(void) xpc_identify_act_IRQ_sender();
			}
			last_IRQ_count = new_IRQ_count;

			xpc_hb_check_timeout = jiffies +
					   (xpc_hb_check_interval * HZ);
		}

		/* wait for IRQ or timeout */
		(void) wait_event_interruptible(xpc_act_IRQ_wq,
			    (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) ||
					jiffies >= xpc_hb_check_timeout ||
						(volatile int) xpc_exiting));
	}

	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.
 */
static int
xpc_initiate_discovery(void *ignore)
{
	daemonize(XPC_DISCOVERY_THREAD_NAME);

	xpc_discovery();

	dev_dbg(xpc_part, "discovery thread is exiting\n");

	/* mark this thread as having exited */
	complete(&xpc_discovery_exited);
	return 0;
}


/*
 * Establish first contact with the remote partititon. This involves pulling
 * the XPC per partition variables from the remote partition and waiting for
 * the remote partition to pull ours.
 */
static enum xpc_retval
xpc_make_first_contact(struct xpc_partition *part)
{
	enum xpc_retval ret;


	while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) {
		if (ret != xpcRetry) {
			XPC_DEACTIVATE_PARTITION(part, ret);
			return ret;
		}

		dev_dbg(xpc_chan, "waiting to make first contact with "
			"partition %d\n", XPC_PARTID(part));

		/* wait a 1/4 of a second or so */
		(void) msleep_interruptible(250);

		if (part->act_state == XPC_P_DEACTIVATING) {
			return part->reason;
		}
	}

	return xpc_mark_partition_active(part);
}


/*
 * The first kthread assigned to a newly activated partition is the one
 * created by XPC HB with which it calls xpc_partition_up(). 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.
 */
static void
xpc_channel_mgr(struct xpc_partition *part)
{
	while (part->act_state != XPC_P_DEACTIVATING ||
			atomic_read(&part->nchannels_active) > 0 ||
					!xpc_partition_disengaged(part)) {

		xpc_process_channel_activity(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 ||
				(volatile u64) part->local_IPI_amo != 0 ||
				((volatile u8) part->act_state ==
							XPC_P_DEACTIVATING &&
				atomic_read(&part->nchannels_active) == 0 &&
				xpc_partition_disengaged(part))));
		atomic_set(&part->channel_mgr_requests, 1);

		// >>> Does it need to wakeup periodically as well? In case we
		// >>> miscalculated the #of kthreads to wakeup or create?
	}
}


/*
 * 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) != xpcSuccess) {
		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) == xpcSuccess) {
		xpc_channel_mgr(part);
	}

	xpc_part_deref(part);

	xpc_teardown_infrastructure(part);
}


static int
xpc_activating(void *__partid)
{
	partid_t partid = (u64) __partid;
	struct xpc_partition *part = &xpc_partitions[partid];
	unsigned long irq_flags;
	struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
	int ret;