lustre/ptlrpc/ptlrpcd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author Peter Braam <braam@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  *
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_RPC
  28
  29 #ifdef __KERNEL__
  30 # include <libcfs/libcfs.h>
  31 #else /* __KERNEL__ */
  32 # include <liblustre.h>
  33 # include <ctype.h>
  34 #endif
  35
  36 #include <libcfs/kp30.h>
  37 #include <lustre_net.h>
  38 # include <lustre_lib.h>
  39
  40 #include <lustre_ha.h>
  41 #include <obd_class.h>   /* for obd_zombie */
  42 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  43 #include <lprocfs_status.h>
  44
  45 #define LIOD_STOP 0
  46 struct ptlrpcd_ctl {
  47         unsigned long             pc_flags;
  48         spinlock_t                pc_lock;
  49         struct completion         pc_starting;
  50         struct completion         pc_finishing;
  51         struct list_head          pc_req_list;
  52         cfs_waitq_t               pc_waitq;
  53         struct ptlrpc_request_set *pc_set;
  54         char                      pc_name[16];
  55 #ifndef __KERNEL__
  56         int                       pc_recurred;
  57         void                     *pc_callback;
  58 #endif
  59 };
  60
  61 static struct ptlrpcd_ctl ptlrpcd_pc;
  62 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  63
  64 struct semaphore ptlrpcd_sem;
  65 static int ptlrpcd_users = 0;
  66
  67 void ptlrpcd_wake(struct ptlrpc_request *req)
  68 {
  69         struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  70
  71         LASSERT(pc != NULL);
  72
  73         cfs_waitq_signal(&pc->pc_waitq);
  74 }
  75
  76 /* requests that are added to the ptlrpcd queue are sent via
  77  * ptlrpcd_check->ptlrpc_check_set() */
  78 void ptlrpcd_add_req(struct ptlrpc_request *req)
  79 {
  80         struct ptlrpcd_ctl *pc;
  81
  82         if (req->rq_send_state == LUSTRE_IMP_FULL)
  83                 pc = &ptlrpcd_pc;
  84         else
  85                 pc = &ptlrpcd_recovery_pc;
  86
  87         req->rq_ptlrpcd_data = pc;
  88         ptlrpc_set_add_new_req(pc->pc_set, req);
  89         wake_up(&pc->pc_waitq);
  90 }
  91
  92 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
  93 {
  94         struct list_head *tmp, *pos;
  95         struct ptlrpc_request *req;
  96         int rc = 0;
  97         ENTRY;
  98
  99         if (test_bit(LIOD_STOP, &pc->pc_flags))
 100                 RETURN(1);
 101
 102         obd_zombie_impexp_cull();
 103
 104         spin_lock(&pc->pc_set->set_new_req_lock);
 105         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 106                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 107                 list_del_init(&req->rq_set_chain);
 108                 ptlrpc_set_add_req(pc->pc_set, req);
 109                 rc = 1; /* need to calculate its timeout */
 110         }
 111         spin_unlock(&pc->pc_set->set_new_req_lock);
 112
 113         if (pc->pc_set->set_remaining) {
 114                 rc = rc | ptlrpc_check_set(pc->pc_set);
 115
 116                 /* XXX our set never completes, so we prune the completed
 117                  * reqs after each iteration. boy could this be smarter. */
 118                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 119                         req = list_entry(pos, struct ptlrpc_request,
 120                                          rq_set_chain);
 121                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 122                                 continue;
 123
 124                         list_del_init(&req->rq_set_chain);
 125                         req->rq_set = NULL;
 126                         ptlrpc_req_finished (req);
 127                 }
 128         }
 129
 130         if (rc == 0) {
 131                 /* If new requests have been added, make sure to wake up */
 132                 spin_lock(&pc->pc_set->set_new_req_lock);
 133                 rc = !list_empty(&pc->pc_set->set_new_requests);
 134                 spin_unlock(&pc->pc_set->set_new_req_lock);
 135         }
 136
 137         RETURN(rc);
 138 }
 139
 140 #ifdef __KERNEL__
 141 /* ptlrpc's code paths like to execute in process context, so we have this
 142  * thread which spins on a set which contains the io rpcs.  llite specifies
 143  * ptlrpcd's set when it pushes pages down into the oscs */
 144 static int ptlrpcd(void *arg)
 145 {
 146         struct ptlrpcd_ctl *pc = arg;
 147         int rc;
 148         ENTRY;
 149
 150         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 151                 complete(&pc->pc_starting);
 152                 return rc;
 153         }
 154
 155         complete(&pc->pc_starting);
 156
 157         /* this mainloop strongly resembles ptlrpc_set_wait except
