xpc_main.c

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