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, struct llog_create_locks *lcl)
 122 {
 123         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 124         LASSERT (rs != NULL);
 125         LASSERT (rs->rs_llog_locks == NULL);
 126
 127         rs->rs_llog_locks = lcl;
 128 }
 129
 130 void
 131 ptlrpc_require_repack(struct ptlrpc_request *req)
 132 {
 133         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 134         LASSERT (rs != NULL);
 135         rs->rs_difficult = 1;
 136 }
 137
 138 void
 139 ptlrpc_save_lock (struct ptlrpc_request *req,
 140                   struct lustre_handle *lock, int mode)
 141 {
 142         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 143         int                        idx;
 144
 145         if (!lock->cookie)
 146                 return;
 147
 148         LASSERT (rs != NULL);
 149         LASSERT (rs->rs_nlocks < RS_MAX_LOCKS);
 150
 151         idx = rs->rs_nlocks++;
 152         rs->rs_locks[idx] = *lock;
 153         rs->rs_modes[idx] = mode;
 154         rs->rs_difficult = 1;
 155 }
 156
 157 void
 158 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 159 {
 160         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 161
 162 #ifdef CONFIG_SMP
 163         LASSERT (spin_is_locked (&svc->srv_lock));
 164 #endif
 165         LASSERT (rs->rs_difficult);
 166         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 167
 168         if (rs->rs_scheduled)                   /* being set up or already notified */
 169                 return;
 170
 171         rs->rs_scheduled = 1;
 172         list_del (&rs->rs_list);
 173         list_add (&rs->rs_list, &svc->srv_reply_queue);
 174         wake_up (&svc->srv_waitq);
 175 }
 176
 177 void
 178 ptlrpc_commit_replies (struct obd_device *obd)
 179 {
 180         struct list_head   *tmp;
 181         struct list_head   *nxt;
 182         unsigned long       flags;
 183
 184         /* Find any replies that have been committed and get their service
 185          * to attend to complete them. */
 186
 187         /* CAVEAT EMPTOR: spinlock ordering!!! */
 188         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
 189
 190         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 191                 struct ptlrpc_reply_state *rs =
 192                         list_entry (tmp, struct ptlrpc_reply_state, rs_obd_list);
 193                 struct llog_create_locks *lcl = rs->rs_llog_locks;
 194
 195                 rs->rs_llog_locks = NULL;
 196                 LASSERT (rs->rs_difficult);
 197
 198                 if (rs->rs_transno <= obd->obd_last_committed) {
 199                         struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 200
 201                         spin_lock (&svc->srv_lock);
 202                         list_del_init (&rs->rs_obd_list);
 203                         ptlrpc_schedule_difficult_reply (rs);
 204                         spin_unlock (&svc->srv_lock);
 205
 206                         if (lcl != NULL)
 207                                 llog_create_lock_free(lcl);
 208                 }
 209         }
 210
 211         spin_unlock_irqrestore (&obd->obd_uncommitted_replies_lock, flags);
 212 }
 213
 214 static long
 215 timeval_sub(struct timeval *large, struct timeval *small)
 216 {
 217         return (large->tv_sec - small->tv_sec) * 1000000 +
 218                 (large->tv_usec - small->tv_usec);
 219 }
 220
 221 static int
 222 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 223 {
 224         struct ptlrpc_srv_ni              *srv_ni;
 225         struct ptlrpc_request_buffer_desc *rqbd;
 226         unsigned long                      flags;
 227         int                                rc;
 228
 229         spin_lock_irqsave(&svc->srv_lock, flags);
 230         if (list_empty (&svc->srv_idle_rqbds)) {
 231                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 232                 return (0);
 233         }
 234
 235         rqbd = list_entry(svc->srv_idle_rqbds.next,
 236                           struct ptlrpc_request_buffer_desc,
 237                           rqbd_list);
 238         list_del (&rqbd->rqbd_list);
 239
 240         /* assume we will post successfully */
 241         srv_ni = rqbd->rqbd_srv_ni;
 242         srv_ni->sni_nrqbd_receiving++;
 243         list_add (&rqbd->rqbd_list, &srv_ni->sni_active_rqbds);
 244
 245         spin_unlock_irqrestore(&svc->srv_lock, flags);
 246
 247         rc = ptlrpc_register_rqbd(rqbd);
 248         if (rc == 0)
 249                 return (1);
 250
 251         spin_lock_irqsave(&svc->srv_lock, flags);
 252
 253         srv_ni->sni_nrqbd_receiving--;
 254         list_del(&rqbd->rqbd_list);
 255         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 256