 158          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 159          * when there are requests in the set.  new requests come in
 160          * on the set's new_req_list and ptlrpcd_check moves them into
 161          * the set. */
 162         while (1) {
 163                 cfs_waitlink_t set_wait;
 164                 struct l_wait_info lwi;
 165                 cfs_duration_t timeout;
 166
 167                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
 168                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 169
 170                 /* ala the pinger, wait on pc's waitqueue and the set's */
 171                 cfs_waitlink_init(&set_wait);
 172                 cfs_waitq_add(&pc->pc_set->set_waitq, &set_wait);
 173                 cfs_waitq_forward(&set_wait, &pc->pc_waitq);
 174                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 175                 cfs_waitq_del(&pc->pc_set->set_waitq, &set_wait);
 176
 177                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 178                         break;
 179         }
 180         /* wait for inflight requests to drain */
 181         if (!list_empty(&pc->pc_set->set_requests))
 182                 ptlrpc_set_wait(pc->pc_set);
 183         complete(&pc->pc_finishing);
 184         return 0;
 185 }
 186
 187 static void ptlrpcd_zombie_impexp_notify(void)
 188 {
 189         cfs_waitq_signal(&ptlrpcd_pc.pc_waitq);
 190 }
 191 #else
 192
 193 int ptlrpcd_check_async_rpcs(void *arg)
 194 {
 195         struct ptlrpcd_ctl *pc = arg;
 196         int                  rc = 0;
 197
 198         /* single threaded!! */
 199         pc->pc_recurred++;
 200
 201         if (pc->pc_recurred == 1) {
 202                 rc = ptlrpcd_check(pc);
 203                 if (!rc)
 204                         ptlrpc_expired_set(pc->pc_set);
 205                 /*XXX send replay requests */
 206                 if (pc == &ptlrpcd_recovery_pc)
 207                         rc = ptlrpcd_check(pc);
 208         }
 209
 210         pc->pc_recurred--;
 211         return rc;
 212 }
 213 #endif
 214
 215 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 216 {
 217         int rc;
 218
 219         ENTRY;
 220         memset(pc, 0, sizeof(*pc));
 221         init_completion(&pc->pc_starting);
 222         init_completion(&pc->pc_finishing);
 223         cfs_waitq_init(&pc->pc_waitq);
 224         pc->pc_flags = 0;
 225         spin_lock_init(&pc->pc_lock);
 226         CFS_INIT_LIST_HEAD(&pc->pc_req_list);
 227         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 228
 229         pc->pc_set = ptlrpc_prep_set();
 230         if (pc->pc_set == NULL)
 231                 RETURN(-ENOMEM);
 232
 233 #ifdef __KERNEL__
 234         /* wake ptlrpcd when zombie imports or exports exist */
 235         obd_zombie_impexp_notify = ptlrpcd_zombie_impexp_notify;
 236
 237         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 238         if (rc < 0)  {
 239                 ptlrpc_set_destroy(pc->pc_set);
 240                 RETURN(rc);
 241         }
 242
 243         wait_for_completion(&pc->pc_starting);
 244 #else
 245         pc->pc_callback =
 246                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 247         (void)rc;
 248 #endif
 249         RETURN(0);
 250 }
 251
 252 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 253 {
 254         set_bit(LIOD_STOP, &pc->pc_flags);
 255         cfs_waitq_signal(&pc->pc_waitq);
 256 #ifdef __KERNEL__
 257         obd_zombie_impexp_notify = NULL;
 258         wait_for_completion(&pc->pc_finishing);
 259 #else
 260         liblustre_deregister_wait_callback(pc->pc_callback);
 261 #endif
 262         ptlrpc_set_destroy(pc->pc_set);
 263 }
 264
 265 int ptlrpcd_addref(void)
 266 {
 267         int rc = 0;
 268         ENTRY;
 269
 270         mutex_down(&ptlrpcd_sem);
 271         if (++ptlrpcd_users != 1)
 272                 GOTO(out, rc);
 273
 274         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 275         if (rc) {
 276                 --ptlrpcd_users;
 277                 GOTO(out, rc);
 278         }
 279
 280         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 281         if (rc) {
 282                 ptlrpcd_stop(&ptlrpcd_pc);
 283                 --ptlrpcd_users;
 284                 GOTO(out, rc);
 285         }
 286 out:
 287         mutex_up(&ptlrpcd_sem);
 288         RETURN(rc);
 289 }
 290
 291 void ptlrpcd_decref(void)
 292 {
 293         mutex_down(&ptlrpcd_sem);
 294         if (--ptlrpcd_users == 0) {
 295                 ptlrpcd_stop(&ptlrpcd_pc);
 296                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 297         }
 298         mutex_up(&ptlrpcd_sem);
 299 }