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;
  41         int idx;
  42         int rc;
  43         ENTRY;
  44
  45         spin_lock(&svc->srv_lock);
  46
  47         if (thread->t_flags & SVC_STOPPING)
  48                 GOTO(out, rc = 1);
  49
  50         LASSERT ((thread->t_flags & SVC_EVENT) == 0);
  51         LASSERT (ptlrpc_ninterfaces > 0);
  52
  53         for (i = 0; i < ptlrpc_ninterfaces; i++) {
  54                 idx = (svc->srv_interface_rover + i) % ptlrpc_ninterfaces;
  55                 srv_ni = &svc->srv_interfaces[idx];
  56
  57                 LASSERT (!PtlHandleEqual (srv_ni->sni_eq_h, PTL_HANDLE_NONE));
  58
  59                 rc = PtlEQGet(srv_ni->sni_eq_h, event);
  60                 switch (rc) {
  61                 case PTL_OK:
  62                         /* next time start with the next interface */
  63                         svc->srv_interface_rover = (idx+1) % ptlrpc_ninterfaces;
  64                         thread->t_flags |= SVC_EVENT;
  65                         GOTO(out, rc = 1);
  66
  67                 case PTL_EQ_EMPTY:
  68                         continue;
  69
  70                 default:
  71                         CERROR("BUG: PtlEQGet returned %d\n", rc);
  72                         LBUG();
  73                 }
  74         }
  75         rc = 0;
  76         EXIT;
  77  out:
  78         spin_unlock(&svc->srv_lock);
  79         return rc;
  80 }
  81
  82 struct ptlrpc_service *
  83 ptlrpc_init_svc(__u32 nevents, __u32 nbufs,
  84                 __u32 bufsize, __u32 max_req_size,
  85                 int req_portal, int rep_portal,
  86                 svc_handler_t handler, char *name,
  87                 struct obd_device *obddev)
  88 {
  89         int i, j, ssize, rc;
  90         struct ptlrpc_service *service;
  91         struct ptlrpc_srv_ni  *srv_ni;
  92         ENTRY;
  93
  94         LASSERT (ptlrpc_ninterfaces > 0);
  95
  96         ssize = offsetof (struct ptlrpc_service,
  97                           srv_interfaces[ptlrpc_ninterfaces]);
  98         OBD_ALLOC(service, ssize);
  99         if (service == NULL)
 100                 RETURN(NULL);
 101
 102         service->srv_name = name;
 103         spin_lock_init(&service->srv_lock);
 104         INIT_LIST_HEAD(&service->srv_threads);
 105         init_waitqueue_head(&service->srv_waitq);
 106
 107         service->srv_max_req_size = max_req_size;
 108         service->srv_buf_size = bufsize;
 109
 110         service->srv_rep_portal = rep_portal;
 111         service->srv_req_portal = req_portal;
 112         service->srv_handler = handler;
 113         service->srv_interface_rover = 0;
 114
 115         /* First initialise enough for early teardown */
 116         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 117                 srv_ni = &service->srv_interfaces[i];
 118
 119                 srv_ni->sni_service = service;
 120                 srv_ni->sni_ni = &ptlrpc_interfaces[i];
 121                 srv_ni->sni_eq_h = PTL_HANDLE_NONE;
 122                 INIT_LIST_HEAD(&srv_ni->sni_rqbds);
 123                 srv_ni->sni_nrqbds = 0;
 124                 atomic_set(&srv_ni->sni_nrqbds_receiving, 0);
 125         }
 126
 127         /* Now allocate the event queue and request buffers, assuming all
 128          * interfaces require the same level of buffering. */
 129         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 130                 srv_ni = &service->srv_interfaces[i];
 131                 CDEBUG (D_NET, "%s: initialising interface %s\n", name,
 132                         srv_ni->sni_ni->pni_name);
 133
 134                 rc = PtlEQAlloc(srv_ni->sni_ni->pni_ni_h, nevents,
 135                                 request_in_callback, &(srv_ni->sni_eq_h));
 136                 if (rc != PTL_OK) {
 137                         CERROR("%s.%d: PtlEQAlloc on %s failed: %d\n",
 138                                name, i, srv_ni->sni_ni->pni_name, rc);
 139                         GOTO (failed, NULL);
 140                 }
 141
 142                 for (j = 0; j < nbufs; j++) {
 143                         struct ptlrpc_request_buffer_desc *rqbd;
 144
 145                         OBD_ALLOC(rqbd, sizeof(*rqbd));
 146                         if (rqbd == NULL) {
 147                                 CERROR ("%s.%d: Can't allocate request "
 148                                         "descriptor %d on %s\n",
