arch/ia64/sn/kernel/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  * Copyright (c) 2004-2006 Silicon Graphics, Inc.  All Rights Reserved.
   7  */
   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  *        . We currently have no way to determine which nasid an IPI came
  30  *          from. Thus, xpc_IPI_send() does a remote AMO write followed by
  31  *          an IPI. The AMO indicates where data is to be pulled from, so
  32  *          after the IPI arrives, the remote partition checks the AMO word.
  33  *          The IPI can actually arrive before the AMO however, so other code
  34  *          must periodically check for this case. Also, remote AMO operations
  35  *          do not reliably time out. Thus we do a remote PIO read solely to
  36  *          know whether the remote partition is down and whether we should
  37  *          stop sending IPIs to it. This remote PIO read operation is set up
  38  *          in a special nofault region so SAL knows to ignore (and cleanup)
  39  *          any errors due to the remote AMO write, PIO read, and/or PIO
  40  *          write operations.
  41  *
  42  *          If/when new hardware solves this IPI problem, we should abandon
  43  *          the current approach.
  44  *
  45  */
  46
  47
  48 #include <linux/kernel.h>
  49 #include <linux/module.h>
  50 #include <linux/init.h>
  51 #include <linux/sched.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/cache.h>
  54 #include <linux/interrupt.h>
  55 #include <linux/slab.h>
  56 #include <linux/delay.h>
  57 #include <linux/reboot.h>
  58 #include <linux/completion.h>
  59 #include <asm/sn/intr.h>
  60 #include <asm/sn/sn_sal.h>
  61 #include <asm/kdebug.h>
  62 #include <asm/uaccess.h>
  63 #include <asm/sn/xpc.h>
  64
  65
  66 /* define two XPC debug device structures to be used with dev_dbg() et al */
  67
  68 struct device_driver xpc_dbg_name = {
  69         .name = "xpc"
  70 };
  71
  72 struct device xpc_part_dbg_subname = {
  73         .bus_id = {0},          /* set to "part" at xpc_init() time */
  74         .driver = &xpc_dbg_name
  75 };
  76
  77 struct device xpc_chan_dbg_subname = {
  78         .bus_id = {0},          /* set to "chan" at xpc_init() time */
  79         .driver = &xpc_dbg_name
  80 };
  81
  82 struct device *xpc_part = &xpc_part_dbg_subname;
  83 struct device *xpc_chan = &xpc_chan_dbg_subname;
  84
  85
  86 static int xpc_kdebug_ignore;
  87
  88
  89 /* systune related variables for /proc/sys directories */
  90
  91 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
  92 static int xpc_hb_min_interval = 1;
  93 static int xpc_hb_max_interval = 10;
  94
  95 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
  96 static int xpc_hb_check_min_interval = 10;
  97 static int xpc_hb_check_max_interval = 120;
  98
  99 int xpc_disengage_request_timelimit = XPC_DISENGAGE_REQUEST_DEFAULT_TIMELIMIT;
 100 static int xpc_disengage_request_min_timelimit = 0;
 101 static int xpc_disengage_request_max_timelimit = 120;
 102
 103 static ctl_table xpc_sys_xpc_hb_dir[] = {
 104         {
 105                 1,
 106                 "hb_interval",
 107                 &xpc_hb_interval,
 108                 sizeof(int),
 109                 0644,
 110                 NULL,
 111                 &proc_dointvec_minmax,
 112                 &sysctl_intvec,
 113                 NULL,
 114                 &xpc_hb_min_interval,
 115                 &xpc_hb_max_interval
 116         },
 117         {
 118                 2,
 119                 "hb_check_interval",
 120                 &xpc_hb_check_interval,
 121                 sizeof(int),
 122                 0644,
 123                 NULL,
 124                 &proc_dointvec_minmax,
 125                 &sysctl_intvec,
 126                 NULL,
 127                 &xpc_hb_check_min_interval,
 128                 &xpc_hb_check_max_interval
 129         },
 130         {0}
 131 };
 132 static ctl_table xpc_sys_xpc_dir[] = {
 133         {
 134                 1,
 135                 "hb",
 136                 NULL,
 137                 0,
 138                 0555,
 139                 xpc_sys_xpc_hb_dir
 140         },
 141         {
 142                 2,
 143                 "disengage_request_timelimit",
 144                 &xpc_disengage_request_timelimit,
 145                 sizeof(int),
 146                 0644,
 147                 NULL,
 148                 &proc_dointvec_minmax,
 149                 &sysctl_intvec,
 150                 NULL,
 151                 &xpc_disengage_request_min_timelimit,
 152                 &xpc_disengage_request_max_timelimit
 153         },
 154         {0}
 155 };
 156 static ctl_table xpc_sys_dir[] = {
 157         {
 158                 1,
 159                 "xpc",
 160                 NULL,
 161                 0,
 162                 0555,
 163                 xpc_sys_xpc_dir
 164         },
 165         {0}
 166 };
 167 static struct ctl_table_header *xpc_sysctl;
 168
 169 /* non-zero if any remote partition disengage request was timed out */
 170 int xpc_disengage_request_timedout;
 171
 172 /* #of IRQs received */
 173 static atomic_t xpc_act_IRQ_rcvd;
 174
 175 /* IRQ handler notifies this wait queue on receipt of an IRQ */
 176 static DECLARE_WAIT_QUEUE_HEAD(xpc_act_IRQ_wq);
 177
 178 static unsigned long xpc_hb_check_timeout;
 179
 180 /* notification that the xpc_hb_checker thread has exited */
 181 static DECLARE_COMPLETION(xpc_hb_checker_exited);
 182
 183 /* notification that the xpc_discovery thread has exited */
 184 static DECLARE_COMPLETION(xpc_discovery_exited);
 185
 186
 187 static struct timer_list xpc_hb_timer;
 188
 189
 190 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
 191
 192
 193 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
 194 static struct notifier_block xpc_reboot_notifier = {
 195         .notifier_call = xpc_system_reboot,
 196 };
 197
 198 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
 199 static struct notifier_block xpc_die_notifier = {
 200         .notifier_call = xpc_system_die,
 201 };
 202
 203
 204 /*
 205  * Timer function to enforce the timelimit on the partition disengage request.
