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) (void *buf, u64 *cookie,
 173                                                   unsigned long *rp_pa,
 174                                                   size_t *len);
 175 enum xp_retval (*xpc_rsvd_page_init) (struct xpc_rsvd_page *rp);
 176 void (*xpc_heartbeat_init) (void);
 177 void (*xpc_heartbeat_exit) (void);
 178 void (*xpc_increment_heartbeat) (void);
 179 void (*xpc_offline_heartbeat) (void);
 180 void (*xpc_online_heartbeat) (void);
 181 enum xp_retval (*xpc_get_remote_heartbeat) (struct xpc_partition *part);
 182
 183 enum xp_retval (*xpc_make_first_contact) (struct xpc_partition *part);
 184 void (*xpc_notify_senders_of_disconnect) (struct xpc_channel *ch);
 185 u64 (*xpc_get_chctl_all_flags) (struct xpc_partition *part);
 186 enum xp_retval (*xpc_allocate_msgqueues) (struct xpc_channel *ch);
 187 void (*xpc_free_msgqueues) (struct xpc_channel *ch);
 188 void (*xpc_process_msg_chctl_flags) (struct xpc_partition *part, int ch_number);
 189 int (*xpc_n_of_deliverable_msgs) (struct xpc_channel *ch);
 190 struct xpc_msg *(*xpc_get_deliverable_msg) (struct xpc_channel *ch);
 191
 192 void (*xpc_request_partition_activation) (struct xpc_rsvd_page *remote_rp,
 193                                           unsigned long remote_rp_pa,
 194                                           int nasid);
 195 void (*xpc_request_partition_reactivation) (struct xpc_partition *part);
 196 void (*xpc_request_partition_deactivation) (struct xpc_partition *part);
 197 void (*xpc_cancel_partition_deactivation_request) (struct xpc_partition *part);
 198
 199 void (*xpc_process_activate_IRQ_rcvd) (int n_IRQs_expected);
 200 enum xp_retval (*xpc_setup_infrastructure) (struct xpc_partition *part);
 201 void (*xpc_teardown_infrastructure) (struct xpc_partition *part);
 202
 203 void (*xpc_indicate_partition_engaged) (struct xpc_partition *part);
 204 int (*xpc_partition_engaged) (short partid);
 205 int (*xpc_any_partition_engaged) (void);
 206 void (*xpc_indicate_partition_disengaged) (struct xpc_partition *part);
 207 void (*xpc_assume_partition_disengaged) (short partid);
 208
 209 void (*xpc_send_chctl_closerequest) (struct xpc_channel *ch,
 210                                      unsigned long *irq_flags);
 211 void (*xpc_send_chctl_closereply) (struct xpc_channel *ch,
 212                                    unsigned long *irq_flags);
 213 void (*xpc_send_chctl_openrequest) (struct xpc_channel *ch,
 214                                     unsigned long *irq_flags);
 215 void (*xpc_send_chctl_openreply) (struct xpc_channel *ch,
 216                                   unsigned long *irq_flags);
 217
 218 enum xp_retval (*xpc_send_msg) (struct xpc_channel *ch, u32 flags,
 219                                 void *payload, u16 payload_size, u8 notify_type,
 220                                 xpc_notify_func func, void *key);
 221 void (*xpc_received_msg) (struct xpc_channel *ch, struct xpc_msg *msg);
 222
 223 /*
 224  * Timer function to enforce the timelimit on the partition disengage.
 225  */
 226 static void
 227 xpc_timeout_partition_disengage(unsigned long data)
 228 {
 229         struct xpc_partition *part = (struct xpc_partition *)data;
 230
 231         DBUG_ON(time_is_after_jiffies(part->disengage_timeout));
 232
 233         (void)xpc_partition_disengaged(part);
 234
 235         DBUG_ON(part->disengage_timeout != 0);
 236         DBUG_ON(xpc_partition_engaged(XPC_PARTID(part)));
 237 }
 238
 239 /*
 240  * Timer to produce the heartbeat.  The timer structures function is
 241  * already set when this is initially called.  A tunable is used to
 242  * specify when the next timeout should occur.
 243  */
 244 static void
 245 xpc_hb_beater(unsigned long dummy)
 246 {
 247         xpc_increment_heartbeat();
 248
 249         if (time_is_before_eq_jiffies(xpc_hb_check_timeout))
 250                 wake_up_interruptible(&xpc_activate_IRQ_wq);
 251
 252         xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ);
 253         add_timer(&xpc_hb_timer);
 254 }
 255
 256 static void
 257 xpc_start_hb_beater(void)
 258 {
 259         xpc_heartbeat_init();
 260         init_timer(&xpc_hb_timer);
 261         xpc_hb_timer.function = xpc_hb_beater;
 262         xpc_hb_beater(0);
 263 }
 264
 265 static void
 266 xpc_stop_hb_beater(void)
 267 {
 268         del_timer_sync(&xpc_hb_timer);
 269         xpc_heartbeat_exit();
 270 }
 271
 272 /*
 273  * At periodic intervals, scan through all active partitions and ensure
 274  * their heartbeat is still active.  If not, the partition is deactivated.
