Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 * This file is subject to the terms and conditions of the GNU General Public
   3 * License.  See the file "COPYING" in the main directory of this archive
   4 * for more details.
   5 *
 
   6 * Copyright (c) 2004-2009 Silicon Graphics, Inc.  All Rights Reserved.
   7 */
   8
   9/*
  10 * Cross Partition Communication (XPC) support - standard version.
  11 *
  12 *	XPC provides a message passing capability that crosses partition
  13 *	boundaries. This module is made up of two parts:
  14 *
  15 *	    partition	This part detects the presence/absence of other
  16 *			partitions. It provides a heartbeat and monitors
  17 *			the heartbeats of other partitions.
  18 *
  19 *	    channel	This part manages the channels and sends/receives
  20 *			messages across them to/from other partitions.
  21 *
  22 *	There are a couple of additional functions residing in XP, which
  23 *	provide an interface to XPC for its users.
  24 *
  25 *
  26 *	Caveats:
  27 *
  28 *	  . Currently on sn2, we have no way to determine which nasid an IRQ
  29 *	    came from. Thus, xpc_send_IRQ_sn2() does a remote amo write
  30 *	    followed by an IPI. The amo indicates where data is to be pulled
  31 *	    from, so after the IPI arrives, the remote partition checks the amo
  32 *	    word. The IPI can actually arrive before the amo however, so other
  33 *	    code must periodically check for this case. Also, remote amo
  34 *	    operations do not reliably time out. Thus we do a remote PIO read
  35 *	    solely to know whether the remote partition is down and whether we
  36 *	    should stop sending IPIs to it. This remote PIO read operation is
  37 *	    set up in a special nofault region so SAL knows to ignore (and
  38 *	    cleanup) any errors due to the remote amo write, PIO read, and/or
  39 *	    PIO write operations.
  40 *
  41 *	    If/when new hardware solves this IPI problem, we should abandon
  42 *	    the current approach.
  43 *
  44 */
  45
  46#include <linux/module.h>
  47#include <linux/slab.h>
  48#include <linux/sysctl.h>
  49#include <linux/device.h>
  50#include <linux/delay.h>
  51#include <linux/reboot.h>
  52#include <linux/kdebug.h>
  53#include <linux/kthread.h>
  54#include "xpc.h"
  55
 
 
 
 
  56/* define two XPC debug device structures to be used with dev_dbg() et al */
  57
  58struct device_driver xpc_dbg_name = {
  59	.name = "xpc"
  60};
  61
  62struct device xpc_part_dbg_subname = {
  63	.init_name = "",	/* set to "part" at xpc_init() time */
  64	.driver = &xpc_dbg_name
  65};
  66
  67struct device xpc_chan_dbg_subname = {
  68	.init_name = "",	/* set to "chan" at xpc_init() time */
  69	.driver = &xpc_dbg_name
  70};
  71
  72struct device *xpc_part = &xpc_part_dbg_subname;
  73struct device *xpc_chan = &xpc_chan_dbg_subname;
  74
  75static int xpc_kdebug_ignore;
  76
  77/* systune related variables for /proc/sys directories */
  78
  79static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
  80static int xpc_hb_min_interval = 1;
  81static int xpc_hb_max_interval = 10;
  82
  83static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
  84static int xpc_hb_check_min_interval = 10;
  85static int xpc_hb_check_max_interval = 120;
  86
  87int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
  88static int xpc_disengage_min_timelimit;	/* = 0 */
  89static int xpc_disengage_max_timelimit = 120;
  90
  91static ctl_table xpc_sys_xpc_hb_dir[] = {
  92	{
  93	 .procname = "hb_interval",
  94	 .data = &xpc_hb_interval,
  95	 .maxlen = sizeof(int),
  96	 .mode = 0644,
  97	 .proc_handler = proc_dointvec_minmax,
  98	 .extra1 = &xpc_hb_min_interval,
  99	 .extra2 = &xpc_hb_max_interval},
 100	{
 101	 .procname = "hb_check_interval",
 102	 .data = &xpc_hb_check_interval,
 103	 .maxlen = sizeof(int),
 104	 .mode = 0644,
 105	 .proc_handler = proc_dointvec_minmax,
 106	 .extra1 = &xpc_hb_check_min_interval,
 107	 .extra2 = &xpc_hb_check_max_interval},
 108	{}
 109};
 110static ctl_table xpc_sys_xpc_dir[] = {
 111	{
 112	 .procname = "hb",
 113	 .mode = 0555,
 114	 .child = xpc_sys_xpc_hb_dir},
 115	{
 116	 .procname = "disengage_timelimit",
 117	 .data = &xpc_disengage_timelimit,
 118	 .maxlen = sizeof(int),
 119	 .mode = 0644,
 120	 .proc_handler = proc_dointvec_minmax,
 121	 .extra1 = &xpc_disengage_min_timelimit,
 122	 .extra2 = &xpc_disengage_max_timelimit},
 123	{}
 124};
 125static ctl_table xpc_sys_dir[] = {
 126	{
 127	 .procname = "xpc",
 128	 .mode = 0555,
 129	 .child = xpc_sys_xpc_dir},
 130	{}
 131};
 132static struct ctl_table_header *xpc_sysctl;
 133
 134/* non-zero if any remote partition disengage was timed out */
 135int xpc_disengage_timedout;
 136
 137/* #of activate IRQs received and not yet processed */
 138int xpc_activate_IRQ_rcvd;
 139DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
 140
 141/* IRQ handler notifies this wait queue on receipt of an IRQ */
 142DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
 143
 144static unsigned long xpc_hb_check_timeout;
 145static struct timer_list xpc_hb_timer;
 146
 147/* notification that the xpc_hb_checker thread has exited */
 148static DECLARE_COMPLETION(xpc_hb_checker_exited);
 149
 150/* notification that the xpc_discovery thread has exited */
 151static DECLARE_COMPLETION(xpc_discovery_exited);
 152
 153static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
 154
 155static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
 156static struct notifier_block xpc_reboot_notifier = {
 157	.notifier_call = xpc_system_reboot,
 158};
 159
 160static int xpc_system_die(struct notifier_block *, unsigned long, void *);
 161static struct notifier_block xpc_die_notifier = {
 162	.notifier_call = xpc_system_die,
 163};
 164
 165struct xpc_arch_operations xpc_arch_ops;
 166
 167/*
 168 * Timer function to enforce the timelimit on the partition disengage.
 169 */
 170static void
 171xpc_timeout_partition_disengage(unsigned long data)
 172{
 173	struct xpc_partition *part = (struct xpc_partition *)data;
 174
 175	DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
 176
 177	(void)xpc_partition_disengaged(part);
 178
 179	DBUG_ON(part->disengage_timeout != 0);
 180	DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
 181}
 182
 183/*
 184 * Timer to produce the heartbeat.  The timer structures function is
 185 * already set when this is initially called.  A tunable is used to
 186 * specify when the next timeout should occur.
 187 */
 188static void
 189xpc_hb_beater(unsigned long dummy)
 190{
 191	xpc_arch_ops.increment_heartbeat();
 192
 193	if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
 194		wake_up_interruptible(&xpc_activate_IRQ_wq);
 195
 196	xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
 197	add_timer(&xpc_hb_timer);
 198}
 199
 200static void
 201xpc_start_hb_beater(void)
 202{
 203	xpc_arch_ops.heartbeat_init();
 204	init_timer(&xpc_hb_timer);
 205	xpc_hb_timer.function = xpc_hb_beater;
 206	xpc_hb_beater(0);
 207}
 208
 209static void
 210xpc_stop_hb_beater(void)
 211{
 212	del_timer_sync(&xpc_hb_timer);
 213	xpc_arch_ops.heartbeat_exit();
 214}
 215
 216/*
 217 * At periodic intervals, scan through all active partitions and ensure
 218 * their heartbeat is still active.  If not, the partition is deactivated.
 219 */
 220static void
 221xpc_check_remote_hb(void)
 222{
 223	struct xpc_partition *part;
 224	short partid;
 225	enum xp_retval ret;
 226
 227	for (partid = 0; partid < xp_max_npartitions; partid++) {
 228
 229		if (xpc_exiting)
 230			break;
 231
 232		if (partid == xp_partition_id)
 233			continue;
 234
 235		part = &xpc_partitions[partid];
 236
 237		if (part->act_state == XPC_P_AS_INACTIVE ||
 238		    part->act_state == XPC_P_AS_DEACTIVATING) {
 239			continue;
 240		}
 241
 242		ret = xpc_arch_ops.get_remote_heartbeat(part);
 243		if (ret != xpSuccess)
 244			XPC_DEACTIVATE_PARTITION(part, ret);
 245	}
 246}
 247
 248/*
 249 * This thread is responsible for nearly all of the partition
 250 * activation/deactivation.
 251 */
 252static int
 253xpc_hb_checker(void *ignore)
 254{
 255	int force_IRQ = 0;
 256
 257	/* this thread was marked active by xpc_hb_init() */
 258
 259	set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
 260
 261	/* set our heartbeating to other partitions into motion */
 262	xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
 263	xpc_start_hb_beater();
 264
 265	while (!xpc_exiting) {
 266
 267		dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
 268			"been received\n",
 269			(int)(xpc_hb_check_timeout - jiffies),
 270			xpc_activate_IRQ_rcvd);
 271
 272		/* checking of remote heartbeats is skewed by IRQ handling */
 273		if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
 274			xpc_hb_check_timeout = jiffies +
 275			    (xpc_hb_check_interval * HZ);
 276
 277			dev_dbg(xpc_part, "checking remote heartbeats\n");
 278			xpc_check_remote_hb();
 279
 280			/*
 281			 * On sn2 we need to periodically recheck to ensure no
 282			 * IRQ/amo pairs have been missed.
 283			 */
 284			if (is_shub())
 285				force_IRQ = 1;
 286		}
 287
 288		/* check for outstanding IRQs */
 289		if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
 290			force_IRQ = 0;
 291			dev_dbg(xpc_part, "processing activate IRQs "
 292				"received\n");
 293			xpc_arch_ops.process_activate_IRQ_rcvd();
 294		}
 295
 296		/* wait for IRQ or timeout */
 297		(void)wait_event_interruptible(xpc_activate_IRQ_wq,
 298					       (time_is_before_eq_jiffies(
 299						xpc_hb_check_timeout) ||
 300						xpc_activate_IRQ_rcvd > 0 ||
 301						xpc_exiting));
 302	}
 303
 304	xpc_stop_hb_beater();
 305
 306	dev_dbg(xpc_part, "heartbeat checker is exiting\n");
 307
 308	/* mark this thread as having exited */
 309	complete(&xpc_hb_checker_exited);
 310	return 0;
 311}
 312
 313/*
 314 * This thread will attempt to discover other partitions to activate
 315 * based on info provided by SAL. This new thread is short lived and
 316 * will exit once discovery is complete.
 317 */
 318static int
 319xpc_initiate_discovery(void *ignore)
 320{
 321	xpc_discovery();
 322
 323	dev_dbg(xpc_part, "discovery thread is exiting\n");
 324
 325	/* mark this thread as having exited */
 326	complete(&xpc_discovery_exited);
 327	return 0;
 328}
 329
 330/*
 331 * The first kthread assigned to a newly activated partition is the one
 332 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
 333 * that kthread until the partition is brought down, at which time that kthread
 334 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
 335 * that XPC has dismantled all communication infrastructure for the associated
 336 * partition.) This kthread becomes the channel manager for that partition.
 337 *
 338 * Each active partition has a channel manager, who, besides connecting and
 339 * disconnecting channels, will ensure that each of the partition's connected
 340 * channels has the required number of assigned kthreads to get the work done.
 341 */
 342static void
 343xpc_channel_mgr(struct xpc_partition *part)
 344{
 345	while (part->act_state != XPC_P_AS_DEACTIVATING ||
 346	       atomic_read(&part->nchannels_active) > 0 ||
 347	       !xpc_partition_disengaged(part)) {
 348
 349		xpc_process_sent_chctl_flags(part);
 350
 351		/*
 352		 * Wait until we've been requested to activate kthreads or
 353		 * all of the channel's message queues have been torn down or
 354		 * a signal is pending.
 355		 *
 356		 * The channel_mgr_requests is set to 1 after being awakened,
 357		 * This is done to prevent the channel mgr from making one pass
 358		 * through the loop for each request, since he will
 359		 * be servicing all the requests in one pass. The reason it's
 360		 * set to 1 instead of 0 is so that other kthreads will know
 361		 * that the channel mgr is running and won't bother trying to
 362		 * wake him up.
 363		 */
 364		atomic_dec(&part->channel_mgr_requests);
 365		(void)wait_event_interruptible(part->channel_mgr_wq,
 366				(atomic_read(&part->channel_mgr_requests) > 0 ||
 367				 part->chctl.all_flags != 0 ||
 368				 (part->act_state == XPC_P_AS_DEACTIVATING &&
 369				 atomic_read(&part->nchannels_active) == 0 &&
 370				 xpc_partition_disengaged(part))));
 371		atomic_set(&part->channel_mgr_requests, 1);
 372	}
 373}
 374
 375/*
 376 * Guarantee that the kzalloc'd memory is cacheline aligned.
 377 */
 378void *
 379xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
 380{
 381	/* see if kzalloc will give us cachline aligned memory by default */
 382	*base = kzalloc(size, flags);
 383	if (*base == NULL)
 384		return NULL;
 385
 386	if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
 387		return *base;
 388
 389	kfree(*base);
 390
 391	/* nope, we'll have to do it ourselves */
 392	*base = kzalloc(size + L1_CACHE_BYTES, flags);
 393	if (*base == NULL)
 394		return NULL;
 395
 396	return (void *)L1_CACHE_ALIGN((u64)*base);
 397}
 398
 399/*
 400 * Setup the channel structures necessary to support XPartition Communication
 401 * between the specified remote partition and the local one.
 402 */
 403static enum xp_retval
 404xpc_setup_ch_structures(struct xpc_partition *part)
 405{
 406	enum xp_retval ret;
 407	int ch_number;
 408	struct xpc_channel *ch;
 409	short partid = XPC_PARTID(part);
 410
 411	/*
 412	 * Allocate all of the channel structures as a contiguous chunk of
 413	 * memory.
 414	 */
 415	DBUG_ON(part->channels != NULL);
 416	part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_MAX_NCHANNELS,
 
