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 <linux/lustre_log.h>
  32 #include <portals/types.h>
  33 #include "ptlrpc_internal.h"
  34
  35 static LIST_HEAD (ptlrpc_all_services);
  36 static spinlock_t ptlrpc_all_services_lock = SPIN_LOCK_UNLOCKED;
  37
  38 static void
  39 ptlrpc_free_server_req (struct ptlrpc_request *req)
  40 {
  41         /* The last request to be received into a request buffer uses space
  42          * in the request buffer descriptor, otherwise requests are
  43          * allocated dynamically in the incoming reply event handler */
  44         if (req == &req->rq_rqbd->rqbd_req)
  45                 return;
  46
  47         OBD_FREE(req, sizeof(*req));
  48 }
  49
  50 static char *
  51 ptlrpc_alloc_request_buffer (int size)
  52 {
  53         char *ptr;
  54
  55         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  56                 OBD_VMALLOC(ptr, size);
  57         else
  58                 OBD_ALLOC(ptr, size);
  59
  60         return (ptr);
  61 }
  62
  63 static void
  64 ptlrpc_free_request_buffer (char *ptr, int size)
  65 {
  66         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  67                 OBD_VFREE(ptr, size);
  68         else
  69                 OBD_FREE(ptr, size);
  70 }
  71
  72 struct ptlrpc_request_buffer_desc *
  73 ptlrpc_alloc_rqbd (struct ptlrpc_srv_ni *srv_ni)
  74 {
  75         struct ptlrpc_service             *svc = srv_ni->sni_service;
  76         unsigned long                      flags;
  77         struct ptlrpc_request_buffer_desc *rqbd;
  78
  79         OBD_ALLOC(rqbd, sizeof (*rqbd));
  80         if (rqbd == NULL)
  81                 return (NULL);
  82
  83         rqbd->rqbd_srv_ni = srv_ni;
  84         rqbd->rqbd_refcount = 0;
  85         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  86         rqbd->rqbd_cbid.cbid_arg = rqbd;
  87         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
  88
  89         if (rqbd->rqbd_buffer == NULL) {
  90                 OBD_FREE(rqbd, sizeof (*rqbd));
  91                 return (NULL);
  92         }
  93
  94         spin_lock_irqsave (&svc->srv_lock, flags);
  95         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
  96         svc->srv_nbufs++;
  97         spin_unlock_irqrestore (&svc->srv_lock, flags);
  98
  99         return (rqbd);
 100 }
 101
 102 void
 103 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 104 {
 105         struct ptlrpc_srv_ni  *sni = rqbd->rqbd_srv_ni;
 106         struct ptlrpc_service *svc = sni->sni_service;
 107         unsigned long          flags;
 108
 109         LASSERT (rqbd->rqbd_refcount == 0);
 110
 111         spin_lock_irqsave(&svc->srv_lock, flags);
 112         list_del(&rqbd->rqbd_list);
 113         svc->srv_nbufs--;
 114         spin_unlock_irqrestore(&svc->srv_lock, flags);
 115
 116         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 117         OBD_FREE (rqbd, sizeof (*rqbd));
 118 }
 119
 120 void
 121 ptlrpc_save_llog_lock (struct ptlrpc_request *req,
 122                        struct llog_create_locks *lcl)
 123 {
 124         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 125         LASSERT (rs != NULL);
 126         LASSERT (rs->rs_llog_locks == NULL);
 127
 128         rs->rs_llog_locks = lcl;
 129 }
 130
 131 void
 132 ptlrpc_save_lock (struct ptlrpc_request *req,
 133                   struct lustre_handle *lock, int mode)
 134 {
 135         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 136         int                        idx;
 137
 138         if (!lock->cookie)
 139                 return;
 140
 141         LASSERT (rs != NULL);
 142         LASSERT (rs->rs_nlocks < RS_MAX_LOCKS);
 143
 144         idx = rs->rs_nlocks++;
 145         rs->rs_locks[idx] = *lock;
 146         rs->rs_modes[idx] = mode;
 147         rs->rs_difficult = 1;
 148 }
 149
 150 void
 151 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 152 {
 153         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 154
 155 #ifdef CONFIG_SMP
 156         LASSERT (spin_is_locked (&svc->srv_lock));
 157 #endif
 158         LASSERT (rs->rs_difficult);
 159         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 160
 161         if (rs->rs_scheduled)                   /* being set up or already notified */
 162                 return;
 163
 164         rs->rs_scheduled = 1;
 165         list_del (&rs->rs_list);
 166         list_add (&rs->rs_list, &svc->srv_reply_queue);