 257         if (srv_ni->sni_nrqbd_receiving == 0) {
 258                 /* This service is off-air on this interface because all
 259                  * its request buffers are busy.  Portals will have started
 260                  * dropping incoming requests until more buffers get
 261                  * posted */
 262                 CERROR("All %s %s request buffers busy\n",
 263                        svc->srv_name, srv_ni->sni_ni->pni_name);
 264         }
 265
 266         spin_unlock_irqrestore (&svc->srv_lock, flags);
 267
 268         return (-1);
 269 }
 270
 271 struct ptlrpc_service *
 272 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size,
 273                 int req_portal, int rep_portal,
 274                 svc_handler_t handler, char *name,
 275                 struct proc_dir_entry *proc_entry)
 276 {
 277         int                                i;
 278         int                                j;
 279         int                                ssize;
 280         struct ptlrpc_service             *service;
 281         struct ptlrpc_srv_ni              *srv_ni;
 282         struct ptlrpc_request_buffer_desc *rqbd;
 283         ENTRY;
 284
 285         LASSERT (ptlrpc_ninterfaces > 0);
 286         LASSERT (nbufs > 0);
 287         LASSERT (bufsize >= max_req_size);
 288
 289         ssize = offsetof (struct ptlrpc_service,
 290                           srv_interfaces[ptlrpc_ninterfaces]);
 291         OBD_ALLOC(service, ssize);
 292         if (service == NULL)
 293                 RETURN(NULL);
 294
 295         service->srv_name = name;
 296         spin_lock_init(&service->srv_lock);
 297         INIT_LIST_HEAD(&service->srv_threads);
 298         init_waitqueue_head(&service->srv_waitq);
 299
 300         service->srv_max_req_size = max_req_size;
 301         service->srv_buf_size = bufsize;
 302         service->srv_rep_portal = rep_portal;
 303         service->srv_req_portal = req_portal;
 304         service->srv_handler = handler;
 305
 306         INIT_LIST_HEAD(&service->srv_request_queue);
 307         INIT_LIST_HEAD(&service->srv_idle_rqbds);
 308         INIT_LIST_HEAD(&service->srv_reply_queue);
 309
 310         /* First initialise enough for early teardown */
 311         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 312                 srv_ni = &service->srv_interfaces[i];
 313
 314                 srv_ni->sni_service = service;
 315                 srv_ni->sni_ni = &ptlrpc_interfaces[i];
 316                 INIT_LIST_HEAD(&srv_ni->sni_active_rqbds);
 317                 INIT_LIST_HEAD(&srv_ni->sni_active_replies);
 318         }
 319
 320         spin_lock (&ptlrpc_all_services_lock);
 321         list_add (&service->srv_list, &ptlrpc_all_services);
 322         spin_unlock (&ptlrpc_all_services_lock);
 323
 324         /* Now allocate the request buffers, assuming all interfaces require
 325          * the same number. */
 326         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 327                 srv_ni = &service->srv_interfaces[i];
 328                 CDEBUG (D_NET, "%s: initialising interface %s\n", name,
 329                         srv_ni->sni_ni->pni_name);
 330
 331                 for (j = 0; j < nbufs; j++) {
 332                         rqbd = ptlrpc_alloc_rqbd (srv_ni);
 333
 334                         if (rqbd == NULL) {
 335                                 CERROR ("%s.%d: Can't allocate request %d "
 336                                         "on %s\n", name, i, j,
 337                                         srv_ni->sni_ni->pni_name);
 338                                 GOTO(failed, NULL);
 339                         }
 340
 341                         /* We shouldn't be under memory pressure at
 342                          * startup, so fail if we can't post all our
 343                          * buffers at this time. */
 344                         if (ptlrpc_server_post_idle_rqbds(service) <= 0)
 345                                 GOTO(failed, NULL);
 346                 }
 347         }
 348
 349         if (proc_entry != NULL)
 350                 ptlrpc_lprocfs_register_service(proc_entry, service);
 351
 352         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 353                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 354
 355         RETURN(service);
 356 failed:
 357         ptlrpc_unregister_service(service);
 358         return NULL;
 359 }
 360
 361 static void
 362 ptlrpc_server_free_request(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 363 {
 364         unsigned long  flags;
 365         int            refcount;
 366
 367         spin_lock_irqsave(&svc->srv_lock, flags);
 368         svc->srv_n_active_reqs--;
 369         refcount = --(req->rq_rqbd->rqbd_refcount);
