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         ENTRY;
 364
 365         spin_lock_irqsave (&svc->srv_lock, flags);
 366         if (list_empty (&svc->srv_request_queue) ||
 367             (svc->srv_n_difficult_replies != 0 &&
 368              svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
 369                 /* If all the other threads are handling requests, I must
 370                  * remain free to handle any 'difficult' reply that might
 371                  * block them */
 372                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 373                 RETURN(0);
 374         }
 375
 376         request = list_entry (svc->srv_request_queue.next,
 377                               struct ptlrpc_request, rq_list);
 378         list_del_init (&request->rq_list);
 379         svc->srv_n_queued_reqs--;
 380         svc->srv_n_active_reqs++;
 381
 382         spin_unlock_irqrestore (&svc->srv_lock, flags);
 383
 384         do_gettimeofday(&work_start);
 385         timediff = timeval_sub(&work_start, &request->rq_arrival_time);
 386         if (svc->srv_stats != NULL) {
 387                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 388                                     timediff);
 389                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 390                                     svc->srv_n_queued_reqs);
 391                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 392                                     svc->srv_n_active_reqs);
 393         }
 394
 395 #if SWAB_PARANOIA
 396         /* Clear request swab mask; this is a new request */
 397         request->rq_req_swab_mask = 0;
 398 #endif
 399         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 400         if (rc != 0) {
 401                 CERROR ("error unpacking request: ptl %d from "LPX64
 402                         " xid "LPU64"\n", svc->srv_req_portal,
 403                        request->rq_peer.peer_nid, request->rq_xid);
 404                 goto out;
 405         }
 406
 407         rc = -EINVAL;
 408         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 409                 CERROR("wrong packet type received (type=%u) from "
 410                        LPX64"\n", request->rq_reqmsg->type,
 411                        request->rq_peer.peer_nid);
 412                 goto out;
 413         }
 414
 415         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 416
 417         /* Discard requests queued for longer than my timeout.  If the
 418          * client's timeout is similar to mine, she'll be timing out this
 419          * REQ anyway (bug 1502) */
 420         if (timediff / 1000000 > (long)obd_timeout) {
 421                 CERROR("Dropping timed-out request from "LPX64
 422                        ": %ld seconds old\n",
 423                        request->rq_peer.peer_nid, timediff / 1000000);
 424                 goto out;
 425         }
 426
 427         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 428
 429         if (request->rq_export) {
 430                 if (request->rq_reqmsg->conn_cnt <
 431                     request->rq_export->exp_conn_cnt) {
 432                         DEBUG_REQ(D_ERROR, request,
 433                                   "DROPPING req from old connection %d < %d",
 434                                   request->rq_reqmsg->conn_cnt,
 435                                   request->rq_export->exp_conn_cnt);
 436                         goto put_conn;
 437                 }
 438
 439                 request->rq_export->exp_last_request_time =
 440                         LTIME_S(CURRENT_TIME);
 441         }
 442
 443         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 444                "%s:%s+%d:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 445                (request->rq_export ?
 446                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 447                (request->rq_export ?
 448                 atomic_read(&request->rq_export->exp_refcount) : -99),
 449                request->rq_reqmsg->status, request->rq_xid,
 450                request->rq_peer.peer_ni->pni_name,
 451                request->rq_peer.peer_nid,
 452                request->rq_reqmsg->opc);
 453
 454         rc = svc->srv_handler(request);
 455         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 456                "%s:%s+%d:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 457                (request->rq_export ?
 458                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 459                (request->rq_export ?