 167         wake_up (&svc->srv_waitq);
 168 }
 169
 170 void
 171 ptlrpc_commit_replies (struct obd_device *obd)
 172 {
 173         struct list_head   *tmp;
 174         struct list_head   *nxt;
 175         unsigned long       flags;
 176
 177         /* Find any replies that have been committed and get their service
 178          * to attend to complete them. */
 179
 180         /* CAVEAT EMPTOR: spinlock ordering!!! */
 181         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
 182
 183         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 184                 struct ptlrpc_reply_state *rs =
 185                         list_entry (tmp, struct ptlrpc_reply_state, rs_obd_list);
 186                 struct llog_create_locks *lcl = rs->rs_llog_locks;
 187
 188                 rs->rs_llog_locks = NULL;
 189                 LASSERT (rs->rs_difficult);
 190
 191                 if (rs->rs_transno <= obd->obd_last_committed) {
 192                         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 193
 194                         spin_lock (&svc->srv_lock);
 195                         list_del_init (&rs->rs_obd_list);
 196                         ptlrpc_schedule_difficult_reply (rs);
 197                         spin_unlock (&svc->srv_lock);
 198
 199                         if (lcl != NULL)
 200                                 llog_create_lock_free(lcl);
 201                 }
 202         }
 203
 204         spin_unlock_irqrestore (&obd->obd_uncommitted_replies_lock, flags);
 205 }
 206
 207 static long
 208 timeval_sub(struct timeval *large, struct timeval *small)
 209 {
 210         return (large->tv_sec - small->tv_sec) * 1000000 +
 211                 (large->tv_usec - small->tv_usec);
 212 }
 213
 214 static int
 215 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 216 {
 217         struct ptlrpc_srv_ni              *srv_ni;
 218         struct ptlrpc_request_buffer_desc *rqbd;
 219         unsigned long                      flags;
 220         int                                rc;
 221
 222         spin_lock_irqsave(&svc->srv_lock, flags);
 223         if (list_empty (&svc->srv_idle_rqbds)) {
 224                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 225                 return (0);
 226         }
 227
 228         rqbd = list_entry(svc->srv_idle_rqbds.next,
 229                           struct ptlrpc_request_buffer_desc,
 230                           rqbd_list);
 231         list_del (&rqbd->rqbd_list);
 232
 233         /* assume we will post successfully */
 234         srv_ni = rqbd->rqbd_srv_ni;
 235         srv_ni->sni_nrqbd_receiving++;
 236         list_add (&rqbd->rqbd_list, &srv_ni->sni_active_rqbds);
 237
 238         spin_unlock_irqrestore(&svc->srv_lock, flags);
 239
 240         rc = ptlrpc_register_rqbd(rqbd);
 241         if (rc == 0)
 242                 return (1);
 243
 244         spin_lock_irqsave(&svc->srv_lock, flags);
 245
 246         srv_ni->sni_nrqbd_receiving--;
 247         list_del(&rqbd->rqbd_list);
 248         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 249
 250         if (srv_ni->sni_nrqbd_receiving == 0) {
 251                 /* This service is off-air on this interface because all
 252                  * its request buffers are busy.  Portals will have started
 253                  * dropping incoming requests until more buffers get
 254                  * posted */
 255                 CERROR("All %s %s request buffers busy\n",
 256                        svc->srv_name, srv_ni->sni_ni->pni_name);
 257         }
 258
 259         spin_unlock_irqrestore (&svc->srv_lock, flags);
 260
 261         return (-1);
 262 }
 263
 264 struct ptlrpc_service *
 265 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size,
 266                 int req_portal, int rep_portal,
 267                 svc_handler_t handler, char *name,
 268                 struct proc_dir_entry *proc_entry)
 269 {
 270         int                                i;
 271         int                                j;
 272         int                                ssize;
 273         struct ptlrpc_service             *service;
 274         struct ptlrpc_srv_ni              *srv_ni;
 275         struct ptlrpc_request_buffer_desc *rqbd;
 276         ENTRY;
 277
 278         LASSERT (ptlrpc_ninterfaces > 0);
 279         LASSERT (nbufs > 0);
 280         LASSERT (bufsize >= max_req_size);
 281
 282         ssize = offsetof (struct ptlrpc_service,
