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