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         if (!lock->cookie)
 127                 return;
 128
 129         LASSERT (rs != NULL);
 130         LASSERT (rs->rs_nlocks < RS_MAX_LOCKS);
 131
 132         idx = rs->rs_nlocks++;
 133         rs->rs_locks[idx] = *lock;
 134         rs->rs_modes[idx] = mode;
 135         rs->rs_difficult = 1;
 136 }
 137
 138 void
 139 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 140 {
 141         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 142
 143 #ifdef CONFIG_SMP
 144         LASSERT (spin_is_locked (&svc->srv_lock));
 145 #endif
 146         LASSERT (rs->rs_difficult);
 147         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 148
 149         if (rs->rs_scheduled)                   /* being set up or already notified */
 150                 return;
 151
 152         rs->rs_scheduled = 1;
 153         list_del (&rs->rs_list);
 154         list_add (&rs->rs_list, &svc->srv_reply_queue);
 155         wake_up (&svc->srv_waitq);
 156 }
 157
 158 void
 159 ptlrpc_commit_replies (struct obd_device *obd)
 160 {
 161         struct list_head   *tmp;
 162         struct list_head   *nxt;
 163         unsigned long       flags;
 164
 165         /* Find any replies that have been committed and get their service
 166          * to attend to complete them. */
 167
 168         /* CAVEAT EMPTOR: spinlock ordering!!! */
 169         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
 170
 171         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 172                 struct ptlrpc_reply_state *rs =
 173                         list_entry (tmp, struct ptlrpc_reply_state, rs_obd_list);
 174
 175                 LASSERT (rs->rs_difficult);
 176
 177                 if (rs->rs_transno <= obd->obd_last_committed) {
 178                         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 179
 180                         spin_lock (&svc->srv_lock);
 181                         list_del_init (&rs->rs_obd_list);
 182                         ptlrpc_schedule_difficult_reply (rs);
 183                         spin_unlock (&svc->srv_lock);
 184                 }
 185         }
 186
 187         spin_unlock_irqrestore (&obd->obd_uncommitted_replies_lock, flags);
 188 }
 189
 190 static long
 191 timeval_sub(struct timeval *large, struct timeval *small)
 192 {
 193         return (large->tv_sec - small->tv_sec) * 1000000 +
 194                 (large->tv_usec - small->tv_usec);
 195 }
 196
 197 static int
 198 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 199 {
 200         struct ptlrpc_srv_ni              *srv_ni;
 201         struct ptlrpc_request_buffer_desc *rqbd;
 202         unsigned long                      flags;
 203         int                                rc;
 204
 205         spin_lock_irqsave(&svc->srv_lock, flags);
 206         if (list_empty (&svc->srv_idle_rqbds)) {
 207                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 208                 return (0);
 209         }
 210
 211         rqbd = list_entry(svc->srv_idle_rqbds.next,
 212                           struct ptlrpc_request_buffer_desc,
 213                           rqbd_list);
 214         list_del (&rqbd->rqbd_list);
 215
 216         /* assume we will post successfully */
 217         srv_ni = rqbd->rqbd_srv_ni;
 218         srv_ni->sni_nrqbd_receiving++;
 219         list_add (&rqbd->rqbd_list, &srv_ni->sni_active_rqbds);
 220
 221         spin_unlock_irqrestore(&svc->srv_lock, flags);
 222
 223         rc = ptlrpc_register_rqbd(rqbd);
 224         if (rc == 0)
 225                 return (1);
 226
 227         spin_lock_irqsave(&svc->srv_lock, flags);
 228
 229         srv_ni->sni_nrqbd_receiving--;
 230         list_del(&rqbd->rqbd_list);
 231         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 232
 233         if (srv_ni->sni_nrqbd_receiving == 0) {
 234                 /* This service is off-air on this interface because all
 235                  * its request buffers are busy.  Portals will have started
 236                  * dropping incoming requests until more buffers get
 237                  * posted */
 238                 CERROR("All %s %s request buffers busy\n",
 239                        svc->srv_name, srv_ni->sni_ni->pni_name);
 240         }
 241
 242         spin_unlock_irqrestore (&svc->srv_lock, flags);
 243
 244         return (-1);
 245 }
 246
 247 struct ptlrpc_service *
 248 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size,
 249                 int req_portal, int rep_portal,
 250                 svc_handler_t handler, char *name,
 251                 struct proc_dir_entry *proc_entry)
