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 == NULL)
  59                 CWARN("replied (%d) w/o rep buffer?\n", req->rq_replied);
  60         if (ptlrpc_client_replied(req) && req->rq_repmsg &&
  61                         req->rq_repmsg->type == PTL_RPC_MSG_ERR)
  62                 return 0;
  63
  64         imp->imp_waiting_ping_reply = 0;
  65         return 0;
  66 }
  67
  68 int ptlrpc_ping(struct obd_import *imp)
  69 {
  70         struct ptlrpc_request *req;
  71         int rc = 0;
  72         ENTRY;
  73
  74         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 0, NULL, NULL);
  75         if (req) {
  76                 DEBUG_REQ(D_HA, req, "pinging %s->%s",
  77                           imp->imp_obd->obd_uuid.uuid,
  78                           imp->imp_target_uuid.uuid);
  79                 req->rq_no_resend = req->rq_no_delay = 1;
  80                 req->rq_replen = lustre_msg_size(0, NULL);
  81                 req->rq_interpret_reply = ptlrpc_ping_interpret;
  82                 req->rq_timeout = obd_timeout / PINGER_RATE;
  83                 imp->imp_waiting_ping_reply = 1;
  84                 imp->imp_last_ping_xid = req->rq_xid;
  85                 ptlrpcd_add_req(req);
  86         } else {
  87                 CERROR("OOM trying to ping %s->%s\n",
  88                        imp->imp_obd->obd_uuid.uuid,
  89                        imp->imp_target_uuid.uuid);
  90                 rc = -ENOMEM;
  91         }
  92
  93         RETURN(rc);
  94 }
  95
  96 #ifdef __KERNEL__
  97 static inline int ptlrpc_next_ping(struct obd_import *imp)
  98 {
  99         return jiffies + (obd_timeout / PINGER_RATE * HZ);
 100 }
 101
 102 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 103 {
 104         if (imp->imp_server_timeout)
 105                 return jiffies + (obd_timeout / 2 * HZ);
 106         else
 107                 return jiffies + (obd_timeout * HZ);
 108 }
 109
 110 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
 111 static wait_queue_head_t suspend_timeouts_waitq;
 112
 113 void ptlrpc_deactivate_timeouts(void)
 114 {
 115         CDEBUG(D_HA, "deactivate timeouts\n");
 116         atomic_inc(&suspend_timeouts);
 117 }
 118
 119 void ptlrpc_activate_timeouts(void)
 120 {
 121         CDEBUG(D_HA, "activate timeouts\n");
 122         LASSERT(atomic_read(&suspend_timeouts) > 0);
 123         if (atomic_dec_and_test(&suspend_timeouts))
 124                 wake_up(&suspend_timeouts_waitq);
 125 }
 126
 127 int ptlrpc_check_suspend(void)
 128 {
 129         if (atomic_read(&suspend_timeouts))
 130                 return 1;
 131         return 0;
 132 }
 133
 134 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 135 {
 136         struct l_wait_info lwi;
 137
 138         if (atomic_read(&suspend_timeouts)) {
 139                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 140                           atomic_read(&suspend_timeouts));
 141                 lwi = LWI_INTR(NULL, NULL);
 142                 l_wait_event(suspend_timeouts_waitq,
 143                              atomic_read(&suspend_timeouts) == 0, &lwi);
 144                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 145                 return 1;
 146         }
 147         return 0;
 148 }
 149
 150 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 151                                          unsigned long this_ping)
 152 {
 153         unsigned long flags;
 154         int force, level;
 155
 156         spin_lock_irqsave(&imp->imp_lock, flags);
 157         level = imp->imp_state;
 158         force = imp->imp_force_verify;
 159         if (force)
 160                 imp->imp_force_verify = 0;
 161         spin_unlock_irqrestore(&imp->imp_lock, flags);
 162
 163         if (imp->imp_next_ping > this_ping && force == 0)
 164                 return;
 165
 166         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 167                 /* wait at least a timeout before trying recovery again */
 168                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 169                 ptlrpc_initiate_recovery(imp);
 170         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov) {
 171                 CDEBUG(D_HA, "not pinging %s (in recovery "
 172                        " or recovery disabled: %s)\n",
 173                        imp->imp_target_uuid.uuid,
 174                        ptlrpc_import_state_name(level));
 175         } else if (level == LUSTRE_IMP_FULL && imp->imp_waiting_ping_reply &&
 176                         imp->imp_next_ping >= this_ping && imp->imp_pingable) {
