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