 417				 GFP_KERNEL);
 418	if (part->channels == NULL) {
 419		dev_err(xpc_chan, "can't get memory for channels\n");
 420		return xpNoMemory;
 421	}
 422
 423	/* allocate the remote open and close args */
 424
 425	part->remote_openclose_args =
 426	    xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
 427					  GFP_KERNEL, &part->
 428					  remote_openclose_args_base);
 429	if (part->remote_openclose_args == NULL) {
 430		dev_err(xpc_chan, "can't get memory for remote connect args\n");
 431		ret = xpNoMemory;
 432		goto out_1;
 433	}
 434
 435	part->chctl.all_flags = 0;
 436	spin_lock_init(&part->chctl_lock);
 437
 438	atomic_set(&part->channel_mgr_requests, 1);
 439	init_waitqueue_head(&part->channel_mgr_wq);
 440
 441	part->nchannels = XPC_MAX_NCHANNELS;
 442
 443	atomic_set(&part->nchannels_active, 0);
 444	atomic_set(&part->nchannels_engaged, 0);
 445
 446	for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
 447		ch = &part->channels[ch_number];
 448
 449		ch->partid = partid;
 450		ch->number = ch_number;
 451		ch->flags = XPC_C_DISCONNECTED;
 452
 453		atomic_set(&ch->kthreads_assigned, 0);
 454		atomic_set(&ch->kthreads_idle, 0);
 455		atomic_set(&ch->kthreads_active, 0);
 456
 457		atomic_set(&ch->references, 0);
 458		atomic_set(&ch->n_to_notify, 0);
 459
 460		spin_lock_init(&ch->lock);
 461		init_completion(&ch->wdisconnect_wait);
 462
 463		atomic_set(&ch->n_on_msg_allocate_wq, 0);
 464		init_waitqueue_head(&ch->msg_allocate_wq);
 465		init_waitqueue_head(&ch->idle_wq);
 466	}
 467
 468	ret = xpc_arch_ops.setup_ch_structures(part);
 469	if (ret != xpSuccess)
 470		goto out_2;
 471
 472	/*
 473	 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
 474	 * we're declaring that this partition is ready to go.
 475	 */
 476	part->setup_state = XPC_P_SS_SETUP;
 477
 478	return xpSuccess;
 479
 480	/* setup of ch structures failed */
 481out_2:
 482	kfree(part->remote_openclose_args_base);
 483	part->remote_openclose_args = NULL;
 484out_1:
 485	kfree(part->channels);
 486	part->channels = NULL;
 487	return ret;
 488}
 489
 490/*
 491 * Teardown the channel structures necessary to support XPartition Communication
 492 * between the specified remote partition and the local one.
 493 */
 494static void
 495xpc_teardown_ch_structures(struct xpc_partition *part)
 496{
 497	DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
 498	DBUG_ON(atomic_read(&part->nchannels_active) != 0);
 499
 500	/*
 501	 * Make this partition inaccessible to local processes by marking it
 502	 * as no longer setup. Then wait before proceeding with the teardown
 503	 * until all existing references cease.
 504	 */
 505	DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
 506	part->setup_state = XPC_P_SS_WTEARDOWN;
 507
 508	wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
 509
 510	/* now we can begin tearing down the infrastructure */
 511
 512	xpc_arch_ops.teardown_ch_structures(part);
 513
 514	kfree(part->remote_openclose_args_base);
 515	part->remote_openclose_args = NULL;
 516	kfree(part->channels);
 517	part->channels = NULL;
 518
 519	part->setup_state = XPC_P_SS_TORNDOWN;
 520}
 521
 522/*
 523 * When XPC HB determines that a partition has come up, it will create a new
 524 * kthread and that kthread will call this function to attempt to set up the
 525 * basic infrastructure used for Cross Partition Communication with the newly
 526 * upped partition.
 527 *
 528 * The kthread that was created by XPC HB and which setup the XPC
 529 * infrastructure will remain assigned to the partition becoming the channel
 530 * manager for that partition until the partition is deactivating, at which
 531 * time the kthread will teardown the XPC infrastructure and then exit.
 532 */
 533static int
 534xpc_activating(void *__partid)
 535{
 536	short partid = (u64)__partid;
 537	struct xpc_partition *part = &xpc_partitions[partid];
 538	unsigned long irq_flags;
 539
 540	DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
 541
 542	spin_lock_irqsave(&part->act_lock, irq_flags);
 543
 544	if (part->act_state == XPC_P_AS_DEACTIVATING) {
 545		part->act_state = XPC_P_AS_INACTIVE;
 546		spin_unlock_irqrestore(&part->act_lock, irq_flags);
 547		part->remote_rp_pa = 0;
 548		return 0;
 549	}
 550
 551	/* indicate the thread is activating */
 552	DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
 553	part->act_state = XPC_P_AS_ACTIVATING;
 554
 555	XPC_SET_REASON(part, 0, 0);
 556	spin_unlock_irqrestore(&part->act_lock, irq_flags);
 557
 558	dev_dbg(xpc_part, "activating partition %d\n", partid);
 559
 560	xpc_arch_ops.allow_hb(partid);
 561
 562	if (xpc_setup_ch_structures(part) == xpSuccess) {
 563		(void)xpc_part_ref(part);	/* this will always succeed */
 564
 565		if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
 566			xpc_mark_partition_active(part);
 567			xpc_channel_mgr(part);
 568			/* won't return until partition is deactivating */
 569		}
 570
 571		xpc_part_deref(part);
 572		xpc_teardown_ch_structures(part);
 573	}
 574
 575	xpc_arch_ops.disallow_hb(partid);
 576	xpc_mark_partition_inactive(part);
 577
 578	if (part->reason == xpReactivating) {
 579		/* interrupting ourselves results in activating partition */
 580		xpc_arch_ops.request_partition_reactivation(part);
 581	}
 582
 583	return 0;
 584}
 585
 586void
 587xpc_activate_partition(struct xpc_partition *part)
 588{
 589	short partid = XPC_PARTID(part);
 590	unsigned long irq_flags;
 591	struct task_struct *kthread;
 592
 593	spin_lock_irqsave(&part->act_lock, irq_flags);
 594
 595	DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
 596
 597	part->act_state = XPC_P_AS_ACTIVATION_REQ;
 598	XPC_SET_REASON(part, xpCloneKThread, __LINE__);
 599
 600	spin_unlock_irqrestore(&part->act_lock, irq_flags);
 601
 602	kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
 603			      partid);
 604	if (IS_ERR(kthread)) {
 605		spin_lock_irqsave(&part->act_lock, irq_flags);
 606		part->act_state = XPC_P_AS_INACTIVE;
 607		XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
 608		spin_unlock_irqrestore(&part->act_lock, irq_flags);
 609	}
 610}
 611
 612void
 613xpc_activate_kthreads(struct xpc_channel *ch, int needed)
 614{
 615	int idle = atomic_read(&ch->kthreads_idle);
 616	int assigned = atomic_read(&ch->kthreads_assigned);
 617	int wakeup;
 618
 619	DBUG_ON(needed <= 0);
 620
 621	if (idle > 0) {
 622		wakeup = (needed > idle) ? idle : needed;
 623		needed -= wakeup;
 624
 625		dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
 626			"channel=%d\n", wakeup, ch->partid, ch->number);
 627
 628		/* only wakeup the requested number of kthreads */
 629		wake_up_nr(&ch->idle_wq, wakeup);
 630	}
 631
 632	if (needed <= 0)
 633		return;
 634
 635	if (needed + assigned > ch->kthreads_assigned_limit) {
 636		needed = ch->kthreads_assigned_limit - assigned;
 637		if (needed <= 0)
 638			return;
 639	}
 640
 641	dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
 642		needed, ch->partid, ch->number);
 643
 644	xpc_create_kthreads(ch, needed, 0);
 645}
 646
 647/*
 648 * This function is where XPC's kthreads wait for messages to deliver.
 649 */
 650static void
 651xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
 652{
 653	int (*n_of_deliverable_payloads) (struct xpc_channel *) =
 654		xpc_arch_ops.n_of_deliverable_payloads;
 655
 656	do {
 657		/* deliver messages to their intended recipients */
 658
 659		while (n_of_deliverable_payloads(ch) > 0 &&
 660		       !(ch->flags & XPC_C_DISCONNECTING)) {
 661			xpc_deliver_payload(ch);
 662		}
 663
 664		if (atomic_inc_return(&ch->kthreads_idle) >
 665		    ch->kthreads_idle_limit) {
 666			/* too many idle kthreads on this channel */
 667			atomic_dec(&ch->kthreads_idle);
 668			break;
 669		}
 670
 671		dev_dbg(xpc_chan, "idle kthread calling "
 672			"wait_event_interruptible_exclusive()\n");
 673
 674		(void)wait_event_interruptible_exclusive(ch->idle_wq,
 675				(n_of_deliverable_payloads(ch) > 0 ||
 676				 (ch->flags & XPC_C_DISCONNECTING)));
 677
 678		atomic_dec(&ch->kthreads_idle);
 679
 680	} while (!(ch->flags & XPC_C_DISCONNECTING));
 681}
 682
 683static int
 684xpc_kthread_start(void *args)
 685{
 686	short partid = XPC_UNPACK_ARG1(args);
 687	u16 ch_number = XPC_UNPACK_ARG2(args);
 688	struct xpc_partition *part = &xpc_partitions[partid];
 689	struct xpc_channel *ch;
 690	int n_needed;
 691	unsigned long irq_flags;
 692	int (*n_of_deliverable_payloads) (struct xpc_channel *) =
 693		xpc_arch_ops.n_of_deliverable_payloads;
 694
 695	dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
 696		partid, ch_number);
 697
 698	ch = &part->channels[ch_number];
 699
 700	if (!(ch->flags & XPC_C_DISCONNECTING)) {
 701
 702		/* let registerer know that connection has been established */
 703
 704		spin_lock_irqsave(&ch->lock, irq_flags);
 705		if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
 706			ch->flags |= XPC_C_CONNECTEDCALLOUT;
 707			spin_unlock_irqrestore(&ch->lock, irq_flags);
 708
 709			xpc_connected_callout(ch);
 710
 711			spin_lock_irqsave(&ch->lock, irq_flags);
 712			ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
 713			spin_unlock_irqrestore(&ch->lock, irq_flags);
 714
 715			/*
 716			 * It is possible that while the callout was being
 717			 * made that the remote partition sent some messages.
 718			 * If that is the case, we may need to activate
 719			 * additional kthreads to help deliver them. We only
 720			 * need one less than total #of messages to deliver.
 721			 */
 722			n_needed = n_of_deliverable_payloads(ch) - 1;
 723			if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
 724				xpc_activate_kthreads(ch, n_needed);
 725
 726		} else {
 727			spin_unlock_irqrestore(&ch->lock, irq_flags);
 728		}
 729
 730		xpc_kthread_waitmsgs(part, ch);
 731	}
 732
 733	/* let registerer know that connection is disconnecting */
 734
 735	spin_lock_irqsave(&ch->lock, irq_flags);
 736	if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
 737	    !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
 738		ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
 739		spin_unlock_irqrestore(&ch->lock, irq_flags);
 740
 741		xpc_disconnect_callout(ch, xpDisconnecting);
 742
 743		spin_lock_irqsave(&ch->lock, irq_flags);
 744		ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
 745	}
 746	spin_unlock_irqrestore(&ch->lock, irq_flags);
 747
 748	if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 749	    atomic_dec_return(&part->nchannels_engaged) == 0) {
 750		xpc_arch_ops.indicate_partition_disengaged(part);
 751	}
 752
 753	xpc_msgqueue_deref(ch);
 754
 755	dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
 756		partid, ch_number);
 757
 758	xpc_part_deref(part);
 759	return 0;
 760}
 761
 762/*
 763 * For each partition that XPC has established communications with, there is
 764 * a minimum of one kernel thread assigned to perform any operation that
 765 * may potentially sleep or block (basically the callouts to the asynchronous
 766 * functions registered via xpc_connect()).
 767 *
 768 * Additional kthreads are created and destroyed by XPC as the workload
 769 * demands.
 770 *
 771 * A kthread is assigned to one of the active channels that exists for a given
 772 * partition.
 773 */
 774void
 775xpc_create_kthreads(struct xpc_channel *ch, int needed,
 776		    int ignore_disconnecting)
 777{
 778	unsigned long irq_flags;
 779	u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
 780	struct xpc_partition *part = &xpc_partitions[ch->partid];
 781	struct task_struct *kthread;
 782	void (*indicate_partition_disengaged) (struct xpc_partition *) =
 783		xpc_arch_ops.indicate_partition_disengaged;
 784
 785	while (needed-- > 0) {
 786
 787		/*
 788		 * The following is done on behalf of the newly created
 789		 * kthread. That kthread is responsible for doing the
 790		 * counterpart to the following before it exits.
 791		 */
 792		if (ignore_disconnecting) {
 793			if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
 794				/* kthreads assigned had gone to zero */
 795				BUG_ON(!(ch->flags &
 796					 XPC_C_DISCONNECTINGCALLOUT_MADE));
 797				break;
 798			}
 799
 800		} else if (ch->flags & XPC_C_DISCONNECTING) {
 801			break;
 802
 803		} else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
 804			   atomic_inc_return(&part->nchannels_engaged) == 1) {
 805			xpc_arch_ops.indicate_partition_engaged(part);
 806		}
 807		(void)xpc_part_ref(part);
 808		xpc_msgqueue_ref(ch);
 809
 810		kthread = kthread_run(xpc_kthread_start, (void *)args,
 811				      "xpc%02dc%d", ch->partid, ch->number);
 812		if (IS_ERR(kthread)) {
 813			/* the fork failed */
 814
 815			/*
 816			 * NOTE: if (ignore_disconnecting &&
 817			 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
 818			 * then we'll deadlock if all other kthreads assigned
 819			 * to this channel are blocked in the channel's
 820			 * registerer, because the only thing that will unblock
 821			 * them is the xpDisconnecting callout that this
 822			 * failed kthread_run() would have made.
 823			 */
 824
 825			if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 826			    atomic_dec_return(&part->nchannels_engaged) == 0) {
 827				indicate_partition_disengaged(part);
 828			}
 829			xpc_msgqueue_deref(ch);
 830			xpc_part_deref(part);
 831
 832			if (atomic_read(&ch->kthreads_assigned) <
 833			    ch->kthreads_idle_limit) {
 834				/*
 835				 * Flag this as an error only if we have an
 836				 * insufficient #of kthreads for the channel
 837				 * to function.
 838				 */
 839				spin_lock_irqsave(&ch->lock, irq_flags);
 840				XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
 841						       &irq_flags);
 842				spin_unlock_irqrestore(&ch->lock, irq_flags);
 843			}
 844			break;
 845		}
 846	}
 847}
 848
 849void
 850xpc_disconnect_wait(int ch_number)
 851{
 852	unsigned long irq_flags;
 853	short partid;
 854	struct xpc_partition *part;
 855	struct xpc_channel *ch;
 856	int wakeup_channel_mgr;
 857
 858	/* now wait for all callouts to the caller's function to cease */
 859	for (partid = 0; partid < xp_max_npartitions; partid++) {
 860		part = &xpc_partitions[partid];
 861
 862		if (!xpc_part_ref(part))
 863			continue;
 864
 865		ch = &part->channels[ch_number];
 866
 867		if (!(ch->flags & XPC_C_WDISCONNECT)) {
 868			xpc_part_deref(part);
 869			continue;
 870		}
 871
 872		wait_for_completion(&ch->wdisconnect_wait);
 873
 874		spin_lock_irqsave(&ch->lock, irq_flags);
 875		DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
 876		wakeup_channel_mgr = 0;
 877
 878		if (ch->delayed_chctl_flags) {
 879			if (part->act_state != XPC_P_AS_DEACTIVATING) {
 880				spin_lock(&part->chctl_lock);
 881				part->chctl.flags[ch->number] |=
 882				    ch->delayed_chctl_flags;
 883				spin_unlock(&part->chctl_lock);
 884				wakeup_channel_mgr = 1;
 885			}
 886			ch->delayed_chctl_flags = 0;
 887		}
 888
 889		ch->flags &= ~XPC_C_WDISCONNECT;
 890		spin_unlock_irqrestore(&ch->lock, irq_flags);
 891
 892		if (wakeup_channel_mgr)
 893			xpc_wakeup_channel_mgr(part);
 894
 895		xpc_part_deref(part);
 896	}
 897}
 898
 899static int
 900xpc_setup_partitions(void)
 901{
 902	short partid;
 903	struct xpc_partition *part;
 904
 905	xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
 906				 xp_max_npartitions, GFP_KERNEL);
 