 460                 atomic_read(&request->rq_export->exp_refcount) : -99),
 461                request->rq_reqmsg->status, request->rq_xid,
 462                request->rq_peer.peer_ni->pni_name,
 463                request->rq_peer.peer_nid,
 464                request->rq_reqmsg->opc);
 465
 466 put_conn:
 467         if (request->rq_export != NULL)
 468                 class_export_put(request->rq_export);
 469
 470  out:
 471         do_gettimeofday(&work_end);
 472
 473         timediff = timeval_sub(&work_end, &work_start);
 474
 475         CDEBUG((timediff / 1000000 > (long)obd_timeout) ? D_ERROR : D_HA,
 476                "request "LPU64" opc %u from NID "LPX64" processed in %ldus "
 477                "(%ldus total)\n", request->rq_xid, request->rq_reqmsg->opc,
 478                request->rq_peer.peer_nid,
 479                timediff, timeval_sub(&work_end, &request->rq_arrival_time));
 480
 481         if (svc->srv_stats != NULL) {
 482                 int opc = opcode_offset(request->rq_reqmsg->opc);
 483                 if (opc > 0) {
 484                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 485                         lprocfs_counter_add(svc->srv_stats,
 486                                             opc + PTLRPC_LAST_CNTR,
 487                                             timediff);
 488                 }
 489         }
 490
 491         ptlrpc_server_free_request(svc, request);
 492
 493         RETURN(1);
 494 }
 495
 496 static int
 497 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 498 {
 499         struct ptlrpc_reply_state *rs;
 500         unsigned long              flags;
 501         struct obd_export         *exp;
 502         struct obd_device         *obd;
 503         int                        nlocks;
 504         int                        been_handled;
 505         ENTRY;
 506
 507         spin_lock_irqsave (&svc->srv_lock, flags);
 508         if (list_empty (&svc->srv_reply_queue)) {
 509                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 510                 RETURN(0);
 511         }
 512
 513         rs = list_entry (svc->srv_reply_queue.next,
 514                          struct ptlrpc_reply_state, rs_list);
 515
 516         exp = rs->rs_export;
 517         obd = exp->exp_obd;
 518
 519         LASSERT (rs->rs_difficult);
 520         LASSERT (rs->rs_scheduled);
 521
 522         list_del_init (&rs->rs_list);
 523
 524         /* Disengage from notifiers carefully (lock ordering!) */
 525         spin_unlock(&svc->srv_lock);
 526
 527         spin_lock (&obd->obd_uncommitted_replies_lock);
 528         /* Noop if removed already */
 529         list_del_init (&rs->rs_obd_list);
 530         spin_unlock (&obd->obd_uncommitted_replies_lock);
 531
 532         spin_lock (&exp->exp_lock);
 533         /* Noop if removed already */
 534         list_del_init (&rs->rs_exp_list);
 535         spin_unlock (&exp->exp_lock);
 536
 537         spin_lock(&svc->srv_lock);
 538
 539         been_handled = rs->rs_handled;
 540         rs->rs_handled = 1;
 541
 542         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 543         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 544
 545         if (nlocks == 0 && !been_handled) {
 546                 /* If we see this, we should already have seen the warning
 547                  * in mds_steal_ack_locks()  */
 548                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 549                       " o%d NID"LPX64"\n",
 550                       rs,
 551                       rs->rs_xid, rs->rs_transno,
 552                       rs->rs_msg.opc, exp->exp_connection->c_peer.peer_nid);
 553         }
 554
 555         if ((!been_handled && rs->rs_on_net) ||
 556             nlocks > 0) {
 557                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 558
 559                 if (!been_handled && rs->rs_on_net) {
 560                         PtlMDUnlink(rs->rs_md_h);
 561                         /* Ignore return code; we're racing with
 562                          * completion... */
 563                 }
 564
 565                 while (nlocks-- > 0)
 566                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 567                                          rs->rs_modes[nlocks]);
 568
 569                 spin_lock_irqsave(&svc->srv_lock, flags);
 570         }
 571
 572         rs->rs_scheduled = 0;
 573
 574         if (!rs->rs_on_net) {
 575                 /* Off the net */
 576                 svc->srv_n_difficult_replies--;
 577                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 578
 579                 class_export_put (exp);
 580                 rs->rs_export = NULL;
 581                 lustre_free_reply_state (rs);
 582                 atomic_dec (&svc->srv_outstanding_replies);
 583                 RETURN(1);
 584         }
 585
 586         /* still on the net; callback will schedule */
 587         spin_unlock_irqrestore (&svc->srv_lock, flags);
 588         RETURN(1);
 589 }
 590
 591 #ifndef __KERNEL__
 592 /* FIXME make use of timeout later */
 593 int
 594 liblustre_check_services (void *arg)
 595 {
 596         int  did_something = 0;
 597         int  rc;
 598         struct list_head *tmp, *nxt;
 599         ENTRY;
 600
 601         /* I'm relying on being single threaded, not to have to lock
 602          * ptlrpc_all_services etc */
 603         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 604                 struct ptlrpc_service *svc =
 605                         list_entry (tmp, struct ptlrpc_service, srv_list);
 606
 607                 if (svc->srv_nthreads != 0)     /* I've recursed */
 608                         continue;
 609
 610                 /* service threads can block for bulk, so this limits us
 611                  * (arbitrarily) to recursing 1 stack frame per service.
