lustre/ptlrpc/service.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_RPC
  24 #ifndef __KERNEL__
  25 #include <liblustre.h>
  26 #include <linux/kp30.h>
  27 #endif
  28 #include <linux/obd_support.h>
  29 #include <linux/obd_class.h>
  30 #include <linux/lustre_net.h>
  31 #include <portals/types.h>
  32 #include "ptlrpc_internal.h"
  33
  34 static LIST_HEAD (ptlrpc_all_services);
  35 static spinlock_t ptlrpc_all_services_lock = SPIN_LOCK_UNLOCKED;
  36
  37 static void
  38 ptlrpc_free_server_req (struct ptlrpc_request *req)
  39 {
  40         /* The last request to be received into a request buffer uses space
  41          * in the request buffer descriptor, otherwise requests are
  42          * allocated dynamically in the incoming reply event handler */
  43         if (req == &req->rq_rqbd->rqbd_req)
  44                 return;
  45
  46         OBD_FREE(req, sizeof(*req));
  47 }
  48
  49 static char *
  50 ptlrpc_alloc_request_buffer (int size)
  51 {
  52         char *ptr;
  53
  54         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  55                 OBD_VMALLOC(ptr, size);
  56         else
  57                 OBD_ALLOC(ptr, size);
  58
  59         return (ptr);
  60 }
  61
  62 static void
  63 ptlrpc_free_request_buffer (char *ptr, int size)
  64 {
  65         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  66                 OBD_VFREE(ptr, size);
  67         else
  68                 OBD_FREE(ptr, size);
  69 }
  70
  71 struct ptlrpc_request_buffer_desc *
  72 ptlrpc_alloc_rqbd (struct ptlrpc_srv_ni *srv_ni)
  73 {
  74         struct ptlrpc_service             *svc = srv_ni->sni_service;
  75         unsigned long                      flags;
  76         struct ptlrpc_request_buffer_desc *rqbd;
  77
  78         OBD_ALLOC(rqbd, sizeof (*rqbd));
  79         if (rqbd == NULL)
  80                 return (NULL);
  81
  82         rqbd->rqbd_srv_ni = srv_ni;
  83         rqbd->rqbd_refcount = 0;
  84         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  85         rqbd->rqbd_cbid.cbid_arg = rqbd;
  86         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
  87
  88         if (rqbd->rqbd_buffer == NULL) {
  89                 OBD_FREE(rqbd, sizeof (*rqbd));
  90                 return (NULL);
  91         }
  92
  93         spin_lock_irqsave (&svc->srv_lock, flags);
  94         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
  95         svc->srv_nbufs++;
  96         spin_unlock_irqrestore (&svc->srv_lock, flags);
  97
  98         return (rqbd);
  99 }
 100
 101 void
 102 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 103 {
 104         struct ptlrpc_srv_ni  *sni = rqbd->rqbd_srv_ni;
 105         struct ptlrpc_service *svc = sni->sni_service;
 106         unsigned long          flags;
 107
 108         LASSERT (rqbd->rqbd_refcount == 0);
 109
 110         spin_lock_irqsave(&svc->srv_lock, flags);
 111         list_del(&rqbd->rqbd_list);
 112         svc->srv_nbufs--;
 113         spin_unlock_irqrestore(&svc->srv_lock, flags);
 114
 115         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 116         OBD_FREE (rqbd, sizeof (*rqbd));
 117 }
 118
 119 void
 120 ptlrpc_save_lock (struct ptlrpc_request *req,
 121                   struct lustre_handle *lock, int mode)
 122 {
 123         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 124         int                        idx;
 125
 126         LASSERT (rs != NULL);
 127         LASSERT (rs->rs_nlocks < RS_MAX_LOCKS);
 128
 129         idx = rs->rs_nlocks++;
 130         rs->rs_locks[idx] = *lock;
 131         rs->rs_modes[idx] = mode;
 132         rs->rs_difficult = 1;
 133 }
 134
 135 void
 136 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 137 {
 138         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 139
 140 #ifdef CONFIG_SMP
 141         LASSERT (spin_is_locked (&svc->srv_lock));
 142 #endif
 143         LASSERT (rs->rs_difficult);
 144         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 145
 146         if (rs->rs_scheduled)                   /* being set up or already notified */
 147                 return;
 148
 149         rs->rs_scheduled = 1;
 150         list_del (&rs->rs_list);
 151         list_add (&rs->rs_list, &svc->srv_reply_queue);
 152         wake_up (&svc->srv_waitq);
 153 }
 154
