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