 275  */
 276 static void
 277 xpc_check_remote_hb(void)
 278 {
 279         struct xpc_partition *part;
 280         short partid;
 281         enum xp_retval ret;
 282
 283         for (partid = 0; partid < xp_max_npartitions; partid++) {
 284
 285                 if (xpc_exiting)
 286                         break;
 287
 288                 if (partid == xp_partition_id)
 289                         continue;
 290
 291                 part = &xpc_partitions[partid];
 292
 293                 if (part->act_state == XPC_P_AS_INACTIVE ||
 294                     part->act_state == XPC_P_AS_DEACTIVATING) {
 295                         continue;
 296                 }
 297
 298                 ret = xpc_get_remote_heartbeat(part);
 299                 if (ret != xpSuccess)
 300                         XPC_DEACTIVATE_PARTITION(part, ret);
 301         }
 302 }
 303
 304 /*
 305  * This thread is responsible for nearly all of the partition
 306  * activation/deactivation.
 307  */
 308 static int
 309 xpc_hb_checker(void *ignore)
 310 {
 311         int last_IRQ_count = 0;
 312         int new_IRQ_count;
 313         int force_IRQ = 0;
 314
 315         /* this thread was marked active by xpc_hb_init() */
 316
 317         set_cpus_allowed_ptr(current, &cpumask_of_cpu(XPC_HB_CHECK_CPU));
 318
 319         /* set our heartbeating to other partitions into motion */
 320         xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ);
 321         xpc_start_hb_beater();
 322
 323         while (!xpc_exiting) {
 324
 325                 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have "
 326                         "been received\n",
 327                         (int)(xpc_hb_check_timeout - jiffies),
 328                         atomic_read(&xpc_activate_IRQ_rcvd) - last_IRQ_count);
 329
 330                 /* checking of remote heartbeats is skewed by IRQ handling */
 331                 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) {
 332                         dev_dbg(xpc_part, "checking remote heartbeats\n");
 333                         xpc_check_remote_hb();
 334
 335                         /*
 336                          * We need to periodically recheck to ensure no
 337                          * IRQ/amo pairs have been missed.  That check
 338                          * must always reset xpc_hb_check_timeout.
 339                          */
 340                         force_IRQ = 1;
 341                 }
 342
 343                 /* check for outstanding IRQs */
 344                 new_IRQ_count = atomic_read(&xpc_activate_IRQ_rcvd);
 345                 if (last_IRQ_count < new_IRQ_count || force_IRQ != 0) {
 346                         force_IRQ = 0;
 347
 348                         dev_dbg(xpc_part, "found an IRQ to process; will be "
 349                                 "resetting xpc_hb_check_timeout\n");
 350
 351                         xpc_process_activate_IRQ_rcvd(new_IRQ_count -
 352                                                       last_IRQ_count);
 353                         last_IRQ_count = new_IRQ_count;
 354
 355                         xpc_hb_check_timeout = jiffies +
 356                             (xpc_hb_check_interval * HZ);
 357                 }
 358
 359                 /* wait for IRQ or timeout */
 360                 (void)wait_event_interruptible(xpc_activate_IRQ_wq,
 361                                                (last_IRQ_count < atomic_read(
 362                                                 &xpc_activate_IRQ_rcvd)
 363                                                 || time_is_before_eq_jiffies(
 364                                                 xpc_hb_check_timeout) ||
 365                                                 xpc_exiting));
 366         }
 367
 368         xpc_stop_hb_beater();
 369
 370         dev_dbg(xpc_part, "heartbeat checker is exiting\n");
 371
 372         /* mark this thread as having exited */
 373         complete(&xpc_hb_checker_exited);
 374         return 0;
 375 }
 376
 377 /*
 378  * This thread will attempt to discover other partitions to activate
 379  * based on info provided by SAL. This new thread is short lived and
 380  * will exit once discovery is complete.
 381  */
 382 static int
 383 xpc_initiate_discovery(void *ignore)
 384 {
 385         xpc_discovery();
 386
 387         dev_dbg(xpc_part, "discovery thread is exiting\n");
 388
 389         /* mark this thread as having exited */
 390         complete(&xpc_discovery_exited);
 391         return 0;
 392 }
 393
 394 /*
 395  * The first kthread assigned to a newly activated partition is the one
 396  * created by XPC HB with which it calls xpc_activating(). XPC hangs on to
 397  * that kthread until the partition is brought down, at which time that kthread
 398  * returns back to XPC HB. (The return of that kthread will signify to XPC HB
 399  * that XPC has dismantled all communication infrastructure for the associated
 400  * partition.) This kthread becomes the channel manager for that partition.
 401  *
 402  * Each active partition has a channel manager, who, besides connecting and
 403  * disconnecting channels, will ensure that each of the partition's connected
 404  * channels has the required number of assigned kthreads to get the work done.
