lnet/klnds/gnilnd/gnilnd_stack.c

   1 /*
   2  * Copyright (C) 2012 Cray, Inc.
   3  *
   4  *   Author: Nic Henke <nic@cray.com>
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22 #include "gnilnd.h"
  23 #if defined(GNILND_USE_RCA)
  24 #include <rsms/rs_sm_states.h>
  25 #endif
  26 /* Advance all timeouts by nap_time seconds. */
  27 void
  28 kgnilnd_bump_timeouts(__u32 nap_time, char *reason)
  29 {
  30         int                     i;
  31         kgn_peer_t             *peer;
  32         kgn_conn_t             *conn;
  33         kgn_tx_t               *tx;
  34         kgn_device_t           *dev;
  35         kgn_dgram_t            *dgram;
  36
  37         LCONSOLE_INFO("%s: bumping all timeouts by %ds\n", reason, nap_time);
  38
  39         LASSERTF(GNILND_IS_QUIESCED, "gnilnd not quiesced %d != %d\n",
  40                  atomic_read(&kgnilnd_data.kgn_nquiesce),
  41                  atomic_read(&kgnilnd_data.kgn_nthreads));
  42
  43         /* requiring that the threads are paused ensures a couple of things:
  44          * - combined code paths for stack reset and quiesce event as stack reset
  45          *   runs with the threads paused
  46          * - prevents traffic to the Gemini during a quiesce period
  47          * - reduces the locking requirements
  48         */
  49
  50         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
  51                 list_for_each_entry(peer, &kgnilnd_data.kgn_peers[i], gnp_list) {
  52
  53                         /* we can reconnect again at any time */
  54                         peer->gnp_reconnect_time = jiffies;
  55                         /* reset now that network is healthy */
  56                         peer->gnp_reconnect_interval = 0;
  57                         /* tell LNet dude is still alive */
  58                         kgnilnd_peer_alive(peer);
  59
  60                         list_for_each_entry(tx, &peer->gnp_tx_queue, tx_list) {
  61                                 tx->tx_qtime = jiffies;
  62                         }
  63
  64                         list_for_each_entry(conn, &peer->gnp_conns, gnc_list) {
  65                                 unsigned long           timeout;
  66
  67                                 timeout = cfs_time_seconds(conn->gnc_timeout);
  68
  69                                 /* bump last_rx/last_rx_cq on all conns - including
  70                                  * closed ones, this will have the effect of
  71                                  * bumping the purgatory timers for those */
  72                                 conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;
  73
  74                                 /* we don't timeout based on old gnc_last_tx, so
  75                                  * we'll back it up and schedule the conn to trigger
  76                                  * a NOOP */
  77                                 conn->gnc_last_tx = jiffies - timeout;
  78                                 if (conn->gnc_state != GNILND_CONN_DONE)
  79                                 kgnilnd_schedule_conn(conn);
  80                         }
  81                 }
  82         }
  83
  84         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
  85                 dev = &kgnilnd_data.kgn_devices[i];
  86                 for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
  87                         list_for_each_entry(dgram, &dev->gnd_dgrams[i], gndg_list) {
  88                                 dgram->gndg_post_time = jiffies;
  89                         }
  90                 }
  91         }
  92 }
  93
  94 /* Quiesce or wake up the stack.  The caller must hold the kgn_quiesce_sem semaphore
  95  * on entry, which holds off any pending stack shutdown.   */
  96 void
  97 kgnilnd_quiesce_wait(char *reason)
  98 {
  99         int             i;
 100
 101         if (kgnilnd_data.kgn_quiesce_trigger) {
 102                 unsigned long   quiesce_deadline, quiesce_to;
 103                 /* FREEZE TAG!!!! */
 104
 105                 /* morning sunshine */
 106                 spin_lock(&kgnilnd_data.kgn_reaper_lock);
 107                 wake_up_all(&kgnilnd_data.kgn_reaper_waitq);
 108                 spin_unlock(&kgnilnd_data.kgn_reaper_lock);
 109
 110                 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 111                         kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
 112
 113                         wake_up_all(&dev->gnd_waitq);
 114                         wake_up_all(&dev->gnd_dgram_waitq);
 115                         wake_up_all(&dev->gnd_dgping_waitq);
