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 # include <linux/ctype.h>
  39 # include <linux/init.h>
  40 #else /* __KERNEL__ */
  41 # include <liblustre.h>
  42 # include <ctype.h>
  43 #endif
  44
  45 #include <linux/kp30.h>
  46 #include <linux/lustre_net.h>
  47
  48 #include <linux/lustre_ha.h>
  49 #include <linux/obd_support.h> /* for OBD_FAIL_CHECK */
  50 #include <linux/lprocfs_status.h>
  51
  52 #define LIOD_STOP 0
  53 struct ptlrpcd_ctl {
  54         unsigned long             pc_flags;
  55         spinlock_t                pc_lock;
  56         struct completion         pc_starting;
  57         struct completion         pc_finishing;
  58         struct list_head          pc_req_list;
  59         wait_queue_head_t         pc_waitq;
  60         struct ptlrpc_request_set *pc_set;
  61 #ifndef __KERNEL__
  62         int                       pc_recurred;
  63         void                     *pc_callback;
  64 #endif
  65 };
  66
  67 static struct ptlrpcd_ctl ptlrpcd_pc;
  68 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  69
  70 static DECLARE_MUTEX(ptlrpcd_sem);
  71 static int ptlrpcd_users = 0;
  72
  73 void ptlrpcd_wake(struct ptlrpc_request *req)
  74 {
  75         struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  76
  77         LASSERT(pc != NULL);
  78
  79         wake_up(&pc->pc_waitq);
  80 }
  81
  82 void ptlrpcd_add_req(struct ptlrpc_request *req)
  83 {
  84         struct ptlrpcd_ctl *pc;
  85
  86         if (req->rq_send_state == LUSTRE_IMP_FULL)
  87                 pc = &ptlrpcd_pc;
  88         else
  89                 pc = &ptlrpcd_recovery_pc;
  90
  91         ptlrpc_set_add_new_req(pc->pc_set, req);
  92         req->rq_ptlrpcd_data = pc;
  93
  94         ptlrpcd_wake(req);
  95 }
  96
  97 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
  98 {
  99         struct list_head *tmp, *pos;
 100         struct ptlrpc_request *req;
 101         unsigned long flags;
 102         int rc = 0;
 103         ENTRY;
 104
 105         if (test_bit(LIOD_STOP, &pc->pc_flags))
 106                 RETURN(1);
 107
 108         spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 109         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 110                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 111                 list_del_init(&req->rq_set_chain);
 112                 ptlrpc_set_add_req(pc->pc_set, req);
 113                 rc = 1; /* need to calculate its timeout */
 114         }
 115         spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 116
 117         if (pc->pc_set->set_remaining) {
 118                 rc = rc | ptlrpc_check_set(pc->pc_set);
 119
 120                 /* XXX our set never completes, so we prune the completed
 121                  * reqs after each iteration. boy could this be smarter. */
 122                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 123                         req = list_entry(pos, struct ptlrpc_request,
 124                                          rq_set_chain);
 125                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 126                                 continue;
 127
 128                         list_del_init(&req->rq_set_chain);
 129                         req->rq_set = NULL;
 130                         ptlrpc_req_finished (req);
 131                 }
 132         }
 133
 134         if (rc == 0) {
 135                 /* If new requests have been added, make sure to wake up */
 136                 spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 137                 rc = !list_empty(&pc->pc_set->set_new_requests);
 138                 spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 139         }
 140
 141         RETURN(rc);
 142 }
 143
 144 #ifdef __KERNEL__
 145 /* ptlrpc's code paths like to execute in process context, so we have this
 146  * thread which spins on a set which contains the io rpcs.  llite specifies
 147  * ptlrpcd's set when it pushes pages down into the oscs */
 148 static int ptlrpcd(void *arg)
 149 {
 150         struct ptlrpcd_ctl *pc = arg;