 206  */
 207 static void
 208 xpc_timeout_partition_disengage_request(unsigned long data)
 209 {
 210         struct xpc_partition *part = (struct xpc_partition *) data;
 211
 212
 213         DBUG_ON(jiffies < part->disengage_request_timeout);
 214
 215         (void) xpc_partition_disengaged(part);
 216
 217         DBUG_ON(part->disengage_request_timeout != 0);
 218         DBUG_ON(xpc_partition_engaged(1UL << XPC_PARTID(part)) != 0);
 219 }
 220
 221
 222 /*
 223  * Notify the heartbeat check thread that an IRQ has been received.
 224  */
 225 static irqreturn_t
 226 xpc_act_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
 227 {
 228         atomic_inc(&xpc_act_IRQ_rcvd);
 229         wake_up_interruptible(&xpc_act_IRQ_wq);
 230         return IRQ_HANDLED;
 231 }
 232
 233
 234 /*
 235  * Timer to produce the heartbeat.  The timer structures function is
 236  * already set when this is initially called.  A tunable is used to
 237  * specify when the next timeout should occur.
 238  */
 239 static void
 240 xpc_hb_beater(unsigned long dummy)
 241 {
 242         xpc_vars->heartbeat++;
 243
 244         if (jiffies >= xpc_hb_check_timeout) {
 245                 wake_up_interruptible(&xpc_act_IRQ_wq);
 246         }
 247
 248         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
 249         add_timer(&xpc_hb_timer);
 250 }
 251
 252
 253 /*
 254  * This thread is responsible for nearly all of the partition
 255  * activation/deactivation.
 256  */
 257 static int
 258 xpc_hb_checker(void *ignore)
 259 {
 260         int last_IRQ_count = 0;
 261         int new_IRQ_count;
 262         int force_IRQ=0;
 263
 264
 265         /* this thread was marked active by xpc_hb_init() */
 266
 267         daemonize(XPC_HB_CHECK_THREAD_NAME);
 268
 269         set_cpus_allowed(current, cpumask_of_cpu(XPC_HB_CHECK_CPU));
 270
 271         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
 272
 273         while (!(volatile int) xpc_exiting) {
 274
 275                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
 276                         "been received\n",
 277                         (int) (xpc_hb_check_timeout - jiffies),
 278                         atomic_read(&xpc_act_IRQ_rcvd) - last_IRQ_count);
 279
 280
 281                 /* checking of remote heartbeats is skewed by IRQ handling */
 282                 if (jiffies >= xpc_hb_check_timeout) {
 283                         dev_dbg(xpc_part, "checking remote heartbeats\n");
 284                         xpc_check_remote_hb();
 285
 286                         /*
 287                          * We need to periodically recheck to ensure no
 288                          * IPI/AMO pairs have been missed.  That check
 289                          * must always reset xpc_hb_check_timeout.
 290                          */
 291                         force_IRQ = 1;
 292                 }
 293
 294
 295                 /* check for outstanding IRQs */
 296                 new_IRQ_count = atomic_read(&xpc_act_IRQ_rcvd);
 297                 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
 298                         force_IRQ = 0;
 299
 300                         dev_dbg(xpc_part, "found an IRQ to process; will be "
 301                                 "resetting xpc_hb_check_timeout\n");
 302
 303                         last_IRQ_count += xpc_identify_act_IRQ_sender();
 304                         if (last_IRQ_count < new_IRQ_count) {
 305                                 /* retry once to help avoid missing AMO */
 306                                 (void) xpc_identify_act_IRQ_sender();
 307                         }
 308                         last_IRQ_count = new_IRQ_count;
 309
 310                         xpc_hb_check_timeout = jiffies +
 311                                            (xpc_hb_check_interval * HZ);
 312                 }
 313
 314                 /* wait for IRQ or timeout */
 315                 (void) wait_event_interruptible(xpc_act_IRQ_wq,
 316                             (last_IRQ_count < atomic_read(&xpc_act_IRQ_rcvd) ||
 317                                         jiffies >= xpc_hb_check_timeout ||
 318                                                 (volatile int) xpc_exiting));
 319         }
 320
 321         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
 322
 323
 324         /* mark this thread as having exited */
 325         complete(&xpc_hb_checker_exited);
 326         return 0;
 327 }
 328
 329
 330 /*
 331  * This thread will attempt to discover other partitions to activate
 332  * based on info provided by SAL. This new thread is short lived and
 333  * will exit once discovery is complete.
 334  */
 335 static int
 336 xpc_initiate_discovery(void *ignore)
 337 {
 338         daemonize(XPC_DISCOVERY_THREAD_NAME);
 339
 340         xpc_discovery();
 341
 342         dev_dbg(xpc_part, "discovery thread is exiting\n");
 343
 344         /* mark this thread as having exited */
 345         complete(&xpc_discovery_exited);
 346         return 0;
 347 }
 348
 349
 350 /*
 351  * Establish first contact with the remote partititon. This involves pulling
 352  * the XPC per partition variables from the remote partition and waiting for
 353  * the remote partition to pull ours.
 354  */
 355 static enum xpc_retval
 356 xpc_make_first_contact(struct xpc_partition *part)
 357 {
 358         enum xpc_retval ret;
 359
 360
 361         while ((ret = xpc_pull_remote_vars_part(part)) != xpcSuccess) {
 362                 if (ret != xpcRetry) {
 363                         XPC_DEACTIVATE_PARTITION(part, ret);
 364                         return ret;
 365                 }
 366
 367                 dev_dbg(xpc_chan, "waiting to make first contact with "
 368                         "partition %d\n", XPC_PARTID(part));
 369
 370                 /* wait a 1/4 of a second or so */
 371                 (void) msleep_interruptible(250);
 372
 373                 if (part->act_state == XPC_P_DEACTIVATING) {
 374                         return part->reason;
 375                 }
 376         }
 377
 378         return xpc_mark_partition_active(part);
 379 }
 380
 381
 382 /*
 383  * The first kthread assigned to a newly activated partition is the one
 384  * created by XPC HB with which it calls xpc_partition_up(). XPC hangs on to
 385  * that kthread until the partition is brought down, at which time that kthread
 386  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
 387  * that XPC has dismantled all communication infrastructure for the associated
 388  * partition.) This kthread becomes the channel manager for that partition.
