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 (c) 2003, 2010, Oracle and/or its affiliates. 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 /** \defgroup ptlrpcd PortalRPC daemon
  40  *
  41  * ptlrpcd is a special thread with its own set where other user might add
  42  * requests when they don't want to wait for their completion.
  43  * PtlRPCD will take care of sending such requests and then processing their
  44  * replies and calling completion callbacks as necessary.
  45  * The callbacks are called directly from ptlrpcd context.
  46  * It is important to never significantly block (esp. on RPCs!) within such
  47  * completion handler or a deadlock might occur where ptlrpcd enters some
  48  * callback that attempts to send another RPC and wait for it to return,
  49  * during which time ptlrpcd is completely blocked, so e.g. if import
  50  * fails, recovery cannot progress because connection requests are also
  51  * sent by ptlrpcd.
  52  *
  53  * @{
  54  */
  55
  56 #define DEBUG_SUBSYSTEM S_RPC
  57
  58 #ifdef __KERNEL__
  59 # include <libcfs/libcfs.h>
  60 #else /* __KERNEL__ */
  61 # include <liblustre.h>
  62 # include <ctype.h>
  63 #endif
  64
  65 #include <lustre_net.h>
  66 # include <lustre_lib.h>
  67
  68 #include <lustre_ha.h>
  69 #include <obd_class.h>   /* for obd_zombie */
  70 #include <obd_support.h> /* for OBD_FAIL_CHECK */
  71 #include <cl_object.h> /* cl_env_{get,put}() */
  72 #include <lprocfs_status.h>
  73
  74 enum pscope_thread {
  75         PT_NORMAL,
  76         PT_RECOVERY,
  77         PT_NR
  78 };
  79
  80 struct ptlrpcd_scope_ctl {
  81         struct ptlrpcd_thread {
  82                 const char        *pt_name;
  83                 struct ptlrpcd_ctl pt_ctl;
  84         } pscope_thread[PT_NR];
  85 };
  86
  87 static struct ptlrpcd_scope_ctl ptlrpcd_scopes[PSCOPE_NR] = {
  88         [PSCOPE_BRW] = {
  89                 .pscope_thread = {
  90                         [PT_NORMAL] = {
  91                                 .pt_name = "ptlrpcd-brw"
  92                         },
  93                         [PT_RECOVERY] = {
  94                                 .pt_name = "ptlrpcd-brw-rcv"
  95                         }
  96                 }
  97         },
  98         [PSCOPE_OTHER] = {
  99                 .pscope_thread = {
 100                         [PT_NORMAL] = {
 101                                 .pt_name = "ptlrpcd"
 102                         },
 103                         [PT_RECOVERY] = {
 104                                 .pt_name = "ptlrpcd-rcv"
 105                         }
 106                 }
 107         }
 108 };
 109
 110 cfs_semaphore_t ptlrpcd_sem;
 111 static int ptlrpcd_users = 0;
 112
 113 void ptlrpcd_wake(struct ptlrpc_request *req)
 114 {
 115         struct ptlrpc_request_set *rq_set = req->rq_set;
 116
 117         LASSERT(rq_set != NULL);
 118
 119         cfs_waitq_signal(&rq_set->set_waitq);
 120 }
 121
 122 /**
 123  * Move all request from an existing request set to the ptlrpcd queue.
 124  * All requests from the set must be in phase RQ_PHASE_NEW.
 125  */
 126 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
 127 {
 128         cfs_list_t *tmp, *pos;
 129
 130         cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
 131                 struct ptlrpc_request *req =
 132                         cfs_list_entry(pos, struct ptlrpc_request,
 133                                        rq_set_chain);
 134
 135                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 136                 cfs_list_del_init(&req->rq_set_chain);
 137                 req->rq_set = NULL;
 138                 ptlrpcd_add_req(req, PSCOPE_OTHER);
 139                 cfs_atomic_dec(&set->set_remaining);
 140         }
 141         LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
 142 }
 143 EXPORT_SYMBOL(ptlrpcd_add_rqset);
 144
 145 /**
 146  * Requests that are added to the ptlrpcd queue are sent via
 147  * ptlrpcd_check->ptlrpc_check_set().
