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 Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_RPC
  25
  26 #ifdef __KERNEL__
  27 # include <linux/version.h>
  28 # include <linux/module.h>
  29 # include <linux/mm.h>
  30 # include <linux/highmem.h>
  31 # include <linux/lustre_dlm.h>
  32 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  33 #  include <linux/workqueue.h>
  34 #  include <linux/smp_lock.h>
  35 # else
  36 #  include <linux/locks.h>
  37 # endif
  38 #else /* __KERNEL__ */
  39 # include <liblustre.h>
  40 #endif
  41
  42 #include <linux/kp30.h>
  43 #include <linux/lustre_net.h>
  44
  45 #ifndef  __CYGWIN__
  46 # include <linux/init.h>
  47 #else
  48 # include <ctype.h>
  49 #endif
  50
  51 #include <linux/lustre_ha.h>
  52 #include <linux/obd_support.h> /* for OBD_FAIL_CHECK */
  53 #include <linux/lprocfs_status.h>
  54
  55 #define LIOD_STOP 0
  56 struct ptlrpcd_ctl {
  57         unsigned long             pc_flags;
  58         spinlock_t                pc_lock;
  59         struct completion         pc_starting;
  60         struct completion         pc_finishing;
  61         struct list_head          pc_req_list;
  62         wait_queue_head_t         pc_waitq;
  63         struct ptlrpc_request_set *pc_set;
  64 #ifndef __KERNEL__
  65         int                       pc_recurred;
  66         void                     *pc_callback;
  67 #endif
  68 };
  69
  70 static struct ptlrpcd_ctl ptlrpcd_pc;
  71 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  72
  73 static DECLARE_MUTEX(ptlrpcd_sem);
  74 static int ptlrpcd_users = 0;
  75
  76 void ptlrpcd_wake(struct ptlrpc_request *req)
  77 {
  78         struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  79
  80         LASSERT(pc != NULL);
  81
  82         wake_up(&pc->pc_waitq);
  83 }
  84
  85 void ptlrpcd_add_req(struct ptlrpc_request *req)
  86 {
  87         struct ptlrpcd_ctl *pc;
  88
  89         if (req->rq_send_state == LUSTRE_IMP_FULL)
  90                 pc = &ptlrpcd_pc;
  91         else
  92                 pc = &ptlrpcd_recovery_pc;
  93
  94         ptlrpc_set_add_new_req(pc->pc_set, req);
  95         req->rq_ptlrpcd_data = pc;
  96
  97         ptlrpcd_wake(req);
  98 }
  99
 100 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
 101 {
 102         struct list_head *tmp, *pos;
 103         struct ptlrpc_request *req;
 104         unsigned long flags;
 105         int rc = 0;
 106         ENTRY;
 107
 108         if (test_bit(LIOD_STOP, &pc->pc_flags))
 109                 RETURN(1);
 110
 111         spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 112         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 113                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 114                 list_del_init(&req->rq_set_chain);
 115                 ptlrpc_set_add_req(pc->pc_set, req);
 116                 rc = 1; /* need to calculate its timeout */
 117         }
 118         spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 119
 120         if (pc->pc_set->set_remaining) {
 121                 rc = rc | ptlrpc_check_set(pc->pc_set);
 122
 123                 /* XXX our set never completes, so we prune the completed
 124                  * reqs after each iteration. boy could this be smarter. */
 125                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 126                         req = list_entry(pos, struct ptlrpc_request,
 127                                          rq_set_chain);
 128                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 129                                 continue;
 130
 131                         list_del_init(&req->rq_set_chain);
 132                         req->rq_set = NULL;
 133                         ptlrpc_req_finished (req);
 134                 }
 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         unsigned long flags;
 148         ENTRY;
 149