 405  */
 406 static void
 407 xpc_channel_mgr(struct xpc_partition *part)
 408 {
 409         while (part->act_state != XPC_P_AS_DEACTIVATING ||
 410                atomic_read(&part->nchannels_active) > 0 ||
 411                !xpc_partition_disengaged(part)) {
 412
 413                 xpc_process_sent_chctl_flags(part);
 414
 415                 /*
 416                  * Wait until we've been requested to activate kthreads or
 417                  * all of the channel's message queues have been torn down or
 418                  * a signal is pending.
 419                  *
 420                  * The channel_mgr_requests is set to 1 after being awakened,
 421                  * This is done to prevent the channel mgr from making one pass
 422                  * through the loop for each request, since he will
 423                  * be servicing all the requests in one pass. The reason it's
 424                  * set to 1 instead of 0 is so that other kthreads will know
 425                  * that the channel mgr is running and won't bother trying to
 426                  * wake him up.
 427                  */
 428                 atomic_dec(&part->channel_mgr_requests);
 429                 (void)wait_event_interruptible(part->channel_mgr_wq,
 430                                 (atomic_read(&part->channel_mgr_requests) > 0 ||
 431                                  part->chctl.all_flags != 0 ||
 432                                  (part->act_state == XPC_P_AS_DEACTIVATING &&
 433                                  atomic_read(&part->nchannels_active) == 0 &&
 434                                  xpc_partition_disengaged(part))));
 435                 atomic_set(&part->channel_mgr_requests, 1);
 436         }
 437 }
 438
 439 /*
 440  * When XPC HB determines that a partition has come up, it will create a new
 441  * kthread and that kthread will call this function to attempt to set up the
 442  * basic infrastructure used for Cross Partition Communication with the newly
 443  * upped partition.
 444  *
 445  * The kthread that was created by XPC HB and which setup the XPC
 446  * infrastructure will remain assigned to the partition becoming the channel
 447  * manager for that partition until the partition is deactivating, at which
 448  * time the kthread will teardown the XPC infrastructure and then exit.
 449  */
 450 static int
 451 xpc_activating(void *__partid)
 452 {
 453         short partid = (u64)__partid;
 454         struct xpc_partition *part = &xpc_partitions[partid];
 455         unsigned long irq_flags;
 456
 457         DBUG_ON(partid < 0 || partid >= xp_max_npartitions);
 458
 459         spin_lock_irqsave(&part->act_lock, irq_flags);
 460
 461         if (part->act_state == XPC_P_AS_DEACTIVATING) {
 462                 part->act_state = XPC_P_AS_INACTIVE;
 463                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 464                 part->remote_rp_pa = 0;
 465                 return 0;
 466         }
 467
 468         /* indicate the thread is activating */
 469         DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ);
 470         part->act_state = XPC_P_AS_ACTIVATING;
 471
 472         XPC_SET_REASON(part, 0, 0);
 473         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 474
 475         dev_dbg(xpc_part, "activating partition %d\n", partid);
 476
 477         xpc_allow_hb(partid);
 478
 479         if (xpc_setup_infrastructure(part) == xpSuccess) {
 480                 (void)xpc_part_ref(part);       /* this will always succeed */
 481
 482                 if (xpc_make_first_contact(part) == xpSuccess) {
 483                         xpc_mark_partition_active(part);
 484                         xpc_channel_mgr(part);
 485                         /* won't return until partition is deactivating */
 486                 }
 487
 488                 xpc_part_deref(part);
 489                 xpc_teardown_infrastructure(part);
 490         }
 491
 492         xpc_disallow_hb(partid);
 493         xpc_mark_partition_inactive(part);
 494
 495         if (part->reason == xpReactivating) {
 496                 /* interrupting ourselves results in activating partition */
 497                 xpc_request_partition_reactivation(part);
 498         }
 499
 500         return 0;
 501 }
 502
 503 void
 504 xpc_activate_partition(struct xpc_partition *part)
 505 {
 506         short partid = XPC_PARTID(part);
 507         unsigned long irq_flags;
 508         struct task_struct *kthread;
 509
 510         spin_lock_irqsave(&part->act_lock, irq_flags);
 511
 512         DBUG_ON(part->act_state != XPC_P_AS_INACTIVE);
 513
 514         part->act_state = XPC_P_AS_ACTIVATION_REQ;
 515         XPC_SET_REASON(part, xpCloneKThread, __LINE__);
 516
 517         spin_unlock_irqrestore(&part->act_lock, irq_flags);
 518
 519         kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d",
 520                               partid);
 521         if (IS_ERR(kthread)) {
 522                 spin_lock_irqsave(&part->act_lock, irq_flags);
 523                 part->act_state = XPC_P_AS_INACTIVE;
 524                 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__);
 525                 spin_unlock_irqrestore(&part->act_lock, irq_flags);
 526         }
 527 }
 528
 529 void
 530 xpc_activate_kthreads(struct xpc_channel *ch, int needed)
 531 {
 532         int idle = atomic_read(&ch->kthreads_idle);
 533         int assigned = atomic_read(&ch->kthreads_assigned);
 534         int wakeup;
 535
 536         DBUG_ON(needed <= 0);
 537
 538         if (idle > 0) {
 539                 wakeup = (needed > idle) ? idle : needed;
 540                 needed -= wakeup;
 541
 542                 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, "
 543                         "channel=%d\n", wakeup, ch->partid, ch->number);
 544
 545                 /* only wakeup the requested number of kthreads */
 546                 wake_up_nr(&ch->idle_wq, wakeup);
 547         }
 548
 549         if (needed <= 0)
 550                 return;
 551
 552         if (needed + assigned > ch->kthreads_assigned_limit) {
 553                 needed = ch->kthreads_assigned_limit - assigned;
 554                 if (needed <= 0)
 555                         return;
 556         }
 557
 558         dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n",
 559                 needed, ch->partid, ch->number);
 560
 561         xpc_create_kthreads(ch, needed, 0);
 562 }
 563
 564 /*
 565  * This function is where XPC's kthreads wait for messages to deliver.
