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 nevents, __u32 nbufs,
  68                 __u32 bufsize, __u32 max_req_size,
  69                 int req_portal, int rep_portal,
  70                 obd_uuid_t uuid, svc_handler_t handler, char *name)
  71 {
  72         int err;
  73         int rc, i;
  74         struct ptlrpc_service *service;
  75         ENTRY;
  76
  77         OBD_ALLOC(service, sizeof(*service));
  78         if (!service) {
  79                 LBUG();
  80                 RETURN(NULL);
  81         }
  82
  83         service->srv_name = name;
  84         spin_lock_init(&service->srv_lock);
  85         INIT_LIST_HEAD(&service->srv_threads);
  86         init_waitqueue_head(&service->srv_waitq);
  87
  88         service->srv_max_req_size = max_req_size;
  89         service->srv_buf_size = bufsize;
  90         INIT_LIST_HEAD (&service->srv_rqbds);
  91         service->srv_nrqbds = 0;
  92         atomic_set (&service->srv_nrqbds_receiving, 0);
  93
  94         service->srv_rep_portal = rep_portal;
  95         service->srv_req_portal = req_portal;
  96         service->srv_handler = handler;
  97
  98         err = kportal_uuid_to_peer(uuid, &service->srv_self);
  99         if (err) {
 100                 CERROR("cannot get peer for uuid '%s'\n", uuid);
 101                 OBD_FREE(service, sizeof(*service));
 102                 RETURN(NULL);
 103         }
 104
 105         rc = PtlEQAlloc(service->srv_self.peer_ni, nevents,
 106                         request_in_callback, &(service->srv_eq_h));
 107
 108         if (rc != PTL_OK) {
 109                 CERROR("PtlEQAlloc failed: %d\n", rc);
 110                 OBD_FREE(service, sizeof(*service));
 111                 RETURN(NULL);
 112         }
 113
 114         for (i = 0; i < nbufs; i++) {
 115                 struct ptlrpc_request_buffer_desc *rqbd;
 116
 117                 OBD_ALLOC (rqbd, sizeof (*rqbd));
 118                 if (rqbd == NULL)
 119                 {
 120                         CERROR ("no memory\n");
 121                         GOTO (failed, NULL);
 122                 }
 123
 124                 rqbd->rqbd_service = service;
 125                 ptl_set_inv_handle (&rqbd->rqbd_me_h);
 126                 atomic_set (&rqbd->rqbd_refcount, 0);
 127                 OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 128                 if (rqbd->rqbd_buffer == NULL) {
 129                         OBD_FREE (rqbd, sizeof (*rqbd));
 130                         CERROR("no memory\n");
 131                         GOTO(failed, NULL);
 132                 }
 133                 list_add (&rqbd->rqbd_list, &service->srv_rqbds);
 134                 service->srv_nrqbds++;
 135
 136                 ptlrpc_link_svc_me(rqbd);
 137         }
 138
 139         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 140                service->srv_req_portal);
 141
 142         RETURN(service);
 143 failed:
 144         ptlrpc_unregister_service(service);
 145         return NULL;
 146 }
 147
 148 static int handle_incoming_request(struct obd_device *obddev,
 149                                    struct ptlrpc_service *svc,
 150                                    ptl_event_t *event,
 151                                    struct ptlrpc_request *request)
 152 {
 153         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 154         int rc;
 155
 156         /* FIXME: If we move to an event-driven model, we should put the request
 157          * on the stack of mds_handle instead. */
 158
 159         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 160         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 161         LASSERT (rqbd->rqbd_service == svc);
 162         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 163         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 164
 165         memset(request, 0, sizeof(*request));
 166         request->rq_svc = svc;
 167         request->rq_obd = obddev;
 168         request->rq_xid = event->match_bits;
 169         request->rq_reqmsg = event->mem_desc.start + event->offset;
 170         request->rq_reqlen = event->mlength;
 171
 172         rc = -EINVAL;
 173
 174         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 175                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "LPD64"\n",
 176                        request->rq_reqlen, svc->srv_req_portal,
 177                        event->initiator.nid, request->rq_xid);
 178                 goto out;
 179         }
 180
 181         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 182                 CERROR("wrong packet type received (type=%u)\n",
 183                        request->rq_reqmsg->type);
 184                 goto out;
 185         }
 186
 187         if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 188                 CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "LPD64"\n",
 189                        request->rq_reqmsg->magic, svc->srv_req_portal,
 190                        event->initiator.nid, request->rq_xid);
 191                 goto out;
 192         }
 193
 194         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 195                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "LPD64"\n",
 196                        request->rq_reqmsg->version, svc->srv_req_portal,
 197                        event->initiator.nid, request->rq_xid);
 198                 goto out;
 199         }
 200
 201         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 202
 203         request->rq_peer.peer_nid = event->initiator.nid;
 204         /* FIXME: this NI should be the incoming NI.