 150         kportal_daemonize("ptlrpcd");
 151
 152         SIGNAL_MASK_LOCK(current, flags);
 153         sigfillset(&current->blocked);
 154         RECALC_SIGPENDING;
 155         SIGNAL_MASK_UNLOCK(current, flags);
 156
 157         complete(&pc->pc_starting);
 158
 159         /* this mainloop strongly resembles ptlrpc_set_wait except
 160          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 161          * when there are requests in the set.  new requests come in
 162          * on the set's new_req_list and ptlrpcd_check moves them into
 163          * the set. */
 164         while (1) {
 165                 wait_queue_t set_wait;
 166                 struct l_wait_info lwi;
 167                 int timeout;
 168
 169                 timeout = ptlrpc_set_next_timeout(pc->pc_set) * HZ;
 170                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 171
 172                 /* ala the pinger, wait on pc's waitqueue and the set's */
 173                 init_waitqueue_entry(&set_wait, current);
 174                 add_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 175                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 176                 remove_wait_queue(&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 #else
 188
 189 int ptlrpcd_check_async_rpcs(void *arg)
 190 {
 191         struct ptlrpcd_ctl *pc = arg;
 192         int                  rc = 0;
 193
 194         /* single threaded!! */
 195         pc->pc_recurred++;
 196
 197         if (pc->pc_recurred == 1)
 198                 rc = ptlrpcd_check(pc);
 199
 200         pc->pc_recurred--;
 201         return rc;
 202 }
 203 #endif
 204
 205 static int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 206 {
 207         int rc = 0;
 208
 209         memset(pc, 0, sizeof(*pc));
 210         init_completion(&pc->pc_starting);
 211         init_completion(&pc->pc_finishing);
 212         init_waitqueue_head(&pc->pc_waitq);
 213         pc->pc_flags = 0;
 214         spin_lock_init(&pc->pc_lock);
 215         INIT_LIST_HEAD(&pc->pc_req_list);
 216
 217         pc->pc_set = ptlrpc_prep_set();
 218         if (pc->pc_set == NULL)
 219                 GOTO(out, rc = -ENOMEM);
 220
 221 #ifdef __KERNEL__
 222         if (kernel_thread(ptlrpcd, pc, 0) < 0)  {
 223                 ptlrpc_set_destroy(pc->pc_set);
 224                 GOTO(out, rc = -ECHILD);
 225         }
 226
 227         wait_for_completion(&pc->pc_starting);
 228 #else
 229         pc->pc_callback =
 230                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 231 #endif
 232 out:
 233         RETURN(rc);
 234 }
 235
 236 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 237 {
 238         set_bit(LIOD_STOP, &pc->pc_flags);
 239         wake_up(&pc->pc_waitq);
 240 #ifdef __KERNEL__
 241         wait_for_completion(&pc->pc_finishing);
 242 #else
 243         liblustre_deregister_wait_callback(pc->pc_callback);
 244 #endif
 245         ptlrpc_set_destroy(pc->pc_set);
 246 }
 247
 248 int ptlrpcd_addref(void)
 249 {
 250         int rc = 0;
 251         ENTRY;
 252
 253         down(&ptlrpcd_sem);
 254         if (++ptlrpcd_users != 1)
 255                 GOTO(out, rc);
 256
 257         rc = ptlrpcd_start(&ptlrpcd_pc);
 258         if (rc) {
 259                 --ptlrpcd_users;
 260                 GOTO(out, rc);
 261         }
 262
 263         rc = ptlrpcd_start(&ptlrpcd_recovery_pc);
 264         if (rc) {
 265                 ptlrpcd_stop(&ptlrpcd_pc);
 266                 --ptlrpcd_users;
 267                 GOTO(out, rc);
 268         }
 269 out:
 270         up(&ptlrpcd_sem);
 271         RETURN(rc);
 272 }
 273
 274 void ptlrpcd_decref(void)
 275 {
 276         down(&ptlrpcd_sem);
 277         if (--ptlrpcd_users == 0) {
 278                 ptlrpcd_stop(&ptlrpcd_pc);
 279                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 280         }
 281         up(&ptlrpcd_sem);
 282 }