 148  */
 149 int ptlrpcd_add_req(struct ptlrpc_request *req, enum ptlrpcd_scope scope)
 150 {
 151         struct ptlrpcd_ctl *pc;
 152         enum pscope_thread  pt;
 153         int rc;
 154
 155         LASSERT(scope < PSCOPE_NR);
 156
 157         cfs_spin_lock(&req->rq_lock);
 158         if (req->rq_invalid_rqset) {
 159                 cfs_duration_t timeout;
 160                 struct l_wait_info lwi;
 161
 162                 req->rq_invalid_rqset = 0;
 163                 cfs_spin_unlock(&req->rq_lock);
 164
 165                 timeout = cfs_time_seconds(5);
 166                 lwi = LWI_TIMEOUT(timeout, back_to_sleep, NULL);
 167                 l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi);
 168         } else if (req->rq_set) {
 169                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
 170                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
 171
 172                 /* ptlrpc_check_set will decrease the count */
 173                 cfs_atomic_inc(&req->rq_set->set_remaining);
 174                 cfs_spin_unlock(&req->rq_lock);
 175
 176                 cfs_waitq_signal(&req->rq_set->set_waitq);
 177         } else {
 178                 cfs_spin_unlock(&req->rq_lock);
 179         }
 180
 181         pt = req->rq_send_state == LUSTRE_IMP_FULL ? PT_NORMAL : PT_RECOVERY;
 182         pc = &ptlrpcd_scopes[scope].pscope_thread[pt].pt_ctl;
 183         rc = ptlrpc_set_add_new_req(pc, req);
 184         /*
 185          * XXX disable this for CLIO: environment is needed for interpreter.
 186          *     add debug temporary to check rc.
 187          */
 188         LASSERTF(rc == 0, "ptlrpcd_add_req failed (rc = %d)\n", rc);
 189         if (rc && 0) {
 190                 /*
 191                  * Thread is probably in stop now so we need to
 192                  * kill this rpc as it was not added. Let's call
 193                  * interpret for it to let know we're killing it
 194                  * so that higher levels might free associated
 195                  * resources.
 196                  */
 197                 ptlrpc_req_interpret(NULL, req, -EBADR);
 198                 req->rq_set = NULL;
 199                 ptlrpc_req_finished(req);
 200         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING) {
 201                 /*
 202                  * The request is for recovery, should be sent ASAP.
 203                  */
 204                 cfs_waitq_signal(&pc->pc_set->set_waitq);
 205         }
 206
 207         return rc;
 208 }
 209
 210 /**
 211  * Check if there is more work to do on ptlrpcd set.
 212  * Returns 1 if yes.
 213  */
 214 static int ptlrpcd_check(const struct lu_env *env, struct ptlrpcd_ctl *pc)
 215 {
 216         cfs_list_t *tmp, *pos;
 217         struct ptlrpc_request *req;
 218         int rc = 0;
 219         ENTRY;
 220
 221         cfs_spin_lock(&pc->pc_set->set_new_req_lock);
 222         cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 223                 req = cfs_list_entry(pos, struct ptlrpc_request, rq_set_chain);
 224                 cfs_list_del_init(&req->rq_set_chain);
 225                 ptlrpc_set_add_req(pc->pc_set, req);
 226                 /*
 227                  * Need to calculate its timeout.
 228                  */
 229                 rc = 1;
 230         }
 231         cfs_spin_unlock(&pc->pc_set->set_new_req_lock);
 232
 233         if (cfs_atomic_read(&pc->pc_set->set_remaining)) {
 234                 rc = rc | ptlrpc_check_set(env, pc->pc_set);
 235
 236                 /*
 237                  * XXX: our set never completes, so we prune the completed
 238                  * reqs after each iteration. boy could this be smarter.
