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