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 static DECLARE_MUTEX(pinger_sem);
  39 static struct list_head pinger_imports = LIST_HEAD_INIT(pinger_imports);
  40
  41 int ptlrpc_ping(struct obd_import *imp)
  42 {
  43         struct ptlrpc_request *req;
  44         int rc = 0;
  45         ENTRY;
  46
  47         req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL,
  48                               NULL);
  49         if (req) {
  50                 DEBUG_REQ(D_HA, req, "pinging %s->%s",
  51                           imp->imp_obd->obd_uuid.uuid,
  52                           imp->imp_target_uuid.uuid);
  53                 req->rq_no_resend = req->rq_no_delay = 1;
  54                 req->rq_replen = lustre_msg_size(0,
  55                                                  NULL);
  56                 ptlrpcd_add_req(req);
  57         } else {
  58                 CERROR("OOM trying to ping %s->%s\n",
  59                           imp->imp_obd->obd_uuid.uuid,
  60                           imp->imp_target_uuid.uuid);
  61                 rc = -ENOMEM;
  62         }
  63
  64         RETURN(rc);
  65 }
  66
  67 #ifdef __KERNEL__
  68 int ptlrpc_next_ping(struct obd_import *imp)
  69 {
  70         if (imp->imp_server_timeout)
  71                 return jiffies + (obd_timeout / 4 * HZ);
  72         else
  73                 return jiffies + (obd_timeout / 2 * HZ);
  74 }
  75
  76 static int ptlrpc_pinger_main(void *arg)
  77 {
  78         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
  79         struct ptlrpc_thread *thread = data->thread;
  80         unsigned long flags;
  81         ENTRY;
  82
  83         lock_kernel();
  84         ptlrpc_daemonize();
  85
  86         SIGNAL_MASK_LOCK(current, flags);
  87         sigfillset(&current->blocked);
  88         RECALC_SIGPENDING;
  89         SIGNAL_MASK_UNLOCK(current, flags);
  90
  91         LASSERTF(strlen(data->name) < sizeof(current->comm),
  92                  "name %d > len %d\n",
  93                  (int)strlen(data->name), (int)sizeof(current->comm));
  94         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
  95         unlock_kernel();
  96
  97         /* Record that the thread is running */
  98         thread->t_flags = SVC_RUNNING;
  99         wake_up(&thread->t_ctl_waitq);
 100
 101         /* And now, loop forever, pinging as needed. */
 102         while (1) {
 103                 unsigned long this_ping = jiffies;
 104                 long time_to_next_ping;
 105                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
 106                                                      NULL, NULL);
 107                 struct list_head *iter;
 108
 109                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 110                 down(&pinger_sem);
 111                 list_for_each(iter, &pinger_imports) {
 112                         struct obd_import *imp =
 113                                 list_entry(iter, struct obd_import,
 114                                            imp_pinger_chain);
 115                         int force, level;
 116                         unsigned long flags;
 117
 118
 119                         spin_lock_irqsave(&imp->imp_lock, flags);
 120                         level = imp->imp_state;
 121                         force = imp->imp_force_verify;
 122                         if (force)
 123                                 imp->imp_force_verify = 0;
 124                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 125
 126                         if (imp->imp_next_ping <= this_ping || force) {
 127                                 if (level == LUSTRE_IMP_DISCON) {
 128                                         /* wait at least a timeout before
 129                                            trying recovery again. */
 130                                         imp->imp_next_ping =
 131                                                 ptlrpc_next_ping(imp);
 132                                         ptlrpc_initiate_recovery(imp);
 133                                 } else if (level != LUSTRE_IMP_FULL ||
 134                                            imp->imp_obd->obd_no_recov) {
 135                                         CDEBUG(D_HA,
 136                                                "not pinging %s (in recovery "
 137                                                " or recovery disabled: %s)\n",
 138                                                imp->imp_target_uuid.uuid,
 139                                                ptlrpc_import_state_name(level));
 140                                 } else if (imp->imp_pingable || force) {
 141                                         ptlrpc_ping(imp);
 142                                 }
 143
 144                         } else if (imp->imp_pingable) {
 145                                 CDEBUG(D_HA, "don't need to ping %s "
 146                                        "(%lu > %lu)\n",
 147                                        imp->imp_target_uuid.uuid,
 148                                        imp->imp_next_ping, this_ping);
 149                         }
 150                         CDEBUG(D_OTHER, "%s: pingable %d, next_ping %lu(%lu)\n",
 151                                 imp->imp_target_uuid.uuid,
 152                                 imp->imp_pingable, imp->imp_next_ping, jiffies);
 153                         if (imp->imp_pingable && imp->imp_next_ping &&
 154                             imp->imp_next_ping - jiffies < time_to_next_ping &&
