lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Portal-RPC reconnection and replay operations, for use in recovery.
   5  *
   6  *  Copyright (c) 2003 Cluster File Systems, Inc.
   7  *   Authors: Phil Schwan <phil@clusterfs.com>
   8  *            Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef __KERNEL__
  27 #include <liblustre.h>
  28 #else
  29 #include <linux/version.h>
  30 #include <asm/semaphore.h>
  31 #define DEBUG_SUBSYSTEM S_RPC
  32 #endif
  33
  34 #include <linux/obd_support.h>
  35 #include <linux/obd_class.h>
  36 #include "ptlrpc_internal.h"
  37
  38 #define PINGER_RATE     3 /* how many pings we'll do in obd_timeout period */
  39
  40 static DECLARE_MUTEX(pinger_sem);
  41 static struct list_head pinger_imports = LIST_HEAD_INIT(pinger_imports);
  42
  43 int ptlrpc_ping(struct obd_import *imp)
  44 {
  45         struct ptlrpc_request *req;
  46         int rc = 0;
  47         ENTRY;
  48
  49         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 0, NULL, NULL);
  50         if (req) {
  51                 DEBUG_REQ(D_HA, req, "pinging %s->%s",
  52                           imp->imp_obd->obd_uuid.uuid,
  53                           imp->imp_target_uuid.uuid);
  54                 req->rq_no_resend = req->rq_no_delay = 1;
  55                 req->rq_replen = lustre_msg_size(0, NULL);
  56                 req->rq_timeout = obd_timeout / PINGER_RATE;
  57                 ptlrpcd_add_req(req);
  58         } else {
  59                 CERROR("OOM trying to ping %s->%s\n",
  60                        imp->imp_obd->obd_uuid.uuid,
  61                        imp->imp_target_uuid.uuid);
  62                 rc = -ENOMEM;
  63         }
  64
  65         RETURN(rc);
  66 }
  67
  68 #ifdef __KERNEL__
  69 static inline int ptlrpc_next_ping(struct obd_import *imp)
  70 {
  71         return jiffies + (obd_timeout / PINGER_RATE * HZ);
  72 }
  73
  74 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
  75 {
  76         if (imp->imp_server_timeout)
  77                 return jiffies + (obd_timeout / 2 * HZ);
  78         else
  79                 return jiffies + (obd_timeout * HZ);
  80 }
  81
  82 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
  83 static wait_queue_head_t suspend_timeouts_waitq;
  84
  85 void ptlrpc_deactivate_timeouts(void)
  86 {
  87         CDEBUG(D_HA, "deactivate timeouts\n");
  88         atomic_inc(&suspend_timeouts);
  89 }
  90
  91 void ptlrpc_activate_timeouts(void)
  92 {
  93         CDEBUG(D_HA, "activate timeouts\n");
  94         LASSERT(atomic_read(&suspend_timeouts) > 0);
  95         if (atomic_dec_and_test(&suspend_timeouts))
  96                 wake_up(&suspend_timeouts_waitq);
  97 }
  98
  99 int ptlrpc_check_suspend(void)
 100 {
 101         if (atomic_read(&suspend_timeouts))
 102                 return 1;
 103         return 0;
 104 }
 105
 106 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 107 {
 108         struct l_wait_info lwi;
 109
 110         if (atomic_read(&suspend_timeouts)) {
 111                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 112                           atomic_read(&suspend_timeouts));
 113                 lwi = LWI_INTR(NULL, NULL);
 114                 l_wait_event(suspend_timeouts_waitq,
 115                              atomic_read(&suspend_timeouts) == 0, &lwi);
 116                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 117                 return 1;
 118         }
 119         return 0;
 120 }
 121
 122 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 123                                          unsigned long this_ping)
 124 {
 125         unsigned long flags;
 126         int force, level;
 127
 128         spin_lock_irqsave(&imp->imp_lock, flags);
 129         level = imp->imp_state;
 130         force = imp->imp_force_verify;
 131         if (force)
 132                 imp->imp_force_verify = 0;
 133         spin_unlock_irqrestore(&imp->imp_lock, flags);
 134
 135         if (imp->imp_next_ping - this_ping > obd_timeout * HZ) {
 136                 CWARN("wrong ping time %lu (current %lu)\n",
 137                       imp->imp_next_ping, this_ping);
 138                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 139         }
 140
 141         if (imp->imp_next_ping > this_ping && force == 0)
 142                 return;
 143
 144         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 145                 /* wait at least a timeout before trying recovery again */
 146                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 147                 ptlrpc_initiate_recovery(imp);
