lustre/ptlrpc/ptlrpcd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/ptlrpcd.c
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_RPC
  40
  41 #ifdef __KERNEL__
  42 # include <libcfs/libcfs.h>
  43 #else /* __KERNEL__ */
  44 # include <liblustre.h>
  45 # include <ctype.h>
  46 #endif
  47
  48 #include <lustre_net.h>
  49 # include <lustre_lib.h>
  50
  51 #include <lustre_ha.h>
  52 #include <obd_class.h>   /* for obd_zombie */
  53 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  54 #include <lprocfs_status.h>
  55
  56 static struct ptlrpcd_ctl ptlrpcd_pc;
  57 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  58
  59 struct semaphore ptlrpcd_sem;
  60 static int ptlrpcd_users = 0;
  61
  62 void ptlrpcd_wake(struct ptlrpc_request *req)
  63 {
  64         struct ptlrpc_request_set *rq_set = req->rq_set;
  65
  66         LASSERT(rq_set != NULL);
  67
  68         cfs_waitq_signal(&rq_set->set_waitq);
  69 }
  70
  71 /*
  72  * Requests that are added to the ptlrpcd queue are sent via
  73  * ptlrpcd_check->ptlrpc_check_set().
  74  */
  75 void ptlrpcd_add_req(struct ptlrpc_request *req)
  76 {
  77         struct ptlrpcd_ctl *pc;
  78         int rc;
  79
  80         if (req->rq_send_state == LUSTRE_IMP_FULL)
  81                 pc = &ptlrpcd_pc;
  82         else
  83                 pc = &ptlrpcd_recovery_pc;
  84
  85         rc = ptlrpc_set_add_new_req(pc, req);
  86         if (rc) {
  87                 int (*interpreter)(struct ptlrpc_request *,
  88                                    void *, int);
  89
  90                 interpreter = req->rq_interpret_reply;
  91
  92                 /*
  93                  * Thread is probably in stop now so we need to
  94                  * kill this rpc as it was not added. Let's call
  95                  * interpret for it to let know we're killing it
  96                  * so that higher levels might free assosiated
  97                  * resources.
  98                  */
  99                 req->rq_status = -EBADR;
 100                 interpreter(req, &req->rq_async_args,
 101                             req->rq_status);
 102                 req->rq_set = NULL;
 103                 ptlrpc_req_finished(req);
 104         }
 105 }
 106
 107 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
 108 {
 109         struct list_head *tmp, *pos;
 110         struct ptlrpc_request *req;
 111         int rc = 0;
 112         ENTRY;
 113
 114         if (test_bit(LIOD_STOP, &pc->pc_flags))
 115                 RETURN(1);
 116
 117         spin_lock(&pc->pc_set->set_new_req_lock);
 118         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 119                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 120                 list_del_init(&req->rq_set_chain);
 121                 ptlrpc_set_add_req(pc->pc_set, req);
 122                 /*
 123                  * Need to calculate its timeout.
 124                  */
 125                 rc = 1;
 126         }
 127         spin_unlock(&pc->pc_set->set_new_req_lock);
 128
 129         if (pc->pc_set->set_remaining) {
 130                 rc = rc | ptlrpc_check_set(pc->pc_set);
 131
 132                 /*
 133                  * XXX: our set never completes, so we prune the completed
 134                  * reqs after each iteration. boy could this be smarter.
 135                  */
 136                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 137                         req = list_entry(pos, struct ptlrpc_request,
 138                                          rq_set_chain);
 139                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 140                                 continue;
 141
 142                         list_del_init(&req->rq_set_chain);
 143                         req->rq_set = NULL;
 144                         ptlrpc_req_finished (req);
 145                 }
 146         }
 147
 148         if (rc == 0) {
 149                 /*
 150                  * If new requests have been added, make sure to wake up.
 151                  */
 152                 spin_lock(&pc->pc_set->set_new_req_lock);
 153                 rc = !list_empty(&pc->pc_set->set_new_requests);
 154                 spin_unlock(&pc->pc_set->set_new_req_lock);
 155         }
 156
 157         RETURN(rc);
 158 }
 159
 160 #ifdef __KERNEL__
 161 /*
 162  * ptlrpc's code paths like to execute in process context, so we have this
 163  * thread which spins on a set which contains the io rpcs. llite specifies
 164  * ptlrpcd's set when it pushes pages down into the oscs.
 165  */
 166 static int ptlrpcd(void *arg)
 167 {
 168         struct ptlrpcd_ctl *pc = arg;
 169         int rc;
 170         ENTRY;
 171
 172         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 173                 complete(&pc->pc_starting);
 174                 goto out;
 175         }
 176
 177         complete(&pc->pc_starting);
 178
 179         /*
 180          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 181          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 182          * there are requests in the set. New requests come in on the set's
 183          * new_req_list and ptlrpcd_check() moves them into the set.
