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 static int ptlrpc_ping_interpret(struct ptlrpc_request *req,
  44                                     void *data, int rc)
  45 {
  46         struct obd_import *imp = req->rq_import;
  47         DEBUG_REQ(D_HA, req, "ping reply");
  48         if (imp->imp_waiting_ping_reply == 0)
  49                 DEBUG_REQ(D_ERROR, req, "late ping reply?");
  50         if (imp->imp_last_ping_xid != req->rq_xid)
  51                 DEBUG_REQ(D_ERROR, req, "uh, wrong ping reply on x%lx",
  52                           imp->imp_last_ping_xid);
  53         else
  54                 imp->imp_last_ping_xid = 0;
  55
  56         /* if ping reply is an error, don't drop "replied" flag
  57          * on import, so pinger will invalidate it */
  58         if (ptlrpc_client_replied(req) && req->rq_repmsg->type == PTL_RPC_MSG_ERR)
  59                 return 0;
  60
  61         imp->imp_waiting_ping_reply = 0;
  62         return 0;
  63 }
  64
  65 int ptlrpc_ping(struct obd_import *imp)
  66 {
  67         struct ptlrpc_request *req;
  68         int rc = 0;
  69         ENTRY;
  70
  71         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 0, NULL, NULL);
  72         if (req) {
  73                 DEBUG_REQ(D_HA, req, "pinging %s->%s",
  74                           imp->imp_obd->obd_uuid.uuid,
  75                           imp->imp_target_uuid.uuid);
  76                 req->rq_no_resend = req->rq_no_delay = 1;
  77                 req->rq_replen = lustre_msg_size(0, NULL);
  78                 req->rq_interpret_reply = ptlrpc_ping_interpret;
  79                 req->rq_timeout = obd_timeout / PINGER_RATE;
  80                 imp->imp_waiting_ping_reply = 1;
  81                 imp->imp_last_ping_xid = req->rq_xid;
  82                 ptlrpcd_add_req(req);
  83         } else {
  84                 CERROR("OOM trying to ping %s->%s\n",
  85                        imp->imp_obd->obd_uuid.uuid,
  86                        imp->imp_target_uuid.uuid);
  87                 rc = -ENOMEM;
  88         }
  89
  90         RETURN(rc);
  91 }
  92
  93 #ifdef __KERNEL__
  94 static inline int ptlrpc_next_ping(struct obd_import *imp)
  95 {
  96         return jiffies + (obd_timeout / PINGER_RATE * HZ);
  97 }
  98
  99 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 100 {
 101         if (imp->imp_server_timeout)
 102                 return jiffies + (obd_timeout / 2 * HZ);
 103         else
 104                 return jiffies + (obd_timeout * HZ);
 105 }
 106
 107 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
 108 static wait_queue_head_t suspend_timeouts_waitq;
 109
 110 void ptlrpc_deactivate_timeouts(void)
 111 {
 112         CDEBUG(D_HA, "deactivate timeouts\n");
 113         atomic_inc(&suspend_timeouts);
 114 }
 115
 116 void ptlrpc_activate_timeouts(void)
 117 {
 118         CDEBUG(D_HA, "activate timeouts\n");
 119         LASSERT(atomic_read(&suspend_timeouts) > 0);
 120         if (atomic_dec_and_test(&suspend_timeouts))
 121                 wake_up(&suspend_timeouts_waitq);
 122 }
 123
 124 int ptlrpc_check_suspend(void)
 125 {
 126         if (atomic_read(&suspend_timeouts))
 127                 return 1;
 128         return 0;
 129 }
 130
 131 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 132 {
 133         struct l_wait_info lwi;
 134
 135         if (atomic_read(&suspend_timeouts)) {
 136                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 137                           atomic_read(&suspend_timeouts));
 138                 lwi = LWI_INTR(NULL, NULL);
 139                 l_wait_event(suspend_timeouts_waitq,
 140                              atomic_read(&suspend_timeouts) == 0, &lwi);
 141                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 142                 return 1;
 143         }
 144         return 0;
 145 }
 146
 147 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 148                                          unsigned long this_ping)
 149 {
 150         unsigned long flags;
 151         int force, level;
 152
 153         spin_lock_irqsave(&imp->imp_lock, flags);
 154         level = imp->imp_state;
 155         force = imp->imp_force_verify;
 156         if (force)
 157                 imp->imp_force_verify = 0;
 158         spin_unlock_irqrestore(&imp->imp_lock, flags);
 159
 160         if (imp->imp_next_ping > this_ping && force == 0)
 161                 return;
 162
 163         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 164                 /* wait at least a timeout before trying recovery again */
 165                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 166                 ptlrpc_initiate_recovery(imp);
 167         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov) {