 148         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov) {
 149                 CDEBUG(D_HA, "not pinging %s (in recovery "
 150                        " or recovery disabled: %s)\n",
 151                        imp->imp_target_uuid.uuid,
 152                        ptlrpc_import_state_name(level));
 153         } else if (imp->imp_pingable || force) {
 154                 imp->imp_next_ping = ptlrpc_next_ping(imp);
 155                 ptlrpc_ping(imp);
 156         }
 157 }
 158
 159 static int ptlrpc_pinger_main(void *arg)
 160 {
 161         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 162         struct ptlrpc_thread *thread = data->thread;
 163         unsigned long flags;
 164         ENTRY;
 165
 166         lock_kernel();
 167         ptlrpc_daemonize();
 168
 169         SIGNAL_MASK_LOCK(current, flags);
 170         sigfillset(&current->blocked);
 171         RECALC_SIGPENDING;
 172         SIGNAL_MASK_UNLOCK(current, flags);
 173
 174         LASSERTF(strlen(data->name) < sizeof(current->comm),
 175                  "name %d > len %d\n",
 176                  (int)strlen(data->name), (int)sizeof(current->comm));
 177         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 178         unlock_kernel();
 179
 180         /* Record that the thread is running */
 181         thread->t_flags = SVC_RUNNING;
 182         wake_up(&thread->t_ctl_waitq);
 183
 184         /* And now, loop forever, pinging as needed. */
 185         while (1) {
 186                 unsigned long this_ping = jiffies;
 187                 long time_to_next_ping;
 188                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
 189                                                      NULL, NULL);
 190                 struct list_head *iter;
 191
 192                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 193                 down(&pinger_sem);
 194                 list_for_each(iter, &pinger_imports) {
 195                         struct obd_import *imp =
 196                                 list_entry(iter, struct obd_import,
 197                                            imp_pinger_chain);
 198
 199                         ptlrpc_pinger_process_import(imp, this_ping);
 200
 201                         CDEBUG(D_OTHER, "%s: pingable %d, next_ping %lu(%lu)\n",
 202                                 imp->imp_target_uuid.uuid,
 203                                 imp->imp_pingable, imp->imp_next_ping, jiffies);
 204
 205                         if (imp->imp_pingable && imp->imp_next_ping &&
 206                             imp->imp_next_ping - jiffies < time_to_next_ping &&
 207                             imp->imp_next_ping > jiffies)
 208                                 time_to_next_ping = imp->imp_next_ping - jiffies;
 209                 }
 210                 up(&pinger_sem);
 211
 212                 /* Wait until the next ping time, or until we're stopped. */
 213                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 214                        this_ping + (obd_timeout * HZ));
 215                 if (time_to_next_ping > 0) {
 216                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 217                         l_wait_event(thread->t_ctl_waitq,
 218                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 219                                      &lwi);
 220                         if (thread->t_flags & SVC_STOPPING) {
 221                                 thread->t_flags &= ~SVC_STOPPING;
 222                                 EXIT;
 223                                 break;
 224                         } else if (thread->t_flags & SVC_EVENT) {
 225                                 /* woken after adding import to reset timer */
 226                                 thread->t_flags &= ~SVC_EVENT;
 227                         }
 228                 }
 229         }
 230
 231         thread->t_flags = SVC_STOPPED;
 232         wake_up(&thread->t_ctl_waitq);
 233
 234         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 235         return 0;
 236 }
 237
 238 static struct ptlrpc_thread *pinger_thread = NULL;
 239
 240 int ptlrpc_start_pinger(void)
 241 {
 242         struct l_wait_info lwi = { 0 };
 243         struct ptlrpc_svc_data d;
 244         int rc;
 245 #ifndef ENABLE_PINGER
 246         return 0;
 247 #endif
 248         ENTRY;
 249
 250         LASSERT(obd_timeout > PINGER_RATE);
 251
 252         if (pinger_thread != NULL)
 253                 RETURN(-EALREADY);
 254
 255         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 256         if (pinger_thread == NULL)
