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 ptlrpc_request_set *pc_set;
  52         char                      pc_name[16];
  53 #ifndef __KERNEL__
  54         int                       pc_recurred;
  55         void                     *pc_wait_callback;
  56         void                     *pc_idle_callback;
  57 #endif
  58 };
  59
  60 static struct ptlrpcd_ctl ptlrpcd_pc;
  61 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  62
  63 struct semaphore ptlrpcd_sem;
  64 static int ptlrpcd_users = 0;
  65
  66 void ptlrpcd_wake(struct ptlrpc_request *req)
  67 {
  68         struct ptlrpc_request_set *rq_set = req->rq_set;
  69
  70         LASSERT(rq_set != NULL);
  71
  72         cfs_waitq_signal(&rq_set->set_waitq);
  73 }
  74
  75 /* requests that are added to the ptlrpcd queue are sent via
  76  * ptlrpcd_check->ptlrpc_check_set() */
  77 void ptlrpcd_add_req(struct ptlrpc_request *req)
  78 {
  79         struct ptlrpcd_ctl *pc;
  80
  81         if (req->rq_send_state == LUSTRE_IMP_FULL)
  82                 pc = &ptlrpcd_pc;
  83         else
  84                 pc = &ptlrpcd_recovery_pc;
  85
  86         ptlrpc_set_add_new_req(pc->pc_set, req);
  87         cfs_waitq_signal(&pc->pc_set->set_waitq);
  88 }
  89
  90 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
  91 {
  92         struct list_head *tmp, *pos;
  93         struct ptlrpc_request *req;
  94         int rc = 0;
  95         ENTRY;
  96
  97         if (test_bit(LIOD_STOP, &pc->pc_flags))
  98                 RETURN(1);
  99
 100         obd_zombie_impexp_cull();
 101
 102         spin_lock(&pc->pc_set->set_new_req_lock);
 103         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 104                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 105                 list_del_init(&req->rq_set_chain);
 106                 ptlrpc_set_add_req(pc->pc_set, req);
 107                 rc = 1; /* need to calculate its timeout */
 108         }
 109         spin_unlock(&pc->pc_set->set_new_req_lock);
 110
 111         if (pc->pc_set->set_remaining) {
 112                 rc = rc | ptlrpc_check_set(pc->pc_set);
 113
 114                 /* XXX our set never completes, so we prune the completed
 115                  * reqs after each iteration. boy could this be smarter. */
 116                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 117                         req = list_entry(pos, struct ptlrpc_request,
 118                                          rq_set_chain);
 119                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 120                                 continue;
 121
 122                         list_del_init(&req->rq_set_chain);
 123                         req->rq_set = NULL;
 124                         ptlrpc_req_finished (req);
 125                 }
 126         }
 127
 128         if (rc == 0) {
 129                 /* If new requests have been added, make sure to wake up */
 130                 spin_lock(&pc->pc_set->set_new_req_lock);
 131                 rc = !list_empty(&pc->pc_set->set_new_requests);
 132                 spin_unlock(&pc->pc_set->set_new_req_lock);
 133         }
 134
 135         RETURN(rc);
 136 }
 137
 138 #ifdef __KERNEL__
 139 /* ptlrpc's code paths like to execute in process context, so we have this
 140  * thread which spins on a set which contains the io rpcs.  llite specifies
 141  * ptlrpcd's set when it pushes pages down into the oscs */
 142 static int ptlrpcd(void *arg)
 143 {
 144         struct ptlrpcd_ctl *pc = arg;
 145         int rc;
 146         ENTRY;
 147
 148         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 149                 complete(&pc->pc_starting);
 150                 return rc;
 151         }
 152
 153         complete(&pc->pc_starting);
 154
 155         /* this mainloop strongly resembles ptlrpc_set_wait except
 156          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 157          * when there are requests in the set.  new requests come in
 158          * on the set's new_req_list and ptlrpcd_check moves them into
 159          * the set. */
 160         while (1) {
 161                 struct l_wait_info lwi;
 162                 cfs_duration_t timeout;
 163
 164                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
 165                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 166
 167                 l_wait_event(pc->pc_set->set_waitq, ptlrpcd_check(pc), &lwi);
 168
 169                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 170                         break;
 171         }
 172         /* wait for inflight requests to drain */
 173         if (!list_empty(&pc->pc_set->set_requests))
 174                 ptlrpc_set_wait(pc->pc_set);
 175         complete(&pc->pc_finishing);
 176         return 0;
 177 }
 178
 179 static void ptlrpcd_zombie_impexp_notify(void)