 168                 CDEBUG(D_HA, "not pinging %s (in recovery "
 169                        " or recovery disabled: %s)\n",
 170                        imp->imp_target_uuid.uuid,
 171                        ptlrpc_import_state_name(level));
 172         } else if (level == LUSTRE_IMP_FULL && imp->imp_waiting_ping_reply &&
 173                         imp->imp_next_ping >= this_ping && imp->imp_pingable) {
 174                 CDEBUG(D_HA, "%s: %s hasn't respond on ping x%lu\n",
 175                        imp->imp_obd->obd_uuid.uuid,
 176                        imp->imp_target_uuid.uuid, imp->imp_last_ping_xid);
 177                 CDEBUG(D_ERROR, "%s: %s hasn't respond on ping x%lu\n",
 178                        imp->imp_obd->obd_uuid.uuid,
 179                        imp->imp_target_uuid.uuid, imp->imp_last_ping_xid);
 180                 ptlrpc_fail_import(imp, 0);
 181         } else if (imp->imp_pingable || force) {
 182                 imp->imp_next_ping = ptlrpc_next_ping(imp);
 183                 ptlrpc_ping(imp);
 184         }
 185 }
 186
 187 static int ptlrpc_pinger_main(void *arg)
 188 {
 189         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 190         struct ptlrpc_thread *thread = data->thread;
 191         unsigned long flags;
 192         ENTRY;
 193
 194         lock_kernel();
 195         ptlrpc_daemonize();
 196
 197         SIGNAL_MASK_LOCK(current, flags);
 198         sigfillset(&current->blocked);
 199         RECALC_SIGPENDING;
 200         SIGNAL_MASK_UNLOCK(current, flags);
 201
 202         LASSERTF(strlen(data->name) < sizeof(current->comm),
 203                  "name %d > len %d\n",
 204                  (int)strlen(data->name), (int)sizeof(current->comm));
 205         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 206         unlock_kernel();
 207
 208         /* Record that the thread is running */
 209         thread->t_flags = SVC_RUNNING;
 210         wake_up(&thread->t_ctl_waitq);
 211
 212         /* And now, loop forever, pinging as needed. */
 213         while (1) {
 214                 unsigned long this_ping = jiffies;
 215                 long time_to_next_ping;
 216                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
 217                                                      NULL, NULL);
 218                 struct list_head *iter;
 219
 220                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 221                 down(&pinger_sem);
 222                 list_for_each(iter, &pinger_imports) {
 223                         struct obd_import *imp =
 224                                 list_entry(iter, struct obd_import,
 225                                            imp_pinger_chain);
 226
 227                         ptlrpc_pinger_process_import(imp, this_ping);
 228
 229                         CDEBUG(D_OTHER, "%s: pingable %d, next_ping %lu(%lu)\n",
 230                                 imp->imp_target_uuid.uuid,
 231                                 imp->imp_pingable, imp->imp_next_ping, jiffies);
 232
 233                         if (imp->imp_pingable && imp->imp_next_ping &&
 234                             imp->imp_next_ping - jiffies < time_to_next_ping &&
 235                             imp->imp_next_ping > jiffies)
 236                                 time_to_next_ping = imp->imp_next_ping - jiffies;
 237                 }
 238                 up(&pinger_sem);
 239
 240                 /* Wait until the next ping time, or until we're stopped. */
 241                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 242                        this_ping + (obd_timeout * HZ));
 243                 if (time_to_next_ping > 0) {
 244                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 245                         l_wait_event(thread->t_ctl_waitq,
 246                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 247                                      &lwi);
 248                         if (thread->t_flags & SVC_STOPPING) {
 249                                 thread->t_flags &= ~SVC_STOPPING;
 250                                 EXIT;
 251                                 break;
 252                         } else if (thread->t_flags & SVC_EVENT) {
 253                                 /* woken after adding import to reset timer */
 254                                 thread->t_flags &= ~SVC_EVENT;
 255                         }
 256                 }
 257         }
 258
 259         thread->t_flags = SVC_STOPPED;
 260         wake_up(&thread->t_ctl_waitq);
 261
 262         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 263         return 0;
 264 }
 265
 266 static struct ptlrpc_thread *pinger_thread = NULL;
 267
 268 int ptlrpc_start_pinger(void)
 269 {
 270         struct l_wait_info lwi = { 0 };
 271         struct ptlrpc_svc_data d;