 907	if (xpc_partitions == NULL) {
 908		dev_err(xpc_part, "can't get memory for partition structure\n");
 909		return -ENOMEM;
 910	}
 911
 912	/*
 913	 * The first few fields of each entry of xpc_partitions[] need to
 914	 * be initialized now so that calls to xpc_connect() and
 915	 * xpc_disconnect() can be made prior to the activation of any remote
 916	 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
 917	 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
 918	 * PARTITION HAS BEEN ACTIVATED.
 919	 */
 920	for (partid = 0; partid < xp_max_npartitions; partid++) {
 921		part = &xpc_partitions[partid];
 922
 923		DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
 924
 925		part->activate_IRQ_rcvd = 0;
 926		spin_lock_init(&part->act_lock);
 927		part->act_state = XPC_P_AS_INACTIVE;
 928		XPC_SET_REASON(part, 0, 0);
 929
 930		init_timer(&part->disengage_timer);
 931		part->disengage_timer.function =
 932		    xpc_timeout_partition_disengage;
 933		part->disengage_timer.data = (unsigned long)part;
 934
 935		part->setup_state = XPC_P_SS_UNSET;
 936		init_waitqueue_head(&part->teardown_wq);
 937		atomic_set(&part->references, 0);
 938	}
 939
 940	return xpc_arch_ops.setup_partitions();
 941}
 942
 943static void
 944xpc_teardown_partitions(void)
 945{
 946	xpc_arch_ops.teardown_partitions();
 947	kfree(xpc_partitions);
 948}
 949
 950static void
 951xpc_do_exit(enum xp_retval reason)
 952{
 953	short partid;
 954	int active_part_count, printed_waiting_msg = 0;
 955	struct xpc_partition *part;
 956	unsigned long printmsg_time, disengage_timeout = 0;
 957
 958	/* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
 959	DBUG_ON(xpc_exiting == 1);
 960
 961	/*
 962	 * Let the heartbeat checker thread and the discovery thread
 963	 * (if one is running) know that they should exit. Also wake up
 964	 * the heartbeat checker thread in case it's sleeping.
 965	 */
 966	xpc_exiting = 1;
 967	wake_up_interruptible(&xpc_activate_IRQ_wq);
 968
 969	/* wait for the discovery thread to exit */
 970	wait_for_completion(&xpc_discovery_exited);
 971
 972	/* wait for the heartbeat checker thread to exit */
 973	wait_for_completion(&xpc_hb_checker_exited);
 974
 975	/* sleep for a 1/3 of a second or so */
 976	(void)msleep_interruptible(300);
 977
 978	/* wait for all partitions to become inactive */
 979
 980	printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
 981	xpc_disengage_timedout = 0;
 982
 983	do {
 984		active_part_count = 0;
 985
 986		for (partid = 0; partid < xp_max_npartitions; partid++) {
 987			part = &xpc_partitions[partid];
 988
 989			if (xpc_partition_disengaged(part) &&
 990			    part->act_state == XPC_P_AS_INACTIVE) {
 991				continue;
 992			}
 993
 994			active_part_count++;
 995
 996			XPC_DEACTIVATE_PARTITION(part, reason);
 997
 998			if (part->disengage_timeout > disengage_timeout)
 999				disengage_timeout = part->disengage_timeout;
1000		}
1001
1002		if (xpc_arch_ops.any_partition_engaged()) {
1003			if (time_is_before_jiffies(printmsg_time)) {
1004				dev_info(xpc_part, "waiting for remote "
1005					 "partitions to deactivate, timeout in "
1006					 "%ld seconds\n", (disengage_timeout -
1007					 jiffies) / HZ);
1008				printmsg_time = jiffies +
1009				    (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
1010				printed_waiting_msg = 1;
1011			}
1012
1013		} else if (active_part_count > 0) {
1014			if (printed_waiting_msg) {
1015				dev_info(xpc_part, "waiting for local partition"
1016					 " to deactivate\n");
1017				printed_waiting_msg = 0;
1018			}
1019
1020		} else {
1021			if (!xpc_disengage_timedout) {
1022				dev_info(xpc_part, "all partitions have "
1023					 "deactivated\n");
1024			}
1025			break;
1026		}
1027
1028		/* sleep for a 1/3 of a second or so */
1029		(void)msleep_interruptible(300);
1030
1031	} while (1);
1032
1033	DBUG_ON(xpc_arch_ops.any_partition_engaged());
1034
1035	xpc_teardown_rsvd_page();
1036
1037	if (reason == xpUnloading) {
1038		(void)unregister_die_notifier(&xpc_die_notifier);
1039		(void)unregister_reboot_notifier(&xpc_reboot_notifier);
1040	}
1041
1042	/* clear the interface to XPC's functions */
1043	xpc_clear_interface();
1044
1045	if (xpc_sysctl)
1046		unregister_sysctl_table(xpc_sysctl);
1047
1048	xpc_teardown_partitions();
1049
1050	if (is_shub())
1051		xpc_exit_sn2();
1052	else if (is_uv())
1053		xpc_exit_uv();
1054}
1055
1056/*
1057 * This function is called when the system is being rebooted.
1058 */
1059static int
1060xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1061{
1062	enum xp_retval reason;
1063
1064	switch (event) {
1065	case SYS_RESTART:
1066		reason = xpSystemReboot;
1067		break;
1068	case SYS_HALT:
1069		reason = xpSystemHalt;
1070		break;
1071	case SYS_POWER_OFF:
1072		reason = xpSystemPoweroff;
1073		break;
1074	default:
1075		reason = xpSystemGoingDown;
1076	}
1077
1078	xpc_do_exit(reason);
1079	return NOTIFY_DONE;
1080}
1081
 