 252 {
 253         int                                i;
 254         int                                j;
 255         int                                ssize;
 256         struct ptlrpc_service             *service;
 257         struct ptlrpc_srv_ni              *srv_ni;
 258         struct ptlrpc_request_buffer_desc *rqbd;
 259         ENTRY;
 260
 261         LASSERT (ptlrpc_ninterfaces > 0);
 262         LASSERT (nbufs > 0);
 263         LASSERT (bufsize >= max_req_size);
 264
 265         ssize = offsetof (struct ptlrpc_service,
 266                           srv_interfaces[ptlrpc_ninterfaces]);
 267         OBD_ALLOC(service, ssize);
 268         if (service == NULL)
 269                 RETURN(NULL);
 270
 271         service->srv_name = name;
 272         spin_lock_init(&service->srv_lock);
 273         INIT_LIST_HEAD(&service->srv_threads);
 274         init_waitqueue_head(&service->srv_waitq);
 275
 276         service->srv_max_req_size = max_req_size;
 277         service->srv_buf_size = bufsize;
 278         service->srv_rep_portal = rep_portal;
 279         service->srv_req_portal = req_portal;
 280         service->srv_handler = handler;
 281
 282         INIT_LIST_HEAD(&service->srv_request_queue);
 283         INIT_LIST_HEAD(&service->srv_idle_rqbds);
 284         INIT_LIST_HEAD(&service->srv_reply_queue);
 285
 286         /* First initialise enough for early teardown */
 287         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 288                 srv_ni = &service->srv_interfaces[i];
 289
 290                 srv_ni->sni_service = service;
 291                 srv_ni->sni_ni = &ptlrpc_interfaces[i];
 292                 INIT_LIST_HEAD(&srv_ni->sni_active_rqbds);
 293                 INIT_LIST_HEAD(&srv_ni->sni_active_replies);
 294         }
 295
 296         spin_lock (&ptlrpc_all_services_lock);
 297         list_add (&service->srv_list, &ptlrpc_all_services);
 298         spin_unlock (&ptlrpc_all_services_lock);
 299
 300         /* Now allocate the request buffers, assuming all interfaces require
 301          * the same number. */
 302         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 303                 srv_ni = &service->srv_interfaces[i];
 304                 CDEBUG (D_NET, "%s: initialising interface %s\n", name,
 305                         srv_ni->sni_ni->pni_name);
 306
 307                 for (j = 0; j < nbufs; j++) {
 308                         rqbd = ptlrpc_alloc_rqbd (srv_ni);
 309
 310                         if (rqbd == NULL) {
 311                                 CERROR ("%s.%d: Can't allocate request %d "
 312                                         "on %s\n", name, i, j,
 313                                         srv_ni->sni_ni->pni_name);
 314                                 GOTO(failed, NULL);
 315                         }
 316
 317                         /* We shouldn't be under memory pressure at
 318                          * startup, so fail if we can't post all our
 319                          * buffers at this time. */
 320                         if (ptlrpc_server_post_idle_rqbds(service) <= 0)
 321                                 GOTO(failed, NULL);
 322                 }
 323         }
 324
 325         if (proc_entry != NULL)
 326                 ptlrpc_lprocfs_register_service(proc_entry, service);
 327
 328         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 329                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 330
 331         RETURN(service);
 332 failed:
 333         ptlrpc_unregister_service(service);
 334         return NULL;
 335 }
 336
 337 static void
 338 ptlrpc_server_free_request(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 339 {
 340         unsigned long  flags;
 341         int            refcount;
 342
 343         spin_lock_irqsave(&svc->srv_lock, flags);
 344         svc->srv_n_active_reqs--;
 345         refcount = --(req->rq_rqbd->rqbd_refcount);
 346         if (refcount == 0) {
 347                 /* request buffer is now idle */
 348                 list_del(&req->rq_rqbd->rqbd_list);
 349                 list_add_tail(&req->rq_rqbd->rqbd_list,
 350                               &svc->srv_idle_rqbds);
 351         }
 352         spin_unlock_irqrestore(&svc->srv_lock, flags);
 353
 354         ptlrpc_free_server_req(req);
 355 }
 356
 357 static int
 358 ptlrpc_server_handle_request (struct ptlrpc_service *svc)
 359 {
 360         struct ptlrpc_request *request;
 361         unsigned long          flags;
 362         struct timeval         work_start;
 363         struct timeval         work_end;
