drivers/misc/sgi-xp/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) 2004-2008 Silicon Graphics, Inc.  All Rights Reserved.
   7  */
   8
   9 /*
  10  * Cross Partition Communication (XPC) support - standard version.
  11  *
  12  *      XPC provides a message passing capability that crosses partition
  13  *      boundaries. This module is made up of two parts:
  14  *
  15  *          partition   This part detects the presence/absence of other
  16  *                      partitions. It provides a heartbeat and monitors
  17  *                      the heartbeats of other partitions.
  18  *
  19  *          channel     This part manages the channels and sends/receives
  20  *                      messages across them to/from other partitions.
  21  *
  22  *      There are a couple of additional functions residing in XP, which
  23  *      provide an interface to XPC for its users.
  24  *
  25  *
  26  *      Caveats:
  27  *
  28  *        . Currently on sn2, we have no way to determine which nasid an IRQ
  29  *          came from. Thus, xpc_send_IRQ_sn2() does a remote amo write
  30  *          followed by an IPI. The amo indicates where data is to be pulled
  31  *          from, so after the IPI arrives, the remote partition checks the amo
  32  *          word. The IPI can actually arrive before the amo however, so other
  33  *          code must periodically check for this case. Also, remote amo
  34  *          operations do not reliably time out. Thus we do a remote PIO read
  35  *          solely to know whether the remote partition is down and whether we
  36  *          should stop sending IPIs to it. This remote PIO read operation is
  37  *          set up in a special nofault region so SAL knows to ignore (and
  38  *          cleanup) any errors due to the remote amo write, PIO read, and/or
  39  *          PIO write operations.
  40  *
  41  *          If/when new hardware solves this IPI problem, we should abandon
  42  *          the current approach.
  43  *
  44  */
  45
  46 #include <linux/module.h>
  47 #include <linux/sysctl.h>
  48 #include <linux/device.h>
  49 #include <linux/delay.h>
  50 #include <linux/reboot.h>
  51 #include <linux/kdebug.h>
  52 #include <linux/kthread.h>
  53 #include "xpc.h"
  54
  55 /* define two XPC debug device structures to be used with dev_dbg() et al */
  56
  57 struct device_driver xpc_dbg_name = {
  58         .name = "xpc"
  59 };
  60
  61 struct device xpc_part_dbg_subname = {
  62         .bus_id = {0},          /* set to "part" at xpc_init() time */
  63         .driver = &xpc_dbg_name
  64 };
  65
  66 struct device xpc_chan_dbg_subname = {
  67         .bus_id = {0},          /* set to "chan" at xpc_init() time */
  68         .driver = &xpc_dbg_name
  69 };
  70
  71 struct device *xpc_part = &xpc_part_dbg_subname;
  72 struct device *xpc_chan = &xpc_chan_dbg_subname;
  73
  74 static int xpc_kdebug_ignore;
  75
  76 /* systune related variables for /proc/sys directories */
  77
  78 static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
  79 static int xpc_hb_min_interval = 1;
  80 static int xpc_hb_max_interval = 10;
  81
  82 static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL;
  83 static int xpc_hb_check_min_interval = 10;
  84 static int xpc_hb_check_max_interval = 120;
  85
  86 int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT;
  87 static int xpc_disengage_min_timelimit; /* = 0 */
  88 static int xpc_disengage_max_timelimit = 120;
  89
  90 static ctl_table xpc_sys_xpc_hb_dir[] = {
  91         {
  92          .ctl_name = CTL_UNNUMBERED,
  93          .procname = "hb_interval",
  94          .data = &xpc_hb_interval,
  95          .maxlen = sizeof(int),
  96          .mode = 0644,
  97          .proc_handler = &proc_dointvec_minmax,
  98          .strategy = &sysctl_intvec,
  99          .extra1 = &xpc_hb_min_interval,
 100          .extra2 = &xpc_hb_max_interval},
 101         {
 102          .ctl_name = CTL_UNNUMBERED,
 103          .procname = "hb_check_interval",
 104          .data = &xpc_hb_check_interval,
 105          .maxlen = sizeof(int),
 106          .mode = 0644,
 107          .proc_handler = &proc_dointvec_minmax,
 108          .strategy = &sysctl_intvec,
 109          .extra1 = &xpc_hb_check_min_interval,
 110          .extra2 = &xpc_hb_check_max_interval},
 111         {}
 112 };
 113 static ctl_table xpc_sys_xpc_dir[] = {
 114         {
 115          .ctl_name = CTL_UNNUMBERED,
 116          .procname = "hb",
 117          .mode = 0555,
 118          .child = xpc_sys_xpc_hb_dir},
 119         {
 120          .ctl_name = CTL_UNNUMBERED,
 121          .procname = "disengage_timelimit",
 122          .data = &xpc_disengage_timelimit,
 123          .maxlen = sizeof(int),
 124          .mode = 0644,
 125          .proc_handler = &proc_dointvec_minmax,
 126          .strategy = &sysctl_intvec,
 127          .extra1 = &xpc_disengage_min_timelimit,
 128          .extra2 = &xpc_disengage_max_timelimit},
 129         {}
