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 > 0) &&
 136                         (imp->imp_next_ping - this_ping > obd_timeout * HZ)) {
 137                 CWARN("wrong ping time %lu (current %lu)\n",
 138                       imp->imp_next_ping, this_ping);
 139                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 140         }
 141
 142         if (imp->imp_next_ping > this_ping && force == 0)
 143                 return;
 144
 145         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 146                 /* wait at least a timeout before trying recovery again */
 147                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 148                 ptlrpc_initiate_recovery(imp);
 149         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov) {
 150                 CDEBUG(D_HA, "not pinging %s (in recovery "
 151                        " or recovery disabled: %s)\n",
 152                        imp->imp_target_uuid.uuid,
 153                        ptlrpc_import_state_name(level));
 154         } else if (imp->imp_pingable || force) {
 155                 imp->imp_next_ping = ptlrpc_next_ping(imp);
 156                 ptlrpc_ping(imp);
 157         }
 158 }
 159
 160 static int ptlrpc_pinger_main(void *arg)
 161 {
 162         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 163         struct ptlrpc_thread *thread = data->thread;
 164         unsigned long flags;
 165         ENTRY;
 166
 167         lock_kernel();
 168         ptlrpc_daemonize();
 169
 170         SIGNAL_MASK_LOCK(current, flags);
 171         sigfillset(&current->blocked);
 172         RECALC_SIGPENDING;
 173         SIGNAL_MASK_UNLOCK(current, flags);
 174
 175         LASSERTF(strlen(data->name) < sizeof(current->comm),
 176                  "name %d > len %d\n",
 177                  (int)strlen(data->name), (int)sizeof(current->comm));
 178         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 179         unlock_kernel();
 180
 181         /* Record that the thread is running */
 182         thread->t_flags = SVC_RUNNING;
 183         wake_up(&thread->t_ctl_waitq);
 184
 185         /* And now, loop forever, pinging as needed. */
 186         while (1) {
 187                 unsigned long this_ping = jiffies;
 188                 long time_to_next_ping;
 189                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
 190                                                      NULL, NULL);
 191                 struct list_head *iter;
 192
 193                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 194                 down(&pinger_sem);
 195                 list_for_each(iter, &pinger_imports) {
 196                         struct obd_import *imp =
 197                                 list_entry(iter, struct obd_import,
 198                                            imp_pinger_chain);
 199
 200                         ptlrpc_pinger_process_import(imp, this_ping);
 201
 202                         CDEBUG(D_OTHER, "%s: pingable %d, next_ping %lu(%lu)\n",
 203                                 imp->imp_target_uuid.uuid,
 204                                 imp->imp_pingable, imp->imp_next_ping, jiffies);
 205
 206                         if (imp->imp_pingable && imp->imp_next_ping &&
 207                             imp->imp_next_ping - jiffies < time_to_next_ping &&
 208                             imp->imp_next_ping > jiffies)
 209                                 time_to_next_ping = imp->imp_next_ping - jiffies;
 210                 }
 211                 up(&pinger_sem);
 212
 213                 /* Wait until the next ping time, or until we're stopped. */
 214                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 215                        this_ping + (obd_timeout * HZ));
 216                 if (time_to_next_ping > 0) {
 217                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 218                         l_wait_event(thread->t_ctl_waitq,
 219                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 220                                      &lwi);
 221                         if (thread->t_flags & SVC_STOPPING) {
 222                                 thread->t_flags &= ~SVC_STOPPING;
 223                                 EXIT;
 224                                 break;
 225                         } else if (thread->t_flags & SVC_EVENT) {
 226                                 /* woken after adding import to reset timer */
 227                                 thread->t_flags &= ~SVC_EVENT;
 228                         }
 229                 }
 230         }
 231
 232         thread->t_flags = SVC_STOPPED;
 233         wake_up(&thread->t_ctl_waitq);
 234
 235         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 236         return 0;
 237 }
 238
 239 static struct ptlrpc_thread *pinger_thread = NULL;