 612                  * Note that the problem with recursion is that we have to
 613                  * unwind completely before our caller can resume. */
 614
 615                 svc->srv_nthreads++;
 616
 617                 do {
 618                         rc = ptlrpc_server_handle_reply(svc);
 619                         rc |= ptlrpc_server_handle_request(svc);
 620                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 621                         did_something |= rc;
 622                 } while (rc);
 623
 624                 svc->srv_nthreads--;
 625         }
 626
 627         RETURN(did_something);
 628 }
 629
 630 #else /* __KERNEL__ */
 631
 632 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 633 void ptlrpc_daemonize(void)
 634 {
 635         exit_mm(current);
 636         lustre_daemonize_helper();
 637         exit_files(current);
 638         reparent_to_init();
 639 }
 640
 641 static int
 642 ptlrpc_retry_rqbds(void *arg)
 643 {
 644         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 645
 646         svc->srv_rqbd_timeout = 0;
 647         return (-ETIMEDOUT);
 648 }
 649
 650 static int ptlrpc_main(void *arg)
 651 {
 652         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 653         struct ptlrpc_service  *svc = data->svc;
 654         struct ptlrpc_thread   *thread = data->thread;
 655         unsigned long           flags;
 656         ENTRY;
 657
 658         lock_kernel();
 659         ptlrpc_daemonize();
 660
 661         SIGNAL_MASK_LOCK(current, flags);
 662         sigfillset(&current->blocked);
 663         RECALC_SIGPENDING;
 664         SIGNAL_MASK_UNLOCK(current, flags);
 665
 666         THREAD_NAME(current->comm, "%s", data->name);
 667         unlock_kernel();
 668
 669         /* Record that the thread is running */
 670         thread->t_flags = SVC_RUNNING;
 671         wake_up(&thread->t_ctl_waitq);
 672
 673         spin_lock_irqsave(&svc->srv_lock, flags);
 674         svc->srv_nthreads++;
 675         spin_unlock_irqrestore(&svc->srv_lock, flags);
 676
 677         /* XXX maintain a list of all managed devices: insert here */
 678
 679         while ((thread->t_flags & SVC_STOPPING) == 0 ||
 680                svc->srv_n_difficult_replies != 0) {
 681                 /* Don't exit while there are replies to be handled */
 682                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
 683                                                      ptlrpc_retry_rqbds, svc);
 684
 685                 l_wait_event_exclusive (svc->srv_waitq,
 686                               ((thread->t_flags & SVC_STOPPING) != 0 &&
 687                                svc->srv_n_difficult_replies == 0) ||
 688                               (!list_empty(&svc->srv_idle_rqbds) &&
 689                                svc->srv_rqbd_timeout == 0) ||
 690                               !list_empty (&svc->srv_reply_queue) ||
 691                               (!list_empty (&svc->srv_request_queue) &&
 692                                (svc->srv_n_difficult_replies == 0 ||
 693                                 svc->srv_n_active_reqs <
 694                                 (svc->srv_nthreads - 1))),
 695                               &lwi);
 696
 697                 if (!list_empty (&svc->srv_reply_queue))
 698                         ptlrpc_server_handle_reply (svc);
 699
 700                 /* only handle requests if there are no difficult replies
 701                  * outstanding, or I'm not the last thread handling
 702                  * requests */
 703                 if (!list_empty (&svc->srv_request_queue) &&
 704                     (svc->srv_n_difficult_replies == 0 ||
 705                      svc->srv_n_active_reqs < (svc->srv_nthreads - 1)))
 706                         ptlrpc_server_handle_request (svc);
 707
 708                 if (!list_empty(&svc->srv_idle_rqbds) &&
 709                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
 710                         /* I just failed to repost request buffers.  Wait
 711                          * for a timeout (unless something else happens)
 712                          * before I try again */
 713                         svc->srv_rqbd_timeout = HZ/10;
 714                 }
 715         }
 716
 717         spin_lock_irqsave(&svc->srv_lock, flags);
 718
 719         svc->srv_nthreads--;                    /* must know immediately */
 720         thread->t_flags = SVC_STOPPED;
 721         wake_up(&thread->t_ctl_waitq);
 722
 723         spin_unlock_irqrestore(&svc->srv_lock, flags);
 724
 725         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 726         return 0;
 727 }
 728