 155 void
 156 ptlrpc_commit_replies (struct obd_device *obd)
 157 {
 158         struct list_head   *tmp;
 159         struct list_head   *nxt;
 160         unsigned long       flags;
 161
 162         /* Find any replies that have been committed and get their service
 163          * to attend to complete them. */
 164
 165         /* CAVEAT EMPTOR: spinlock ordering!!! */
 166         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
 167
 168         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 169                 struct ptlrpc_reply_state *rs =
 170                         list_entry (tmp, struct ptlrpc_reply_state, rs_obd_list);
 171
 172                 LASSERT (rs->rs_difficult);
 173
 174                 if (rs->rs_transno <= obd->obd_last_committed) {
 175                         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 176
 177                         spin_lock (&svc->srv_lock);
 178                         list_del_init (&rs->rs_obd_list);
 179                         ptlrpc_schedule_difficult_reply (rs);
 180                         spin_unlock (&svc->srv_lock);
 181                 }
 182         }
 183
 184         spin_unlock_irqrestore (&obd->obd_uncommitted_replies_lock, flags);
 185 }
 186
 187 static long
 188 timeval_sub(struct timeval *large, struct timeval *small)
 189 {
 190         return (large->tv_sec - small->tv_sec) * 1000000 +
 191                 (large->tv_usec - small->tv_usec);
 192 }
 193
 194 static int
 195 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 196 {
 197         struct ptlrpc_srv_ni              *srv_ni;
 198         struct ptlrpc_request_buffer_desc *rqbd;
 199         unsigned long                      flags;
 200         int                                rc;
 201
 202         spin_lock_irqsave(&svc->srv_lock, flags);
 203         if (list_empty (&svc->srv_idle_rqbds)) {
 204                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 205                 return (0);
 206         }
 207
 208         rqbd = list_entry(svc->srv_idle_rqbds.next,
 209                           struct ptlrpc_request_buffer_desc,
 210                           rqbd_list);
 211         list_del (&rqbd->rqbd_list);
 212
 213         /* assume we will post successfully */
 214         srv_ni = rqbd->rqbd_srv_ni;
 215         srv_ni->sni_nrqbd_receiving++;
 216         list_add (&rqbd->rqbd_list, &srv_ni->sni_active_rqbds);
 217
 218         spin_unlock_irqrestore(&svc->srv_lock, flags);
 219
 220         rc = ptlrpc_register_rqbd(rqbd);
 221         if (rc == 0)
 222                 return (1);
 223
 224         spin_lock_irqsave(&svc->srv_lock, flags);
 225
 226         srv_ni->sni_nrqbd_receiving--;
 227         list_del(&rqbd->rqbd_list);
 228         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 229
 230         if (srv_ni->sni_nrqbd_receiving == 0) {
 231                 /* This service is off-air on this interface because all
 232                  * its request buffers are busy.  Portals will have started
 233                  * dropping incoming requests until more buffers get
 234                  * posted */
 235                 CERROR("All %s %s request buffers busy\n",
 236                        svc->srv_name, srv_ni->sni_ni->pni_name);
 237         }
 238
 239         spin_unlock_irqrestore (&svc->srv_lock, flags);
 240
 241         return (-1);
 242 }
 243
 244 struct ptlrpc_service *
 245 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size,
 246                 int req_portal, int rep_portal,
 247                 svc_handler_t handler, char *name,
 248                 struct proc_dir_entry *proc_entry)
 249 {
 250         int                                i;
 251         int                                j;
 252         int                                ssize;
 253         struct ptlrpc_service             *service;
 254         struct ptlrpc_srv_ni              *srv_ni;
 255         struct ptlrpc_request_buffer_desc *rqbd;
 256         ENTRY;
 257
 258         LASSERT (ptlrpc_ninterfaces > 0);
 259         LASSERT (nbufs > 0);
 260         LASSERT (bufsize >= max_req_size);
 261
 262         ssize = offsetof (struct ptlrpc_service,
 263                           srv_interfaces[ptlrpc_ninterfaces]);
 264         OBD_ALLOC(service, ssize);
 265         if (service == NULL)