 240
 241 int ptlrpc_start_pinger(void)
 242 {
 243         struct l_wait_info lwi = { 0 };
 244         struct ptlrpc_svc_data d;
 245         int rc;
 246 #ifndef ENABLE_PINGER
 247         return 0;
 248 #endif
 249         ENTRY;
 250
 251         LASSERT(obd_timeout > PINGER_RATE);
 252
 253         if (pinger_thread != NULL)
 254                 RETURN(-EALREADY);
 255
 256         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 257         if (pinger_thread == NULL)
 258                 RETURN(-ENOMEM);
 259         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 260         init_waitqueue_head(&suspend_timeouts_waitq);
 261
 262         d.name = "ll_ping";
 263         d.thread = pinger_thread;
 264
 265         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 266          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 267         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 268         if (rc < 0) {
 269                 CERROR("cannot start thread: %d\n", rc);
 270                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 271                 RETURN(rc);
 272         }
 273         l_wait_event(pinger_thread->t_ctl_waitq,
 274                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 275
 276         RETURN(rc);
 277 }
 278
 279 int ptlrpc_stop_pinger(void)
 280 {
 281         struct l_wait_info lwi = { 0 };
 282         int rc = 0;
 283 #ifndef ENABLE_PINGER
 284         return 0;
 285 #endif
 286         ENTRY;
 287
 288         if (pinger_thread == NULL)
 289                 RETURN(-EALREADY);
 290         down(&pinger_sem);
 291         pinger_thread->t_flags = SVC_STOPPING;
 292         wake_up(&pinger_thread->t_ctl_waitq);
 293         up(&pinger_sem);
 294
 295         l_wait_event(pinger_thread->t_ctl_waitq,
 296                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 297
 298         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 299         pinger_thread = NULL;
 300         RETURN(rc);
 301 }
 302
 303 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 304 {
 305         down(&pinger_sem);
 306         imp->imp_next_ping = ptlrpc_next_ping(imp);
 307         up(&pinger_sem);
 308 }
 309
 310 int ptlrpc_pinger_add_import(struct obd_import *imp)
 311 {
 312         ENTRY;
 313         if (!list_empty(&imp->imp_pinger_chain))
 314                 RETURN(-EALREADY);
 315
 316         down(&pinger_sem);
 317         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 318                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 319         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 320         /* XXX sort, blah blah */
 321         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 322         class_import_get(imp);
 323
 324         ptlrpc_pinger_wake_up();
 325         up(&pinger_sem);
 326
 327         RETURN(0);
 328 }
 329
 330 int ptlrpc_pinger_del_import(struct obd_import *imp)
 331 {
 332         ENTRY;
 333         if (list_empty(&imp->imp_pinger_chain))
 334                 RETURN(-ENOENT);
 335
 336         down(&pinger_sem);
 337         list_del_init(&imp->imp_pinger_chain);
 338         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 339                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 340         class_import_put(imp);
 341         up(&pinger_sem);
 342         RETURN(0);
 343 }
 344
 345 void ptlrpc_pinger_wake_up()
 346 {
 347 #ifdef ENABLE_PINGER
 348         pinger_thread->t_flags |= SVC_EVENT;
 349         wake_up(&pinger_thread->t_ctl_waitq);
 350 #endif
 351 }
 352
 353 #else /* !__KERNEL__ */
 354
 355 /* XXX
 356  * the current implementation of pinger in liblustre is not optimized
 357  */
 358
 359 static struct pinger_data {
 360         int             pd_recursion;
 361         unsigned long   pd_this_ping;
 362         unsigned long   pd_next_ping;
 363         int             pd_force_check;
 364 } pinger_args;
 365
 366 static int pinger_check_rpcs(void *arg)
 367 {
 368         unsigned long curtime = time(NULL);
 369         struct list_head *iter;
 370         struct pinger_data *pd = &pinger_args;
 371
 372         /* prevent recursion */
 373         if (pd->pd_recursion++) {
 374                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 375                 pd->pd_recursion--;
 376                 return 0;
 377         }
 378
 379         /* have we reached ping point? */
 380         if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
 381                 pd->pd_recursion--;
 382                 return 0;
 383         }
 384
 385         if (pd->pd_force_check)
 386                 pd->pd_force_check = 0;
 387
 388         pd->pd_this_ping = curtime;
 389
 390         /* add rpcs into set */
 391         down(&pinger_sem);
 392         list_for_each(iter, &pinger_imports) {
 393                 struct obd_import *imp =
