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