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", err);
  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                 RETURN(NULL);
  80
  81         service->srv_name = name;
  82         spin_lock_init(&service->srv_lock);
  83         INIT_LIST_HEAD(&service->srv_threads);
  84         init_waitqueue_head(&service->srv_waitq);
  85
  86         service->srv_max_req_size = max_req_size;
  87         service->srv_buf_size = bufsize;
  88         INIT_LIST_HEAD(&service->srv_rqbds);
  89         service->srv_nrqbds = 0;
  90         atomic_set(&service->srv_nrqbds_receiving, 0);
  91
  92         service->srv_rep_portal = rep_portal;
  93         service->srv_req_portal = req_portal;
  94         service->srv_handler = handler;
  95
  96         err = kportal_uuid_to_peer(uuid, &service->srv_self);
  97         if (err) {
  98                 CERROR("%s: cannot get peer for uuid '%s'\n", name, uuid);
  99                 OBD_FREE(service, sizeof(*service));
 100                 RETURN(NULL);
 101         }
 102
 103         rc = PtlEQAlloc(service->srv_self.peer_ni, nevents,
 104                         request_in_callback, &(service->srv_eq_h));
 105
 106         if (rc != PTL_OK) {
 107                 CERROR("%s: PtlEQAlloc failed: %d\n", name, rc);
 108                 OBD_FREE(service, sizeof(*service));
 109                 RETURN(NULL);
 110         }
 111
 112         for (i = 0; i < nbufs; i++) {
 113                 struct ptlrpc_request_buffer_desc *rqbd;
 114
 115                 OBD_ALLOC(rqbd, sizeof(*rqbd));
 116                 if (rqbd == NULL)
 117                         GOTO(failed, NULL);
 118
 119                 rqbd->rqbd_service = service;
 120                 ptl_set_inv_handle(&rqbd->rqbd_me_h);
 121                 atomic_set(&rqbd->rqbd_refcount, 0);
 122                 OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 123                 if (rqbd->rqbd_buffer == NULL) {
 124                         OBD_FREE(rqbd, sizeof(*rqbd));
 125                         GOTO(failed, NULL);
 126                 }
 127                 list_add(&rqbd->rqbd_list, &service->srv_rqbds);
 128                 service->srv_nrqbds++;
 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
 153         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 154         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 155         LASSERT (rqbd->rqbd_service == svc);
 156         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 157         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 158
 159         memset(request, 0, sizeof(*request));
 160         request->rq_svc = svc;
 161         request->rq_obd = obddev;
 162         request->rq_xid = event->match_bits;
 163         request->rq_reqmsg = event->mem_desc.start + event->offset;
 164         request->rq_reqlen = event->mlength;
 165
 166         rc = -EINVAL;
 167
 168         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 169                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "
 170                        LPD64"\n",
 171                        request->rq_reqlen, svc->srv_req_portal,
 172                        event->initiator.nid, request->rq_xid);
 173                 goto out;
 174         }
 175
 176         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 177                 CERROR("wrong packet type received (type=%u)\n",
 178                        request->rq_reqmsg->type);
 179                 goto out;
 180         }
 181
 182         if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 183                 CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "
 184                        LPD64"\n",
 185                        request->rq_reqmsg->magic, svc->srv_req_portal,
 186                        event->initiator.nid, request->rq_xid);
 187                 goto out;
 188         }
 189
 190         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 191                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "
 192                        LPD64"\n",
 193                        request->rq_reqmsg->version, svc->srv_req_portal,
 194                        event->initiator.nid, request->rq_xid);
 195                 goto out;
 196         }
 197
 198         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 199
 200         request->rq_peer.peer_nid = event->initiator.nid;
 201         /* FIXME: this NI should be the incoming NI.
 202          * We don't know how to find that from here. */
 203         request->rq_peer.peer_ni = svc->srv_self.peer_ni;
 204
 205         request->rq_export = class_conn2export((struct lustre_handle *)
 206                                                request->rq_reqmsg);
 207
 208         if (request->rq_export) {
 209                 request->rq_connection = request->rq_export->exp_connection;
 210                 ptlrpc_connection_addref(request->rq_connection);
 211         } else {
 212                 /* create a (hopefully temporary) connection that will be used
 213                  * to send the reply if this call doesn't create an export.