 130 };
 131 static ctl_table xpc_sys_dir[] = {
 132         {
 133          .ctl_name = CTL_UNNUMBERED,
 134          .procname = "xpc",
 135          .mode = 0555,
 136          .child = xpc_sys_xpc_dir},
 137         {}
 138 };
 139 static struct ctl_table_header *xpc_sysctl;
 140
 141 /* non-zero if any remote partition disengage was timed out */
 142 int xpc_disengage_timedout;
 143
 144 /* #of activate IRQs received */
 145 atomic_t xpc_activate_IRQ_rcvd = ATOMIC_INIT(0);
 146
 147 /* IRQ handler notifies this wait queue on receipt of an IRQ */
 148 DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq);
 149
 150 static unsigned long xpc_hb_check_timeout;
 151 static struct timer_list xpc_hb_timer;
 152 void *xpc_heartbeating_to_mask;
 153
 154 /* notification that the xpc_hb_checker thread has exited */
 155 static DECLARE_COMPLETION(xpc_hb_checker_exited);
 156
 157 /* notification that the xpc_discovery thread has exited */
 158 static DECLARE_COMPLETION(xpc_discovery_exited);
 159
 160 static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *);
 161
 162 static int xpc_system_reboot(struct notifier_block *, unsigned long, void *);
 163 static struct notifier_block xpc_reboot_notifier = {
 164         .notifier_call = xpc_system_reboot,
 165 };
 166
 167 static int xpc_system_die(struct notifier_block *, unsigned long, void *);
 168 static struct notifier_block xpc_die_notifier = {
 169         .notifier_call = xpc_system_die,
 170 };
 171
 172 enum xp_retval (*xpc_get_partition_rsvd_page_pa) (u64 buf, u64 *cookie,
 173                                                   u64 *paddr, size_t *len);
 174 enum xp_retval (*xpc_rsvd_page_init) (struct xpc_rsvd_page *rp);
 175 void (*xpc_heartbeat_init) (void);
 176 void (*xpc_heartbeat_exit) (void);
 177 void (*xpc_increment_heartbeat) (void);
 178 void (*xpc_offline_heartbeat) (void);
 179 void (*xpc_online_heartbeat) (void);
 180 void (*xpc_check_remote_hb) (void);
 181
 182 enum xp_retval (*xpc_make_first_contact) (struct xpc_partition *part);
 183 void (*xpc_notify_senders_of_disconnect) (struct xpc_channel *ch);
 184 u64 (*xpc_get_chctl_all_flags) (struct xpc_partition *part);
 185 enum xp_retval (*xpc_allocate_msgqueues) (struct xpc_channel *ch);
 186 void (*xpc_free_msgqueues) (struct xpc_channel *ch);
 187 void (*xpc_process_msg_chctl_flags) (struct xpc_partition *part, int ch_number);
 188 int (*xpc_n_of_deliverable_msgs) (struct xpc_channel *ch);
 189 struct xpc_msg *(*xpc_get_deliverable_msg) (struct xpc_channel *ch);
 190
 191 void (*xpc_request_partition_activation) (struct xpc_rsvd_page *remote_rp,
 192                                           u64 remote_rp_pa, int nasid);
 193 void (*xpc_request_partition_reactivation) (struct xpc_partition *part);
 194 void (*xpc_request_partition_deactivation) (struct xpc_partition *part);
 195 void (*xpc_cancel_partition_deactivation_request) (struct xpc_partition *part);
 196
 197 void (*xpc_process_activate_IRQ_rcvd) (int n_IRQs_expected);
 198 enum xp_retval (*xpc_setup_infrastructure) (struct xpc_partition *part);
 199 void (*xpc_teardown_infrastructure) (struct xpc_partition *part);
 200
 201 void (*xpc_indicate_partition_engaged) (struct xpc_partition *part);
 202 int (*xpc_partition_engaged) (short partid);
 203 int (*xpc_any_partition_engaged) (void);
 204 void (*xpc_indicate_partition_disengaged) (struct xpc_partition *part);
 205 void (*xpc_assume_partition_disengaged) (short partid);
 206
 207 void (*xpc_send_chctl_closerequest) (struct xpc_channel *ch,
 208                                      unsigned long *irq_flags);
 209 void (*xpc_send_chctl_closereply) (struct xpc_channel *ch,
 210                                    unsigned long *irq_flags);
 211 void (*xpc_send_chctl_openrequest) (struct xpc_channel *ch,
 212                                     unsigned long *irq_flags);
 213 void (*xpc_send_chctl_openreply) (struct xpc_channel *ch,
 214                                   unsigned long *irq_flags);
 215
 216 enum xp_retval (*xpc_send_msg) (struct xpc_channel *ch, u32 flags,
 217                                 void *payload, u16 payload_size, u8 notify_type,
 218                                 xpc_notify_func func, void *key);
 219 void (*xpc_received_msg) (struct xpc_channel *ch, struct xpc_msg *msg);
 220
 221 /*
 222  * Timer function to enforce the timelimit on the partition disengage.
 223  */
 224 static void
 225 xpc_timeout_partition_disengage(unsigned long data)
 226 {
 227         struct xpc_partition *part = (struct xpc_partition *)data;
 228
 229         DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
 230
 231         (void)xpc_partition_disengaged(part);
 232
 233         DBUG_ON(part->disengage_timeout != 0);
 234         DBUG_ON(xpc_partition_engaged(XPC_PARTID(part)));
 235 }
 236
 237 /*
 238  * Timer to produce the heartbeat.  The timer structures function is
 239  * already set when this is initially called.  A tunable is used to
 240  * specify when the next timeout should occur.
