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 extern int request_in_callback(ptl_event_t *ev);
  35
  36 static int ptlrpc_check_event(struct ptlrpc_service *svc,
  37                               struct ptlrpc_thread *thread, ptl_event_t *event)
  38 {
  39         struct ptlrpc_srv_ni *srv_ni;
  40         int i, idx, rc;
  41         ENTRY;
  42
  43         spin_lock(&svc->srv_lock);
  44
  45         if (thread->t_flags & SVC_STOPPING)
  46                 GOTO(out, rc = 1);
  47
  48         LASSERT ((thread->t_flags & SVC_EVENT) == 0);
  49         LASSERT (ptlrpc_ninterfaces > 0);
  50
  51         for (i = 0; i < ptlrpc_ninterfaces; i++) {
  52                 idx = (svc->srv_interface_rover + i) % ptlrpc_ninterfaces;
  53                 srv_ni = &svc->srv_interfaces[idx];
  54
  55                 LASSERT (!PtlHandleEqual (srv_ni->sni_eq_h, PTL_HANDLE_NONE));
  56
  57                 rc = PtlEQGet(srv_ni->sni_eq_h, event);
  58                 switch (rc) {
  59                 case PTL_OK:
  60                         /* next time start with the next interface */
  61                         svc->srv_interface_rover = (idx+1) % ptlrpc_ninterfaces;
  62                         thread->t_flags |= SVC_EVENT;
  63                         GOTO(out, rc = 1);
  64
  65                 case PTL_EQ_EMPTY:
  66                         continue;
  67
  68                 case PTL_EQ_DROPPED:
  69                         CWARN("Event queue overflow (bug 2125): timeouts will "
  70                               "follow.\n");
  71                         continue;
  72
  73                 default:
  74                         CERROR("BUG: PtlEQGet returned %d\n", rc);
  75                         LBUG();
  76                 }
  77         }
  78         rc = 0;
  79         EXIT;
  80  out:
  81         spin_unlock(&svc->srv_lock);
  82         return rc;
  83 }
  84
  85 struct ptlrpc_service * ptlrpc_init_svc(__u32 nevents, __u32 nbufs,
  86                                         __u32 bufsize, __u32 max_req_size,
  87                                         int req_portal, int rep_portal,
  88                                         svc_handler_t handler, char *name,
  89                                         struct proc_dir_entry *proc_entry)
  90 {
  91         int i, j, ssize, rc;
  92         struct ptlrpc_service *service;
  93         struct ptlrpc_srv_ni  *srv_ni;
  94         ENTRY;
  95
  96         LASSERT (ptlrpc_ninterfaces > 0);
  97
  98         ssize = offsetof (struct ptlrpc_service,
  99                           srv_interfaces[ptlrpc_ninterfaces]);
 100         OBD_ALLOC(service, ssize);
 101         if (service == NULL)
 102                 RETURN(NULL);
 103
 104         service->srv_name = name;
 105         spin_lock_init(&service->srv_lock);
 106         INIT_LIST_HEAD(&service->srv_threads);
 107         init_waitqueue_head(&service->srv_waitq);
 108
 109         service->srv_max_req_size = max_req_size;
 110         service->srv_buf_size = bufsize;
 111
 112         service->srv_rep_portal = rep_portal;
 113         service->srv_req_portal = req_portal;
 114         service->srv_handler = handler;
 115         service->srv_interface_rover = 0;
 116
 117         /* First initialise enough for early teardown */
 118         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 119                 srv_ni = &service->srv_interfaces[i];
 120
 121                 srv_ni->sni_service = service;
 122                 srv_ni->sni_ni = &ptlrpc_interfaces[i];
 123                 srv_ni->sni_eq_h = PTL_HANDLE_NONE;
 124                 INIT_LIST_HEAD(&srv_ni->sni_rqbds);
 125                 srv_ni->sni_nrqbds = 0;
 126                 atomic_set(&srv_ni->sni_nrqbds_receiving, 0);
 127         }
 128
 129         /* Now allocate the event queue and request buffers, assuming all
 130          * interfaces require the same level of buffering. */
 131         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 132                 srv_ni = &service->srv_interfaces[i];
 133                 CDEBUG (D_NET, "%s: initialising interface %s\n", name,
 134                         srv_ni->sni_ni->pni_name);
 135
 136                 rc = PtlEQAlloc(srv_ni->sni_ni->pni_ni_h, nevents,
 137                                 request_in_callback, &(srv_ni->sni_eq_h));
 138                 if (rc != PTL_OK) {
 139                         CERROR("%s.%d: PtlEQAlloc on %s failed: %d\n",
 140                                name, i, srv_ni->sni_ni->pni_name, rc);
 141                         GOTO (failed, NULL);
 142                 }
 143
 144                 for (j = 0; j < nbufs; j++) {
 145                         struct ptlrpc_request_buffer_desc *rqbd;
 146
 147                         OBD_ALLOC_WAIT(rqbd, sizeof(*rqbd));
 148                         if (rqbd == NULL) {
 149                                 CERROR ("%s.%d: Can't allocate request "