 729 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 730                                struct ptlrpc_thread *thread)
 731 {
 732         struct l_wait_info lwi = { 0 };
 733         unsigned long      flags;
 734
 735         spin_lock_irqsave(&svc->srv_lock, flags);
 736         thread->t_flags = SVC_STOPPING;
 737         spin_unlock_irqrestore(&svc->srv_lock, flags);
 738
 739         wake_up_all(&svc->srv_waitq);
 740         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 741                      &lwi);
 742
 743         spin_lock_irqsave(&svc->srv_lock, flags);
 744         list_del(&thread->t_link);
 745         spin_unlock_irqrestore(&svc->srv_lock, flags);
 746
 747         OBD_FREE(thread, sizeof(*thread));
 748 }
 749
 750 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 751 {
 752         unsigned long flags;
 753         struct ptlrpc_thread *thread;
 754
 755         spin_lock_irqsave(&svc->srv_lock, flags);
 756         while (!list_empty(&svc->srv_threads)) {
 757                 thread = list_entry(svc->srv_threads.next,
 758                                     struct ptlrpc_thread, t_link);
 759
 760                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 761                 ptlrpc_stop_thread(svc, thread);
 762                 spin_lock_irqsave(&svc->srv_lock, flags);
 763         }
 764
 765         spin_unlock_irqrestore(&svc->srv_lock, flags);
 766 }
 767
 768 int ptlrpc_start_n_threads(struct obd_device *dev, struct ptlrpc_service *svc,
 769                            int num_threads, char *base_name)
 770 {
 771         int i, rc = 0;
 772         ENTRY;
 773
 774         for (i = 0; i < num_threads; i++) {
 775                 char name[32];
 776                 sprintf(name, "%s_%02d", base_name, i);
 777                 rc = ptlrpc_start_thread(dev, svc, name);
 778                 if (rc) {
 779                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
 780                                i, rc);
 781                         ptlrpc_stop_all_threads(svc);
 782                 }
 783         }
 784         RETURN(rc);
 785 }
 786
 787 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 788                         char *name)
 789 {
 790         struct l_wait_info lwi = { 0 };
 791         struct ptlrpc_svc_data d;
 792         struct ptlrpc_thread *thread;
 793         unsigned long flags;
 794         int rc;
 795         ENTRY;
 796
 797         OBD_ALLOC(thread, sizeof(*thread));
 798         if (thread == NULL)
 799                 RETURN(-ENOMEM);
 800         init_waitqueue_head(&thread->t_ctl_waitq);
 801
 802         d.dev = dev;
 803         d.svc = svc;
 804         d.name = name;
 805         d.thread = thread;
 806
 807         spin_lock_irqsave(&svc->srv_lock, flags);
 808         list_add(&thread->t_link, &svc->srv_threads);
 809         spin_unlock_irqrestore(&svc->srv_lock, flags);
 810
 811         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 812          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 813          */
 814         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 815         if (rc < 0) {
 816                 CERROR("cannot start thread: %d\n", rc);
 817                 OBD_FREE(thread, sizeof(*thread));
 818                 RETURN(rc);
 819         }
 820         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 821
 822         RETURN(0);
 823 }
 824 #endif
 825
 826 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 827 {
 828         int                   i;
 829         int                   rc;
 830         unsigned long         flags;
 831         struct ptlrpc_srv_ni *srv_ni;
 832         struct l_wait_info    lwi;
 833         struct list_head     *tmp;
 834
 835         LASSERT(list_empty(&service->srv_threads));
 836
 837         spin_lock (&ptlrpc_all_services_lock);
 838         list_del_init (&service->srv_list);
 839         spin_unlock (&ptlrpc_all_services_lock);
 840
 841         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 842                 srv_ni = &service->srv_interfaces[i];
 843                 CDEBUG(D_NET, "%s: tearing down interface %s\n",
 844                        service->srv_name, srv_ni->sni_ni->pni_name);
 845
 846                 /* Unlink all the request buffers.  This forces a 'final'
 847                  * event with its 'unlink' flag set for each posted rqbd */
 848                 list_for_each(tmp, &srv_ni->sni_active_rqbds) {
 849                         struct ptlrpc_request_buffer_desc *rqbd =
 850                                 list_entry(tmp, struct ptlrpc_request_buffer_desc,
 851                                            rqbd_list);