 241  */
 242 static void
 243 xpc_hb_beater(unsigned long dummy)
 244 {
 245         xpc_increment_heartbeat();
 246
 247         if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
 248                 wake_up_interruptible(&xpc_activate_IRQ_wq);
 249
 250         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
 251         add_timer(&xpc_hb_timer);
 252 }
 253
 254 static void
 255 xpc_start_hb_beater(void)
 256 {
 257         xpc_heartbeat_init();
 258         init_timer(&xpc_hb_timer);
 259         xpc_hb_timer.function = xpc_hb_beater;
 260         xpc_hb_beater(0);
 261 }
 262
 263 static void
 264 xpc_stop_hb_beater(void)
 265 {
 266         del_timer_sync(&xpc_hb_timer);
 267         xpc_heartbeat_exit();
 268 }
 269
 270 /*
 271  * This thread is responsible for nearly all of the partition
 272  * activation/deactivation.
 273  */
 274 static int
 275 xpc_hb_checker(void *ignore)
 276 {
 277         int last_IRQ_count = 0;
 278         int new_IRQ_count;
 279         int force_IRQ = 0;
 280
 281         /* this thread was marked active by xpc_hb_init() */
 282
 283         set_cpus_allowed_ptr(current, &cpumask_of_cpu(XPC_HB_CHECK_CPU));
 284
 285         /* set our heartbeating to other partitions into motion */
 286         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
 287         xpc_start_hb_beater();
 288
 289         while (!xpc_exiting) {
 290
 291                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
 292                         "been received\n",
 293                         (int)(xpc_hb_check_timeout - jiffies),
 294                         atomic_read(&xpc_activate_IRQ_rcvd) - last_IRQ_count);
 295
 296                 /* checking of remote heartbeats is skewed by IRQ handling */
 297                 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
 298                         dev_dbg(xpc_part, "checking remote heartbeats\n");
 299                         xpc_check_remote_hb();
 300
 301                         /*
 302                          * We need to periodically recheck to ensure no
 303                          * IRQ/amo pairs have been missed.  That check
 304                          * must always reset xpc_hb_check_timeout.
 305                          */
 306                         force_IRQ = 1;
 307                 }
 308
 309                 /* check for outstanding IRQs */
 310                 new_IRQ_count = atomic_read(&xpc_activate_IRQ_rcvd);
 311                 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
 312                         force_IRQ = 0;
 313
 314                         dev_dbg(xpc_part, "found an IRQ to process; will be "
 315                                 "resetting xpc_hb_check_timeout\n");
 316
 317                         xpc_process_activate_IRQ_rcvd(new_IRQ_count -
 318                                                       last_IRQ_count);
 319                         last_IRQ_count = new_IRQ_count;
 320
 321                         xpc_hb_check_timeout = jiffies +
 322                             (xpc_hb_check_interval * HZ);
 323                 }
 324
 325                 /* wait for IRQ or timeout */
 326                 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
 327                                                (last_IRQ_count < atomic_read(
 328                                                 &xpc_activate_IRQ_rcvd)
 329                                                 || time_is_before_eq_jiffies(
 330                                                 xpc_hb_check_timeout) ||
 331                                                 xpc_exiting));
 332         }
 333
 334         xpc_stop_hb_beater();
 335
 336         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
 337
 338         /* mark this thread as having exited */
 339         complete(&xpc_hb_checker_exited);
 340         return 0;
 341 }
 342
 343 /*
 344  * This thread will attempt to discover other partitions to activate
 345  * based on info provided by SAL. This new thread is short lived and
 346  * will exit once discovery is complete.
 347  */
 348 static int
 349 xpc_initiate_discovery(void *ignore)
 350 {
 351         xpc_discovery();
 352
 353         dev_dbg(xpc_part, "discovery thread is exiting\n");
 354
 355         /* mark this thread as having exited */
 356         complete(&xpc_discovery_exited);
 357         return 0;
 358 }
 359
 360 /*
 361  * The first kthread assigned to a newly activated partition is the one
 362  * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
 363  * that kthread until the partition is brought down, at which time that kthread
 364  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
 365  * that XPC has dismantled all communication infrastructure for the associated
 366  * partition.) This kthread becomes the channel manager for that partition.
 367  *
 368  * Each active partition has a channel manager, who, besides connecting and
 369  * disconnecting channels, will ensure that each of the partition's connected
 370  * channels has the required number of assigned kthreads to get the work done.
 371  */
 372 static void
 373 xpc_channel_mgr(struct xpc_partition *part)
 374 {
 375         while (part->act_state != XPC_P_DEACTIVATING ||
 376                atomic_read(&part->nchannels_active) > 0 ||
 377                !xpc_partition_disengaged(part)) {
 378
 379                 xpc_process_sent_chctl_flags(part);
 380
 381                 /*
 382                  * Wait until we've been requested to activate kthreads or
 383                  * all of the channel's message queues have been torn down or
 384                  * a signal is pending.
 385                  *
 386                  * The channel_mgr_requests is set to 1 after being awakened,
 387                  * This is done to prevent the channel mgr from making one pass
 388                  * through the loop for each request, since he will
 389                  * be servicing all the requests in one pass. The reason it's
 390                  * set to 1 instead of 0 is so that other kthreads will know
 391                  * that the channel mgr is running and won't bother trying to
 392                  * wake him up.