 
 
1082/*
1083 * Notify other partitions to deactivate from us by first disengaging from all
1084 * references to our memory.
1085 */
1086static void
1087xpc_die_deactivate(void)
1088{
1089	struct xpc_partition *part;
1090	short partid;
1091	int any_engaged;
1092	long keep_waiting;
1093	long wait_to_print;
1094
 
 
 
1095	/* keep xpc_hb_checker thread from doing anything (just in case) */
1096	xpc_exiting = 1;
1097
1098	xpc_arch_ops.disallow_all_hbs();   /*indicate we're deactivated */
1099
1100	for (partid = 0; partid < xp_max_npartitions; partid++) {
1101		part = &xpc_partitions[partid];
1102
1103		if (xpc_arch_ops.partition_engaged(partid) ||
1104		    part->act_state != XPC_P_AS_INACTIVE) {
1105			xpc_arch_ops.request_partition_deactivation(part);
1106			xpc_arch_ops.indicate_partition_disengaged(part);
1107		}
1108	}
1109
1110	/*
1111	 * Though we requested that all other partitions deactivate from us,
1112	 * we only wait until they've all disengaged or we've reached the
1113	 * defined timelimit.
1114	 *
1115	 * Given that one iteration through the following while-loop takes
1116	 * approximately 200 microseconds, calculate the #of loops to take
1117	 * before bailing and the #of loops before printing a waiting message.
1118	 */
1119	keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1120	wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1121
1122	while (1) {
1123		any_engaged = xpc_arch_ops.any_partition_engaged();
1124		if (!any_engaged) {
1125			dev_info(xpc_part, "all partitions have deactivated\n");
1126			break;
1127		}
1128
1129		if (!keep_waiting--) {
1130			for (partid = 0; partid < xp_max_npartitions;
1131			     partid++) {
1132				if (xpc_arch_ops.partition_engaged(partid)) {
1133					dev_info(xpc_part, "deactivate from "
1134						 "remote partition %d timed "
1135						 "out\n", partid);
1136				}
1137			}
1138			break;
1139		}
1140
1141		if (!wait_to_print--) {
1142			dev_info(xpc_part, "waiting for remote partitions to "
1143				 "deactivate, timeout in %ld seconds\n",
1144				 keep_waiting / (1000 * 5));
1145			wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1146			    1000 * 5;
1147		}
1148
1149		udelay(200);
1150	}
1151}
1152
1153/*
1154 * This function is called when the system is being restarted or halted due
1155 * to some sort of system failure. If this is the case we need to notify the
1156 * other partitions to disengage from all references to our memory.
1157 * This function can also be called when our heartbeater could be offlined
1158 * for a time. In this case we need to notify other partitions to not worry
1159 * about the lack of a heartbeat.
1160 */
1161static int
1162xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1163{
1164#ifdef CONFIG_IA64		/* !!! temporary kludge */
1165	switch (event) {
1166	case DIE_MACHINE_RESTART:
1167	case DIE_MACHINE_HALT:
1168		xpc_die_deactivate();
1169		break;
1170
1171	case DIE_KDEBUG_ENTER:
1172		/* Should lack of heartbeat be ignored by other partitions? */
1173		if (!xpc_kdebug_ignore)
1174			break;
1175
1176		/* fall through */
1177	case DIE_MCA_MONARCH_ENTER:
1178	case DIE_INIT_MONARCH_ENTER:
1179		xpc_arch_ops.offline_heartbeat();
1180		break;
1181
1182	case DIE_KDEBUG_LEAVE:
1183		/* Is lack of heartbeat being ignored by other partitions? */
1184		if (!xpc_kdebug_ignore)
1185			break;
1186
1187		/* fall through */
1188	case DIE_MCA_MONARCH_LEAVE:
1189	case DIE_INIT_MONARCH_LEAVE:
1190		xpc_arch_ops.online_heartbeat();
1191		break;
1192	}
1193#else
1194	xpc_die_deactivate();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1195#endif
1196
1197	return NOTIFY_DONE;
1198}
1199
1200int __init
1201xpc_init(void)
1202{
1203	int ret;
1204	struct task_struct *kthread;
1205
1206	dev_set_name(xpc_part, "part");
1207	dev_set_name(xpc_chan, "chan");
1208
1209	if (is_shub()) {
1210		/*
1211		 * The ia64-sn2 architecture supports at most 64 partitions.
1212		 * And the inability to unregister remote amos restricts us
1213		 * further to only support exactly 64 partitions on this
1214		 * architecture, no less.
1215		 */
1216		if (xp_max_npartitions != 64) {
1217			dev_err(xpc_part, "max #of partitions not set to 64\n");
1218			ret = -EINVAL;
1219		} else {
1220			ret = xpc_init_sn2();
1221		}
1222
1223	} else if (is_uv()) {
1224		ret = xpc_init_uv();
1225
1226	} else {
1227		ret = -ENODEV;
1228	}
1229
1230	if (ret != 0)
1231		return ret;
1232
1233	ret = xpc_setup_partitions();
1234	if (ret != 0) {
1235		dev_err(xpc_part, "can't get memory for partition structure\n");
1236		goto out_1;
1237	}
1238
1239	xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1240
1241	/*
1242	 * Fill the partition reserved page with the information needed by
1243	 * other partitions to discover we are alive and establish initial
1244	 * communications.
1245	 */
1246	ret = xpc_setup_rsvd_page();
1247	if (ret != 0) {
1248		dev_err(xpc_part, "can't setup our reserved page\n");
1249		goto out_2;
1250	}
1251
1252	/* add ourselves to the reboot_notifier_list */
1253	ret = register_reboot_notifier(&xpc_reboot_notifier);
1254	if (ret != 0)
1255		dev_warn(xpc_part, "can't register reboot notifier\n");
1256
1257	/* add ourselves to the die_notifier list */
1258	ret = register_die_notifier(&xpc_die_notifier);
1259	if (ret != 0)
1260		dev_warn(xpc_part, "can't register die notifier\n");
1261
1262	/*
1263	 * The real work-horse behind xpc.  This processes incoming
1264	 * interrupts and monitors remote heartbeats.
1265	 */
1266	kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1267	if (IS_ERR(kthread)) {
1268		dev_err(xpc_part, "failed while forking hb check thread\n");
1269		ret = -EBUSY;
1270		goto out_3;
1271	}
1272
1273	/*
1274	 * Startup a thread that will attempt to discover other partitions to
1275	 * activate based on info provided by SAL. This new thread is short
1276	 * lived and will exit once discovery is complete.
1277	 */
1278	kthread = kthread_run(xpc_initiate_discovery, NULL,
1279			      XPC_DISCOVERY_THREAD_NAME);
1280	if (IS_ERR(kthread)) {
1281		dev_err(xpc_part, "failed while forking discovery thread\n");
1282
1283		/* mark this new thread as a non-starter */
1284		complete(&xpc_discovery_exited);
1285
1286		xpc_do_exit(xpUnloading);
1287		return -EBUSY;
1288	}
1289
1290	/* set the interface to point at XPC's functions */
1291	xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1292			  xpc_initiate_send, xpc_initiate_send_notify,
1293			  xpc_initiate_received, xpc_initiate_partid_to_nasids);
1294
1295	return 0;
1296
1297	/* initialization was not successful */
1298out_3:
1299	xpc_teardown_rsvd_page();
1300
1301	(void)unregister_die_notifier(&xpc_die_notifier);
1302	(void)unregister_reboot_notifier(&xpc_reboot_notifier);
1303out_2:
1304	if (xpc_sysctl)
1305		unregister_sysctl_table(xpc_sysctl);
1306
1307	xpc_teardown_partitions();
1308out_1:
1309	if (is_shub())
1310		xpc_exit_sn2();
1311	else if (is_uv())
1312		xpc_exit_uv();
1313	return ret;
1314}
1315
1316module_init(xpc_init);
1317
1318void __exit
1319xpc_exit(void)
1320{
1321	xpc_do_exit(xpUnloading);
1322}
1323
1324module_exit(xpc_exit);
1325
1326MODULE_AUTHOR("Silicon Graphics, Inc.");
1327MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1328MODULE_LICENSE("GPL");
1329
1330module_param(xpc_hb_interval, int, 0);
1331MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1332		 "heartbeat increments.");
1333
1334module_param(xpc_hb_check_interval, int, 0);
1335MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1336		 "heartbeat checks.");
1337
1338module_param(xpc_disengage_timelimit, int, 0);
1339MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1340		 "for disengage to complete.");
1341
1342module_param(xpc_kdebug_ignore, int, 0);
1343MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1344		 "other partitions when dropping into kdebug.");
v6.2
   1/*
   2 * This file is subject to the terms and conditions of the GNU General Public
   3 * License.  See the file "COPYING" in the main directory of this archive
   4 * for more details.
   5 *
   6 * (C) Copyright 2020 Hewlett Packard Enterprise Development LP
   7 * Copyright (c) 2004-2009 Silicon Graphics, Inc.  All Rights Reserved.
   8 */
   9
  10/*
  11 * Cross Partition Communication (XPC) support - standard version.
  12 *
  13 *	XPC provides a message passing capability that crosses partition
  14 *	boundaries. This module is made up of two parts:
  15 *
  16 *	    partition	This part detects the presence/absence of other
  17 *			partitions. It provides a heartbeat and monitors
  18 *			the heartbeats of other partitions.
  19 *
  20 *	    channel	This part manages the channels and sends/receives
  21 *			messages across them to/from other partitions.
  22 *
  23 *	There are a couple of additional functions residing in XP, which
  24 *	provide an interface to XPC for its users.
  25 *
  26 *
  27 *	Caveats:
  28 *
  29 *	  . Currently on sn2, we have no way to determine which nasid an IRQ
  30 *	    came from. Thus, xpc_send_IRQ_sn2() does a remote amo write
  31 *	    followed by an IPI. The amo indicates where data is to be pulled
  32 *	    from, so after the IPI arrives, the remote partition checks the amo
  33 *	    word. The IPI can actually arrive before the amo however, so other
  34 *	    code must periodically check for this case. Also, remote amo
  35 *	    operations do not reliably time out. Thus we do a remote PIO read
  36 *	    solely to know whether the remote partition is down and whether we
  37 *	    should stop sending IPIs to it. This remote PIO read operation is
  38 *	    set up in a special nofault region so SAL knows to ignore (and
  39 *	    cleanup) any errors due to the remote amo write, PIO read, and/or
  40 *	    PIO write operations.
  41 *
  42 *	    If/when new hardware solves this IPI problem, we should abandon
  43 *	    the current approach.
  44 *
  45 */
  46
  47#include <linux/module.h>
  48#include <linux/slab.h>
  49#include <linux/sysctl.h>
  50#include <linux/device.h>
  51#include <linux/delay.h>
  52#include <linux/reboot.h>
  53#include <linux/kdebug.h>
  54#include <linux/kthread.h>
  55#include "xpc.h"
  56
  57#ifdef CONFIG_X86_64
  58#include <asm/traps.h>
  59#endif
  60
  61/* define two XPC debug device structures to be used with dev_dbg() et al */
  62
  63static struct device_driver xpc_dbg_name = {
  64	.name = "xpc"
  65};
  66
  67static struct device xpc_part_dbg_subname = {
  68	.init_name = "",	/* set to "part" at xpc_init() time */
  69	.driver = &xpc_dbg_name
  70};
  71
  72static struct device xpc_chan_dbg_subname = {
  73	.init_name = "",	/* set to "chan" at xpc_init() time */
  74	.driver = &xpc_dbg_name
  75};
  76
  77struct device *xpc_part = &xpc_part_dbg_subname;
  78struct device *xpc_chan = &xpc_chan_dbg_subname;
  79
  80static int xpc_kdebug_ignore;
  81
  82/* systune related variables for /proc/sys directories */
  83
  84static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
  85static int xpc_hb_min_interval = 1;
  86static int xpc_hb_max_interval = 10;
  87
  88static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
  89static int xpc_hb_check_min_interval = 10;
  90static int xpc_hb_check_max_interval = 120;
  91
  92int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
  93static int xpc_disengage_min_timelimit;	/* = 0 */
  94static int xpc_disengage_max_timelimit = 120;
  95
  96static struct ctl_table xpc_sys_xpc_hb_dir[] = {
  97	{
  98	 .procname = "hb_interval",
  99	 .data = &xpc_hb_interval,
 100	 .maxlen = sizeof(int),
 101	 .mode = 0644,
 102	 .proc_handler = proc_dointvec_minmax,
 103	 .extra1 = &xpc_hb_min_interval,
 104	 .extra2 = &xpc_hb_max_interval},
 105	{
 106	 .procname = "hb_check_interval",
 107	 .data = &xpc_hb_check_interval,
 108	 .maxlen = sizeof(int),
 109	 .mode = 0644,
 110	 .proc_handler = proc_dointvec_minmax,
 111	 .extra1 = &xpc_hb_check_min_interval,
 112	 .extra2 = &xpc_hb_check_max_interval},
 113	{}
 114};
 115static struct ctl_table xpc_sys_xpc_dir[] = {
 116	{
 117	 .procname = "hb",
 118	 .mode = 0555,
 119	 .child = xpc_sys_xpc_hb_dir},
 120	{
 121	 .procname = "disengage_timelimit",
 122	 .data = &xpc_disengage_timelimit,
 123	 .maxlen = sizeof(int),
 124	 .mode = 0644,
 125	 .proc_handler = proc_dointvec_minmax,
 126	 .extra1 = &xpc_disengage_min_timelimit,
 127	 .extra2 = &xpc_disengage_max_timelimit},
 128	{}
 129};
 130static struct ctl_table xpc_sys_dir[] = {
 131	{
 132	 .procname = "xpc",
 133	 .mode = 0555,
 134	 .child = xpc_sys_xpc_dir},
 135	{}
 136};
 137static struct ctl_table_header *xpc_sysctl;
 138
 139/* non-zero if any remote partition disengage was timed out */
 140int xpc_disengage_timedout;
 141
 142/* #of activate IRQs received and not yet processed */
 143int xpc_activate_IRQ_rcvd;
 144DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock);
 145
 146/* IRQ handler notifies this wait queue on receipt of an IRQ */
 147DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
 148
 149static unsigned long xpc_hb_check_timeout;
 150static struct timer_list xpc_hb_timer;
 151
 152/* notification that the xpc_hb_checker thread has exited */
 153static DECLARE_COMPLETION(xpc_hb_checker_exited);
 154
 155/* notification that the xpc_discovery thread has exited */
 156static DECLARE_COMPLETION(xpc_discovery_exited);
 157
 158static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
 159
 160static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
 161static struct notifier_block xpc_reboot_notifier = {
 162	.notifier_call = xpc_system_reboot,
 163};
 164
 165static int xpc_system_die(struct notifier_block *, unsigned long, void *);
 166static struct notifier_block xpc_die_notifier = {
 167	.notifier_call = xpc_system_die,
 168};
 169
 170struct xpc_arch_operations xpc_arch_ops;
 171
 172/*
 173 * Timer function to enforce the timelimit on the partition disengage.
 174 */
 175static void
 176xpc_timeout_partition_disengage(struct timer_list *t)
 177{
 178	struct xpc_partition *part = from_timer(part, t, disengage_timer);
 179
 180	DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
 181
 182	xpc_partition_disengaged_from_timer(part);
 183
 184	DBUG_ON(part->disengage_timeout != 0);
 185	DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part)));
 186}
 187
 188/*
 189 * Timer to produce the heartbeat.  The timer structures function is
 190 * already set when this is initially called.  A tunable is used to
 191 * specify when the next timeout should occur.
 192 */
 193static void
 194xpc_hb_beater(struct timer_list *unused)
 195{
 196	xpc_arch_ops.increment_heartbeat();
 197
 198	if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
 199		wake_up_interruptible(&xpc_activate_IRQ_wq);
 200
 201	xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
 202	add_timer(&xpc_hb_timer);
 203}
 204
 205static void
 206xpc_start_hb_beater(void)
 207{
 208	xpc_arch_ops.heartbeat_init();
 209	timer_setup(&xpc_hb_timer, xpc_hb_beater, 0);
 210	xpc_hb_beater(NULL);
 