 370         if (refcount == 0) {
 371                 /* request buffer is now idle */
 372                 list_del(&req->rq_rqbd->rqbd_list);
 373                 list_add_tail(&req->rq_rqbd->rqbd_list,
 374                               &svc->srv_idle_rqbds);
 375         }
 376         spin_unlock_irqrestore(&svc->srv_lock, flags);
 377
 378         ptlrpc_free_server_req(req);
 379 }
 380
 381 static int
 382 ptlrpc_server_handle_request (struct ptlrpc_service *svc)
 383 {
 384         struct obd_export     *export = NULL;
 385         struct ptlrpc_request *request;
 386         unsigned long          flags;
 387         struct timeval         work_start;
 388         struct timeval         work_end;
 389         long                   timediff;
 390         int                    rc;
 391         char                   str[PTL_NALFMT_SIZE];
 392         ENTRY;
 393
 394         spin_lock_irqsave (&svc->srv_lock, flags);
 395         if (list_empty (&svc->srv_request_queue) ||
 396             (svc->srv_n_difficult_replies != 0 &&
 397              svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
 398                 /* If all the other threads are handling requests, I must
 399                  * remain free to handle any 'difficult' reply that might
 400                  * block them */
 401                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 402                 RETURN(0);
 403         }
 404
 405         request = list_entry (svc->srv_request_queue.next,
 406                               struct ptlrpc_request, rq_list);
 407         list_del_init (&request->rq_list);
 408         svc->srv_n_queued_reqs--;
 409         svc->srv_n_active_reqs++;
 410
 411         spin_unlock_irqrestore (&svc->srv_lock, flags);
 412
 413         do_gettimeofday(&work_start);
 414         timediff = timeval_sub(&work_start, &request->rq_arrival_time);
 415         if (svc->srv_stats != NULL) {
 416                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 417                                     timediff);
 418                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 419                                     svc->srv_n_queued_reqs);
 420                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 421                                     svc->srv_n_active_reqs);
 422         }
 423
 424 #if SWAB_PARANOIA
 425         /* Clear request swab mask; this is a new request */
 426         request->rq_req_swab_mask = 0;
 427 #endif
 428         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 429         if (rc != 0) {
 430                 CERROR ("error unpacking request: ptl %d from %s"
 431                         " xid "LPU64"\n", svc->srv_req_portal,
 432                         ptlrpc_peernid2str(&request->rq_peer, str),
 433                        request->rq_xid);
 434                 goto out;
 435         }
 436
 437         rc = -EINVAL;
 438         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 439                 CERROR("wrong packet type received (type=%u) from %s\n",
 440                        request->rq_reqmsg->type,
 441                        ptlrpc_peernid2str(&request->rq_peer, str));
 442                 goto out;
 443         }
 444
 445         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 446
 447         /* Discard requests queued for longer than my timeout.  If the
 448          * client's timeout is similar to mine, she'll be timing out this
 449          * REQ anyway (bug 1502) */
 450         if (timediff / 1000000 > (long)obd_timeout) {
 451                 CERROR("Dropping timed-out request from %s: %ld seconds old\n",
 452                        ptlrpc_peernid2str(&request->rq_peer, str),
 453                        timediff / 1000000);
 454                 goto out;
 455         }
 456
 457         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 458
 459         if (request->rq_export) {
 460                 if (request->rq_reqmsg->conn_cnt <
 461                     request->rq_export->exp_conn_cnt) {
 462                         DEBUG_REQ(D_ERROR, request,
 463                                   "DROPPING req from old connection %d < %d",
 464                                   request->rq_reqmsg->conn_cnt,
 465                                   request->rq_export->exp_conn_cnt);
 466                         goto put_conn;
 467                 }
 468
 469                 export = class_export_rpc_get(request->rq_export);
 470                 request->rq_export->exp_last_request_time =
 471                         LTIME_S(CURRENT_TIME);
 472         }
 473
 474         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 475                "%s:%s+%d:%d:"LPU64":%s:%s:%d\n", current->comm,
 476                (request->rq_export ?