 257                 RETURN(-ENOMEM);
 258         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 259         init_waitqueue_head(&suspend_timeouts_waitq);
 260
 261         d.name = "ll_ping";
 262         d.thread = pinger_thread;
 263
 264         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 265          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 266         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 267         if (rc < 0) {
 268                 CERROR("cannot start thread: %d\n", rc);
 269                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 270                 RETURN(rc);
 271         }
 272         l_wait_event(pinger_thread->t_ctl_waitq,
 273                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 274
 275         RETURN(rc);
 276 }
 277
 278 int ptlrpc_stop_pinger(void)
 279 {
 280         struct l_wait_info lwi = { 0 };
 281         int rc = 0;
 282 #ifndef ENABLE_PINGER
 283         return 0;
 284 #endif
 285         ENTRY;
 286
 287         if (pinger_thread == NULL)
 288                 RETURN(-EALREADY);
 289         down(&pinger_sem);
 290         pinger_thread->t_flags = SVC_STOPPING;
 291         wake_up(&pinger_thread->t_ctl_waitq);
 292         up(&pinger_sem);
 293
 294         l_wait_event(pinger_thread->t_ctl_waitq,
 295                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 296
 297         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 298         pinger_thread = NULL;
 299         RETURN(rc);
 300 }
 301
 302 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 303 {
 304         down(&pinger_sem);
 305         imp->imp_next_ping = ptlrpc_next_ping(imp);
 306         up(&pinger_sem);
 307 }
 308
 309 int ptlrpc_pinger_add_import(struct obd_import *imp)
 310 {
 311         ENTRY;
 312         if (!list_empty(&imp->imp_pinger_chain))
 313                 RETURN(-EALREADY);
 314
 315         down(&pinger_sem);
 316         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 317                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 318         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 319         /* XXX sort, blah blah */
 320         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 321         class_import_get(imp);
 322
 323         ptlrpc_pinger_wake_up();
 324         up(&pinger_sem);
 325
 326         RETURN(0);
 327 }
 328
 329 int ptlrpc_pinger_del_import(struct obd_import *imp)
 330 {
 331         ENTRY;
 332         if (list_empty(&imp->imp_pinger_chain))
 333                 RETURN(-ENOENT);
 334
 335         down(&pinger_sem);
 336         list_del_init(&imp->imp_pinger_chain);
 337         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 338                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 339         class_import_put(imp);
 340         up(&pinger_sem);
 341         RETURN(0);
 342 }
 343
 344 void ptlrpc_pinger_wake_up()
 345 {
 346 #ifdef ENABLE_PINGER
 347         pinger_thread->t_flags |= SVC_EVENT;
 348         wake_up(&pinger_thread->t_ctl_waitq);
 349 #endif
 350 }
 351
 352 #else /* !__KERNEL__ */
 353
 354 /* XXX
 355  * the current implementation of pinger in liblustre is not optimized
 356  */
 357
 358 static struct pinger_data {
 359         int             pd_recursion;
 360         unsigned long   pd_this_ping;
 361         unsigned long   pd_next_ping;
 362         int             pd_force_check;
 363 } pinger_args;
 364
 365 static int pinger_check_rpcs(void *arg)
 366 {
 367         unsigned long curtime = time(NULL);
 368         struct list_head *iter;
 369         struct pinger_data *pd = &pinger_args;
 370
 371         /* prevent recursion */
 372         if (pd->pd_recursion++) {
 373                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 374                 pd->pd_recursion--;
 375                 return 0;
 376         }
 377
 378         /* have we reached ping point? */
 379         if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
 380                 pd->pd_recursion--;
 381                 return 0;
 382         }
 383
 384         if (pd->pd_force_check)
 385                 pd->pd_force_check = 0;
 386
 387         pd->pd_this_ping = curtime;
 388
 389         /* add rpcs into set */
 390         down(&pinger_sem);
 391         list_for_each(iter, &pinger_imports) {
 392                 struct obd_import *imp =
 393                         list_entry(iter, struct obd_import,
 394                                    imp_pinger_chain);
 395                 int level, force;
 396                 unsigned long flags;
 397
 398