 393                  */
 394                 atomic_dec(&part->channel_mgr_requests);
 395                 (void)wait_event_interruptible(part->channel_mgr_wq,
 396                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
 397                                  part->chctl.all_flags != 0 ||
 398                                  (part->act_state == XPC_P_DEACTIVATING &&
 399                                  atomic_read(&part->nchannels_active) == 0 &&
 400                                  xpc_partition_disengaged(part))));
 401                 atomic_set(&part->channel_mgr_requests, 1);
 402         }
 403 }
 404
 405 /*
 406  * When XPC HB determines that a partition has come up, it will create a new
 407  * kthread and that kthread will call this function to attempt to set up the
 408  * basic infrastructure used for Cross Partition Communication with the newly
 409  * upped partition.
 410  *
 411  * The kthread that was created by XPC HB and which setup the XPC
 412  * infrastructure will remain assigned to the partition becoming the channel
 413  * manager for that partition until the partition is deactivating, at which
 414  * time the kthread will teardown the XPC infrastructure and then exit.
 415  */
 416 static int
 417 xpc_activating(void *__partid)
 418 {
 419         short partid = (u64)__partid;
 420         struct xpc_partition *part = &xpc_partitions[partid];
 421         unsigned long irq_flags;
 422
 423         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
 424
 425         spin_lock_irqsave(&part->act_lock, irq_flags);
 426
 427         if (part->act_state == XPC_P_DEACTIVATING) {
 428                 part->act_state = XPC_P_INACTIVE;
 429                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 430                 part->remote_rp_pa = 0;
 431                 return 0;
 432         }
 433
 434         /* indicate the thread is activating */
 435         DBUG_ON(part->act_state != XPC_P_ACTIVATION_REQ);
 436         part->act_state = XPC_P_ACTIVATING;
 437
 438         XPC_SET_REASON(part, 0, 0);
 439         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 440
 441         dev_dbg(xpc_part, "activating partition %d\n", partid);
 442
 443         xpc_allow_hb(partid);
 444
 445         if (xpc_setup_infrastructure(part) == xpSuccess) {
 446                 (void)xpc_part_ref(part);       /* this will always succeed */
 447
 448                 if (xpc_make_first_contact(part) == xpSuccess) {
 449                         xpc_mark_partition_active(part);
 450                         xpc_channel_mgr(part);
 451                         /* won't return until partition is deactivating */
 452                 }
 453
 454                 xpc_part_deref(part);
 455                 xpc_teardown_infrastructure(part);
 456         }
 457
 458         xpc_disallow_hb(partid);
 459         xpc_mark_partition_inactive(part);
 460
 461         if (part->reason == xpReactivating) {
 462                 /* interrupting ourselves results in activating partition */
 463                 xpc_request_partition_reactivation(part);
 464         }
 465
 466         return 0;
 467 }
 468
 469 void
 470 xpc_activate_partition(struct xpc_partition *part)
 471 {
 472         short partid = XPC_PARTID(part);
 473         unsigned long irq_flags;
 474         struct task_struct *kthread;
 475
 476         spin_lock_irqsave(&part->act_lock, irq_flags);
 477
 478         DBUG_ON(part->act_state != XPC_P_INACTIVE);
 479
 480         part->act_state = XPC_P_ACTIVATION_REQ;
 481         XPC_SET_REASON(part, xpCloneKThread, __LINE__);
 482
 483         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 484
 485         kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
 486                               partid);
 487         if (IS_ERR(kthread)) {
 488                 spin_lock_irqsave(&part->act_lock, irq_flags);
 489                 part->act_state = XPC_P_INACTIVE;
 490                 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
 491                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 492         }
 493 }
 494
 495 void
 496 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
 497 {
 498         int idle = atomic_read(&ch->kthreads_idle);
 499         int assigned = atomic_read(&ch->kthreads_assigned);
 500         int wakeup;
 501
 502         DBUG_ON(needed <= 0);
 503
 504         if (idle > 0) {
 505                 wakeup = (needed > idle) ? idle : needed;
 506                 needed -= wakeup;
 507
 508                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
 509                         "channel=%d\n", wakeup, ch->partid, ch->number);
 510
 511                 /* only wakeup the requested number of kthreads */
 512                 wake_up_nr(&ch->idle_wq, wakeup);
 513         }
 514
 515         if (needed <= 0)
 516                 return;
 517
 518         if (needed + assigned > ch->kthreads_assigned_limit) {
 519                 needed = ch->kthreads_assigned_limit - assigned;
 520                 if (needed <= 0)
 521                         return;
 522         }
 523
 524         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
 525                 needed, ch->partid, ch->number);
 526
 527         xpc_create_kthreads(ch, needed, 0);
 528 }
 529
 530 /*
 531  * This function is where XPC's kthreads wait for messages to deliver.