 149                                         name, i, srv_ni->sni_nrqbds,
 150                                         srv_ni->sni_ni->pni_name);
 151                                 GOTO(failed, NULL);
 152                         }
 153
 154                         rqbd->rqbd_srv_ni = srv_ni;
 155                         rqbd->rqbd_me_h = PTL_HANDLE_NONE;
 156                         atomic_set(&rqbd->rqbd_refcount, 0);
 157
 158                         OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 159                         if (rqbd->rqbd_buffer == NULL) {
 160                                 CERROR ("%s.%d: Can't allocate request "
 161                                         "buffer %d on %s\n",
 162                                         name, i, srv_ni->sni_nrqbds,
 163                                         srv_ni->sni_ni->pni_name);
 164                                 OBD_FREE(rqbd, sizeof(*rqbd));
 165                                 GOTO(failed, NULL);
 166                         }
 167                         list_add(&rqbd->rqbd_list, &srv_ni->sni_rqbds);
 168                         srv_ni->sni_nrqbds++;
 169
 170                         ptlrpc_link_svc_me(rqbd);
 171                 }
 172         }
 173
 174         ptlrpc_lprocfs_register_service(obddev, service);
 175
 176         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 177                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 178
 179         RETURN(service);
 180 failed:
 181         ptlrpc_unregister_service(service);
 182         return NULL;
 183 }
 184
 185 static int handle_incoming_request(struct obd_device *obddev,
 186                                    struct ptlrpc_service *svc,
 187                                    ptl_event_t *event,
 188                                    struct ptlrpc_request *request)
 189 {
 190         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 191         int rc;
 192
 193         /* FIXME: If we move to an event-driven model, we should put the request
 194          * on the stack of mds_handle instead. */
 195
 196         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 197         LASSERT ((event->mem_desc.options & (PTL_MD_IOV | PTL_MD_KIOV)) == 0);
 198         LASSERT (rqbd->rqbd_srv_ni->sni_service == svc);
 199         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 200         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 201
 202         memset(request, 0, sizeof(*request));
 203         spin_lock_init (&request->rq_lock);
 204         INIT_LIST_HEAD(&request->rq_list);
 205         request->rq_svc = svc;
 206         request->rq_obd = obddev;
 207         request->rq_xid = event->match_bits;
 208         request->rq_reqmsg = event->mem_desc.start + event->offset;
 209         request->rq_reqlen = event->mlength;
 210
 211 #if SWAB_PARANOIA
 212         /* Clear request swab mask; this is a new request */
 213         request->rq_req_swab_mask = 0;
 214 #endif
 215         rc = lustre_unpack_msg (request->rq_reqmsg, request->rq_reqlen);
 216         if (rc != 0) {
 217                 CERROR ("error unpacking request: ptl %d from "LPX64
 218                         " xid "LPU64"\n", svc->srv_req_portal,
 219                        event->initiator.nid, request->rq_xid);
 220                 goto out;
 221         }
 222         rc = -EINVAL;
 223         if (request->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 224                 CERROR("wrong packet type received (type=%u)\n",
 225                        request->rq_reqmsg->type);
 226                 goto out;
 227         }
 228
 229         CDEBUG(D_NET, "got req "LPD64" (md: %p + %d)\n", request->rq_xid,
 230                event->mem_desc.start, event->offset);
 231
 232         request->rq_peer.peer_nid = event->initiator.nid;
 233         request->rq_peer.peer_ni = rqbd->rqbd_srv_ni->sni_ni;
 234
 235         request->rq_export = class_conn2export(&request->rq_reqmsg->handle);
 236
 237         if (request->rq_export) {
 238                 request->rq_connection = request->rq_export->exp_connection;
 239                 ptlrpc_connection_addref(request->rq_connection);
 240                 request->rq_export->exp_last_request_time =
 241                         LTIME_S(CURRENT_TIME);
 242         } else {
 243                 /* create a (hopefully temporary) connection that will be used
 244                  * to send the reply if this call doesn't create an export.
