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_support.h> /* for OBD_FAIL_CHECK */
  42 #include <lprocfs_status.h>
  43
  44 #define LIOD_STOP 0
  45 struct ptlrpcd_ctl {
  46         unsigned long             pc_flags;
  47         spinlock_t                pc_lock;
  48         struct completion         pc_starting;
  49         struct completion         pc_finishing;
  50         struct list_head          pc_req_list;
  51         cfs_waitq_t               pc_waitq;
  52         struct ptlrpc_request_set *pc_set;
  53         char                      pc_name[16];
  54 #ifndef __KERNEL__
  55         int                       pc_recurred;
  56         void                     *pc_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 ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  69
  70         LASSERT(pc != NULL);
  71
  72         cfs_waitq_signal(&pc->pc_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         req->rq_ptlrpcd_data = pc;
  87         ptlrpc_set_add_new_req(pc->pc_set, req);
  88         wake_up(&pc->pc_waitq);
  89 }
  90
  91 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
  92 {
  93         struct list_head *tmp, *pos;
  94         struct ptlrpc_request *req;
  95         unsigned long flags;
  96         int rc = 0;
  97         ENTRY;
  98
  99         if (test_bit(LIOD_STOP, &pc->pc_flags))
 100                 RETURN(1);
 101
 102         spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 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_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 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_irqsave(&pc->pc_set->set_new_req_lock, flags);
 131                 rc = !list_empty(&pc->pc_set->set_new_requests);
 132                 spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 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         ENTRY;
 146
 147         cfs_daemonize(pc->pc_name);
 148
 149         complete(&pc->pc_starting);
 150
 151         /* this mainloop strongly resembles ptlrpc_set_wait except
 152          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 153          * when there are requests in the set.  new requests come in
 154          * on the set's new_req_list and ptlrpcd_check moves them into
 155          * the set. */
 156         while (1) {
 157                 cfs_waitlink_t set_wait;
 158                 struct l_wait_info lwi;
 159                 cfs_duration_t timeout;
 160
 161                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
 162                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 163
 164                 /* ala the pinger, wait on pc's waitqueue and the set's */
 165                 cfs_waitlink_init(&set_wait);
 166                 cfs_waitq_add(&pc->pc_set->set_waitq, &set_wait);
 167                 cfs_waitq_forward(&set_wait, &pc->pc_waitq);
 168                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 169                 cfs_waitq_del(&pc->pc_set->set_waitq, &set_wait);
 170
 171                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 172                         break;
 173         }
 174         /* wait for inflight requests to drain */
 175         if (!list_empty(&pc->pc_set->set_requests))
 176                 ptlrpc_set_wait(pc->pc_set);
 177         complete(&pc->pc_finishing);
 178         return 0;
 179 }
 180 #else
 181
 182 int ptlrpcd_check_async_rpcs(void *arg)
 183 {
 184         struct ptlrpcd_ctl *pc = arg;
 185         int                  rc = 0;
 186
 187         /* single threaded!! */
 188         pc->pc_recurred++;
 189
 190         if (pc->pc_recurred == 1) {
 191                 rc = ptlrpcd_check(pc);
 192                 if (!rc)
 193                         ptlrpc_expired_set(pc->pc_set);
 194         }
 195
 196         pc->pc_recurred--;
 197         return rc;
 198 }
 199 #endif
 200
 201 static int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 202 {
 203         int rc;
 204
 205         ENTRY;
 206         memset(pc, 0, sizeof(*pc));
 207         init_completion(&pc->pc_starting);
 208         init_completion(&pc->pc_finishing);
 209         cfs_waitq_init(&pc->pc_waitq);
 210         pc->pc_flags = 0;
 211         spin_lock_init(&pc->pc_lock);
 212         CFS_INIT_LIST_HEAD(&pc->pc_req_list);
 213         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 214
 215         pc->pc_set = ptlrpc_prep_set();
 216         if (pc->pc_set == NULL)
 217                 RETURN(-ENOMEM);
 218
 219 #ifdef __KERNEL__
 220         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 221         if (rc < 0)  {
 222                 ptlrpc_set_destroy(pc->pc_set);
 223                 RETURN(rc);
 224         }
 225
 226         wait_for_completion(&pc->pc_starting);
 227 #else
 228         pc->pc_callback =
 229                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 230         (void)rc;
 231 #endif
 232         RETURN(0);
 233 }
 234
 235 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 236 {
 237         set_bit(LIOD_STOP, &pc->pc_flags);
 238         cfs_waitq_signal(&pc->pc_waitq);
 239 #ifdef __KERNEL__
 240         wait_for_completion(&pc->pc_finishing);
 241 #else
 242         liblustre_deregister_wait_callback(pc->pc_callback);
 243 #endif
 244         ptlrpc_set_destroy(pc->pc_set);
 245 }
 246
 247 int ptlrpcd_addref(void)
 248 {
 249         int rc = 0;
 250         ENTRY;
 251
 252         mutex_down(&ptlrpcd_sem);
 253         if (++ptlrpcd_users != 1)
 254                 GOTO(out, rc);
 255
 256         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 257         if (rc) {
 258                 --ptlrpcd_users;
 259                 GOTO(out, rc);
 260         }
 261
 262         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 263         if (rc) {
 264                 ptlrpcd_stop(&ptlrpcd_pc);
 265                 --ptlrpcd_users;
 266                 GOTO(out, rc);
 267         }
 268 out:
 269         mutex_up(&ptlrpcd_sem);
 270         RETURN(rc);
 271 }
 272
 273 void ptlrpcd_decref(void)
 274 {
 275         mutex_down(&ptlrpcd_sem);
 276         if (--ptlrpcd_users == 0) {
 277                 ptlrpcd_stop(&ptlrpcd_pc);
 278                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 279         }
 280         mutex_up(&ptlrpcd_sem);
 281 }