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_wait_callback;
  58         void                     *pc_idle_callback;
  59 #endif
  60 };
  61
  62 static struct ptlrpcd_ctl ptlrpcd_pc;
  63 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  64
  65 struct semaphore ptlrpcd_sem;
  66 static int ptlrpcd_users = 0;
  67
  68 void ptlrpcd_wake(struct ptlrpc_request *req)
  69 {
  70         struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  71
  72         LASSERT(pc != NULL);
  73
  74         cfs_waitq_signal(&pc->pc_waitq);
  75 }
  76
  77 /* requests that are added to the ptlrpcd queue are sent via
  78  * ptlrpcd_check->ptlrpc_check_set() */
  79 void ptlrpcd_add_req(struct ptlrpc_request *req)
  80 {
  81         struct ptlrpcd_ctl *pc;
  82
  83         if (req->rq_send_state == LUSTRE_IMP_FULL)
  84                 pc = &ptlrpcd_pc;
  85         else
  86                 pc = &ptlrpcd_recovery_pc;
  87
  88         req->rq_ptlrpcd_data = pc;
  89         ptlrpc_set_add_new_req(pc->pc_set, req);
  90         wake_up(&pc->pc_waitq);
  91 }
  92
  93 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
  94 {
  95         struct list_head *tmp, *pos;
  96         struct ptlrpc_request *req;
  97         int rc = 0;
  98         ENTRY;
  99
 100         if (test_bit(LIOD_STOP, &pc->pc_flags))
 101                 RETURN(1);
 102
 103         obd_zombie_impexp_cull();
 104
 105         spin_lock(&pc->pc_set->set_new_req_lock);
 106         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 107                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 108                 list_del_init(&req->rq_set_chain);
 109                 ptlrpc_set_add_req(pc->pc_set, req);
 110                 rc = 1; /* need to calculate its timeout */
 111         }
 112         spin_unlock(&pc->pc_set->set_new_req_lock);
 113
 114         if (pc->pc_set->set_remaining) {
 115                 rc = rc | ptlrpc_check_set(pc->pc_set);
 116
 117                 /* XXX our set never completes, so we prune the completed
 118                  * reqs after each iteration. boy could this be smarter. */
 119                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 120                         req = list_entry(pos, struct ptlrpc_request,
 121                                          rq_set_chain);
 122                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 123                                 continue;
 124
 125                         list_del_init(&req->rq_set_chain);
 126                         req->rq_set = NULL;
 127                         ptlrpc_req_finished (req);
 128                 }
 129         }
 130
 131         if (rc == 0) {
 132                 /* If new requests have been added, make sure to wake up */
 133                 spin_lock(&pc->pc_set->set_new_req_lock);
 134                 rc = !list_empty(&pc->pc_set->set_new_requests);
 135                 spin_unlock(&pc->pc_set->set_new_req_lock);
 136         }
 137
 138         RETURN(rc);
 139 }
 140
 141 #ifdef __KERNEL__
 142 /* ptlrpc's code paths like to execute in process context, so we have this
 143  * thread which spins on a set which contains the io rpcs.  llite specifies
 144  * ptlrpcd's set when it pushes pages down into the oscs */
 145 static int ptlrpcd(void *arg)
 146 {
 147         struct ptlrpcd_ctl *pc = arg;
 148         int rc;
 149         ENTRY;
 150
 151         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 152                 complete(&pc->pc_starting);
 153                 return rc;
 154         }
 155
 156         complete(&pc->pc_starting);
 157
 158         /* this mainloop strongly resembles ptlrpc_set_wait except
 159          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 160          * when there are requests in the set.  new requests come in
 161          * on the set's new_req_list and ptlrpcd_check moves them into
 162          * the set. */
 163         while (1) {
 164                 cfs_waitlink_t set_wait;
 165                 struct l_wait_info lwi;
 166                 cfs_duration_t timeout;
 167
 168                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
 169                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 170
 171                 /* ala the pinger, wait on pc's waitqueue and the set's */
 172                 cfs_waitlink_init(&set_wait);
 173                 cfs_waitq_add(&pc->pc_set->set_waitq, &set_wait);
 174                 cfs_waitq_forward(&set_wait, &pc->pc_waitq);