 211}
 212
 213static void
 214xpc_stop_hb_beater(void)
 215{
 216	del_timer_sync(&xpc_hb_timer);
 217	xpc_arch_ops.heartbeat_exit();
 218}
 219
 220/*
 221 * At periodic intervals, scan through all active partitions and ensure
 222 * their heartbeat is still active.  If not, the partition is deactivated.
 223 */
 224static void
 225xpc_check_remote_hb(void)
 226{
 227	struct xpc_partition *part;
 228	short partid;
 229	enum xp_retval ret;
 230
 231	for (partid = 0; partid < xp_max_npartitions; partid++) {
 232
 233		if (xpc_exiting)
 234			break;
 235
 236		if (partid == xp_partition_id)
 237			continue;
 238
 239		part = &xpc_partitions[partid];
 240
 241		if (part->act_state == XPC_P_AS_INACTIVE ||
 242		    part->act_state == XPC_P_AS_DEACTIVATING) {
 243			continue;
 244		}
 245
 246		ret = xpc_arch_ops.get_remote_heartbeat(part);
 247		if (ret != xpSuccess)
 248			XPC_DEACTIVATE_PARTITION(part, ret);
 249	}
 250}
 251
 252/*
 253 * This thread is responsible for nearly all of the partition
 254 * activation/deactivation.
 255 */
 256static int
 257xpc_hb_checker(void *ignore)
 258{
 259	int force_IRQ = 0;
 260
 261	/* this thread was marked active by xpc_hb_init() */
 262
 263	set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU));
 264
 265	/* set our heartbeating to other partitions into motion */
 266	xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
 267	xpc_start_hb_beater();
 268
 269	while (!xpc_exiting) {
 270
 271		dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
 272			"been received\n",
 273			(int)(xpc_hb_check_timeout - jiffies),
 274			xpc_activate_IRQ_rcvd);
 275
 276		/* checking of remote heartbeats is skewed by IRQ handling */
 277		if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
 278			xpc_hb_check_timeout = jiffies +
 279			    (xpc_hb_check_interval * HZ);
 280
 281			dev_dbg(xpc_part, "checking remote heartbeats\n");
 282			xpc_check_remote_hb();
 
 
 
 
 
 
 
