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
  25 #include <linux/obd_support.h>
  26 #include <linux/obd_class.h>
  27 #include <linux/lustre_net.h>
  28
  29 extern int request_in_callback(ptl_event_t *ev);
  30
  31 static int ptlrpc_check_event(struct ptlrpc_service *svc,
  32                               struct ptlrpc_thread *thread, ptl_event_t *event)
  33 {
  34         int rc = 0;
  35         ENTRY;
  36
  37         spin_lock(&svc->srv_lock);
  38         if (thread->t_flags & SVC_STOPPING)
  39                 GOTO(out, rc = 1);
  40
  41         LASSERT ((thread->t_flags & SVC_EVENT) == 0);
  42
  43         if (ptl_is_valid_handle(&svc->srv_eq_h)) {
  44                 int err;
  45                 err = PtlEQGet(svc->srv_eq_h, event);
  46
  47                 if (err == PTL_OK) {
  48                         thread->t_flags |= SVC_EVENT;
  49                         GOTO(out, rc = 1);
  50                 }
  51
  52                 if (err != PTL_EQ_EMPTY) {
  53                         CERROR("BUG: PtlEQGet returned %d\n", rc);
  54                         LBUG();
  55                 }
  56
  57                 GOTO(out, rc = 0);
  58         }
  59
  60         EXIT;
  61  out:
  62         spin_unlock(&svc->srv_lock);
  63         return rc;
  64 }
  65
  66 struct ptlrpc_service *
  67 ptlrpc_init_svc(__u32 bufsize, int nbuffs, int req_portal, int rep_portal,
  68                 obd_uuid_t uuid, svc_handler_t handler, char *name)
  69 {
  70         int err;
  71         int rc, i;
  72         struct ptlrpc_service *service;
  73         ENTRY;
  74
  75         OBD_ALLOC(service, sizeof(*service));
  76         if (!service) {
  77                 LBUG();
  78                 RETURN(NULL);
  79         }
  80
  81         service->srv_name = name;
  82         spin_lock_init(&service->srv_lock);
  83         INIT_LIST_HEAD(&service->srv_reqs);
  84         INIT_LIST_HEAD(&service->srv_threads);
  85         init_waitqueue_head(&service->srv_waitq);
  86
  87         service->srv_buf_size = bufsize;
  88         service->srv_nbuffs = nbuffs;
  89         service->srv_rep_portal = rep_portal;
  90         service->srv_req_portal = req_portal;
  91         service->srv_handler = handler;
  92
  93         err = kportal_uuid_to_peer(uuid, &service->srv_self);
  94         if (err) {
  95                 CERROR("cannot get peer for uuid '%s'\n", uuid);
  96                 OBD_FREE(service, sizeof(*service));
  97                 RETURN(NULL);
  98         }
  99
 100         /* NB We need exactly 1 event for each buffer we queue */
 101         rc = PtlEQAlloc(service->srv_self.peer_ni, service->srv_nbuffs,
 102                         request_in_callback, &(service->srv_eq_h));
 103
 104         if (rc != PTL_OK) {
 105                 CERROR("PtlEQAlloc failed: %d\n", rc);
 106                 LBUG();
 107                 OBD_FREE(service, sizeof(*service));
 108                 RETURN(NULL);
 109         }
 110
 111         OBD_ALLOC(service->srv_rqbds, service->srv_nbuffs *
 112                   sizeof(struct ptlrpc_request_buffer_desc));
 113         if (service->srv_rqbds == NULL) {
 114                 CERROR("no memory\n");
 115                 LBUG();
 116                 GOTO(failed, NULL);
 117         }
 118
 119         for (i = 0; i < service->srv_nbuffs; i++) {
 120                 struct ptlrpc_request_buffer_desc *rqbd =&service->srv_rqbds[i];
 121
 122                 rqbd->rqbd_service = service;
 123                 ptl_set_inv_handle (&rqbd->rqbd_me_h);
 124                 OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 125                 if (rqbd->rqbd_buffer == NULL) {
 126                         CERROR("no memory\n");
 127                         LBUG();
 128                         GOTO(failed, NULL);
 129                 }
 130                 ptlrpc_link_svc_me(rqbd);
 131         }
 132
 133         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 134                service->srv_req_portal);
 135
 136         RETURN(service);
 137 failed:
 138         ptlrpc_unregister_service(service);
 139         return NULL;
 140 }
 141
 142 static int handle_incoming_request(struct obd_device *obddev,
 143                                    struct ptlrpc_service *svc,
 144                                    ptl_event_t *event,
 145                                    struct ptlrpc_request *request)
 146 {
 147         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 148         int rc;
 149
 150         /* FIXME: If we move to an event-driven model, we should put the request
 151          * on the stack of mds_handle instead. */