 151         unsigned long flags;
 152         ENTRY;
 153
 154         kportal_daemonize("ptlrpcd");
 155
 156         SIGNAL_MASK_LOCK(current, flags);
 157         sigfillset(&current->blocked);
 158         RECALC_SIGPENDING;
 159         SIGNAL_MASK_UNLOCK(current, flags);
 160
 161         complete(&pc->pc_starting);
 162
 163         /* this mainloop strongly resembles ptlrpc_set_wait except
 164          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 165          * when there are requests in the set.  new requests come in
 166          * on the set's new_req_list and ptlrpcd_check moves them into
 167          * the set. */
 168         while (1) {
 169                 wait_queue_t set_wait;
 170                 struct l_wait_info lwi;
 171                 int timeout;
 172
 173                 timeout = ptlrpc_set_next_timeout(pc->pc_set) * HZ;
 174                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 175
 176                 /* ala the pinger, wait on pc's waitqueue and the set's */
 177                 init_waitqueue_entry(&set_wait, current);
 178                 add_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 179                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 180                 remove_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 181
 182                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 183                         break;
 184         }
 185         /* wait for inflight requests to drain */
 186         if (!list_empty(&pc->pc_set->set_requests))
 187                 ptlrpc_set_wait(pc->pc_set);
 188         complete(&pc->pc_finishing);
 189         return 0;
 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
 204         pc->pc_recurred--;
 205         return rc;
 206 }
 207 #endif
 208
 209 static int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 210 {
 211         int rc = 0;
 212
 213         memset(pc, 0, sizeof(*pc));
 214         init_completion(&pc->pc_starting);
 215         init_completion(&pc->pc_finishing);
 216         init_waitqueue_head(&pc->pc_waitq);
 217         pc->pc_flags = 0;
 218         spin_lock_init(&pc->pc_lock);
 219         INIT_LIST_HEAD(&pc->pc_req_list);
 220
 221         pc->pc_set = ptlrpc_prep_set();
 222         if (pc->pc_set == NULL)
 223                 GOTO(out, rc = -ENOMEM);
 224
 225 #ifdef __KERNEL__
 226         if (kernel_thread(ptlrpcd, pc, 0) < 0)  {
 227                 ptlrpc_set_destroy(pc->pc_set);
 228                 GOTO(out, rc = -ECHILD);
 229         }
 230
 231         wait_for_completion(&pc->pc_starting);
 232 #else
 233         pc->pc_callback =
 234                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 235 #endif
 236 out:
 237         RETURN(rc);
 238 }
 239
 240 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 241 {
 242         set_bit(LIOD_STOP, &pc->pc_flags);
 243         wake_up(&pc->pc_waitq);
 244 #ifdef __KERNEL__
 245         wait_for_completion(&pc->pc_finishing);
 246 #else
 247         liblustre_deregister_wait_callback(pc->pc_callback);
 248 #endif
 249         ptlrpc_set_destroy(pc->pc_set);
 250 }
 251
 252 int ptlrpcd_addref(void)
 253 {
 254         int rc = 0;
 255         ENTRY;
 256
 257         down(&ptlrpcd_sem);
 258         if (++ptlrpcd_users != 1)
 259                 GOTO(out, rc);
 260
 261         rc = ptlrpcd_start(&ptlrpcd_pc);
 262         if (rc) {
 263                 --ptlrpcd_users;
 264                 GOTO(out, rc);
 265         }
 266
 267         rc = ptlrpcd_start(&ptlrpcd_recovery_pc);
 268         if (rc) {
 269                 ptlrpcd_stop(&ptlrpcd_pc);
 270                 --ptlrpcd_users;
 271                 GOTO(out, rc);
 272         }
 273 out:
 274         up(&ptlrpcd_sem);
 275         RETURN(rc);
 276 }
 277
 278 void ptlrpcd_decref(void)
 279 {
 280         down(&ptlrpcd_sem);
 281         if (--ptlrpcd_users == 0) {
 282                 ptlrpcd_stop(&ptlrpcd_pc);
 283                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 284         }
 285         up(&ptlrpcd_sem);
 286 }