 532  */
 533 static void
 534 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
 535 {
 536         do {
 537                 /* deliver messages to their intended recipients */
 538
 539                 while (xpc_n_of_deliverable_msgs(ch) > 0 &&
 540                        !(ch->flags & XPC_C_DISCONNECTING)) {
 541                         xpc_deliver_msg(ch);
 542                 }
 543
 544                 if (atomic_inc_return(&ch->kthreads_idle) >
 545                     ch->kthreads_idle_limit) {
 546                         /* too many idle kthreads on this channel */
 547                         atomic_dec(&ch->kthreads_idle);
 548                         break;
 549                 }
 550
 551                 dev_dbg(xpc_chan, "idle kthread calling "
 552                         "wait_event_interruptible_exclusive()\n");
 553
 554                 (void)wait_event_interruptible_exclusive(ch->idle_wq,
 555                                 (xpc_n_of_deliverable_msgs(ch) > 0 ||
 556                                  (ch->flags & XPC_C_DISCONNECTING)));
 557
 558                 atomic_dec(&ch->kthreads_idle);
 559
 560         } while (!(ch->flags & XPC_C_DISCONNECTING));
 561 }
 562
 563 static int
 564 xpc_kthread_start(void *args)
 565 {
 566         short partid = XPC_UNPACK_ARG1(args);
 567         u16 ch_number = XPC_UNPACK_ARG2(args);
 568         struct xpc_partition *part = &xpc_partitions[partid];
 569         struct xpc_channel *ch;
 570         int n_needed;
 571         unsigned long irq_flags;
 572
 573         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
 574                 partid, ch_number);
 575
 576         ch = &part->channels[ch_number];
 577
 578         if (!(ch->flags & XPC_C_DISCONNECTING)) {
 579
 580                 /* let registerer know that connection has been established */
 581
 582                 spin_lock_irqsave(&ch->lock, irq_flags);
 583                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
 584                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
 585                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 586
 587                         xpc_connected_callout(ch);
 588
 589                         spin_lock_irqsave(&ch->lock, irq_flags);
 590                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
 591                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 592
 593                         /*
 594                          * It is possible that while the callout was being
 595                          * made that the remote partition sent some messages.
 596                          * If that is the case, we may need to activate
 597                          * additional kthreads to help deliver them. We only
 598                          * need one less than total #of messages to deliver.
 599                          */
 600                         n_needed = xpc_n_of_deliverable_msgs(ch) - 1;
 601                         if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
 602                                 xpc_activate_kthreads(ch, n_needed);
 603
 604                 } else {
 605                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 606                 }
 607
 608                 xpc_kthread_waitmsgs(part, ch);
 609         }
 610
 611         /* let registerer know that connection is disconnecting */
 612
 613         spin_lock_irqsave(&ch->lock, irq_flags);
 614         if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
 615             !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
 616                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
 617                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 618
 619                 xpc_disconnect_callout(ch, xpDisconnecting);
 620
 621                 spin_lock_irqsave(&ch->lock, irq_flags);
 622                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
 623         }
 624         spin_unlock_irqrestore(&ch->lock, irq_flags);
 625
 626         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 627             atomic_dec_return(&part->nchannels_engaged) == 0) {
 628                 xpc_indicate_partition_disengaged(part);
 629         }
 630
 631         xpc_msgqueue_deref(ch);
 632
 633         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
 634                 partid, ch_number);
 635
 636         xpc_part_deref(part);
 637         return 0;
 638 }
 639
 640 /*
 641  * For each partition that XPC has established communications with, there is
 642  * a minimum of one kernel thread assigned to perform any operation that
 643  * may potentially sleep or block (basically the callouts to the asynchronous
 644  * functions registered via xpc_connect()).
 645  *
 646  * Additional kthreads are created and destroyed by XPC as the workload
 647  * demands.
 648  *
 649  * A kthread is assigned to one of the active channels that exists for a given
 650  * partition.
 651  */
 652 void
 653 xpc_create_kthreads(struct xpc_channel *ch, int needed,
 654                     int ignore_disconnecting)
 655 {
 656         unsigned long irq_flags;
 657         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
 658         struct xpc_partition *part = &xpc_partitions[ch->partid];
 659         struct task_struct *kthread;
 660
 661         while (needed-- > 0) {
 662
 663                 /*
 664                  * The following is done on behalf of the newly created
 665                  * kthread. That kthread is responsible for doing the
 666                  * counterpart to the following before it exits.
 667                  */
 668                 if (ignore_disconnecting) {
 669                         if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
 670                                 /* kthreads assigned had gone to zero */
 671                                 BUG_ON(!(ch->flags &
 672                                          XPC_C_DISCONNECTINGCALLOUT_MADE));
 673                                 break;
 674                         }
 675
 676                 } else if (ch->flags & XPC_C_DISCONNECTING) {
 677                         break;
 678
 679                 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
 680                            atomic_inc_return(&part->nchannels_engaged) == 1) {
 681                                 xpc_indicate_partition_engaged(part);
 682                 }
 683                 (void)xpc_part_ref(part);
 684                 xpc_msgqueue_ref(ch);
 685
 686                 kthread = kthread_run(xpc_kthread_start, (void *)args,
 687                                       "xpc%02dc%d", ch->partid, ch->number);
 688                 if (IS_ERR(kthread)) {
 689                         /* the fork failed */
 690
 691                         /*
 692                          * NOTE: if (ignore_disconnecting &&
 693                          * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
 694                          * then we'll deadlock if all other kthreads assigned
 695                          * to this channel are blocked in the channel's
 696                          * registerer, because the only thing that will unblock
 697                          * them is the xpDisconnecting callout that this
 698                          * failed kthread_run() would have made.
 699                          */
 700
 701                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 702                             atomic_dec_return(&part->nchannels_engaged) == 0) {
 703                                 xpc_indicate_partition_disengaged(part);
 704                         }
 705                         xpc_msgqueue_deref(ch);
 706                         xpc_part_deref(part);
 707
 708                         if (atomic_read(&ch->kthreads_assigned) <
 709                             ch->kthreads_idle_limit) {
 710                                 /*
 711                                  * Flag this as an error only if we have an
 712                                  * insufficient #of kthreads for the channel
 713                                  * to function.