 364         long                   timediff;
 365         int                    rc;
 366         ENTRY;
 367
 368         spin_lock_irqsave (&svc->srv_lock, flags);
 369         if (list_empty (&svc->srv_request_queue) ||
 370             (svc->srv_n_difficult_replies != 0 &&
 371              svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
 372                 /* If all the other threads are handling requests, I must
 373                  * remain free to handle any 'difficult' reply that might
 374                  * block them */
 375                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 376                 RETURN(0);
 377         }
 378
 379         request = list_entry (svc->srv_request_queue.next,
 380                               struct ptlrpc_request, rq_list);
 381         list_del_init (&request->rq_list);
 382         svc->srv_n_queued_reqs--;
 383         svc->srv_n_active_reqs++;
 384
 385         spin_unlock_irqrestore (&svc->srv_lock, flags);
 386
 387         do_gettimeofday(&work_start);
 388         timediff = timeval_sub(&work_start, &request->rq_arrival_time);
 389         if (svc->srv_stats != NULL) {
 390                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 391                                     timediff);
 392                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 393                                     svc->srv_n_queued_reqs);
 394                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 395                                     svc->srv_n_active_reqs);
 396         }
 397
 398 #if SWAB_PARANOIA
 399         /* Clear request swab mask; this is a new request */
 400         request->rq_req_swab_mask = 0;
 401 #endif
 402         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 403         if (rc != 0) {
 404                 CERROR ("error unpacking request: ptl %d from "LPX64
 405                         " xid "LPU64"\n", svc->srv_req_portal,
 406                        request->rq_peer.peer_nid, request->rq_xid);
 407                 goto out;
 408         }
 409
 410         rc = -EINVAL;
 411         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 412                 CERROR("wrong packet type received (type=%u) from "
 413                        LPX64"\n", request->rq_reqmsg->type,
 414                        request->rq_peer.peer_nid);
 415                 goto out;
 416         }
 417
 418         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 419
 420         /* Discard requests queued for longer than my timeout.  If the
 421          * client's timeout is similar to mine, she'll be timing out this
 422          * REQ anyway (bug 1502) */
 423         if (timediff / 1000000 > (long)obd_timeout) {
 424                 CERROR("Dropping timed-out request from "LPX64
 425                        ": %ld seconds old\n",
 426                        request->rq_peer.peer_nid, timediff / 1000000);
 427                 goto out;
 428         }
 429
 430         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 431
 432         if (request->rq_export) {
 433                 if (request->rq_reqmsg->conn_cnt <
 434                     request->rq_export->exp_conn_cnt) {
 435                         DEBUG_REQ(D_ERROR, request,
 436                                   "DROPPING req from old connection %d < %d",
 437                                   request->rq_reqmsg->conn_cnt,
 438                                   request->rq_export->exp_conn_cnt);
 439                         goto put_conn;
 440                 }
 441
 442                 request->rq_export->exp_last_request_time =
 443                         LTIME_S(CURRENT_TIME);
 444         }
 445
 446         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 447                "%s:%s+%d:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 448                (request->rq_export ?
 449                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 450                (request->rq_export ?
 451                 atomic_read(&request->rq_export->exp_refcount) : -99),
 452                request->rq_reqmsg->status, request->rq_xid,
 453                request->rq_peer.peer_ni->pni_name,
 454                request->rq_peer.peer_nid,
 455                request->rq_reqmsg->opc);
 456
 457         rc = svc->srv_handler(request);
 458         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 459                "%s:%s+%d:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 460                (request->rq_export ?
 461                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 462                (request->rq_export ?