 283		}
 284
 285		/* check for outstanding IRQs */
 286		if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) {
 287			force_IRQ = 0;
 288			dev_dbg(xpc_part, "processing activate IRQs "
 289				"received\n");
 290			xpc_arch_ops.process_activate_IRQ_rcvd();
 291		}
 292
 293		/* wait for IRQ or timeout */
 294		(void)wait_event_interruptible(xpc_activate_IRQ_wq,
 295					       (time_is_before_eq_jiffies(
 296						xpc_hb_check_timeout) ||
 297						xpc_activate_IRQ_rcvd > 0 ||
 298						xpc_exiting));
 299	}
 300
 301	xpc_stop_hb_beater();
 302
 303	dev_dbg(xpc_part, "heartbeat checker is exiting\n");
 304
 305	/* mark this thread as having exited */
 306	complete(&xpc_hb_checker_exited);
 307	return 0;
 308}
 309
 310/*
 311 * This thread will attempt to discover other partitions to activate
 312 * based on info provided by SAL. This new thread is short lived and
 313 * will exit once discovery is complete.
 314 */
 315static int
 316xpc_initiate_discovery(void *ignore)
 317{
 318	xpc_discovery();
 319
 320	dev_dbg(xpc_part, "discovery thread is exiting\n");
 321
 322	/* mark this thread as having exited */
 323	complete(&xpc_discovery_exited);
 324	return 0;
 325}
 326
 327/*
 328 * The first kthread assigned to a newly activated partition is the one
 329 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
 330 * that kthread until the partition is brought down, at which time that kthread
 331 * returns back to XPC HB. (The return of that kthread will signify to XPC HB
 332 * that XPC has dismantled all communication infrastructure for the associated
 333 * partition.) This kthread becomes the channel manager for that partition.
 334 *
 335 * Each active partition has a channel manager, who, besides connecting and
 336 * disconnecting channels, will ensure that each of the partition's connected
 337 * channels has the required number of assigned kthreads to get the work done.
 338 */
 339static void
 340xpc_channel_mgr(struct xpc_partition *part)
 341{
 342	while (part->act_state != XPC_P_AS_DEACTIVATING ||
 343	       atomic_read(&part->nchannels_active) > 0 ||
 344	       !xpc_partition_disengaged(part)) {
 345
 346		xpc_process_sent_chctl_flags(part);
 347
 348		/*
 349		 * Wait until we've been requested to activate kthreads or
 350		 * all of the channel's message queues have been torn down or
 351		 * a signal is pending.
 352		 *
 353		 * The channel_mgr_requests is set to 1 after being awakened,
 354		 * This is done to prevent the channel mgr from making one pass
 355		 * through the loop for each request, since he will
 356		 * be servicing all the requests in one pass. The reason it's
 357		 * set to 1 instead of 0 is so that other kthreads will know
 358		 * that the channel mgr is running and won't bother trying to
 359		 * wake him up.
 360		 */
 361		atomic_dec(&part->channel_mgr_requests);
 362		(void)wait_event_interruptible(part->channel_mgr_wq,
 363				(atomic_read(&part->channel_mgr_requests) > 0 ||
 364				 part->chctl.all_flags != 0 ||
 365				 (part->act_state == XPC_P_AS_DEACTIVATING &&
 366				 atomic_read(&part->nchannels_active) == 0 &&
 367				 xpc_partition_disengaged(part))));
 368		atomic_set(&part->channel_mgr_requests, 1);
 369	}
 370}
 371
 372/*
 373 * Guarantee that the kzalloc'd memory is cacheline aligned.
 374 */
 375void *
 376xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
 377{
 378	/* see if kzalloc will give us cachline aligned memory by default */
 379	*base = kzalloc(size, flags);
 380	if (*base == NULL)
 381		return NULL;
 382
 383	if ((u64)*base == L1_CACHE_ALIGN((u64)*base))
 384		return *base;
 385
 386	kfree(*base);
 387
 388	/* nope, we'll have to do it ourselves */
 389	*base = kzalloc(size + L1_CACHE_BYTES, flags);
 390	if (*base == NULL)
 391		return NULL;
 392
 393	return (void *)L1_CACHE_ALIGN((u64)*base);
 394}
 395
 396/*
 397 * Setup the channel structures necessary to support XPartition Communication
 398 * between the specified remote partition and the local one.
 399 */
 400static enum xp_retval
 401xpc_setup_ch_structures(struct xpc_partition *part)
 402{
 403	enum xp_retval ret;
 404	int ch_number;
 405	struct xpc_channel *ch;
 406	short partid = XPC_PARTID(part);
 407
 408	/*
 409	 * Allocate all of the channel structures as a contiguous chunk of
 410	 * memory.
 411	 */
 412	DBUG_ON(part->channels != NULL);
 413	part->channels = kcalloc(XPC_MAX_NCHANNELS,
 414				 sizeof(struct xpc_channel),
 415				 GFP_KERNEL);
 416	if (part->channels == NULL) {
 417		dev_err(xpc_chan, "can't get memory for channels\n");
 418		return xpNoMemory;
 419	}
 420
 421	/* allocate the remote open and close args */
 422
 423	part->remote_openclose_args =
 424	    xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE,
 425					  GFP_KERNEL, &part->
 426					  remote_openclose_args_base);
 427	if (part->remote_openclose_args == NULL) {
 428		dev_err(xpc_chan, "can't get memory for remote connect args\n");
 429		ret = xpNoMemory;
 430		goto out_1;
 431	}
 432
 433	part->chctl.all_flags = 0;
 434	spin_lock_init(&part->chctl_lock);
 435
 436	atomic_set(&part->channel_mgr_requests, 1);
 437	init_waitqueue_head(&part->channel_mgr_wq);
 438
 439	part->nchannels = XPC_MAX_NCHANNELS;
 440
 441	atomic_set(&part->nchannels_active, 0);
 442	atomic_set(&part->nchannels_engaged, 0);
 443
 444	for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
 445		ch = &part->channels[ch_number];
 446
 447		ch->partid = partid;
 448		ch->number = ch_number;
 449		ch->flags = XPC_C_DISCONNECTED;
 450
 451		atomic_set(&ch->kthreads_assigned, 0);
 452		atomic_set(&ch->kthreads_idle, 0);
 453		atomic_set(&ch->kthreads_active, 0);
 454
 455		atomic_set(&ch->references, 0);
 456		atomic_set(&ch->n_to_notify, 0);
 457
 458		spin_lock_init(&ch->lock);
 459		init_completion(&ch->wdisconnect_wait);
 460
 461		atomic_set(&ch->n_on_msg_allocate_wq, 0);
 462		init_waitqueue_head(&ch->msg_allocate_wq);
 463		init_waitqueue_head(&ch->idle_wq);
 464	}
 465
 466	ret = xpc_arch_ops.setup_ch_structures(part);
 467	if (ret != xpSuccess)
 468		goto out_2;
 469
 470	/*
 471	 * With the setting of the partition setup_state to XPC_P_SS_SETUP,
 472	 * we're declaring that this partition is ready to go.
 473	 */
 474	part->setup_state = XPC_P_SS_SETUP;
 475
 476	return xpSuccess;
 477
 478	/* setup of ch structures failed */
 479out_2:
 480	kfree(part->remote_openclose_args_base);
 481	part->remote_openclose_args = NULL;
 482out_1:
 483	kfree(part->channels);
 484	part->channels = NULL;
 485	return ret;
 486}
 487
 488/*
 489 * Teardown the channel structures necessary to support XPartition Communication
 490 * between the specified remote partition and the local one.
 491 */
 492static void
 493xpc_teardown_ch_structures(struct xpc_partition *part)
 494{
 495	DBUG_ON(atomic_read(&part->nchannels_engaged) != 0);
 496	DBUG_ON(atomic_read(&part->nchannels_active) != 0);
 497
 498	/*
 499	 * Make this partition inaccessible to local processes by marking it
 500	 * as no longer setup. Then wait before proceeding with the teardown
 501	 * until all existing references cease.
 502	 */
 503	DBUG_ON(part->setup_state != XPC_P_SS_SETUP);
 504	part->setup_state = XPC_P_SS_WTEARDOWN;
 505
 506	wait_event(part->teardown_wq, (atomic_read(&part->references) == 0));
 507
 508	/* now we can begin tearing down the infrastructure */
 509
 510	xpc_arch_ops.teardown_ch_structures(part);
 511
 512	kfree(part->remote_openclose_args_base);
 513	part->remote_openclose_args = NULL;
 514	kfree(part->channels);
 515	part->channels = NULL;
 516
 517	part->setup_state = XPC_P_SS_TORNDOWN;
 518}
 519
 520/*
 521 * When XPC HB determines that a partition has come up, it will create a new
 522 * kthread and that kthread will call this function to attempt to set up the
 523 * basic infrastructure used for Cross Partition Communication with the newly
 524 * upped partition.
 525 *
 526 * The kthread that was created by XPC HB and which setup the XPC
 527 * infrastructure will remain assigned to the partition becoming the channel
 528 * manager for that partition until the partition is deactivating, at which
 529 * time the kthread will teardown the XPC infrastructure and then exit.
 530 */
 531static int
 532xpc_activating(void *__partid)
 533{
 534	short partid = (u64)__partid;
 535	struct xpc_partition *part = &xpc_partitions[partid];
 536	unsigned long irq_flags;
 537
 538	DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
 539
 540	spin_lock_irqsave(&part->act_lock, irq_flags);
 541
 542	if (part->act_state == XPC_P_AS_DEACTIVATING) {
 543		part->act_state = XPC_P_AS_INACTIVE;
 544		spin_unlock_irqrestore(&part->act_lock, irq_flags);
 545		part->remote_rp_pa = 0;
 546		return 0;
 547	}
 548
 549	/* indicate the thread is activating */
 550	DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
 551	part->act_state = XPC_P_AS_ACTIVATING;
 552
 553	XPC_SET_REASON(part, 0, 0);
 554	spin_unlock_irqrestore(&part->act_lock, irq_flags);
 555
 556	dev_dbg(xpc_part, "activating partition %d\n", partid);
 557
 558	xpc_arch_ops.allow_hb(partid);
 559
 560	if (xpc_setup_ch_structures(part) == xpSuccess) {
 561		(void)xpc_part_ref(part);	/* this will always succeed */
 562
 563		if (xpc_arch_ops.make_first_contact(part) == xpSuccess) {
 564			xpc_mark_partition_active(part);
 565			xpc_channel_mgr(part);
 566			/* won't return until partition is deactivating */
 567		}
 568
 569		xpc_part_deref(part);
 570		xpc_teardown_ch_structures(part);
 571	}
 572
 573	xpc_arch_ops.disallow_hb(partid);
 574	xpc_mark_partition_inactive(part);
 575
 576	if (part->reason == xpReactivating) {
 577		/* interrupting ourselves results in activating partition */
 578		xpc_arch_ops.request_partition_reactivation(part);
 579	}
 580
 581	return 0;
 582}
 583
 584void
 585xpc_activate_partition(struct xpc_partition *part)
 586{
 587	short partid = XPC_PARTID(part);
 588	unsigned long irq_flags;
 589	struct task_struct *kthread;
 590
 591	spin_lock_irqsave(&part->act_lock, irq_flags);
 592
 593	DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
 594
 595	part->act_state = XPC_P_AS_ACTIVATION_REQ;
 596	XPC_SET_REASON(part, xpCloneKThread, __LINE__);
 597
 598	spin_unlock_irqrestore(&part->act_lock, irq_flags);
 599
 600	kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
 601			      partid);
 602	if (IS_ERR(kthread)) {
 603		spin_lock_irqsave(&part->act_lock, irq_flags);
 604		part->act_state = XPC_P_AS_INACTIVE;
 605		XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
 606		spin_unlock_irqrestore(&part->act_lock, irq_flags);
 607	}
 608}
 609
 610void
 611xpc_activate_kthreads(struct xpc_channel *ch, int needed)
 612{
 613	int idle = atomic_read(&ch->kthreads_idle);
 614	int assigned = atomic_read(&ch->kthreads_assigned);
 615	int wakeup;
 616
 617	DBUG_ON(needed <= 0);
 618
 619	if (idle > 0) {
 620		wakeup = (needed > idle) ? idle : needed;
 621		needed -= wakeup;
 622
 623		dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
 624			"channel=%d\n", wakeup, ch->partid, ch->number);
 625
 626		/* only wakeup the requested number of kthreads */
 627		wake_up_nr(&ch->idle_wq, wakeup);
 628	}
 629
 630	if (needed <= 0)
 631		return;
 632
 633	if (needed + assigned > ch->kthreads_assigned_limit) {
 634		needed = ch->kthreads_assigned_limit - assigned;
 635		if (needed <= 0)
 636			return;
 637	}
 638
 639	dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
 640		needed, ch->partid, ch->number);
 641
 642	xpc_create_kthreads(ch, needed, 0);
 643}
 644
 645/*
 646 * This function is where XPC's kthreads wait for messages to deliver.
 647 */
 648static void
 649xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
 650{
 651	int (*n_of_deliverable_payloads) (struct xpc_channel *) =
 652		xpc_arch_ops.n_of_deliverable_payloads;
 653
 654	do {
 655		/* deliver messages to their intended recipients */
 656
 657		while (n_of_deliverable_payloads(ch) > 0 &&
 658		       !(ch->flags & XPC_C_DISCONNECTING)) {
 659			xpc_deliver_payload(ch);
 660		}
 661
 662		if (atomic_inc_return(&ch->kthreads_idle) >
 663		    ch->kthreads_idle_limit) {
 664			/* too many idle kthreads on this channel */
 665			atomic_dec(&ch->kthreads_idle);
 666			break;
 667		}
 668
 669		dev_dbg(xpc_chan, "idle kthread calling "
 670			"wait_event_interruptible_exclusive()\n");
 671
 672		(void)wait_event_interruptible_exclusive(ch->idle_wq,
 673				(n_of_deliverable_payloads(ch) > 0 ||
 674				 (ch->flags & XPC_C_DISCONNECTING)));
 675
 676		atomic_dec(&ch->kthreads_idle);
 677
 678	} while (!(ch->flags & XPC_C_DISCONNECTING));
 679}
 680
 681static int
 682xpc_kthread_start(void *args)
 683{
 684	short partid = XPC_UNPACK_ARG1(args);
 685	u16 ch_number = XPC_UNPACK_ARG2(args);
 686	struct xpc_partition *part = &xpc_partitions[partid];
 687	struct xpc_channel *ch;
 688	int n_needed;
 689	unsigned long irq_flags;
 690	int (*n_of_deliverable_payloads) (struct xpc_channel *) =
 691		xpc_arch_ops.n_of_deliverable_payloads;
 692
 693	dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
 694		partid, ch_number);
 695
 696	ch = &part->channels[ch_number];
 697
 698	if (!(ch->flags & XPC_C_DISCONNECTING)) {
 699
 700		/* let registerer know that connection has been established */
 701
 702		spin_lock_irqsave(&ch->lock, irq_flags);
 703		if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
 704			ch->flags |= XPC_C_CONNECTEDCALLOUT;
 705			spin_unlock_irqrestore(&ch->lock, irq_flags);
 706
 707			xpc_connected_callout(ch);
 708
 709			spin_lock_irqsave(&ch->lock, irq_flags);
 710			ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
 711			spin_unlock_irqrestore(&ch->lock, irq_flags);
 712
 713			/*
 714			 * It is possible that while the callout was being
 715			 * made that the remote partition sent some messages.
 716			 * If that is the case, we may need to activate
 717			 * additional kthreads to help deliver them. We only
 718			 * need one less than total #of messages to deliver.
 719			 */
 720			n_needed = n_of_deliverable_payloads(ch) - 1;
 721			if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
 722				xpc_activate_kthreads(ch, n_needed);
 723
 724		} else {
 725			spin_unlock_irqrestore(&ch->lock, irq_flags);
 726		}
 727
 728		xpc_kthread_waitmsgs(part, ch);
 729	}
 730
 731	/* let registerer know that connection is disconnecting */
 732
 733	spin_lock_irqsave(&ch->lock, irq_flags);
 734	if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
 735	    !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
 736		ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
 737		spin_unlock_irqrestore(&ch->lock, irq_flags);
 738
 739		xpc_disconnect_callout(ch, xpDisconnecting);
 740
 741		spin_lock_irqsave(&ch->lock, irq_flags);
 742		ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
 743	}
 744	spin_unlock_irqrestore(&ch->lock, irq_flags);
 745
 746	if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 747	    atomic_dec_return(&part->nchannels_engaged) == 0) {
 748		xpc_arch_ops.indicate_partition_disengaged(part);
 749	}
 750
 751	xpc_msgqueue_deref(ch);
 752
 753	dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
 754		partid, ch_number);
 755
 756	xpc_part_deref(part);
 757	return 0;
 758}
 759
 760/*
 761 * For each partition that XPC has established communications with, there is
 762 * a minimum of one kernel thread assigned to perform any operation that
 763 * may potentially sleep or block (basically the callouts to the asynchronous
 764 * functions registered via xpc_connect()).
 765 *
 766 * Additional kthreads are created and destroyed by XPC as the workload
 767 * demands.
 768 *
 769 * A kthread is assigned to one of the active channels that exists for a given
 770 * partition.
 771 */
 772void
 773xpc_create_kthreads(struct xpc_channel *ch, int needed,
 774		    int ignore_disconnecting)
 775{
 776	unsigned long irq_flags;
 777	u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
 778	struct xpc_partition *part = &xpc_partitions[ch->partid];
 779	struct task_struct *kthread;
 780	void (*indicate_partition_disengaged) (struct xpc_partition *) =
 781		xpc_arch_ops.indicate_partition_disengaged;
 782
 783	while (needed-- > 0) {
 784
 785		/*
 786		 * The following is done on behalf of the newly created
 787		 * kthread. That kthread is responsible for doing the
 788		 * counterpart to the following before it exits.
 789		 */
 790		if (ignore_disconnecting) {
 791			if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
 792				/* kthreads assigned had gone to zero */
 793				BUG_ON(!(ch->flags &
 794					 XPC_C_DISCONNECTINGCALLOUT_MADE));
 795				break;
 796			}
 797
 798		} else if (ch->flags & XPC_C_DISCONNECTING) {
 799			break;
 800
 801		} else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
 802			   atomic_inc_return(&part->nchannels_engaged) == 1) {
 803			xpc_arch_ops.indicate_partition_engaged(part);
 804		}
 805		(void)xpc_part_ref(part);
 806		xpc_msgqueue_ref(ch);
 807
 808		kthread = kthread_run(xpc_kthread_start, (void *)args,
 809				      "xpc%02dc%d", ch->partid, ch->number);
 810		if (IS_ERR(kthread)) {
 811			/* the fork failed */
 812
 813			/*
 814			 * NOTE: if (ignore_disconnecting &&
 815			 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
 816			 * then we'll deadlock if all other kthreads assigned
 817			 * to this channel are blocked in the channel's
 818			 * registerer, because the only thing that will unblock
 819			 * them is the xpDisconnecting callout that this
 820			 * failed kthread_run() would have made.
 821			 */
 822
 823			if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 824			    atomic_dec_return(&part->nchannels_engaged) == 0) {
 825				indicate_partition_disengaged(part);
 826			}
 827			xpc_msgqueue_deref(ch);
 828			xpc_part_deref(part);
 829
 830			if (atomic_read(&ch->kthreads_assigned) <
 831			    ch->kthreads_idle_limit) {
 832				/*
 833				 * Flag this as an error only if we have an
 834				 * insufficient #of kthreads for the channel
 835				 * to function.
 836				 */
 837				spin_lock_irqsave(&ch->lock, irq_flags);
 838				XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
 839						       &irq_flags);
 840				spin_unlock_irqrestore(&ch->lock, irq_flags);
 841			}
 842			break;
 843		}
 844	}
 845}
 846
 847void
 848xpc_disconnect_wait(int ch_number)
 849{
 850	unsigned long irq_flags;
 851	short partid;
 852	struct xpc_partition *part;
 853	struct xpc_channel *ch;
 854	int wakeup_channel_mgr;
 855
 856	/* now wait for all callouts to the caller's function to cease */
 857	for (partid = 0; partid < xp_max_npartitions; partid++) {
 858		part = &xpc_partitions[partid];
 859
 860		if (!xpc_part_ref(part))
 861			continue;
 862
 863		ch = &part->channels[ch_number];
 864
 865		if (!(ch->flags & XPC_C_WDISCONNECT)) {
 866			xpc_part_deref(part);
 867			continue;
 868		}
 869
 870		wait_for_completion(&ch->wdisconnect_wait);
 871
 872		spin_lock_irqsave(&ch->lock, irq_flags);
 873		DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
 874		wakeup_channel_mgr = 0;
 875
 876		if (ch->delayed_chctl_flags) {
 877			if (part->act_state != XPC_P_AS_DEACTIVATING) {
 878				spin_lock(&part->chctl_lock);
 879				part->chctl.flags[ch->number] |=
 880				    ch->delayed_chctl_flags;
 881				spin_unlock(&part->chctl_lock);
 882				wakeup_channel_mgr = 1;
 883			}
 884			ch->delayed_chctl_flags = 0;
 885		}
 886
 887		ch->flags &= ~XPC_C_WDISCONNECT;
 888		spin_unlock_irqrestore(&ch->lock, irq_flags);
 889
 890		if (wakeup_channel_mgr)
 891			xpc_wakeup_channel_mgr(part);
 892
 893		xpc_part_deref(part);
 894	}
 895}
 896
 897static int
 898xpc_setup_partitions(void)
 899{
 900	short partid;
 901	struct xpc_partition *part;
 902
 903	xpc_partitions = kcalloc(xp_max_npartitions,
 904				 sizeof(struct xpc_partition),
 905				 GFP_KERNEL);
 906	if (xpc_partitions == NULL) {
 907		dev_err(xpc_part, "can't get memory for partition structure\n");
 908		return -ENOMEM;
 909	}
 910
 911	/*
 912	 * The first few fields of each entry of xpc_partitions[] need to
 913	 * be initialized now so that calls to xpc_connect() and
 914	 * xpc_disconnect() can be made prior to the activation of any remote
 915	 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
 916	 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
 917	 * PARTITION HAS BEEN ACTIVATED.
 918	 */
 919	for (partid = 0; partid < xp_max_npartitions; partid++) {
 920		part = &xpc_partitions[partid];
 921
 922		DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
 923
 924		part->activate_IRQ_rcvd = 0;
 925		spin_lock_init(&part->act_lock);
 926		part->act_state = XPC_P_AS_INACTIVE;
 927		XPC_SET_REASON(part, 0, 0);
 928
 929		timer_setup(&part->disengage_timer,
 930			    xpc_timeout_partition_disengage, 0);
 