 175                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 176                 cfs_waitq_del(&pc->pc_set->set_waitq, &set_wait);
 177
 178                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 179                         break;
 180         }
 181         /* wait for inflight requests to drain */
 182         if (!list_empty(&pc->pc_set->set_requests))
 183                 ptlrpc_set_wait(pc->pc_set);
 184         complete(&pc->pc_finishing);
 185         return 0;
 186 }
 187
 188 static void ptlrpcd_zombie_impexp_notify(void)
 189 {
 190         cfs_waitq_signal(&ptlrpcd_pc.pc_waitq);
 191 }
 192 #else
 193
 194 int ptlrpcd_check_async_rpcs(void *arg)
 195 {
 196         struct ptlrpcd_ctl *pc = arg;
 197         int                  rc = 0;
 198
 199         /* single threaded!! */
 200         pc->pc_recurred++;
 201
 202         if (pc->pc_recurred == 1) {
 203                 rc = ptlrpcd_check(pc);
 204                 if (!rc)
 205                         ptlrpc_expired_set(pc->pc_set);
 206                 /*XXX send replay requests */
 207                 if (pc == &ptlrpcd_recovery_pc)
 208                         rc = ptlrpcd_check(pc);
 209         }
 210
 211         pc->pc_recurred--;
 212         return rc;
 213 }
 214
 215 int ptlrpcd_idle(void *arg)
 216 {
 217         struct ptlrpcd_ctl *pc = arg;
 218
 219         return (list_empty(&pc->pc_set->set_new_requests) &&
 220                 pc->pc_set->set_remaining == 0);
 221 }
 222
 223 #endif
 224
 225 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 226 {
 227         int rc;
 228
 229         ENTRY;
 230         memset(pc, 0, sizeof(*pc));
 231         init_completion(&pc->pc_starting);
 232         init_completion(&pc->pc_finishing);
 233         cfs_waitq_init(&pc->pc_waitq);
 234         pc->pc_flags = 0;
 235         spin_lock_init(&pc->pc_lock);
 236         CFS_INIT_LIST_HEAD(&pc->pc_req_list);
 237         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 238
 239         pc->pc_set = ptlrpc_prep_set();
 240         if (pc->pc_set == NULL)
 241                 RETURN(-ENOMEM);
 242
 243 #ifdef __KERNEL__
 244         /* wake ptlrpcd when zombie imports or exports exist */
 245         obd_zombie_impexp_notify = ptlrpcd_zombie_impexp_notify;
 246
 247         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 248         if (rc < 0)  {
 249                 ptlrpc_set_destroy(pc->pc_set);
 250                 RETURN(rc);
 251         }
 252
 253         wait_for_completion(&pc->pc_starting);
 254 #else
 255         pc->pc_wait_callback =
 256                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 257                                                  &ptlrpcd_check_async_rpcs, pc);
 258         pc->pc_idle_callback =
 259                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 260                                                  &ptlrpcd_idle, pc);
 261         (void)rc;
 262 #endif
 263         RETURN(0);
 264 }
 265
 266 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 267 {
 268         set_bit(LIOD_STOP, &pc->pc_flags);
 269         cfs_waitq_signal(&pc->pc_waitq);
 270 #ifdef __KERNEL__
 271         obd_zombie_impexp_notify = NULL;
 272         wait_for_completion(&pc->pc_finishing);
 273 #else
 274         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 275         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 276 #endif
 277         ptlrpc_set_destroy(pc->pc_set);
 278 }
 279
 280 int ptlrpcd_addref(void)
 281 {
 282         int rc = 0;
 283         ENTRY;
 284
 285         mutex_down(&ptlrpcd_sem);
 286         if (++ptlrpcd_users != 1)
 287                 GOTO(out, rc);
 288
 289         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 290         if (rc) {
 291                 --ptlrpcd_users;
 292                 GOTO(out, rc);
 293         }
 294
 295         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 296         if (rc) {
 297                 ptlrpcd_stop(&ptlrpcd_pc);
 298                 --ptlrpcd_users;
 299                 GOTO(out, rc);
 300         }
 301 out:
 302         mutex_up(&ptlrpcd_sem);
 303         RETURN(rc);
 304 }
 305
 306 void ptlrpcd_decref(void)
 307 {
 308         mutex_down(&ptlrpcd_sem);
 309         if (--ptlrpcd_users == 0) {
 310                 ptlrpcd_stop(&ptlrpcd_pc);
 311                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 312         }
 313         mutex_up(&ptlrpcd_sem);
 314 }