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 <cl_object.h> /* cl_env_{get,put}() */
  55 #include <lprocfs_status.h>
  56
  57 enum pscope_thread {
  58         PT_NORMAL,
  59         PT_RECOVERY,
  60         PT_NR
  61 };
  62
  63 struct ptlrpcd_scope_ctl {
  64         struct ptlrpcd_thread {
  65                 const char        *pt_name;
  66                 struct ptlrpcd_ctl pt_ctl;
  67         } pscope_thread[PT_NR];
  68 };
  69
  70 static struct ptlrpcd_scope_ctl ptlrpcd_scopes[PSCOPE_NR] = {
  71         [PSCOPE_BRW] = {
  72                 .pscope_thread = {
  73                         [PT_NORMAL] = {
  74                                 .pt_name = "ptlrpcd-brw"
  75                         },
  76                         [PT_RECOVERY] = {
  77                                 .pt_name = "ptlrpcd-brw-rcv"
  78                         }
  79                 }
  80         },
  81         [PSCOPE_OTHER] = {
  82                 .pscope_thread = {
  83                         [PT_NORMAL] = {
  84                                 .pt_name = "ptlrpcd"
  85                         },
  86                         [PT_RECOVERY] = {
  87                                 .pt_name = "ptlrpcd-rcv"
  88                         }
  89                 }
  90         }
  91 };
  92
  93 struct semaphore ptlrpcd_sem;
  94 static int ptlrpcd_users = 0;
  95
  96 void ptlrpcd_wake(struct ptlrpc_request *req)
  97 {
  98         struct ptlrpc_request_set *rq_set = req->rq_set;
  99
 100         LASSERT(rq_set != NULL);
 101
 102         cfs_waitq_signal(&rq_set->set_waitq);
 103 }
 104
 105 /*
 106  * Requests that are added to the ptlrpcd queue are sent via
 107  * ptlrpcd_check->ptlrpc_check_set().
 108  */
 109 void ptlrpcd_add_req(struct ptlrpc_request *req, enum ptlrpcd_scope scope)
 110 {
 111         struct ptlrpcd_ctl *pc;
 112         enum pscope_thread  pt;
 113         int rc;
 114
 115         LASSERT(scope < PSCOPE_NR);
 116         pt = req->rq_send_state == LUSTRE_IMP_FULL ? PT_NORMAL : PT_RECOVERY;
 117         pc = &ptlrpcd_scopes[scope].pscope_thread[pt].pt_ctl;
 118         rc = ptlrpc_set_add_new_req(pc, req);
 119         /*
 120          * XXX disable this for CLIO: environment is needed for interpreter.
 121          */
 122         if (rc && 0) {
 123                 ptlrpc_interpterer_t interpreter;
 124
 125                 interpreter = req->rq_interpret_reply;
 126
 127                 /*
 128                  * Thread is probably in stop now so we need to
 129                  * kill this rpc as it was not added. Let's call
 130                  * interpret for it to let know we're killing it
 131                  * so that higher levels might free assosiated
 132                  * resources.
 133                  */
 134                 req->rq_status = -EBADR;
 135                 interpreter(NULL, req, &req->rq_async_args,
 136                             req->rq_status);
 137                 req->rq_set = NULL;
 138                 ptlrpc_req_finished(req);
 139         }
 140 }
 141
 142 static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc)
 143 {
 144         struct list_head *tmp, *pos;
 145         struct ptlrpc_request *req;
 146         int rc = 0;
 147         ENTRY;
 148
 149         spin_lock(&pc->pc_set->set_new_req_lock);
 150         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 151                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 152                 list_del_init(&req->rq_set_chain);
 153                 ptlrpc_set_add_req(pc->pc_set, req);
 154                 /*
 155                  * Need to calculate its timeout.
 156                  */
 157                 rc = 1;
 158         }
 159         spin_unlock(&pc->pc_set->set_new_req_lock);
 160
 161         if (pc->pc_set->set_remaining) {
 162                 rc = rc | ptlrpc_check_set(env, pc->pc_set);
 163
 164                 /*
 165                  * XXX: our set never completes, so we prune the completed
 166                  * reqs after each iteration. boy could this be smarter.
 167                  */
 168                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 169                         req = list_entry(pos, struct ptlrpc_request,
 170                                          rq_set_chain);
 171                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 172                                 continue;
 173
 174                         list_del_init(&req->rq_set_chain);
 175                         req->rq_set = NULL;
 176                         ptlrpc_req_finished (req);
 177                 }
 178         }
 179
 180         if (rc == 0) {
 181                 /*
 182                  * If new requests have been added, make sure to wake up.