 152         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 153         LASSERT (rqbd->rqbd_service == svc);
 154         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 155         LASSERT (event->offset == 0);
 156
 157         memset(request, 0, sizeof(*request));
 158         request->rq_svc = svc;
 159         request->rq_obd = obddev;
 160         request->rq_xid = event->match_bits;
 161         request->rq_reqmsg = event->mem_desc.start + event->offset;
 162         request->rq_reqlen = event->mem_desc.length;
 163
 164         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 165                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "LPD64"\n",
 166                        request->rq_reqlen, svc->srv_req_portal,
 167                        event->initiator.nid, request->rq_xid);
 168                 spin_unlock(&svc->srv_lock);
 169                 RETURN(-EINVAL);
 170         }
 171
 172         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 173                 CERROR("wrong packet type received (type=%u)\n",
 174                        request->rq_reqmsg->type);
 175                 LBUG();
 176                 spin_unlock(&svc->srv_lock);
 177                 RETURN(-EINVAL);
 178         }
 179
 180         if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 181                 CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "LPD64"\n",
 182                        request->rq_reqmsg->magic, svc->srv_req_portal,
 183                        event->initiator.nid, request->rq_xid);
 184                 spin_unlock(&svc->srv_lock);
 185                 RETURN(-EINVAL);
 186         }
 187
 188         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 189                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "LPD64"\n",
 190                        request->rq_reqmsg->version, svc->srv_req_portal,
 191                        event->initiator.nid, request->rq_xid);
 192                 spin_unlock(&svc->srv_lock);
 193                 RETURN(-EINVAL);
 194         }
 195
 196         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 197
 198         request->rq_peer.peer_nid = event->initiator.nid;
 199         /* FIXME: this NI should be the incoming NI.
 200          * We don't know how to find that from here. */
 201         request->rq_peer.peer_ni = svc->srv_self.peer_ni;
 202
 203         request->rq_export = class_conn2export((struct lustre_handle *)
 204                                                request->rq_reqmsg);
 205
 206         if (request->rq_export) {
 207                 request->rq_connection = request->rq_export->exp_connection;
 208                 ptlrpc_connection_addref(request->rq_connection);
 209         }
 210
 211         spin_unlock(&svc->srv_lock);
 212
 213         rc = svc->srv_handler(request);
 214         ptlrpc_put_connection(request->rq_connection);
 215
 216         ptlrpc_link_svc_me (rqbd);
 217         return rc;
 218 }
 219
 220 static int ptlrpc_main(void *arg)
 221 {
 222         int rc;
 223         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 224         struct obd_device *obddev = data->dev;
 225         struct ptlrpc_service *svc = data->svc;
 226         struct ptlrpc_thread *thread = data->thread;
 227         struct ptlrpc_request *request;
 228         ptl_event_t *event;
 229
 230         ENTRY;
 231
 232         lock_kernel();
 233         daemonize();
 234         spin_lock_irq(&current->sigmask_lock);
 235         sigfillset(&current->blocked);
 236         recalc_sigpending(current);
 237         spin_unlock_irq(&current->sigmask_lock);
 238
 239         sprintf(current->comm, data->name);
 240         unlock_kernel();
 241
 242         /* Record that the thread is running */
 243         thread->t_flags = SVC_RUNNING;
 244         wake_up(&thread->t_ctl_waitq);
 245
 246         /* XXX maintain a list of all managed devices: insert here */
 247
 248         OBD_ALLOC(event, sizeof(*event));
 249         LASSERT(event);
 250         OBD_ALLOC(request, sizeof(*request));
 251         LASSERT(request);
 252
 253         /* And now, loop forever on requests */
 254         while (1) {
 255                 wait_event(svc->srv_waitq,
 256                            ptlrpc_check_event(svc, thread, event));
 257
 258                 spin_lock(&svc->srv_lock);
 259
 260                 if (thread->t_flags & SVC_STOPPING) {
 261                         thread->t_flags &= ~SVC_STOPPING;
 262                         spin_unlock(&svc->srv_lock);
 263                         EXIT;
 264                         break;
 265                 }
 266
 267                 if (thread->t_flags & SVC_EVENT) {