 266                 RETURN(NULL);
 267
 268         service->srv_name = name;
 269         spin_lock_init(&service->srv_lock);
 270         INIT_LIST_HEAD(&service->srv_threads);
 271         init_waitqueue_head(&service->srv_waitq);
 272
 273         service->srv_max_req_size = max_req_size;
 274         service->srv_buf_size = bufsize;
 275         service->srv_rep_portal = rep_portal;
 276         service->srv_req_portal = req_portal;
 277         service->srv_handler = handler;
 278
 279         INIT_LIST_HEAD(&service->srv_request_queue);
 280         INIT_LIST_HEAD(&service->srv_idle_rqbds);
 281         INIT_LIST_HEAD(&service->srv_reply_queue);
 282
 283         /* First initialise enough for early teardown */
 284         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 285                 srv_ni = &service->srv_interfaces[i];
 286
 287                 srv_ni->sni_service = service;
 288                 srv_ni->sni_ni = &ptlrpc_interfaces[i];
 289                 INIT_LIST_HEAD(&srv_ni->sni_active_rqbds);
 290                 INIT_LIST_HEAD(&srv_ni->sni_active_replies);
 291         }
 292
 293         spin_lock (&ptlrpc_all_services_lock);
 294         list_add (&service->srv_list, &ptlrpc_all_services);
 295         spin_unlock (&ptlrpc_all_services_lock);
 296
 297         /* Now allocate the request buffers, assuming all interfaces require
 298          * the same number. */
 299         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 300                 srv_ni = &service->srv_interfaces[i];
 301                 CDEBUG (D_NET, "%s: initialising interface %s\n", name,
 302                         srv_ni->sni_ni->pni_name);
 303
 304                 for (j = 0; j < nbufs; j++) {
 305                         rqbd = ptlrpc_alloc_rqbd (srv_ni);
 306
 307                         if (rqbd == NULL) {
 308                                 CERROR ("%s.%d: Can't allocate request %d "
 309                                         "on %s\n", name, i, j,
 310                                         srv_ni->sni_ni->pni_name);
 311                                 GOTO(failed, NULL);
 312                         }
 313
 314                         /* We shouldn't be under memory pressure at
 315                          * startup, so fail if we can't post all our
 316                          * buffers at this time. */
 317                         if (ptlrpc_server_post_idle_rqbds(service) <= 0)
 318                                 GOTO(failed, NULL);
 319                 }
 320         }
 321
 322         if (proc_entry != NULL)
 323                 ptlrpc_lprocfs_register_service(proc_entry, service);
 324
 325         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 326                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 327
 328         RETURN(service);
 329 failed:
 330         ptlrpc_unregister_service(service);
 331         return NULL;
 332 }
 333
 334 static void
 335 ptlrpc_server_free_request(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 336 {
 337         unsigned long  flags;
 338         int            refcount;
 339
 340         spin_lock_irqsave(&svc->srv_lock, flags);
 341         svc->srv_n_active_reqs--;
 342         refcount = --(req->rq_rqbd->rqbd_refcount);
 343         if (refcount == 0) {
 344                 /* request buffer is now idle */
 345                 list_del(&req->rq_rqbd->rqbd_list);
 346                 list_add_tail(&req->rq_rqbd->rqbd_list,
 347                               &svc->srv_idle_rqbds);
 348         }
 349         spin_unlock_irqrestore(&svc->srv_lock, flags);
 350
 351         ptlrpc_free_server_req(req);
 352 }
 353
 354 static int
 355 ptlrpc_server_handle_request (struct ptlrpc_service *svc)
 356 {
 357         struct ptlrpc_request *request;
 358         unsigned long          flags;
 359         struct timeval         work_start;
 360         struct timeval         work_end;
 361         long                   timediff;
 362         int                    rc;
 363         char                   str[PTL_NALFMT_SIZE];
 364         ENTRY;
 365
 366         spin_lock_irqsave (&svc->srv_lock, flags);
 367         if (list_empty (&svc->srv_request_queue) ||
 368             (svc->srv_n_difficult_replies != 0 &&
 369              svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
