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                 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                         CERROR ("no memory\n");
 120                         GOTO (failed, NULL);
 121                 }
 122
 123                 rqbd->rqbd_service = service;
 124                 ptl_set_inv_handle (&rqbd->rqbd_me_h);
 125                 atomic_set (&rqbd->rqbd_refcount, 0);
 126                 OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 127                 if (rqbd->rqbd_buffer == NULL) {
 128                         OBD_FREE (rqbd, sizeof (*rqbd));
 129                         CERROR("no memory\n");
 130                         GOTO(failed, NULL);
 131                 }
 132                 list_add (&rqbd->rqbd_list, &service->srv_rqbds);
 133                 service->srv_nrqbds++;
 134
 135                 ptlrpc_link_svc_me(rqbd);
 136         }
 137
 138         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 139                service->srv_req_portal);
 140
 141         RETURN(service);
 142 failed:
 143         ptlrpc_unregister_service(service);
 144         return NULL;
 145 }
 146
 147 static int handle_incoming_request(struct obd_device *obddev,
 148                                    struct ptlrpc_service *svc,
 149                                    ptl_event_t *event,
 150                                    struct ptlrpc_request *request)
 151 {
 152         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 153         int rc;
 154
 155         /* FIXME: If we move to an event-driven model, we should put the request
 156          * on the stack of mds_handle instead. */
 157
 158         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 159         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 160         LASSERT (rqbd->rqbd_service == svc);
 161         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 162         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 163
 164         memset(request, 0, sizeof(*request));
 165         request->rq_svc = svc;
 166         request->rq_obd = obddev;
 167         request->rq_xid = event->match_bits;
 168         request->rq_reqmsg = event->mem_desc.start + event->offset;
 169         request->rq_reqlen = event->mlength;
 170
 171         rc = -EINVAL;
 172
 173         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 174                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "
 175                        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 "
 189                        LPD64"\n",
 190                        request->rq_reqmsg->magic, svc->srv_req_portal,
 191                        event->initiator.nid, request->rq_xid);
 192                 goto out;
 193         }
 194
 195         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 196                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "
 197                        LPD64"\n",
 198                        request->rq_reqmsg->version, svc->srv_req_portal,
 199                        event->initiator.nid, request->rq_xid);
 200                 goto out;
 201         }
 202
 203         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 204
 205         request->rq_peer.peer_nid = event->initiator.nid;
 206         /* FIXME: this NI should be the incoming NI.
 207          * We don't know how to find that from here. */
 208         request->rq_peer.peer_ni = svc->srv_self.peer_ni;
 209
 210         request->rq_export = class_conn2export((struct lustre_handle *)
 211                                                request->rq_reqmsg);
 212
 213         if (request->rq_export) {
 214                 request->rq_connection = request->rq_export->exp_connection;
 215                 ptlrpc_connection_addref(request->rq_connection);
 216         } else {
 217                 /* create a (hopefully temporary) connection that will be used
 218                  * to send the reply if this call doesn't create an export.
 219                  * XXX revisit this when we revamp ptlrpc */
 220                 request->rq_connection =
 221                         ptlrpc_get_connection(&request->rq_peer, NULL);
 222         }
 223
 224         rc = svc->srv_handler(request);
 225         ptlrpc_put_connection(request->rq_connection);
 226
 227  out:
 228         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 229                 ptlrpc_link_svc_me (rqbd);
 230
 231         return rc;
 232 }
 233
 234 static int ptlrpc_main(void *arg)
 235 {
 236         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 237         struct obd_device *obddev = data->dev;
 238         struct ptlrpc_service *svc = data->svc;
 239         struct ptlrpc_thread *thread = data->thread;
 240         struct ptlrpc_request *request;
 241         ptl_event_t *event;
 242         int rc = 0;
 243
 244         ENTRY;
 245
 246         lock_kernel();
 247         daemonize();
 248         spin_lock_irq(&current->sigmask_lock);
 249         sigfillset(&current->blocked);
 250         recalc_sigpending(current);
 251         spin_unlock_irq(&current->sigmask_lock);
 252
 253         strcpy(current->comm, data->name);
 254         unlock_kernel();
 255
 256         OBD_ALLOC(event, sizeof(*event));
 257         if (!event)
 258                 GOTO(out, rc = -ENOMEM);
 259         OBD_ALLOC(request, sizeof(*request));
 260         if (!request)
 261                 GOTO(out_event, rc = -ENOMEM);
 262
 263         /* Record that the thread is running */
 264         thread->t_flags = SVC_RUNNING;
 265         wake_up(&thread->t_ctl_waitq);
 266
