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