 239                  */
 240                 cfs_list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 241                         req = cfs_list_entry(pos, struct ptlrpc_request,
 242                                          rq_set_chain);
 243                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 244                                 continue;
 245
 246                         cfs_list_del_init(&req->rq_set_chain);
 247                         req->rq_set = NULL;
 248                         ptlrpc_req_finished (req);
 249                 }
 250         }
 251
 252         if (rc == 0) {
 253                 /*
 254                  * If new requests have been added, make sure to wake up.
 255                  */
 256                 cfs_spin_lock(&pc->pc_set->set_new_req_lock);
 257                 rc = !cfs_list_empty(&pc->pc_set->set_new_requests);
 258                 cfs_spin_unlock(&pc->pc_set->set_new_req_lock);
 259         }
 260
 261         RETURN(rc);
 262 }
 263
 264 #ifdef __KERNEL__
 265 /**
 266  * Main ptlrpcd thread.
 267  * ptlrpc's code paths like to execute in process context, so we have this
 268  * thread which spins on a set which contains the rpcs and sends them.
 269  *
 270  */
 271 static int ptlrpcd(void *arg)
 272 {
 273         struct ptlrpcd_ctl *pc = arg;
 274         struct lu_env env = { .le_ses = NULL };
 275         int rc, exit = 0;
 276         ENTRY;
 277
 278         rc = cfs_daemonize_ctxt(pc->pc_name);
 279         if (rc == 0) {
 280                 /*
 281                  * XXX So far only "client" ptlrpcd uses an environment. In
 282                  * the future, ptlrpcd thread (or a thread-set) has to given
 283                  * an argument, describing its "scope".
 284                  */
 285                 rc = lu_context_init(&env.le_ctx,
 286                                      LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF);
 287         }
 288
 289         cfs_complete(&pc->pc_starting);
 290
 291         if (rc != 0)
 292                 RETURN(rc);
 293         env.le_ctx.lc_cookie = 0x7;
 294
 295         /*
 296          * This mainloop strongly resembles ptlrpc_set_wait() except that our
 297          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
 298          * there are requests in the set. New requests come in on the set's
 299          * new_req_list and ptlrpcd_check() moves them into the set.
 300          */
 301         do {
 302                 struct l_wait_info lwi;
 303                 int timeout;
 304
 305                 rc = lu_env_refill(&env);
 306                 if (rc != 0) {
 307                         /*
 308                          * XXX This is very awkward situation, because
 309                          * execution can neither continue (request
 310                          * interpreters assume that env is set up), nor repeat
 311                          * the loop (as this potentially results in a tight
 312                          * loop of -ENOMEM's).
 313                          *
 314                          * Fortunately, refill only ever does something when
 315                          * new modules are loaded, i.e., early during boot up.
 316                          */
 317                         CERROR("Failure to refill session: %d\n", rc);
 318                         continue;
 319                 }
 320
 321                 timeout = ptlrpc_set_next_timeout(pc->pc_set);
 322                 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
 323                                   ptlrpc_expired_set, pc->pc_set);
 324
 325                 lu_context_enter(&env.le_ctx);
 326                 l_wait_event(pc->pc_set->set_waitq,
 327                              ptlrpcd_check(&env, pc), &lwi);
 328                 lu_context_exit(&env.le_ctx);
 329
 330                 /*
 331                  * Abort inflight rpcs for forced stop case.
 332                  */
 333                 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags)) {
 334                         if (cfs_test_bit(LIOD_FORCE, &pc->pc_flags))
 335                                 ptlrpc_abort_set(pc->pc_set);
 336                         exit++;
 337                 }
 338
 339                 /*
 340                  * Let's make one more loop to make sure that ptlrpcd_check()
 341                  * copied all raced new rpcs into the set so we can kill them.
