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