 463                 atomic_read(&request->rq_export->exp_refcount) : -99),
 464                request->rq_reqmsg->status, request->rq_xid,
 465                request->rq_peer.peer_ni->pni_name,
 466                request->rq_peer.peer_nid,
 467                request->rq_reqmsg->opc);
 468
 469 put_conn:
 470         if (request->rq_export != NULL)
 471                 class_export_put(request->rq_export);
 472
 473  out:
 474         do_gettimeofday(&work_end);
 475
 476         timediff = timeval_sub(&work_end, &work_start);
 477
 478         CDEBUG((timediff / 1000000 > (long)obd_timeout) ? D_ERROR : D_HA,
 479                "request "LPU64" opc %u from NID "LPX64" processed in %ldus "
 480                "(%ldus total)\n", request->rq_xid, request->rq_reqmsg->opc,
 481                request->rq_peer.peer_nid,
 482                timediff, timeval_sub(&work_end, &request->rq_arrival_time));
 483
 484         if (svc->srv_stats != NULL) {
 485                 int opc = opcode_offset(request->rq_reqmsg->opc);
 486                 if (opc > 0) {
 487                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 488                         lprocfs_counter_add(svc->srv_stats,
 489                                             opc + PTLRPC_LAST_CNTR,
 490                                             timediff);
 491                 }
 492         }
 493
 494         ptlrpc_server_free_request(svc, request);
 495
 496         RETURN(1);
 497 }
 498
 499 static int
 500 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 501 {
 502         struct ptlrpc_reply_state *rs;
 503         unsigned long              flags;
 504         struct obd_export         *exp;
 505         struct obd_device         *obd;
 506         int                        nlocks;
 507         int                        been_handled;
 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"LPX64"\n",
 553                       rs,
 554                       rs->rs_xid, rs->rs_transno,
 555                       rs->rs_msg.opc, exp->exp_connection->c_peer.peer_nid);
 556         }
 557
 558         if ((!been_handled && rs->rs_on_net) ||
 559             nlocks > 0) {
 560                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 561
 562                 if (!been_handled && rs->rs_on_net) {
 563                         PtlMDUnlink(rs->rs_md_h);
 564                         /* Ignore return code; we're racing with
 565                          * completion... */
 566                 }
 567
 568                 while (nlocks-- > 0)
 569                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 570                                          rs->rs_modes[nlocks]);
 571
 572                 spin_lock_irqsave(&svc->srv_lock, flags);
 573         }
 574
 575         rs->rs_scheduled = 0;
 576
 577         if (!rs->rs_on_net) {
 578                 /* Off the net */
 579                 svc->srv_n_difficult_replies--;
 580                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 581
 582                 class_export_put (exp);
 583                 rs->rs_export = NULL;
 584                 lustre_free_reply_state (rs);
 585                 atomic_dec (&svc->srv_outstanding_replies);
 586                 RETURN(1);
 587         }
 588
 589         /* still on the net; callback will schedule */
 590         spin_unlock_irqrestore (&svc->srv_lock, flags);
 591         RETURN(1);
 592 }
 593
 594 #ifndef __KERNEL__
 595 /* FIXME make use of timeout later */
 596 int
 597 liblustre_check_services (void *arg)
 598 {
 599         int  did_something = 0;
 600         int  rc;
 601         struct list_head *tmp, *nxt;
 602         ENTRY;
 603
 604         /* I'm relying on being single threaded, not to have to lock
 605          * ptlrpc_all_services etc */
 606         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 607                 struct ptlrpc_service *svc =
 608                         list_entry (tmp, struct ptlrpc_service, srv_list);
 609
 610                 if (svc->srv_nthreads != 0)     /* I've recursed */
 611                         continue;
 612
 613                 /* service threads can block for bulk, so this limits us
 614                  * (arbitrarily) to recursing 1 stack frame per service.