 389  *
 390  * Each active partition has a channel manager, who, besides connecting and
 391  * disconnecting channels, will ensure that each of the partition's connected
 392  * channels has the required number of assigned kthreads to get the work done.
 393  */
 394 static void
 395 xpc_channel_mgr(struct xpc_partition *part)
 396 {
 397         while (part->act_state != XPC_P_DEACTIVATING ||
 398                         atomic_read(&part->nchannels_active) > 0 ||
 399                                         !xpc_partition_disengaged(part)) {
 400
 401                 xpc_process_channel_activity(part);
 402
 403
 404                 /*
 405                  * Wait until we've been requested to activate kthreads or
 406                  * all of the channel's message queues have been torn down or
 407                  * a signal is pending.
 408                  *
 409                  * The channel_mgr_requests is set to 1 after being awakened,
 410                  * This is done to prevent the channel mgr from making one pass
 411                  * through the loop for each request, since he will
 412                  * be servicing all the requests in one pass. The reason it's
 413                  * set to 1 instead of 0 is so that other kthreads will know
 414                  * that the channel mgr is running and won't bother trying to
 415                  * wake him up.
 416                  */
 417                 atomic_dec(&part->channel_mgr_requests);
 418                 (void) wait_event_interruptible(part->channel_mgr_wq,
 419                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
 420                                 (volatile u64) part->local_IPI_amo != 0 ||
 421                                 ((volatile u8) part->act_state ==
 422                                                         XPC_P_DEACTIVATING &&
 423                                 atomic_read(&part->nchannels_active) == 0 &&
 424                                 xpc_partition_disengaged(part))));
 425                 atomic_set(&part->channel_mgr_requests, 1);
 426
 427                 // >>> Does it need to wakeup periodically as well? In case we
 428                 // >>> miscalculated the #of kthreads to wakeup or create?
 429         }
 430 }
 431
 432
 433 /*
 434  * When XPC HB determines that a partition has come up, it will create a new
 435  * kthread and that kthread will call this function to attempt to set up the
 436  * basic infrastructure used for Cross Partition Communication with the newly
 437  * upped partition.
 438  *
 439  * The kthread that was created by XPC HB and which setup the XPC
 440  * infrastructure will remain assigned to the partition until the partition
 441  * goes down. At which time the kthread will teardown the XPC infrastructure
 442  * and then exit.
 443  *
 444  * XPC HB will put the remote partition's XPC per partition specific variables
 445  * physical address into xpc_partitions[partid].remote_vars_part_pa prior to
 446  * calling xpc_partition_up().
 447  */
 448 static void
 449 xpc_partition_up(struct xpc_partition *part)
 450 {
 451         DBUG_ON(part->channels != NULL);
 452
 453         dev_dbg(xpc_chan, "activating partition %d\n", XPC_PARTID(part));
 454
 455         if (xpc_setup_infrastructure(part) != xpcSuccess) {
 456                 return;
 457         }
 458
 459         /*
 460          * The kthread that XPC HB called us with will become the
 461          * channel manager for this partition. It will not return
 462          * back to XPC HB until the partition's XPC infrastructure
 463          * has been dismantled.
 464          */
 465
 466         (void) xpc_part_ref(part);      /* this will always succeed */
 467
 468         if (xpc_make_first_contact(part) == xpcSuccess) {
 469                 xpc_channel_mgr(part);
 470         }
 471
 472         xpc_part_deref(part);
 473
 474         xpc_teardown_infrastructure(part);
 475 }
 476
 477
 478 static int
 479 xpc_activating(void *__partid)
 480 {
 481         partid_t partid = (u64) __partid;
 482         struct xpc_partition *part = &xpc_partitions[partid];
 483         unsigned long irq_flags;
 484         struct sched_param param = { sched_priority: MAX_RT_PRIO - 1 };
 485         int ret;
 486
 487
 488         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
 489
 490         spin_lock_irqsave(&part->act_lock, irq_flags);
 491
 492         if (part->act_state == XPC_P_DEACTIVATING) {
 493                 part->act_state = XPC_P_INACTIVE;
 494                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 495                 part->remote_rp_pa = 0;
 496                 return 0;
 497         }
 498
 499         /* indicate the thread is activating */
 500         DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
 501         part->act_state = XPC_P_ACTIVATING;
 502
 503         XPC_SET_REASON(part, 0, 0);
 504         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 505
 506         dev_dbg(xpc_part, "bringing partition %d up\n", partid);
 507
 508         daemonize("xpc%02d", partid);
 509
 510         /*
 511          * This thread needs to run at a realtime priority to prevent a
 512          * significant performance degradation.
 513          */
 514         ret = sched_setscheduler(current, SCHED_FIFO, &param);
 515         if (ret != 0) {
 516                 dev_warn(xpc_part, "unable to set pid %d to a realtime "
 517                         "priority, ret=%d\n", current->pid, ret);
 518         }
 519
 520         /* allow this thread and its children to run on any CPU */
 521         set_cpus_allowed(current, CPU_MASK_ALL);
 522
 523         /*
 524          * Register the remote partition's AMOs with SAL so it can handle
 525          * and cleanup errors within that address range should the remote
 526          * partition go down. We don't unregister this range because it is
 527          * difficult to tell when outstanding writes to the remote partition
 528          * are finished and thus when it is safe to unregister. This should
 529          * not result in wasted space in the SAL xp_addr_region table because
 530          * we should get the same page for remote_amos_page_pa after module
 531          * reloads and system reboots.