 283                           srv_interfaces[ptlrpc_ninterfaces]);
 284         OBD_ALLOC(service, ssize);
 285         if (service == NULL)
 286                 RETURN(NULL);
 287
 288         service->srv_name = name;
 289         spin_lock_init(&service->srv_lock);
 290         INIT_LIST_HEAD(&service->srv_threads);
 291         init_waitqueue_head(&service->srv_waitq);
 292
 293         service->srv_max_req_size = max_req_size;
 294         service->srv_buf_size = bufsize;
 295         service->srv_rep_portal = rep_portal;
 296         service->srv_req_portal = req_portal;
 297         service->srv_handler = handler;
 298
 299         INIT_LIST_HEAD(&service->srv_request_queue);
 300         INIT_LIST_HEAD(&service->srv_idle_rqbds);
 301         INIT_LIST_HEAD(&service->srv_reply_queue);
 302
 303         /* First initialise enough for early teardown */
 304         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 305                 srv_ni = &service->srv_interfaces[i];
 306
 307                 srv_ni->sni_service = service;
 308                 srv_ni->sni_ni = &ptlrpc_interfaces[i];
 309                 INIT_LIST_HEAD(&srv_ni->sni_active_rqbds);
 310                 INIT_LIST_HEAD(&srv_ni->sni_active_replies);
 311         }
 312
 313         spin_lock (&ptlrpc_all_services_lock);
 314         list_add (&service->srv_list, &ptlrpc_all_services);
 315         spin_unlock (&ptlrpc_all_services_lock);
 316
 317         /* Now allocate the request buffers, assuming all interfaces require
 318          * the same number. */
 319         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 320                 srv_ni = &service->srv_interfaces[i];
 321                 CDEBUG (D_NET, "%s: initialising interface %s\n", name,
 322                         srv_ni->sni_ni->pni_name);
 323
 324                 for (j = 0; j < nbufs; j++) {
 325                         rqbd = ptlrpc_alloc_rqbd (srv_ni);
 326
 327                         if (rqbd == NULL) {
 328                                 CERROR ("%s.%d: Can't allocate request %d "
 329                                         "on %s\n", name, i, j,
 330                                         srv_ni->sni_ni->pni_name);
 331                                 GOTO(failed, NULL);
 332                         }
 333
 334                         /* We shouldn't be under memory pressure at
 335                          * startup, so fail if we can't post all our
 336                          * buffers at this time. */
 337                         if (ptlrpc_server_post_idle_rqbds(service) <= 0)
 338                                 GOTO(failed, NULL);
 339                 }
 340         }
 341
 342         if (proc_entry != NULL)
 343                 ptlrpc_lprocfs_register_service(proc_entry, service);
 344
 345         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 346                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 347
 348         RETURN(service);
 349 failed:
 350         ptlrpc_unregister_service(service);
 351         return NULL;
 352 }
 353
 354 static void
 355 ptlrpc_server_free_request(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 356 {
 357         unsigned long  flags;
 358         int            refcount;
 359
 360         spin_lock_irqsave(&svc->srv_lock, flags);
 361         svc->srv_n_active_reqs--;
 362         refcount = --(req->rq_rqbd->rqbd_refcount);
 363         if (refcount == 0) {
 364                 /* request buffer is now idle */
 365                 list_del(&req->rq_rqbd->rqbd_list);
 366                 list_add_tail(&req->rq_rqbd->rqbd_list,
 367                               &svc->srv_idle_rqbds);
 368         }
 369         spin_unlock_irqrestore(&svc->srv_lock, flags);
 370
 371         ptlrpc_free_server_req(req);
 372 }
 373
 374 static int
 375 ptlrpc_server_handle_request (struct ptlrpc_service *svc)
 376 {
 377         struct obd_export     *export = NULL;
 378         struct ptlrpc_request *request;
 379         unsigned long          flags;
 380         struct timeval         work_start;
 381         struct timeval         work_end;
 382         long                   timediff;
 383         int                    rc;
 384         char                   str[PTL_NALFMT_SIZE];
 385         ENTRY;
 386
 387         spin_lock_irqsave (&svc->srv_lock, flags);
 388         if (list_empty (&svc->srv_request_queue) ||
 389             (svc->srv_n_difficult_replies != 0 &&
