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 cfs_semaphore_t 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  * Move all request from an existing request set to the ptlrpcd queue.
 107  * All requests from the set must be in phase RQ_PHASE_NEW.
 108  */
 109 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 110 {
 111         cfs_list_t *tmp, *pos;
 112
 113         cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 114                 struct ptlrpc_request *req =
 115                         cfs_list_entry(pos, struct ptlrpc_request,
 116                                        rq_set_chain);
 117
 118                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 119                 cfs_list_del_init(&req->rq_set_chain);
 120                 req->rq_set = NULL;
 121                 ptlrpcd_add_req(req, PSCOPE_OTHER);
 122                 cfs_atomic_dec(&set->set_remaining);
 123         }
 124         LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
 125 }
 126 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 127
 128 /*
 129  * Requests that are added to the ptlrpcd queue are sent via
 130  * ptlrpcd_check->ptlrpc_check_set().
 131  */
 132 int ptlrpcd_add_req(struct ptlrpc_request *req, enum ptlrpcd_scope scope)
 133 {
 134         struct ptlrpcd_ctl *pc;
 135         enum pscope_thread  pt;
 136         int rc;
 137
 138         LASSERT(scope < PSCOPE_NR);
 139
 140         cfs_spin_lock(&req->rq_lock);
 141         if (req->rq_invalid_rqset) {
 142                 cfs_duration_t timeout;
 143                 struct l_wait_info lwi;
 144
 145                 req->rq_invalid_rqset = 0;
 146                 cfs_spin_unlock(&req->rq_lock);
 147
 148                 timeout = cfs_time_seconds(5);
 149                 lwi = LWI_TIMEOUT(timeout, back_to_sleep, NULL);
 150                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 151         } else if (req->rq_set) {
 152                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 153                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 154
 155                 /* ptlrpc_check_set will decrease the count */
 156                 cfs_atomic_inc(&req->rq_set->set_remaining);
 157                 cfs_spin_unlock(&req->rq_lock);
 158
 159                 cfs_waitq_signal(&req->rq_set->set_waitq);
 160         } else {
 161                 cfs_spin_unlock(&req->rq_lock);
 162         }
 163
 164         pt = req->rq_send_state == LUSTRE_IMP_FULL ? PT_NORMAL : PT_RECOVERY;
 165         pc = &ptlrpcd_scopes[scope].pscope_thread[pt].pt_ctl;
 166         rc = ptlrpc_set_add_new_req(pc, req);
 167         /*
 168          * XXX disable this for CLIO: environment is needed for interpreter.
 169          *     add debug temporary to check rc.
 170          */
 171         LASSERTF(rc == 0, "ptlrpcd_add_req failed (rc = %d)\n", rc);
 172         if (rc && 0) {
 173                 /*
 174                  * Thread is probably in stop now so we need to
 175                  * kill this rpc as it was not added. Let's call
 176                  * interpret for it to let know we're killing it
 177                  * so that higher levels might free associated
 178                  * resources.
 179                  */
 180                 ptlrpc_req_interpret(NULL, req, -EBADR);
 181                 req->rq_set = NULL;
 182                 ptlrpc_req_finished(req);
 183         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING) {
 184                 /*
 185                  * The request is for recovery, should be sent ASAP.
 186                  */
 187                 cfs_waitq_signal(&pc->pc_set->set_waitq);
 188         }
 189
 190         return rc;
 191 }
 192
 193 static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc)
 194 {
 195         cfs_list_t *tmp, *pos;
 196         struct ptlrpc_request *req;
 197         int rc = 0;
 198         ENTRY;
 199
 200         cfs_spin_lock(&pc->pc_set->set_new_req_lock);
 201         cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 202                 req = cfs_list_entry(pos, struct ptlrpc_request, rq_set_chain);
 203                 cfs_list_del_init(&req->rq_set_chain);
 204                 ptlrpc_set_add_req(pc->pc_set, req);
 205                 /*
 206                  * Need to calculate its timeout.
 207                  */
 208                 rc = 1;
 209         }
 210         cfs_spin_unlock(&pc->pc_set->set_new_req_lock);
 211
 212         if (cfs_atomic_read(&pc->pc_set->set_remaining)) {
 213                 rc = rc | ptlrpc_check_set(env, pc->pc_set);
 214
 215                 /*
 216                  * XXX: our set never completes, so we prune the completed
 217                  * reqs after each iteration. boy could this be smarter.
 218                  */
 219                 cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 220                         req = cfs_list_entry(pos, struct ptlrpc_request,
 221                                          rq_set_chain);
 222                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 223                                 continue;
 224
 225                         cfs_list_del_init(&req->rq_set_chain);
 226                         req->rq_set = NULL;
 227                         ptlrpc_req_finished (req);
 228                 }
 229         }
 230
 231         if (rc == 0) {
 232                 /*
 233                  * If new requests have been added, make sure to wake up.
