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 (c) 2003, 2010, Oracle and/or its affiliates. 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  * Move all request from an existing request set to the ptlrpcd queue.
  74  * All requests from the set must be in phase RQ_PHASE_NEW.
  75  */
  76 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
  77 {
  78         struct list_head *tmp, *pos;
  79
  80         list_for_each_safe(pos, tmp, &set->set_requests) {
  81                 struct ptlrpc_request *req =
  82                         list_entry(pos, struct ptlrpc_request, rq_set_chain);
  83
  84                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
  85                 list_del_init(&req->rq_set_chain);
  86                 req->rq_set = NULL;
  87                 ptlrpcd_add_req(req);
  88                 atomic_dec(&set->set_remaining);
  89         }
  90         LASSERT(atomic_read(&set->set_remaining) == 0);
  91 }
  92 EXPORT_SYMBOL(ptlrpcd_add_rqset);
  93
  94 /*
  95  * Requests that are added to the ptlrpcd queue are sent via
  96  * ptlrpcd_check->ptlrpc_check_set().
  97  */
  98 int ptlrpcd_add_req(struct ptlrpc_request *req)
  99 {
 100         struct ptlrpcd_ctl *pc;
 101         int rc;
 102
 103         spin_lock(&req->rq_lock);
 104         if (req->rq_invalid_rqset) {
 105                 cfs_duration_t timeout;
 106                 struct l_wait_info lwi;
 107
 108                 req->rq_invalid_rqset = 0;
 109                 spin_unlock(&req->rq_lock);
 110
 111                 timeout = cfs_time_seconds(5);
 112                 lwi = LWI_TIMEOUT(timeout, back_to_sleep, NULL);
 113                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 114         } else if (req->rq_set) {
 115                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 116                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 117
 118                 /* ptlrpc_check_set will decrease the count */
 119                 atomic_inc(&req->rq_set->set_remaining);
 120                 spin_unlock(&req->rq_lock);
 121
 122                 cfs_waitq_signal(&req->rq_set->set_waitq);
 123                 return 0;
 124         } else {
 125                 spin_unlock(&req->rq_lock);
 126         }
 127
 128         if (req->rq_send_state == LUSTRE_IMP_FULL)
 129                 pc = &ptlrpcd_pc;
 130         else
 131                 pc = &ptlrpcd_recovery_pc;
 132         rc = ptlrpc_set_add_new_req(pc, req);
 133         if (rc) {
 134                 /*
 135                  * Thread is probably in stop now so we need to
 136                  * kill this rpc as it was not added. Let's call
 137                  * interpret for it to let know we're killing it
 138                  * so that higher levels might free assosiated
 139                  * resources.
 140                 */
 141
 142                 ptlrpc_req_interpret(req, -EBADR);
 143                 req->rq_set = NULL;
 144                 ptlrpc_req_finished(req);
 145         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING) {
 146                /*
 147                 * The request is for recovery, should be sent ASAP.
 148                 */
 149                cfs_waitq_signal(&pc->pc_set->set_waitq);
 150         }
 151
 152         return rc;
 153 }
 154
 155 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
 156 {
 157         struct list_head *tmp, *pos;
 158         struct ptlrpc_request *req;
 159         int rc = 0;
 160         ENTRY;
 161
 162         spin_lock(&pc->pc_set->set_new_req_lock);
 163         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 164                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 165                 list_del_init(&req->rq_set_chain);
 166                 ptlrpc_set_add_req(pc->pc_set, req);
 167                 /*
 168                  * Need to calculate its timeout.
 169                  */
 170                 rc = 1;
 171         }
 172         spin_unlock(&pc->pc_set->set_new_req_lock);
 173
 174         if (atomic_read(&pc->pc_set->set_remaining)) {
 175                 rc = rc | ptlrpc_check_set(pc->pc_set);
 176
 177                 /*
 178                  * XXX: our set never completes, so we prune the completed
 179                  * reqs after each iteration. boy could this be smarter.
 180                  */
 181                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 182                         req = list_entry(pos, struct ptlrpc_request,
 183                                          rq_set_chain);
 184                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 185                                 continue;
 186
 187                         list_del_init(&req->rq_set_chain);
 188                         req->rq_set = NULL;
 189                         ptlrpc_req_finished (req);
 190                 }
 191         }
 192
 193         if (rc == 0) {
 194                 /*
 195                  * If new requests have been added, make sure to wake up.
 196                  */
 197                 spin_lock(&pc->pc_set->set_new_req_lock);
 198                 rc = !list_empty(&pc->pc_set->set_new_requests);
 199                 spin_unlock(&pc->pc_set->set_new_req_lock);
 200         }
 201
 202         RETURN(rc);
 203 }
 204
 205 #ifdef __KERNEL__
 206 /*
 207  * ptlrpc's code paths like to execute in process context, so we have this
 208  * thread which spins on a set which contains the io rpcs. llite specifies
 209  * ptlrpcd's set when it pushes pages down into the oscs.
 210  */
 211 static int ptlrpcd(void *arg)
 212 {
 213         struct ptlrpcd_ctl *pc = arg;
 214         int rc, exit = 0;
 215         ENTRY;
 216
 217         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 218                 complete(&pc->pc_starting);
 219                 goto out;
 220         }
 221
 222         complete(&pc->pc_starting);
 223
 224         /*
 225          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 226          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 227          * there are requests in the set. New requests come in on the set's
 228          * new_req_list and ptlrpcd_check() moves them into the set.