 272         int rc;
 273 #ifndef ENABLE_PINGER
 274         return 0;
 275 #endif
 276         ENTRY;
 277
 278         LASSERT(obd_timeout > PINGER_RATE);
 279
 280         if (pinger_thread != NULL)
 281                 RETURN(-EALREADY);
 282
 283         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 284         if (pinger_thread == NULL)
 285                 RETURN(-ENOMEM);
 286         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 287         init_waitqueue_head(&suspend_timeouts_waitq);
 288
 289         d.name = "ll_ping";
 290         d.thread = pinger_thread;
 291
 292         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 293          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 294         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 295         if (rc < 0) {
 296                 CERROR("cannot start thread: %d\n", rc);
 297                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 298                 RETURN(rc);
 299         }
 300         l_wait_event(pinger_thread->t_ctl_waitq,
 301                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 302
 303         RETURN(rc);
 304 }
 305
 306 int ptlrpc_stop_pinger(void)
 307 {
 308         struct l_wait_info lwi = { 0 };
 309         int rc = 0;
 310 #ifndef ENABLE_PINGER
 311         return 0;
 312 #endif
 313         ENTRY;
 314
 315         if (pinger_thread == NULL)
 316                 RETURN(-EALREADY);
 317         down(&pinger_sem);
 318         pinger_thread->t_flags = SVC_STOPPING;
 319         wake_up(&pinger_thread->t_ctl_waitq);
 320         up(&pinger_sem);
 321
 322         l_wait_event(pinger_thread->t_ctl_waitq,
 323                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 324
 325         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 326         pinger_thread = NULL;
 327         RETURN(rc);
 328 }
 329
 330 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 331 {
 332         down(&pinger_sem);
 333         imp->imp_next_ping = ptlrpc_next_ping(imp);
 334         up(&pinger_sem);
 335 }
 336
 337 int ptlrpc_pinger_add_import(struct obd_import *imp)
 338 {
 339         ENTRY;
 340         if (!list_empty(&imp->imp_pinger_chain))
 341                 RETURN(-EALREADY);
 342
 343         down(&pinger_sem);
 344         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 345                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 346         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 347         /* XXX sort, blah blah */
 348         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 349         class_import_get(imp);
 350
 351         ptlrpc_pinger_wake_up();
 352         up(&pinger_sem);
 353
 354         RETURN(0);
 355 }
 356
 357 int ptlrpc_pinger_del_import(struct obd_import *imp)
 358 {
 359         ENTRY;
 360         if (list_empty(&imp->imp_pinger_chain))
 361                 RETURN(-ENOENT);
 362
 363         down(&pinger_sem);
 364         list_del_init(&imp->imp_pinger_chain);
 365         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 366                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 367         class_import_put(imp);
 368         up(&pinger_sem);
 369         RETURN(0);
 370 }
 371
 372 void ptlrpc_pinger_wake_up()
 373 {
 374 #ifdef ENABLE_PINGER
 375         pinger_thread->t_flags |= SVC_EVENT;
 376         wake_up(&pinger_thread->t_ctl_waitq);
 377 #endif
 378 }
 379
 380 #else /* !__KERNEL__ */
 381
 382 /* XXX
 383  * the current implementation of pinger in liblustre is not optimized
 384  */
 385
 386 static struct pinger_data {
 387         int             pd_recursion;
 388         unsigned long   pd_this_ping;
 389         unsigned long   pd_next_ping;
 390         int             pd_force_check;
 391 } pinger_args;
 392
 393 static int pinger_check_rpcs(void *arg)
 394 {
 395         unsigned long curtime = time(NULL);
 396         struct list_head *iter;
 397         struct pinger_data *pd = &pinger_args;
 398
 399         /* prevent recursion */
 400         if (pd->pd_recursion++) {
 401                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 402                 pd->pd_recursion--;
 403                 return 0;
 404         }
 405
 406         /* have we reached ping point? */
 407         if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
 408                 pd->pd_recursion--;
 409                 return 0;
 410         }
 411
 412         if (pd->pd_force_check)
 413                 pd->pd_force_check = 0;
 414
 415         pd->pd_this_ping = curtime;
 416
 417         /* add rpcs into set */
 418         down(&pinger_sem);
 419         list_for_each(iter, &pinger_imports) {
 420                 struct obd_import *imp =
 421                         list_entry(iter, struct obd_import,