 477                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 478                (request->rq_export ?
 479                 atomic_read(&request->rq_export->exp_refcount) : -99),
 480                request->rq_reqmsg->status, request->rq_xid,
 481                request->rq_peer.peer_ni->pni_name,
 482                ptlrpc_peernid2str(&request->rq_peer, str),
 483                request->rq_reqmsg->opc);
 484
 485         rc = svc->srv_handler(request);
 486         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 487                "%s:%s+%d:%d:"LPU64":%s:%s:%d\n", current->comm,
 488                (request->rq_export ?
 489                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 490                (request->rq_export ?
 491                 atomic_read(&request->rq_export->exp_refcount) : -99),
 492                request->rq_reqmsg->status, request->rq_xid,
 493                request->rq_peer.peer_ni->pni_name,
 494                ptlrpc_peernid2str(&request->rq_peer, str),
 495                request->rq_reqmsg->opc);
 496
 497         if (export != NULL)
 498                 class_export_rpc_put(export);
 499
 500 put_conn:
 501         if (request->rq_export != NULL)
 502                 class_export_put(request->rq_export);
 503
 504  out:
 505         do_gettimeofday(&work_end);
 506
 507         timediff = timeval_sub(&work_end, &work_start);
 508
 509         CDEBUG((timediff / 1000000 > (long)obd_timeout) ? D_ERROR : D_HA,
 510                "request "LPU64" opc %u from NID %s processed in %ldus "
 511                "(%ldus total)\n", request->rq_xid, request->rq_reqmsg->opc,
 512                ptlrpc_peernid2str(&request->rq_peer, str),
 513                timediff, timeval_sub(&work_end, &request->rq_arrival_time));
 514
 515         if (svc->srv_stats != NULL) {
 516                 int opc = opcode_offset(request->rq_reqmsg->opc);
 517                 if (opc > 0) {
 518                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 519                         lprocfs_counter_add(svc->srv_stats,
 520                                             opc + PTLRPC_LAST_CNTR,
 521                                             timediff);
 522                 }
 523         }
 524
 525         ptlrpc_server_free_request(svc, request);
 526
 527         RETURN(1);
 528 }
 529
 530 static int
 531 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 532 {
 533         struct ptlrpc_reply_state *rs;
 534         unsigned long              flags;
 535         struct obd_export         *exp;
 536         struct obd_device         *obd;
 537         struct llog_create_locks  *lcl;
 538         int                        nlocks;
 539         int                        been_handled;
 540         ENTRY;
 541
 542         spin_lock_irqsave (&svc->srv_lock, flags);
 543         if (list_empty (&svc->srv_reply_queue)) {
 544                 spin_unlock_irqrestore (&svc->srv_lock, flags);
 545                 RETURN(0);
 546         }
 547
 548         rs = list_entry (svc->srv_reply_queue.next,
 549                          struct ptlrpc_reply_state, rs_list);
 550
 551         exp = rs->rs_export;
 552         obd = exp->exp_obd;
 553
 554         LASSERT (rs->rs_difficult);
 555         LASSERT (rs->rs_scheduled);
 556
 557         list_del_init (&rs->rs_list);
 558
 559         /* Disengage from notifiers carefully (lock ordering!) */
 560         spin_unlock(&svc->srv_lock);
 561
 562         spin_lock (&obd->obd_uncommitted_replies_lock);
 563         /* Noop if removed already */
