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/lustre_net.h>
  26
  27 extern int request_in_callback(ptl_event_t *ev);
  28 extern int ptl_handled_rpc(struct ptlrpc_service *service, void *start);
  29
  30 static int ptlrpc_check_event(struct ptlrpc_service *svc,
  31                               struct ptlrpc_thread *thread, ptl_event_t *event)
  32 {
  33         int rc = 0;
  34         ENTRY;
  35
  36         spin_lock(&svc->srv_lock);
  37         if (sigismember(&(current->pending.signal), SIGKILL) ||
  38             sigismember(&(current->pending.signal), SIGTERM) ||
  39             sigismember(&(current->pending.signal), SIGINT)) {
  40                 thread->t_flags |= SVC_KILLED;
  41                 GOTO(out, rc = 1);
  42         }
  43
  44         if (thread->t_flags & SVC_STOPPING)
  45                 GOTO(out, rc = 1);
  46
  47         if (ptl_is_valid_handle(&svc->srv_eq_h)) {
  48                 int err;
  49                 err = PtlEQGet(svc->srv_eq_h, event);
  50
  51                 if (err == PTL_OK) {
  52                         thread->t_flags |= SVC_EVENT;
  53                         GOTO(out, rc = 1);
  54                 }
  55
  56                 if (err != PTL_EQ_EMPTY) {
  57                         CDEBUG(D_NET, "BUG: PtlEQGet returned %d\n", rc);
  58                         LBUG();
  59                 }
  60
  61                 GOTO(out, rc = 0);
  62         }
  63
  64         EXIT;
  65  out:
  66         spin_unlock(&svc->srv_lock);
  67         return rc;
  68 }
  69
  70 struct ptlrpc_service *
  71 ptlrpc_init_svc(__u32 bufsize, int req_portal, int rep_portal, char *uuid,
  72                 svc_handler_t handler)
  73 {
  74         int err;
  75         int rc, i;
  76         struct ptlrpc_service *service;
  77         ENTRY;
  78
  79         OBD_ALLOC(service, sizeof(*service));
  80         if (!service) {
  81                 LBUG();
  82                 RETURN(NULL);
  83         }
  84
  85         spin_lock_init(&service->srv_lock);
  86         INIT_LIST_HEAD(&service->srv_reqs);
  87         INIT_LIST_HEAD(&service->srv_threads);
  88         init_waitqueue_head(&service->srv_waitq);
  89
  90         service->srv_buf_size = bufsize;
  91         service->srv_rep_portal = rep_portal;
  92         service->srv_req_portal = req_portal;
  93         service->srv_handler = handler;
  94
  95         err = kportal_uuid_to_peer(uuid, &service->srv_self);
  96         if (err) {
  97                 CERROR("cannot get peer for uuid '%s'", uuid);
  98                 OBD_FREE(service, sizeof(*service));
  99                 RETURN(NULL);
 100         }
 101
 102         service->srv_ring_length = RPC_RING_LENGTH;
 103
 104         rc = PtlEQAlloc(service->srv_self.peer_ni, 128, request_in_callback,
 105                         &(service->srv_eq_h));
 106
 107         if (rc != PTL_OK) {
 108                 CERROR("PtlEQAlloc failed: %d\n", rc);
 109                 LBUG();
 110                 OBD_FREE(service, sizeof(*service));
 111                 RETURN(NULL);
 112         }
 113
 114         for (i = 0; i < service->srv_ring_length; i++) {
 115                 OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
 116                 if (service->srv_buf[i] == NULL) {
 117                         CERROR("no memory\n");
 118                         LBUG();
 119                         GOTO(err_ring, NULL);
 120                 }
 121                 service->srv_ref_count[i] = 0;
 122                 ptlrpc_link_svc_me(service, i);
 123         }
 124
 125         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 126                service->srv_req_portal);
 127
 128         RETURN(service);
 129 err_ring:
 130         service->srv_ring_length = i;
 131         ptlrpc_unregister_service(service);
 132         return NULL;
 133 }
 134
 135 static int handle_incoming_request(struct obd_device *obddev,
 136                                    struct ptlrpc_service *svc,
 137                                    ptl_event_t *event)
 138 {
 139         struct ptlrpc_request request;
 140         struct lustre_peer peer;
 141         void *start;
 142         int rc;
 143
 144         /* FIXME: If we move to an event-driven model, we should put the request
 145          * on the stack of mds_handle instead. */
 146         start = event->mem_desc.start;
 147         memset(&request, 0, sizeof(request));
 148         request.rq_obd = obddev;
 149         request.rq_xid = event->match_bits;
 150         request.rq_reqmsg = event->mem_desc.start + event->offset;
 151         request.rq_reqlen = event->mem_desc.length;
 152
 153         CDEBUG(D_NET, "got req %Ld\n", request.rq_xid);
 154
 155         peer.peer_nid = event->initiator.nid;
 156         /* FIXME: this NI should be the incoming NI.