 116                 }
 117
 118                 kgnilnd_wakeup_rca_thread();
 119
 120                 /* we'll wait for 10x the timeout for the threads to pause */
 121                 quiesce_to = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * 10);
 122                 quiesce_deadline = (long) jiffies + quiesce_to;
 123
 124                 /* wait for everyone to check-in as quiesced */
 125                 i = 1;
 126                 while (!GNILND_IS_QUIESCED) {
 127                         i++;
 128                         LCONSOLE((((i) & (-i)) == i) ? D_WARNING : D_NET,
 129                                  "%s: Waiting for %d threads to pause\n",
 130                                  reason,
 131                                  atomic_read(&kgnilnd_data.kgn_nthreads) -
 132                                  atomic_read(&kgnilnd_data.kgn_nquiesce));
 133                         CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);
 134                         cfs_pause(cfs_time_seconds(1 * i));
 135
 136                         LASSERTF(quiesce_deadline > jiffies,
 137                                  "couldn't quiesce threads in %lu seconds, falling over now\n",
 138                                  cfs_duration_sec(quiesce_to));
 139                 }
 140
 141                 LCONSOLE_WARN("%s: All threads paused!\n", reason);
 142                 /* XXX Nic: Is there a set of counters we can grab here to
 143                  * ensure that there is no traffic until quiesce is over ?*/
 144         } else {
 145                 /* GO! GO! GO! */
 146
 147                 for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 148                         kgn_device_t *dev = &kgnilnd_data.kgn_devices[i];
 149                         kgnilnd_schedule_dgram(dev);
 150                 }
 151
 152                 /* wait for everyone to check-in as running - they will be spinning
 153                  * and looking, so no need to poke any waitq */
 154                 i = 1;
 155                 while (atomic_read(&kgnilnd_data.kgn_nquiesce) > 0) {
 156                         i++;
 157                         LCONSOLE((((i) & (-i)) == i) ? D_WARNING : D_NET,
 158                                  "%s: Waiting for %d threads to wake up\n",
 159                                   reason,
 160                                   atomic_read(&kgnilnd_data.kgn_nquiesce));
 161                         cfs_pause(cfs_time_seconds(1 * i));
 162                 }
 163
 164                 LCONSOLE_WARN("%s: All threads awake!\n", reason);
 165         }
 166 }
 167
 168 /* Reset the stack.  */
 169 void
 170 kgnilnd_reset_stack(void)
 171 {
 172         int              i, rc = 0;
 173         kgn_net_t       *net;
 174         kgn_peer_t      *peer, *peerN;
 175         LIST_HEAD        (souls);
 176         char            *reason = "critical hardware error";
 177         __u32            seconds;
 178         unsigned long    start, end;
 179         ENTRY;
 180
 181         /* Race with del_peer and its atomics */
 182         CFS_RACE(CFS_FAIL_GNI_RACE_RESET);
 183
 184         if (kgnilnd_data.kgn_init != GNILND_INIT_ALL) {
 185                 CERROR("can't reset the stack, gnilnd is not initialized\n");
 186                 RETURN_EXIT;
 187         }
 188
 189         /* First make sure we are not already quiesced - we panic if so,
 190          * as that could leave software in a bad state */
 191         LASSERTF(kgnilnd_data.kgn_quiesce_trigger == GNILND_QUIESCE_IDLE,
 192                 "can't reset the stack, already doing so: trigger %d\n",
 193                  kgnilnd_data.kgn_quiesce_trigger);
 194
 195         set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_RESET);
 196
 197         /* wake up the dgram waitq thread - but after trigger set to make sure it
 198          * goes into quiesce */
 199         CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);
 200         /* same for scheduler that is dropping state transitiosn */
 201         CFS_RACE(CFS_FAIL_GNI_DROP_CLOSING);
 202         CFS_RACE(CFS_FAIL_GNI_DROP_DESTROY_EP);
 203
 204         kgnilnd_quiesce_wait(reason);
 205
 206         start = jiffies;
 207
 208         kgnilnd_data.kgn_in_reset = 1;
 209         kgnilnd_data.kgn_nresets++;
 210         LCONSOLE_WARN("%s: resetting all resources (count %d)\n",
 211                       reason, kgnilnd_data.kgn_nresets);
 212
 213         for (i = 0; i < *kgnilnd_tunables.kgn_net_hash_size; i++) {
 214                 list_for_each_entry(net, &kgnilnd_data.kgn_nets[i], gnn_list) {
