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