 390              svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
 391                 /* If all the other threads are handling requests, I must
 392                  * remain free to handle any 'difficult' reply that might
 393                  * block them */
 394                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 395                 RETURN(0);
 396         }
 397
 398         request = list_entry (svc->srv_request_queue.next,
 399                               struct ptlrpc_request, rq_list);
 400         list_del_init (&request->rq_list);
 401         svc->srv_n_queued_reqs--;
 402         svc->srv_n_active_reqs++;
 403
 404         spin_unlock_irqrestore (&svc->srv_lock, flags);
 405
 406         do_gettimeofday(&work_start);
 407         timediff = timeval_sub(&work_start, &request->rq_arrival_time);
 408         if (svc->srv_stats != NULL) {
 409                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 410                                     timediff);
 411                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 412                                     svc->srv_n_queued_reqs);
 413                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 414                                     svc->srv_n_active_reqs);
 415         }
 416
 417 #if SWAB_PARANOIA
 418         /* Clear request swab mask; this is a new request */
 419         request->rq_req_swab_mask = 0;
 420 #endif
 421         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 422         if (rc != 0) {
 423                 CERROR ("error unpacking request: ptl %d from %s"
 424                         " xid "LPU64"\n", svc->srv_req_portal,
 425                         ptlrpc_peernid2str(&request->rq_peer, str),
 426                        request->rq_xid);
 427                 goto out;
 428         }
 429
 430         rc = -EINVAL;
 431         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 432                 CERROR("wrong packet type received (type=%u) from %s\n",
 433                        request->rq_reqmsg->type,
 434                        ptlrpc_peernid2str(&request->rq_peer, str));
 435                 goto out;
 436         }
 437
 438         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 439
 440         /* Discard requests queued for longer than my timeout.  If the
 441          * client's timeout is similar to mine, she'll be timing out this
 442          * REQ anyway (bug 1502) */
 443         if (timediff / 1000000 > (long)obd_timeout) {
 444                 CERROR("Dropping timed-out request from %s: %ld seconds old\n",
 445                        ptlrpc_peernid2str(&request->rq_peer, str),
 446                        timediff / 1000000);
 447                 goto out;
 448         }
 449
 450         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 451
 452         if (request->rq_export) {
 453                 if (request->rq_reqmsg->conn_cnt <
 454                     request->rq_export->exp_conn_cnt) {
 455                         DEBUG_REQ(D_ERROR, request,
 456                                   "DROPPING req from old connection %d < %d",
 457                                   request->rq_reqmsg->conn_cnt,
 458                                   request->rq_export->exp_conn_cnt);
 459                         goto put_conn;
 460                 }
 461
 462                 export = class_export_rpc_get(request->rq_export);
 463                 request->rq_export->exp_last_request_time =
 464                         LTIME_S(CURRENT_TIME);
 465         }
 466
 467         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 468                "%s:%s+%d:%d:"LPU64":%s:%s:%d\n", current->comm,
 469                (request->rq_export ?
 470                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 471                (request->rq_export ?
 472                 atomic_read(&request->rq_export->exp_refcount) : -99),
 473                request->rq_reqmsg->status, request->rq_xid,
 474                request->rq_peer.peer_ni->pni_name,
 475                ptlrpc_peernid2str(&request->rq_peer, str),
 476                request->rq_reqmsg->opc);
 477
 478         rc = svc->srv_handler(request);
 479         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 480                "%s:%s+%d:%d:"LPU64":%s:%s:%d\n", current->comm,
 481                (request->rq_export ?
 482                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 483                (request->rq_export ?