 215                         rc = kgnilnd_cancel_net_dgrams(net);
 216                         LASSERTF(rc == 0, "couldn't cleanup datagrams: %d\n", rc);
 217                 }
 218         }
 219
 220         /* error -ENOTRECOVERABLE is stack reset */
 221         kgnilnd_del_conn_or_peer(NULL, LNET_NID_ANY, GNILND_DEL_CONN, -ENOTRECOVERABLE);
 222
 223         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 224                 kgn_device_t    *dev = &kgnilnd_data.kgn_devices[i];
 225                 kgnilnd_cancel_wc_dgrams(dev);
 226                 kgnilnd_wait_for_canceled_dgrams(dev);
 227         }
 228
 229         /* manually do some conn processing ala kgnilnd_process_conns */
 230         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 231                 kgn_device_t    *dev = &kgnilnd_data.kgn_devices[i];
 232                 kgn_conn_t      *conn;
 233                 int              conn_sched;
 234
 235                 /* go find all the closed conns that need to be nuked - the
 236                  * scheduler thread isn't running to do this for us */
 237
 238                 CDEBUG(D_NET, "will try to clear up %d ready_conns\n",
 239                         kgnilnd_count_list(&dev->gnd_ready_conns));
 240
 241                 /* use while/list_first_entry loop to ensure we can handle any
 242                  * DESTROY_EP conns added from kgnilnd_complete_closed_conn */
 243                 while (!list_empty(&dev->gnd_ready_conns)) {
 244                         conn = list_first_entry(&dev->gnd_ready_conns,
 245                                                 kgn_conn_t, gnc_schedlist);
 246                         conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_PROCESS);
 247
 248                         LASSERTF(conn_sched != GNILND_CONN_IDLE &&
 249                                  conn_sched != GNILND_CONN_PROCESS,
 250                                  "conn %p on ready list but in bad state: %d\n",
 251                                  conn, conn_sched);
 252
 253                         list_del_init(&conn->gnc_schedlist);
 254
 255                         if (conn->gnc_state == GNILND_CONN_CLOSING) {
 256                                 /* bump to CLOSED to fake out send of CLOSE */
 257                                 conn->gnc_state = GNILND_CONN_CLOSED;
 258                                 conn->gnc_close_sent = 1;
 259                         }
 260
 261                         if (conn->gnc_state == GNILND_CONN_DESTROY_EP) {
 262                                 kgnilnd_destroy_conn_ep(conn);
 263                         } else {
 264                                 kgnilnd_complete_closed_conn(conn);
 265                         }
 266
 267                         /* there really shouldn't be any other states here -
 268                          * they would have been cleared out in the del_peer_or_conn or the dgram
 269                          * aborts above.
 270                          * there is an LASSERTF in kgnilnd_complete_closed_conn that will take
 271                          * care of catching anything else for us */
 272
 273                         kgnilnd_schedule_process_conn(conn, -1);
 274
 275                         kgnilnd_conn_decref(conn);
 276                 }
 277         }
 278
 279         /* don't let the little weasily purgatory conns hide from us */
 280         for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
 281                 list_for_each_entry_safe(peer, peerN, &kgnilnd_data.kgn_peers[i], gnp_list) {
 282                         kgn_conn_t       *conn, *connN;
 283
 284                         list_for_each_entry_safe(conn, connN, &peer->gnp_conns, gnc_list) {
 285                                 kgnilnd_detach_purgatory_locked(conn, &souls);
 286                         }
 287                 }
 288         }
 289
 290         CDEBUG(D_NET, "about to release %d purgatory entries\n",
 291                 kgnilnd_count_list(&souls));
 292
 293         kgnilnd_release_purgatory_list(&souls);
 294
 295         /* validate we are now clean */
 296         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 297                 kgn_device_t    *dev = &kgnilnd_data.kgn_devices[i];
 298
 299                 /* now all the cons/mboxes should be cleaned up, including purgatory
 300                  * so go through and release the MDDs for our persistent PHYS fma_blks
 301                  */
 302                 kgnilnd_unmap_fma_blocks(dev);
 303
 304                 LASSERTF(atomic_read(&dev->gnd_nfmablk) == 0,