 268                         LASSERT (event->sequence != 0);
 269                         rc = handle_incoming_request(obddev, svc, event,
 270                                                      request);
 271                         thread->t_flags &= ~SVC_EVENT;
 272                         continue;
 273                 }
 274
 275                 CERROR("unknown break in service");
 276                 spin_unlock(&svc->srv_lock);
 277                 EXIT;
 278                 break;
 279         }
 280
 281         OBD_FREE(event, sizeof(*event));
 282         OBD_FREE(request, sizeof(*request));
 283
 284         thread->t_flags = SVC_STOPPED;
 285         wake_up(&thread->t_ctl_waitq);
 286         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 287         return 0;
 288 }
 289
 290 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 291                                struct ptlrpc_thread *thread)
 292 {
 293         spin_lock(&svc->srv_lock);
 294         thread->t_flags = SVC_STOPPING;
 295         spin_unlock(&svc->srv_lock);
 296
 297         wake_up(&svc->srv_waitq);
 298         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 299 }
 300
 301 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 302 {
 303         spin_lock(&svc->srv_lock);
 304         while (!list_empty(&svc->srv_threads)) {
 305                 struct ptlrpc_thread *thread;
 306                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 307                                     t_link);
 308                 spin_unlock(&svc->srv_lock);
 309                 ptlrpc_stop_thread(svc, thread);
 310                 spin_lock(&svc->srv_lock);
 311                 list_del(&thread->t_link);
 312                 OBD_FREE(thread, sizeof(*thread));
 313         }
 314         spin_unlock(&svc->srv_lock);
 315 }
 316
 317 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 318                         char *name)
 319 {
 320         struct ptlrpc_svc_data d;
 321         struct ptlrpc_thread *thread;
 322         int rc;
 323         ENTRY;
 324
 325         OBD_ALLOC(thread, sizeof(*thread));
 326         if (thread == NULL) {
 327                 LBUG();
 328                 RETURN(-ENOMEM);
 329         }
 330         init_waitqueue_head(&thread->t_ctl_waitq);
 331
 332         d.dev = dev;
 333         d.svc = svc;
 334         d.name = name;
 335         d.thread = thread;
 336
 337         spin_lock(&svc->srv_lock);
 338         list_add(&thread->t_link, &svc->srv_threads);
 339         spin_unlock(&svc->srv_lock);
 340
 341         rc = kernel_thread(ptlrpc_main, (void *) &d,
 342                            CLONE_VM | CLONE_FS | CLONE_FILES);
 343         if (rc < 0) {
 344                 CERROR("cannot start thread\n");
 345                 OBD_FREE(thread, sizeof(*thread));
 346                 RETURN(-EINVAL);
 347         }
 348         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 349
 350         RETURN(0);
 351 }
 352
 353 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 354 {
 355         int rc, i;
 356
 357         /* NB service->srv_nbuffs gets set before we attempt (and possibly
 358          * fail) to allocate srv_rqbds.
 359          */
 360         if (service->srv_rqbds != NULL) {
 361                 for (i = 0; i < service->srv_nbuffs; i++) {
 362                         struct ptlrpc_request_buffer_desc *rqbd =
 363                                 &service->srv_rqbds[i];
 364
 365                         if (rqbd->rqbd_buffer == NULL) /* no buffer allocated */
 366                                 continue;             /* => never initialised */
 367
 368                         /* Buffer allocated => got linked */
 369                         LASSERT (ptl_is_valid_handle (&rqbd->rqbd_me_h));
 370
 371                         rc = PtlMEUnlink(rqbd->rqbd_me_h);
 372                         if (rc)
 373                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 374
 375                         OBD_FREE(rqbd->rqbd_buffer, service->srv_buf_size);
 376                 }
 377
 378                 OBD_FREE(service->srv_rqbds, service->srv_nbuffs *
 379                          sizeof (struct ptlrpc_request_buffer_desc));
 380         }
 381
 382         rc = PtlEQFree(service->srv_eq_h);
 383         if (rc)
 384                 CERROR("PtlEQFree failed: %d\n", rc);
 385
 386         if (!list_empty(&service->srv_reqs)) {
 387                 // XXX reply with errors and clean up
 388                 CERROR("Request list not empty!\n");
 389                 rc = -EBUSY;
 390         }
 391
 392         OBD_FREE(service, sizeof(*service));
 393         if (rc)
 394                 LBUG();
 395         return rc;
 396 }