 370                 /* If all the other threads are handling requests, I must
 371                  * remain free to handle any 'difficult' reply that might
 372                  * block them */
 373                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 374                 RETURN(0);
 375         }
 376
 377         request = list_entry (svc->srv_request_queue.next,
 378                               struct ptlrpc_request, rq_list);
 379         list_del_init (&request->rq_list);
 380         svc->srv_n_queued_reqs--;
 381         svc->srv_n_active_reqs++;
 382
 383         spin_unlock_irqrestore (&svc->srv_lock, flags);
 384
 385         do_gettimeofday(&work_start);
 386         timediff = timeval_sub(&work_start, &request->rq_arrival_time);
 387         if (svc->srv_stats != NULL) {
 388                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 389                                     timediff);
 390                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 391                                     svc->srv_n_queued_reqs);
 392                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 393                                     svc->srv_n_active_reqs);
 394         }
 395
 396 #if SWAB_PARANOIA
 397         /* Clear request swab mask; this is a new request */
 398         request->rq_req_swab_mask = 0;
 399 #endif
 400         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 401         if (rc != 0) {
 402                 CERROR ("error unpacking request: ptl %d from %s"
 403                         " xid "LPU64"\n", svc->srv_req_portal,
 404                         ptlrpc_peernid2str(&request->rq_peer, str),
 405                        request->rq_xid);
 406                 goto out;
 407         }
 408
 409         rc = -EINVAL;
 410         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 411                 CERROR("wrong packet type received (type=%u) from %s\n",
 412                        request->rq_reqmsg->type,
 413                        ptlrpc_peernid2str(&request->rq_peer, str));
 414                 goto out;
 415         }
 416
 417         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 418
 419         /* Discard requests queued for longer than my timeout.  If the
 420          * client's timeout is similar to mine, she'll be timing out this
 421          * REQ anyway (bug 1502) */
 422         if (timediff / 1000000 > (long)obd_timeout) {
 423                 CERROR("Dropping timed-out request from %s: %ld seconds old\n",
 424                        ptlrpc_peernid2str(&request->rq_peer, str),
 425                        timediff / 1000000);
 426                 goto out;
 427         }
 428
 429         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 430
 431         if (request->rq_export) {
 432                 if (request->rq_reqmsg->conn_cnt <
 433                     request->rq_export->exp_conn_cnt) {
 434                         DEBUG_REQ(D_ERROR, request,
 435                                   "DROPPING req from old connection %d < %d",
 436                                   request->rq_reqmsg->conn_cnt,
 437                                   request->rq_export->exp_conn_cnt);
 438                         goto put_conn;
 439                 }
 440
 441                 request->rq_export->exp_last_request_time =
 442                         LTIME_S(CURRENT_TIME);
 443         }
 444
 445         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 446                "%s:%s+%d:%d:"LPU64":%s:%s:%d\n", current->comm,
 447                (request->rq_export ?
 448                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 449                (request->rq_export ?
 450                 atomic_read(&request->rq_export->exp_refcount) : -99),
 451                request->rq_reqmsg->status, request->rq_xid,
 452                request->rq_peer.peer_ni->pni_name,
 453                ptlrpc_peernid2str(&request->rq_peer, str),
 454                request->rq_reqmsg->opc);
 455
 456         rc = svc->srv_handler(request);
 457         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 458                "%s:%s+%d:%d:"LPU64":%s:%s:%d\n", current->comm,
 459                (request->rq_export ?
 460                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 461                (request->rq_export ?