 305                         "reset failed: fma blocks still live %d\n",
 306                         atomic_read(&dev->gnd_nfmablk));
 307
 308                 LASSERTF(atomic_read(&dev->gnd_neps) == 0,
 309                         "reset failed: EP handles still live %d\n",
 310                         atomic_read(&dev->gnd_neps));
 311         }
 312
 313         LASSERTF(atomic_read(&kgnilnd_data.kgn_nconns) == 0,
 314                 "reset failed: conns left %d\n",
 315                 atomic_read(&kgnilnd_data.kgn_nconns));
 316
 317         /* fine to have peers left - they are waiting for new conns
 318          * but should not be holding any open HW resources */
 319
 320         /* like the last part of kgnilnd_base_shutdown() */
 321
 322         CFS_RACE(CFS_FAIL_GNI_SR_DOWN_RACE);
 323
 324         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 325                 kgnilnd_dev_fini(&kgnilnd_data.kgn_devices[i]);
 326         }
 327
 328         /* no need to free and recreate the TX descriptors
 329          * we nuked all the ones that could be using HW resources in
 330          * kgnilnd_close_matching_conns and asserted it worked in
 331          * kgnilnd_dev_fini */
 332
 333         /* At this point, all HW is torn down, start to reset */
 334
 335         /* only reset our known devs */
 336         for (i = 0; i < kgnilnd_data.kgn_ndevs; i++) {
 337                 kgn_device_t    *dev = &kgnilnd_data.kgn_devices[i];
 338                 rc = kgnilnd_dev_init(dev);
 339                 LASSERTF(rc == 0, "dev_init failed for dev %d\n", i);
 340                 kgnilnd_map_phys_fmablk(dev);
 341                 LASSERTF(rc == 0, "map_phys_fmablk failed for dev %d\n", i);
 342                 rc = kgnilnd_setup_wildcard_dgram(dev);
 343                 LASSERTF(rc == 0, "couldnt setup datagrams on dev %d: %d\n",
 344                         i, rc);
 345         }
 346
 347         /* Now the fun restarts... - release the hounds! */
 348
 349         end = jiffies;
 350         seconds = cfs_duration_sec((long)end - start);
 351         kgnilnd_bump_timeouts(seconds, reason);
 352
 353         kgnilnd_data.kgn_in_reset = 0;
 354         set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
 355         kgnilnd_quiesce_wait(reason);
 356         LCONSOLE_WARN("%s reset of all hardware resources\n",
 357                 rc ? "failed" : "successful");
 358
 359         RETURN_EXIT;
 360 }
 361
 362 /* A thread that handles quiece and reset hardware events.
 363  * We do the same thing regardless of which device reported the event. */
 364 int
 365 kgnilnd_ruhroh_thread(void *arg)
 366 {
 367         int                i = 1;
 368         DEFINE_WAIT(wait);
 369
 370         cfs_block_allsigs();
 371         set_user_nice(current, *kgnilnd_tunables.kgn_nice);
 372         kgnilnd_data.kgn_ruhroh_running = 1;
 373
 374         while (1) {
 375
 376                 /* Block until there's a request..  A reset request could come in
 377                  * while we're handling a quiesce one, or vice versa.
 378                  * Keep processing requests until there are none.*/
 379                 prepare_to_wait(&kgnilnd_data.kgn_ruhroh_waitq, &wait, TASK_INTERRUPTIBLE);
 380                 while (!(kgnilnd_data.kgn_ruhroh_shutdown ||
 381                                 kgnilnd_data.kgn_needs_reset || kgnilnd_data.kgn_needs_pause))
 382                         schedule();
 383                 finish_wait(&kgnilnd_data.kgn_ruhroh_waitq, &wait);
 384
 385                /* Exit if the driver is shutting down. */
 386                 if (kgnilnd_data.kgn_ruhroh_shutdown)
 387                         break;
 388
 389                 /* Serialize with driver startup and shutdown. */
 390                 mutex_lock(&kgnilnd_data.kgn_quiesce_mutex);
 391
 392                CDEBUG(D_NET, "trigger %d reset %d to_bump %d pause %d\n",
 393                         kgnilnd_data.kgn_quiesce_trigger,
 394                         kgnilnd_data.kgn_needs_reset,
 395                         kgnilnd_data.kgn_bump_info_rdy,
 396                         kgnilnd_data.kgn_needs_pause);
 397
 398                 /* Do we need to do a pause/quiesce? */
 399                 if (kgnilnd_data.kgn_needs_pause) {
 400
 401                         /* Pause all other kgnilnd threads. */
 402                         set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_HW_QUIESCE);
 403                         kgnilnd_quiesce_wait("hardware quiesce flag");
 404
 405                         /* If the hardware quiesce flag is set, wait for it to clear.