 714                                  */
 715                                 spin_lock_irqsave(&ch->lock, irq_flags);
 716                                 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
 717                                                        &irq_flags);
 718                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 719                         }
 720                         break;
 721                 }
 722         }
 723 }
 724
 725 void
 726 xpc_disconnect_wait(int ch_number)
 727 {
 728         unsigned long irq_flags;
 729         short partid;
 730         struct xpc_partition *part;
 731         struct xpc_channel *ch;
 732         int wakeup_channel_mgr;
 733
 734         /* now wait for all callouts to the caller's function to cease */
 735         for (partid = 0; partid < xp_max_npartitions; partid++) {
 736                 part = &xpc_partitions[partid];
 737
 738                 if (!xpc_part_ref(part))
 739                         continue;
 740
 741                 ch = &part->channels[ch_number];
 742
 743                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
 744                         xpc_part_deref(part);
 745                         continue;
 746                 }
 747
 748                 wait_for_completion(&ch->wdisconnect_wait);
 749
 750                 spin_lock_irqsave(&ch->lock, irq_flags);
 751                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
 752                 wakeup_channel_mgr = 0;
 753
 754                 if (ch->delayed_chctl_flags) {
 755                         if (part->act_state != XPC_P_DEACTIVATING) {
 756                                 spin_lock(&part->chctl_lock);
 757                                 part->chctl.flags[ch->number] |=
 758                                     ch->delayed_chctl_flags;
 759                                 spin_unlock(&part->chctl_lock);
 760                                 wakeup_channel_mgr = 1;
 761                         }
 762                         ch->delayed_chctl_flags = 0;
 763                 }
 764
 765                 ch->flags &= ~XPC_C_WDISCONNECT;
 766                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 767
 768                 if (wakeup_channel_mgr)
 769                         xpc_wakeup_channel_mgr(part);
 770
 771                 xpc_part_deref(part);
 772         }
 773 }
 774
 775 static void
 776 xpc_do_exit(enum xp_retval reason)
 777 {
 778         short partid;
 779         int active_part_count, printed_waiting_msg = 0;
 780         struct xpc_partition *part;
 781         unsigned long printmsg_time, disengage_timeout = 0;
 782
 783         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
 784         DBUG_ON(xpc_exiting == 1);
 785
 786         /*
 787          * Let the heartbeat checker thread and the discovery thread
 788          * (if one is running) know that they should exit. Also wake up
 789          * the heartbeat checker thread in case it's sleeping.
 790          */
 791         xpc_exiting = 1;
 792         wake_up_interruptible(&xpc_activate_IRQ_wq);
 793
 794         /* wait for the discovery thread to exit */
 795         wait_for_completion(&xpc_discovery_exited);
 796
 797         /* wait for the heartbeat checker thread to exit */
 798         wait_for_completion(&xpc_hb_checker_exited);
 799
 800         /* sleep for a 1/3 of a second or so */
 801         (void)msleep_interruptible(300);
 802
 803         /* wait for all partitions to become inactive */
 804
 805         printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
 806         xpc_disengage_timedout = 0;
 807
 808         do {
 809                 active_part_count = 0;
 810
 811                 for (partid = 0; partid < xp_max_npartitions; partid++) {
 812                         part = &xpc_partitions[partid];
 813
 814                         if (xpc_partition_disengaged(part) &&
 815                             part->act_state == XPC_P_INACTIVE) {
 816                                 continue;
 817                         }
 818
 819                         active_part_count++;
 820
 821                         XPC_DEACTIVATE_PARTITION(part, reason);
 822
 823                         if (part->disengage_timeout > disengage_timeout)
 824                                 disengage_timeout = part->disengage_timeout;
 825                 }
 826
 827                 if (xpc_any_partition_engaged()) {
 828                         if (time_is_before_jiffies(printmsg_time)) {
 829                                 dev_info(xpc_part, "waiting for remote "
 830                                          "partitions to deactivate, timeout in "
 831                                          "%ld seconds\n", (disengage_timeout -
 832                                          jiffies) / HZ);
 833                                 printmsg_time = jiffies +
 834                                     (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
 835                                 printed_waiting_msg = 1;
 836                         }
 837
 838                 } else if (active_part_count > 0) {
 839                         if (printed_waiting_msg) {
 840                                 dev_info(xpc_part, "waiting for local partition"
 841                                          " to deactivate\n");
 842                                 printed_waiting_msg = 0;
 843                         }
 844
 845                 } else {
 846                         if (!xpc_disengage_timedout) {
 847                                 dev_info(xpc_part, "all partitions have "
 848                                          "deactivated\n");
 849                         }
 850                         break;
 851                 }
 852
 853                 /* sleep for a 1/3 of a second or so */
 854                 (void)msleep_interruptible(300);
 855
 856         } while (1);
 857
 858         DBUG_ON(xpc_any_partition_engaged());
 859         DBUG_ON(xpc_any_hbs_allowed() != 0);
 860
 861         /* a zero timestamp indicates our rsvd page is not initialized */
 862         xpc_rsvd_page->ts_jiffies = 0;
 863
 864         if (reason == xpUnloading) {
 865                 (void)unregister_die_notifier(&xpc_die_notifier);
 866                 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
 867         }
 868
 869         /* clear the interface to XPC's functions */
 870         xpc_clear_interface();
 871
 872         if (xpc_sysctl)
 873                 unregister_sysctl_table(xpc_sysctl);
 874
 875         kfree(xpc_partitions);
 876
 877         if (is_shub())
 878                 xpc_exit_sn2();
 879         else
 880                 xpc_exit_uv();
 881 }
 882
 883 /*
 884  * This function is called when the system is being rebooted.