 245                  * XXX revisit this when we revamp ptlrpc */
 246                 request->rq_connection =
 247                         ptlrpc_get_connection(&request->rq_peer, NULL);
 248         }
 249
 250         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid:pid:xid:ni:nid:opc %s:%s:%d:"
 251                LPU64":%s:"LPX64":%d\n",
 252                current->comm,
 253                (request->rq_export ?
 254                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 255                request->rq_reqmsg->status, request->rq_xid,
 256                rqbd->rqbd_srv_ni->sni_ni->pni_name, event->initiator.nid,
 257                request->rq_reqmsg->opc);
 258
 259         rc = svc->srv_handler(request);
 260         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid:pid:xid:ni:nid:opc %s:%s:%d:"
 261                LPU64":%s:"LPX64":%d\n",
 262                current->comm,
 263                (request->rq_export ?
 264                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 265                request->rq_reqmsg->status, request->rq_xid,
 266                rqbd->rqbd_srv_ni->sni_ni->pni_name, event->initiator.nid,
 267                request->rq_reqmsg->opc);
 268
 269         ptlrpc_put_connection(request->rq_connection);
 270         if (request->rq_export != NULL)
 271                 class_export_put(request->rq_export);
 272
 273  out:
 274         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 275                 ptlrpc_link_svc_me (rqbd);
 276
 277         return rc;
 278 }
 279
 280 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 281 void ptlrpc_daemonize(void)
 282 {
 283         exit_mm(current);
 284
 285         current->session = 1;
 286         current->pgrp = 1;
 287         current->tty = NULL;
 288
 289         exit_files(current);
 290         reparent_to_init();
 291 }
 292
 293 static int ptlrpc_main(void *arg)
 294 {
 295         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 296         struct obd_device *obddev = data->dev;
 297         struct ptlrpc_service *svc = data->svc;
 298         struct ptlrpc_thread *thread = data->thread;
 299         struct ptlrpc_request *request;
 300         ptl_event_t *event;
 301         int rc = 0;
 302         unsigned long flags;
 303         cycles_t workdone_time;
 304         cycles_t svc_workcycles;
 305         ENTRY;
 306
 307         lock_kernel();
 308         ptlrpc_daemonize();
 309
 310         SIGNAL_MASK_LOCK(current, flags);
 311         sigfillset(&current->blocked);
 312         RECALC_SIGPENDING;
 313         SIGNAL_MASK_UNLOCK(current, flags);
 314
 315 #if defined(__arch_um__) && (LINUX_VERSION_CODE < KERNEL_VERSION(2,4,20))
 316         sprintf(current->comm, "%s|%d", data->name,current->thread.extern_pid);
 317 #elif defined(__arch_um__) && (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 318         sprintf(current->comm, "%s|%d", data->name,
 319                 current->thread.mode.tt.extern_pid);
 320 #else
 321         strcpy(current->comm, data->name);
 322 #endif
 323         unlock_kernel();
 324
 325         OBD_ALLOC(event, sizeof(*event));
 326         if (!event)
 327                 GOTO(out, rc = -ENOMEM);
 328         OBD_ALLOC(request, sizeof(*request));
 329         if (!request)
 330                 GOTO(out_event, rc = -ENOMEM);
 331
 332         /* Record that the thread is running */
 333         thread->t_flags = SVC_RUNNING;
 334         svc_workcycles = workdone_time = 0;
 335         wake_up(&thread->t_ctl_waitq);
 336
 337         /* XXX maintain a list of all managed devices: insert here */