 615                  * Note that the problem with recursion is that we have to
 616                  * unwind completely before our caller can resume. */
 617
 618                 svc->srv_nthreads++;
 619
 620                 do {
 621                         rc = ptlrpc_server_handle_reply(svc);
 622                         rc |= ptlrpc_server_handle_request(svc);
 623                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 624                         did_something |= rc;
 625                 } while (rc);
 626
 627                 svc->srv_nthreads--;
 628         }
 629
 630         RETURN(did_something);
 631 }
 632
 633 #else /* __KERNEL__ */
 634
 635 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 636 void ptlrpc_daemonize(void)
 637 {
 638         exit_mm(current);
 639         lustre_daemonize_helper();
 640         exit_files(current);
 641         reparent_to_init();
 642 }
 643
 644 static int
 645 ptlrpc_retry_rqbds(void *arg)
 646 {
 647         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 648
 649         svc->srv_rqbd_timeout = 0;
 650         return (-ETIMEDOUT);
 651 }
 652
 653 static int ptlrpc_main(void *arg)
 654 {
 655         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 656         struct ptlrpc_service  *svc = data->svc;
 657         struct ptlrpc_thread   *thread = data->thread;
 658         unsigned long           flags;
 659         ENTRY;
 660
 661         lock_kernel();
 662         ptlrpc_daemonize();
 663
 664         SIGNAL_MASK_LOCK(current, flags);
 665         sigfillset(&current->blocked);
 666         RECALC_SIGPENDING;
 667         SIGNAL_MASK_UNLOCK(current, flags);
 668
 669         THREAD_NAME(current->comm, "%s", data->name);
 670         unlock_kernel();
 671
 672         /* Record that the thread is running */
 673         thread->t_flags = SVC_RUNNING;
 674         wake_up(&thread->t_ctl_waitq);
 675
 676         spin_lock_irqsave(&svc->srv_lock, flags);
 677         svc->srv_nthreads++;
 678         spin_unlock_irqrestore(&svc->srv_lock, flags);
 679
 680         /* XXX maintain a list of all managed devices: insert here */
 681
 682         while ((thread->t_flags & SVC_STOPPING) == 0 ||
 683                svc->srv_n_difficult_replies != 0) {
 684                 /* Don't exit while there are replies to be handled */
 685                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
 686                                                      ptlrpc_retry_rqbds, svc);
 687
 688                 l_wait_event_exclusive (svc->srv_waitq,
 689                               ((thread->t_flags & SVC_STOPPING) != 0 &&
 690                                svc->srv_n_difficult_replies == 0) ||
 691                               (!list_empty(&svc->srv_idle_rqbds) &&
 692                                svc->srv_rqbd_timeout == 0) ||
 693                               !list_empty (&svc->srv_reply_queue) ||
 694                               (!list_empty (&svc->srv_request_queue) &&
 695                                (svc->srv_n_difficult_replies == 0 ||
 696                                 svc->srv_n_active_reqs <
 697                                 (svc->srv_nthreads - 1))),
 698                               &lwi);
 699
 700                 if (!list_empty (&svc->srv_reply_queue))
 701                         ptlrpc_server_handle_reply (svc);
 702
 703                 /* only handle requests if there are no difficult replies
 704                  * outstanding, or I'm not the last thread handling
 705                  * requests */
 706                 if (!list_empty (&svc->srv_request_queue) &&
 707                     (svc->srv_n_difficult_replies == 0 ||
 708                      svc->srv_n_active_reqs < (svc->srv_nthreads - 1)))
 709                         ptlrpc_server_handle_request (svc);
 710
 711                 if (!list_empty(&svc->srv_idle_rqbds) &&
 712                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
 713                         /* I just failed to repost request buffers.  Wait
 714                          * for a timeout (unless something else happens)
 715                          * before I try again */
 716                         svc->srv_rqbd_timeout = HZ/10;
 717                 }
 718         }
 719
 720         spin_lock_irqsave(&svc->srv_lock, flags);
 721
 722         svc->srv_nthreads--;                    /* must know immediately */
 723         thread->t_flags = SVC_STOPPED;
 724         wake_up(&thread->t_ctl_waitq);
 725
 726         spin_unlock_irqrestore(&svc->srv_lock, flags);
 727
 728         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 729         return 0;
 730 }
 731
