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 <linux/version.h>
  31 # include <linux/module.h>
  32 # include <linux/mm.h>
  33 # include <linux/highmem.h>
  34 # include <linux/lustre_dlm.h>
  35 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  36 #  include <linux/workqueue.h>
  37 #  include <linux/smp_lock.h>
  38 # else
  39 #  include <linux/locks.h>
  40 # endif
  41 # include <linux/ctype.h>
  42 # include <linux/init.h>
  43 #else /* __KERNEL__ */
  44 # include <liblustre.h>
  45 # include <ctype.h>
  46 #endif
  47
  48 #include <libcfs/kp30.h>
  49 #include <linux/lustre_net.h>
  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         char                      pc_name[16];
  65 #ifndef __KERNEL__
  66         int                       pc_recurred;
  67         void                     *pc_callback;
  68 #endif
  69 };
  70
  71 static struct ptlrpcd_ctl ptlrpcd_pc;
  72 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  73
  74 static DECLARE_MUTEX(ptlrpcd_sem);
  75 static int ptlrpcd_users = 0;
  76
  77 void ptlrpcd_wake(struct ptlrpc_request *req)
  78 {
  79         struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  80
  81         LASSERT(pc != NULL);
  82
  83         wake_up(&pc->pc_waitq);
  84 }
  85
  86 /* requests that are added to the ptlrpcd queue are sent via
  87  * ptlrpcd_check->ptlrpc_check_set() */
  88 void ptlrpcd_add_req(struct ptlrpc_request *req)
  89 {
  90         struct ptlrpcd_ctl *pc;
  91
  92         if (req->rq_send_state == LUSTRE_IMP_FULL)
  93                 pc = &ptlrpcd_pc;
  94         else
  95                 pc = &ptlrpcd_recovery_pc;
  96
  97         ptlrpc_set_add_new_req(pc->pc_set, req);
  98         req->rq_ptlrpcd_data = pc;
  99
 100         ptlrpcd_wake(req);
 101 }
 102
 103 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
 104 {
 105         struct list_head *tmp, *pos;
 106         struct ptlrpc_request *req;
 107         unsigned long flags;
 108         int rc = 0;
 109         ENTRY;
 110
 111         if (test_bit(LIOD_STOP, &pc->pc_flags))
 112                 RETURN(1);
 113
 114         spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 115         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 116                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 117                 list_del_init(&req->rq_set_chain);
 118                 ptlrpc_set_add_req(pc->pc_set, req);
 119                 rc = 1; /* need to calculate its timeout */
 120         }
 121         spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 122
 123         if (pc->pc_set->set_remaining) {
 124                 rc = rc | ptlrpc_check_set(pc->pc_set);
 125
 126                 /* XXX our set never completes, so we prune the completed
 127                  * reqs after each iteration. boy could this be smarter. */
 128                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 129                         req = list_entry(pos, struct ptlrpc_request,
 130                                          rq_set_chain);
 131                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 132                                 continue;
 133
 134                         list_del_init(&req->rq_set_chain);
 135                         req->rq_set = NULL;
 136                         ptlrpc_req_finished (req);
 137                 }
 138         }
 139
 140         if (rc == 0) {
 141                 /* If new requests have been added, make sure to wake up */
 142                 spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 143                 rc = !list_empty(&pc->pc_set->set_new_requests);
 144                 spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 145         }
 146
 147         RETURN(rc);
 148 }
 149
 150 #ifdef __KERNEL__
 151 /* ptlrpc's code paths like to execute in process context, so we have this
 152  * thread which spins on a set which contains the io rpcs.  llite specifies
 153  * ptlrpcd's set when it pushes pages down into the oscs */
 154 static int ptlrpcd(void *arg)
 155 {
 156         struct ptlrpcd_ctl *pc = arg;
 157         unsigned long flags;
 158         ENTRY;
 159
 160         libcfs_daemonize(pc->pc_name);
 161
 162         SIGNAL_MASK_LOCK(current, flags);