 267         /* XXX maintain a list of all managed devices: insert here */
 268
 269         /* And now, loop forever on requests */
 270         while (1) {
 271                 wait_event(svc->srv_waitq,
 272                            ptlrpc_check_event(svc, thread, event));
 273
 274                 if (thread->t_flags & SVC_STOPPING) {
 275                         spin_lock(&svc->srv_lock);
 276                         thread->t_flags &= ~SVC_STOPPING;
 277                         spin_unlock(&svc->srv_lock);
 278
 279                         EXIT;
 280                         break;
 281                 }
 282
 283                 if (thread->t_flags & SVC_EVENT) {
 284                         spin_lock(&svc->srv_lock);
 285                         thread->t_flags &= ~SVC_EVENT;
 286                         spin_unlock(&svc->srv_lock);
 287
 288                         rc = handle_incoming_request(obddev, svc, event,
 289                                                      request);
 290                         continue;
 291                 }
 292
 293                 CERROR("unknown break in service");
 294                 LBUG();
 295                 EXIT;
 296                 break;
 297         }
 298
 299         OBD_FREE(request, sizeof(*request));
 300 out_event:
 301         OBD_FREE(event, sizeof(*event));
 302 out:
 303         thread->t_flags = SVC_STOPPED;
 304         wake_up(&thread->t_ctl_waitq);
 305
 306         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 307                current->pid, rc);
 308         return rc;
 309 }
 310
 311 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 312                                struct ptlrpc_thread *thread)
 313 {
 314         spin_lock(&svc->srv_lock);
 315         thread->t_flags = SVC_STOPPING;
 316         spin_unlock(&svc->srv_lock);
 317
 318         wake_up(&svc->srv_waitq);
 319         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 320 }
 321
 322 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 323 {
 324         spin_lock(&svc->srv_lock);
 325         while (!list_empty(&svc->srv_threads)) {
 326                 struct ptlrpc_thread *thread;
 327                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 328                                     t_link);
 329                 spin_unlock(&svc->srv_lock);
 330                 ptlrpc_stop_thread(svc, thread);
 331                 spin_lock(&svc->srv_lock);
 332                 list_del(&thread->t_link);
 333                 OBD_FREE(thread, sizeof(*thread));
 334         }
 335         spin_unlock(&svc->srv_lock);
 336 }
 337
 338 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 339                         char *name)
 340 {
 341         struct ptlrpc_svc_data d;
 342         struct ptlrpc_thread *thread;
 343         int rc;
 344         ENTRY;
 345
 346         OBD_ALLOC(thread, sizeof(*thread));
 347         if (thread == NULL) {
 348                 LBUG();
 349                 RETURN(-ENOMEM);
 350         }
 351         init_waitqueue_head(&thread->t_ctl_waitq);
 352
 353         d.dev = dev;
 354         d.svc = svc;
 355         d.name = name;
 356         d.thread = thread;
 357
 358         spin_lock(&svc->srv_lock);
 359         list_add(&thread->t_link, &svc->srv_threads);
 360         spin_unlock(&svc->srv_lock);
 361
 362         /* XXX should we really be cloning open file handles here? */
 363         rc = kernel_thread(ptlrpc_main, (void *) &d,
 364                            CLONE_VM | CLONE_FS | CLONE_FILES);
 365         if (rc < 0) {
 366                 CERROR("cannot start thread\n");
 367                 OBD_FREE(thread, sizeof(*thread));
 368                 RETURN(rc);
 369         }
 370         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 371
 372         RETURN(0);
 373 }
 374
 375 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 376 {
 377         int rc;
 378
 379         LASSERT (list_empty (&service->srv_threads));
 380
 381         /* XXX We could reply (with failure) to all buffered requests
 382          * _after_ unlinking _all_ the request buffers, but _before_
 383          * freeing them.
 384          */
 385
 386         while (!list_empty (&service->srv_rqbds)) {
 387                 struct ptlrpc_request_buffer_desc *rqbd =
 388                         list_entry (service->srv_rqbds.next,
 389                                     struct ptlrpc_request_buffer_desc,
 390                                     rqbd_list);
 391
 392                 list_del (&rqbd->rqbd_list);
 393
 394                 LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 395                 /* refcount could be anything; it's possible for the
 396                  * buffers to continued to get filled after all the server
 397                  * threads exited.  But we know they _have_ exited.
 398                  */
 399
 400                 (void) PtlMEUnlink(rqbd->rqbd_me_h);
 401                 /* The callback handler could have unlinked this ME already
 402                  * (we're racing with her) but it's safe to ensure it _has_
 403                  * been unlinked.
 404                  */
 405
 406                 OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 407                 OBD_FREE (rqbd, sizeof (*rqbd));
 408                 service->srv_nrqbds--;
 409         }
 410
 411         LASSERT (service->srv_nrqbds == 0);
 412
 413         rc = PtlEQFree(service->srv_eq_h);
 414         if (rc)
 415                 CERROR("PtlEQFree failed: %d\n", rc);
 416
 417         OBD_FREE(service, sizeof(*service));
 418         if (rc)
 419                 LBUG();
 420         return rc;
 421 }