 394                         list_entry(iter, struct obd_import,
 395                                    imp_pinger_chain);
 396                 int level, force;
 397                 unsigned long flags;
 398
 399
 400                 spin_lock_irqsave(&imp->imp_lock, flags);
 401                 level = imp->imp_state;
 402                 force = imp->imp_force_verify;
 403                 if (force)
 404                         imp->imp_force_verify = 0;
 405                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 406
 407                 if (imp->imp_next_ping <= pd->pd_this_ping || force) {
 408                         if (level == LUSTRE_IMP_DISCON) {
 409                                 /* wait at least a timeout before
 410                                    trying recovery again. */
 411                                 unsigned long timeout = obd_timeout;
 412                                 if (imp->imp_server_timeout)
 413                                         timeout = obd_timeout / 2;
 414                                 imp->imp_next_ping = time(NULL) +
 415                                         (timeout * HZ);
 416                                 ptlrpc_initiate_recovery(imp);
 417                         }
 418                         else if (level != LUSTRE_IMP_FULL ||
 419                                  imp->imp_obd->obd_no_recov) {
 420                                 CDEBUG(D_HA,
 421                                        "not pinging %s (in recovery "
 422                                        " or recovery disabled: %s)\n",
 423                                        imp->imp_target_uuid.uuid,
 424                                        ptlrpc_import_state_name(level));
 425                         }
 426                         else if (imp->imp_pingable || force) {
 427                                 ptlrpc_ping(imp);
 428                         }
 429
 430                 } else {
 431                         if (imp->imp_pingable) {
 432                                 CDEBUG(D_HA, "don't need to ping %s "
 433                                        "(%lu > %lu)\n",
 434                                        imp->imp_target_uuid.uuid,
 435                                        imp->imp_next_ping, pd->pd_this_ping);
 436                         }
 437                 }
 438         }
 439
 440         up(&pinger_sem);
 441
 442         pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
 443
 444         CDEBUG(D_HA, "finished a round ping\n");
 445         pd->pd_recursion--;
 446         return 0;
 447 }
 448
 449 static void *pinger_callback = NULL;
 450
 451 int ptlrpc_start_pinger(void)
 452 {
 453         memset(&pinger_args, 0, sizeof(pinger_args));
 454 #ifdef ENABLE_PINGER
 455         pinger_callback =
 456                 liblustre_register_wait_callback(&pinger_check_rpcs, &pinger_args);
 457 #endif
 458         return 0;
 459 }
 460
 461 int ptlrpc_stop_pinger(void)
 462 {
 463 #ifdef ENABLE_PINGER
 464         if (pinger_callback)
 465                 liblustre_deregister_wait_callback(pinger_callback);
 466 #endif
 467         return 0;
 468 }
 469
 470 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 471 {
 472         down(&pinger_sem);
 473         imp->imp_next_ping = time(NULL) + obd_timeout;
 474         if (pinger_args.pd_next_ping > imp->imp_next_ping) {
 475                 CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
 476                         imp->imp_next_ping, time(NULL));
 477                 pinger_args.pd_next_ping = imp->imp_next_ping;
 478         }
 479         up(&pinger_sem);
 480 }
 481
 482 int ptlrpc_pinger_add_import(struct obd_import *imp)
 483 {
 484         ENTRY;
 485         if (!list_empty(&imp->imp_pinger_chain))
 486                 RETURN(-EALREADY);
 487
 488         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 489                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 490         ptlrpc_pinger_sending_on_import(imp);
 491
 492         down(&pinger_sem);
 493         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 494         class_import_get(imp);
 495         up(&pinger_sem);
 496
 497         RETURN(0);
 498 }
 499
 500 int ptlrpc_pinger_del_import(struct obd_import *imp)
 501 {
 502         ENTRY;
 503         if (list_empty(&imp->imp_pinger_chain))
 504                 RETURN(-ENOENT);
 505
 506         down(&pinger_sem);
 507         list_del_init(&imp->imp_pinger_chain);
 508         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 509                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 510         class_import_put(imp);
 511         up(&pinger_sem);
 512         RETURN(0);
 513 }
 514
 515 void ptlrpc_pinger_wake_up()
 516 {
 517         pinger_args.pd_force_check = 1;
 518 }
 519 #endif /* !__KERNEL__ */