 885  */
 886 static int
 887 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
 888 {
 889         enum xp_retval reason;
 890
 891         switch (event) {
 892         case SYS_RESTART:
 893                 reason = xpSystemReboot;
 894                 break;
 895         case SYS_HALT:
 896                 reason = xpSystemHalt;
 897                 break;
 898         case SYS_POWER_OFF:
 899                 reason = xpSystemPoweroff;
 900                 break;
 901         default:
 902                 reason = xpSystemGoingDown;
 903         }
 904
 905         xpc_do_exit(reason);
 906         return NOTIFY_DONE;
 907 }
 908
 909 /*
 910  * Notify other partitions to deactivate from us by first disengaging from all
 911  * references to our memory.
 912  */
 913 static void
 914 xpc_die_deactivate(void)
 915 {
 916         struct xpc_partition *part;
 917         short partid;
 918         int any_engaged;
 919         long keep_waiting;
 920         long wait_to_print;
 921
 922         /* keep xpc_hb_checker thread from doing anything (just in case) */
 923         xpc_exiting = 1;
 924
 925         xpc_disallow_all_hbs(); /*indicate we're deactivated */
 926
 927         for (partid = 0; partid < xp_max_npartitions; partid++) {
 928                 part = &xpc_partitions[partid];
 929
 930                 if (xpc_partition_engaged(partid) ||
 931                     part->act_state != XPC_P_INACTIVE) {
 932                         xpc_request_partition_deactivation(part);
 933                         xpc_indicate_partition_disengaged(part);
 934                 }
 935         }
 936
 937         /*
 938          * Though we requested that all other partitions deactivate from us,
 939          * we only wait until they've all disengaged or we've reached the
 940          * defined timelimit.
 941          *
 942          * Given that one iteration through the following while-loop takes
 943          * approximately 200 microseconds, calculate the #of loops to take
 944          * before bailing and the #of loops before printing a waiting message.
 945          */
 946         keep_waiting = xpc_disengage_timelimit * 1000 * 5;
 947         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
 948
 949         while (1) {
 950                 any_engaged = xpc_any_partition_engaged();
 951                 if (!any_engaged) {
 952                         dev_info(xpc_part, "all partitions have deactivated\n");
 953                         break;
 954                 }
 955
 956                 if (!keep_waiting--) {
 957                         for (partid = 0; partid < xp_max_npartitions;
 958                              partid++) {
 959                                 if (xpc_partition_engaged(partid)) {
 960                                         dev_info(xpc_part, "deactivate from "
 961                                                  "remote partition %d timed "
 962                                                  "out\n", partid);
 963                                 }
 964                         }
 965                         break;
 966                 }
 967
 968                 if (!wait_to_print--) {
 969                         dev_info(xpc_part, "waiting for remote partitions to "
 970                                  "deactivate, timeout in %ld seconds\n",
 971                                  keep_waiting / (1000 * 5));
 972                         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
 973                             1000 * 5;
 974                 }
 975
 976                 udelay(200);
 977         }
 978 }
 979
 980 /*
 981  * This function is called when the system is being restarted or halted due
 982  * to some sort of system failure. If this is the case we need to notify the
 983  * other partitions to disengage from all references to our memory.
 984  * This function can also be called when our heartbeater could be offlined
 985  * for a time. In this case we need to notify other partitions to not worry
 986  * about the lack of a heartbeat.
 987  */
 988 static int
 989 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
 990 {
 991 #ifdef CONFIG_IA64              /* !!! temporary kludge */
 992         switch (event) {
 993         case DIE_MACHINE_RESTART:
 994         case DIE_MACHINE_HALT:
 995                 xpc_die_deactivate();
 996                 break;
 997
 998         case DIE_KDEBUG_ENTER:
 999                 /* Should lack of heartbeat be ignored by other partitions? */
1000                 if (!xpc_kdebug_ignore)
1001                         break;
1002
1003                 /* fall through */
1004         case DIE_MCA_MONARCH_ENTER:
1005         case DIE_INIT_MONARCH_ENTER:
1006                 xpc_offline_heartbeat();
1007                 break;
1008
1009         case DIE_KDEBUG_LEAVE:
1010                 /* Is lack of heartbeat being ignored by other partitions? */
1011                 if (!xpc_kdebug_ignore)
1012                         break;
1013
1014                 /* fall through */
1015         case DIE_MCA_MONARCH_LEAVE:
1016         case DIE_INIT_MONARCH_LEAVE:
1017                 xpc_online_heartbeat();
1018                 break;
1019         }
1020 #else
1021         xpc_die_deactivate();
1022 #endif
1023
1024         return NOTIFY_DONE;
1025 }
1026
1027 int __init
1028 xpc_init(void)
1029 {
1030         int ret;
1031         short partid;
1032         struct xpc_partition *part;
1033         struct task_struct *kthread;
1034
1035         snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1036         snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1037
1038         if (is_shub()) {
1039                 /*
1040                  * The ia64-sn2 architecture supports at most 64 partitions.
1041                  * And the inability to unregister remote amos restricts us
1042                  * further to only support exactly 64 partitions on this
1043                  * architecture, no less.
