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
  32 extern int request_in_callback(ptl_event_t *ev);
  33
  34 static int ptlrpc_check_event(struct ptlrpc_service *svc,
  35                               struct ptlrpc_thread *thread, ptl_event_t *event)
  36 {
  37         struct ptlrpc_srv_ni *srv_ni;
  38         int i;
  39         int idx;
  40         int 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 (ptl_is_valid_handle (&srv_ni->sni_eq_h));
  56
  57                 rc = PtlEQGet(srv_ni->sni_eq_h, event);
  58                 switch (rc)
  59                 {
  60                 case PTL_OK:
  61                         /* next time start with the next interface */
  62                         svc->srv_interface_rover = (idx+1) % ptlrpc_ninterfaces;
  63                         thread->t_flags |= SVC_EVENT;
  64                         GOTO(out, rc = 1);
  65
  66                 case PTL_EQ_EMPTY:
  67                         continue;
  68
  69                 default:
  70                         CERROR("BUG: PtlEQGet returned %d\n", rc);
  71                         LBUG();
  72                 }
  73         }
  74         rc = 0;
  75  out:
  76         spin_unlock(&svc->srv_lock);
  77         return rc;
  78 }
  79
  80 struct ptlrpc_service *
  81 ptlrpc_init_svc(__u32 nevents, __u32 nbufs,
  82                 __u32 bufsize, __u32 max_req_size,
  83                 int req_portal, int rep_portal,
  84                 svc_handler_t handler, char *name)
  85 {
  86         int ssize;
  87         int rc;
  88         int i;
  89         int j;
  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                 ptl_set_inv_handle (&srv_ni->sni_eq_h);
 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                         ptl_set_inv_handle(&rqbd->rqbd_me_h);
 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         CDEBUG(D_NET, "%s: Started on %d interfaces, listening on portal %d\n",
 175                service->srv_name, ptlrpc_ninterfaces, service->srv_req_portal);
 176
 177         RETURN(service);
 178 failed:
 179         ptlrpc_unregister_service(service);
 180         return NULL;
 181 }
 182
 183 static int handle_incoming_request(struct obd_device *obddev,
 184                                    struct ptlrpc_service *svc,
 185                                    ptl_event_t *event,
 186                                    struct ptlrpc_request *request)
 187 {
 188         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 189         int rc;
 190
 191         /* FIXME: If we move to an event-driven model, we should put the request
 192          * on the stack of mds_handle instead. */
 193
 194         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 195         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 196         LASSERT (rqbd->rqbd_srv_ni->sni_service == svc);
 197         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 198         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 199
 200         memset(request, 0, sizeof(*request));
 201         INIT_LIST_HEAD(&request->rq_list);
 202         request->rq_svc = svc;
 203         request->rq_obd = obddev;
 204         request->rq_xid = event->match_bits;
 205         request->rq_reqmsg = event->mem_desc.start + event->offset;
 206         request->rq_reqlen = event->mlength;
 207
 208         rc = -EINVAL;
 209
 210         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 211                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "
 212                        LPU64"\n",
 213                        request->rq_reqlen, svc->srv_req_portal,
 214                        event->initiator.nid, request->rq_xid);
 215                 goto out;
 216         }
 217
 218         CDEBUG(D_RPCTRACE, "Handling RPC ni:pid:xid:nid:opc %d:%d:"LPU64":"
 219                LPX64":%d\n", rqbd->rqbd_srv_ni - &svc->srv_interfaces[0],
 220                NTOH__u32(request->rq_reqmsg->status), request->rq_xid,
 221                event->initiator.nid, NTOH__u32(request->rq_reqmsg->opc));
 222
 223         if (NTOH__u32(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         if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 230                 CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "
 231                        LPD64"\n",
 232                        request->rq_reqmsg->magic, svc->srv_req_portal,
 233                        event->initiator.nid, request->rq_xid);
 234                 goto out;
 235         }
 236
 237         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 238                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "
 239                        LPD64"\n",
 240                        request->rq_reqmsg->version, svc->srv_req_portal,
 241                        event->initiator.nid, request->rq_xid);
 242                 goto out;
 243         }
 244
 245         CDEBUG(D_NET, "got req "LPD64" (md: %p + %d)\n", request->rq_xid,
 246                event->mem_desc.start, event->offset);
 247
 248         request->rq_peer.peer_nid = event->initiator.nid;
 249         request->rq_peer.peer_ni = rqbd->rqbd_srv_ni->sni_ni;
 250
 251         request->rq_export = class_conn2export((struct lustre_handle *)
 252                                                request->rq_reqmsg);
 253
 254         if (request->rq_export) {
 255                 request->rq_connection = request->rq_export->exp_connection;
 256                 ptlrpc_connection_addref(request->rq_connection);
 257         } else {
 258                 /* create a (hopefully temporary) connection that will be used
 259                  * to send the reply if this call doesn't create an export.
