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 <libcfs/kp30.h>
  49 #include <lustre_net.h>
  50 # include <lustre_lib.h>
  51
  52 #include <lustre_ha.h>
  53 #include <obd_class.h>   /* for obd_zombie */
  54 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  55 #include <lprocfs_status.h>
  56
  57 static struct ptlrpcd_ctl ptlrpcd_pc;
  58 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  59
  60 struct semaphore ptlrpcd_sem;
  61 static int ptlrpcd_users = 0;
  62
  63 void ptlrpcd_wake(struct ptlrpc_request *req)
  64 {
  65         struct ptlrpc_request_set *rq_set = req->rq_set;
  66
  67         LASSERT(rq_set != NULL);
  68
  69         cfs_waitq_signal(&rq_set->set_waitq);
  70 }
  71
  72 /*
  73  * Requests that are added to the ptlrpcd queue are sent via
  74  * ptlrpcd_check->ptlrpc_check_set().
  75  */
  76 void ptlrpcd_add_req(struct ptlrpc_request *req)
  77 {
  78         struct ptlrpcd_ctl *pc;
  79         int rc;
  80
  81         if (req->rq_send_state == LUSTRE_IMP_FULL)
  82                 pc = &ptlrpcd_pc;
  83         else
  84                 pc = &ptlrpcd_recovery_pc;
  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         spin_lock(&pc->pc_set->set_new_req_lock);
 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                 /*
 120                  * Need to calculate its timeout.
 121                  */
 122                 rc = 1;
 123         }
 124         spin_unlock(&pc->pc_set->set_new_req_lock);
 125
 126         if (pc->pc_set->set_remaining) {
 127                 rc = rc | ptlrpc_check_set(pc->pc_set);
 128
 129                 /*
 130                  * XXX: our set never completes, so we prune the completed
 131                  * reqs after each iteration. boy could this be smarter.
 132                  */
 133                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 134                         req = list_entry(pos, struct ptlrpc_request,
 135                                          rq_set_chain);
 136                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 137                                 continue;
 138
 139                         list_del_init(&req->rq_set_chain);
 140                         req->rq_set = NULL;
 141                         ptlrpc_req_finished (req);
 142                 }
 143         }
 144
 145         if (rc == 0) {
 146                 /*
 147                  * If new requests have been added, make sure to wake up.
 148                  */
 149                 spin_lock(&pc->pc_set->set_new_req_lock);
 150                 rc = !list_empty(&pc->pc_set->set_new_requests);
 151                 spin_unlock(&pc->pc_set->set_new_req_lock);
 152         }
 153
 154         RETURN(rc);
 155 }
 156
 157 #ifdef __KERNEL__
 158 /*
 159  * ptlrpc's code paths like to execute in process context, so we have this
 160  * thread which spins on a set which contains the io rpcs. llite specifies
 161  * ptlrpcd's set when it pushes pages down into the oscs.
 162  */
 163 static int ptlrpcd(void *arg)
 164 {
 165         struct ptlrpcd_ctl *pc = arg;
 166         int rc, exit = 0;
 167         ENTRY;
 168
 169         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 170                 complete(&pc->pc_starting);
 171                 goto out;
 172         }
 173
 174         complete(&pc->pc_starting);
 175
 176         /*
 177          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 178          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 179          * there are requests in the set. New requests come in on the set's
 180          * new_req_list and ptlrpcd_check() moves them into the set.
 181          */
 182         do {
 183                 struct l_wait_info lwi;
 184                 int timeout;
 185
 186                 timeout = ptlrpc_set_next_timeout(pc->pc_set);
 187                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 188                                   ptlrpc_expired_set, pc->pc_set);
 189
 190                 l_wait_event(pc->pc_set->set_waitq, ptlrpcd_check(pc), &lwi);
 191
 192                 /*
 193                  * Abort inflight rpcs for forced stop case.
 194                  */
 195                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 196                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 197                                 ptlrpc_abort_set(pc->pc_set);
 198                         exit++;
 199                 }
 200
 201                 /*
 202                  * Let's make one more loop to make sure that ptlrpcd_check()
 203                  * copied all raced new rpcs into the set so we can kill them.
 204                  */
 205         } while (exit < 2);
 206
 207         /*
 208          * Wait for inflight requests to drain.