 406                          * This should happen relatively quickly, so we wait for it.
 407                          * This will hold up the eventd thread, but on everything but
 408                          * the simulator, this is ok-- there is one thread per core.
 409                          *
 410                          * Handle (possibly multiple) quiesce events while we wait. The
 411                          * memory barrier ensures that the core doesn't start fetching
 412                          * kgn_bump_info_rdy before it fetches kgn_needs_pause, and
 413                          * matches the second mb in kgnilnd_quiesce_end_callback(). */
 414                         smp_rmb();
 415                         while (kgnilnd_hw_in_quiesce() || kgnilnd_data.kgn_bump_info_rdy) {
 416
 417                                 i++;
 418                                 LCONSOLE((((i) & (-i)) == i) ? D_WARNING : D_NET,
 419                                                 "Waiting for hardware quiesce flag to clear\n");
 420                                 cfs_pause(cfs_time_seconds(1 * i));
 421
 422                                 /* If we got a quiesce event with bump info, DO THE BUMP!. */
 423                                 if (kgnilnd_data.kgn_bump_info_rdy) {
 424                                         /* reset console rate limiting for each event */
 425                                         i = 1;
 426
 427                                         /* Make sure the core doesn't start fetching
 428                                          * kgni_quiesce_seconds until after it sees
 429                                          * kgn_bump_info_rdy set.  This is the match to the
 430                                          * first mb in kgnilnd_quiesce_end_callback(). */
 431                                         smp_rmb();
 432                                         (void) kgnilnd_bump_timeouts(kgnilnd_data.kgn_quiesce_secs,
 433                                                                "hardware quiesce callback");
 434                                         set_mb(kgnilnd_data.kgn_quiesce_secs, 0);
 435                                         set_mb(kgnilnd_data.kgn_bump_info_rdy, 0);
 436                                 }
 437                       }
 438
 439                         /* Reset the kgn_needs_pause flag before coming out of
 440                          * the pause.  This ordering avoids a race with the
 441                          * setting of this flag in kgnilnd_pause_threads().  */
 442                         set_mb(kgnilnd_data.kgn_needs_pause, 0);
 443
 444                         /* ok, let the kids back into the pool */
 445                         set_mb(kgnilnd_data.kgn_quiesce_trigger, GNILND_QUIESCE_IDLE);
 446                         kgnilnd_quiesce_wait("hardware quiesce");
 447                 }
 448
 449                 /* Do a stack reset if needed. */
 450                 if (kgnilnd_data.kgn_needs_reset) {
 451                         kgnilnd_reset_stack();
 452                         set_mb(kgnilnd_data.kgn_needs_reset, 0);
 453                 }
 454
 455                 mutex_unlock(&kgnilnd_data.kgn_quiesce_mutex);
 456         }
 457
 458         kgnilnd_data.kgn_ruhroh_running = 0;
 459         return 0;
 460 }
 461
 462 /* Set pause request flag.  Any functions that
 463  * call this one are responsible for ensuring that
 464  * variables they set up are visible on other cores before
 465  * this flag setting.  This executes in interrupt or kernel
 466  * thread context.  */
 467 void
 468 kgnilnd_pause_threads(void)
 469 {
 470         /* only device 0 gets the handle, see kgnilnd_dev_init */
 471         kgn_device_t  *dev = &kgnilnd_data.kgn_devices[0];
 472         LASSERTF(dev != NULL, "dev 0 is NULL\n");
 473
 474         /* If we're currently in a pause triggered by the pause flag,
 475          * there's no need to set it again.  We clear the kgn_needs_pause
 476          * flag before we reset kgn_quiesce_trigger to avoid a race.  The
 477          * read memory barrier matches the setmb() on the trigger in
 478          * kgnilnd_ruhroh_task().                                       */
 479         smp_rmb();
 480         if (!(kgnilnd_data.kgn_quiesce_trigger == GNILND_QUIESCE_HW_QUIESCE &&
 481                         GNILND_IS_QUIESCED)) {
 482                  CDEBUG(D_NET, "requesting thread pause\n");
 483
 484                 kgnilnd_data.kgn_needs_pause = 1;