 150                                         "descriptor %d on %s\n",
 151                                         name, i, srv_ni->sni_nrqbds,
 152                                         srv_ni->sni_ni->pni_name);
 153                                 GOTO(failed, NULL);
 154                         }
 155
 156                         rqbd->rqbd_srv_ni = srv_ni;
 157                         rqbd->rqbd_me_h = PTL_HANDLE_NONE;
 158                         atomic_set(&rqbd->rqbd_refcount, 0);
 159
 160                         OBD_ALLOC_WAIT(rqbd->rqbd_buffer, service->srv_buf_size);
 161                         if (rqbd->rqbd_buffer == NULL) {
 162                                 CERROR ("%s.%d: Can't allocate request "
 163                                         "buffer %d on %s\n",
 164                                         name, i, srv_ni->sni_nrqbds,
 165                                         srv_ni->sni_ni->pni_name);
 166                                 OBD_FREE(rqbd, sizeof(*rqbd));
 167                                 GOTO(failed, NULL);
 168                         }
 169                         list_add(&rqbd->rqbd_list, &srv_ni->sni_rqbds);
 170                         srv_ni->sni_nrqbds++;
 171
 172                         ptlrpc_link_svc_me(rqbd);
 173                 }
 174         }
 175
 176         if (proc_entry != NULL)
 177                 ptlrpc_lprocfs_register_service(proc_entry, service);
 178
 179         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 180                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 181
 182         RETURN(service);
 183 failed:
 184         ptlrpc_unregister_service(service);
 185         return NULL;
 186 }
 187
 188 static int handle_incoming_request(struct obd_device *obddev,
 189                                    struct ptlrpc_service *svc,
 190                                    ptl_event_t *event,
 191                                    struct ptlrpc_request *request)
 192 {
 193         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 194         int rc;
 195
 196         /* FIXME: If we move to an event-driven model, we should put the request
 197          * on the stack of mds_handle instead. */
 198
 199         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 200         LASSERT ((event->mem_desc.options & (PTL_MD_IOV | PTL_MD_KIOV)) == 0);
 201         LASSERT (rqbd->rqbd_srv_ni->sni_service == svc);
 202         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 203         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 204
 205         memset(request, 0, sizeof(*request));
 206         spin_lock_init (&request->rq_lock);
 207         INIT_LIST_HEAD(&request->rq_list);
 208         request->rq_svc = svc;
 209         request->rq_xid = event->match_bits;
 210         request->rq_reqmsg = event->mem_desc.start + event->offset;
 211         request->rq_reqlen = event->mlength;
 212
 213 #if SWAB_PARANOIA
 214         /* Clear request swab mask; this is a new request */
 215         request->rq_req_swab_mask = 0;
 216 #endif
 217         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 218         if (rc != 0) {
 219                 CERROR ("error unpacking request: ptl %d from "LPX64
 220                         " xid "LPU64"\n", svc->srv_req_portal,
 221                        event->initiator.nid, request->rq_xid);
 222                 goto out;
 223         }
 224         rc = -EINVAL;
 225         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 226                 CERROR("wrong packet type received (type=%u)\n",
 227                        request->rq_reqmsg->type);
 228                 goto out;
 229         }
 230
 231         CDEBUG(D_NET, "got req "LPD64" (md: %p + %d)\n", request->rq_xid,
 232                event->mem_desc.start, event->offset);
 233
 234         request->rq_peer.peer_nid = event->initiator.nid;
 235         request->rq_peer.peer_ni = rqbd->rqbd_srv_ni->sni_ni;
 236
 237         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 238
 239         if (request->rq_export) {
 240                 if (request->rq_reqmsg->conn_cnt <
 241                     request->rq_export->exp_conn_cnt) {
 242                         DEBUG_REQ(D_ERROR, request,
 243                                   "DROPPING req from old connection %d < %d",
 244                                   request->rq_reqmsg->conn_cnt,
 245                                   request->rq_export->exp_conn_cnt);
 246                         goto put_conn;
 247                 }
 248
 249                 request->rq_export->exp_last_request_time =
 250                         LTIME_S(CURRENT_TIME);
 251         }
 252
 253         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 254                "%s:%s+%d:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 255                (request->rq_export ?