 214                  * XXX revisit this when we revamp ptlrpc */
 215                 request->rq_connection =
 216                         ptlrpc_get_connection(&request->rq_peer, NULL);
 217         }
 218
 219         rc = svc->srv_handler(request);
 220         ptlrpc_put_connection(request->rq_connection);
 221
 222  out:
 223         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 224                 ptlrpc_link_svc_me (rqbd);
 225
 226         return rc;
 227 }
 228
 229 static int ptlrpc_main(void *arg)
 230 {
 231         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 232         struct obd_device *obddev = data->dev;
 233         struct ptlrpc_service *svc = data->svc;
 234         struct ptlrpc_thread *thread = data->thread;
 235         struct ptlrpc_request *request;
 236         ptl_event_t *event;
 237         int rc = 0;
 238
 239         ENTRY;
 240
 241         lock_kernel();
 242         daemonize();
 243         spin_lock_irq(&current->sigmask_lock);
 244         sigfillset(&current->blocked);
 245         recalc_sigpending(current);
 246         spin_unlock_irq(&current->sigmask_lock);
 247
 248         strcpy(current->comm, data->name);
 249         unlock_kernel();
 250
 251         OBD_ALLOC(event, sizeof(*event));
 252         if (!event)
 253                 GOTO(out, rc = -ENOMEM);
 254         OBD_ALLOC(request, sizeof(*request));
 255         if (!request)
 256                 GOTO(out_event, rc = -ENOMEM);
 257
 258         /* Record that the thread is running */
 259         thread->t_flags = SVC_RUNNING;
 260         wake_up(&thread->t_ctl_waitq);
 261
 262         /* XXX maintain a list of all managed devices: insert here */
 263
 264         /* And now, loop forever on requests */
 265         while (1) {
 266                 wait_event(svc->srv_waitq,
 267                            ptlrpc_check_event(svc, thread, event));
 268
 269                 if (thread->t_flags & SVC_STOPPING) {
 270                         spin_lock(&svc->srv_lock);
 271                         thread->t_flags &= ~SVC_STOPPING;
 272                         spin_unlock(&svc->srv_lock);
 273
 274                         EXIT;
 275                         break;
 276                 }
 277
 278                 if (thread->t_flags & SVC_EVENT) {
 279                         spin_lock(&svc->srv_lock);
 280                         thread->t_flags &= ~SVC_EVENT;
 281                         spin_unlock(&svc->srv_lock);
 282
 283                         rc = handle_incoming_request(obddev, svc, event,
 284                                                      request);
 285                         continue;
 286                 }
 287
 288                 CERROR("unknown break in service");
 289                 LBUG();
 290                 EXIT;
 291                 break;
 292         }
 293
 294         OBD_FREE(request, sizeof(*request));
 295 out_event:
 296         OBD_FREE(event, sizeof(*event));
 297 out:
 298         thread->t_flags = SVC_STOPPED;
 299         wake_up(&thread->t_ctl_waitq);
 300
 301         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 302                current->pid, rc);
 303         return rc;
 304 }
 305
 306 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 307                                struct ptlrpc_thread *thread)
 308 {
 309         spin_lock(&svc->srv_lock);
 310         thread->t_flags = SVC_STOPPING;
 311         spin_unlock(&svc->srv_lock);
 312
 313         wake_up(&svc->srv_waitq);
 314         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 315 }
 316
 317 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 318 {
 319         spin_lock(&svc->srv_lock);
 320         while (!list_empty(&svc->srv_threads)) {
 321                 struct ptlrpc_thread *thread;
 322                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 323                                     t_link);
 324                 spin_unlock(&svc->srv_lock);
 325                 ptlrpc_stop_thread(svc, thread);
 326                 spin_lock(&svc->srv_lock);
 327                 list_del(&thread->t_link);
 328                 OBD_FREE(thread, sizeof(*thread));
 329         }
 330         spin_unlock(&svc->srv_lock);
 331 }
 332
 333 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 334                         char *name)
 335 {
 336         struct ptlrpc_svc_data d;
 337         struct ptlrpc_thread *thread;
 338         int rc;
 339         ENTRY;
 340
 341         OBD_ALLOC(thread, sizeof(*thread));
 342         if (thread == NULL) {
 343                 LBUG();
 344                 RETURN(-ENOMEM);
 345         }
 346         init_waitqueue_head(&thread->t_ctl_waitq);
 347
 348         d.dev = dev;
 349         d.svc = svc;
 350         d.name = name;
 351         d.thread = thread;
 352
 353         spin_lock(&svc->srv_lock);
 354         list_add(&thread->t_link, &svc->srv_threads);
 355         spin_unlock(&svc->srv_lock);
 356
 357         /* XXX should we really be cloning open file handles here? */
 358         rc = kernel_thread(ptlrpc_main, (void *) &d,
 359                            CLONE_VM | CLONE_FS | CLONE_FILES);
 360         if (rc < 0) {
 361                 CERROR("cannot start thread\n");
 362                 OBD_FREE(thread, sizeof(*thread));
 363                 RETURN(rc);
 364         }
 365         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 366
 367         RETURN(0);
 368 }
 369
 370 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 371 {
 372         int rc;
 373
 374         LASSERT (list_empty (&service->srv_threads));
 375
 376         /* XXX We could reply (with failure) to all buffered requests
 377          * _after_ unlinking _all_ the request buffers, but _before_
 378          * freeing them.
 379          */
 380
 381         while (!list_empty (&service->srv_rqbds)) {
 382                 struct ptlrpc_request_buffer_desc *rqbd =
 383                         list_entry (service->srv_rqbds.next,
 384                                     struct ptlrpc_request_buffer_desc,
 385                                     rqbd_list);
 386
 387                 list_del (&rqbd->rqbd_list);
 388
 389                 LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 390                 /* refcount could be anything; it's possible for the
 391                  * buffers to continued to get filled after all the server
 392                  * threads exited.  But we know they _have_ exited.
 393                  */
 394
 395                 (void) PtlMEUnlink(rqbd->rqbd_me_h);
 396                 /* The callback handler could have unlinked this ME already
 397                  * (we're racing with her) but it's safe to ensure it _has_
 398                  * been unlinked.
 399                  */
 400
 401                 OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 402                 OBD_FREE (rqbd, sizeof (*rqbd));
 403                 service->srv_nrqbds--;
 404         }
 405
 406         LASSERT (service->srv_nrqbds == 0);
 407
 408         rc = PtlEQFree(service->srv_eq_h);
 409         if (rc)
 410                 CERROR("PtlEQFree failed: %d\n", rc);
 411
 412         OBD_FREE(service, sizeof(*service));
 413         if (rc)
 414                 LBUG();
 415         return rc;
 416 }