 183                  */
 184                 spin_lock(&pc->pc_set->set_new_req_lock);
 185                 rc = !list_empty(&pc->pc_set->set_new_requests);
 186                 spin_unlock(&pc->pc_set->set_new_req_lock);
 187         }
 188
 189         RETURN(rc);
 190 }
 191
 192 #ifdef __KERNEL__
 193 /*
 194  * ptlrpc's code paths like to execute in process context, so we have this
 195  * thread which spins on a set which contains the io rpcs. llite specifies
 196  * ptlrpcd's set when it pushes pages down into the oscs.
 197  */
 198 static int ptlrpcd(void *arg)
 199 {
 200         struct ptlrpcd_ctl *pc = arg;
 201         struct lu_env env = { .le_ses = NULL };
 202         int rc, exit = 0;
 203         ENTRY;
 204
 205         rc = cfs_daemonize_ctxt(pc->pc_name);
 206         if (rc == 0) {
 207                 /*
 208                  * XXX So far only "client" ptlrpcd uses an environment. In
 209                  * the future, ptlrpcd thread (or a thread-set) has to given
 210                  * an argument, describing its "scope".
 211                  */
 212                 rc = lu_context_init(&env.le_ctx,
 213                                      LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 214         }
 215
 216         complete(&pc->pc_starting);
 217
 218         if (rc != 0)
 219                 RETURN(rc);
 220         env.le_ctx.lc_cookie = 0x7;
 221
 222         /*
 223          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 224          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 225          * there are requests in the set. New requests come in on the set's
 226          * new_req_list and ptlrpcd_check() moves them into the set.
 227          */
 228         do {
 229                 struct l_wait_info lwi;
 230                 int timeout;
 231
 232                 rc = lu_env_refill(&env);
 233                 if (rc != 0) {
 234                         /*
 235                          * XXX This is very awkward situation, because
 236                          * execution can neither continue (request
 237                          * interpreters assume that env is set up), nor repeat
 238                          * the loop (as this potentially results in a tight
 239                          * loop of -ENOMEM's).
 240                          *
 241                          * Fortunately, refill only ever does something when
 242                          * new modules are loaded, i.e., early during boot up.
 243                          */
 244                         CERROR("Failure to refill session: %d\n", rc);
 245                         continue;
 246                 }
 247
 248                 timeout = ptlrpc_set_next_timeout(pc->pc_set);
 249                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 250                                   ptlrpc_expired_set, pc->pc_set);
 251
 252                 lu_context_enter(&env.le_ctx);
 253                 l_wait_event(pc->pc_set->set_waitq,
 254                              ptlrpcd_check(&env, pc), &lwi);
 255                 lu_context_exit(&env.le_ctx);
 256
 257                 /*
 258                  * Abort inflight rpcs for forced stop case.
 259                  */
 260                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
 261                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
 262                                 ptlrpc_abort_set(pc->pc_set);
 263                         exit++;
 264                 }
 265
 266                 /*
 267                  * Let's make one more loop to make sure that ptlrpcd_check()
 268                  * copied all raced new rpcs into the set so we can kill them.
 269                  */
 270         } while (exit < 2);
 271
 272         /*
 273          * Wait for inflight requests to drain.
 274          */
 275         if (!list_empty(&pc->pc_set->set_requests))
 276                 ptlrpc_set_wait(pc->pc_set);
 277         lu_context_fini(&env.le_ctx);
 278         complete(&pc->pc_finishing);
 279
 280         clear_bit(LIOD_START, &pc->pc_flags);
 281         clear_bit(LIOD_STOP, &pc->pc_flags);
 282         clear_bit(LIOD_FORCE, &pc->pc_flags);
 283         return 0;
 284 }
 285
 286 #else /* !__KERNEL__ */
 287
 288 int ptlrpcd_check_async_rpcs(void *arg)
 289 {
 290         struct ptlrpcd_ctl *pc = arg;
 291         int                 rc = 0;
 292
 293         /*
 294          * Single threaded!!
 295          */
 296         pc->pc_recurred++;
 297
 298         if (pc->pc_recurred == 1) {
 299                 lu_context_enter(&pc->pc_env.le_ctx);
 300                 rc = ptlrpcd_check(&pc->pc_env, pc);
 301                 lu_context_exit(&pc->pc_env.le_ctx);
 302                 if (!rc)
 303                         ptlrpc_expired_set(pc->pc_set);
 304                 /*
 305                  * XXX: send replay requests.
 306                  */
 307                 if (test_bit(LIOD_RECOVERY, &pc->pc_flags))
 308                         rc = ptlrpcd_check(&pc->pc_env, pc);
 309         }
 310
 311         pc->pc_recurred--;
 312         return rc;
 313 }
 314
 315 int ptlrpcd_idle(void *arg)
 316 {
 317         struct ptlrpcd_ctl *pc = arg;
 318
 319         return (list_empty(&pc->pc_set->set_new_requests) &&
 320                 pc->pc_set->set_remaining == 0);
 321 }
 322
 323 #endif
 324
 325 int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc)
 326 {
 327         int rc;
 328         ENTRY;
 329
 330         /*
 331          * Do not allow start second thread for one pc.