 256                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 257                (request->rq_export ?
 258                 atomic_read(&request->rq_export->exp_refcount) : -99),
 259                request->rq_reqmsg->status, request->rq_xid,
 260                rqbd->rqbd_srv_ni->sni_ni->pni_name, event->initiator.nid,
 261                request->rq_reqmsg->opc);
 262
 263         rc = svc->srv_handler(request);
 264         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:ni:nid:opc "
 265                "%s:%s+%d:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 266                (request->rq_export ?
 267                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 268                (request->rq_export ?
 269                 atomic_read(&request->rq_export->exp_refcount) : -99),
 270                request->rq_reqmsg->status, request->rq_xid,
 271                rqbd->rqbd_srv_ni->sni_ni->pni_name, event->initiator.nid,
 272                request->rq_reqmsg->opc);
 273
 274 put_conn:
 275         if (request->rq_export != NULL)
 276                 class_export_put(request->rq_export);
 277
 278  out:
 279         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 280                 ptlrpc_link_svc_me (rqbd);
 281
 282         return rc;
 283 }
 284
 285 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 286 void ptlrpc_daemonize(void)
 287 {
 288         exit_mm(current);
 289         lustre_daemonize_helper();
 290         exit_files(current);
 291         reparent_to_init();
 292 }
 293
 294 static long timeval_sub(struct timeval *large, struct timeval *small)
 295 {
 296         return (large->tv_sec - small->tv_sec) * 1000000 +
 297                 (large->tv_usec - small->tv_usec);
 298 }
 299
 300 static int ptlrpc_main(void *arg)
 301 {
 302         struct ptlrpc_svc_data *data = arg;
 303         struct obd_device *obddev = data->dev;
 304         struct ptlrpc_service *svc = data->svc;
 305         struct ptlrpc_thread *thread = data->thread;
 306         struct ptlrpc_request *request;
 307         ptl_event_t *event;
 308         unsigned long flags;
 309         struct timeval start_time, finish_time;
 310         long total;
 311         int rc = 0;
 312         ENTRY;
 313
 314         lock_kernel();
 315         ptlrpc_daemonize();
 316
 317         SIGNAL_MASK_LOCK(current, flags);
 318         sigfillset(&current->blocked);
 319         RECALC_SIGPENDING;
 320         SIGNAL_MASK_UNLOCK(current, flags);
 321
 322         THREAD_NAME(current->comm, "%s", data->name);
 323         unlock_kernel();
 324
 325         OBD_ALLOC(event, sizeof(*event));
 326         if (event == NULL)
 327                 GOTO(out, rc = -ENOMEM);
 328         OBD_ALLOC(request, sizeof(*request));
 329         if (request == NULL)
 330                 GOTO(out_event, rc = -ENOMEM);
 331
 332         /* Record that the thread is running */
 333         thread->t_flags = SVC_RUNNING;
 334         wake_up(&thread->t_ctl_waitq);
 335
 336         /* XXX maintain a list of all managed devices: insert here */
 337
 338         do_gettimeofday(&finish_time);
 339         /* And now, loop forever on requests */
 340         while (1) {
 341                 struct l_wait_info lwi = { 0 };
 342                 l_wait_event_exclusive(svc->srv_waitq,
 343                                        ptlrpc_check_event(svc, thread, event),
 344                                        &lwi);
 345
 346                 spin_lock(&svc->srv_lock);
 347                 if (thread->t_flags & SVC_STOPPING) {
 348                         thread->t_flags &= ~SVC_STOPPING;
 349                         spin_unlock(&svc->srv_lock);
 350
 351                         EXIT;
 352                         break;
 353                 }
 354
 355                 if (!(thread->t_flags & SVC_EVENT)) {
 356                         CERROR("unknown flag in service");
 357                         spin_unlock(&svc->srv_lock);
 358                         LBUG();
 359                         EXIT;
 360                         break;
 361                 }
 362
 363                 thread->t_flags &= ~SVC_EVENT;
 364                 spin_unlock(&svc->srv_lock);
 365
 366                 do_gettimeofday(&start_time);
 367                 total = timeval_sub(&start_time, &event->arrival_time);
