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