 177                 CDEBUG(D_HA, "%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                 CDEBUG(D_ERROR, "%s: %s hasn't respond on ping x%lu\n",
 181                        imp->imp_obd->obd_uuid.uuid,
 182                        imp->imp_target_uuid.uuid, imp->imp_last_ping_xid);
 183                 ptlrpc_fail_import(imp, 0);
 184         } else if (imp->imp_pingable || force) {
 185                 imp->imp_next_ping = ptlrpc_next_ping(imp);
 186                 ptlrpc_ping(imp);
 187         }
 188 }
 189
 190 static int ptlrpc_pinger_main(void *arg)
 191 {
 192         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 193         struct ptlrpc_thread *thread = data->thread;
 194         unsigned long flags;
 195         ENTRY;
 196
 197         lock_kernel();
 198         ptlrpc_daemonize();
 199
 200         SIGNAL_MASK_LOCK(current, flags);
 201         sigfillset(&current->blocked);
 202         RECALC_SIGPENDING;
 203         SIGNAL_MASK_UNLOCK(current, flags);
 204
 205         LASSERTF(strlen(data->name) < sizeof(current->comm),
 206                  "name %d > len %d\n",
 207                  (int)strlen(data->name), (int)sizeof(current->comm));
 208         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 209         unlock_kernel();
 210
 211         /* Record that the thread is running */
 212         thread->t_flags = SVC_RUNNING;
 213         wake_up(&thread->t_ctl_waitq);
 214
 215         /* And now, loop forever, pinging as needed. */
 216         while (1) {
 217                 unsigned long this_ping = jiffies;
 218                 long time_to_next_ping;
 219                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
 220                                                      NULL, NULL);
 221                 struct list_head *iter;
 222
 223                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 224                 down(&pinger_sem);
 225                 list_for_each(iter, &pinger_imports) {
 226                         struct obd_import *imp =
 227                                 list_entry(iter, struct obd_import,
 228                                            imp_pinger_chain);
 229
 230                         ptlrpc_pinger_process_import(imp, this_ping);
 231
 232                         CDEBUG(D_OTHER, "%s: pingable %d, next_ping %lu(%lu)\n",
 233                                 imp->imp_target_uuid.uuid,
 234                                 imp->imp_pingable, imp->imp_next_ping, jiffies);
 235
 236                         if (imp->imp_pingable && imp->imp_next_ping &&
 237                             imp->imp_next_ping - jiffies < time_to_next_ping &&
 238                             imp->imp_next_ping > jiffies)
 239                                 time_to_next_ping = imp->imp_next_ping - jiffies;
 240                 }
 241                 up(&pinger_sem);
 242
 243                 /* Wait until the next ping time, or until we're stopped. */
 244                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 245                        this_ping + (obd_timeout * HZ));
 246                 if (time_to_next_ping > 0) {
 247                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 248                         l_wait_event(thread->t_ctl_waitq,
 249                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 250                                      &lwi);
 251                         if (thread->t_flags & SVC_STOPPING) {
 252                                 thread->t_flags &= ~SVC_STOPPING;
 253                                 EXIT;
 254                                 break;
 255                         } else if (thread->t_flags & SVC_EVENT) {
 256                                 /* woken after adding import to reset timer */
 257                                 thread->t_flags &= ~SVC_EVENT;
 258                         }
 259                 }
 260         }
 261
 262         thread->t_flags = SVC_STOPPED;
 263         wake_up(&thread->t_ctl_waitq);
 264
 265         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 266         return 0;
 267 }
 268
 269 static struct ptlrpc_thread *pinger_thread = NULL;
 270
 271 int ptlrpc_start_pinger(void)
 272 {
 273         struct l_wait_info lwi = { 0 };
 274         struct ptlrpc_svc_data d;
 275         int rc;
 276 #ifndef ENABLE_PINGER
 277         return 0;
 278 #endif
 279         ENTRY;
 280
 281         LASSERT(obd_timeout > PINGER_RATE);
 282
 283         if (pinger_thread != NULL)
 284                 RETURN(-EALREADY);
 285
 286         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 287         if (pinger_thread == NULL)