 485
 486                 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
 487         } else {
 488             CDEBUG(D_NET, "thread pause already underway\n");
 489         }
 490 }
 491
 492 /* Return non-zero if the GNI hardware quiesce flag is set */
 493 int
 494 kgnilnd_hw_in_quiesce(void)
 495 {
 496         /* only device 0 gets the handle, see kgnilnd_dev_init */
 497         kgn_device_t      *dev0 = &kgnilnd_data.kgn_devices[0];
 498
 499         LASSERTF(dev0 != NULL, "dev 0 is NULL\n");
 500
 501         smp_rmb();
 502         return kgnilnd_get_quiesce_status(dev0->gnd_handle) != 0;
 503 }
 504
 505
 506 /* If the GNI hardware quiesce flag is set, initiate our pause and
 507  * return non-zero.  Also return non-zero if the stack is shutting down. */
 508 int
 509 kgnilnd_check_hw_quiesce(void)
 510 {
 511         if (likely(!kgnilnd_hw_in_quiesce()))
 512                 return 0;
 513
 514         if (!kgnilnd_data.kgn_ruhroh_shutdown) {
 515                 CDEBUG(D_NET, "initiating thread pause\n");
 516                 kgnilnd_pause_threads();
 517         } else {
 518                 CDEBUG(D_NET, "thread pause bypassed because of shutdown\n");
 519         }
 520
 521         return 1;
 522 }
 523
 524 /* Callback from kngi with the quiesce duration.  This executes
 525  * in interrupt context.                                        */
 526 void
 527 kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs)
 528 {
 529         /* only device 0 gets the handle, see kgnilnd_dev_init */
 530         kgn_device_t  *dev = &kgnilnd_data.kgn_devices[0];
 531         LASSERTF(dev != NULL, "dev 0 is NULL\n");
 532
 533         if (!kgnilnd_data.kgn_ruhroh_shutdown) {
 534
 535                 CDEBUG(D_NET, "requesting timeout bump by "LPD64" msecs\n", msecs);
 536
 537                 /* Save the bump interval and request the bump.
 538                  * The memory barrier ensures that the interval is in place before
 539                  * the bump flag can be seen (in case a core is already running the
 540                  * ruhroh task), and that the bump request flag in place before
 541                  * the pause request can be seen (to ensure a core doesn't miss the bump
 542                  * request flag).       */
 543                 /* If another callback occurred before the ruhroh task
 544                  * finished processing the first bump request, we'd over-write its info.
 545                  * Nic says that callbacks occur so slowly that this isn't an issue.    */
 546                 set_mb(kgnilnd_data.kgn_quiesce_secs, msecs / MSEC_PER_SEC);
 547                 set_mb(kgnilnd_data.kgn_bump_info_rdy, 1);
 548                 kgnilnd_pause_threads();
 549         } else {
 550                 CDEBUG(D_NET, "timeout bump bypassed because of shutdown\n");
 551         }
 552 }
 553
 554 void
 555 kgnilnd_critical_error(struct gni_err *err_handle)
 556 {
 557         /* only device 0 gets the handle, see kgnilnd_dev_init */
 558         kgn_device_t  *dev = &kgnilnd_data.kgn_devices[0];
 559         LASSERTF(dev != NULL, "dev 0 is NULL\n");
 560
 561         if (!kgnilnd_data.kgn_ruhroh_shutdown) {
 562                 CDEBUG(D_NET, "requesting stack reset\n");
 563                 kgnilnd_data.kgn_needs_reset = 1;
 564                 wake_up(&kgnilnd_data.kgn_ruhroh_waitq);
 565         } else {
 566                 CDEBUG(D_NET, "stack reset bypassed because of shutdown\n");
 567         }
 568 }
 569
 570 #if defined(GNILND_USE_RCA)
 571 #include <krca_lib.h>
 572 #define RCA_EVENTS 3
 573 /* RCA ticket is needed for krca_wakeup_wait_event() */
 574 static krca_ticket_t rca_krt = KRCA_NULL_TICKET;
 575 struct rcadata {
 576         rca_ticket_t ticket;
 577         int subscribed;
 578         rs_event_code_t ec;
 579 };
 580 static struct rcadata rd[RCA_EVENTS] = {
 581         {0, 0, ec_node_unavailable},
 582         {0, 0, ec_node_available},
 583         {0, 0, ec_node_failed}
 584 };
 585
 586 /* thread for receiving rca events */
 587 int
 588 kgnilnd_rca(void *arg)
 589 {
 590         int        i, rc;
 591         int        retry_count;
 592         rs_event_t event;
 593         lnet_nid_t nid;
 594
 595         cfs_block_allsigs();
 596
 597         /* all gnilnd threads need to run fairly urgently */
 598         set_user_nice(current, *kgnilnd_tunables.kgn_nice);
 599
 600         /*
 601          * Register our service with RCA and subscribe to events
 602          * of interest.