 532          */
 533         if (sn_register_xp_addr_region(part->remote_amos_page_pa,
 534                                                         PAGE_SIZE, 1) < 0) {
 535                 dev_warn(xpc_part, "xpc_partition_up(%d) failed to register "
 536                         "xp_addr region\n", partid);
 537
 538                 spin_lock_irqsave(&part->act_lock, irq_flags);
 539                 part->act_state = XPC_P_INACTIVE;
 540                 XPC_SET_REASON(part, xpcPhysAddrRegFailed, __LINE__);
 541                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 542                 part->remote_rp_pa = 0;
 543                 return 0;
 544         }
 545
 546         xpc_allow_hb(partid, xpc_vars);
 547         xpc_IPI_send_activated(part);
 548
 549
 550         /*
 551          * xpc_partition_up() holds this thread and marks this partition as
 552          * XPC_P_ACTIVE by calling xpc_hb_mark_active().
 553          */
 554         (void) xpc_partition_up(part);
 555
 556         xpc_disallow_hb(partid, xpc_vars);
 557         xpc_mark_partition_inactive(part);
 558
 559         if (part->reason == xpcReactivating) {
 560                 /* interrupting ourselves results in activating partition */
 561                 xpc_IPI_send_reactivate(part);
 562         }
 563
 564         return 0;
 565 }
 566
 567
 568 void
 569 xpc_activate_partition(struct xpc_partition *part)
 570 {
 571         partid_t partid = XPC_PARTID(part);
 572         unsigned long irq_flags;
 573         pid_t pid;
 574
 575
 576         spin_lock_irqsave(&part->act_lock, irq_flags);
 577
 578         DBUG_ON(part->act_state != XPC_P_INACTIVE);
 579
 580         part->act_state = XPC_P_ACTIVATION_REQ;
 581         XPC_SET_REASON(part, xpcCloneKThread, __LINE__);
 582
 583         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 584
 585         pid = kernel_thread(xpc_activating, (void *) ((u64) partid), 0);
 586
 587         if (unlikely(pid <= 0)) {
 588                 spin_lock_irqsave(&part->act_lock, irq_flags);
 589                 part->act_state = XPC_P_INACTIVE;
 590                 XPC_SET_REASON(part, xpcCloneKThreadFailed, __LINE__);
 591                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 592         }
 593 }
 594
 595
 596 /*
 597  * Handle the receipt of a SGI_XPC_NOTIFY IRQ by seeing whether the specified
 598  * partition actually sent it. Since SGI_XPC_NOTIFY IRQs may be shared by more
 599  * than one partition, we use an AMO_t structure per partition to indicate
 600  * whether a partition has sent an IPI or not.  >>> If it has, then wake up the
 601  * associated kthread to handle it.
 602  *
 603  * All SGI_XPC_NOTIFY IRQs received by XPC are the result of IPIs sent by XPC
 604  * running on other partitions.
 605  *
 606  * Noteworthy Arguments:
 607  *
 608  *      irq - Interrupt ReQuest number. NOT USED.
 609  *
 610  *      dev_id - partid of IPI's potential sender.
 611  *
 612  *      regs - processor's context before the processor entered
 613  *             interrupt code. NOT USED.
 614  */
 615 irqreturn_t
 616 xpc_notify_IRQ_handler(int irq, void *dev_id, struct pt_regs *regs)
 617 {
 618         partid_t partid = (partid_t) (u64) dev_id;
 619         struct xpc_partition *part = &xpc_partitions[partid];
 620
 621
 622         DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
 623
 624         if (xpc_part_ref(part)) {
 625                 xpc_check_for_channel_activity(part);
 626
 627                 xpc_part_deref(part);
 628         }
 629         return IRQ_HANDLED;
 630 }
 631
 632
 633 /*
 634  * Check to see if xpc_notify_IRQ_handler() dropped any IPIs on the floor
 635  * because the write to their associated IPI amo completed after the IRQ/IPI
 636  * was received.
 637  */
 638 void
 639 xpc_dropped_IPI_check(struct xpc_partition *part)
 640 {
 641         if (xpc_part_ref(part)) {
 642                 xpc_check_for_channel_activity(part);
 643
 644                 part->dropped_IPI_timer.expires = jiffies +
 645                                                         XPC_P_DROPPED_IPI_WAIT;
 646                 add_timer(&part->dropped_IPI_timer);
 647                 xpc_part_deref(part);
 648         }
 649 }
 650
 651
 652 void
 653 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
 654 {
 655         int idle = atomic_read(&ch->kthreads_idle);
 656         int assigned = atomic_read(&ch->kthreads_assigned);
 657         int wakeup;
 658
 659
 660         DBUG_ON(needed <= 0);
 661
 662         if (idle > 0) {
 663                 wakeup = (needed > idle) ? idle : needed;
 664                 needed -= wakeup;
 665
 666                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
 667                         "channel=%d\n", wakeup, ch->partid, ch->number);
 668
 669                 /* only wakeup the requested number of kthreads */
 670                 wake_up_nr(&ch->idle_wq, wakeup);
 671         }
 672
 673         if (needed <= 0) {
 674                 return;
 675         }
 676
 677         if (needed + assigned > ch->kthreads_assigned_limit) {
 678                 needed = ch->kthreads_assigned_limit - assigned;
 679                 // >>>should never be less than 0
 680                 if (needed <= 0) {
 681                         return;
 682                 }
 683         }
 684
 685         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
 686                 needed, ch->partid, ch->number);
 687
 688         xpc_create_kthreads(ch, needed);
 689 }
 690
 691
 692 /*
 693  * This function is where XPC's kthreads wait for messages to deliver.