 260                  * XXX revisit this when we revamp ptlrpc */
 261                 request->rq_connection =
 262                         ptlrpc_get_connection(&request->rq_peer, NULL);
 263         }
 264
 265         rc = svc->srv_handler(request);
 266         ptlrpc_put_connection(request->rq_connection);
 267
 268  out:
 269         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 270                 ptlrpc_link_svc_me (rqbd);
 271
 272         return rc;
 273 }
 274
 275 /* Don't use daemonize, it removes fs struct from new thread  (bug 418) */
 276 static void ptlrpc_daemonize(void)
 277 {
 278         exit_mm(current);
 279
 280         current->session = 1;
 281         current->pgrp = 1;
 282         current->tty = NULL;
 283
 284         exit_files(current);
 285         reparent_to_init();
 286 }
 287
 288 static int ptlrpc_main(void *arg)
 289 {
 290         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 291         struct obd_device *obddev = data->dev;
 292         struct ptlrpc_service *svc = data->svc;
 293         struct ptlrpc_thread *thread = data->thread;
 294         struct ptlrpc_request *request;
 295         ptl_event_t *event;
 296         int rc = 0;
 297         unsigned long flags;
 298         ENTRY;
 299
 300         lock_kernel();
 301         ptlrpc_daemonize();
 302
 303 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
 304         sigfillset(&current->blocked);
 305         recalc_sigpending();
 306 #else
 307         spin_lock_irqsave(&current->sigmask_lock, flags);
 308         sigfillset(&current->blocked);
 309         recalc_sigpending(current);
 310         spin_unlock_irqrestore(&current->sigmask_lock, flags);
 311 #endif
 312
 313 #ifdef __arch_um__
 314 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 315         sprintf(current->comm, "%s|%d", data->name,current->thread.extern_pid);
 316 #endif
 317 #else
 318         strcpy(current->comm, data->name);
 319 #endif
 320         unlock_kernel();
 321
 322         OBD_ALLOC(event, sizeof(*event));
 323         if (!event)
 324                 GOTO(out, rc = -ENOMEM);
 325         OBD_ALLOC(request, sizeof(*request));
 326         if (!request)
 327                 GOTO(out_event, rc = -ENOMEM);
 328
 329         /* Record that the thread is running */
 330         thread->t_flags = SVC_RUNNING;
 331         wake_up(&thread->t_ctl_waitq);
 332
 333         /* XXX maintain a list of all managed devices: insert here */
 334
 335         /* And now, loop forever on requests */
 336         while (1) {
 337                 wait_event(svc->srv_waitq,
 338                            ptlrpc_check_event(svc, thread, event));
 339
 340                 if (thread->t_flags & SVC_STOPPING) {
 341                         spin_lock(&svc->srv_lock);
 342                         thread->t_flags &= ~SVC_STOPPING;
 343                         spin_unlock(&svc->srv_lock);
 344
 345                         EXIT;
 346                         break;
 347                 }
 348
 349                 if (thread->t_flags & SVC_EVENT) {
 350                         spin_lock(&svc->srv_lock);
 351                         thread->t_flags &= ~SVC_EVENT;
 352                         spin_unlock(&svc->srv_lock);
 353
 354                         rc = handle_incoming_request(obddev, svc, event,
 355                                                      request);
 356                         continue;
 357                 }
 358
 359                 CERROR("unknown break in service");
 360                 LBUG();
 361                 EXIT;
 362                 break;
 363         }
 364
 365         OBD_FREE(request, sizeof(*request));
 366 out_event:
 367         OBD_FREE(event, sizeof(*event));
 368 out:
 369         thread->t_flags = SVC_STOPPED;
 370         wake_up(&thread->t_ctl_waitq);
 371
 372         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 373                current->pid, rc);
 374         return rc;
 375 }
 376
 377 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 378                                struct ptlrpc_thread *thread)
 379 {
 380         spin_lock(&svc->srv_lock);
 381         thread->t_flags = SVC_STOPPING;
 382         spin_unlock(&svc->srv_lock);
 383
 384         wake_up(&svc->srv_waitq);
 385         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 386 }
 387
 388 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 389 {
 390         spin_lock(&svc->srv_lock);
 391         while (!list_empty(&svc->srv_threads)) {