 332          */
 333         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 334                 CERROR("Starting second thread (%s) for same pc %p\n",
 335                        name, pc);
 336                 RETURN(-EALREADY);
 337         }
 338
 339         init_completion(&pc->pc_starting);
 340         init_completion(&pc->pc_finishing);
 341         spin_lock_init(&pc->pc_lock);
 342         snprintf (pc->pc_name, sizeof (pc->pc_name), name);
 343         pc->pc_set = ptlrpc_prep_set();
 344         if (pc->pc_set == NULL)
 345                 GOTO(out, rc = -ENOMEM);
 346         /*
 347          * So far only "client" ptlrpcd uses an environment. In the future,
 348          * ptlrpcd thread (or a thread-set) has to be given an argument,
 349          * describing its "scope".
 350          */
 351         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 352         if (rc != 0) {
 353                 ptlrpc_set_destroy(pc->pc_set);
 354                 GOTO(out, rc);
 355         }
 356
 357 #ifdef __KERNEL__
 358         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 359         if (rc < 0)  {
 360                 lu_context_fini(&pc->pc_env.le_ctx);
 361                 ptlrpc_set_destroy(pc->pc_set);
 362                 GOTO(out, rc);
 363         }
 364         rc = 0;
 365         wait_for_completion(&pc->pc_starting);
 366 #else
 367         pc->pc_wait_callback =
 368                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 369                                                  &ptlrpcd_check_async_rpcs, pc);
 370         pc->pc_idle_callback =
 371                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 372                                                  &ptlrpcd_idle, pc);
 373 #endif
 374 out:
 375         if (rc)
 376                 clear_bit(LIOD_START, &pc->pc_flags);
 377         RETURN(rc);
 378 }
 379
 380 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 381 {
 382         if (!test_bit(LIOD_START, &pc->pc_flags)) {
 383                 CERROR("Thread for pc %p was not started\n", pc);
 384                 return;
 385         }
 386
 387         set_bit(LIOD_STOP, &pc->pc_flags);
 388         if (force)
 389                 set_bit(LIOD_FORCE, &pc->pc_flags);
 390         cfs_waitq_signal(&pc->pc_set->set_waitq);
 391 #ifdef __KERNEL__
 392         wait_for_completion(&pc->pc_finishing);
 393 #else
 394         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 395         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 396 #endif
 397         lu_context_fini(&pc->pc_env.le_ctx);
 398         ptlrpc_set_destroy(pc->pc_set);
 399 }
 400
 401 void ptlrpcd_fini(void)
 402 {
 403         int i;
 404         int j;
 405
 406         ENTRY;
 407
 408         for (i = 0; i < PSCOPE_NR; ++i) {
 409                 for (j = 0; j < PT_NR; ++j) {
 410                         struct ptlrpcd_ctl *pc;
 411
 412                         pc = &ptlrpcd_scopes[i].pscope_thread[j].pt_ctl;
 413
 414                         if (test_bit(LIOD_START, &pc->pc_flags))
 415                                 ptlrpcd_stop(pc, 0);
 416                 }
 417         }
 418         EXIT;
 419 }
 420
 421 int ptlrpcd_addref(void)
 422 {
 423         int rc = 0;
 424         int i;
 425         int j;
 426         ENTRY;
 427
 428         mutex_down(&ptlrpcd_sem);
 429         if (++ptlrpcd_users == 1) {
 430                 for (i = 0; rc == 0 && i < PSCOPE_NR; ++i) {
 431                         for (j = 0; rc == 0 && j < PT_NR; ++j) {
 432                                 struct ptlrpcd_thread *pt;
 433                                 struct ptlrpcd_ctl    *pc;
 434
 435                                 pt = &ptlrpcd_scopes[i].pscope_thread[j];
 436                                 pc = &pt->pt_ctl;
 437                                 if (j == PT_RECOVERY)
 438                                         set_bit(LIOD_RECOVERY, &pc->pc_flags);
 439                                 rc = ptlrpcd_start(pt->pt_name, pc);
 440                         }
 441                 }
 442                 if (rc != 0) {
 443                         --ptlrpcd_users;
 444                         ptlrpcd_fini();
 445                 }
 446         }
 447         mutex_up(&ptlrpcd_sem);
 448         RETURN(rc);
 449 }
 450
 451 void ptlrpcd_decref(void)
 452 {
 453         mutex_down(&ptlrpcd_sem);
 454         if (--ptlrpcd_users == 0)
 455                 ptlrpcd_fini();
 456         mutex_up(&ptlrpcd_sem);
 457 }