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 #ifdef __KERNEL__
  39
  40 static struct ptlrpc_thread *pinger_thread = NULL;
  41 static DECLARE_MUTEX(pinger_sem);
  42 static struct list_head pinger_imports = LIST_HEAD_INIT(pinger_imports);
  43
  44 static int ptlrpc_pinger_main(void *arg)
  45 {
  46         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
  47         struct ptlrpc_thread *thread = data->thread;
  48         unsigned long flags;
  49         ENTRY;
  50
  51         lock_kernel();
  52         ptlrpc_daemonize();
  53
  54         SIGNAL_MASK_LOCK(current, flags);
  55         sigfillset(&current->blocked);
  56         RECALC_SIGPENDING;
  57         SIGNAL_MASK_UNLOCK(current, flags);
  58
  59         THREAD_NAME(current->comm, "%s", data->name);
  60         unlock_kernel();
  61
  62         /* Record that the thread is running */
  63         thread->t_flags = SVC_RUNNING;
  64         wake_up(&thread->t_ctl_waitq);
  65
  66         /* And now, loop forever, pinging as needed. */
  67         while (1) {
  68                 unsigned long this_ping = jiffies;
  69                 long time_to_next_ping;
  70                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
  71                                                      NULL, NULL);
  72                 struct ptlrpc_request_set *set;
  73                 struct ptlrpc_request *req;
  74                 struct list_head *iter;
  75                 wait_queue_t set_wait;
  76                 int rc;
  77
  78                 set = ptlrpc_prep_set();
  79                 down(&pinger_sem);
  80                 list_for_each(iter, &pinger_imports) {
  81                         struct obd_import *imp =
  82                                 list_entry(iter, struct obd_import,
  83                                            imp_pinger_chain);
  84                         int generation, level;
  85                         unsigned long flags;
  86
  87                         if (imp->imp_next_ping <= this_ping) {
  88                                 /* Add a ping. */
  89                                 spin_lock_irqsave(&imp->imp_lock, flags);
  90                                 generation = imp->imp_generation;
  91                                 level = imp->imp_state;
  92                                 spin_unlock_irqrestore(&imp->imp_lock, flags);
  93
  94                                 if (level != LUSTRE_IMP_FULL) {
  95                                         CDEBUG(D_HA,
  96                                                "not pinging %s (in recovery)\n",
  97                                                imp->imp_target_uuid.uuid);
  98                                         continue;
  99                                 }
 100
 101                                 req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL,
 102                                                       NULL);
 103                                 if (!req) {
 104                                         CERROR("OOM trying to ping\n");
 105                                         break;
 106                                 }
 107                                 req->rq_no_resend = 1;
 108                                 req->rq_replen = lustre_msg_size(0, NULL);
 109                                 req->rq_send_state = LUSTRE_IMP_FULL;
 110                                 req->rq_phase = RQ_PHASE_RPC;
 111                                 req->rq_import_generation = generation;
 112                                 ptlrpc_set_add_req(set, req);
 113                         } else {
 114                                 CDEBUG(D_HA, "don't need to ping %s (%lu > "
 115                                        "%lu)\n", imp->imp_target_uuid.uuid,
 116                                        imp->imp_next_ping, this_ping);
 117                         }
 118                 }
 119                 up(&pinger_sem);
 120
 121                 /* Might be empty, that's OK. */
 122                 if (set->set_remaining == 0)
 123                         CDEBUG(D_HA, "nothing to ping\n");
 124                 list_for_each(iter, &set->set_requests) {
 125                         struct ptlrpc_request *req =
 126                                 list_entry(iter, struct ptlrpc_request,
 127                                            rq_set_chain);
 128                         DEBUG_REQ(D_HA, req, "pinging %s->%s",
 129                                   req->rq_import->imp_obd->obd_uuid.uuid,
 130                                   req->rq_import->imp_target_uuid.uuid);
 131                         (void)ptl_send_rpc(req);
 132                 }
 133
 134                 /* Have to wait on both the thread's queue and the set's. */
 135                 init_waitqueue_entry(&set_wait, current);
 136                 add_wait_queue(&set->set_waitq, &set_wait);
 137                 rc = l_wait_event(thread->t_ctl_waitq,
 138                                   thread->t_flags & SVC_STOPPING ||
 139                                   ptlrpc_check_set(set),
 140                                   &lwi);
 141                 remove_wait_queue(&set->set_waitq, &set_wait);
 142                 CDEBUG(D_HA, "ping complete (%lu)\n", jiffies);
 143
 144                 if (thread->t_flags & SVC_STOPPING) {
 145                         thread->t_flags &= ~SVC_STOPPING;
 146                         list_for_each(iter, &set->set_requests) {
 147                                 req = list_entry(iter, struct ptlrpc_request,
 148                                                  rq_set_chain);
 149                                 if (!req->rq_replied)
 150                                         ptlrpc_unregister_reply(req);
 151                         }
 152                         ptlrpc_set_destroy(set);
 153                         EXIT;
 154                         break;
 155                 }
 156
 157                 /* Expire all the requests that didn't come back. */
 158                 down(&pinger_sem);