 180 {
 181         LASSERT(ptlrpcd_pc.pc_set != NULL); // call before ptlrpcd inited ?
 182
 183         cfs_waitq_signal(&ptlrpcd_pc.pc_set->set_waitq);
 184 }
 185
 186 #else
 187
 188 int ptlrpcd_check_async_rpcs(void *arg)
 189 {
 190         struct ptlrpcd_ctl *pc = arg;
 191         int                  rc = 0;
 192
 193         /* single threaded!! */
 194         pc->pc_recurred++;
 195
 196         if (pc->pc_recurred == 1) {
 197                 rc = ptlrpcd_check(pc);
 198                 if (!rc)
 199                         ptlrpc_expired_set(pc->pc_set);
 200                 /*XXX send replay requests */
 201                 if (pc == &ptlrpcd_recovery_pc)
 202                         rc = ptlrpcd_check(pc);
 203         }
 204
 205         pc->pc_recurred--;
 206         return rc;
 207 }
 208
 209 int ptlrpcd_idle(void *arg)
 210 {
 211         struct ptlrpcd_ctl *pc = arg;
 212
 213         return (list_empty(&pc->pc_set->set_new_requests) &&
 214                 pc->pc_set->set_remaining == 0);
 215 }
 216
 217 #endif
 218
 219 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 220 {
 221         int rc;
 222
 223         ENTRY;
 224         memset(pc, 0, sizeof(*pc));
 225         init_completion(&pc->pc_starting);
 226         init_completion(&pc->pc_finishing);
 227         pc->pc_flags = 0;
 228         spin_lock_init(&pc->pc_lock);
 229         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 230
 231         pc->pc_set = ptlrpc_prep_set();
 232         if (pc->pc_set == NULL)
 233                 RETURN(-ENOMEM);
 234
 235 #ifdef __KERNEL__
 236         /* wake ptlrpcd when zombie imports or exports exist */
 237         obd_zombie_impexp_notify = ptlrpcd_zombie_impexp_notify;
 238
 239         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 240         if (rc < 0)  {
 241                 ptlrpc_set_destroy(pc->pc_set);
 242                 RETURN(rc);
 243         }
 244
 245         wait_for_completion(&pc->pc_starting);
 246 #else
 247         pc->pc_wait_callback =
 248                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 249                                                  &ptlrpcd_check_async_rpcs, pc);
 250         pc->pc_idle_callback =
 251                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 252                                                  &ptlrpcd_idle, pc);
 253         (void)rc;
 254 #endif
 255         RETURN(0);
 256 }
 257
 258 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 259 {
 260         set_bit(LIOD_STOP, &pc->pc_flags);
 261         cfs_waitq_signal(&pc->pc_set->set_waitq);
 262 #ifdef __KERNEL__
 263         obd_zombie_impexp_notify = NULL;
 264         wait_for_completion(&pc->pc_finishing);
 265 #else
 266         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 267         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 268 #endif
 269         ptlrpc_set_destroy(pc->pc_set);
 270 }
 271
 272 int ptlrpcd_addref(void)
 273 {
 274         int rc = 0;
 275         ENTRY;
 276
 277         mutex_down(&ptlrpcd_sem);
 278         if (++ptlrpcd_users != 1)
 279                 GOTO(out, rc);
 280
 281         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 282         if (rc) {
 283                 --ptlrpcd_users;
 284                 GOTO(out, rc);
 285         }
 286
 287         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 288         if (rc) {
 289                 ptlrpcd_stop(&ptlrpcd_pc);
 290                 --ptlrpcd_users;
 291                 GOTO(out, rc);
 292         }
 293 out:
 294         mutex_up(&ptlrpcd_sem);
 295         RETURN(rc);
 296 }
 297
 298 void ptlrpcd_decref(void)
 299 {
 300         mutex_down(&ptlrpcd_sem);
 301         if (--ptlrpcd_users == 0) {
 302                 ptlrpcd_stop(&ptlrpcd_pc);
 303                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 304         }
 305         mutex_up(&ptlrpcd_sem);
 306 }