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