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