 163         sigfillset(&current->blocked);
 164         RECALC_SIGPENDING;
 165         SIGNAL_MASK_UNLOCK(current, flags);
 166
 167         complete(&pc->pc_starting);
 168
 169         /* this mainloop strongly resembles ptlrpc_set_wait except
 170          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 171          * when there are requests in the set.  new requests come in
 172          * on the set's new_req_list and ptlrpcd_check moves them into
 173          * the set. */
 174         while (1) {
 175                 wait_queue_t set_wait;
 176                 struct l_wait_info lwi;
 177                 int timeout;
 178
 179                 timeout = ptlrpc_set_next_timeout(pc->pc_set) * HZ;
 180                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 181
 182                 /* ala the pinger, wait on pc's waitqueue and the set's */
 183                 init_waitqueue_entry(&set_wait, current);
 184                 add_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 185                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 186                 remove_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 187
 188                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 189                         break;
 190         }
 191         /* wait for inflight requests to drain */
 192         if (!list_empty(&pc->pc_set->set_requests))
 193                 ptlrpc_set_wait(pc->pc_set);
 194         complete(&pc->pc_finishing);
 195         return 0;
 196 }
 197 #else
 198
 199 int ptlrpcd_check_async_rpcs(void *arg)
 200 {
 201         struct ptlrpcd_ctl *pc = arg;
 202         int                  rc = 0;
 203
 204         /* single threaded!! */
 205         pc->pc_recurred++;
 206
 207         if (pc->pc_recurred == 1) {
 208                 rc = ptlrpcd_check(pc);
 209                 if (!rc)
 210                         ptlrpc_expired_set(pc->pc_set);
 211         }
 212
 213         pc->pc_recurred--;
 214         return rc;
 215 }
 216 #endif
 217
 218 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 219 {
 220         int rc;
 221
 222         memset(pc, 0, sizeof(*pc));
 223         init_completion(&pc->pc_starting);
 224         init_completion(&pc->pc_finishing);
 225         init_waitqueue_head(&pc->pc_waitq);
 226         pc->pc_flags = 0;
 227         spin_lock_init(&pc->pc_lock);
 228         INIT_LIST_HEAD(&pc->pc_req_list);
 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         rc = kernel_thread(ptlrpcd, pc, 0);
 237         if (rc < 0)  {
 238                 ptlrpc_set_destroy(pc->pc_set);
 239                 RETURN(rc);
 240         }
 241
 242         wait_for_completion(&pc->pc_starting);
 243 #else
 244         pc->pc_callback =
 245                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 246         (void)rc;
 247 #endif
 248         RETURN(0);
 249 }
 250
 251 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 252 {
 253         set_bit(LIOD_STOP, &pc->pc_flags);
 254         wake_up(&pc->pc_waitq);
 255 #ifdef __KERNEL__
 256         wait_for_completion(&pc->pc_finishing);
 257 #else
 258         liblustre_deregister_wait_callback(pc->pc_callback);
 259 #endif
 260         ptlrpc_set_destroy(pc->pc_set);
 261 }
 262
 263 int ptlrpcd_addref(void)
 264 {
 265         int rc = 0;
 266         ENTRY;
 267
 268         down(&ptlrpcd_sem);
 269         if (++ptlrpcd_users != 1)
 270                 GOTO(out, rc);
 271
 272         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 273         if (rc) {
 274                 --ptlrpcd_users;
 275                 GOTO(out, rc);
 276         }
 277
 278         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 279         if (rc) {
 280                 ptlrpcd_stop(&ptlrpcd_pc);
 281                 --ptlrpcd_users;
 282                 GOTO(out, rc);
 283         }
 284 out:
 285         up(&ptlrpcd_sem);
 286         RETURN(rc);
 287 }
 288
 289 void ptlrpcd_decref(void)
 290 {
 291         down(&ptlrpcd_sem);
 292         if (--ptlrpcd_users == 0) {
 293                 ptlrpcd_stop(&ptlrpcd_pc);
 294                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 295         }
 296         up(&ptlrpcd_sem);
 297 }