 234                  */
 235                 cfs_spin_lock(&pc->pc_set->set_new_req_lock);
 236                 rc = !cfs_list_empty(&pc->pc_set->set_new_requests);
 237                 cfs_spin_unlock(&pc->pc_set->set_new_req_lock);
 238         }
 239
 240         RETURN(rc);
 241 }
 242
 243 #ifdef __KERNEL__
 244 /*
 245  * ptlrpc's code paths like to execute in process context, so we have this
 246  * thread which spins on a set which contains the io rpcs. llite specifies
 247  * ptlrpcd's set when it pushes pages down into the oscs.
 248  */
 249 static int ptlrpcd(void *arg)
 250 {
 251         struct ptlrpcd_ctl *pc = arg;
 252         struct lu_env env = { .le_ses = NULL };
 253         int rc, exit = 0;
 254         ENTRY;
 255
 256         rc = cfs_daemonize_ctxt(pc->pc_name);
 257         if (rc == 0) {
 258                 /*
 259                  * XXX So far only "client" ptlrpcd uses an environment. In
 260                  * the future, ptlrpcd thread (or a thread-set) has to given
 261                  * an argument, describing its "scope".
 262                  */
 263                 rc = lu_context_init(&env.le_ctx,
 264                                      LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 265         }
 266
 267         cfs_complete(&pc->pc_starting);
 268
 269         if (rc != 0)
 270                 RETURN(rc);
 271         env.le_ctx.lc_cookie = 0x7;
 272
 273         /*
 274          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 275          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 276          * there are requests in the set. New requests come in on the set's
 277          * new_req_list and ptlrpcd_check() moves them into the set.
 278          */
 279         do {
 280                 struct l_wait_info lwi;
 281                 int timeout;
 282
 283                 rc = lu_env_refill(&env);
 284                 if (rc != 0) {
 285                         /*
 286                          * XXX This is very awkward situation, because
 287                          * execution can neither continue (request
 288                          * interpreters assume that env is set up), nor repeat
 289                          * the loop (as this potentially results in a tight
 290                          * loop of -ENOMEM's).
 291                          *
 292                          * Fortunately, refill only ever does something when
 293                          * new modules are loaded, i.e., early during boot up.
 294                          */
 295                         CERROR("Failure to refill session: %d\n", rc);
 296                         continue;
 297                 }
 298
 299                 timeout = ptlrpc_set_next_timeout(pc->pc_set);
 300                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 301                                   ptlrpc_expired_set, pc->pc_set);
 302
 303                 lu_context_enter(&env.le_ctx);
 304                 l_wait_event(pc->pc_set->set_waitq,
 305                              ptlrpcd_check(&env, pc), &lwi);
 306                 lu_context_exit(&env.le_ctx);
 307
 308                 /*
 309                  * Abort inflight rpcs for forced stop case.
 310                  */
 311                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 312                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 313                                 ptlrpc_abort_set(pc->pc_set);
 314                         exit++;
 315                 }
 316
 317                 /*
 318                  * Let's make one more loop to make sure that ptlrpcd_check()
 319                  * copied all raced new rpcs into the set so we can kill them.
 320                  */
 321         } while (exit < 2);
 322
 323         /*
 324          * Wait for inflight requests to drain.
 325          */
 326         if (!cfs_list_empty(&pc->pc_set->set_requests))
 327                 ptlrpc_set_wait(pc->pc_set);
 328         lu_context_fini(&env.le_ctx);
 329         cfs_complete(&pc->pc_finishing);
 330
 331         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 332         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 333         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 334         return 0;
 335 }
 336
 337 #else /* !__KERNEL__ */
 338
 339 int ptlrpcd_check_async_rpcs(void *arg)
 340 {
 341         struct ptlrpcd_ctl *pc = arg;
 342         int                 rc = 0;
 343
 344         /*
 345          * Single threaded!!
 346          */
 347         pc->pc_recurred++;
 348
 349         if (pc->pc_recurred == 1) {
 350                 rc = lu_env_refill(&pc->pc_env);
 351                 if (rc == 0) {
 352                         lu_context_enter(&pc->pc_env.le_ctx);
 353                         rc = ptlrpcd_check(&pc->pc_env, pc);
 354                         lu_context_exit(&pc->pc_env.le_ctx);
 355                         if (!rc)
 356                                 ptlrpc_expired_set(pc->pc_set);
 357                         /*
 358                          * XXX: send replay requests.