 694  */
 695 static void
 696 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
 697 {
 698         do {
 699                 /* deliver messages to their intended recipients */
 700
 701                 while ((volatile s64) ch->w_local_GP.get <
 702                                 (volatile s64) ch->w_remote_GP.put &&
 703                                         !((volatile u32) ch->flags &
 704                                                 XPC_C_DISCONNECTING)) {
 705                         xpc_deliver_msg(ch);
 706                 }
 707
 708                 if (atomic_inc_return(&ch->kthreads_idle) >
 709                                                 ch->kthreads_idle_limit) {
 710                         /* too many idle kthreads on this channel */
 711                         atomic_dec(&ch->kthreads_idle);
 712                         break;
 713                 }
 714
 715                 dev_dbg(xpc_chan, "idle kthread calling "
 716                         "wait_event_interruptible_exclusive()\n");
 717
 718                 (void) wait_event_interruptible_exclusive(ch->idle_wq,
 719                                 ((volatile s64) ch->w_local_GP.get <
 720                                         (volatile s64) ch->w_remote_GP.put ||
 721                                 ((volatile u32) ch->flags &
 722                                                 XPC_C_DISCONNECTING)));
 723
 724                 atomic_dec(&ch->kthreads_idle);
 725
 726         } while (!((volatile u32) ch->flags & XPC_C_DISCONNECTING));
 727 }
 728
 729
 730 static int
 731 xpc_daemonize_kthread(void *args)
 732 {
 733         partid_t partid = XPC_UNPACK_ARG1(args);
 734         u16 ch_number = XPC_UNPACK_ARG2(args);
 735         struct xpc_partition *part = &xpc_partitions[partid];
 736         struct xpc_channel *ch;
 737         int n_needed;
 738         unsigned long irq_flags;
 739
 740
 741         daemonize("xpc%02dc%d", partid, ch_number);
 742
 743         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
 744                 partid, ch_number);
 745
 746         ch = &part->channels[ch_number];
 747
 748         if (!(ch->flags & XPC_C_DISCONNECTING)) {
 749
 750                 /* let registerer know that connection has been established */
 751
 752                 spin_lock_irqsave(&ch->lock, irq_flags);
 753                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
 754                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
 755                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 756
 757                         xpc_connected_callout(ch);
 758
 759                         spin_lock_irqsave(&ch->lock, irq_flags);
 760                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
 761                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 762
 763                         /*
 764                          * It is possible that while the callout was being
 765                          * made that the remote partition sent some messages.
 766                          * If that is the case, we may need to activate
 767                          * additional kthreads to help deliver them. We only
 768                          * need one less than total #of messages to deliver.
 769                          */
 770                         n_needed = ch->w_remote_GP.put - ch->w_local_GP.get - 1;
 771                         if (n_needed > 0 &&
 772                                         !(ch->flags & XPC_C_DISCONNECTING)) {
 773                                 xpc_activate_kthreads(ch, n_needed);
 774                         }
 775                 } else {
 776                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 777                 }
 778
 779                 xpc_kthread_waitmsgs(part, ch);
 780         }
 781
 782         if (atomic_dec_return(&ch->kthreads_assigned) == 0) {
 783                 spin_lock_irqsave(&ch->lock, irq_flags);
 784                 if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
 785                                 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
 786                         ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
 787                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 788
 789                         xpc_disconnect_callout(ch, xpcDisconnecting);
 790
 791                         spin_lock_irqsave(&ch->lock, irq_flags);
 792                         ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
 793                 }
 794                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 795                 if (atomic_dec_return(&part->nchannels_engaged) == 0) {
 796                         xpc_mark_partition_disengaged(part);
 797                         xpc_IPI_send_disengage(part);
 798                 }
 799         }
 800
 801
 802         xpc_msgqueue_deref(ch);
 803
 804         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
 805                 partid, ch_number);
 806
 807         xpc_part_deref(part);
 808         return 0;
 809 }
 810
 811
 812 /*
 813  * For each partition that XPC has established communications with, there is
 814  * a minimum of one kernel thread assigned to perform any operation that
 815  * may potentially sleep or block (basically the callouts to the asynchronous
 816  * functions registered via xpc_connect()).
 817  *
 818  * Additional kthreads are created and destroyed by XPC as the workload
 819  * demands.
 820  *
 821  * A kthread is assigned to one of the active channels that exists for a given
 822  * partition.
 823  */
 824 void
 825 xpc_create_kthreads(struct xpc_channel *ch, int needed)
 826 {
 827         unsigned long irq_flags;
 828         pid_t pid;
 829         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
 830         struct xpc_partition *part = &xpc_partitions[ch->partid];
 831
 832
 833         while (needed-- > 0) {
 834
 835                 /*
 836                  * The following is done on behalf of the newly created
 837                  * kthread. That kthread is responsible for doing the
 838                  * counterpart to the following before it exits.
 839                  */
 840                 (void) xpc_part_ref(part);
 841                 xpc_msgqueue_ref(ch);
 842                 if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
 843                     atomic_inc_return(&part->nchannels_engaged) == 1) {
 844                         xpc_mark_partition_engaged(part);
 845                 }
 846
 847                 pid = kernel_thread(xpc_daemonize_kthread, (void *) args, 0);
 848                 if (pid < 0) {
 849                         /* the fork failed */
 850                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 851                             atomic_dec_return(&part->nchannels_engaged) == 0) {
 852                                 xpc_mark_partition_disengaged(part);
 853                                 xpc_IPI_send_disengage(part);
 854                         }
 855                         xpc_msgqueue_deref(ch);
 856                         xpc_part_deref(part);
 857
 858                         if (atomic_read(&ch->kthreads_assigned) <
 859                                                 ch->kthreads_idle_limit) {
 860                                 /*
 861                                  * Flag this as an error only if we have an
 862                                  * insufficient #of kthreads for the channel
 863                                  * to function.
 864                                  *
 865                                  * No xpc_msgqueue_ref() is needed here since
 866                                  * the channel mgr is doing this.
 867                                  */
 868                                 spin_lock_irqsave(&ch->lock, irq_flags);
 869                                 XPC_DISCONNECT_CHANNEL(ch, xpcLackOfResources,
 870                                                                 &irq_flags);
 871                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 872                         }
 873                         break;
 874                 }
 875
 876                 ch->kthreads_created++; // >>> temporary debug only!!!