 338
 339         /* And now, loop forever on requests */
 340         while (1) {
 341                 struct l_wait_info lwi = { 0 };
 342                 l_wait_event(svc->srv_waitq,
 343                              ptlrpc_check_event(svc, thread, event), &lwi);
 344
 345                 if (thread->t_flags & SVC_STOPPING) {
 346                         spin_lock(&svc->srv_lock);
 347                         thread->t_flags &= ~SVC_STOPPING;
 348                         spin_unlock(&svc->srv_lock);
 349
 350                         EXIT;
 351                         break;
 352                 }
 353
 354                 if (thread->t_flags & SVC_EVENT) {
 355                         cycles_t  workstart_time;
 356                         spin_lock(&svc->srv_lock);
 357                         thread->t_flags &= ~SVC_EVENT;
 358                         /* Update Service Statistics */
 359                         workstart_time = get_cycles();
 360                         if (workdone_time && (svc->svc_counters != NULL)) {
 361                                 /* Stats for req(n) are updated just before
 362                                  * req(n+1) is executed. This avoids need to
 363                                  * reacquire svc->srv_lock after
 364                                  * call to handling_request().
 365                                  */
 366                                 int opc_offset;
 367                                 /* req_waittime */
 368                                 LPROCFS_COUNTER_INCR(&svc->svc_counters->cntr[PTLRPC_REQWAIT_CNTR],
 369                                                      (workstart_time -
 370                                                       event->arrival_time));
 371                                 /* svc_eqdepth */
 372                                 LPROCFS_COUNTER_INCR(&svc->svc_counters->cntr[PTLRPC_SVCEQDEPTH_CNTR],
 373                                                      0); /* Wait for b_eq branch */
 374                                 /* svc_idletime */
 375                                 LPROCFS_COUNTER_INCR(&svc->svc_counters->cntr[PTLRPC_SVCIDLETIME_CNTR],
 376                                                      (workstart_time -
 377                                                       workdone_time));
 378                                 /* previous request */
 379                                 opc_offset =
 380                                         opcode_offset(request->rq_reqmsg->opc);
 381                                 if (opc_offset >= 0) {
 382                                         LASSERT(opc_offset < LUSTRE_MAX_OPCODES);
 383                                         LPROCFS_COUNTER_INCR(&svc->svc_counters->cntr[PTLRPC_LAST_CNTR+opc_offset], svc_workcycles);
 384                                 }
 385                         }
 386                         spin_unlock(&svc->srv_lock);
 387
 388                         rc = handle_incoming_request(obddev, svc, event,
 389                                                      request);
 390                         workdone_time = get_cycles();
 391                         svc_workcycles = workdone_time - workstart_time;
 392                         continue;
 393                 }
 394
 395                 CERROR("unknown break in service");
 396                 LBUG();
 397                 EXIT;
 398                 break;
 399         }
 400
 401         /* NB should wait for all SENT callbacks to complete before exiting
 402          * here.  Unfortunately at this time there is no way to track this
 403          * state.