 368                 if (svc->srv_stats != NULL) {
 369                         lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 370                                             total);
 371                         lprocfs_counter_add(svc->srv_stats,
 372                                             PTLRPC_SVCIDLETIME_CNTR,
 373                                             timeval_sub(&start_time,
 374                                                         &finish_time));
 375 #if 0 /* Wait for b_eq branch */
 376                         lprocfs_counter_add(svc->srv_stats,
 377                                             PTLRPC_SVCEQDEPTH_CNTR, 0);
 378 #endif
 379                 }
 380
 381                 if (total / 1000000 > (long)obd_timeout) {
 382                         CERROR("Dropping request from NID "LPX64" because it's "
 383                                "%ld seconds old.\n", event->initiator.nid,
 384                                total / 1000000); /* bug 1502 */
 385                 } else {
 386                         CDEBUG(D_HA, "request from NID "LPX64" noticed after "
 387                                "%ldus\n", event->initiator.nid, total);
 388                         rc = handle_incoming_request(obddev, svc, event,
 389                                                      request);
 390                 }
 391                 do_gettimeofday(&finish_time);
 392                 total = timeval_sub(&finish_time, &start_time);
 393
 394                 CDEBUG((total / 1000000 > (long)obd_timeout) ? D_ERROR : D_HA,
 395                        "request "LPU64" from NID "LPX64" processed in %ldus "
 396                        "(%ldus total)\n", request->rq_xid, event->initiator.nid,
 397                        total, timeval_sub(&finish_time, &event->arrival_time));
 398
 399                 if (svc->srv_stats != NULL) {
 400                         int opc = opcode_offset(request->rq_reqmsg->opc);
 401                         if (opc > 0) {
 402                                 LASSERT(opc < LUSTRE_MAX_OPCODES);
 403                                 lprocfs_counter_add(svc->srv_stats,
 404                                                     opc + PTLRPC_LAST_CNTR,
 405                                                     total);
 406                         }
 407                 }
 408         }
 409
 410         /* NB should wait for all SENT callbacks to complete before exiting
 411          * here.  Unfortunately at this time there is no way to track this
 412          * state. */
 413         OBD_FREE(request, sizeof(*request));
 414 out_event:
 415         OBD_FREE(event, sizeof(*event));
 416 out:
 417         thread->t_flags = SVC_STOPPED;
 418         wake_up(&thread->t_ctl_waitq);
 419
 420         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 421                current->pid, rc);
 422         return rc;
 423 }
 424
 425 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 426                                struct ptlrpc_thread *thread)
 427 {
 428         struct l_wait_info lwi = { 0 };
 429
 430         spin_lock(&svc->srv_lock);
 431         thread->t_flags = SVC_STOPPING;
 432         spin_unlock(&svc->srv_lock);
 433
 434         wake_up_all(&svc->srv_waitq);
 435         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 436                      &lwi);
 437 }
 438
 439 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 440 {
 441         spin_lock(&svc->srv_lock);
 442         while (!list_empty(&svc->srv_threads)) {
 443                 struct ptlrpc_thread *thread;
 444                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 445                                     t_link);
 446                 spin_unlock(&svc->srv_lock);
 447                 ptlrpc_stop_thread(svc, thread);
 448                 spin_lock(&svc->srv_lock);
 449                 list_del(&thread->t_link);
 450                 OBD_FREE(thread, sizeof(*thread));
 451         }
 452         spin_unlock(&svc->srv_lock);
 453 }
 454
 455 int ptlrpc_start_n_threads(struct obd_device *dev, struct ptlrpc_service *svc,
 456                            int num_threads, char *base_name)
 457 {
 458         int i, rc = 0;
 459         ENTRY;
 460
 461         for (i = 0; i < num_threads; i++) {
 462                 char name[32];
 463                 sprintf(name, "%s_%02d", base_name, i);