 392                 struct ptlrpc_thread *thread;
 393                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 394                                     t_link);
 395                 spin_unlock(&svc->srv_lock);
 396                 ptlrpc_stop_thread(svc, thread);
 397                 spin_lock(&svc->srv_lock);
 398                 list_del(&thread->t_link);
 399                 OBD_FREE(thread, sizeof(*thread));
 400         }
 401         spin_unlock(&svc->srv_lock);
 402 }
 403
 404 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 405                         char *name)
 406 {
 407         struct ptlrpc_svc_data d;
 408         struct ptlrpc_thread *thread;
 409         int rc;
 410         ENTRY;
 411
 412         OBD_ALLOC(thread, sizeof(*thread));
 413         if (thread == NULL) {
 414                 LBUG();
 415                 RETURN(-ENOMEM);
 416         }
 417         init_waitqueue_head(&thread->t_ctl_waitq);
 418
 419         d.dev = dev;
 420         d.svc = svc;
 421         d.name = name;
 422         d.thread = thread;
 423
 424         spin_lock(&svc->srv_lock);
 425         list_add(&thread->t_link, &svc->srv_threads);
 426         spin_unlock(&svc->srv_lock);
 427
 428         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 429          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 430          */
 431         rc = kernel_thread(ptlrpc_main, (void *) &d, CLONE_VM | CLONE_FILES);
 432         if (rc < 0) {
 433                 CERROR("cannot start thread\n");
 434                 OBD_FREE(thread, sizeof(*thread));
 435                 RETURN(rc);
 436         }
 437         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 438
 439         RETURN(0);
 440 }
 441
 442 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 443 {
 444         int i;
 445         int rc;
 446         struct ptlrpc_srv_ni *srv_ni;
 447
 448         LASSERT (list_empty (&service->srv_threads));
 449
 450         /* XXX We could reply (with failure) to all buffered requests
 451          * _after_ unlinking _all_ the request buffers, but _before_
 452          * freeing them.
 453          */
 454
 455         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 456                 srv_ni = &service->srv_interfaces[i];
 457                 CDEBUG (D_NET, "%s: tearing down interface %s\n",
 458                         service->srv_name, srv_ni->sni_ni->pni_name);
 459
 460                 while (!list_empty (&srv_ni->sni_rqbds)) {
 461                         struct ptlrpc_request_buffer_desc *rqbd =
 462                                 list_entry (srv_ni->sni_rqbds.next,
 463                                             struct ptlrpc_request_buffer_desc,
 464                                             rqbd_list);
 465
 466                         list_del (&rqbd->rqbd_list);
 467
 468                         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 469                         /* refcount could be anything; it's possible for
 470                          * the buffers to continued to get filled after all
 471                          * the server threads exited.  But we know they
 472                          * _have_ exited.
 473                          */
 474
 475                         (void) PtlMEUnlink(rqbd->rqbd_me_h);
 476                         /* The callback handler could have unlinked this ME
 477                          * already (we're racing with her) but it's safe to
 478                          * ensure it _has_ been unlinked.
 479                          */
 480
 481                         OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 482                         OBD_FREE (rqbd, sizeof (*rqbd));
 483                         srv_ni->sni_nrqbds--;
 484                 }
 485
 486                 LASSERT (srv_ni->sni_nrqbds == 0);
 487
 488                 if (ptl_is_valid_handle (&srv_ni->sni_eq_h)) {
 489                         rc = PtlEQFree(srv_ni->sni_eq_h);
 490                         if (rc)
 491                                 CERROR("%s.%d: PtlEQFree failed on %s: %d\n",
 492                                        service->srv_name, i,
 493                                        srv_ni->sni_ni->pni_name, rc);
 494                 }
 495         }
 496
 497         OBD_FREE(service,
 498                  offsetof (struct ptlrpc_service,
 499                            srv_interfaces[ptlrpc_ninterfaces]));
 500         return 0;
 501 }