 566  */
 567 static void
 568 xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch)
 569 {
 570         do {
 571                 /* deliver messages to their intended recipients */
 572
 573                 while (xpc_n_of_deliverable_msgs(ch) > 0 &&
 574                        !(ch->flags & XPC_C_DISCONNECTING)) {
 575                         xpc_deliver_msg(ch);
 576                 }
 577
 578                 if (atomic_inc_return(&ch->kthreads_idle) >
 579                     ch->kthreads_idle_limit) {
 580                         /* too many idle kthreads on this channel */
 581                         atomic_dec(&ch->kthreads_idle);
 582                         break;
 583                 }
 584
 585                 dev_dbg(xpc_chan, "idle kthread calling "
 586                         "wait_event_interruptible_exclusive()\n");
 587
 588                 (void)wait_event_interruptible_exclusive(ch->idle_wq,
 589                                 (xpc_n_of_deliverable_msgs(ch) > 0 ||
 590                                  (ch->flags & XPC_C_DISCONNECTING)));
 591
 592                 atomic_dec(&ch->kthreads_idle);
 593
 594         } while (!(ch->flags & XPC_C_DISCONNECTING));
 595 }
 596
 597 static int
 598 xpc_kthread_start(void *args)
 599 {
 600         short partid = XPC_UNPACK_ARG1(args);
 601         u16 ch_number = XPC_UNPACK_ARG2(args);
 602         struct xpc_partition *part = &xpc_partitions[partid];
 603         struct xpc_channel *ch;
 604         int n_needed;
 605         unsigned long irq_flags;
 606
 607         dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n",
 608                 partid, ch_number);
 609
 610         ch = &part->channels[ch_number];
 611
 612         if (!(ch->flags & XPC_C_DISCONNECTING)) {
 613
 614                 /* let registerer know that connection has been established */
 615
 616                 spin_lock_irqsave(&ch->lock, irq_flags);
 617                 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) {
 618                         ch->flags |= XPC_C_CONNECTEDCALLOUT;
 619                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 620
 621                         xpc_connected_callout(ch);
 622
 623                         spin_lock_irqsave(&ch->lock, irq_flags);
 624                         ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE;
 625                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 626
 627                         /*
 628                          * It is possible that while the callout was being
 629                          * made that the remote partition sent some messages.
 630                          * If that is the case, we may need to activate
 631                          * additional kthreads to help deliver them. We only
 632                          * need one less than total #of messages to deliver.
 633                          */
 634                         n_needed = xpc_n_of_deliverable_msgs(ch) - 1;
 635                         if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING))
 636                                 xpc_activate_kthreads(ch, n_needed);
 637
 638                 } else {
 639                         spin_unlock_irqrestore(&ch->lock, irq_flags);
 640                 }
 641
 642                 xpc_kthread_waitmsgs(part, ch);
 643         }
 644
 645         /* let registerer know that connection is disconnecting */
 646
 647         spin_lock_irqsave(&ch->lock, irq_flags);
 648         if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) &&
 649             !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) {
 650                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT;
 651                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 652
 653                 xpc_disconnect_callout(ch, xpDisconnecting);
 654
 655                 spin_lock_irqsave(&ch->lock, irq_flags);
 656                 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE;
 657         }
 658         spin_unlock_irqrestore(&ch->lock, irq_flags);
 659
 660         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 661             atomic_dec_return(&part->nchannels_engaged) == 0) {
 662                 xpc_indicate_partition_disengaged(part);
 663         }
 664
 665         xpc_msgqueue_deref(ch);
 666
 667         dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n",
 668                 partid, ch_number);
 669
 670         xpc_part_deref(part);
 671         return 0;
 672 }
 673
 674 /*
 675  * For each partition that XPC has established communications with, there is
 676  * a minimum of one kernel thread assigned to perform any operation that
 677  * may potentially sleep or block (basically the callouts to the asynchronous
 678  * functions registered via xpc_connect()).