1044                  */
1045                 if (xp_max_npartitions != 64)
1046                         return -EINVAL;
1047
1048                 ret = xpc_init_sn2();
1049                 if (ret != 0)
1050                         return ret;
1051
1052         } else if (is_uv()) {
1053                 xpc_init_uv();
1054
1055         } else {
1056                 return -ENODEV;
1057         }
1058
1059         xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
1060                                  xp_max_npartitions, GFP_KERNEL);
1061         if (xpc_partitions == NULL) {
1062                 dev_err(xpc_part, "can't get memory for partition structure\n");
1063                 ret = -ENOMEM;
1064                 goto out_1;
1065         }
1066
1067         /*
1068          * The first few fields of each entry of xpc_partitions[] need to
1069          * be initialized now so that calls to xpc_connect() and
1070          * xpc_disconnect() can be made prior to the activation of any remote
1071          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1072          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1073          * PARTITION HAS BEEN ACTIVATED.
1074          */
1075         for (partid = 0; partid < xp_max_npartitions; partid++) {
1076                 part = &xpc_partitions[partid];
1077
1078                 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
1079
1080                 part->activate_IRQ_rcvd = 0;
1081                 spin_lock_init(&part->act_lock);
1082                 part->act_state = XPC_P_INACTIVE;
1083                 XPC_SET_REASON(part, 0, 0);
1084
1085                 init_timer(&part->disengage_timer);
1086                 part->disengage_timer.function =
1087                     xpc_timeout_partition_disengage;
1088                 part->disengage_timer.data = (unsigned long)part;
1089
1090                 part->setup_state = XPC_P_UNSET;
1091                 init_waitqueue_head(&part->teardown_wq);
1092                 atomic_set(&part->references, 0);
1093         }
1094
1095         xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1096
1097         /*
1098          * Fill the partition reserved page with the information needed by
1099          * other partitions to discover we are alive and establish initial
1100          * communications.
1101          */
1102         xpc_rsvd_page = xpc_setup_rsvd_page();
1103         if (xpc_rsvd_page == NULL) {
1104                 dev_err(xpc_part, "can't setup our reserved page\n");
1105                 ret = -EBUSY;
1106                 goto out_2;
1107         }
1108
1109         /* add ourselves to the reboot_notifier_list */
1110         ret = register_reboot_notifier(&xpc_reboot_notifier);
1111         if (ret != 0)
1112                 dev_warn(xpc_part, "can't register reboot notifier\n");
1113
1114         /* add ourselves to the die_notifier list */
1115         ret = register_die_notifier(&xpc_die_notifier);
1116         if (ret != 0)
1117                 dev_warn(xpc_part, "can't register die notifier\n");
1118
1119         /*
1120          * The real work-horse behind xpc.  This processes incoming
1121          * interrupts and monitors remote heartbeats.
1122          */
1123         kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1124         if (IS_ERR(kthread)) {
1125                 dev_err(xpc_part, "failed while forking hb check thread\n");
1126                 ret = -EBUSY;
1127                 goto out_3;
1128         }
1129
1130         /*
1131          * Startup a thread that will attempt to discover other partitions to
1132          * activate based on info provided by SAL. This new thread is short
1133          * lived and will exit once discovery is complete.
1134          */
1135         kthread = kthread_run(xpc_initiate_discovery, NULL,
1136                               XPC_DISCOVERY_THREAD_NAME);
1137         if (IS_ERR(kthread)) {
1138                 dev_err(xpc_part, "failed while forking discovery thread\n");
1139
1140                 /* mark this new thread as a non-starter */
1141                 complete(&xpc_discovery_exited);
1142
1143                 xpc_do_exit(xpUnloading);
1144                 return -EBUSY;
1145         }
1146
1147         /* set the interface to point at XPC's functions */
1148         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1149                           xpc_initiate_send, xpc_initiate_send_notify,
1150                           xpc_initiate_received, xpc_initiate_partid_to_nasids);
1151
1152         return 0;
1153
1154         /* initialization was not successful */
1155 out_3:
1156         /* a zero timestamp indicates our rsvd page is not initialized */
1157         xpc_rsvd_page->ts_jiffies = 0;
1158
1159         (void)unregister_die_notifier(&xpc_die_notifier);
1160         (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1161 out_2:
1162         if (xpc_sysctl)
1163                 unregister_sysctl_table(xpc_sysctl);
1164         kfree(xpc_partitions);
1165 out_1:
1166         if (is_shub())
1167                 xpc_exit_sn2();
1168         else
1169                 xpc_exit_uv();
1170         return ret;
1171 }
1172
1173 module_init(xpc_init);
1174
1175 void __exit
1176 xpc_exit(void)
1177 {
1178         xpc_do_exit(xpUnloading);
1179 }
1180
1181 module_exit(xpc_exit);
1182
1183 MODULE_AUTHOR("Silicon Graphics, Inc.");
1184 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1185 MODULE_LICENSE("GPL");
1186
1187 module_param(xpc_hb_interval, int, 0);
1188 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1189                  "heartbeat increments.");
1190
1191 module_param(xpc_hb_check_interval, int, 0);
1192 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1193                  "heartbeat checks.");
1194
1195 module_param(xpc_disengage_timelimit, int, 0);
1196 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1197                  "for disengage to complete.");
1198
1199 module_param(xpc_kdebug_ignore, int, 0);
1200 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1201                  "other partitions when dropping into kdebug.");