 157          * We don't know how to find that from here. */
 158         peer.peer_ni = svc->srv_self.peer_ni;
 159
 160         if (request.rq_reqmsg->conn) {
 161                 request.rq_connection =
 162                         (void *)(unsigned long)request.rq_reqmsg->conn;
 163                 if (request.rq_reqmsg->token !=
 164                     request.rq_connection->c_token) {
 165                         struct ptlrpc_connection *tmp;
 166                         tmp = ptlrpc_get_connection(&peer);
 167                         CERROR("rq_reqmsg->conn: %p\n", request.rq_connection);
 168                         CERROR("real connection: %p\n", tmp);
 169                         CERROR("rq_reqmsg->token: %Lu\n",
 170                                (unsigned long long)request.rq_reqmsg->token);
 171                         CERROR("real token      : %Lu\n",
 172                                (unsigned long long)tmp->c_token);
 173                         LBUG();
 174                 }
 175                 ptlrpc_connection_addref(request.rq_connection);
 176         } else {
 177                 /*
 178                   PHIL? should we perhaps only do this when
 179                    we get an incoming connmgr_connect request?
 180                 */
 181                 request.rq_connection = ptlrpc_get_connection(&peer);
 182                 if (!request.rq_connection)
 183                         LBUG();
 184         }
 185
 186         spin_unlock(&svc->srv_lock);
 187         rc = svc->srv_handler(obddev, svc, &request);
 188         ptlrpc_put_connection(request.rq_connection);
 189         ptl_handled_rpc(svc, start);
 190         return rc;
 191 }
 192
 193 void ptlrpc_rotate_reqbufs(struct ptlrpc_service *service,
 194                             ptl_event_t *ev)
 195 {
 196         int index;
 197
 198         for (index = 0; index < service->srv_ring_length; index++)
 199                 if (service->srv_buf[index] == ev->mem_desc.start)
 200                         break;
 201
 202         if (index == service->srv_ring_length)
 203                 LBUG();
 204
 205         service->srv_ref_count[index]++;
 206
 207         if (ptl_is_valid_handle(&ev->unlinked_me)) {
 208                 int idx;
 209
 210                 for (idx = 0; idx < service->srv_ring_length; idx++)
 211                         if (service->srv_me_h[idx].handle_idx ==
 212                             ev->unlinked_me.handle_idx)
 213                                 break;
 214                 if (idx == service->srv_ring_length)
 215                         LBUG();
 216
 217                 CDEBUG(D_NET, "unlinked %d\n", idx);
 218                 ptl_set_inv_handle(&(service->srv_me_h[idx]));
 219
 220                 if (service->srv_ref_count[idx] == 0)
 221                         ptlrpc_link_svc_me(service, idx);
 222         }
 223 }
 224
 225 static int ptlrpc_main(void *arg)
 226 {
 227         int rc;
 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
 233         ENTRY;
 234
 235         lock_kernel();
 236         daemonize();
 237         spin_lock_irq(&current->sigmask_lock);
 238         sigfillset(&current->blocked);
 239         recalc_sigpending(current);
 240         spin_unlock_irq(&current->sigmask_lock);
 241
 242         sprintf(current->comm, data->name);
 243
 244         /* Record that the thread is running */
 245         thread->t_flags = SVC_RUNNING;
 246         wake_up(&thread->t_ctl_waitq);
 247
 248         /* XXX maintain a list of all managed devices: insert here */
 249
 250         /* And now, loop forever on requests */
 251         while (1) {
 252                 ptl_event_t event;
 253
 254                 wait_event(svc->srv_waitq,
 255                            ptlrpc_check_event(svc, thread, &event));
 256
 257                 spin_lock(&svc->srv_lock);
 258                 if (thread->t_flags & SVC_SIGNAL) {
 259                         thread->t_flags &= ~SVC_SIGNAL;
 260                         spin_unlock(&svc->srv_lock);
 261                         EXIT;
 262                         break;
 263                 }
 264
 265                 if (thread->t_flags & SVC_STOPPING) {