 
 931
 932		part->setup_state = XPC_P_SS_UNSET;
 933		init_waitqueue_head(&part->teardown_wq);
 934		atomic_set(&part->references, 0);
 935	}
 936
 937	return xpc_arch_ops.setup_partitions();
 938}
 939
 940static void
 941xpc_teardown_partitions(void)
 942{
 943	xpc_arch_ops.teardown_partitions();
 944	kfree(xpc_partitions);
 945}
 946
 947static void
 948xpc_do_exit(enum xp_retval reason)
 949{
 950	short partid;
 951	int active_part_count, printed_waiting_msg = 0;
 952	struct xpc_partition *part;
 953	unsigned long printmsg_time, disengage_timeout = 0;
 954
 955	/* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
 956	DBUG_ON(xpc_exiting == 1);
 957
 958	/*
 959	 * Let the heartbeat checker thread and the discovery thread
 960	 * (if one is running) know that they should exit. Also wake up
 961	 * the heartbeat checker thread in case it's sleeping.
 962	 */
 963	xpc_exiting = 1;
 964	wake_up_interruptible(&xpc_activate_IRQ_wq);
 965
 966	/* wait for the discovery thread to exit */
 967	wait_for_completion(&xpc_discovery_exited);
 968
 969	/* wait for the heartbeat checker thread to exit */
 970	wait_for_completion(&xpc_hb_checker_exited);
 971
 972	/* sleep for a 1/3 of a second or so */
 973	(void)msleep_interruptible(300);
 974
 975	/* wait for all partitions to become inactive */
 976
 977	printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
 978	xpc_disengage_timedout = 0;
 979
 980	do {
 981		active_part_count = 0;
 982
 983		for (partid = 0; partid < xp_max_npartitions; partid++) {
 984			part = &xpc_partitions[partid];
 985
 986			if (xpc_partition_disengaged(part) &&
 987			    part->act_state == XPC_P_AS_INACTIVE) {
 988				continue;
 989			}
 990
 991			active_part_count++;
 992
 993			XPC_DEACTIVATE_PARTITION(part, reason);
 994
 995			if (part->disengage_timeout > disengage_timeout)
 996				disengage_timeout = part->disengage_timeout;
 997		}
 998
 999		if (xpc_arch_ops.any_partition_engaged()) {
1000			if (time_is_before_jiffies(printmsg_time)) {
1001				dev_info(xpc_part, "waiting for remote "
1002					 "partitions to deactivate, timeout in "
1003					 "%ld seconds\n", (disengage_timeout -
1004					 jiffies) / HZ);
1005				printmsg_time = jiffies +
1006				    (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
1007				printed_waiting_msg = 1;
1008			}
1009
1010		} else if (active_part_count > 0) {
1011			if (printed_waiting_msg) {
1012				dev_info(xpc_part, "waiting for local partition"
1013					 " to deactivate\n");
1014				printed_waiting_msg = 0;
1015			}
1016
1017		} else {
1018			if (!xpc_disengage_timedout) {
1019				dev_info(xpc_part, "all partitions have "
1020					 "deactivated\n");
1021			}
1022			break;
1023		}
1024
1025		/* sleep for a 1/3 of a second or so */
1026		(void)msleep_interruptible(300);
1027
1028	} while (1);
1029
1030	DBUG_ON(xpc_arch_ops.any_partition_engaged());
1031
1032	xpc_teardown_rsvd_page();
1033
1034	if (reason == xpUnloading) {
1035		(void)unregister_die_notifier(&xpc_die_notifier);
1036		(void)unregister_reboot_notifier(&xpc_reboot_notifier);
1037	}
1038
1039	/* clear the interface to XPC's functions */
1040	xpc_clear_interface();
1041
1042	if (xpc_sysctl)
1043		unregister_sysctl_table(xpc_sysctl);
1044
1045	xpc_teardown_partitions();
1046
1047	if (is_uv_system())
 