 679  *
 680  * Additional kthreads are created and destroyed by XPC as the workload
 681  * demands.
 682  *
 683  * A kthread is assigned to one of the active channels that exists for a given
 684  * partition.
 685  */
 686 void
 687 xpc_create_kthreads(struct xpc_channel *ch, int needed,
 688                     int ignore_disconnecting)
 689 {
 690         unsigned long irq_flags;
 691         u64 args = XPC_PACK_ARGS(ch->partid, ch->number);
 692         struct xpc_partition *part = &xpc_partitions[ch->partid];
 693         struct task_struct *kthread;
 694
 695         while (needed-- > 0) {
 696
 697                 /*
 698                  * The following is done on behalf of the newly created
 699                  * kthread. That kthread is responsible for doing the
 700                  * counterpart to the following before it exits.
 701                  */
 702                 if (ignore_disconnecting) {
 703                         if (!atomic_inc_not_zero(&ch->kthreads_assigned)) {
 704                                 /* kthreads assigned had gone to zero */
 705                                 BUG_ON(!(ch->flags &
 706                                          XPC_C_DISCONNECTINGCALLOUT_MADE));
 707                                 break;
 708                         }
 709
 710                 } else if (ch->flags & XPC_C_DISCONNECTING) {
 711                         break;
 712
 713                 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 &&
 714                            atomic_inc_return(&part->nchannels_engaged) == 1) {
 715                                 xpc_indicate_partition_engaged(part);
 716                 }
 717                 (void)xpc_part_ref(part);
 718                 xpc_msgqueue_ref(ch);
 719
 720                 kthread = kthread_run(xpc_kthread_start, (void *)args,
 721                                       "xpc%02dc%d", ch->partid, ch->number);
 722                 if (IS_ERR(kthread)) {
 723                         /* the fork failed */
 724
 725                         /*
 726                          * NOTE: if (ignore_disconnecting &&
 727                          * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true,
 728                          * then we'll deadlock if all other kthreads assigned
 729                          * to this channel are blocked in the channel's
 730                          * registerer, because the only thing that will unblock
 731                          * them is the xpDisconnecting callout that this
 732                          * failed kthread_run() would have made.
 733                          */
 734
 735                         if (atomic_dec_return(&ch->kthreads_assigned) == 0 &&
 736                             atomic_dec_return(&part->nchannels_engaged) == 0) {
 737                                 xpc_indicate_partition_disengaged(part);
 738                         }
 739                         xpc_msgqueue_deref(ch);
 740                         xpc_part_deref(part);
 741
 742                         if (atomic_read(&ch->kthreads_assigned) <
 743                             ch->kthreads_idle_limit) {
 744                                 /*
 745                                  * Flag this as an error only if we have an
 746                                  * insufficient #of kthreads for the channel
 747                                  * to function.
 748                                  */
 749                                 spin_lock_irqsave(&ch->lock, irq_flags);
 750                                 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources,
 751                                                        &irq_flags);
 752                                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 753                         }
 754                         break;
 755                 }
 756         }
 757 }
 758
 759 void
 760 xpc_disconnect_wait(int ch_number)
 761 {
 762         unsigned long irq_flags;
 763         short partid;
 764         struct xpc_partition *part;
 765         struct xpc_channel *ch;
 766         int wakeup_channel_mgr;
 767
 768         /* now wait for all callouts to the caller's function to cease */
 769         for (partid = 0; partid < xp_max_npartitions; partid++) {
 770                 part = &xpc_partitions[partid];
 771
 772                 if (!xpc_part_ref(part))
 773                         continue;
 774
 775                 ch = &part->channels[ch_number];
 776
 777                 if (!(ch->flags & XPC_C_WDISCONNECT)) {
 778                         xpc_part_deref(part);
 779                         continue;
 780                 }
 781
 782                 wait_for_completion(&ch->wdisconnect_wait);
 783
 784                 spin_lock_irqsave(&ch->lock, irq_flags);
 785                 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED));
 786                 wakeup_channel_mgr = 0;
 787
 788                 if (ch->delayed_chctl_flags) {
 789                         if (part->act_state != XPC_P_AS_DEACTIVATING) {
 790                                 spin_lock(&part->chctl_lock);
 791                                 part->chctl.flags[ch->number] |=
 792                                     ch->delayed_chctl_flags;
 793                                 spin_unlock(&part->chctl_lock);
 794                                 wakeup_channel_mgr = 1;
 795                         }
 796                         ch->delayed_chctl_flags = 0;
 797                 }
 798
 799                 ch->flags &= ~XPC_C_WDISCONNECT;
 800                 spin_unlock_irqrestore(&ch->lock, irq_flags);
 801
 802                 if (wakeup_channel_mgr)
 803                         xpc_wakeup_channel_mgr(part);
 804
 805                 xpc_part_deref(part);
 806         }
 807 }
 808
 809 static void
 810 xpc_do_exit(enum xp_retval reason)
 811 {
 812         short partid;
 813         int active_part_count, printed_waiting_msg = 0;
 814         struct xpc_partition *part;
 815         unsigned long printmsg_time, disengage_timeout = 0;
 816
 817         /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */
 818         DBUG_ON(xpc_exiting == 1);
 819
 820         /*
 821          * Let the heartbeat checker thread and the discovery thread
 822          * (if one is running) know that they should exit. Also wake up
 823          * the heartbeat checker thread in case it's sleeping.