 462                 atomic_read(&request->rq_export->exp_refcount) : -99),
 463                request->rq_reqmsg->status, request->rq_xid,
 464                request->rq_peer.peer_ni->pni_name,
 465                ptlrpc_peernid2str(&request->rq_peer, str),
 466                request->rq_reqmsg->opc);
 467
 468 put_conn:
 469         if (request->rq_export != NULL)
 470                 class_export_put(request->rq_export);
 471
 472  out:
 473         do_gettimeofday(&work_end);
 474
 475         timediff = timeval_sub(&work_end, &work_start);
 476
 477         CDEBUG((timediff / 1000000 > (long)obd_timeout) ? D_ERROR : D_HA,
 478                "request "LPU64" opc %u from NID %s processed in %ldus "
 479                "(%ldus total)\n", request->rq_xid, request->rq_reqmsg->opc,
 480                ptlrpc_peernid2str(&request->rq_peer, str),
 481                timediff, timeval_sub(&work_end, &request->rq_arrival_time));
 482
 483         if (svc->srv_stats != NULL) {
 484                 int opc = opcode_offset(request->rq_reqmsg->opc);
 485                 if (opc > 0) {
 486                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 487                         lprocfs_counter_add(svc->srv_stats,
 488                                             opc + PTLRPC_LAST_CNTR,
 489                                             timediff);
 490                 }
 491         }
 492
 493         ptlrpc_server_free_request(svc, request);
 494
 495         RETURN(1);
 496 }
 497
 498 static int
 499 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 500 {
 501         struct ptlrpc_reply_state *rs;
 502         unsigned long              flags;
 503         struct obd_export         *exp;
 504         struct obd_device         *obd;
 505         int                        nlocks;
 506         int                        been_handled;
 507         char                       str[PTL_NALFMT_SIZE];
 508         ENTRY;
 509
 510         spin_lock_irqsave (&svc->srv_lock, flags);
 511         if (list_empty (&svc->srv_reply_queue)) {
 512                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 513                 RETURN(0);
 514         }
 515
 516         rs = list_entry (svc->srv_reply_queue.next,
 517                          struct ptlrpc_reply_state, rs_list);
 518
 519         exp = rs->rs_export;
 520         obd = exp->exp_obd;
 521
 522         LASSERT (rs->rs_difficult);
 523         LASSERT (rs->rs_scheduled);
 524
 525         list_del_init (&rs->rs_list);
 526
 527         /* Disengage from notifiers carefully (lock ordering!) */
 528         spin_unlock(&svc->srv_lock);
 529
 530         spin_lock (&obd->obd_uncommitted_replies_lock);
 531         /* Noop if removed already */
 532         list_del_init (&rs->rs_obd_list);
 533         spin_unlock (&obd->obd_uncommitted_replies_lock);
 534
 535         spin_lock (&exp->exp_lock);
 536         /* Noop if removed already */
 537         list_del_init (&rs->rs_exp_list);
 538         spin_unlock (&exp->exp_lock);
 539
 540         spin_lock(&svc->srv_lock);
 541
 542         been_handled = rs->rs_handled;
 543         rs->rs_handled = 1;
 544
 545         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 546         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 547
 548         if (nlocks == 0 && !been_handled) {
 549                 /* If we see this, we should already have seen the warning
 550                  * in mds_steal_ack_locks()  */
 551                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 552                       " o%d NID %s\n",
 553                       rs,
 554                       rs->rs_xid, rs->rs_transno,
 555                       rs->rs_msg.opc,
 556                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 557         }
 558
 559         if ((!been_handled && rs->rs_on_net) ||
 560             nlocks > 0) {
 561                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 562
 563                 if (!been_handled && rs->rs_on_net) {
 564                         PtlMDUnlink(rs->rs_md_h);
 565                         /* Ignore return code; we're racing with
 566                          * completion... */
 567                 }
 568
 569                 while (nlocks-- > 0)
 570                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 571                                          rs->rs_modes[nlocks]);
 572
 573                 spin_lock_irqsave(&svc->srv_lock, flags);
 574         }
 575
 576         rs->rs_scheduled = 0;
 577
 578         if (!rs->rs_on_net) {
 579                 /* Off the net */
 580                 svc->srv_n_difficult_replies--;
 581                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 582
 583                 class_export_put (exp);
 584                 rs->rs_export = NULL;
 585                 lustre_free_reply_state (rs);
 586                 atomic_dec (&svc->srv_outstanding_replies);
 587                 RETURN(1);
 588         }
 589
 590         /* still on the net; callback will schedule */
 591         spin_unlock_irqrestore (&svc->srv_lock, flags);
 592         RETURN(1);
 593 }
 594
 595 #ifndef __KERNEL__
 596 /* FIXME make use of timeout later */
 597 int
 598 liblustre_check_services (void *arg)
 599 {
 600         int  did_something = 0;
 601         int  rc;
 602         struct list_head *tmp, *nxt;
 603         ENTRY;
 604
 605         /* I'm relying on being single threaded, not to have to lock
 606          * ptlrpc_all_services etc */
 607         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 608                 struct ptlrpc_service *svc =
 609                         list_entry (tmp, struct ptlrpc_service, srv_list);
 610
 611                 if (svc->srv_nthreads != 0)     /* I've recursed */
 612                         continue;
 613
 614                 /* service threads can block for bulk, so this limits us
 615                  * (arbitrarily) to recursing 1 stack frame per service.
