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_svc = svc;
 149         request.rq_obd = obddev;
 150         request.rq_xid = event->match_bits;
 151         request.rq_reqmsg = event->mem_desc.start + event->offset;
 152         request.rq_reqlen = event->mem_desc.length;
 153
 154         CDEBUG(D_NET, "got req %Ld\n", request.rq_xid);
 155
 156         peer.peer_nid = event->initiator.nid;
 157         /* FIXME: this NI should be the incoming NI.
 158          * We don't know how to find that from here. */
 159         peer.peer_ni = svc->srv_self.peer_ni;
 160
 161         if (request.rq_reqmsg->conn) {
 162                 request.rq_connection =
 163                         (void *)(unsigned long)request.rq_reqmsg->conn;
 164                 if (request.rq_reqmsg->token !=
 165                     request.rq_connection->c_token) {
 166                         struct ptlrpc_connection *tmp;
 167                         tmp = ptlrpc_get_connection(&peer);
 168                         CERROR("rq_reqmsg->conn: %p\n", request.rq_connection);
 169                         CERROR("real connection: %p\n", tmp);
 170                         CERROR("rq_reqmsg->token: %Lu\n",
 171                                (unsigned long long)request.rq_reqmsg->token);
 172                         CERROR("real token      : %Lu\n",
 173                                (unsigned long long)tmp->c_token);
 174                         LBUG();
 175                 }
 176                 ptlrpc_connection_addref(request.rq_connection);
 177         } else {
 178                 /*
 179                   PHIL? should we perhaps only do this when
 180                    we get an incoming connmgr_connect request?
 181                 */
 182                 request.rq_connection = ptlrpc_get_connection(&peer);
 183                 if (!request.rq_connection)
 184                         LBUG();
 185         }
 186
 187         spin_unlock(&svc->srv_lock);
 188         rc = svc->srv_handler(&request);
 189         ptlrpc_put_connection(request.rq_connection);
 190         ptl_handled_rpc(svc, start);
 191         return rc;
 192 }
 193
 194 void ptlrpc_rotate_reqbufs(struct ptlrpc_service *service,
 195                             ptl_event_t *ev)
 196 {
 197         int index;
 198
 199         for (index = 0; index < service->srv_ring_length; index++)
 200                 if (service->srv_buf[index] == ev->mem_desc.start)
 201                         break;
 202
 203         if (index == service->srv_ring_length)
 204                 LBUG();
 205
 206         service->srv_ref_count[index]++;
 207
 208         if (ptl_is_valid_handle(&ev->unlinked_me)) {
 209                 int idx;
 210
 211                 for (idx = 0; idx < service->srv_ring_length; idx++)
 212                         if (service->srv_me_h[idx].handle_idx ==
 213                             ev->unlinked_me.handle_idx)
 214                                 break;
 215                 if (idx == service->srv_ring_length)
 216                         LBUG();
 217
 218                 CDEBUG(D_NET, "unlinked %d\n", idx);
 219                 ptl_set_inv_handle(&(service->srv_me_h[idx]));
 220
 221                 if (service->srv_ref_count[idx] == 0)
 222                         ptlrpc_link_svc_me(service, idx);
 223         }
 224 }
 225
 226 static int ptlrpc_main(void *arg)
 227 {
 228         int rc;
 229         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 230         struct obd_device *obddev = data->dev;
 231         struct ptlrpc_service *svc = data->svc;
 232         struct ptlrpc_thread *thread = data->thread;
 233
 234         ENTRY;
 235
 236         lock_kernel();
 237         daemonize();
 238         spin_lock_irq(&current->sigmask_lock);
 239         sigfillset(&current->blocked);
 240         recalc_sigpending(current);
 241         spin_unlock_irq(&current->sigmask_lock);
 242
 243         sprintf(current->comm, data->name);
 244
 245         /* Record that the thread is running */
 246         thread->t_flags = SVC_RUNNING;
 247         wake_up(&thread->t_ctl_waitq);
 248
 249         /* XXX maintain a list of all managed devices: insert here */
 250
 251         /* And now, loop forever on requests */
 252         while (1) {
 253                 ptl_event_t event;
 254
 255                 wait_event(svc->srv_waitq,
 256                            ptlrpc_check_event(svc, thread, &event));
 257
 258                 spin_lock(&svc->srv_lock);