 564         list_del_init (&rs->rs_obd_list);
 565         spin_unlock (&obd->obd_uncommitted_replies_lock);
 566
 567         spin_lock (&exp->exp_lock);
 568         /* Noop if removed already */
 569         list_del_init (&rs->rs_exp_list);
 570         spin_unlock (&exp->exp_lock);
 571
 572         spin_lock(&svc->srv_lock);
 573
 574         been_handled = rs->rs_handled;
 575         rs->rs_handled = 1;
 576
 577         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 578         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 579
 580         lcl = rs->rs_llog_locks;
 581         rs->rs_llog_locks = NULL;
 582
 583         if (nlocks == 0 && !been_handled) {
 584                 /* If we see this, we should already have seen the warning
 585                  * in mds_steal_ack_locks()  */
 586 #if 0
 587                 char str[PTL_NALFMT_SIZE];
 588                 /* CMD may ask to save request with no DLM locks -bzzz */
 589                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 590                       " o%d NID %s\n",
 591                       rs,
 592                       rs->rs_xid, rs->rs_transno,
 593                       rs->rs_msg.opc,
 594                       ptlrpc_peernid2str(&exp->exp_connection->c_peer, str));
 595 #endif
 596         }
 597
 598         if ((!been_handled && rs->rs_on_net) ||
 599             nlocks > 0 || lcl != NULL) {
 600                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 601
 602                 if (!been_handled && rs->rs_on_net) {
 603                         PtlMDUnlink(rs->rs_md_h);
 604                         /* Ignore return code; we're racing with
 605                          * completion... */
 606                 }
 607
 608                 while (nlocks-- > 0)
 609                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 610                                          rs->rs_modes[nlocks]);
 611
 612                 if (lcl != NULL)
 613                         llog_create_lock_free(lcl);
 614
 615                 spin_lock_irqsave(&svc->srv_lock, flags);
 616         }
 617
 618         rs->rs_scheduled = 0;
 619
 620         if (!rs->rs_on_net) {
 621                 /* Off the net */
 622                 svc->srv_n_difficult_replies--;
 623                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 624
 625                 class_export_put (exp);
 626                 rs->rs_export = NULL;
 627                 lustre_free_reply_state (rs);
 628                 atomic_dec (&svc->srv_outstanding_replies);
 629                 RETURN(1);
 630         }
 631
 632         /* still on the net; callback will schedule */
 633         spin_unlock_irqrestore (&svc->srv_lock, flags);
 634         RETURN(1);
 635 }
 636
 637 #ifndef __KERNEL__
 638 /* FIXME make use of timeout later */
 639 int
 640 liblustre_check_services (void *arg)
 641 {
 642         int  did_something = 0;
 643         int  rc;
 644         struct list_head *tmp, *nxt;
 645         ENTRY;
 646
 647         /* I'm relying on being single threaded, not to have to lock
 648          * ptlrpc_all_services etc */
 649         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 650                 struct ptlrpc_service *svc =
 651                         list_entry (tmp, struct ptlrpc_service, srv_list);
 652
 653                 if (svc->srv_nthreads != 0)     /* I've recursed */
 654                         continue;
 655
 656                 /* service threads can block for bulk, so this limits us
 657                  * (arbitrarily) to recursing 1 stack frame per service.