 184          */
 185         while (1) {
 186                 struct l_wait_info lwi;
 187                 cfs_duration_t timeout;
 188
 189                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
 190                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 191
 192                 l_wait_event(pc->pc_set->set_waitq, ptlrpcd_check(pc), &lwi);
 193
 194                 /*
 195                  * Abort inflight rpcs for forced stop case.
 196                  */
 197                 if (test_bit(LIOD_STOP_FORCE, &pc->pc_flags))
 198                         ptlrpc_abort_set(pc->pc_set);
 199
 200                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 201                         break;
 202         }
 203
 204         /*
 205          * Wait for inflight requests to drain.
 206          */
 207         if (!list_empty(&pc->pc_set->set_requests))
 208                 ptlrpc_set_wait(pc->pc_set);
 209
 210         complete(&pc->pc_finishing);
 211 out:
 212         clear_bit(LIOD_START, &pc->pc_flags);
 213         clear_bit(LIOD_STOP, &pc->pc_flags);
 214         return 0;
 215 }
 216
 217 #else
 218
 219 int ptlrpcd_check_async_rpcs(void *arg)
 220 {
 221         struct ptlrpcd_ctl *pc = arg;
 222         int                  rc = 0;
 223
 224         /*
 225          * Single threaded!!
 226          */
 227         pc->pc_recurred++;
 228
 229         if (pc->pc_recurred == 1) {
 230                 rc = ptlrpcd_check(pc);
 231                 if (!rc)
 232                         ptlrpc_expired_set(pc->pc_set);
 233                 /*
 234                  * XXX: send replay requests.
 235                  */
 236                 if (pc == &ptlrpcd_recovery_pc)
 237                         rc = ptlrpcd_check(pc);
 238         }
 239
 240         pc->pc_recurred--;
 241         return rc;
 242 }
 243
 244 int ptlrpcd_idle(void *arg)
 245 {
 246         struct ptlrpcd_ctl *pc = arg;
 247
 248         return (list_empty(&pc->pc_set->set_new_requests) &&
 249                 pc->pc_set->set_remaining == 0);
 250 }
 251
 252 #endif
 253
 254 int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 255 {
 256         int rc = 0;
 257         ENTRY;
 258
 259         /*
 260          * Do not allow start second thread for one pc.
 261          */
 262         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 263                 CERROR("Starting second thread (%s) for same pc %p\n",
 264                        name, pc);
 265                 RETURN(-EALREADY);
 266         }
 267
 268         init_completion(&pc->pc_starting);
 269         init_completion(&pc->pc_finishing);
 270         spin_lock_init(&pc->pc_lock);
 271         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 272
 273         pc->pc_set = ptlrpc_prep_set();
 274         if (pc->pc_set == NULL)
 275                 GOTO(out, rc = -ENOMEM);
 276
 277 #ifdef __KERNEL__
 278         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 279         if (rc < 0)  {
 280                 ptlrpc_set_destroy(pc->pc_set);
 281                 GOTO(out, rc);
 282         }
 283         rc = 0;
 284         wait_for_completion(&pc->pc_starting);
 285 #else
 286         pc->pc_wait_callback =
 287                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 288                                                  &ptlrpcd_check_async_rpcs, pc);
 289         pc->pc_idle_callback =
 290                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 291                                                  &ptlrpcd_idle, pc);
 292 #endif
 293 out:
 294         if (rc)
 295                 clear_bit(LIOD_START, &pc->pc_flags);
 296         RETURN(rc);
 297 }
 298
 299 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 300 {
 301         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 302                 CERROR("Thread for pc %p was not started\n", pc);
 303                 return;
 304         }
 305
 306         set_bit(LIOD_STOP, &pc->pc_flags);
 307         if (force)
 308                 set_bit(LIOD_STOP_FORCE, &pc->pc_flags);
 309         cfs_waitq_signal(&pc->pc_set->set_waitq);
 310 #ifdef __KERNEL__
 311         wait_for_completion(&pc->pc_finishing);
 312 #else
 313         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 314         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 315 #endif
 316         ptlrpc_set_destroy(pc->pc_set);
 317 }
 318
 319 int ptlrpcd_addref(void)
 320 {
 321         int rc = 0;
 322         ENTRY;
 323
 324         mutex_down(&ptlrpcd_sem);
 325         if (++ptlrpcd_users != 1)
 326                 GOTO(out, rc);
 327
 328         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 329         if (rc) {
 330                 --ptlrpcd_users;
 331                 GOTO(out, rc);
 332         }
 333
 334         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 335         if (rc) {
 336                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 337                 --ptlrpcd_users;
 338                 GOTO(out, rc);
 339         }
 340 out:
 341         mutex_up(&ptlrpcd_sem);
 342         RETURN(rc);
 343 }
 344
 345 void ptlrpcd_decref(void)
 346 {
 347         mutex_down(&ptlrpcd_sem);
 348         if (--ptlrpcd_users == 0) {
 349                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 350                 ptlrpcd_stop(&ptlrpcd_recovery_pc, 0);
 351         }
 352         mutex_up(&ptlrpcd_sem);
 353 }