 229          */
 230         do {
 231                 struct l_wait_info lwi;
 232                 int timeout;
 233
 234                 timeout = ptlrpc_set_next_timeout(pc->pc_set);
 235                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 236                                   ptlrpc_expired_set, pc->pc_set);
 237
 238                 l_wait_event(pc->pc_set->set_waitq, ptlrpcd_check(pc), &lwi);
 239
 240                 /*
 241                  * Abort inflight rpcs for forced stop case.
 242                  */
 243                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 244                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 245                                 ptlrpc_abort_set(pc->pc_set);
 246                         exit++;
 247                 }
 248
 249                 /*
 250                  * Let's make one more loop to make sure that ptlrpcd_check()
 251                  * copied all raced new rpcs into the set so we can kill them.
 252                  */
 253         } while (exit < 2);
 254
 255         /*
 256          * Wait for inflight requests to drain.
 257          */
 258         if (!list_empty(&pc->pc_set->set_requests))
 259                 ptlrpc_set_wait(pc->pc_set);
 260
 261         complete(&pc->pc_finishing);
 262 out:
 263         clear_bit(LIOD_START, &pc->pc_flags);
 264         clear_bit(LIOD_STOP, &pc->pc_flags);
 265         clear_bit(LIOD_FORCE, &pc->pc_flags);
 266         return 0;
 267 }
 268
 269 #else
 270
 271 int ptlrpcd_check_async_rpcs(void *arg)
 272 {
 273         struct ptlrpcd_ctl *pc = arg;
 274         int                  rc = 0;
 275
 276         /*
 277          * Single threaded!!
 278          */
 279         pc->pc_recurred++;
 280
 281         if (pc->pc_recurred == 1) {
 282                 rc = ptlrpcd_check(pc);
 283                 if (!rc)
 284                         ptlrpc_expired_set(pc->pc_set);
 285                 /*
 286                  * XXX: send replay requests.
 287                  */
 288                 if (pc == &ptlrpcd_recovery_pc)
 289                         rc = ptlrpcd_check(pc);
 290         }
 291
 292         pc->pc_recurred--;
 293         return rc;
 294 }
 295
 296 int ptlrpcd_idle(void *arg)
 297 {
 298         struct ptlrpcd_ctl *pc = arg;
 299
 300         return (list_empty(&pc->pc_set->set_new_requests) &&
 301                 atomic_read(&pc->pc_set->set_remaining) == 0);
 302 }
 303
 304 #endif
 305
 306 int ptlrpcd_start(char *name, struct ptlrpcd_ctl *pc)
 307 {
 308         int rc = 0;
 309         ENTRY;
 310
 311         /*
 312          * Do not allow start second thread for one pc.
 313          */
 314         if (test_bit(LIOD_START, &pc->pc_flags)) {
 315                 CERROR("Starting second thread (%s) for same pc %p\n",
 316                        name, pc);
 317                 RETURN(-EALREADY);
 318         }
 319
 320         set_bit(LIOD_START, &pc->pc_flags);
 321         init_completion(&pc->pc_starting);
 322         init_completion(&pc->pc_finishing);
 323         spin_lock_init(&pc->pc_lock);
 324         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 325
 326         pc->pc_set = ptlrpc_prep_set();
 327         if (pc->pc_set == NULL)
 328                 GOTO(out, rc = -ENOMEM);
 329
 330 #ifdef __KERNEL__
 331         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 332         if (rc < 0)  {
 333                 ptlrpc_set_destroy(pc->pc_set);
 334                 GOTO(out, rc);
 335         }
 336         rc = 0;
 337         wait_for_completion(&pc->pc_starting);
 338 #else
 339         pc->pc_wait_callback =
 340                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 341                                                  &ptlrpcd_check_async_rpcs, pc);
 342         pc->pc_idle_callback =
 343                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 344                                                  &ptlrpcd_idle, pc);
 345 #endif
 346 out:
 347         if (rc)
 348                 clear_bit(LIOD_START, &pc->pc_flags);
 349         RETURN(rc);
 350 }
 351
 352 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 353 {
 354         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 355                 CERROR("Thread for pc %p was not started\n", pc);
 356                 return;
 357         }
 358
 359         set_bit(LIOD_STOP, &pc->pc_flags);
 360         if (force)
 361                 set_bit(LIOD_FORCE, &pc->pc_flags);
 362         cfs_waitq_signal(&pc->pc_set->set_waitq);
 363 #ifdef __KERNEL__
 364         wait_for_completion(&pc->pc_finishing);
 365 #else
 366         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 367         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 368 #endif
 369         ptlrpc_set_destroy(pc->pc_set);
 370 }
 371
 372 int ptlrpcd_addref(void)
 373 {
 374         int rc = 0;
 375         ENTRY;
 376
 377         mutex_down(&ptlrpcd_sem);
 378         if (++ptlrpcd_users != 1)
 379                 GOTO(out, rc);
 380
 381         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 382         if (rc) {
 383                 --ptlrpcd_users;
 384                 GOTO(out, rc);
 385         }
 386
 387         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 388         if (rc) {
 389                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 390                 --ptlrpcd_users;
 391                 GOTO(out, rc);
 392         }
 393 out:
 394         mutex_up(&ptlrpcd_sem);
 395         RETURN(rc);
 396 }
 397
 398 void ptlrpcd_decref(void)
 399 {
 400         mutex_down(&ptlrpcd_sem);
 401         if (--ptlrpcd_users == 0) {
 402                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 403                 ptlrpcd_stop(&ptlrpcd_recovery_pc, 0);
 404         }
 405         mutex_up(&ptlrpcd_sem);
 406 }