 266                         thread->t_flags &= ~SVC_STOPPING;
 267                         spin_unlock(&svc->srv_lock);
 268                         EXIT;
 269                         break;
 270                 }
 271
 272                 if (thread->t_flags & SVC_EVENT) {
 273                         thread->t_flags &= ~SVC_EVENT;
 274                         ptlrpc_rotate_reqbufs(svc, &event);
 275
 276                         rc = handle_incoming_request(obddev, svc, &event);
 277                         thread->t_flags &= ~SVC_EVENT;
 278                         continue;
 279                 }
 280
 281                 CERROR("unknown break in service");
 282                 spin_unlock(&svc->srv_lock);
 283                 EXIT;
 284                 break;
 285         }
 286
 287         thread->t_flags = SVC_STOPPED;
 288         wake_up(&thread->t_ctl_waitq);
 289         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 290         return 0;
 291 }
 292
 293 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 294                                struct ptlrpc_thread *thread)
 295 {
 296         thread->t_flags = SVC_STOPPING;
 297
 298         wake_up(&svc->srv_waitq);
 299         wait_event_interruptible(thread->t_ctl_waitq,
 300                                  (thread->t_flags & SVC_STOPPED));
 301 }
 302
 303 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 304 {
 305         spin_lock(&svc->srv_lock);
 306         while (!list_empty(&svc->srv_threads)) {
 307                 struct ptlrpc_thread *thread;
 308                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 309                                     t_link);
 310                 spin_unlock(&svc->srv_lock);
 311                 ptlrpc_stop_thread(svc, thread);
 312                 spin_lock(&svc->srv_lock);
 313                 list_del(&thread->t_link);
 314                 OBD_FREE(thread, sizeof(*thread));
 315         }
 316         spin_unlock(&svc->srv_lock);
 317 }
 318
 319 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 320                         char *name)
 321 {
 322         struct ptlrpc_svc_data d;
 323         struct ptlrpc_thread *thread;
 324         int rc;
 325         ENTRY;
 326
 327         OBD_ALLOC(thread, sizeof(*thread));
 328         if (thread == NULL) {
 329                 LBUG();
 330                 RETURN(-ENOMEM);
 331         }
 332         init_waitqueue_head(&thread->t_ctl_waitq);
 333
 334         d.dev = dev;
 335         d.svc = svc;
 336         d.name = name;
 337         d.thread = thread;
 338
 339         spin_lock(&svc->srv_lock);
 340         list_add(&thread->t_link, &svc->srv_threads);
 341         spin_unlock(&svc->srv_lock);
 342
 343         rc = kernel_thread(ptlrpc_main, (void *) &d,
 344                            CLONE_VM | CLONE_FS | CLONE_FILES);
 345         if (rc < 0) {
 346                 CERROR("cannot start thread\n");
 347                 RETURN(-EINVAL);
 348         }
 349         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 350
 351         RETURN(0);
 352 }
 353
 354 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 355 {
 356         int rc, i;
 357
 358         for (i = 0; i < service->srv_ring_length; i++) {
 359                 if (ptl_is_valid_handle(&(service->srv_me_h[i]))) {
 360                         rc = PtlMEUnlink(service->srv_me_h[i]);
 361                         if (rc)
 362                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 363                         ptl_set_inv_handle(&(service->srv_me_h[i]));
 364                 }
 365
 366                 if (service->srv_buf[i] != NULL)
 367                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 368                 service->srv_buf[i] = NULL;
 369         }
 370
 371         rc = PtlEQFree(service->srv_eq_h);
 372         if (rc)
 373                 CERROR("PtlEQFree failed: %d\n", rc);
 374
 375         if (!list_empty(&service->srv_reqs)) {
 376                 // XXX reply with errors and clean up
 377                 CERROR("Request list not empty!\n");
 378         }
 379
 380         OBD_FREE(service, sizeof(*service));
 381         return 0;
 382 }