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