 824          */
 825         xpc_exiting = 1;
 826         wake_up_interruptible(&xpc_activate_IRQ_wq);
 827
 828         /* wait for the discovery thread to exit */
 829         wait_for_completion(&xpc_discovery_exited);
 830
 831         /* wait for the heartbeat checker thread to exit */
 832         wait_for_completion(&xpc_hb_checker_exited);
 833
 834         /* sleep for a 1/3 of a second or so */
 835         (void)msleep_interruptible(300);
 836
 837         /* wait for all partitions to become inactive */
 838
 839         printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
 840         xpc_disengage_timedout = 0;
 841
 842         do {
 843                 active_part_count = 0;
 844
 845                 for (partid = 0; partid < xp_max_npartitions; partid++) {
 846                         part = &xpc_partitions[partid];
 847
 848                         if (xpc_partition_disengaged(part) &&
 849                             part->act_state == XPC_P_AS_INACTIVE) {
 850                                 continue;
 851                         }
 852
 853                         active_part_count++;
 854
 855                         XPC_DEACTIVATE_PARTITION(part, reason);
 856
 857                         if (part->disengage_timeout > disengage_timeout)
 858                                 disengage_timeout = part->disengage_timeout;
 859                 }
 860
 861                 if (xpc_any_partition_engaged()) {
 862                         if (time_is_before_jiffies(printmsg_time)) {
 863                                 dev_info(xpc_part, "waiting for remote "
 864                                          "partitions to deactivate, timeout in "
 865                                          "%ld seconds\n", (disengage_timeout -
 866                                          jiffies) / HZ);
 867                                 printmsg_time = jiffies +
 868                                     (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ);
 869                                 printed_waiting_msg = 1;
 870                         }
 871
 872                 } else if (active_part_count > 0) {
 873                         if (printed_waiting_msg) {
 874                                 dev_info(xpc_part, "waiting for local partition"
 875                                          " to deactivate\n");
 876                                 printed_waiting_msg = 0;
 877                         }
 878
 879                 } else {
 880                         if (!xpc_disengage_timedout) {
 881                                 dev_info(xpc_part, "all partitions have "
 882                                          "deactivated\n");
 883                         }
 884                         break;
 885                 }
 886
 887                 /* sleep for a 1/3 of a second or so */
 888                 (void)msleep_interruptible(300);
 889
 890         } while (1);
 891
 892         DBUG_ON(xpc_any_partition_engaged());
 893         DBUG_ON(xpc_any_hbs_allowed() != 0);
 894
 895         /* a zero timestamp indicates our rsvd page is not initialized */
 896         xpc_rsvd_page->ts_jiffies = 0;
 897
 898         if (reason == xpUnloading) {
 899                 (void)unregister_die_notifier(&xpc_die_notifier);
 900                 (void)unregister_reboot_notifier(&xpc_reboot_notifier);
 901         }
 902
 903         /* clear the interface to XPC's functions */
 904         xpc_clear_interface();
 905
 906         if (xpc_sysctl)
 907                 unregister_sysctl_table(xpc_sysctl);
 908
 909         kfree(xpc_partitions);
 910
 911         if (is_shub())
 912                 xpc_exit_sn2();
 913         else
 914                 xpc_exit_uv();
 915 }
 916
 917 /*
 918  * This function is called when the system is being rebooted.
 919  */
 920 static int
 921 xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused)
 922 {
 923         enum xp_retval reason;
 924
 925         switch (event) {
 926         case SYS_RESTART:
 927                 reason = xpSystemReboot;
 928                 break;
 929         case SYS_HALT:
 930                 reason = xpSystemHalt;
 931                 break;
 932         case SYS_POWER_OFF:
 933                 reason = xpSystemPoweroff;
 934                 break;
 935         default:
 936                 reason = xpSystemGoingDown;
 937         }
 938
 939         xpc_do_exit(reason);
 940         return NOTIFY_DONE;
 941 }
 942
 943 /*
 944  * Notify other partitions to deactivate from us by first disengaging from all
 945  * references to our memory.
 946  */
 947 static void
 948 xpc_die_deactivate(void)
 949 {
 950         struct xpc_partition *part;
 951         short partid;
 952         int any_engaged;
 953         long keep_waiting;
 954         long wait_to_print;
 955
 956         /* keep xpc_hb_checker thread from doing anything (just in case) */
 957         xpc_exiting = 1;
 958
 959         xpc_disallow_all_hbs(); /*indicate we're deactivated */
 960
 961         for (partid = 0; partid < xp_max_npartitions; partid++) {
 962                 part = &xpc_partitions[partid];
 963
 964                 if (xpc_partition_engaged(partid) ||
 965                     part->act_state != XPC_P_AS_INACTIVE) {
 966                         xpc_request_partition_deactivation(part);
 967                         xpc_indicate_partition_disengaged(part);
 968                 }
 969         }
 970
 971         /*
 972          * Though we requested that all other partitions deactivate from us,
 973          * we only wait until they've all disengaged or we've reached the
 974          * defined timelimit.