 877         }
 878 }
 879
 880
 881 void
 882 xpc_disconnect_wait(int ch_number)
 883 {
 884         unsigned long irq_flags;
 885         partid_t partid;
 886         struct xpc_partition *part;
 887         struct xpc_channel *ch;
 888         int wakeup_channel_mgr;
 889
 890
 891         /* now wait for all callouts to the caller's function to cease */
 892         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
 893                 part = &xpc_partitions[partid];
 894
 895                 if (!xpc_part_ref(part)) {
 896                         continue;
 897                 }
 898
 899                 ch = &part->channels[ch_number];
 900
 901                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
 902                         xpc_part_deref(part);
 903                         continue;
 904                 }
 905
 906                 wait_for_completion(&ch->wdisconnect_wait);
 907
 908                 spin_lock_irqsave(&ch->lock, irq_flags);
 909                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
 910                 wakeup_channel_mgr = 0;
 911
 912                 if (ch->delayed_IPI_flags) {
 913                         if (part->act_state != XPC_P_DEACTIVATING) {
 914                                 spin_lock(&part->IPI_lock);
 915                                 XPC_SET_IPI_FLAGS(part->local_IPI_amo,
 916                                         ch->number, ch->delayed_IPI_flags);
 917                                 spin_unlock(&part->IPI_lock);
 918                                 wakeup_channel_mgr = 1;
 919                         }
 920                         ch->delayed_IPI_flags = 0;
 921                 }
 922
 923                 ch->flags &= ~XPC_C_WDISCONNECT;
 924                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 925
 926                 if (wakeup_channel_mgr) {
 927                         xpc_wakeup_channel_mgr(part);
 928                 }
 929
 930                 xpc_part_deref(part);
 931         }
 932 }
 933
 934
 935 static void
 936 xpc_do_exit(enum xpc_retval reason)
 937 {
 938         partid_t partid;
 939         int active_part_count, printed_waiting_msg = 0;
 940         struct xpc_partition *part;
 941         unsigned long printmsg_time, disengage_request_timeout = 0;
 942
 943
 944         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
 945         DBUG_ON(xpc_exiting == 1);
 946
 947         /*
 948          * Let the heartbeat checker thread and the discovery thread
 949          * (if one is running) know that they should exit. Also wake up
 950          * the heartbeat checker thread in case it's sleeping.
 951          */
 952         xpc_exiting = 1;
 953         wake_up_interruptible(&xpc_act_IRQ_wq);
 954
 955         /* ignore all incoming interrupts */
 956         free_irq(SGI_XPC_ACTIVATE, NULL);
 957
 958         /* wait for the discovery thread to exit */
 959         wait_for_completion(&xpc_discovery_exited);
 960
 961         /* wait for the heartbeat checker thread to exit */
 962         wait_for_completion(&xpc_hb_checker_exited);
 963
 964
 965         /* sleep for a 1/3 of a second or so */
 966         (void) msleep_interruptible(300);
 967
 968
 969         /* wait for all partitions to become inactive */
 970
 971         printmsg_time = jiffies + (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
 972         xpc_disengage_request_timedout = 0;
 973
 974         do {
 975                 active_part_count = 0;
 976
 977                 for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
 978                         part = &xpc_partitions[partid];
 979
 980                         if (xpc_partition_disengaged(part) &&
 981                                         part->act_state == XPC_P_INACTIVE) {
 982                                 continue;
 983                         }
 984
 985                         active_part_count++;
 986
 987                         XPC_DEACTIVATE_PARTITION(part, reason);
 988
 989                         if (part->disengage_request_timeout >
 990                                                 disengage_request_timeout) {
 991                                 disengage_request_timeout =
 992                                                 part->disengage_request_timeout;
 993                         }
 994                 }
 995
 996                 if (xpc_partition_engaged(-1UL)) {
 997                         if (time_after(jiffies, printmsg_time)) {
 998                                 dev_info(xpc_part, "waiting for remote "
 999                                         "partitions to disengage, timeout in "
1000                                         "%ld seconds\n",
1001                                         (disengage_request_timeout - jiffies)
1002                                                                         / HZ);
1003                                 printmsg_time = jiffies +
1004                                         (XPC_DISENGAGE_PRINTMSG_INTERVAL * HZ);
1005                                 printed_waiting_msg = 1;
1006                         }
1007
1008                 } else if (active_part_count > 0) {
1009                         if (printed_waiting_msg) {
1010                                 dev_info(xpc_part, "waiting for local partition"
1011                                         " to disengage\n");
1012                                 printed_waiting_msg = 0;
1013                         }
1014
1015                 } else {
1016                         if (!xpc_disengage_request_timedout) {
1017                                 dev_info(xpc_part, "all partitions have "
1018                                         "disengaged\n");
1019                         }
1020                         break;
1021                 }
1022
1023                 /* sleep for a 1/3 of a second or so */
1024                 (void) msleep_interruptible(300);
1025
1026         } while (1);
1027
1028         DBUG_ON(xpc_partition_engaged(-1UL));
1029
1030
1031         /* indicate to others that our reserved page is uninitialized */
1032         xpc_rsvd_page->vars_pa = 0;
1033
1034         /* now it's time to eliminate our heartbeat */
1035         del_timer_sync(&xpc_hb_timer);
1036         DBUG_ON(xpc_vars->heartbeating_to_mask != 0);
1037
1038         if (reason == xpcUnloading) {
1039                 /* take ourselves off of the reboot_notifier_list */
1040                 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1041
1042                 /* take ourselves off of the die_notifier list */
1043                 (void) unregister_die_notifier(&xpc_die_notifier);
1044         }
1045
1046         /* close down protections for IPI operations */
1047         xpc_restrict_IPI_ops();
1048
1049
1050         /* clear the interface to XPC's functions */
1051         xpc_clear_interface();
1052
1053         if (xpc_sysctl) {
1054                 unregister_sysctl_table(xpc_sysctl);
1055         }
1056 }
1057
1058
1059 /*
1060  * This function is called when the system is being rebooted.
1061  */
1062 static int
1063 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
1064 {
1065         enum xpc_retval reason;
1066
1067
1068         switch (event) {
1069         case SYS_RESTART:
1070                 reason = xpcSystemReboot;
1071                 break;
1072         case SYS_HALT:
1073                 reason = xpcSystemHalt;
1074                 break;
1075         case SYS_POWER_OFF:
1076                 reason = xpcSystemPoweroff;
1077                 break;
1078         default:
1079                 reason = xpcSystemGoingDown;
1080         }
1081
1082         xpc_do_exit(reason);
1083         return NOTIFY_DONE;
1084 }
1085
1086
1087 /*
1088  * Notify other partitions to disengage from all references to our memory.