 658                  * Note that the problem with recursion is that we have to
 659                  * unwind completely before our caller can resume. */
 660
 661                 svc->srv_nthreads++;
 662
 663                 do {
 664                         rc = ptlrpc_server_handle_reply(svc);
 665                         rc |= ptlrpc_server_handle_request(svc);
 666                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 667                         did_something |= rc;
 668                 } while (rc);
 669
 670                 svc->srv_nthreads--;
 671         }
 672
 673         RETURN(did_something);
 674 }
 675
 676 #else /* __KERNEL__ */
 677
 678 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 679 void ptlrpc_daemonize(void)
 680 {
 681         exit_mm(current);
 682         lustre_daemonize_helper();
 683         exit_files(current);
 684         reparent_to_init();
 685 }
 686
 687 static int
 688 ptlrpc_retry_rqbds(void *arg)
 689 {
 690         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 691
 692         svc->srv_rqbd_timeout = 0;
 693         return (-ETIMEDOUT);
 694 }
 695
 696 static int ptlrpc_main(void *arg)
 697 {
 698         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 699         struct ptlrpc_service  *svc = data->svc;
 700         struct ptlrpc_thread   *thread = data->thread;
 701         unsigned long           flags;
 702         ENTRY;
 703
 704         lock_kernel();
 705         ptlrpc_daemonize();
 706
 707         SIGNAL_MASK_LOCK(current, flags);
 708         sigfillset(&current->blocked);
 709         RECALC_SIGPENDING;
 710         SIGNAL_MASK_UNLOCK(current, flags);
 711
 712         LASSERTF(strlen(data->name) < sizeof(current->comm),
 713                  "name %d > len %d\n",
 714                  (int)strlen(data->name), (int)sizeof(current->comm));
 715         THREAD_NAME(current->comm, sizeof(current->comm) - 1, "%s", data->name);
 716
 717         unlock_kernel();
 718
 719         /* Record that the thread is running */
 720         thread->t_flags = SVC_RUNNING;
 721         wake_up(&thread->t_ctl_waitq);
 722
 723         spin_lock_irqsave(&svc->srv_lock, flags);
 724         svc->srv_nthreads++;
 725         spin_unlock_irqrestore(&svc->srv_lock, flags);
 726
 727         /* XXX maintain a list of all managed devices: insert here */
 728
 729         while ((thread->t_flags & SVC_STOPPING) == 0 ||
 730                svc->srv_n_difficult_replies != 0) {
 731                 /* Don't exit while there are replies to be handled */
 732                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
 733                                                      ptlrpc_retry_rqbds, svc);
 734
 735                 l_wait_event_exclusive (svc->srv_waitq,
 736                               ((thread->t_flags & SVC_STOPPING) != 0 &&
 737                                svc->srv_n_difficult_replies == 0) ||
 738                               (!list_empty(&svc->srv_idle_rqbds) &&
 739                                svc->srv_rqbd_timeout == 0) ||
 740                               !list_empty (&svc->srv_reply_queue) ||
 741                               (!list_empty (&svc->srv_request_queue) &&
 742                                (svc->srv_n_difficult_replies == 0 ||
 743                                 svc->srv_n_active_reqs <
 744                                 (svc->srv_nthreads - 1))),
 745                               &lwi);
 746
 747                 if (!list_empty (&svc->srv_reply_queue))
 748                         ptlrpc_server_handle_reply (svc);
 749
 750                 /* only handle requests if there are no difficult replies
 751                  * outstanding, or I'm not the last thread handling
 752                  * requests */
 753                 if (!list_empty (&svc->srv_request_queue) &&
 754                     (svc->srv_n_difficult_replies == 0 ||
 755                      svc->srv_n_active_reqs < (svc->srv_nthreads - 1)))
 756                         ptlrpc_server_handle_request (svc);
 757
 758                 if (!list_empty(&svc->srv_idle_rqbds) &&
 759                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
 760                         /* I just failed to repost request buffers.  Wait
 761                          * for a timeout (unless something else happens)
 762                          * before I try again */
 763                         svc->srv_rqbd_timeout = HZ/10;
 764                 }
 765         }
 766
 767         spin_lock_irqsave(&svc->srv_lock, flags);
 768
 769         svc->srv_nthreads--;                    /* must know immediately */
 770         thread->t_flags = SVC_STOPPED;
 771         wake_up(&thread->t_ctl_waitq);
 772
 773         spin_unlock_irqrestore(&svc->srv_lock, flags);