 205          * We don't know how to find that from here. */
 206         request->rq_peer.peer_ni = svc->srv_self.peer_ni;
 207
 208         request->rq_export = class_conn2export((struct lustre_handle *)
 209                                                request->rq_reqmsg);
 210
 211         if (request->rq_export) {
 212                 request->rq_connection = request->rq_export->exp_connection;
 213                 ptlrpc_connection_addref(request->rq_connection);
 214         }
 215
 216         rc = svc->srv_handler(request);
 217         ptlrpc_put_connection(request->rq_connection);
 218
 219  out:
 220         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 221                 ptlrpc_link_svc_me (rqbd);
 222
 223         return rc;
 224 }
 225
 226 static int ptlrpc_main(void *arg)
 227 {
 228         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 229         struct obd_device *obddev = data->dev;
 230         struct ptlrpc_service *svc = data->svc;
 231         struct ptlrpc_thread *thread = data->thread;
 232         struct ptlrpc_request *request;
 233         ptl_event_t *event;
 234         int rc = 0;
 235
 236         ENTRY;
 237
 238         lock_kernel();
 239         daemonize();
 240         spin_lock_irq(&current->sigmask_lock);
 241         sigfillset(&current->blocked);
 242         recalc_sigpending(current);
 243         spin_unlock_irq(&current->sigmask_lock);
 244
 245         strcpy(current->comm, data->name);
 246         unlock_kernel();
 247
 248         OBD_ALLOC(event, sizeof(*event));
 249         if (!event)
 250                 GOTO(out, rc = -ENOMEM);
 251         OBD_ALLOC(request, sizeof(*request));
 252         if (!request)
 253                 GOTO(out_event, rc = -ENOMEM);
 254
 255         /* Record that the thread is running */
 256         thread->t_flags = SVC_RUNNING;
 257         wake_up(&thread->t_ctl_waitq);
 258
 259         /* XXX maintain a list of all managed devices: insert here */
 260
 261         /* And now, loop forever on requests */
 262         while (1) {
 263                 wait_event(svc->srv_waitq,
 264                            ptlrpc_check_event(svc, thread, event));
 265
 266                 if (thread->t_flags & SVC_STOPPING) {
 267                         spin_lock(&svc->srv_lock);
 268                         thread->t_flags &= ~SVC_STOPPING;
 269                         spin_unlock(&svc->srv_lock);
 270
 271                         EXIT;
 272                         break;
 273                 }
 274
 275                 if (thread->t_flags & SVC_EVENT) {
 276                         spin_lock(&svc->srv_lock);
 277                         thread->t_flags &= ~SVC_EVENT;
 278                         spin_unlock(&svc->srv_lock);
 279
 280                         rc = handle_incoming_request(obddev, svc, event,
 281                                                      request);
 282                         continue;
 283                 }
 284
 285                 CERROR("unknown break in service");
 286                 LBUG();
 287                 EXIT;
 288                 break;
 289         }
 290
 291         OBD_FREE(request, sizeof(*request));
 292 out_event:
 293         OBD_FREE(event, sizeof(*event));
 294 out:
 295         thread->t_flags = SVC_STOPPED;
 296         wake_up(&thread->t_ctl_waitq);
 297
 298         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 299                current->pid, rc);
 300         return rc;
 301 }
 302
 303 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 304                                struct ptlrpc_thread *thread)
 305 {
 306         spin_lock(&svc->srv_lock);
 307         thread->t_flags = SVC_STOPPING;
 308         spin_unlock(&svc->srv_lock);
 309
 310         wake_up(&svc->srv_waitq);