 484                 atomic_read(&request->rq_export->exp_refcount) : -99),
 485                request->rq_reqmsg->status, request->rq_xid,
 486                request->rq_peer.peer_ni->pni_name,
 487                ptlrpc_peernid2str(&request->rq_peer, str),
 488                request->rq_reqmsg->opc);
 489
 490         if (export != NULL)
 491                 class_export_rpc_put(export);
 492
 493 put_conn:
 494         if (request->rq_export != NULL)
 495                 class_export_put(request->rq_export);
 496
 497  out:
 498         do_gettimeofday(&work_end);
 499
 500         timediff = timeval_sub(&work_end, &work_start);
 501
 502         CDEBUG((timediff / 1000000 > (long)obd_timeout) ? D_ERROR : D_HA,
 503                "request "LPU64" opc %u from NID %s processed in %ldus "
 504                "(%ldus total)\n", request->rq_xid, request->rq_reqmsg->opc,
 505                ptlrpc_peernid2str(&request->rq_peer, str),
 506                timediff, timeval_sub(&work_end, &request->rq_arrival_time));
 507
 508         if (svc->srv_stats != NULL) {
 509                 int opc = opcode_offset(request->rq_reqmsg->opc);
 510                 if (opc > 0) {
 511                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 512                         lprocfs_counter_add(svc->srv_stats,
 513                                             opc + PTLRPC_LAST_CNTR,
 514                                             timediff);
 515                 }
 516         }
 517
 518         ptlrpc_server_free_request(svc, request);
 519
 520         RETURN(1);
 521 }
 522
 523 static int
 524 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 525 {
 526         struct ptlrpc_reply_state *rs;
 527         unsigned long              flags;
 528         struct obd_export         *exp;
 529         struct obd_device         *obd;
 530         struct llog_create_locks  *lcl;
 531         int                        nlocks;
 532         int                        been_handled;
 533         char                       str[PTL_NALFMT_SIZE];
 534         ENTRY;
 535
 536         spin_lock_irqsave (&svc->srv_lock, flags);
 537         if (list_empty (&svc->srv_reply_queue)) {
 538                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 539                 RETURN(0);
 540         }
 541
 542         rs = list_entry (svc->srv_reply_queue.next,
 543                          struct ptlrpc_reply_state, rs_list);
 544
 545         exp = rs->rs_export;
 546         obd = exp->exp_obd;
 547
 548         LASSERT (rs->rs_difficult);
 549         LASSERT (rs->rs_scheduled);
 550
 551         list_del_init (&rs->rs_list);
 552
 553         /* Disengage from notifiers carefully (lock ordering!) */
 554         spin_unlock(&svc->srv_lock);
 555
 556         spin_lock (&obd->obd_uncommitted_replies_lock);
 557         /* Noop if removed already */
 558         list_del_init (&rs->rs_obd_list);
 559         spin_unlock (&obd->obd_uncommitted_replies_lock);
 560
 561         spin_lock (&exp->exp_lock);
 562         /* Noop if removed already */
 563         list_del_init (&rs->rs_exp_list);
 564         spin_unlock (&exp->exp_lock);
 565
 566         spin_lock(&svc->srv_lock);
 567
 568         been_handled = rs->rs_handled;
 569         rs->rs_handled = 1;
 570
 571         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 572         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 573
 574         lcl = rs->rs_llog_locks;
 575         rs->rs_llog_locks = NULL;
 576
 577         if (nlocks == 0 && !been_handled) {
 578                 /* If we see this, we should already have seen the warning
 579                  * in mds_steal_ack_locks()  */
 580                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 581                       " o%d NID %s\n",
 582                       rs,
 583                       rs->rs_xid, rs->rs_transno,
 584                       rs->rs_msg.opc,
 585                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 586         }
 587
 588         if ((!been_handled && rs->rs_on_net) ||
 589             nlocks > 0 || lcl != NULL) {
 590                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 591
 592                 if (!been_handled && rs->rs_on_net) {
 593                         PtlMDUnlink(rs->rs_md_h);
 594                         /* Ignore return code; we're racing with
 595                          * completion... */
 596                 }
 597
 598                 while (nlocks-- > 0)
 599                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 600                                          rs->rs_modes[nlocks]);
 601
 602                 if (lcl != NULL)
 603                         llog_create_lock_free(lcl);
 604
 605                 spin_lock_irqsave(&svc->srv_lock, flags);
 606         }
 607
 608         rs->rs_scheduled = 0;
 609
 610         if (!rs->rs_on_net) {
 611                 /* Off the net */
 612                 svc->srv_n_difficult_replies--;
 613                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 614
 615                 class_export_put (exp);
 616                 rs->rs_export = NULL;
 617                 lustre_free_reply_state (rs);
 618                 atomic_dec (&svc->srv_outstanding_replies);
 619                 RETURN(1);
 620         }
 621
 622         /* still on the net; callback will schedule */
 623         spin_unlock_irqrestore (&svc->srv_lock, flags);
 624         RETURN(1);
 625 }
 626
 627 #ifndef __KERNEL__
 628 /* FIXME make use of timeout later */
 629 int
 630 liblustre_check_services (void *arg)
 631 {
 632         int  did_something = 0;
 633         int  rc;
 634         struct list_head *tmp, *nxt;
 635         ENTRY;
 636
 637         /* I'm relying on being single threaded, not to have to lock
 638          * ptlrpc_all_services etc */
 639         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 640                 struct ptlrpc_service *svc =
 641                         list_entry (tmp, struct ptlrpc_service, srv_list);
 642
 643                 if (svc->srv_nthreads != 0)     /* I've recursed */
 644                         continue;
 645
 646                 /* service threads can block for bulk, so this limits us
 647                  * (arbitrarily) to recursing 1 stack frame per service.