 155                             imp->imp_next_ping > jiffies)
 156                                 time_to_next_ping = imp->imp_next_ping - jiffies;
 157                 }
 158                 up(&pinger_sem);
 159
 160                 /* Wait until the next ping time, or until we're stopped. */
 161                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 162                        this_ping + (obd_timeout * HZ));
 163                 if (time_to_next_ping > 0) {
 164                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 165                         l_wait_event(thread->t_ctl_waitq,
 166                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 167                                      &lwi);
 168                         if (thread->t_flags & SVC_STOPPING) {
 169                                 thread->t_flags &= ~SVC_STOPPING;
 170                                 EXIT;
 171                                 break;
 172                         } else if (thread->t_flags & SVC_EVENT) {
 173                                 /* woken after adding import to reset timer */
 174                                 thread->t_flags &= ~SVC_EVENT;
 175                         }
 176                 }
 177         }
 178
 179         thread->t_flags = SVC_STOPPED;
 180         wake_up(&thread->t_ctl_waitq);
 181
 182         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 183         return 0;
 184 }
 185
 186 static struct ptlrpc_thread *pinger_thread = NULL;
 187
 188 int ptlrpc_start_pinger(void)
 189 {
 190         struct l_wait_info lwi = { 0 };
 191         struct ptlrpc_svc_data d;
 192         int rc;
 193 #ifndef ENABLE_PINGER
 194         return 0;
 195 #endif
 196         ENTRY;
 197
 198         if (pinger_thread != NULL)
 199                 RETURN(-EALREADY);
 200
 201         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 202         if (pinger_thread == NULL)
 203                 RETURN(-ENOMEM);
 204         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 205
 206         d.name = "ll_ping";
 207         d.thread = pinger_thread;
 208
 209         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 210          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 211         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 212         if (rc < 0) {
 213                 CERROR("cannot start thread: %d\n", rc);
 214                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 215                 RETURN(rc);
 216         }
 217         l_wait_event(pinger_thread->t_ctl_waitq,
 218                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 219
 220         RETURN(rc);
 221 }
 222
 223 int ptlrpc_stop_pinger(void)
 224 {
 225         struct l_wait_info lwi = { 0 };
 226         int rc = 0;
 227 #ifndef ENABLE_PINGER
 228         return 0;
 229 #endif
 230         ENTRY;
 231
 232         if (pinger_thread == NULL)
 233                 RETURN(-EALREADY);
 234         down(&pinger_sem);
 235         pinger_thread->t_flags = SVC_STOPPING;
 236         wake_up(&pinger_thread->t_ctl_waitq);
 237         up(&pinger_sem);
 238
 239         l_wait_event(pinger_thread->t_ctl_waitq,
 240                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 241
 242         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 243         pinger_thread = NULL;
 244         RETURN(rc);
 245 }
 246
 247 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 248 {
 249         down(&pinger_sem);
 250         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 251         up(&pinger_sem);
 252 }
 253
 254 int ptlrpc_pinger_add_import(struct obd_import *imp)
 255 {
 256         ENTRY;
 257         if (!list_empty(&imp->imp_pinger_chain))
 258                 RETURN(-EALREADY);
 259
 260         down(&pinger_sem);
 261         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 262                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 263         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 264         /* XXX sort, blah blah */
 265         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 266         class_import_get(imp);
 267
 268         ptlrpc_pinger_wake_up();
 269         up(&pinger_sem);
 270
 271         RETURN(0);
 272 }
 273
 274 int ptlrpc_pinger_del_import(struct obd_import *imp)
 275 {
 276         ENTRY;
 277         if (list_empty(&imp->imp_pinger_chain))
 278                 RETURN(-ENOENT);
 279
 280         down(&pinger_sem);
 281         list_del_init(&imp->imp_pinger_chain);
 282         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 283                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 284         class_import_put(imp);
 285         up(&pinger_sem);
 286         RETURN(0);
 287 }
 288
 289 void ptlrpc_pinger_wake_up()
 290 {
 291 #ifdef ENABLE_PINGER
 292         pinger_thread->t_flags |= SVC_EVENT;
 293         wake_up(&pinger_thread->t_ctl_waitq);
 294 #endif
 295 }
 296
 297 #else /* !__KERNEL__ */
 298
 299 /* XXX
 300  * the current implementation of pinger in liblustre is not optimized
 301  */
 302
 303 static struct pinger_data {
 304         int             pd_recursion;
 305         unsigned long   pd_this_ping;
 306         unsigned long   pd_next_ping;
 307         int             pd_force_check;
 308 } pinger_args;
 309
 310 static int pinger_check_rpcs(void *arg)
 311 {
 312         unsigned long curtime = time(NULL);
 313         struct list_head *iter;
 314         struct pinger_data *pd = &pinger_args;
 315
 316         /* prevent recursion */
