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, char *uuid,
  68                 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,
 112                   service->srv_nbuffs * 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 %Lx xid %Ld\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 %Lx xid %Ld\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 %Lx xid %Ld\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 %Ld\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 *) request.rq_reqmsg);
 204
 205         if (request.rq_export) {
 206                 request.rq_connection = request.rq_export->exp_connection;
 207                 ptlrpc_connection_addref(request.rq_connection);
 208         }
 209
 210         spin_unlock(&svc->srv_lock);
 211
 212         rc = svc->srv_handler(&request);
 213         ptlrpc_put_connection(request.rq_connection);
 214
 215         ptlrpc_link_svc_me (rqbd);
 216         return rc;
 217 }
 218
 219 static int ptlrpc_main(void *arg)
 220 {
 221         int rc;
 222         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 223         struct obd_device *obddev = data->dev;
 224         struct ptlrpc_service *svc = data->svc;
 225         struct ptlrpc_thread *thread = data->thread;
 226
 227         ENTRY;
 228
 229         lock_kernel();
 230         daemonize();
 231         spin_lock_irq(&current->sigmask_lock);
 232         sigfillset(&current->blocked);
 233         recalc_sigpending(current);
 234         spin_unlock_irq(&current->sigmask_lock);
 235
 236         sprintf(current->comm, data->name);
 237         unlock_kernel();
 238
 239         /* Record that the thread is running */
 240         thread->t_flags = SVC_RUNNING;
 241         wake_up(&thread->t_ctl_waitq);
 242
 243         /* XXX maintain a list of all managed devices: insert here */
 244
 245         /* And now, loop forever on requests */
 246         while (1) {
 247                 ptl_event_t event;
 248
 249                 wait_event(svc->srv_waitq,
 250                            ptlrpc_check_event(svc, thread, &event));
 251
 252                 spin_lock(&svc->srv_lock);
 253
 254                 if (thread->t_flags & SVC_STOPPING) {
 255                         thread->t_flags &= ~SVC_STOPPING;
 256                         spin_unlock(&svc->srv_lock);
 257                         EXIT;
 258                         break;
 259                 }
 260
 261                 if (thread->t_flags & SVC_EVENT) {
 262                         LASSERT (event.sequence != 0);
 263                         rc = handle_incoming_request(obddev, svc, &event);
 264                         thread->t_flags &= ~SVC_EVENT;
 265                         continue;
 266                 }
 267
 268                 CERROR("unknown break in service");
 269                 spin_unlock(&svc->srv_lock);
 270                 EXIT;
 271                 break;
 272         }
 273
 274         thread->t_flags = SVC_STOPPED;
 275         wake_up(&thread->t_ctl_waitq);
 276         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 277         return 0;
 278 }
 279
 280 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 281                                struct ptlrpc_thread *thread)
 282 {
 283         spin_lock(&svc->srv_lock);
 284         thread->t_flags = SVC_STOPPING;
 285         spin_unlock(&svc->srv_lock);
 286
 287         wake_up(&svc->srv_waitq);
 288         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 289 }
 290
 291 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 292 {
 293         spin_lock(&svc->srv_lock);
 294         while (!list_empty(&svc->srv_threads)) {
 295                 struct ptlrpc_thread *thread;
 296                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 297                                     t_link);
 298                 spin_unlock(&svc->srv_lock);
 299                 ptlrpc_stop_thread(svc, thread);
 300                 spin_lock(&svc->srv_lock);
 301                 list_del(&thread->t_link);
 302                 OBD_FREE(thread, sizeof(*thread));
 303         }
 304         spin_unlock(&svc->srv_lock);
 305 }
 306
 307 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 308                         char *name)
 309 {
 310         struct ptlrpc_svc_data d;
 311         struct ptlrpc_thread *thread;
 312         int rc;
 313         ENTRY;
 314
 315         OBD_ALLOC(thread, sizeof(*thread));
 316         if (thread == NULL) {
 317                 LBUG();
 318                 RETURN(-ENOMEM);
 319         }
 320         init_waitqueue_head(&thread->t_ctl_waitq);
 321
 322         d.dev = dev;
 323         d.svc = svc;
 324         d.name = name;
 325         d.thread = thread;
 326
 327         spin_lock(&svc->srv_lock);
 328         list_add(&thread->t_link, &svc->srv_threads);
 329         spin_unlock(&svc->srv_lock);
 330
 331         rc = kernel_thread(ptlrpc_main, (void *) &d,
 332                            CLONE_VM | CLONE_FS | CLONE_FILES);
 333         if (rc < 0) {
 334                 CERROR("cannot start thread\n");
 335                 OBD_FREE(thread, sizeof(*thread));
 336                 RETURN(-EINVAL);
 337         }
 338         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 339
 340         RETURN(0);
 341 }
 342
 343 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 344 {
 345         int rc, i;
 346
 347         if (service->srv_rqbds != NULL)
 348         {
 349                 for (i = 0; i < service->srv_nbuffs; i++) {
 350                         struct ptlrpc_request_buffer_desc *rqbd = &service->srv_rqbds[i];
 351
 352                         if (rqbd->rqbd_buffer == NULL)  /* no buffer allocated */
 353                                 continue;               /* => never initialised */
 354
 355                         /* Buffer allocated => got linked */
 356                         LASSERT (ptl_is_valid_handle (&rqbd->rqbd_me_h));
 357
 358                         rc = PtlMEUnlink(rqbd->rqbd_me_h);
 359                         if (rc)
 360                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 361
 362                         OBD_FREE(rqbd->rqbd_buffer, service->srv_buf_size);
 363                 }
 364
 365                 OBD_FREE(service->srv_rqbds,
 366                          service->srv_nbuffs * sizeof (struct ptlrpc_request_buffer_desc));
 367         }
 368
 369         rc = PtlEQFree(service->srv_eq_h);
 370         if (rc)
 371                 CERROR("PtlEQFree failed: %d\n", rc);
 372
 373         if (!list_empty(&service->srv_reqs)) {
 374                 // XXX reply with errors and clean up
 375                 CERROR("Request list not empty!\n");
 376                 rc = -EBUSY;
 377         }
 378
 379         OBD_FREE(service, sizeof(*service));
 380         if (rc)
 381                 LBUG();
 382         return rc;
 383 }