 648                  * Note that the problem with recursion is that we have to
 649                  * unwind completely before our caller can resume. */
 650
 651                 svc->srv_nthreads++;
 652
 653                 do {
 654                         rc = ptlrpc_server_handle_reply(svc);
 655                         rc |= ptlrpc_server_handle_request(svc);
 656                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 657                         did_something |= rc;
 658                 } while (rc);
 659
 660                 svc->srv_nthreads--;
 661         }
 662
 663         RETURN(did_something);
 664 }
 665
 666 #else /* __KERNEL__ */
 667
 668 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 669 void ptlrpc_daemonize(void)
 670 {
 671         exit_mm(current);
 672         lustre_daemonize_helper();
 673         exit_files(current);
 674         reparent_to_init();
 675 }
 676
 677 static int
 678 ptlrpc_retry_rqbds(void *arg)
 679 {
 680         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 681
 682         svc->srv_rqbd_timeout = 0;
 683         return (-ETIMEDOUT);
 684 }
 685
 686 static int ptlrpc_main(void *arg)
 687 {
 688         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 689         struct ptlrpc_service  *svc = data->svc;
 690         struct ptlrpc_thread   *thread = data->thread;
 691         unsigned long           flags;
 692         ENTRY;
 693
 694         lock_kernel();
 695         ptlrpc_daemonize();
 696
 697         SIGNAL_MASK_LOCK(current, flags);
 698         sigfillset(&current->blocked);
 699         RECALC_SIGPENDING;
 700         SIGNAL_MASK_UNLOCK(current, flags);
 701
 702         LASSERTF(strlen(data->name) < sizeof(current->comm),
 703                  "name %d > len %d\n",
 704                  (int)strlen(data->name), (int)sizeof(current->comm));
 705         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 706
 707         unlock_kernel();
 708
 709         /* Record that the thread is running */
 710         thread->t_flags = SVC_RUNNING;
 711         wake_up(&thread->t_ctl_waitq);
 712
 713         spin_lock_irqsave(&svc->srv_lock, flags);
 714         svc->srv_nthreads++;
 715         spin_unlock_irqrestore(&svc->srv_lock, flags);
 716
 717         /* XXX maintain a list of all managed devices: insert here */
 718
 719         while ((thread->t_flags & SVC_STOPPING) == 0 ||
 720                svc->srv_n_difficult_replies != 0) {
 721                 /* Don't exit while there are replies to be handled */
 722                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
 723                                                      ptlrpc_retry_rqbds, svc);
 724
 725                 l_wait_event_exclusive (svc->srv_waitq,
 726                               ((thread->t_flags & SVC_STOPPING) != 0 &&
 727                                svc->srv_n_difficult_replies == 0) ||
 728                               (!list_empty(&svc->srv_idle_rqbds) &&
 729                                svc->srv_rqbd_timeout == 0) ||
 730                               !list_empty (&svc->srv_reply_queue) ||
 731                               (!list_empty (&svc->srv_request_queue) &&
 732                                (svc->srv_n_difficult_replies == 0 ||
 733                                 svc->srv_n_active_reqs <
 734                                 (svc->srv_nthreads - 1))),
 735                               &lwi);
 736
 737                 if (!list_empty (&svc->srv_reply_queue))
 738                         ptlrpc_server_handle_reply (svc);
 739
 740                 /* only handle requests if there are no difficult replies
 741                  * outstanding, or I'm not the last thread handling
 742                  * requests */
 743                 if (!list_empty (&svc->srv_request_queue) &&
 744                     (svc->srv_n_difficult_replies == 0 ||
 745                      svc->srv_n_active_reqs < (svc->srv_nthreads - 1)))
 746                         ptlrpc_server_handle_request (svc);
 747
 748                 if (!list_empty(&svc->srv_idle_rqbds) &&
 749                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
 750                         /* I just failed to repost request buffers.  Wait
 751                          * for a timeout (unless something else happens)
 752                          * before I try again */
 753                         svc->srv_rqbd_timeout = HZ/10;
 754                 }
 755         }
 756
 757         spin_lock_irqsave(&svc->srv_lock, flags);
 758
 759         svc->srv_nthreads--;                    /* must know immediately */
 760         thread->t_flags = SVC_STOPPED;
 761         wake_up(&thread->t_ctl_waitq);
 762
 763         spin_unlock_irqrestore(&svc->srv_lock, flags);
