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 {
 146         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 147         struct ptlrpc_request request;
 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
 228         ENTRY;
 229
 230         lock_kernel();
 231         daemonize();
 232         spin_lock_irq(&current->sigmask_lock);
 233         sigfillset(&current->blocked);
 234         recalc_sigpending(current);
 235         spin_unlock_irq(&current->sigmask_lock);
 236
 237         sprintf(current->comm, data->name);
 238         unlock_kernel();
 239
 240         /* Record that the thread is running */
 241         thread->t_flags = SVC_RUNNING;
 242         wake_up(&thread->t_ctl_waitq);
 243
 244         /* XXX maintain a list of all managed devices: insert here */
 245
 246         /* And now, loop forever on requests */
 247         while (1) {
 248                 ptl_event_t event;
 249
 250                 wait_event(svc->srv_waitq,
 251                            ptlrpc_check_event(svc, thread, &event));
 252
 253                 spin_lock(&svc->srv_lock);
 254
 255                 if (thread->t_flags & SVC_STOPPING) {
 256                         thread->t_flags &= ~SVC_STOPPING;
 257                         spin_unlock(&svc->srv_lock);
 258                         EXIT;
 259                         break;
 260                 }
 261
 262                 if (thread->t_flags & SVC_EVENT) {
 263                         LASSERT (event.sequence != 0);
 264                         rc = handle_incoming_request(obddev, svc, &event);
 265                         thread->t_flags &= ~SVC_EVENT;
 266                         continue;
 267                 }
 268
 269                 CERROR("unknown break in service");
 270                 spin_unlock(&svc->srv_lock);
 271                 EXIT;
 272                 break;
 273         }
 274
 275         thread->t_flags = SVC_STOPPED;
 276         wake_up(&thread->t_ctl_waitq);
 277         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 278         return 0;
 279 }
 280
 281 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 282                                struct ptlrpc_thread *thread)
 283 {
 284         spin_lock(&svc->srv_lock);
 285         thread->t_flags = SVC_STOPPING;
 286         spin_unlock(&svc->srv_lock);
 287
 288         wake_up(&svc->srv_waitq);
 289         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 290 }
 291
 292 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 293 {
 294         spin_lock(&svc->srv_lock);
 295         while (!list_empty(&svc->srv_threads)) {
 296                 struct ptlrpc_thread *thread;
 297                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 298                                     t_link);
 299                 spin_unlock(&svc->srv_lock);
 300                 ptlrpc_stop_thread(svc, thread);
 301                 spin_lock(&svc->srv_lock);
 302                 list_del(&thread->t_link);
 303                 OBD_FREE(thread, sizeof(*thread));
 304         }
 305         spin_unlock(&svc->srv_lock);
 306 }
 307
 308 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 309                         char *name)
 310 {
 311         struct ptlrpc_svc_data d;
 312         struct ptlrpc_thread *thread;
 313         int rc;
 314         ENTRY;
 315
 316         OBD_ALLOC(thread, sizeof(*thread));
 317         if (thread == NULL) {
 318                 LBUG();
 319                 RETURN(-ENOMEM);
 320         }
 321         init_waitqueue_head(&thread->t_ctl_waitq);
 322
 323         d.dev = dev;
 324         d.svc = svc;
 325         d.name = name;
 326         d.thread = thread;
 327
 328         spin_lock(&svc->srv_lock);
 329         list_add(&thread->t_link, &svc->srv_threads);
 330         spin_unlock(&svc->srv_lock);
 331
 332         rc = kernel_thread(ptlrpc_main, (void *) &d,
 333                            CLONE_VM | CLONE_FS | CLONE_FILES);
 334         if (rc < 0) {
 335                 CERROR("cannot start thread\n");
 336                 OBD_FREE(thread, sizeof(*thread));
 337                 RETURN(-EINVAL);
 338         }
 339         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 340
 341         RETURN(0);
 342 }
 343
 344 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 345 {
 346         int rc, i;
 347
 348         /* NB service->srv_nbuffs gets set before we attempt (and possibly
 349          * fail) to allocate srv_rqbds.
 350          */
 351         if (service->srv_rqbds != NULL) {
 352                 for (i = 0; i < service->srv_nbuffs; i++) {
 353                         struct ptlrpc_request_buffer_desc *rqbd =
 354                                 &service->srv_rqbds[i];
 355
 356                         if (rqbd->rqbd_buffer == NULL) /* no buffer allocated */
 357                                 continue;             /* => never initialised */
 358
 359                         /* Buffer allocated => got linked */
 360                         LASSERT (ptl_is_valid_handle (&rqbd->rqbd_me_h));
 361
 362                         rc = PtlMEUnlink(rqbd->rqbd_me_h);
 363                         if (rc)
 364                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 365
 366                         OBD_FREE(rqbd->rqbd_buffer, service->srv_buf_size);
 367                 }
 368
 369                 OBD_FREE(service->srv_rqbds, service->srv_nbuffs *
 370                          sizeof (struct ptlrpc_request_buffer_desc));
 371         }
 372
 373         rc = PtlEQFree(service->srv_eq_h);
 374         if (rc)
 375                 CERROR("PtlEQFree failed: %d\n", rc);
 376
 377         if (!list_empty(&service->srv_reqs)) {
 378                 // XXX reply with errors and clean up
 379                 CERROR("Request list not empty!\n");
 380                 rc = -EBUSY;
 381         }
 382
 383         OBD_FREE(service, sizeof(*service));
 384         if (rc)
 385                 LBUG();
 386         return rc;
 387 }