 342                  */
 343         } while (exit < 2);
 344
 345         /*
 346          * Wait for inflight requests to drain.
 347          */
 348         if (!cfs_list_empty(&pc->pc_set->set_requests))
 349                 ptlrpc_set_wait(pc->pc_set);
 350         lu_context_fini(&env.le_ctx);
 351         cfs_complete(&pc->pc_finishing);
 352
 353         cfs_clear_bit(LIOD_START, &pc->pc_flags);
 354         cfs_clear_bit(LIOD_STOP, &pc->pc_flags);
 355         cfs_clear_bit(LIOD_FORCE, &pc->pc_flags);
 356         return 0;
 357 }
 358
 359 #else /* !__KERNEL__ */
 360
 361 /**
 362  * In liblustre we do not have separate threads, so this function
 363  * is called from time to time all across common code to see
 364  * if something needs to be processed on ptlrpcd set.
 365  */
 366 int ptlrpcd_check_async_rpcs(void *arg)
 367 {
 368         struct ptlrpcd_ctl *pc = arg;
 369         int                 rc = 0;
 370
 371         /*
 372          * Single threaded!!
 373          */
 374         pc->pc_recurred++;
 375
 376         if (pc->pc_recurred == 1) {
 377                 rc = lu_env_refill(&pc->pc_env);
 378                 if (rc == 0) {
 379                         lu_context_enter(&pc->pc_env.le_ctx);
 380                         rc = ptlrpcd_check(&pc->pc_env, pc);
 381                         lu_context_exit(&pc->pc_env.le_ctx);
 382                         if (!rc)
 383                                 ptlrpc_expired_set(pc->pc_set);
 384                         /*
 385                          * XXX: send replay requests.
 386                          */
 387                         if (cfs_test_bit(LIOD_RECOVERY, &pc->pc_flags))
 388                                 rc = ptlrpcd_check(&pc->pc_env, pc);
 389                 }
 390         }
 391
 392         pc->pc_recurred--;
 393         return rc;
 394 }
 395
 396 int ptlrpcd_idle(void *arg)
 397 {
 398         struct ptlrpcd_ctl *pc = arg;
 399
 400         return (cfs_list_empty(&pc->pc_set->set_new_requests) &&
 401                 cfs_atomic_read(&pc->pc_set->set_remaining) == 0);
 402 }
 403
 404 #endif
 405
 406 int ptlrpcd_start(const char *name, struct ptlrpcd_ctl *pc)
 407 {
 408         int rc;
 409         ENTRY;
 410
 411         /*
 412          * Do not allow start second thread for one pc.
 413          */
 414         if (cfs_test_and_set_bit(LIOD_START, &pc->pc_flags)) {
 415                 CERROR("Starting second thread (%s) for same pc %p\n",
 416                        name, pc);
 417                 RETURN(-EALREADY);
 418         }
 419
 420         cfs_init_completion(&pc->pc_starting);
 421         cfs_init_completion(&pc->pc_finishing);
 422         cfs_spin_lock_init(&pc->pc_lock);
 423         strncpy(pc->pc_name, name, sizeof(pc->pc_name) - 1);
 424         pc->pc_set = ptlrpc_prep_set();
 425         if (pc->pc_set == NULL)
 426                 GOTO(out, rc = -ENOMEM);
 427         /*
 428          * So far only "client" ptlrpcd uses an environment. In the future,
 429          * ptlrpcd thread (or a thread-set) has to be given an argument,
 430          * describing its "scope".