 764
 765         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 766         return 0;
 767 }
 768
 769 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 770                                struct ptlrpc_thread *thread)
 771 {
 772         struct l_wait_info lwi = { 0 };
 773         unsigned long      flags;
 774
 775         spin_lock_irqsave(&svc->srv_lock, flags);
 776         thread->t_flags = SVC_STOPPING;
 777         spin_unlock_irqrestore(&svc->srv_lock, flags);
 778
 779         wake_up_all(&svc->srv_waitq);
 780         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 781                      &lwi);
 782
 783         spin_lock_irqsave(&svc->srv_lock, flags);
 784         list_del(&thread->t_link);
 785         spin_unlock_irqrestore(&svc->srv_lock, flags);
 786
 787         OBD_FREE(thread, sizeof(*thread));
 788 }
 789
 790 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 791 {
 792         unsigned long flags;
 793         struct ptlrpc_thread *thread;
 794
 795         spin_lock_irqsave(&svc->srv_lock, flags);
 796         while (!list_empty(&svc->srv_threads)) {
 797                 thread = list_entry(svc->srv_threads.next,
 798                                     struct ptlrpc_thread, t_link);
 799
 800                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 801                 ptlrpc_stop_thread(svc, thread);
 802                 spin_lock_irqsave(&svc->srv_lock, flags);
 803         }
 804
 805         spin_unlock_irqrestore(&svc->srv_lock, flags);
 806 }
 807
 808 /* @base_name should be 12 characters or less - 3 will be added on */
 809 int ptlrpc_start_n_threads(struct obd_device *dev, struct ptlrpc_service *svc,
 810                            int num_threads, char *base_name)
 811 {
 812         int i, rc = 0;
 813         ENTRY;
 814
 815         for (i = 0; i < num_threads; i++) {
 816                 char name[32];
 817                 sprintf(name, "%s_%02d", base_name, i);
 818                 rc = ptlrpc_start_thread(dev, svc, name);
 819                 if (rc) {
 820                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
 821                                i, rc);
 822                         ptlrpc_stop_all_threads(svc);
 823                 }
 824         }
 825         RETURN(rc);
 826 }
 827
 828 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 829                         char *name)
 830 {
 831         struct l_wait_info lwi = { 0 };
 832         struct ptlrpc_svc_data d;
 833         struct ptlrpc_thread *thread;
 834         unsigned long flags;
 835         int rc;
 836         ENTRY;
 837
 838         OBD_ALLOC(thread, sizeof(*thread));
 839         if (thread == NULL)
 840                 RETURN(-ENOMEM);
 841         init_waitqueue_head(&thread->t_ctl_waitq);
 842
 843         d.dev = dev;
 844         d.svc = svc;
 845         d.name = name;
 846         d.thread = thread;
 847
 848         spin_lock_irqsave(&svc->srv_lock, flags);
 849         list_add(&thread->t_link, &svc->srv_threads);
 850         spin_unlock_irqrestore(&svc->srv_lock, flags);
 851
 852         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 853          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 854          */
 855         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 856         if (rc < 0) {
 857                 CERROR("cannot start thread: %d\n", rc);
 858                 OBD_FREE(thread, sizeof(*thread));
 859                 RETURN(rc);
 860         }
 861         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 862
 863         RETURN(0);
 864 }
 865 #endif
 866
 867 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 868 {
 869         int                   i;
 870         int                   rc;
 871         unsigned long         flags;
 872         struct ptlrpc_srv_ni *srv_ni;
 873         struct l_wait_info    lwi;
 874         struct list_head     *tmp;
 875
 876         LASSERT(list_empty(&service->srv_threads));
 877
 878         spin_lock (&ptlrpc_all_services_lock);
 879         list_del_init (&service->srv_list);
 880         spin_unlock (&ptlrpc_all_services_lock);
 881
 882         ptlrpc_lprocfs_unregister_service(service);
 883
 884         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 885                 srv_ni = &service->srv_interfaces[i];
 886                 CDEBUG(D_NET, "%s: tearing down interface %s\n",
 887                        service->srv_name, srv_ni->sni_ni->pni_name);
 888
 889                 /* Unlink all the request buffers.  This forces a 'final'
 890                  * event with its 'unlink' flag set for each posted rqbd */
 891                 list_for_each(tmp, &srv_ni->sni_active_rqbds) {
 892                         struct ptlrpc_request_buffer_desc *rqbd =