 209          */
 210         if (!list_empty(&pc->pc_set->set_requests))
 211                 ptlrpc_set_wait(pc->pc_set);
 212
 213         complete(&pc->pc_finishing);
 214 out:
 215         clear_bit(LIOD_START, &pc->pc_flags);
 216         clear_bit(LIOD_STOP, &pc->pc_flags);
 217         clear_bit(LIOD_FORCE, &pc->pc_flags);
 218         return 0;
 219 }
 220
 221 #else
 222
 223 int ptlrpcd_check_async_rpcs(void *arg)
 224 {
 225         struct ptlrpcd_ctl *pc = arg;
 226         int                  rc = 0;
 227
 228         /*
 229          * Single threaded!!
 230          */
 231         pc->pc_recurred++;
 232
 233         if (pc->pc_recurred == 1) {
 234                 rc = ptlrpcd_check(pc);
 235                 if (!rc)
 236                         ptlrpc_expired_set(pc->pc_set);
 237                 /*
 238                  * XXX: send replay requests.
 239                  */
 240                 if (pc == &ptlrpcd_recovery_pc)
 241                         rc = ptlrpcd_check(pc);
 242         }
 243
 244         pc->pc_recurred--;
 245         return rc;
 246 }
 247
 248 int ptlrpcd_idle(void *arg)
 249 {
 250         struct ptlrpcd_ctl *pc = arg;
 251
 252         return (list_empty(&pc->pc_set->set_new_requests) &&
 253                 pc->pc_set->set_remaining == 0);
 254 }
 255
 256 #endif
 257
 258 int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 259 {
 260         int rc = 0;
 261         ENTRY;
 262
 263         /*
 264          * Do not allow start second thread for one pc.
 265          */
 266         if (test_bit(LIOD_START, &pc->pc_flags)) {
 267                 CERROR("Starting second thread (%s) for same pc %p\n",
 268                        name, pc);
 269                 RETURN(-EALREADY);
 270         }
 271
 272         set_bit(LIOD_START, &pc->pc_flags);
 273         init_completion(&pc->pc_starting);
 274         init_completion(&pc->pc_finishing);
 275         spin_lock_init(&pc->pc_lock);
 276         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 277
 278         pc->pc_set = ptlrpc_prep_set();
 279         if (pc->pc_set == NULL)
 280                 GOTO(out, rc = -ENOMEM);
 281
 282 #ifdef __KERNEL__
 283         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 284         if (rc < 0)  {
 285                 ptlrpc_set_destroy(pc->pc_set);
 286                 GOTO(out, rc);
 287         }
 288         rc = 0;
 289         wait_for_completion(&pc->pc_starting);
 290 #else
 291         pc->pc_wait_callback =
 292                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 293                                                  &ptlrpcd_check_async_rpcs, pc);
 294         pc->pc_idle_callback =
 295                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 296                                                  &ptlrpcd_idle, pc);
 297 #endif
 298 out:
 299         if (rc)
 300                 clear_bit(LIOD_START, &pc->pc_flags);
 301         RETURN(rc);
 302 }
 303
 304 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 305 {
 306         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 307                 CERROR("Thread for pc %p was not started\n", pc);
 308                 return;
 309         }
 310
 311         set_bit(LIOD_STOP, &pc->pc_flags);
 312         if (force)
 313                 set_bit(LIOD_FORCE, &pc->pc_flags);
 314         cfs_waitq_signal(&pc->pc_set->set_waitq);
 315 #ifdef __KERNEL__
 316         wait_for_completion(&pc->pc_finishing);
 317 #else
 318         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 319         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 320 #endif
 321         ptlrpc_set_destroy(pc->pc_set);
 322 }
 323
 324 int ptlrpcd_addref(void)
 325 {
 326         int rc = 0;
 327         ENTRY;
 328
 329         mutex_down(&ptlrpcd_sem);
 330         if (++ptlrpcd_users != 1)
 331                 GOTO(out, rc);
 332
 333         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 334         if (rc) {
 335                 --ptlrpcd_users;
 336                 GOTO(out, rc);
 337         }
 338
 339         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 340         if (rc) {
 341                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 342                 --ptlrpcd_users;
 343                 GOTO(out, rc);
 344         }
 345 out:
 346         mutex_up(&ptlrpcd_sem);
 347         RETURN(rc);
 348 }
 349
 350 void ptlrpcd_decref(void)
 351 {
 352         mutex_down(&ptlrpcd_sem);
 353         if (--ptlrpcd_users == 0) {
 354                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 355                 ptlrpcd_stop(&ptlrpcd_recovery_pc, 0);
 356         }
 357         mutex_up(&ptlrpcd_sem);
 358 }