 852
 853                         rc = PtlMDUnlink(rqbd->rqbd_md_h);
 854                         LASSERT (rc == PTL_OK || rc == PTL_MD_INVALID);
 855                 }
 856
 857                 /* Wait for the network to release any buffers it's
 858                  * currently filling */
 859                 for (;;) {
 860                         spin_lock_irqsave(&service->srv_lock, flags);
 861                         rc = srv_ni->sni_nrqbd_receiving;
 862                         spin_unlock_irqrestore(&service->srv_lock, flags);
 863
 864                         if (rc == 0)
 865                                 break;
 866
 867                         /* Network access will complete in finite time but
 868                          * the HUGE timeout lets us CWARN for visibility of
 869                          * sluggish NALs */
 870                         lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
 871                         rc = l_wait_event(service->srv_waitq,
 872                                           srv_ni->sni_nrqbd_receiving == 0,
 873                                           &lwi);
 874                         if (rc == -ETIMEDOUT)
 875                                 CWARN("Waiting for request buffers on "
 876                                       "service %s on interface %s ",
 877                                       service->srv_name, srv_ni->sni_ni->pni_name);
 878                 }
 879
 880                 /* schedule all outstanding replies to terminate them */
 881                 spin_lock_irqsave(&service->srv_lock, flags);
 882                 while (!list_empty(&srv_ni->sni_active_replies)) {
 883                         struct ptlrpc_reply_state *rs =
 884                                 list_entry(srv_ni->sni_active_replies.next,
 885                                            struct ptlrpc_reply_state,
 886                                            rs_list);
 887                         ptlrpc_schedule_difficult_reply(rs);
 888                 }
 889                 spin_unlock_irqrestore(&service->srv_lock, flags);
 890         }
 891
 892         /* purge the request queue.  NB No new replies (rqbds all unlinked)
 893          * and no service threads, so I'm the only thread noodling the
 894          * request queue now */
 895         while (!list_empty(&service->srv_request_queue)) {
 896                 struct ptlrpc_request *req =
 897                         list_entry(service->srv_request_queue.next,
 898                                    struct ptlrpc_request,
 899                                    rq_list);
 900
 901                 list_del(&req->rq_list);
 902                 service->srv_n_queued_reqs--;
 903                 service->srv_n_active_reqs++;
 904
 905                 ptlrpc_server_free_request(service, req);
 906         }
 907         LASSERT(service->srv_n_queued_reqs == 0);
 908         LASSERT(service->srv_n_active_reqs == 0);
 909
 910         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 911                 srv_ni = &service->srv_interfaces[i];
 912                 LASSERT(list_empty(&srv_ni->sni_active_rqbds));
 913         }
 914
 915         /* Now free all the request buffers since nothing references them
 916          * any more... */
 917         while (!list_empty(&service->srv_idle_rqbds)) {
 918                 struct ptlrpc_request_buffer_desc *rqbd =
 919                         list_entry(service->srv_idle_rqbds.next,
 920                                    struct ptlrpc_request_buffer_desc,
 921                                    rqbd_list);
 922
 923                 ptlrpc_free_rqbd(rqbd);
 924         }
 925
 926         /* wait for all outstanding replies to complete (they were
 927          * scheduled having been flagged to abort above) */
 928         while (atomic_read(&service->srv_outstanding_replies) != 0) {
 929                 struct l_wait_info lwi = LWI_TIMEOUT(10 * HZ, NULL, NULL);
 930
 931                 rc = l_wait_event(service->srv_waitq,
 932                                   !list_empty(&service->srv_reply_queue), &lwi);
 933                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 934
 935                 if (rc == 0) {
 936                         ptlrpc_server_handle_reply(service);
 937                         continue;
 938                 }
 939                 CWARN("Unexpectedly long timeout %p\n", service);
 940         }
 941
 942         ptlrpc_lprocfs_unregister_service(service);
 943
 944         OBD_FREE(service,
 945                  offsetof(struct ptlrpc_service,
 946                           srv_interfaces[ptlrpc_ninterfaces]));
 947         return 0;
 948 }