 603          */
 604         rca_krt = KRCA_NULL_TICKET;
 605         rc = krca_register(&rca_krt, RCA_SVCTYPE_GNILND, current->pid, 0);
 606         if (rc < 0) {
 607                 CNETERR("krca_register(%x) returned %d\n", current->pid, rc);
 608                 goto done;
 609         }
 610
 611         for (i = 0; i < RCA_EVENTS; i++) {
 612                 retry_count = 0;
 613 subscribe_retry:
 614                 rc = krca_subscribe(&rca_krt, rd[i].ec, RCA_RX_SVC_ANY,
 615                                     &rd[i].ticket);
 616
 617                 if ((rc == -EINTR) && !retry_count) {
 618                         retry_count++;
 619                         CNETERR("krca_subscribe returned %d - retrying\n", rc);
 620                         goto subscribe_retry;
 621                 }
 622
 623                 if (rc < 0) {
 624                         CNETERR("rca subscription failed (%d)\n", rc);
 625                         goto done;
 626                 }
 627
 628                 rd[i].subscribed = 1;
 629         }
 630
 631         while (!kgnilnd_data.kgn_shutdown) {
 632                 if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
 633                         KGNILND_SPIN_QUIESCE;
 634                 }
 635                 /* wait here for a subscribed event */
 636                 rc = krca_wait_event(&rca_krt);
 637
 638                 /* RCA return values:
 639                  * 0 indicates krca_wakeup_wait_event caused krca_wait_event
 640                  *   return.
 641                  * -ERESTARTSYS indicates krca_wait_event returned because of a
 642                  *   signal.
 643                  * -ENOSPC indicates no space available to create an rcad_reg_t
 644                  * 1 indicates a message is waiting.
 645                  */
 646                 if (rc <= 0) {
 647                         continue;
 648                 }
 649
 650                 if (krca_get_message(&rca_krt, &event) == 0) {
 651                         int node_down = GNILND_RCA_NODE_UNKNOWN;
 652                         rs_state_t state;
 653                         LIST_HEAD(zombies);
 654
 655                         /* Compute nodes don't care about other compute nodes
 656                          * so we don't need to create a peer.
 657                          */
 658                         if (GNILND_COMPUTE &&
 659                             !RSN_GET_FLD(event.ev_gen.svid_node.rs_node_flat,
 660                                         IS_SVC)) {
 661                                 continue;
 662                         }
 663
 664                         /* Only care about compute and service nodes not GPUs */
 665                         if (RSN_GET_FLD(event.ev_gen.svid_node.rs_node_flat,
 666                                         TYPE) != rt_node) {
 667                                 continue;
 668                         }
 669
 670                         switch (event.ev_id) {
 671                         case ec_node_available:
 672                                 CDEBUG(D_INFO, "ec_node_available\n");
 673                                 node_down = GNILND_RCA_NODE_UP;
 674                                 break;
 675                         case ec_node_failed:
 676                                 CDEBUG(D_INFO, "ec_node_failed\n");
 677                                 if (event.ev_len > 0) {
 678                                         CDEBUG(D_ERROR,
 679                                                 "ec_node_failed ignored\n");
 680                                         break;
 681                                 }
 682                                 node_down = GNILND_RCA_NODE_DOWN;
 683                                 break;
 684                         case ec_node_unavailable:
 685                                 state = RSN_GET_FLD(event.ev_gen.svid_node.rsn_intval, STATE);
 686
 687                                 CDEBUG(D_INFO, "ec_node_unavailable\n");
 688
 689                                 /*
 690                                  * Ignore overloaded ec_node_unavailable events
 691                                  * generated by 'xtcli set_reserve'.
