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                 ptlrpc_interpterer_t interpreter;
  88
  89                 interpreter = req->rq_interpret_reply;
  90
  91                 /*
  92                  * Thread is probably in stop now so we need to
  93                  * kill this rpc as it was not added. Let's call
  94                  * interpret for it to let know we're killing it
  95                  * so that higher levels might free assosiated
  96                  * resources.
  97                  */
  98                 req->rq_status = -EBADR;
  99                 interpreter(NULL, req, &req->rq_async_args,
 100                             req->rq_status);
 101                 req->rq_set = NULL;
 102                 ptlrpc_req_finished(req);
 103         }
 104 }
 105
 106 static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc)
 107 {
 108         struct list_head *tmp, *pos;
 109         struct ptlrpc_request *req;
 110         int rc = 0;
 111         ENTRY;
 112
 113         if (test_bit(LIOD_STOP, &pc->pc_flags))
 114                 RETURN(1);
 115
 116         spin_lock(&pc->pc_set->set_new_req_lock);
 117         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 118                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 119                 list_del_init(&req->rq_set_chain);
 120                 ptlrpc_set_add_req(pc->pc_set, req);
 121                 /*
 122                  * Need to calculate its timeout.
 123                  */
 124                 rc = 1;
 125         }
 126         spin_unlock(&pc->pc_set->set_new_req_lock);
 127
 128         if (pc->pc_set->set_remaining) {
 129                 rc = rc | ptlrpc_check_set(env, pc->pc_set);
 130
 131                 /*
 132                  * XXX: our set never completes, so we prune the completed
 133                  * reqs after each iteration. boy could this be smarter.
 134                  */
 135                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 136                         req = list_entry(pos, struct ptlrpc_request,
 137                                          rq_set_chain);
 138                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 139                                 continue;
 140
 141                         list_del_init(&req->rq_set_chain);
 142                         req->rq_set = NULL;
 143                         ptlrpc_req_finished (req);
 144                 }
 145         }
 146
 147         if (rc == 0) {
 148                 /*
 149                  * If new requests have been added, make sure to wake up.
 150                  */
 151                 spin_lock(&pc->pc_set->set_new_req_lock);
 152                 rc = !list_empty(&pc->pc_set->set_new_requests);
 153                 spin_unlock(&pc->pc_set->set_new_req_lock);
 154         }
 155
 156         RETURN(rc);
 157 }
 158
 159 #ifdef __KERNEL__
 160 /*
 161  * ptlrpc's code paths like to execute in process context, so we have this
 162  * thread which spins on a set which contains the io rpcs. llite specifies
 163  * ptlrpcd's set when it pushes pages down into the oscs.
 164  */
 165 static int ptlrpcd(void *arg)
 166 {
 167         struct ptlrpcd_ctl *pc = arg;
 168         int rc;
 169         ENTRY;
 170
 171         if ((rc = cfs_daemonize_ctxt(pc->pc_name))) {
 172                 complete(&pc->pc_starting);
 173                 goto out;
 174         }
 175
 176         complete(&pc->pc_starting);
 177
 178         /*
 179          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 180          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 181          * there are requests in the set. New requests come in on the set's
 182          * new_req_list and ptlrpcd_check() moves them into the set.
 183          */
 184         while (1) {
 185                 struct l_wait_info lwi;
 186                 cfs_duration_t timeout;
 187
 188                 timeout = cfs_time_seconds(ptlrpc_set_next_timeout(pc->pc_set));
 189                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 190
 191                 lu_context_enter(&pc->pc_env.le_ctx);
 192                 l_wait_event(pc->pc_set->set_waitq,
 193                              ptlrpcd_check(&pc->pc_env, pc), &lwi);
 194                 lu_context_exit(&pc->pc_env.le_ctx);
 195
 196                 /*
 197                  * Abort inflight rpcs for forced stop case.
 198                  */
 199                 if (test_bit(LIOD_STOP_FORCE, &pc->pc_flags))
 200                         ptlrpc_abort_set(pc->pc_set);
 201
 202                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 203                         break;
 204         }
 205
 206         /*
 207          * Wait for inflight requests to drain.