 774
 775         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 776         return 0;
 777 }
 778
 779 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 780                                struct ptlrpc_thread *thread)
 781 {
 782         struct l_wait_info lwi = { 0 };
 783         unsigned long      flags;
 784
 785         spin_lock_irqsave(&svc->srv_lock, flags);
 786         thread->t_flags = SVC_STOPPING;
 787         spin_unlock_irqrestore(&svc->srv_lock, flags);
 788
 789         wake_up_all(&svc->srv_waitq);
 790         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 791                      &lwi);
 792
 793         spin_lock_irqsave(&svc->srv_lock, flags);
 794         list_del(&thread->t_link);
 795         spin_unlock_irqrestore(&svc->srv_lock, flags);
 796
 797         OBD_FREE(thread, sizeof(*thread));
 798 }
 799
 800 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 801 {
 802         unsigned long flags;
 803         struct ptlrpc_thread *thread;
 804
 805         spin_lock_irqsave(&svc->srv_lock, flags);
 806         while (!list_empty(&svc->srv_threads)) {
 807                 thread = list_entry(svc->srv_threads.next,
 808                                     struct ptlrpc_thread, t_link);
 809
 810                 spin_unlock_irqrestore(&svc->srv_lock, flags);
 811                 ptlrpc_stop_thread(svc, thread);
 812                 spin_lock_irqsave(&svc->srv_lock, flags);
 813         }
 814
 815         spin_unlock_irqrestore(&svc->srv_lock, flags);
 816 }
 817
 818 /* @base_name should be 12 characters or less - 3 will be added on */
 819 int ptlrpc_start_n_threads(struct obd_device *dev, struct ptlrpc_service *svc,
 820                            int num_threads, char *base_name)
 821 {
 822         int i, rc = 0;
 823         ENTRY;
 824
 825         for (i = 0; i < num_threads; i++) {
 826                 char name[32];
 827                 sprintf(name, "%s_%02d", base_name, i);
 828                 rc = ptlrpc_start_thread(dev, svc, name);
 829                 if (rc) {
 830                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
 831                                i, rc);
 832                         ptlrpc_stop_all_threads(svc);
 833                 }
 834         }
 835         RETURN(rc);
 836 }
 837
 838 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 839                         char *name)
 840 {
 841         struct l_wait_info lwi = { 0 };
 842         struct ptlrpc_svc_data d;
 843         struct ptlrpc_thread *thread;
 844         unsigned long flags;
 845         int rc;
 846         ENTRY;
 847
 848         OBD_ALLOC(thread, sizeof(*thread));
 849         if (thread == NULL)
 850                 RETURN(-ENOMEM);
 851         init_waitqueue_head(&thread->t_ctl_waitq);
 852
 853         d.dev = dev;
 854         d.svc = svc;
 855         d.name = name;
 856         d.thread = thread;
 857
 858         spin_lock_irqsave(&svc->srv_lock, flags);
 859         list_add(&thread->t_link, &svc->srv_threads);
 860         spin_unlock_irqrestore(&svc->srv_lock, flags);
 861
 862         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 863          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 864          */
 865         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 866         if (rc < 0) {
 867                 CERROR("cannot start thread: %d\n", rc);
 868                 OBD_FREE(thread, sizeof(*thread));
 869                 RETURN(rc);
 870         }
 871         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 872
 873         RETURN(0);
 874 }
 875 #endif
 876
 877 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 878 {
 879         int                   i;
 880         int                   rc;
 881         unsigned long         flags;
 882         struct ptlrpc_srv_ni *srv_ni;
 883         struct l_wait_info    lwi;
 884         struct list_head     *tmp;
 885
 886         LASSERT(list_empty(&service->srv_threads));
 887
 888         spin_lock (&ptlrpc_all_services_lock);
 889         list_del_init (&service->srv_list);
 890         spin_unlock (&ptlrpc_all_services_lock);
 891
 892         ptlrpc_lprocfs_unregister_service(service);
 893
 894         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 895                 srv_ni = &service->srv_interfaces[i];
 896                 CDEBUG(D_NET, "%s: tearing down interface %s\n",
 897                        service->srv_name, srv_ni->sni_ni->pni_name);
 898
 899                 /* Unlink all the request buffers.  This forces a 'final'
 900                  * event with its 'unlink' flag set for each posted rqbd */
 901                 list_for_each(tmp, &srv_ni->sni_active_rqbds) {
 902                         struct ptlrpc_request_buffer_desc *rqbd =