 159                 list_for_each(iter, &set->set_requests) {
 160                         req = list_entry(iter, struct ptlrpc_request,
 161                                          rq_set_chain);
 162
 163                         if (req->rq_replied)
 164                                 continue;
 165
 166                         req->rq_phase = RQ_PHASE_COMPLETE;
 167                         set->set_remaining--;
 168                         /* If it was disconnected, don't sweat it. */
 169                         if (list_empty(&req->rq_import->imp_pinger_chain)) {
 170                                 ptlrpc_unregister_reply(req);
 171                                 continue;
 172                         }
 173
 174                         ptlrpc_expire_one_request(req);
 175                 }
 176                 up(&pinger_sem);
 177                 ptlrpc_set_destroy(set);
 178
 179                 /* Wait until the next ping time, or until we're stopped. */
 180                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 181                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 182                        this_ping + (obd_timeout * HZ));
 183                 if (time_to_next_ping > 0) {
 184                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 185                         l_wait_event(thread->t_ctl_waitq,
 186                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 187                                      &lwi);
 188                         if (thread->t_flags & SVC_STOPPING) {
 189                                 thread->t_flags &= ~SVC_STOPPING;
 190                                 EXIT;
 191                                 break;
 192                         } else if (thread->t_flags & SVC_EVENT) {
 193                                 /* woken after adding import to reset timer */
 194                                 thread->t_flags &= ~SVC_EVENT;
 195                         }
 196                 }
 197         }
 198
 199         thread->t_flags = SVC_STOPPED;
 200         wake_up(&thread->t_ctl_waitq);
 201
 202         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 203         return 0;
 204 }
 205
 206 int ptlrpc_start_pinger(void)
 207 {
 208         struct l_wait_info lwi = { 0 };
 209         struct ptlrpc_svc_data d;
 210         int rc;
 211 #ifndef ENABLE_PINGER
 212         return 0;
 213 #endif
 214         ENTRY;
 215
 216         if (pinger_thread != NULL)
 217                 RETURN(-EALREADY);
 218
 219         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 220         if (pinger_thread == NULL)
 221                 RETURN(-ENOMEM);
 222         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 223
 224         d.name = "ll_ping";
 225         d.thread = pinger_thread;
 226
 227         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 228          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 229         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 230         if (rc < 0) {
 231                 CERROR("cannot start thread: %d\n", rc);
 232                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 233                 RETURN(rc);
 234         }
 235         l_wait_event(pinger_thread->t_ctl_waitq,
 236                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 237
 238         RETURN(rc);
 239 }
 240
 241 int ptlrpc_stop_pinger(void)
 242 {
 243         struct l_wait_info lwi = { 0 };
 244         int rc = 0;
 245 #ifndef ENABLE_PINGER
 246         return 0;
 247 #endif
 248         ENTRY;
 249
 250         if (pinger_thread == NULL)
 251                 RETURN(-EALREADY);
 252         down(&pinger_sem);
 253         pinger_thread->t_flags = SVC_STOPPING;
 254         wake_up(&pinger_thread->t_ctl_waitq);
 255         up(&pinger_sem);
 256
 257         l_wait_event(pinger_thread->t_ctl_waitq,
 258                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 259
 260         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 261         pinger_thread = NULL;
 262         RETURN(rc);
 263 }
 264
 265 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 266 {
 267         down(&pinger_sem);
 268         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 269         up(&pinger_sem);
 270 }
 271
 272 int ptlrpc_pinger_add_import(struct obd_import *imp)
 273 {
 274         ENTRY;
 275         if (!list_empty(&imp->imp_pinger_chain))
 276                 RETURN(-EALREADY);
 277
 278         down(&pinger_sem);
 279         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 280                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 281         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 282         /* XXX sort, blah blah */
 283         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 284         class_import_get(imp);
 285
 286 #ifdef ENABLE_PINGER
 287         pinger_thread->t_flags |= SVC_EVENT;
 288         wake_up(&pinger_thread->t_ctl_waitq);
 289 #endif
 290         up(&pinger_sem);
 291
 292         RETURN(0);
 293 }
 294
 295 int ptlrpc_pinger_del_import(struct obd_import *imp)
 296 {
 297         ENTRY;
 298         if (list_empty(&imp->imp_pinger_chain))
 299                 RETURN(-ENOENT);
 300
 301         down(&pinger_sem);
 302         list_del_init(&imp->imp_pinger_chain);
 303         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 304                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 305         class_import_put(imp);
 306         up(&pinger_sem);
 307         RETURN(0);
 308 }
 309
 310 #else /* !__KERNEL__ */
 311
 312 int ptlrpc_start_pinger(void)
 313 {
 314         return 0;
 315 }
 316
 317 int ptlrpc_stop_pinger(void)
 318 {
 319         return 0;
 320 }
 321
 322 int ptlrpc_pinger_add_import(struct obd_import *imp)
 323 {
 324         return 0;
 325 }
 326
 327 int ptlrpc_pinger_del_import(struct obd_import *imp)
 328 {
 329         return 0;
 330 }
 331
 332 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 333 {
 334 }
 335
 336 #endif