 616                  * Note that the problem with recursion is that we have to
 617                  * unwind completely before our caller can resume. */
 618
 619                 svc->srv_nthreads++;
 620
 621                 do {
 622                         rc = ptlrpc_server_handle_reply(svc);
 623                         rc |= ptlrpc_server_handle_request(svc);
 624                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 625                         did_something |= rc;
 626                 } while (rc);
 627
 628                 svc->srv_nthreads--;
 629         }
 630
 631         RETURN(did_something);
 632 }
 633
 634 #else /* __KERNEL__ */
 635
 636 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 637 void ptlrpc_daemonize(void)
 638 {
 639         exit_mm(current);
 640         lustre_daemonize_helper();
 641         exit_files(current);
 642         reparent_to_init();
 643 }
 644
 645 static int
 646 ptlrpc_retry_rqbds(void *arg)
 647 {
 648         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 649
 650         svc->srv_rqbd_timeout = 0;
 651         return (-ETIMEDOUT);
 652 }
 653
 654 static int ptlrpc_main(void *arg)
 655 {
 656         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 657         struct ptlrpc_service  *svc = data->svc;
 658         struct ptlrpc_thread   *thread = data->thread;
 659         unsigned long           flags;
 660         ENTRY;
 661
 662         lock_kernel();
 663         ptlrpc_daemonize();
 664
 665         SIGNAL_MASK_LOCK(current, flags);
 666         sigfillset(&current->blocked);
 667         RECALC_SIGPENDING;
 668         SIGNAL_MASK_UNLOCK(current, flags);
 669
 670         LASSERTF(strlen(data->name) < sizeof(current->comm),
 671                  "name %d > len %d\n",
 672                  (int)strlen(data->name), (int)sizeof(current->comm));
 673         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 674         unlock_kernel();
 675
 676         /* Record that the thread is running */
 677         thread->t_flags = SVC_RUNNING;
 678         wake_up(&thread->t_ctl_waitq);
 679
 680         spin_lock_irqsave(&svc->srv_lock, flags);
 681         svc->srv_nthreads++;
 682         spin_unlock_irqrestore(&svc->srv_lock, flags);
 683
 684         /* XXX maintain a list of all managed devices: insert here */
 685
 686         while ((thread->t_flags & SVC_STOPPING) == 0 ||
 687                svc->srv_n_difficult_replies != 0) {
 688                 /* Don't exit while there are replies to be handled */
 689                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
 690                                                      ptlrpc_retry_rqbds, svc);
 691
 692                 l_wait_event_exclusive (svc->srv_waitq,
 693                               ((thread->t_flags & SVC_STOPPING) != 0 &&
 694                                svc->srv_n_difficult_replies == 0) ||
 695                               (!list_empty(&svc->srv_idle_rqbds) &&
 696                                svc->srv_rqbd_timeout == 0) ||
 697                               !list_empty (&svc->srv_reply_queue) ||
 698                               (!list_empty (&svc->srv_request_queue) &&
 699                                (svc->srv_n_difficult_replies == 0 ||
 700                                 svc->srv_n_active_reqs <
 701                                 (svc->srv_nthreads - 1))),
 702                               &lwi);
 703
 704                 if (!list_empty (&svc->srv_reply_queue))
 705                         ptlrpc_server_handle_reply (svc);
 706
 707                 /* only handle requests if there are no difficult replies
 708                  * outstanding, or I'm not the last thread handling
 709                  * requests */
 710                 if (!list_empty (&svc->srv_request_queue) &&
 711                     (svc->srv_n_difficult_replies == 0 ||
 712                      svc->srv_n_active_reqs < (svc->srv_nthreads - 1)))
 713                         ptlrpc_server_handle_request (svc);
 714
 715                 if (!list_empty(&svc->srv_idle_rqbds) &&
 716                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
 717                         /* I just failed to repost request buffers.  Wait
 718                          * for a timeout (unless something else happens)
 719                          * before I try again */
 720                         svc->srv_rqbd_timeout = HZ/10;
 721                 }
 722         }
 723
 724         spin_lock_irqsave(&svc->srv_lock, flags);
 725
 726         svc->srv_nthreads--;                    /* must know immediately */
 727         thread->t_flags = SVC_STOPPED;
 728         wake_up(&thread->t_ctl_waitq);
 729
 730         spin_unlock_irqrestore(&svc->srv_lock, flags);
 731