 903                                 list_entry(tmp, struct ptlrpc_request_buffer_desc,
 904                                            rqbd_list);
 905
 906                         rc = PtlMDUnlink(rqbd->rqbd_md_h);
 907                         LASSERT (rc == PTL_OK || rc == PTL_MD_INVALID);
 908                 }
 909
 910                 /* Wait for the network to release any buffers it's
 911                  * currently filling */
 912                 for (;;) {
 913                         spin_lock_irqsave(&service->srv_lock, flags);
 914                         rc = srv_ni->sni_nrqbd_receiving;
 915                         spin_unlock_irqrestore(&service->srv_lock, flags);
 916
 917                         if (rc == 0)
 918                                 break;
 919
 920                         /* Network access will complete in finite time but
 921                          * the HUGE timeout lets us CWARN for visibility of
 922                          * sluggish NALs */
 923                         lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
 924                         rc = l_wait_event(service->srv_waitq,
 925                                           srv_ni->sni_nrqbd_receiving == 0,
 926                                           &lwi);
 927                         if (rc == -ETIMEDOUT)
 928                                 CWARN("Waiting for request buffers on "
 929                                       "service %s on interface %s ",
 930                                       service->srv_name, srv_ni->sni_ni->pni_name);
 931                 }
 932
 933                 /* schedule all outstanding replies to terminate them */
 934                 spin_lock_irqsave(&service->srv_lock, flags);
 935                 while (!list_empty(&srv_ni->sni_active_replies)) {
 936                         struct ptlrpc_reply_state *rs =
 937                                 list_entry(srv_ni->sni_active_replies.next,
 938                                            struct ptlrpc_reply_state,
 939                                            rs_list);
 940                         ptlrpc_schedule_difficult_reply(rs);
 941                 }
 942                 spin_unlock_irqrestore(&service->srv_lock, flags);
 943         }
 944
 945         /* purge the request queue.  NB No new replies (rqbds all unlinked)
 946          * and no service threads, so I'm the only thread noodling the
 947          * request queue now */
 948         while (!list_empty(&service->srv_request_queue)) {
 949                 struct ptlrpc_request *req =
 950                         list_entry(service->srv_request_queue.next,
 951                                    struct ptlrpc_request,
 952                                    rq_list);
 953
 954                 list_del(&req->rq_list);
 955                 service->srv_n_queued_reqs--;
 956                 service->srv_n_active_reqs++;
 957
 958                 ptlrpc_server_free_request(service, req);
 959         }
 960         LASSERT(service->srv_n_queued_reqs == 0);
 961         LASSERT(service->srv_n_active_reqs == 0);
 962
 963         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 964                 srv_ni = &service->srv_interfaces[i];
 965                 LASSERT(list_empty(&srv_ni->sni_active_rqbds));
 966         }
 967
 968         /* Now free all the request buffers since nothing references them
 969          * any more... */
 970         while (!list_empty(&service->srv_idle_rqbds)) {
 971                 struct ptlrpc_request_buffer_desc *rqbd =
 972                         list_entry(service->srv_idle_rqbds.next,
 973                                    struct ptlrpc_request_buffer_desc,
 974                                    rqbd_list);
 975
 976                 ptlrpc_free_rqbd(rqbd);
 977         }
 978
 979         /* wait for all outstanding replies to complete (they were
 980          * scheduled having been flagged to abort above) */
 981         while (atomic_read(&service->srv_outstanding_replies) != 0) {
 982                 struct l_wait_info lwi = LWI_TIMEOUT(10 * HZ, NULL, NULL);
 983
 984                 rc = l_wait_event(service->srv_waitq,
 985                                   !list_empty(&service->srv_reply_queue), &lwi);
 986                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 987
 988                 if (rc == 0) {
 989                         ptlrpc_server_handle_reply(service);
 990                         continue;
 991                 }
 992                 CWARN("Unexpectedly long timeout %p\n", service);
 993         }
 994
 995         OBD_FREE(service,
 996                  offsetof(struct ptlrpc_service,
 997                           srv_interfaces[ptlrpc_ninterfaces]));
 998         return 0;
 999 }