1089  */
1090 static void
1091 xpc_die_disengage(void)
1092 {
1093         struct xpc_partition *part;
1094         partid_t partid;
1095         unsigned long engaged;
1096         long time, printmsg_time, disengage_request_timeout;
1097
1098
1099         /* keep xpc_hb_checker thread from doing anything (just in case) */
1100         xpc_exiting = 1;
1101
1102         xpc_vars->heartbeating_to_mask = 0;  /* indicate we're deactivated */
1103
1104         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1105                 part = &xpc_partitions[partid];
1106
1107                 if (!XPC_SUPPORTS_DISENGAGE_REQUEST(part->
1108                                                         remote_vars_version)) {
1109
1110                         /* just in case it was left set by an earlier XPC */
1111                         xpc_clear_partition_engaged(1UL << partid);
1112                         continue;
1113                 }
1114
1115                 if (xpc_partition_engaged(1UL << partid) ||
1116                                         part->act_state != XPC_P_INACTIVE) {
1117                         xpc_request_partition_disengage(part);
1118                         xpc_mark_partition_disengaged(part);
1119                         xpc_IPI_send_disengage(part);
1120                 }
1121         }
1122
1123         time = rtc_time();
1124         printmsg_time = time +
1125                 (XPC_DISENGAGE_PRINTMSG_INTERVAL * sn_rtc_cycles_per_second);
1126         disengage_request_timeout = time +
1127                 (xpc_disengage_request_timelimit * sn_rtc_cycles_per_second);
1128
1129         /* wait for all other partitions to disengage from us */
1130
1131         while (1) {
1132                 engaged = xpc_partition_engaged(-1UL);
1133                 if (!engaged) {
1134                         dev_info(xpc_part, "all partitions have disengaged\n");
1135                         break;
1136                 }
1137
1138                 time = rtc_time();
1139                 if (time >= disengage_request_timeout) {
1140                         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1141                                 if (engaged & (1UL << partid)) {
1142                                         dev_info(xpc_part, "disengage from "
1143                                                 "remote partition %d timed "
1144                                                 "out\n", partid);
1145                                 }
1146                         }
1147                         break;
1148                 }
1149
1150                 if (time >= printmsg_time) {
1151                         dev_info(xpc_part, "waiting for remote partitions to "
1152                                 "disengage, timeout in %ld seconds\n",
1153                                 (disengage_request_timeout - time) /
1154                                                 sn_rtc_cycles_per_second);
1155                         printmsg_time = time +
1156                                         (XPC_DISENGAGE_PRINTMSG_INTERVAL *
1157                                                 sn_rtc_cycles_per_second);
1158                 }
1159         }
1160 }
1161
1162
1163 /*
1164  * This function is called when the system is being restarted or halted due
1165  * to some sort of system failure. If this is the case we need to notify the
1166  * other partitions to disengage from all references to our memory.
1167  * This function can also be called when our heartbeater could be offlined
1168  * for a time. In this case we need to notify other partitions to not worry
1169  * about the lack of a heartbeat.
1170  */
1171 static int
1172 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1173 {
1174         switch (event) {
1175         case DIE_MACHINE_RESTART:
1176         case DIE_MACHINE_HALT:
1177                 xpc_die_disengage();
1178                 break;
1179
1180         case DIE_KDEBUG_ENTER:
1181                 /* Should lack of heartbeat be ignored by other partitions? */
1182                 if (!xpc_kdebug_ignore) {
1183                         break;
1184                 }
1185                 /* fall through */
1186         case DIE_MCA_MONARCH_ENTER:
1187         case DIE_INIT_MONARCH_ENTER:
1188                 xpc_vars->heartbeat++;
1189                 xpc_vars->heartbeat_offline = 1;
1190                 break;
1191
1192         case DIE_KDEBUG_LEAVE:
1193                 /* Is lack of heartbeat being ignored by other partitions? */
1194                 if (!xpc_kdebug_ignore) {
1195                         break;
1196                 }
1197                 /* fall through */
1198         case DIE_MCA_MONARCH_LEAVE:
1199         case DIE_INIT_MONARCH_LEAVE:
1200                 xpc_vars->heartbeat++;
1201                 xpc_vars->heartbeat_offline = 0;
1202                 break;
1203         }
1204
1205         return NOTIFY_DONE;
1206 }
1207
1208
1209 int __init
1210 xpc_init(void)
1211 {
1212         int ret;
1213         partid_t partid;
1214         struct xpc_partition *part;
1215         pid_t pid;
1216
1217
1218         if (!ia64_platform_is("sn2")) {
1219                 return -ENODEV;
1220         }
1221
1222         /*
1223          * xpc_remote_copy_buffer is used as a temporary buffer for bte_copy'ng
1224          * various portions of a partition's reserved page. Its size is based
1225          * on the size of the reserved page header and part_nasids mask. So we
1226          * need to ensure that the other items will fit as well.
1227          */
1228         if (XPC_RP_VARS_SIZE > XPC_RP_HEADER_SIZE + XP_NASID_MASK_BYTES) {
1229                 dev_err(xpc_part, "xpc_remote_copy_buffer is not big enough\n");
1230                 return -EPERM;
1231         }
1232         DBUG_ON((u64) xpc_remote_copy_buffer !=
1233                                 L1_CACHE_ALIGN((u64) xpc_remote_copy_buffer));
1234
1235         snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1236         snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1237
1238         xpc_sysctl = register_sysctl_table(xpc_sys_dir, 1);
1239
1240         /*
1241          * The first few fields of each entry of xpc_partitions[] need to
1242          * be initialized now so that calls to xpc_connect() and
1243          * xpc_disconnect() can be made prior to the activation of any remote
1244          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1245          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1246          * PARTITION HAS BEEN ACTIVATED.