 975          *
 976          * Given that one iteration through the following while-loop takes
 977          * approximately 200 microseconds, calculate the #of loops to take
 978          * before bailing and the #of loops before printing a waiting message.
 979          */
 980         keep_waiting = xpc_disengage_timelimit * 1000 * 5;
 981         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5;
 982
 983         while (1) {
 984                 any_engaged = xpc_any_partition_engaged();
 985                 if (!any_engaged) {
 986                         dev_info(xpc_part, "all partitions have deactivated\n");
 987                         break;
 988                 }
 989
 990                 if (!keep_waiting--) {
 991                         for (partid = 0; partid < xp_max_npartitions;
 992                              partid++) {
 993                                 if (xpc_partition_engaged(partid)) {
 994                                         dev_info(xpc_part, "deactivate from "
 995                                                  "remote partition %d timed "
 996                                                  "out\n", partid);
 997                                 }
 998                         }
 999                         break;
1000                 }
1001
1002                 if (!wait_to_print--) {
1003                         dev_info(xpc_part, "waiting for remote partitions to "
1004                                  "deactivate, timeout in %ld seconds\n",
1005                                  keep_waiting / (1000 * 5));
1006                         wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL *
1007                             1000 * 5;
1008                 }
1009
1010                 udelay(200);
1011         }
1012 }
1013
1014 /*
1015  * This function is called when the system is being restarted or halted due
1016  * to some sort of system failure. If this is the case we need to notify the
1017  * other partitions to disengage from all references to our memory.
1018  * This function can also be called when our heartbeater could be offlined
1019  * for a time. In this case we need to notify other partitions to not worry
1020  * about the lack of a heartbeat.
1021  */
1022 static int
1023 xpc_system_die(struct notifier_block *nb, unsigned long event, void *unused)
1024 {
1025 #ifdef CONFIG_IA64              /* !!! temporary kludge */
1026         switch (event) {
1027         case DIE_MACHINE_RESTART:
1028         case DIE_MACHINE_HALT:
1029                 xpc_die_deactivate();
1030                 break;
1031
1032         case DIE_KDEBUG_ENTER:
1033                 /* Should lack of heartbeat be ignored by other partitions? */
1034                 if (!xpc_kdebug_ignore)
1035                         break;
1036
1037                 /* fall through */
1038         case DIE_MCA_MONARCH_ENTER:
1039         case DIE_INIT_MONARCH_ENTER:
1040                 xpc_offline_heartbeat();
1041                 break;
1042
1043         case DIE_KDEBUG_LEAVE:
1044                 /* Is lack of heartbeat being ignored by other partitions? */
1045                 if (!xpc_kdebug_ignore)
1046                         break;
1047
1048                 /* fall through */
1049         case DIE_MCA_MONARCH_LEAVE:
1050         case DIE_INIT_MONARCH_LEAVE:
1051                 xpc_online_heartbeat();
1052                 break;
1053         }
1054 #else
1055         xpc_die_deactivate();
1056 #endif
1057
1058         return NOTIFY_DONE;
1059 }
1060
1061 int __init
1062 xpc_init(void)
1063 {
1064         int ret;
1065         short partid;
1066         struct xpc_partition *part;
1067         struct task_struct *kthread;
1068
1069         snprintf(xpc_part->bus_id, BUS_ID_SIZE, "part");
1070         snprintf(xpc_chan->bus_id, BUS_ID_SIZE, "chan");
1071
1072         if (is_shub()) {
1073                 /*
1074                  * The ia64-sn2 architecture supports at most 64 partitions.
1075                  * And the inability to unregister remote amos restricts us
1076                  * further to only support exactly 64 partitions on this
1077                  * architecture, no less.
1078                  */
1079                 if (xp_max_npartitions != 64)
1080                         return -EINVAL;
1081
1082                 ret = xpc_init_sn2();
1083                 if (ret != 0)
1084                         return ret;
1085
1086         } else if (is_uv()) {
1087                 xpc_init_uv();
1088
1089         } else {
1090                 return -ENODEV;
1091         }
1092
1093         xpc_partitions = kzalloc(sizeof(struct xpc_partition) *
1094                                  xp_max_npartitions, GFP_KERNEL);
1095         if (xpc_partitions == NULL) {
1096                 dev_err(xpc_part, "can't get memory for partition structure\n");
1097                 ret = -ENOMEM;
1098                 goto out_1;
1099         }
1100
1101         /*
1102          * The first few fields of each entry of xpc_partitions[] need to
1103          * be initialized now so that calls to xpc_connect() and
1104          * xpc_disconnect() can be made prior to the activation of any remote
1105          * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE
1106          * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING
1107          * PARTITION HAS BEEN ACTIVATED.