 404          */
 405         OBD_FREE(request, sizeof(*request));
 406 out_event:
 407         OBD_FREE(event, sizeof(*event));
 408 out:
 409         thread->t_flags = SVC_STOPPED;
 410         wake_up(&thread->t_ctl_waitq);
 411
 412         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 413                current->pid, rc);
 414         return rc;
 415 }
 416
 417 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 418                                struct ptlrpc_thread *thread)
 419 {
 420         struct l_wait_info lwi = { 0 };
 421
 422         spin_lock(&svc->srv_lock);
 423         thread->t_flags = SVC_STOPPING;
 424         spin_unlock(&svc->srv_lock);
 425
 426         wake_up(&svc->srv_waitq);
 427         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
 428                      &lwi);
 429 }
 430
 431 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 432 {
 433         spin_lock(&svc->srv_lock);
 434         while (!list_empty(&svc->srv_threads)) {
 435                 struct ptlrpc_thread *thread;
 436                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 437                                     t_link);
 438                 spin_unlock(&svc->srv_lock);
 439                 ptlrpc_stop_thread(svc, thread);
 440                 spin_lock(&svc->srv_lock);
 441                 list_del(&thread->t_link);
 442                 OBD_FREE(thread, sizeof(*thread));
 443         }
 444         spin_unlock(&svc->srv_lock);
 445 }
 446
 447 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 448                         char *name)
 449 {
 450         struct l_wait_info lwi = { 0 };
 451         struct ptlrpc_svc_data d;
 452         struct ptlrpc_thread *thread;
 453         int rc;
 454         ENTRY;
 455
 456         OBD_ALLOC(thread, sizeof(*thread));
 457         if (thread == NULL)
 458                 RETURN(-ENOMEM);
 459         init_waitqueue_head(&thread->t_ctl_waitq);
 460
 461         d.dev = dev;
 462         d.svc = svc;
 463         d.name = name;
 464         d.thread = thread;
 465
 466         spin_lock(&svc->srv_lock);
 467         list_add(&thread->t_link, &svc->srv_threads);
 468         spin_unlock(&svc->srv_lock);
 469
 470         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 471          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 472          */
 473         rc = kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
 474         if (rc < 0) {
 475                 CERROR("cannot start thread: %d\n", rc);
 476                 OBD_FREE(thread, sizeof(*thread));
 477                 RETURN(rc);
 478         }
 479         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING, &lwi);
 480
 481         RETURN(0);
 482 }
 483
 484 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 485 {
 486         int i, rc;
 487         struct ptlrpc_srv_ni *srv_ni;
 488
 489         LASSERT (list_empty (&service->srv_threads));
 490
 491         /* XXX We could reply (with failure) to all buffered requests
 492          * _after_ unlinking _all_ the request buffers, but _before_
 493          * freeing them.
 494          */
 495
 496         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 497                 srv_ni = &service->srv_interfaces[i];
 498                 CDEBUG (D_NET, "%s: tearing down interface %s\n",
 499                         service->srv_name, srv_ni->sni_ni->pni_name);
 500
 501                 while (!list_empty (&srv_ni->sni_rqbds)) {
 502                         struct ptlrpc_request_buffer_desc *rqbd =
 503                                 list_entry (srv_ni->sni_rqbds.next,
 504                                             struct ptlrpc_request_buffer_desc,
 505                                             rqbd_list);
 506
 507                         list_del (&rqbd->rqbd_list);
 508
 509                         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 510                         /* refcount could be anything; it's possible for
 511                          * the buffers to continued to get filled after all
 512                          * the server threads exited.  But we know they
 513                          * _have_ exited.
 514                          */
 515
 516                         (void) PtlMEUnlink(rqbd->rqbd_me_h);
 517                         /* The callback handler could have unlinked this ME
 518                          * already (we're racing with her) but it's safe to
 519                          * ensure it _has_ been unlinked.
 520                          */
 521
 522                         OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 523                         OBD_FREE (rqbd, sizeof (*rqbd));
 524                         srv_ni->sni_nrqbds--;
 525                 }
 526
 527                 LASSERT (srv_ni->sni_nrqbds == 0);
 528
 529                 if (!PtlHandleEqual (srv_ni->sni_eq_h, PTL_HANDLE_NONE)) {
 530                         rc = PtlEQFree(srv_ni->sni_eq_h);
 531                         if (rc)
 532                                 CERROR("%s.%d: PtlEQFree failed on %s: %d\n",
 533                                        service->srv_name, i,
 534                                        srv_ni->sni_ni->pni_name, rc);
 535                 }
 536         }
 537
 538         ptlrpc_lprocfs_unregister_service(service);
 539
 540         OBD_FREE(service,
 541                  offsetof (struct ptlrpc_service,
 542                            srv_interfaces[ptlrpc_ninterfaces]));
 543         return 0;
 544 }