 732         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 733         return 0;
 734 }
 735
 736 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 737                                struct ptlrpc_thread *thread)
 738 {
 739         struct l_wait_info lwi = { 0 };
 740         unsigned long      flags;
 741
 742         spin_lock_irqsave(&svc->srv_lock, flags);
 743         thread->t_flags = SVC_STOPPING;
 744         spin_unlock_irqrestore(&svc->srv_lock, flags);
 745
 746         wake_up_all(&svc->srv_waitq);
 747         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 748                      &lwi);
 749
 750         spin_lock_irqsave(&svc->srv_lock, flags);
 751         list_del(&thread->t_link);
 752         spin_unlock_irqrestore(&svc->srv_lock, flags);
 753
 754         OBD_FREE(thread, sizeof(*thread));
 755 }
 756
 757 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 758 {
 759         unsigned long flags;
 760         struct ptlrpc_thread *thread;
 761
 762         spin_lock_irqsave(&svc->srv_lock, flags);
 763         while (!list_empty(&svc->srv_threads)) {
 764                 thread = list_entry(svc->srv_threads.next,
 765                                     struct ptlrpc_thread, t_link);
 766
 767                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 768                 ptlrpc_stop_thread(svc, thread);
 769                 spin_lock_irqsave(&svc->srv_lock, flags);
 770         }
 771
 772         spin_unlock_irqrestore(&svc->srv_lock, flags);
 773 }
 774
 775 /* @base_name should be 12 characters or less - 3 will be added on */
 776 int ptlrpc_start_n_threads(struct obd_device *dev, struct ptlrpc_service *svc,
 777                            int num_threads, char *base_name)
 778 {
 779         int i, rc = 0;
 780         ENTRY;
 781
 782         for (i = 0; i < num_threads; i++) {
 783                 char name[32];
 784                 sprintf(name, "%s_%02d", base_name, i);
 785                 rc = ptlrpc_start_thread(dev, svc, name);
 786                 if (rc) {
 787                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
 788                                i, rc);
 789                         ptlrpc_stop_all_threads(svc);
 790                 }
 791         }
 792         RETURN(rc);
 793 }
 794
 795 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 796                         char *name)
 797 {
 798         struct l_wait_info lwi = { 0 };
 799         struct ptlrpc_svc_data d;
 800         struct ptlrpc_thread *thread;
 801         unsigned long flags;
 802         int rc;
 803         ENTRY;
 804
 805         OBD_ALLOC(thread, sizeof(*thread));
 806         if (thread == NULL)
 807                 RETURN(-ENOMEM);
 808         init_waitqueue_head(&thread->t_ctl_waitq);
 809
 810         d.dev = dev;
 811         d.svc = svc;
 812         d.name = name;
 813         d.thread = thread;
 814
 815         spin_lock_irqsave(&svc->srv_lock, flags);
 816         list_add(&thread->t_link, &svc->srv_threads);
 817         spin_unlock_irqrestore(&svc->srv_lock, flags);
 818
 819         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 820          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 821          */
 822         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 823         if (rc < 0) {
 824                 CERROR("cannot start thread: %d\n", rc);
 825                 OBD_FREE(thread, sizeof(*thread));
 826                 RETURN(rc);
 827         }
 828         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 829
 830         RETURN(0);
 831 }
 832 #endif
 833
 834 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 835 {
 836         int                   i;
 837         int                   rc;
 838         unsigned long         flags;
 839         struct ptlrpc_srv_ni *srv_ni;
 840         struct l_wait_info    lwi;
 841         struct list_head     *tmp;
 842
 843         LASSERT(list_empty(&service->srv_threads));
 844
 845         spin_lock (&ptlrpc_all_services_lock);
 846         list_del_init (&service->srv_list);
 847         spin_unlock (&ptlrpc_all_services_lock);
 848
 849         ptlrpc_lprocfs_unregister_service(service);
 850
 851         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 852                 srv_ni = &service->srv_interfaces[i];
 853                 CDEBUG(D_NET, "%s: tearing down interface %s\n",
 854                        service->srv_name, srv_ni->sni_ni->pni_name);
 855
 856                 /* Unlink all the request buffers.  This forces a 'final'
 857                  * event with its 'unlink' flag set for each posted rqbd */
 858                 list_for_each(tmp, &srv_ni->sni_active_rqbds) {
 859                         struct ptlrpc_request_buffer_desc *rqbd =