1108          */
1109         for (partid = 0; partid < xp_max_npartitions; partid++) {
1110                 part = &xpc_partitions[partid];
1111
1112                 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part));
1113
1114                 part->activate_IRQ_rcvd = 0;
1115                 spin_lock_init(&part->act_lock);
1116                 part->act_state = XPC_P_AS_INACTIVE;
1117                 XPC_SET_REASON(part, 0, 0);
1118
1119                 init_timer(&part->disengage_timer);
1120                 part->disengage_timer.function =
1121                     xpc_timeout_partition_disengage;
1122                 part->disengage_timer.data = (unsigned long)part;
1123
1124                 part->setup_state = XPC_P_SS_UNSET;
1125                 init_waitqueue_head(&part->teardown_wq);
1126                 atomic_set(&part->references, 0);
1127         }
1128
1129         xpc_sysctl = register_sysctl_table(xpc_sys_dir);
1130
1131         /*
1132          * Fill the partition reserved page with the information needed by
1133          * other partitions to discover we are alive and establish initial
1134          * communications.
1135          */
1136         xpc_rsvd_page = xpc_setup_rsvd_page();
1137         if (xpc_rsvd_page == NULL) {
1138                 dev_err(xpc_part, "can't setup our reserved page\n");
1139                 ret = -EBUSY;
1140                 goto out_2;
1141         }
1142
1143         /* add ourselves to the reboot_notifier_list */
1144         ret = register_reboot_notifier(&xpc_reboot_notifier);
1145         if (ret != 0)
1146                 dev_warn(xpc_part, "can't register reboot notifier\n");
1147
1148         /* add ourselves to the die_notifier list */
1149         ret = register_die_notifier(&xpc_die_notifier);
1150         if (ret != 0)
1151                 dev_warn(xpc_part, "can't register die notifier\n");
1152
1153         /*
1154          * The real work-horse behind xpc.  This processes incoming
1155          * interrupts and monitors remote heartbeats.
1156          */
1157         kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME);
1158         if (IS_ERR(kthread)) {
1159                 dev_err(xpc_part, "failed while forking hb check thread\n");
1160                 ret = -EBUSY;
1161                 goto out_3;
1162         }
1163
1164         /*
1165          * Startup a thread that will attempt to discover other partitions to
1166          * activate based on info provided by SAL. This new thread is short
1167          * lived and will exit once discovery is complete.
1168          */
1169         kthread = kthread_run(xpc_initiate_discovery, NULL,
1170                               XPC_DISCOVERY_THREAD_NAME);
1171         if (IS_ERR(kthread)) {
1172                 dev_err(xpc_part, "failed while forking discovery thread\n");
1173
1174                 /* mark this new thread as a non-starter */
1175                 complete(&xpc_discovery_exited);
1176
1177                 xpc_do_exit(xpUnloading);
1178                 return -EBUSY;
1179         }
1180
1181         /* set the interface to point at XPC's functions */
1182         xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect,
1183                           xpc_initiate_send, xpc_initiate_send_notify,
1184                           xpc_initiate_received, xpc_initiate_partid_to_nasids);
1185
1186         return 0;
1187
1188         /* initialization was not successful */
1189 out_3:
1190         /* a zero timestamp indicates our rsvd page is not initialized */
1191         xpc_rsvd_page->ts_jiffies = 0;
1192
1193         (void)unregister_die_notifier(&xpc_die_notifier);
1194         (void)unregister_reboot_notifier(&xpc_reboot_notifier);
1195 out_2:
1196         if (xpc_sysctl)
1197                 unregister_sysctl_table(xpc_sysctl);
1198         kfree(xpc_partitions);
1199 out_1:
1200         if (is_shub())
1201                 xpc_exit_sn2();
1202         else
1203                 xpc_exit_uv();
1204         return ret;
1205 }
1206
1207 module_init(xpc_init);
1208
1209 void __exit
1210 xpc_exit(void)
1211 {
1212         xpc_do_exit(xpUnloading);
1213 }
1214
1215 module_exit(xpc_exit);
1216
1217 MODULE_AUTHOR("Silicon Graphics, Inc.");
1218 MODULE_DESCRIPTION("Cross Partition Communication (XPC) support");
1219 MODULE_LICENSE("GPL");
1220
1221 module_param(xpc_hb_interval, int, 0);
1222 MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between "
1223                  "heartbeat increments.");
1224
1225 module_param(xpc_hb_check_interval, int, 0);
1226 MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between "
1227                  "heartbeat checks.");
1228
1229 module_param(xpc_disengage_timelimit, int, 0);
1230 MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait "
1231                  "for disengage to complete.");
1232
1233 module_param(xpc_kdebug_ignore, int, 0);
1234 MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by "
1235                  "other partitions when dropping into kdebug.");