 359                          */
 360                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 361                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 362                 }
 363         }
 364
 365         pc->pc_recurred--;
 366         return rc;
 367 }
 368
 369 int ptlrpcd_idle(void *arg)
 370 {
 371         struct ptlrpcd_ctl *pc = arg;
 372
 373         return (cfs_list_empty(&pc->pc_set->set_new_requests) &&
 374                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 375 }
 376
 377 #endif
 378
 379 int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc)
 380 {
 381         int rc;
 382         ENTRY;
 383
 384         /*
 385          * Do not allow start second thread for one pc.
 386          */
 387         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 388                 CERROR("Starting second thread (%s) for same pc %p\n",
 389                        name, pc);
 390                 RETURN(-EALREADY);
 391         }
 392
 393         cfs_init_completion(&pc->pc_starting);
 394         cfs_init_completion(&pc->pc_finishing);
 395         cfs_spin_lock_init(&pc->pc_lock);
 396         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 397         pc->pc_set = ptlrpc_prep_set();
 398         if (pc->pc_set == NULL)
 399                 GOTO(out, rc = -ENOMEM);
 400         /*
 401          * So far only "client" ptlrpcd uses an environment. In the future,
 402          * ptlrpcd thread (or a thread-set) has to be given an argument,
 403          * describing its "scope".
 404          */
 405         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 406         if (rc != 0) {
 407                 ptlrpc_set_destroy(pc->pc_set);
 408                 GOTO(out, rc);
 409         }
 410
 411 #ifdef __KERNEL__
 412         rc = cfs_kernel_thread(ptlrpcd, pc, 0);
 413         if (rc < 0)  {
 414                 lu_context_fini(&pc->pc_env.le_ctx);
 415                 ptlrpc_set_destroy(pc->pc_set);
 416                 GOTO(out, rc);
 417         }
 418         rc = 0;
 419         cfs_wait_for_completion(&pc->pc_starting);
 420 #else
 421         pc->pc_wait_callback =
 422                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 423                                                  &ptlrpcd_check_async_rpcs, pc);
 424         pc->pc_idle_callback =
 425                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 426                                                  &ptlrpcd_idle, pc);
 427 #endif
 428 out:
 429         if (rc)
 430                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 431         RETURN(rc);
 432 }
 433
 434 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 435 {
 436         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 437                 CERROR("Thread for pc %p was not started\n", pc);
 438                 return;
 439         }
 440
 441         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 442         if (force)
 443                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 444         cfs_waitq_signal(&pc->pc_set->set_waitq);
 445 #ifdef __KERNEL__
 446         cfs_wait_for_completion(&pc->pc_finishing);
 447 #else
 448         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 449         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 450 #endif
 451         lu_context_fini(&pc->pc_env.le_ctx);
 452         ptlrpc_set_destroy(pc->pc_set);
 453 }
 454
 455 void ptlrpcd_fini(void)
 456 {
 457         int i;
 458         int j;
 459
 460         ENTRY;
 461
 462         for (i = 0; i < PSCOPE_NR; ++i) {
 463                 for (j = 0; j < PT_NR; ++j) {
 464                         struct ptlrpcd_ctl *pc;
 465
 466                         pc = &ptlrpcd_scopes[i].pscope_thread[j].pt_ctl;
 467
 468                         if (cfs_test_bit(LIOD_START, &pc->pc_flags))
 469                                 ptlrpcd_stop(pc, 0);
 470                 }
 471         }
 472         EXIT;
 473 }
 474
 475 int ptlrpcd_addref(void)
 476 {
 477         int rc = 0;
 478         int i;
 479         int j;
 480         ENTRY;
 481
 482         cfs_mutex_down(&ptlrpcd_sem);
 483         if (++ptlrpcd_users == 1) {
 484                 for (i = 0; rc == 0 && i < PSCOPE_NR; ++i) {
 485                         for (j = 0; rc == 0 && j < PT_NR; ++j) {
 486                                 struct ptlrpcd_thread *pt;
 487                                 struct ptlrpcd_ctl    *pc;
 488
 489                                 pt = &ptlrpcd_scopes[i].pscope_thread[j];
 490                                 pc = &pt->pt_ctl;
 491                                 if (j == PT_RECOVERY)
 492                                         cfs_set_bit(LIOD_RECOVERY, &pc->pc_flags);
 493                                 rc = ptlrpcd_start(pt->pt_name, pc);
 494                         }
 495                 }
 496                 if (rc != 0) {
 497                         --ptlrpcd_users;
 498                         ptlrpcd_fini();
 499                 }
 500         }
 501         cfs_mutex_up(&ptlrpcd_sem);
 502         RETURN(rc);
 503 }
 504
 505 void ptlrpcd_decref(void)
 506 {
 507         cfs_mutex_down(&ptlrpcd_sem);
 508         if (--ptlrpcd_users == 0)
 509                 ptlrpcd_fini();
 510         cfs_mutex_up(&ptlrpcd_sem);
 511 }