 399                 spin_lock_irqsave(&imp->imp_lock, flags);
 400                 level = imp->imp_state;
 401                 force = imp->imp_force_verify;
 402                 if (force)
 403                         imp->imp_force_verify = 0;
 404                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 405
 406                 if (imp->imp_next_ping <= pd->pd_this_ping || force) {
 407                         if (level == LUSTRE_IMP_DISCON) {
 408                                 /* wait at least a timeout before
 409                                    trying recovery again. */
 410                                 unsigned long timeout = obd_timeout;
 411                                 if (imp->imp_server_timeout)
 412                                         timeout = obd_timeout / 2;
 413                                 imp->imp_next_ping = time(NULL) +
 414                                         (timeout * HZ);
 415                                 ptlrpc_initiate_recovery(imp);
 416                         }
 417                         else if (level != LUSTRE_IMP_FULL ||
 418                                  imp->imp_obd->obd_no_recov) {
 419                                 CDEBUG(D_HA,
 420                                        "not pinging %s (in recovery "
 421                                        " or recovery disabled: %s)\n",
 422                                        imp->imp_target_uuid.uuid,
 423                                        ptlrpc_import_state_name(level));
 424                         }
 425                         else if (imp->imp_pingable || force) {
 426                                 ptlrpc_ping(imp);
 427                         }
 428
 429                 } else {
 430                         if (imp->imp_pingable) {
 431                                 CDEBUG(D_HA, "don't need to ping %s "
 432                                        "(%lu > %lu)\n",
 433                                        imp->imp_target_uuid.uuid,
 434                                        imp->imp_next_ping, pd->pd_this_ping);
 435                         }
 436                 }
 437         }
 438
 439         up(&pinger_sem);
 440
 441         pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
 442
 443         CDEBUG(D_HA, "finished a round ping\n");
 444         pd->pd_recursion--;
 445         return 0;
 446 }
 447
 448 static void *pinger_callback = NULL;
 449
 450 int ptlrpc_start_pinger(void)
 451 {
 452         memset(&pinger_args, 0, sizeof(pinger_args));
 453 #ifdef ENABLE_PINGER
 454         pinger_callback =
 455                 liblustre_register_wait_callback(&pinger_check_rpcs, &pinger_args);
 456 #endif
 457         return 0;
 458 }
 459
 460 int ptlrpc_stop_pinger(void)
 461 {
 462 #ifdef ENABLE_PINGER
 463         if (pinger_callback)
 464                 liblustre_deregister_wait_callback(pinger_callback);
 465 #endif
 466         return 0;
 467 }
 468
 469 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 470 {
 471         down(&pinger_sem);
 472         imp->imp_next_ping = time(NULL) + obd_timeout;
 473         if (pinger_args.pd_next_ping > imp->imp_next_ping) {
 474                 CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
 475                         imp->imp_next_ping, time(NULL));
 476                 pinger_args.pd_next_ping = imp->imp_next_ping;
 477         }
 478         up(&pinger_sem);
 479 }
 480
 481 int ptlrpc_pinger_add_import(struct obd_import *imp)
 482 {
 483         ENTRY;
 484         if (!list_empty(&imp->imp_pinger_chain))
 485                 RETURN(-EALREADY);
 486
 487         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 488                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 489         ptlrpc_pinger_sending_on_import(imp);
 490
 491         down(&pinger_sem);
 492         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 493         class_import_get(imp);
 494         up(&pinger_sem);
 495
 496         RETURN(0);
 497 }
 498
 499 int ptlrpc_pinger_del_import(struct obd_import *imp)
 500 {
 501         ENTRY;
 502         if (list_empty(&imp->imp_pinger_chain))
 503                 RETURN(-ENOENT);
 504
 505         down(&pinger_sem);
 506         list_del_init(&imp->imp_pinger_chain);
 507         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 508                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 509         class_import_put(imp);
 510         up(&pinger_sem);
 511         RETURN(0);
 512 }
 513
 514 void ptlrpc_pinger_wake_up()
 515 {
 516         pinger_args.pd_force_check = 1;
 517 }
 518 #endif /* !__KERNEL__ */