1247          */
1248         for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
1249                 part = &xpc_partitions[partid];
1250
1251                 DBUG_ON((u64) part != L1_CACHE_ALIGN((u64) part));
1252
1253                 part->act_IRQ_rcvd = 0;
1254                 spin_lock_init(&part->act_lock);
1255                 part->act_state = XPC_P_INACTIVE;
1256                 XPC_SET_REASON(part, 0, 0);
1257
1258                 init_timer(&part->disengage_request_timer);
1259                 part->disengage_request_timer.function =
1260                                 xpc_timeout_partition_disengage_request;
1261                 part->disengage_request_timer.data = (unsigned long) part;
1262
1263                 part->setup_state = XPC_P_UNSET;
1264                 init_waitqueue_head(&part->teardown_wq);
1265                 atomic_set(&part->references, 0);
1266         }
1267
1268         /*
1269          * Open up protections for IPI operations (and AMO operations on
1270          * Shub 1.1 systems).
1271          */
1272         xpc_allow_IPI_ops();
1273
1274         /*
1275          * Interrupts being processed will increment this atomic variable and
1276          * awaken the heartbeat thread which will process the interrupts.
1277          */
1278         atomic_set(&xpc_act_IRQ_rcvd, 0);
1279
1280         /*
1281          * This is safe to do before the xpc_hb_checker thread has started
1282          * because the handler releases a wait queue.  If an interrupt is
1283          * received before the thread is waiting, it will not go to sleep,
1284          * but rather immediately process the interrupt.
1285          */
1286         ret = request_irq(SGI_XPC_ACTIVATE, xpc_act_IRQ_handler, 0,
1287                                                         "xpc hb", NULL);
1288         if (ret != 0) {
1289                 dev_err(xpc_part, "can't register ACTIVATE IRQ handler, "
1290                         "errno=%d\n", -ret);
1291
1292                 xpc_restrict_IPI_ops();
1293
1294                 if (xpc_sysctl) {
1295                         unregister_sysctl_table(xpc_sysctl);
1296                 }
1297                 return -EBUSY;
1298         }
1299
1300         /*
1301          * Fill the partition reserved page with the information needed by
1302          * other partitions to discover we are alive and establish initial
1303          * communications.
1304          */
1305         xpc_rsvd_page = xpc_rsvd_page_init();
1306         if (xpc_rsvd_page == NULL) {
1307                 dev_err(xpc_part, "could not setup our reserved page\n");
1308
1309                 free_irq(SGI_XPC_ACTIVATE, NULL);
1310                 xpc_restrict_IPI_ops();
1311
1312                 if (xpc_sysctl) {
1313                         unregister_sysctl_table(xpc_sysctl);
1314                 }
1315                 return -EBUSY;
1316         }
1317
1318
1319         /* add ourselves to the reboot_notifier_list */
1320         ret = register_reboot_notifier(&xpc_reboot_notifier);
1321         if (ret != 0) {
1322                 dev_warn(xpc_part, "can't register reboot notifier\n");
1323         }
1324
1325         /* add ourselves to the die_notifier list (i.e., ia64die_chain) */
1326         ret = register_die_notifier(&xpc_die_notifier);
1327         if (ret != 0) {
1328                 dev_warn(xpc_part, "can't register die notifier\n");
1329         }
1330
1331
1332         /*
1333          * Set the beating to other partitions into motion.  This is
1334          * the last requirement for other partitions' discovery to
1335          * initiate communications with us.
1336          */
1337         init_timer(&xpc_hb_timer);
1338         xpc_hb_timer.function = xpc_hb_beater;
1339         xpc_hb_beater(0);
1340
1341
1342         /*
1343          * The real work-horse behind xpc.  This processes incoming
1344          * interrupts and monitors remote heartbeats.
1345          */
1346         pid = kernel_thread(xpc_hb_checker, NULL, 0);
1347         if (pid < 0) {
1348                 dev_err(xpc_part, "failed while forking hb check thread\n");
1349
1350                 /* indicate to others that our reserved page is uninitialized */
1351                 xpc_rsvd_page->vars_pa = 0;
1352
1353                 /* take ourselves off of the reboot_notifier_list */
1354                 (void) unregister_reboot_notifier(&xpc_reboot_notifier);
1355
1356                 /* take ourselves off of the die_notifier list */
1357                 (void) unregister_die_notifier(&xpc_die_notifier);
1358
1359                 del_timer_sync(&xpc_hb_timer);
1360                 free_irq(SGI_XPC_ACTIVATE, NULL);
1361                 xpc_restrict_IPI_ops();
1362
1363                 if (xpc_sysctl) {
1364                         unregister_sysctl_table(xpc_sysctl);
1365                 }
1366                 return -EBUSY;
1367         }
1368
1369
1370         /*
1371          * Startup a thread that will attempt to discover other partitions to
1372          * activate based on info provided by SAL. This new thread is short
1373          * lived and will exit once discovery is complete.
1374          */
1375         pid = kernel_thread(xpc_initiate_discovery, NULL, 0);
1376         if (pid < 0) {
1377                 dev_err(xpc_part, "failed while forking discovery thread\n");
1378
1379                 /* mark this new thread as a non-starter */
1380                 complete(&xpc_discovery_exited);
1381
1382                 xpc_do_exit(xpcUnloading);
1383                 return -EBUSY;
1384         }
1385
1386
1387         /* set the interface to point at XPC's functions */
1388         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1389                           xpc_initiate_allocate, xpc_initiate_send,
1390                           xpc_initiate_send_notify, xpc_initiate_received,
1391                           xpc_initiate_partid_to_nasids);
1392
1393         return 0;
1394 }
1395 module_init(xpc_init);
1396
1397
1398 void __exit
1399 xpc_exit(void)
1400 {
1401         xpc_do_exit(xpcUnloading);
1402 }
1403 module_exit(xpc_exit);
1404
1405
1406 MODULE_AUTHOR("Silicon Graphics, Inc.");
1407 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1408 MODULE_LICENSE("GPL");
1409
1410 module_param(xpc_hb_interval, int, 0);
1411 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1412                 "heartbeat increments.");
1413
1414 module_param(xpc_hb_check_interval, int, 0);
1415 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1416                 "heartbeat checks.");
1417
1418 module_param(xpc_disengage_request_timelimit, int, 0);
1419 MODULE_PARM_DESC(xpc_disengage_request_timelimit, "Number of seconds to wait "
1420                 "for disengage request to complete.");
1421
1422 module_param(xpc_kdebug_ignore, int, 0);
1423 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1424                 "other partitions when dropping into kdebug.");
1425