 422                                    imp_pinger_chain);
 423                 int level, force;
 424                 unsigned long flags;
 425
 426
 427                 spin_lock_irqsave(&imp->imp_lock, flags);
 428                 level = imp->imp_state;
 429                 force = imp->imp_force_verify;
 430                 if (force)
 431                         imp->imp_force_verify = 0;
 432                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 433
 434                 if (imp->imp_next_ping <= pd->pd_this_ping || force) {
 435                         if (level == LUSTRE_IMP_DISCON) {
 436                                 /* wait at least a timeout before
 437                                    trying recovery again. */
 438                                 unsigned long timeout = obd_timeout;
 439                                 if (imp->imp_server_timeout)
 440                                         timeout = obd_timeout / 2;
 441                                 imp->imp_next_ping = time(NULL) +
 442                                         (timeout * HZ);
 443                                 ptlrpc_initiate_recovery(imp);
 444                         }
 445                         else if (level != LUSTRE_IMP_FULL ||
 446                                  imp->imp_obd->obd_no_recov) {
 447                                 CDEBUG(D_HA,
 448                                        "not pinging %s (in recovery "
 449                                        " or recovery disabled: %s)\n",
 450                                        imp->imp_target_uuid.uuid,
 451                                        ptlrpc_import_state_name(level));
 452                         }
 453                         else if (imp->imp_pingable || force) {
 454                                 ptlrpc_ping(imp);
 455                         }
 456
 457                 } else {
 458                         if (imp->imp_pingable) {
 459                                 CDEBUG(D_HA, "don't need to ping %s "
 460                                        "(%lu > %lu)\n",
 461                                        imp->imp_target_uuid.uuid,
 462                                        imp->imp_next_ping, pd->pd_this_ping);
 463                         }
 464                 }
 465         }
 466
 467         up(&pinger_sem);
 468
 469         pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
 470
 471         CDEBUG(D_HA, "finished a round ping\n");
 472         pd->pd_recursion--;
 473         return 0;
 474 }
 475
 476 static void *pinger_callback = NULL;
 477
 478 int ptlrpc_start_pinger(void)
 479 {
 480         memset(&pinger_args, 0, sizeof(pinger_args));
 481 #ifdef ENABLE_PINGER
 482         pinger_callback =
 483                 liblustre_register_wait_callback(&pinger_check_rpcs, &pinger_args);
 484 #endif
 485         return 0;
 486 }
 487
 488 int ptlrpc_stop_pinger(void)
 489 {
 490 #ifdef ENABLE_PINGER
 491         if (pinger_callback)
 492                 liblustre_deregister_wait_callback(pinger_callback);
 493 #endif
 494         return 0;
 495 }
 496
 497 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 498 {
 499         down(&pinger_sem);
 500         imp->imp_next_ping = time(NULL) + obd_timeout;
 501         if (pinger_args.pd_next_ping > imp->imp_next_ping) {
 502                 CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
 503                         imp->imp_next_ping, time(NULL));
 504                 pinger_args.pd_next_ping = imp->imp_next_ping;
 505         }
 506         up(&pinger_sem);
 507 }
 508
 509 int ptlrpc_pinger_add_import(struct obd_import *imp)
 510 {
 511         ENTRY;
 512         if (!list_empty(&imp->imp_pinger_chain))
 513                 RETURN(-EALREADY);
 514
 515         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 516                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 517         ptlrpc_pinger_sending_on_import(imp);
 518
 519         down(&pinger_sem);
 520         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 521         class_import_get(imp);
 522         up(&pinger_sem);
 523
 524         RETURN(0);
 525 }
 526
 527 int ptlrpc_pinger_del_import(struct obd_import *imp)
 528 {
 529         ENTRY;
 530         if (list_empty(&imp->imp_pinger_chain))
 531                 RETURN(-ENOENT);
 532
 533         down(&pinger_sem);
 534         list_del_init(&imp->imp_pinger_chain);
 535         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 536                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 537         class_import_put(imp);
 538         up(&pinger_sem);
 539         RETURN(0);
 540 }
 541
 542 void ptlrpc_pinger_wake_up()
 543 {
 544         pinger_args.pd_force_check = 1;
 545 }
 546 #endif /* !__KERNEL__ */