 288                 RETURN(-ENOMEM);
 289         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 290         init_waitqueue_head(&suspend_timeouts_waitq);
 291
 292         d.name = "ll_ping";
 293         d.thread = pinger_thread;
 294
 295         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 296          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 297         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 298         if (rc < 0) {
 299                 CERROR("cannot start thread: %d\n", rc);
 300                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 301                 RETURN(rc);
 302         }
 303         l_wait_event(pinger_thread->t_ctl_waitq,
 304                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 305
 306         RETURN(rc);
 307 }
 308
 309 int ptlrpc_stop_pinger(void)
 310 {
 311         struct l_wait_info lwi = { 0 };
 312         int rc = 0;
 313 #ifndef ENABLE_PINGER
 314         return 0;
 315 #endif
 316         ENTRY;
 317
 318         if (pinger_thread == NULL)
 319                 RETURN(-EALREADY);
 320         down(&pinger_sem);
 321         pinger_thread->t_flags = SVC_STOPPING;
 322         wake_up(&pinger_thread->t_ctl_waitq);
 323         up(&pinger_sem);
 324
 325         l_wait_event(pinger_thread->t_ctl_waitq,
 326                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 327
 328         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 329         pinger_thread = NULL;
 330         RETURN(rc);
 331 }
 332
 333 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 334 {
 335         down(&pinger_sem);
 336         imp->imp_next_ping = ptlrpc_next_ping(imp);
 337         up(&pinger_sem);
 338 }
 339
 340 int ptlrpc_pinger_add_import(struct obd_import *imp)
 341 {
 342         ENTRY;
 343         if (!list_empty(&imp->imp_pinger_chain))
 344                 RETURN(-EALREADY);
 345
 346         down(&pinger_sem);
 347         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 348                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 349         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 350         /* XXX sort, blah blah */
 351         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 352         class_import_get(imp);
 353
 354         ptlrpc_pinger_wake_up();
 355         up(&pinger_sem);
 356
 357         RETURN(0);
 358 }
 359
 360 int ptlrpc_pinger_del_import(struct obd_import *imp)
 361 {
 362         ENTRY;
 363         if (list_empty(&imp->imp_pinger_chain))
 364                 RETURN(-ENOENT);
 365
 366         down(&pinger_sem);
 367         list_del_init(&imp->imp_pinger_chain);
 368         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 369                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 370         class_import_put(imp);
 371         up(&pinger_sem);
 372         RETURN(0);
 373 }
 374
 375 void ptlrpc_pinger_wake_up()
 376 {
 377 #ifdef ENABLE_PINGER
 378         pinger_thread->t_flags |= SVC_EVENT;
 379         wake_up(&pinger_thread->t_ctl_waitq);
 380 #endif
 381 }
 382
 383 #else /* !__KERNEL__ */
 384
 385 /* XXX
 386  * the current implementation of pinger in liblustre is not optimized
 387  */
 388
 389 static struct pinger_data {
 390         int             pd_recursion;
 391         unsigned long   pd_this_ping;
 392         unsigned long   pd_next_ping;
 393         int             pd_force_check;
 394 } pinger_args;
 395
 396 static int pinger_check_rpcs(void *arg)
 397 {
 398         unsigned long curtime = time(NULL);
 399         struct list_head *iter;
 400         struct pinger_data *pd = &pinger_args;
 401
 402         /* prevent recursion */
 403         if (pd->pd_recursion++) {
 404                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 405                 pd->pd_recursion--;
 406                 return 0;
 407         }
 408
 409         /* have we reached ping point? */
 410         if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
 411                 pd->pd_recursion--;
 412                 return 0;
 413         }
 414
 415         if (pd->pd_force_check)
 416                 pd->pd_force_check = 0;
 417
 418         pd->pd_this_ping = curtime;
 419
 420         /* add rpcs into set */
 421         down(&pinger_sem);
 422         list_for_each(iter, &pinger_imports) {
 423                 struct obd_import *imp =
 424                         list_entry(iter, struct obd_import,
 425                                    imp_pinger_chain);
 426                 int level, force;
 427                 unsigned long flags;
 428
 429