 860                                 list_entry(tmp, struct ptlrpc_request_buffer_desc,
 861                                            rqbd_list);
 862
 863                         rc = PtlMDUnlink(rqbd->rqbd_md_h);
 864                         LASSERT (rc == PTL_OK || rc == PTL_MD_INVALID);
 865                 }
 866
 867                 /* Wait for the network to release any buffers it's
 868                  * currently filling */
 869                 for (;;) {
 870                         spin_lock_irqsave(&service->srv_lock, flags);
 871                         rc = srv_ni->sni_nrqbd_receiving;
 872                         spin_unlock_irqrestore(&service->srv_lock, flags);
 873
 874                         if (rc == 0)
 875                                 break;
 876
 877                         /* Network access will complete in finite time but
 878                          * the HUGE timeout lets us CWARN for visibility of
 879                          * sluggish NALs */
 880                         lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
 881                         rc = l_wait_event(service->srv_waitq,
 882                                           srv_ni->sni_nrqbd_receiving == 0,
 883                                           &lwi);
 884                         if (rc == -ETIMEDOUT)
 885                                 CWARN("Waiting for request buffers on "
 886                                       "service %s on interface %s ",
 887                                       service->srv_name, srv_ni->sni_ni->pni_name);
 888                 }
 889
 890                 /* schedule all outstanding replies to terminate them */
 891                 spin_lock_irqsave(&service->srv_lock, flags);
 892                 while (!list_empty(&srv_ni->sni_active_replies)) {
 893                         struct ptlrpc_reply_state *rs =
 894                                 list_entry(srv_ni->sni_active_replies.next,
 895                                            struct ptlrpc_reply_state,
 896                                            rs_list);
 897                         ptlrpc_schedule_difficult_reply(rs);
 898                 }
 899                 spin_unlock_irqrestore(&service->srv_lock, flags);
 900         }
 901
 902         /* purge the request queue.  NB No new replies (rqbds all unlinked)
 903          * and no service threads, so I'm the only thread noodling the
 904          * request queue now */
 905         while (!list_empty(&service->srv_request_queue)) {
 906                 struct ptlrpc_request *req =
 907                         list_entry(service->srv_request_queue.next,
 908                                    struct ptlrpc_request,
 909                                    rq_list);
 910
 911                 list_del(&req->rq_list);
 912                 service->srv_n_queued_reqs--;
 913                 service->srv_n_active_reqs++;
 914
 915                 ptlrpc_server_free_request(service, req);
 916         }
 917         LASSERT(service->srv_n_queued_reqs == 0);
 918         LASSERT(service->srv_n_active_reqs == 0);
 919
 920         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 921                 srv_ni = &service->srv_interfaces[i];
 922                 LASSERT(list_empty(&srv_ni->sni_active_rqbds));
 923         }
 924
 925         /* Now free all the request buffers since nothing references them
 926          * any more... */
 927         while (!list_empty(&service->srv_idle_rqbds)) {
 928                 struct ptlrpc_request_buffer_desc *rqbd =
 929                         list_entry(service->srv_idle_rqbds.next,
 930                                    struct ptlrpc_request_buffer_desc,
 931                                    rqbd_list);
 932
 933                 ptlrpc_free_rqbd(rqbd);
 934         }
 935
 936         /* wait for all outstanding replies to complete (they were
 937          * scheduled having been flagged to abort above) */
 938         while (atomic_read(&service->srv_outstanding_replies) != 0) {
 939                 struct l_wait_info lwi = LWI_TIMEOUT(10 * HZ, NULL, NULL);
 940
 941                 rc = l_wait_event(service->srv_waitq,
 942                                   !list_empty(&service->srv_reply_queue), &lwi);
 943                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 944
 945                 if (rc == 0) {
 946                         ptlrpc_server_handle_reply(service);
 947                         continue;
 948                 }
 949                 CWARN("Unexpectedly long timeout %p\n", service);
 950         }
 951
 952         OBD_FREE(service,
 953                  offsetof(struct ptlrpc_service,
 954                           srv_interfaces[ptlrpc_ninterfaces]));
 955         return 0;
 956 }