 692                                  */
 693                                 if (RS_GET_CS_STATE(state) == RS_CS_READY) {
 694                                         CDEBUG(D_INFO, "ignoring "
 695                                                 "ec_node_unavailable event with"
 696                                                 " RS_CS_READY state\n");
 697                                         break;
 698                                 }
 699                                 node_down = GNILND_RCA_NODE_DOWN;
 700                                 break;
 701                         default:
 702                                 CDEBUG(D_INFO, "unknown event\n");
 703                                 break;
 704                         }
 705
 706                         /* if we get an event we don't know about, just go ahead
 707                          * and wait for another event */
 708                         if (node_down == GNILND_RCA_NODE_UNKNOWN) {
 709                                 continue;
 710                         }
 711
 712                         nid = RSN_GET_FLD(event.ev_gen.svid_node.rs_node_flat,
 713                                           NID);
 714                         CDEBUG(D_INFO,"kgnilnd_rca() reporting nid %d %s\n",
 715                                (int)nid, node_down ? "down" : "up");
 716                         kgnilnd_report_node_state(nid, node_down);
 717
 718                 } else {
 719                         CNETERR("krca_get_message failed\n");
 720                 }
 721         }
 722
 723 done:
 724         CDEBUG(D_INFO, "done\n");
 725
 726         for (i = 0; i < RCA_EVENTS; i++) {
 727                 if (rd[i].subscribed) {
 728                         rc = krca_unsubscribe(&rca_krt, rd[i].ticket);
 729
 730                         if (rc) {
 731                                 CNETERR("rca unsubscribe failed (%d)\n", rc);
 732                         }
 733
 734                         rd[i].subscribed = 0;
 735                 }
 736         }
 737
 738         krca_unregister(&rca_krt);
 739         kgnilnd_thread_fini();
 740         return 0;
 741
 742 }
 743
 744 int
 745 kgnilnd_start_rca_thread(void)
 746 {
 747         return kgnilnd_thread_start(kgnilnd_rca, NULL, "kgnilnd_rca", 0);
 748 }
 749
 750 void
 751 kgnilnd_wakeup_rca_thread(void)
 752 {
 753         int ret;
 754
 755         ret = krca_wakeup_wait_event(&rca_krt);
 756
 757         if (ret) {
 758                 CDEBUG(D_ERROR, "krca_wakeup_wait_event failed\n");
 759         }
 760 }
 761
 762 int
 763 kgnilnd_get_node_state(__u32 nid)
 764 {
 765         int i;
 766         int rc = GNILND_RCA_NODE_UNKNOWN;
 767         int ret;
 768         rs_node_array_t nlist;
 769         rs_node_t       *na = NULL;
 770
 771         if ((ret = krca_get_sysnodes(&nlist)) < 0) {
 772                 CDEBUG(D_NETERROR, "krca_get_sysnodes failed %d\n", ret);
 773                 goto ns_done;
 774         }
 775
 776         na = nlist.na_ids;
 777
 778         for (i = 0; i < nlist.na_len; i++) {
 779                 if ((rca_nid_t)RSN_GET_FLD(na[i].rs_node_flat, NID) == nid) {
 780                         rc = RSN_GET_FLD(na[i].rs_node_flat, STATE) == RS_CS_READY ?
 781                                 GNILND_RCA_NODE_UP : GNILND_RCA_NODE_DOWN;
 782                         break;
 783                 }
 784         }
 785
 786 ns_done:
 787         kfree(na);
 788         CDEBUG(D_NET, "nid %d rc %d (0=up)\n", nid, rc);
 789         return rc;
 790 }
 791
 792 #else /* GNILND_USE_RCA */
 793
 794 int
 795 kgnilnd_start_rca_thread(void)
 796 {
 797         return 0;
 798 }
 799
 800 void
 801 kgnilnd_wakeup_rca_thread(void)
 802 {
 803 }
 804
 805 int
 806 kgnilnd_get_node_state(__u32 nid)
 807 {
 808         return GNILND_RCA_NODE_UP;
 809 }
 810 #endif /* GNILND_USE_RCA */