 208          */
 209         if (!list_empty(&pc->pc_set->set_requests))
 210                 ptlrpc_set_wait(pc->pc_set);
 211
 212         complete(&pc->pc_finishing);
 213 out:
 214         clear_bit(LIOD_START, &pc->pc_flags);
 215         clear_bit(LIOD_STOP, &pc->pc_flags);
 216         return 0;
 217 }
 218
 219 #else /* !__KERNEL__ */
 220
 221 int ptlrpcd_check_async_rpcs(void *arg)
 222 {
 223         struct ptlrpcd_ctl *pc = arg;
 224         int                  rc = 0;
 225
 226         /*
 227          * Single threaded!!
 228          */
 229         pc->pc_recurred++;
 230
 231         if (pc->pc_recurred == 1) {
 232                 lu_context_enter(&pc->pc_env.le_ctx);
 233                 rc = ptlrpcd_check(&pc->pc_env, pc);
 234                 lu_context_exit(&pc->pc_env.le_ctx);
 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->pc_env, 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;
 261         ENTRY;
 262
 263         /*
 264          * Do not allow start second thread for one pc.
 265          */
 266         if (test_and_set_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         init_completion(&pc->pc_starting);
 273         init_completion(&pc->pc_finishing);
 274         spin_lock_init(&pc->pc_lock);
 275         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 276         pc->pc_set = ptlrpc_prep_set();
 277         if (pc->pc_set == NULL)
 278                 GOTO(out, rc = -ENOMEM);
 279         /*
 280          * So far only "client" ptlrpcd uses an environment. In the future,
 281          * ptlrpcd thread (or a thread-set) has to be given an argument,
 282          * describing its "scope".
 283          */
 284         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 285         if (rc != 0) {
 286                 ptlrpc_set_destroy(pc->pc_set);
 287                 GOTO(out, rc);
 288         }
 289
 290 #ifdef __KERNEL__
 291         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 292         if (rc < 0)  {
 293                 lu_context_fini(&pc->pc_env.le_ctx);
 294                 ptlrpc_set_destroy(pc->pc_set);
 295                 GOTO(out, rc);
 296         }
 297         rc = 0;
 298         wait_for_completion(&pc->pc_starting);
 299 #else
 300         pc->pc_wait_callback =
 301                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 302                                                  &ptlrpcd_check_async_rpcs, pc);
 303         pc->pc_idle_callback =
 304                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 305                                                  &ptlrpcd_idle, pc);
 306 #endif
 307 out:
 308         if (rc)
 309                 clear_bit(LIOD_START, &pc->pc_flags);
 310         RETURN(rc);
 311 }
 312
 313 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 314 {
 315         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 316                 CERROR("Thread for pc %p was not started\n", pc);
 317                 return;
 318         }
 319
 320         set_bit(LIOD_STOP, &pc->pc_flags);
 321         if (force)
 322                 set_bit(LIOD_STOP_FORCE, &pc->pc_flags);
 323         cfs_waitq_signal(&pc->pc_set->set_waitq);
 324 #ifdef __KERNEL__
 325         wait_for_completion(&pc->pc_finishing);
 326 #else
 327         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 328         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 329 #endif
 330         lu_context_fini(&pc->pc_env.le_ctx);
 331         ptlrpc_set_destroy(pc->pc_set);
 332 }
 333
 334 int ptlrpcd_addref(void)
 335 {
 336         int rc = 0;
 337         ENTRY;
 338
 339         mutex_down(&ptlrpcd_sem);
 340         if (++ptlrpcd_users != 1)
 341                 GOTO(out, rc);
 342
 343         rc = ptlrpcd_start("ptlrpcd", &ptlrpcd_pc);
 344         if (rc) {
 345                 --ptlrpcd_users;
 346                 GOTO(out, rc);
 347         }
 348
 349         rc = ptlrpcd_start("ptlrpcd-recov", &ptlrpcd_recovery_pc);
 350         if (rc) {
 351                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 352                 --ptlrpcd_users;
 353                 GOTO(out, rc);
 354         }
 355 out:
 356         mutex_up(&ptlrpcd_sem);
 357         RETURN(rc);
 358 }
 359
 360 void ptlrpcd_decref(void)
 361 {
 362         mutex_down(&ptlrpcd_sem);
 363         if (--ptlrpcd_users == 0) {
 364                 ptlrpcd_stop(&ptlrpcd_pc, 0);
 365                 ptlrpcd_stop(&ptlrpcd_recovery_pc, 0);
 366         }
 367         mutex_up(&ptlrpcd_sem);
 368 }