 732 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 733                                struct ptlrpc_thread *thread)
 734 {
 735         struct l_wait_info lwi = { 0 };
 736         unsigned long      flags;
 737
 738         spin_lock_irqsave(&svc->srv_lock, flags);
 739         thread->t_flags = SVC_STOPPING;
 740         spin_unlock_irqrestore(&svc->srv_lock, flags);
 741
 742         wake_up_all(&svc->srv_waitq);
 743         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 744                      &lwi);
 745
 746         spin_lock_irqsave(&svc->srv_lock, flags);
 747         list_del(&thread->t_link);
 748         spin_unlock_irqrestore(&svc->srv_lock, flags);
 749
 750         OBD_FREE(thread, sizeof(*thread));
 751 }
 752
 753 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 754 {
 755         unsigned long flags;
 756         struct ptlrpc_thread *thread;
 757
 758         spin_lock_irqsave(&svc->srv_lock, flags);
 759         while (!list_empty(&svc->srv_threads)) {
 760                 thread = list_entry(svc->srv_threads.next,
 761                                     struct ptlrpc_thread, t_link);
 762
 763                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 764                 ptlrpc_stop_thread(svc, thread);
 765                 spin_lock_irqsave(&svc->srv_lock, flags);
 766         }
 767
 768         spin_unlock_irqrestore(&svc->srv_lock, flags);
 769 }
 770
 771 int ptlrpc_start_n_threads(struct obd_device *dev, struct ptlrpc_service *svc,
 772                            int num_threads, char *base_name)
 773 {
 774         int i, rc = 0;
 775         ENTRY;
 776
 777         for (i = 0; i < num_threads; i++) {
 778                 char name[32];
 779                 sprintf(name, "%s_%02d", base_name, i);
 780                 rc = ptlrpc_start_thread(dev, svc, name);
 781                 if (rc) {
 782                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
 783                                i, rc);
 784                         ptlrpc_stop_all_threads(svc);
 785                 }
 786         }
 787         RETURN(rc);
 788 }
 789
 790 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 791                         char *name)
 792 {
 793         struct l_wait_info lwi = { 0 };
 794         struct ptlrpc_svc_data d;
 795         struct ptlrpc_thread *thread;
 796         unsigned long flags;
 797         int rc;
 798         ENTRY;
 799
 800         OBD_ALLOC(thread, sizeof(*thread));
 801         if (thread == NULL)
 802                 RETURN(-ENOMEM);
 803         init_waitqueue_head(&thread->t_ctl_waitq);
 804
 805         d.dev = dev;
 806         d.svc = svc;
 807         d.name = name;
 808         d.thread = thread;
 809
 810         spin_lock_irqsave(&svc->srv_lock, flags);
 811         list_add(&thread->t_link, &svc->srv_threads);
 812         spin_unlock_irqrestore(&svc->srv_lock, flags);
 813
 814         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 815          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 816          */
 817         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 818         if (rc < 0) {
 819                 CERROR("cannot start thread: %d\n", rc);
 820                 OBD_FREE(thread, sizeof(*thread));
 821                 RETURN(rc);
 822         }
 823         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 824
 825         RETURN(0);
 826 }
 827 #endif
 828
 829 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 830 {
 831         int                   i;
 832         int                   rc;
 833         unsigned long         flags;
 834         struct ptlrpc_srv_ni *srv_ni;
 835         struct l_wait_info    lwi;
 836         struct list_head     *tmp;
 837
 838         LASSERT(list_empty(&service->srv_threads));
 839
 840         spin_lock (&ptlrpc_all_services_lock);
 841         list_del_init (&service->srv_list);
 842         spin_unlock (&ptlrpc_all_services_lock);
 843
 844         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 845                 srv_ni = &service->srv_interfaces[i];
 846                 CDEBUG(D_NET, "%s: tearing down interface %s\n",
 847                        service->srv_name, srv_ni->sni_ni->pni_name);
 848
 849                 /* Unlink all the request buffers.  This forces a 'final'
 850                  * event with its 'unlink' flag set for each posted rqbd */
 851                 list_for_each(tmp, &srv_ni->sni_active_rqbds) {
 852                         struct ptlrpc_request_buffer_desc *rqbd =
 853                                 list_entry(tmp, struct ptlrpc_request_buffer_desc,
 854                                            rqbd_list);