 464                 rc = ptlrpc_start_thread(dev, svc, name);
 465                 if (rc) {
 466                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
 467                                i, rc);
 468                         ptlrpc_stop_all_threads(svc);
 469                 }
 470         }
 471         RETURN(rc);
 472 }
 473
 474 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 475                         char *name)
 476 {
 477         struct l_wait_info lwi = { 0 };
 478         struct ptlrpc_svc_data d;
 479         struct ptlrpc_thread *thread;
 480         int rc;
 481         ENTRY;
 482
 483         OBD_ALLOC(thread, sizeof(*thread));
 484         if (thread == NULL)
 485                 RETURN(-ENOMEM);
 486         init_waitqueue_head(&thread->t_ctl_waitq);
 487
 488         d.dev = dev;
 489         d.svc = svc;
 490         d.name = name;
 491         d.thread = thread;
 492
 493         spin_lock(&svc->srv_lock);
 494         list_add(&thread->t_link, &svc->srv_threads);
 495         spin_unlock(&svc->srv_lock);
 496
 497         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 498          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 499          */
 500         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 501         if (rc < 0) {
 502                 CERROR("cannot start thread: %d\n", rc);
 503                 OBD_FREE(thread, sizeof(*thread));
 504                 RETURN(rc);
 505         }
 506         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 507
 508         RETURN(0);
 509 }
 510
 511 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 512 {
 513         int i, rc;
 514         struct ptlrpc_srv_ni *srv_ni;
 515
 516         LASSERT (list_empty (&service->srv_threads));
 517
 518         /* XXX We could reply (with failure) to all buffered requests
 519          * _after_ unlinking _all_ the request buffers, but _before_
 520          * freeing them.
 521          */
 522
 523         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 524                 srv_ni = &service->srv_interfaces[i];
 525                 CDEBUG (D_NET, "%s: tearing down interface %s\n",
 526                         service->srv_name, srv_ni->sni_ni->pni_name);
 527
 528                 while (!list_empty (&srv_ni->sni_rqbds)) {
 529                         struct ptlrpc_request_buffer_desc *rqbd =
 530                                 list_entry (srv_ni->sni_rqbds.next,
 531                                             struct ptlrpc_request_buffer_desc,
 532                                             rqbd_list);
 533
 534                         list_del (&rqbd->rqbd_list);
 535
 536                         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 537                         /* refcount could be anything; it's possible for
 538                          * the buffers to continued to get filled after all
 539                          * the server threads exited.  But we know they
 540                          * _have_ exited.
 541                          */
 542
 543                         (void) PtlMEUnlink(rqbd->rqbd_me_h);
 544                         /* The callback handler could have unlinked this ME
 545                          * already (we're racing with her) but it's safe to
 546                          * ensure it _has_ been unlinked.
 547                          */
 548
 549                         OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 550                         OBD_FREE (rqbd, sizeof (*rqbd));
 551                         srv_ni->sni_nrqbds--;
 552                 }
 553
 554                 LASSERT (srv_ni->sni_nrqbds == 0);
 555
 556                 if (!PtlHandleEqual (srv_ni->sni_eq_h, PTL_HANDLE_NONE)) {
 557                         rc = PtlEQFree(srv_ni->sni_eq_h);
 558                         if (rc)
 559                                 CERROR("%s.%d: PtlEQFree failed on %s: %d\n",
 560                                        service->srv_name, i,
 561                                        srv_ni->sni_ni->pni_name, rc);
 562                 }
 563         }
 564
 565         ptlrpc_lprocfs_unregister_service(service);
 566
 567         OBD_FREE(service,
 568                  offsetof (struct ptlrpc_service,
 569                            srv_interfaces[ptlrpc_ninterfaces]));
 570         return 0;
 571 }