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/ctype.h>
  47 # include <linux/init.h>
  48 #else
  49 # include <ctype.h>
  50 #endif
  51
  52 #include <linux/lustre_ha.h>
  53 #include <linux/obd_support.h> /* for OBD_FAIL_CHECK */
  54 #include <linux/lprocfs_status.h>
  55
  56 #define LIOD_STOP 0
  57 static struct ptlrpcd_ctl {
  58         unsigned long             pc_flags;
  59         spinlock_t                pc_lock;
  60         struct completion         pc_starting;
  61         struct completion         pc_finishing;
  62         struct list_head          pc_req_list;
  63         wait_queue_head_t         pc_waitq;
  64         struct ptlrpc_request_set *pc_set;
  65 } ptlrpcd_pc;
  66
  67 static DECLARE_MUTEX(ptlrpcd_sem);
  68 static int ptlrpcd_users = 0;
  69
  70 void ptlrpcd_add_req(struct ptlrpc_request *req)
  71 {
  72         struct ptlrpcd_ctl *pc = &ptlrpcd_pc;
  73
  74         ptlrpc_set_add_new_req(pc->pc_set, req);
  75         wake_up(&pc->pc_waitq);
  76 }
  77
  78 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
  79 {
  80         struct list_head *tmp, *pos;
  81         struct ptlrpc_request *req;
  82         unsigned long flags;
  83         int rc = 0;
  84         ENTRY;
  85
  86         if (test_bit(LIOD_STOP, &pc->pc_flags))
  87                 RETURN(1);
  88
  89         spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
  90         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
  91                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
  92                 list_del_init(&req->rq_set_chain);
  93                 ptlrpc_set_add_req(pc->pc_set, req);
  94                 rc = 1; /* need to calculate its timeout */
  95         }
  96         spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
  97
  98         if (pc->pc_set->set_remaining) {
  99                 rc = rc | ptlrpc_check_set(pc->pc_set);
 100
 101                 /* XXX our set never completes, so we prune the completed
 102                  * reqs after each iteration. boy could this be smarter. */
 103                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 104                         req = list_entry(pos, struct ptlrpc_request,
 105                                          rq_set_chain);
 106                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 107                                 continue;
 108
 109                         list_del_init(&req->rq_set_chain);
 110                         req->rq_set = NULL;
 111                         ptlrpc_req_finished (req);
 112                 }
 113         }
 114
 115         RETURN(rc);
 116 }
 117
 118 #ifdef __KERNEL__
 119 /* ptlrpc's code paths like to execute in process context, so we have this
 120  * thread which spins on a set which contains the io rpcs.  llite specifies
 121  * ptlrpcd's set when it pushes pages down into the oscs */
 122 static int ptlrpcd(void *arg)
 123 {
 124         struct ptlrpcd_ctl *pc = arg;
 125         unsigned long flags;
 126         ENTRY;
 127
 128         kportal_daemonize("ptlrpcd");
 129
 130         SIGNAL_MASK_LOCK(current, flags);
 131         sigfillset(&current->blocked);
 132         RECALC_SIGPENDING;
 133         SIGNAL_MASK_UNLOCK(current, flags);
 134
 135         complete(&pc->pc_starting);
 136
 137         /* like kswapd */
 138         current->flags |= PF_MEMALLOC;
 139
 140         /* this mainloop strongly resembles ptlrpc_set_wait except
 141          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 142          * when there are requests in the set.  new requests come in
 143          * on the set's new_req_list and ptlrpcd_check moves them into
 144          * the set. */
 145         while (1) {
 146                 wait_queue_t set_wait;
 147                 struct l_wait_info lwi;
 148                 int timeout;
 149
 150                 timeout = ptlrpc_set_next_timeout(pc->pc_set) * HZ;
 151                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 152
 153                 /* ala the pinger, wait on pc's waitqueue and the set's */
 154                 init_waitqueue_entry(&set_wait, current);
 155                 add_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 156                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 157                 remove_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 158
 159                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 160                         break;
 161         }
 162         /* XXX should be making sure we don't have anything in flight */
 163         complete(&pc->pc_finishing);
 164         return 0;
 165 }
 166 #else
 167 static int ptlrpcd_recurred = 0;
 168 static void *ptlrpcd_callback;
 169
 170 int ptlrpcd_check_async_rpcs(void *arg)
 171 {
 172         struct ptlrpcd_ctl *pc = arg;
 173         int                  rc = 0;
 174
 175         /* single threaded!! */
 176         ptlrpcd_recurred++;
 177
 178         if (ptlrpcd_recurred == 1)
 179                 rc = ptlrpcd_check(pc);
 180
 181         ptlrpcd_recurred--;
 182         return rc;
 183 }
 184 #endif
 185
 186 int ptlrpcd_addref(void)
 187 {
 188         struct ptlrpcd_ctl *pc = &ptlrpcd_pc;
 189         int rc = 0;
 190         ENTRY;
 191
 192         down(&ptlrpcd_sem);
 193         if (++ptlrpcd_users != 1)
 194                 GOTO(out, rc);
 195
 196         memset(pc, 0, sizeof(*pc));
 197         init_completion(&pc->pc_starting);
 198         init_completion(&pc->pc_finishing);
 199         init_waitqueue_head(&pc->pc_waitq);
 200         pc->pc_flags = 0;
 201         spin_lock_init(&pc->pc_lock);
 202         INIT_LIST_HEAD(&pc->pc_req_list);
 203
 204         pc->pc_set = ptlrpc_prep_set();
 205         if (pc->pc_set == NULL)
 206                 GOTO(out, rc = -ENOMEM);
 207
 208 #ifdef __KERNEL__
 209         if (kernel_thread(ptlrpcd, pc, 0) < 0)  {
 210                 ptlrpc_set_destroy(pc->pc_set);
 211                 GOTO(out, rc = -ECHILD);
 212         }
 213
 214         wait_for_completion(&pc->pc_starting);
 215 #else
 216         ptlrpcd_callback =
 217                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 218 #endif
 219 out:
 220         up(&ptlrpcd_sem);
 221         RETURN(rc);
 222 }
 223
 224 void ptlrpcd_decref(void)
 225 {
 226         struct ptlrpcd_ctl *pc = &ptlrpcd_pc;
 227
 228         down(&ptlrpcd_sem);
 229         if (--ptlrpcd_users == 0) {
 230                 set_bit(LIOD_STOP, &pc->pc_flags);
 231                 wake_up(&pc->pc_waitq);
 232 #ifdef __KERNEL__
 233                 wait_for_completion(&pc->pc_finishing);
 234 #else
 235                 liblustre_deregister_wait_callback(ptlrpcd_callback);
 236 #endif
 237                 ptlrpc_set_destroy(pc->pc_set);
 238         }
 239         up(&ptlrpcd_sem);
 240 }