 311         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 312 }
 313
 314 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 315 {
 316         spin_lock(&svc->srv_lock);
 317         while (!list_empty(&svc->srv_threads)) {
 318                 struct ptlrpc_thread *thread;
 319                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 320                                     t_link);
 321                 spin_unlock(&svc->srv_lock);
 322                 ptlrpc_stop_thread(svc, thread);
 323                 spin_lock(&svc->srv_lock);
 324                 list_del(&thread->t_link);
 325                 OBD_FREE(thread, sizeof(*thread));
 326         }
 327         spin_unlock(&svc->srv_lock);
 328 }
 329
 330 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 331                         char *name)
 332 {
 333         struct ptlrpc_svc_data d;
 334         struct ptlrpc_thread *thread;
 335         int rc;
 336         ENTRY;
 337
 338         OBD_ALLOC(thread, sizeof(*thread));
 339         if (thread == NULL) {
 340                 LBUG();
 341                 RETURN(-ENOMEM);
 342         }
 343         init_waitqueue_head(&thread->t_ctl_waitq);
 344
 345         d.dev = dev;
 346         d.svc = svc;
 347         d.name = name;
 348         d.thread = thread;
 349
 350         spin_lock(&svc->srv_lock);
 351         list_add(&thread->t_link, &svc->srv_threads);
 352         spin_unlock(&svc->srv_lock);
 353
 354         /* XXX should we really be cloning open file handles here? */
 355         rc = kernel_thread(ptlrpc_main, (void *) &d,
 356                            CLONE_VM | CLONE_FS | CLONE_FILES);
 357         if (rc < 0) {
 358                 CERROR("cannot start thread\n");
 359                 OBD_FREE(thread, sizeof(*thread));
 360                 RETURN(rc);
 361         }
 362         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 363
 364         RETURN(0);
 365 }
 366
 367 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 368 {
 369         int rc;
 370
 371         LASSERT (list_empty (&service->srv_threads));
 372
 373         /* XXX We could reply (with failure) to all buffered requests
 374          * _after_ unlinking _all_ the request buffers, but _before_
 375          * freeing them.
 376          */
 377
 378         while (!list_empty (&service->srv_rqbds))
 379         {
 380                 struct ptlrpc_request_buffer_desc *rqbd =
 381                         list_entry (service->srv_rqbds.next,
 382                                     struct ptlrpc_request_buffer_desc,
 383                                     rqbd_list);
 384
 385                 list_del (&rqbd->rqbd_list);
 386
 387                 LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 388                 /* refcount could be anything; it's possible for the
 389                  * buffers to continued to get filled after all the server
 390                  * threads exited.  But we know they _have_ exited.
 391                  */
 392
 393                 (void) PtlMEUnlink(rqbd->rqbd_me_h);
 394                 /* The callback handler could have unlinked this ME already
 395                  * (we're racing with her) but it's safe to ensure it _has_
 396                  * been unlinked.
 397                  */
 398
 399                 OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 400                 OBD_FREE (rqbd, sizeof (*rqbd));
 401                 service->srv_nrqbds--;
 402         }
 403
 404         LASSERT (service->srv_nrqbds == 0);
 405
 406         rc = PtlEQFree(service->srv_eq_h);
 407         if (rc)
 408                 CERROR("PtlEQFree failed: %d\n", rc);
 409
 410         OBD_FREE(service, sizeof(*service));
 411         if (rc)
 412                 LBUG();
 413         return rc;
 414 }