 855
 856                         rc = PtlMDUnlink(rqbd->rqbd_md_h);
 857                         LASSERT (rc == PTL_OK || rc == PTL_MD_INVALID);
 858                 }
 859
 860                 /* Wait for the network to release any buffers it's
 861                  * currently filling */
 862                 for (;;) {
 863                         spin_lock_irqsave(&service->srv_lock, flags);
 864                         rc = srv_ni->sni_nrqbd_receiving;
 865                         spin_unlock_irqrestore(&service->srv_lock, flags);
 866
 867                         if (rc == 0)
 868                                 break;
 869
 870                         /* Network access will complete in finite time but
 871                          * the HUGE timeout lets us CWARN for visibility of
 872                          * sluggish NALs */
 873                         lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
 874                         rc = l_wait_event(service->srv_waitq,
 875                                           srv_ni->sni_nrqbd_receiving == 0,
 876                                           &lwi);
 877                         if (rc == -ETIMEDOUT)
 878                                 CWARN("Waiting for request buffers on "
 879                                       "service %s on interface %s ",
 880                                       service->srv_name, srv_ni->sni_ni->pni_name);
 881                 }
 882
 883                 /* schedule all outstanding replies to terminate them */
 884                 spin_lock_irqsave(&service->srv_lock, flags);
 885                 while (!list_empty(&srv_ni->sni_active_replies)) {
 886                         struct ptlrpc_reply_state *rs =
 887                                 list_entry(srv_ni->sni_active_replies.next,
 888                                            struct ptlrpc_reply_state,
 889                                            rs_list);
 890                         ptlrpc_schedule_difficult_reply(rs);
 891                 }
 892                 spin_unlock_irqrestore(&service->srv_lock, flags);
 893         }
 894
 895         /* purge the request queue.  NB No new replies (rqbds all unlinked)
 896          * and no service threads, so I'm the only thread noodling the
 897          * request queue now */
 898         while (!list_empty(&service->srv_request_queue)) {
 899                 struct ptlrpc_request *req =
 900                         list_entry(service->srv_request_queue.next,
 901                                    struct ptlrpc_request,
 902                                    rq_list);
 903
 904                 list_del(&req->rq_list);
 905                 service->srv_n_queued_reqs--;
 906                 service->srv_n_active_reqs++;
 907
 908                 ptlrpc_server_free_request(service, req);
 909         }
 910         LASSERT(service->srv_n_queued_reqs == 0);
 911         LASSERT(service->srv_n_active_reqs == 0);
 912
 913         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 914                 srv_ni = &service->srv_interfaces[i];
 915                 LASSERT(list_empty(&srv_ni->sni_active_rqbds));
 916         }
 917
 918         /* Now free all the request buffers since nothing references them
 919          * any more... */
 920         while (!list_empty(&service->srv_idle_rqbds)) {
 921                 struct ptlrpc_request_buffer_desc *rqbd =
 922                         list_entry(service->srv_idle_rqbds.next,
 923                                    struct ptlrpc_request_buffer_desc,
 924                                    rqbd_list);
 925
 926                 ptlrpc_free_rqbd(rqbd);
 927         }
 928
 929         /* wait for all outstanding replies to complete (they were
 930          * scheduled having been flagged to abort above) */
 931         while (atomic_read(&service->srv_outstanding_replies) != 0) {
 932                 struct l_wait_info lwi = LWI_TIMEOUT(10 * HZ, NULL, NULL);
 933
 934                 rc = l_wait_event(service->srv_waitq,
 935                                   !list_empty(&service->srv_reply_queue), &lwi);
 936                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 937
 938                 if (rc == 0) {
 939                         ptlrpc_server_handle_reply(service);
 940                         continue;
 941                 }
 942                 CWARN("Unexpectedly long timeout %p\n", service);
 943         }
 944
 945         ptlrpc_lprocfs_unregister_service(service);
 946
 947         OBD_FREE(service,
 948                  offsetof(struct ptlrpc_service,
 949                           srv_interfaces[ptlrpc_ninterfaces]));
 950         return 0;
 951 }