 259                 if (thread->t_flags & SVC_SIGNAL) {
 260                         thread->t_flags &= ~SVC_SIGNAL;
 261                         spin_unlock(&svc->srv_lock);
 262                         EXIT;
 263                         break;
 264                 }
 265
 266                 if (thread->t_flags & SVC_STOPPING) {
 267                         thread->t_flags &= ~SVC_STOPPING;
 268                         spin_unlock(&svc->srv_lock);
 269                         EXIT;
 270                         break;
 271                 }
 272
 273                 if (thread->t_flags & SVC_EVENT) {
 274                         thread->t_flags &= ~SVC_EVENT;
 275                         ptlrpc_rotate_reqbufs(svc, &event);
 276
 277                         rc = handle_incoming_request(obddev, svc, &event);
 278                         thread->t_flags &= ~SVC_EVENT;
 279                         continue;
 280                 }
 281
 282                 CERROR("unknown break in service");
 283                 spin_unlock(&svc->srv_lock);
 284                 EXIT;
 285                 break;
 286         }
 287
 288         thread->t_flags = SVC_STOPPED;
 289         wake_up(&thread->t_ctl_waitq);
 290         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 291         return 0;
 292 }
 293
 294 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 295                                struct ptlrpc_thread *thread)
 296 {
 297         thread->t_flags = SVC_STOPPING;
 298
 299         wake_up(&svc->srv_waitq);
 300         wait_event_interruptible(thread->t_ctl_waitq,
 301                                  (thread->t_flags & SVC_STOPPED));
 302 }
 303
 304 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 305 {
 306         spin_lock(&svc->srv_lock);
 307         while (!list_empty(&svc->srv_threads)) {
 308                 struct ptlrpc_thread *thread;
 309                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 310                                     t_link);
 311                 spin_unlock(&svc->srv_lock);
 312                 ptlrpc_stop_thread(svc, thread);
 313                 spin_lock(&svc->srv_lock);
 314                 list_del(&thread->t_link);
 315                 OBD_FREE(thread, sizeof(*thread));
 316         }
 317         spin_unlock(&svc->srv_lock);
 318 }
 319
 320 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 321                         char *name)
 322 {
 323         struct ptlrpc_svc_data d;
 324         struct ptlrpc_thread *thread;
 325         int rc;
 326         ENTRY;
 327
 328         OBD_ALLOC(thread, sizeof(*thread));
 329         if (thread == NULL) {
 330                 LBUG();
 331                 RETURN(-ENOMEM);
 332         }
 333         init_waitqueue_head(&thread->t_ctl_waitq);
 334
 335         d.dev = dev;
 336         d.svc = svc;
 337         d.name = name;
 338         d.thread = thread;
 339
 340         spin_lock(&svc->srv_lock);
 341         list_add(&thread->t_link, &svc->srv_threads);
 342         spin_unlock(&svc->srv_lock);
 343
 344         rc = kernel_thread(ptlrpc_main, (void *) &d,
 345                            CLONE_VM | CLONE_FS | CLONE_FILES);
 346         if (rc < 0) {
 347                 CERROR("cannot start thread\n");
 348                 RETURN(-EINVAL);
 349         }
 350         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 351
 352         RETURN(0);
 353 }
 354
 355 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 356 {
 357         int rc, i;
 358
 359         for (i = 0; i < service->srv_ring_length; i++) {
 360                 if (ptl_is_valid_handle(&(service->srv_me_h[i]))) {
 361                         rc = PtlMEUnlink(service->srv_me_h[i]);
 362                         if (rc)
 363                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 364                         ptl_set_inv_handle(&(service->srv_me_h[i]));
 365                 }
 366
 367                 if (service->srv_buf[i] != NULL)
 368                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 369                 service->srv_buf[i] = NULL;
 370         }
 371
 372         rc = PtlEQFree(service->srv_eq_h);
 373         if (rc)
 374                 CERROR("PtlEQFree failed: %d\n", rc);
 375
 376         if (!list_empty(&service->srv_reqs)) {
 377                 // XXX reply with errors and clean up
 378                 CERROR("Request list not empty!\n");
 379         }
 380
 381         OBD_FREE(service, sizeof(*service));
 382         return 0;
 383 }