 430                 spin_lock_irqsave(&imp->imp_lock, flags);
 431                 level = imp->imp_state;
 432                 force = imp->imp_force_verify;
 433                 if (force)
 434                         imp->imp_force_verify = 0;
 435                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 436
 437                 if (imp->imp_next_ping <= pd->pd_this_ping || force) {
 438                         if (level == LUSTRE_IMP_DISCON) {
 439                                 /* wait at least a timeout before
 440                                    trying recovery again. */
 441                                 unsigned long timeout = obd_timeout;
 442                                 if (imp->imp_server_timeout)
 443                                         timeout = obd_timeout / 2;
 444                                 imp->imp_next_ping = time(NULL) +
 445                                         (timeout * HZ);
 446                                 ptlrpc_initiate_recovery(imp);
 447                         }
 448                         else if (level != LUSTRE_IMP_FULL ||
 449                                  imp->imp_obd->obd_no_recov) {
 450                                 CDEBUG(D_HA,
 451                                        "not pinging %s (in recovery "
 452                                        " or recovery disabled: %s)\n",
 453                                        imp->imp_target_uuid.uuid,
 454                                        ptlrpc_import_state_name(level));
 455                         }
 456                         else if (imp->imp_pingable || force) {
 457                                 ptlrpc_ping(imp);
 458                         }
 459
 460                 } else {
 461                         if (imp->imp_pingable) {
 462                                 CDEBUG(D_HA, "don't need to ping %s "
 463                                        "(%lu > %lu)\n",
 464                                        imp->imp_target_uuid.uuid,
 465                                        imp->imp_next_ping, pd->pd_this_ping);
 466                         }
 467                 }
 468         }
 469
 470         up(&pinger_sem);
 471
 472         pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
 473
 474         CDEBUG(D_HA, "finished a round ping\n");
 475         pd->pd_recursion--;
 476         return 0;
 477 }
 478
 479 static void *pinger_callback = NULL;
 480
 481 int ptlrpc_start_pinger(void)
 482 {
 483         memset(&pinger_args, 0, sizeof(pinger_args));
 484 #ifdef ENABLE_PINGER
 485         pinger_callback =
 486                 liblustre_register_wait_callback(&pinger_check_rpcs, &pinger_args);
 487 #endif
 488         return 0;
 489 }
 490
 491 int ptlrpc_stop_pinger(void)
 492 {
 493 #ifdef ENABLE_PINGER
 494         if (pinger_callback)
 495                 liblustre_deregister_wait_callback(pinger_callback);
 496 #endif
 497         return 0;
 498 }
 499
 500 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 501 {
 502         down(&pinger_sem);
 503         imp->imp_next_ping = time(NULL) + obd_timeout;
 504         if (pinger_args.pd_next_ping > imp->imp_next_ping) {
 505                 CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
 506                         imp->imp_next_ping, time(NULL));
 507                 pinger_args.pd_next_ping = imp->imp_next_ping;
 508         }
 509         up(&pinger_sem);
 510 }
 511
 512 int ptlrpc_pinger_add_import(struct obd_import *imp)
 513 {
 514         ENTRY;
 515         if (!list_empty(&imp->imp_pinger_chain))
 516                 RETURN(-EALREADY);
 517
 518         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 519                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 520         ptlrpc_pinger_sending_on_import(imp);
 521
 522         down(&pinger_sem);
 523         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 524         class_import_get(imp);
 525         up(&pinger_sem);
 526
 527         RETURN(0);
 528 }
 529
 530 int ptlrpc_pinger_del_import(struct obd_import *imp)
 531 {
 532         ENTRY;
 533         if (list_empty(&imp->imp_pinger_chain))
 534                 RETURN(-ENOENT);
 535
 536         down(&pinger_sem);
 537         list_del_init(&imp->imp_pinger_chain);
 538         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 539                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 540         class_import_put(imp);
 541         up(&pinger_sem);
 542         RETURN(0);
 543 }
 544
 545 void ptlrpc_pinger_wake_up()
 546 {
 547         pinger_args.pd_force_check = 1;
 548 }
 549 #endif /* !__KERNEL__ */