 893                                 list_entry(tmp, struct ptlrpc_request_buffer_desc,
 894                                            rqbd_list);
 895
 896                         rc = PtlMDUnlink(rqbd->rqbd_md_h);
 897                         LASSERT (rc == PTL_OK || rc == PTL_MD_INVALID);
 898                 }
 899
 900                 /* Wait for the network to release any buffers it's
 901                  * currently filling */
 902                 for (;;) {
 903                         spin_lock_irqsave(&service->srv_lock, flags);
 904                         rc = srv_ni->sni_nrqbd_receiving;
 905                         spin_unlock_irqrestore(&service->srv_lock, flags);
 906
 907                         if (rc == 0)
 908                                 break;
 909
 910                         /* Network access will complete in finite time but
 911                          * the HUGE timeout lets us CWARN for visibility of
 912                          * sluggish NALs */
 913                         lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
 914                         rc = l_wait_event(service->srv_waitq,
 915                                           srv_ni->sni_nrqbd_receiving == 0,
 916                                           &lwi);
 917                         if (rc == -ETIMEDOUT)
 918                                 CWARN("Waiting for request buffers on "
 919                                       "service %s on interface %s ",
 920                                       service->srv_name, srv_ni->sni_ni->pni_name);
 921                 }
 922
 923                 /* schedule all outstanding replies to terminate them */
 924                 spin_lock_irqsave(&service->srv_lock, flags);
 925                 while (!list_empty(&srv_ni->sni_active_replies)) {
 926                         struct ptlrpc_reply_state *rs =
 927                                 list_entry(srv_ni->sni_active_replies.next,
 928                                            struct ptlrpc_reply_state,
 929                                            rs_list);
 930                         ptlrpc_schedule_difficult_reply(rs);
 931                 }
 932                 spin_unlock_irqrestore(&service->srv_lock, flags);
 933         }
 934
 935         /* purge the request queue.  NB No new replies (rqbds all unlinked)
 936          * and no service threads, so I'm the only thread noodling the
 937          * request queue now */
 938         while (!list_empty(&service->srv_request_queue)) {
 939                 struct ptlrpc_request *req =
 940                         list_entry(service->srv_request_queue.next,
 941                                    struct ptlrpc_request,
 942                                    rq_list);
 943
 944                 list_del(&req->rq_list);
 945                 service->srv_n_queued_reqs--;
 946                 service->srv_n_active_reqs++;
 947
 948                 ptlrpc_server_free_request(service, req);
 949         }
 950         LASSERT(service->srv_n_queued_reqs == 0);
 951         LASSERT(service->srv_n_active_reqs == 0);
 952
 953         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 954                 srv_ni = &service->srv_interfaces[i];
 955                 LASSERT(list_empty(&srv_ni->sni_active_rqbds));
 956         }
 957
 958         /* Now free all the request buffers since nothing references them
 959          * any more... */
 960         while (!list_empty(&service->srv_idle_rqbds)) {
 961                 struct ptlrpc_request_buffer_desc *rqbd =
 962                         list_entry(service->srv_idle_rqbds.next,
 963                                    struct ptlrpc_request_buffer_desc,
 964                                    rqbd_list);
 965
 966                 ptlrpc_free_rqbd(rqbd);
 967         }
 968
 969         /* wait for all outstanding replies to complete (they were
 970          * scheduled having been flagged to abort above) */
 971         while (atomic_read(&service->srv_outstanding_replies) != 0) {
 972                 struct l_wait_info lwi = LWI_TIMEOUT(10 * HZ, NULL, NULL);
 973
 974                 rc = l_wait_event(service->srv_waitq,
 975                                   !list_empty(&service->srv_reply_queue), &lwi);
 976                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 977
 978                 if (rc == 0) {
 979                         ptlrpc_server_handle_reply(service);
 980                         continue;
 981                 }
 982                 CWARN("Unexpectedly long timeout %p\n", service);
 983         }
 984
 985         OBD_FREE(service,
 986                  offsetof(struct ptlrpc_service,
 987                           srv_interfaces[ptlrpc_ninterfaces]));
 988         return 0;
 989 }