 431          */
 432         rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER);
 433         if (rc != 0) {
 434                 ptlrpc_set_destroy(pc->pc_set);
 435                 GOTO(out, rc);
 436         }
 437
 438 #ifdef __KERNEL__
 439         rc = cfs_create_thread(ptlrpcd, pc, 0);
 440         if (rc < 0)  {
 441                 lu_context_fini(&pc->pc_env.le_ctx);
 442                 ptlrpc_set_destroy(pc->pc_set);
 443                 GOTO(out, rc);
 444         }
 445         rc = 0;
 446         cfs_wait_for_completion(&pc->pc_starting);
 447 #else
 448         pc->pc_wait_callback =
 449                 liblustre_register_wait_callback("ptlrpcd_check_async_rpcs",
 450                                                  &ptlrpcd_check_async_rpcs, pc);
 451         pc->pc_idle_callback =
 452                 liblustre_register_idle_callback("ptlrpcd_check_idle_rpcs",
 453                                                  &ptlrpcd_idle, pc);
 454 #endif
 455 out:
 456         if (rc)
 457                 cfs_clear_bit(LIOD_START, &pc->pc_flags);
 458         RETURN(rc);
 459 }
 460
 461 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
 462 {
 463         if (!cfs_test_bit(LIOD_START, &pc->pc_flags)) {
 464                 CERROR("Thread for pc %p was not started\n", pc);
 465                 return;
 466         }
 467
 468         cfs_set_bit(LIOD_STOP, &pc->pc_flags);
 469         if (force)
 470                 cfs_set_bit(LIOD_FORCE, &pc->pc_flags);
 471         cfs_waitq_signal(&pc->pc_set->set_waitq);
 472 #ifdef __KERNEL__
 473         cfs_wait_for_completion(&pc->pc_finishing);
 474 #else
 475         liblustre_deregister_wait_callback(pc->pc_wait_callback);
 476         liblustre_deregister_idle_callback(pc->pc_idle_callback);
 477 #endif
 478         lu_context_fini(&pc->pc_env.le_ctx);
 479         ptlrpc_set_destroy(pc->pc_set);
 480 }
 481
 482 void ptlrpcd_fini(void)
 483 {
 484         int i;
 485         int j;
 486
 487         ENTRY;
 488
 489         for (i = 0; i < PSCOPE_NR; ++i) {
 490                 for (j = 0; j < PT_NR; ++j) {
 491                         struct ptlrpcd_ctl *pc;
 492
 493                         pc = &ptlrpcd_scopes[i].pscope_thread[j].pt_ctl;
 494
 495                         if (cfs_test_bit(LIOD_START, &pc->pc_flags))
 496                                 ptlrpcd_stop(pc, 0);
 497                 }
 498         }
 499         EXIT;
 500 }
 501
 502 int ptlrpcd_addref(void)
 503 {
 504         int rc = 0;
 505         int i;
 506         int j;
 507         ENTRY;
 508
 509         cfs_mutex_down(&ptlrpcd_sem);
 510         if (++ptlrpcd_users == 1) {
 511                 for (i = 0; rc == 0 && i < PSCOPE_NR; ++i) {
 512                         for (j = 0; rc == 0 && j < PT_NR; ++j) {
 513                                 struct ptlrpcd_thread *pt;
 514                                 struct ptlrpcd_ctl    *pc;
 515
 516                                 pt = &ptlrpcd_scopes[i].pscope_thread[j];
 517                                 pc = &pt->pt_ctl;
 518                                 if (j == PT_RECOVERY)
 519                                         cfs_set_bit(LIOD_RECOVERY, &pc->pc_flags);
 520                                 rc = ptlrpcd_start(pt->pt_name, pc);
 521                         }
 522                 }
 523                 if (rc != 0) {
 524                         --ptlrpcd_users;
 525                         ptlrpcd_fini();
 526                 }
 527         }
 528         cfs_mutex_up(&ptlrpcd_sem);
 529         RETURN(rc);
 530 }
 531
 532 void ptlrpcd_decref(void)
 533 {
 534         cfs_mutex_down(&ptlrpcd_sem);
 535         if (--ptlrpcd_users == 0)
 536                 ptlrpcd_fini();
 537         cfs_mutex_up(&ptlrpcd_sem);
 538 }
 539 /** @} ptlrpcd */