 317         if (pd->pd_recursion++) {
 318                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 319                 pd->pd_recursion--;
 320                 return 0;
 321         }
 322
 323         /* have we reached ping point? */
 324         if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
 325                 pd->pd_recursion--;
 326                 return 0;
 327         }
 328
 329         if (pd->pd_force_check)
 330                 pd->pd_force_check = 0;
 331
 332         pd->pd_this_ping = curtime;
 333
 334         /* add rpcs into set */
 335         down(&pinger_sem);
 336         list_for_each(iter, &pinger_imports) {
 337                 struct obd_import *imp =
 338                         list_entry(iter, struct obd_import,
 339                                    imp_pinger_chain);
 340                 int level, force;
 341                 unsigned long flags;
 342
 343
 344                 spin_lock_irqsave(&imp->imp_lock, flags);
 345                 level = imp->imp_state;
 346                 force = imp->imp_force_verify;
 347                 if (force)
 348                         imp->imp_force_verify = 0;
 349                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 350
 351                 if (imp->imp_next_ping <= pd->pd_this_ping || force) {
 352                         if (level == LUSTRE_IMP_DISCON) {
 353                                 /* wait at least a timeout before
 354                                    trying recovery again. */
 355                                 unsigned long timeout = obd_timeout;
 356                                 if (imp->imp_server_timeout)
 357                                         timeout = obd_timeout / 2;
 358                                 imp->imp_next_ping = time(NULL) +
 359                                         (timeout * HZ);
 360                                 ptlrpc_initiate_recovery(imp);
 361                         }
 362                         else if (level != LUSTRE_IMP_FULL ||
 363                                  imp->imp_obd->obd_no_recov) {
 364                                 CDEBUG(D_HA,
 365                                        "not pinging %s (in recovery "
 366                                        " or recovery disabled: %s)\n",
 367                                        imp->imp_target_uuid.uuid,
 368                                        ptlrpc_import_state_name(level));
 369                         }
 370                         else if (imp->imp_pingable || force) {
 371                                 ptlrpc_ping(imp);
 372                         }
 373
 374                 } else {
 375                         if (imp->imp_pingable) {
 376                                 CDEBUG(D_HA, "don't need to ping %s "
 377                                        "(%lu > %lu)\n",
 378                                        imp->imp_target_uuid.uuid,
 379                                        imp->imp_next_ping, pd->pd_this_ping);
 380                         }
 381                 }
 382         }
 383
 384         up(&pinger_sem);
 385
 386         pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
 387
 388         CDEBUG(D_HA, "finished a round ping\n");
 389         pd->pd_recursion--;
 390         return 0;
 391 }
 392
 393 static void *pinger_callback = NULL;
 394
 395 int ptlrpc_start_pinger(void)
 396 {
 397         memset(&pinger_args, 0, sizeof(pinger_args));
 398 #ifdef ENABLE_PINGER
 399         pinger_callback =
 400                 liblustre_register_wait_callback(&pinger_check_rpcs, &pinger_args);
 401 #endif
 402         return 0;
 403 }
 404
 405 int ptlrpc_stop_pinger(void)
 406 {
 407 #ifdef ENABLE_PINGER
 408         if (pinger_callback)
 409                 liblustre_deregister_wait_callback(pinger_callback);
 410 #endif
 411         return 0;
 412 }
 413
 414 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 415 {
 416         down(&pinger_sem);
 417         imp->imp_next_ping = time(NULL) + obd_timeout;
 418         if (pinger_args.pd_next_ping > imp->imp_next_ping) {
 419                 CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
 420                         imp->imp_next_ping, time(NULL));
 421                 pinger_args.pd_next_ping = imp->imp_next_ping;
 422         }
 423         up(&pinger_sem);
 424 }
 425
 426 int ptlrpc_pinger_add_import(struct obd_import *imp)
 427 {
 428         ENTRY;
 429         if (!list_empty(&imp->imp_pinger_chain))
 430                 RETURN(-EALREADY);
 431
 432         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 433                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 434         ptlrpc_pinger_sending_on_import(imp);
 435
 436         down(&pinger_sem);
 437         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 438         class_import_get(imp);
 439         up(&pinger_sem);
 440
 441         RETURN(0);
 442 }
 443
 444 int ptlrpc_pinger_del_import(struct obd_import *imp)
 445 {
 446         ENTRY;
 447         if (list_empty(&imp->imp_pinger_chain))
 448                 RETURN(-ENOENT);
 449
 450         down(&pinger_sem);
 451         list_del_init(&imp->imp_pinger_chain);
 452         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 453                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 454         class_import_put(imp);
 455         up(&pinger_sem);
 456         RETURN(0);
 457 }
 458
 459 void ptlrpc_pinger_wake_up()
 460 {
 461         pinger_args.pd_force_check = 1;
 462 }
 463 #endif /* !__KERNEL__ */