 
1048		xpc_exit_uv();
1049}
1050
1051/*
1052 * This function is called when the system is being rebooted.
1053 */
1054static int
1055xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1056{
1057	enum xp_retval reason;
1058
1059	switch (event) {
1060	case SYS_RESTART:
1061		reason = xpSystemReboot;
1062		break;
1063	case SYS_HALT:
1064		reason = xpSystemHalt;
1065		break;
1066	case SYS_POWER_OFF:
1067		reason = xpSystemPoweroff;
1068		break;
1069	default:
1070		reason = xpSystemGoingDown;
1071	}
1072
1073	xpc_do_exit(reason);
1074	return NOTIFY_DONE;
1075}
1076
1077/* Used to only allow one cpu to complete disconnect */
1078static unsigned int xpc_die_disconnecting;
1079
1080/*
1081 * Notify other partitions to deactivate from us by first disengaging from all
1082 * references to our memory.
1083 */
1084static void
1085xpc_die_deactivate(void)
1086{
1087	struct xpc_partition *part;
1088	short partid;
1089	int any_engaged;
1090	long keep_waiting;
1091	long wait_to_print;
1092
1093	if (cmpxchg(&xpc_die_disconnecting, 0, 1))
1094		return;
1095
1096	/* keep xpc_hb_checker thread from doing anything (just in case) */
1097	xpc_exiting = 1;
1098
1099	xpc_arch_ops.disallow_all_hbs();   /*indicate we're deactivated */
1100
1101	for (partid = 0; partid < xp_max_npartitions; partid++) {
1102		part = &xpc_partitions[partid];
1103
1104		if (xpc_arch_ops.partition_engaged(partid) ||
1105		    part->act_state != XPC_P_AS_INACTIVE) {
1106			xpc_arch_ops.request_partition_deactivation(part);
1107			xpc_arch_ops.indicate_partition_disengaged(part);
1108		}
1109	}
1110
1111	/*
1112	 * Though we requested that all other partitions deactivate from us,
1113	 * we only wait until they've all disengaged or we've reached the
1114	 * defined timelimit.
1115	 *
1116	 * Given that one iteration through the following while-loop takes
1117	 * approximately 200 microseconds, calculate the #of loops to take
1118	 * before bailing and the #of loops before printing a waiting message.
1119	 */
1120	keep_waiting = xpc_disengage_timelimit * 1000 * 5;
1121	wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
1122
1123	while (1) {
1124		any_engaged = xpc_arch_ops.any_partition_engaged();
1125		if (!any_engaged) {
1126			dev_info(xpc_part, "all partitions have deactivated\n");
1127			break;
1128		}
1129
1130		if (!keep_waiting--) {
1131			for (partid = 0; partid < xp_max_npartitions;
1132			     partid++) {
1133				if (xpc_arch_ops.partition_engaged(partid)) {
1134					dev_info(xpc_part, "deactivate from "
1135						 "remote partition %d timed "
1136						 "out\n", partid);
1137				}
1138			}
1139			break;
1140		}
1141
1142		if (!wait_to_print--) {
1143			dev_info(xpc_part, "waiting for remote partitions to "
1144				 "deactivate, timeout in %ld seconds\n",
1145				 keep_waiting / (1000 * 5));
1146			wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1147			    1000 * 5;
1148		}
1149
1150		udelay(200);
1151	}
1152}
1153
1154/*
1155 * This function is called when the system is being restarted or halted due
1156 * to some sort of system failure. If this is the case we need to notify the
1157 * other partitions to disengage from all references to our memory.
1158 * This function can also be called when our heartbeater could be offlined
1159 * for a time. In this case we need to notify other partitions to not worry
1160 * about the lack of a heartbeat.
1161 */
1162static int
1163xpc_system_die(struct notifier_block *nb, unsigned long event, void *_die_args)
1164{
1165#ifdef CONFIG_IA64		/* !!! temporary kludge */
1166	switch (event) {
1167	case DIE_MACHINE_RESTART:
1168	case DIE_MACHINE_HALT:
1169		xpc_die_deactivate();
1170		break;
1171
1172	case DIE_KDEBUG_ENTER:
1173		/* Should lack of heartbeat be ignored by other partitions? */
1174		if (!xpc_kdebug_ignore)
1175			break;
1176
1177		fallthrough;
1178	case DIE_MCA_MONARCH_ENTER:
1179	case DIE_INIT_MONARCH_ENTER:
1180		xpc_arch_ops.offline_heartbeat();
1181		break;
1182
1183	case DIE_KDEBUG_LEAVE:
1184		/* Is lack of heartbeat being ignored by other partitions? */
1185		if (!xpc_kdebug_ignore)
1186			break;
1187
1188		fallthrough;
1189	case DIE_MCA_MONARCH_LEAVE:
1190	case DIE_INIT_MONARCH_LEAVE:
1191		xpc_arch_ops.online_heartbeat();
1192		break;
1193	}
1194#else
1195	struct die_args *die_args = _die_args;
1196
1197	switch (event) {
1198	case DIE_TRAP:
1199		if (die_args->trapnr == X86_TRAP_DF)
1200			xpc_die_deactivate();
1201
1202		if (((die_args->trapnr == X86_TRAP_MF) ||
1203		     (die_args->trapnr == X86_TRAP_XF)) &&
1204		    !user_mode(die_args->regs))
1205			xpc_die_deactivate();
1206
1207		break;
1208	case DIE_INT3:
1209	case DIE_DEBUG:
1210		break;
1211	case DIE_OOPS:
1212	case DIE_GPF:
1213	default:
1214		xpc_die_deactivate();
1215	}
1216#endif
1217
1218	return NOTIFY_DONE;
1219}
1220
1221static int __init
1222xpc_init(void)
1223{
1224	int ret;
1225	struct task_struct *kthread;
1226
1227	dev_set_name(xpc_part, "part");
1228	dev_set_name(xpc_chan, "chan");
1229
1230	if (is_uv_system()) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1231		ret = xpc_init_uv();
1232
1233	} else {
1234		ret = -ENODEV;
1235	}
1236
1237	if (ret != 0)
1238		return ret;
1239
1240	ret = xpc_setup_partitions();
1241	if (ret != 0) {
1242		dev_err(xpc_part, "can't get memory for partition structure\n");
1243		goto out_1;
1244	}
1245
1246	xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1247
1248	/*
1249	 * Fill the partition reserved page with the information needed by
1250	 * other partitions to discover we are alive and establish initial
1251	 * communications.
1252	 */
1253	ret = xpc_setup_rsvd_page();
1254	if (ret != 0) {
1255		dev_err(xpc_part, "can't setup our reserved page\n");
1256		goto out_2;
1257	}
1258
1259	/* add ourselves to the reboot_notifier_list */
1260	ret = register_reboot_notifier(&xpc_reboot_notifier);
1261	if (ret != 0)
1262		dev_warn(xpc_part, "can't register reboot notifier\n");
1263
1264	/* add ourselves to the die_notifier list */
1265	ret = register_die_notifier(&xpc_die_notifier);
1266	if (ret != 0)
1267		dev_warn(xpc_part, "can't register die notifier\n");
1268
1269	/*
1270	 * The real work-horse behind xpc.  This processes incoming
1271	 * interrupts and monitors remote heartbeats.
1272	 */
1273	kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1274	if (IS_ERR(kthread)) {
1275		dev_err(xpc_part, "failed while forking hb check thread\n");
1276		ret = -EBUSY;
1277		goto out_3;
1278	}
1279
1280	/*
1281	 * Startup a thread that will attempt to discover other partitions to
1282	 * activate based on info provided by SAL. This new thread is short
1283	 * lived and will exit once discovery is complete.
1284	 */
1285	kthread = kthread_run(xpc_initiate_discovery, NULL,
1286			      XPC_DISCOVERY_THREAD_NAME);
1287	if (IS_ERR(kthread)) {
1288		dev_err(xpc_part, "failed while forking discovery thread\n");
1289
1290		/* mark this new thread as a non-starter */
1291		complete(&xpc_discovery_exited);
1292
1293		xpc_do_exit(xpUnloading);
1294		return -EBUSY;
1295	}
1296
1297	/* set the interface to point at XPC's functions */
1298	xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1299			  xpc_initiate_send, xpc_initiate_send_notify,
1300			  xpc_initiate_received, xpc_initiate_partid_to_nasids);
1301
1302	return 0;
1303
1304	/* initialization was not successful */
1305out_3:
1306	xpc_teardown_rsvd_page();
1307
1308	(void)unregister_die_notifier(&xpc_die_notifier);
1309	(void)unregister_reboot_notifier(&xpc_reboot_notifier);
1310out_2:
1311	if (xpc_sysctl)
1312		unregister_sysctl_table(xpc_sysctl);
1313
1314	xpc_teardown_partitions();
1315out_1:
1316	if (is_uv_system())
 
 
1317		xpc_exit_uv();
1318	return ret;
1319}
1320
1321module_init(xpc_init);
1322
1323static void __exit
1324xpc_exit(void)
1325{
1326	xpc_do_exit(xpUnloading);
1327}
1328
1329module_exit(xpc_exit);
1330
1331MODULE_AUTHOR("Silicon Graphics, Inc.");
1332MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1333MODULE_LICENSE("GPL");
1334
1335module_param(xpc_hb_interval, int, 0);
1336MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1337		 "heartbeat increments.");
1338
1339module_param(xpc_hb_check_interval, int, 0);
1340MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1341		 "heartbeat checks.");
1342
1343module_param(xpc_disengage_timelimit, int, 0);
1344MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1345		 "for disengage to complete.");
1346
1347module_param(xpc_kdebug_ignore, int, 0);
1348MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1349		 "other partitions when dropping into kdebug.");