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
 148         memset(&request, 0, sizeof(request));
 149         request.rq_svc = svc;
 150         request.rq_obd = obddev;
 151         request.rq_xid = event->match_bits;
 152         request.rq_reqmsg = event->mem_desc.start + event->offset;
 153         request.rq_reqlen = event->mem_desc.length;
 154
 155         if (request.rq_reqlen < sizeof(struct lustre_msg)) {
 156                 CERROR("incomplete request: ptl %d from %Lx xid %Ld\n",
 157                        svc->srv_req_portal, event->initiator.nid,
 158                        request.rq_xid);
 159                 spin_unlock(&svc->srv_lock);
 160                 return -EINVAL;
 161         }
 162
 163         if (request.rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 164                 CERROR("wrong lustre_msg magic: ptl %d from %Lx xid %Ld\n",
 165                        svc->srv_req_portal, event->initiator.nid,
 166                        request.rq_xid);
 167                 spin_unlock(&svc->srv_lock);
 168                 return -EINVAL;
 169         }
 170
 171         if (request.rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 172                 CERROR("wrong lustre_msg version: ptl %d from %Lx xid %Ld\n",
 173                        svc->srv_req_portal, event->initiator.nid,
 174                        request.rq_xid);
 175                 spin_unlock(&svc->srv_lock);
 176                 return -EINVAL;
 177         }
 178
 179         CDEBUG(D_NET, "got req %Ld\n", request.rq_xid);
 180
 181         peer.peer_nid = event->initiator.nid;
 182         /* FIXME: this NI should be the incoming NI.
 183          * We don't know how to find that from here. */
 184         peer.peer_ni = svc->srv_self.peer_ni;
 185
 186         request.rq_export = class_conn2export((struct lustre_handle *) request.rq_reqmsg);
 187
 188         if (request.rq_export) {
 189                 request.rq_connection = request.rq_export->export_connection;
 190                 ptlrpc_connection_addref(request.rq_connection);
 191         } else {
 192                 request.rq_connection = ptlrpc_get_connection(&peer);
 193         }
 194
 195         spin_unlock(&svc->srv_lock);
 196         rc = svc->srv_handler(&request);
 197         ptlrpc_put_connection(request.rq_connection);
 198         ptl_handled_rpc(svc, start);
 199         return rc;
 200 }
 201
 202 void ptlrpc_rotate_reqbufs(struct ptlrpc_service *service,
 203                             ptl_event_t *ev)
 204 {
 205         int index;
 206
 207         for (index = 0; index < service->srv_ring_length; index++)
 208                 if (service->srv_buf[index] == ev->mem_desc.start)
 209                         break;
 210
 211         if (index == service->srv_ring_length)
 212                 LBUG();
 213
 214         service->srv_ref_count[index]++;
 215
 216         if (ptl_is_valid_handle(&ev->unlinked_me)) {
 217                 int idx;
 218
 219                 for (idx = 0; idx < service->srv_ring_length; idx++)
 220                         if (service->srv_me_h[idx].handle_idx ==
 221                             ev->unlinked_me.handle_idx)
 222                                 break;
 223                 if (idx == service->srv_ring_length)
 224                         LBUG();
 225
 226                 CDEBUG(D_NET, "unlinked %d\n", idx);
 227                 ptl_set_inv_handle(&(service->srv_me_h[idx]));
 228
 229                 if (service->srv_ref_count[idx] == 0)
 230                         ptlrpc_link_svc_me(service, idx);
 231         }
 232 }
 233
 234 static int ptlrpc_main(void *arg)
 235 {
 236         int rc;
 237         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 238         struct obd_device *obddev = data->dev;
 239         struct ptlrpc_service *svc = data->svc;
 240         struct ptlrpc_thread *thread = data->thread;
 241
 242         ENTRY;
 243
 244         lock_kernel();
 245         daemonize();
 246         spin_lock_irq(&current->sigmask_lock);
 247         sigfillset(&current->blocked);
 248         recalc_sigpending(current);
 249         spin_unlock_irq(&current->sigmask_lock);
 250
 251         sprintf(current->comm, data->name);
 252
 253         /* Record that the thread is running */
 254         thread->t_flags = SVC_RUNNING;
 255         wake_up(&thread->t_ctl_waitq);
 256
 257         /* XXX maintain a list of all managed devices: insert here */
 258
 259         /* And now, loop forever on requests */
 260         while (1) {
 261                 ptl_event_t event;
 262
 263                 wait_event(svc->srv_waitq,
 264                            ptlrpc_check_event(svc, thread, &event));
 265
 266                 spin_lock(&svc->srv_lock);
 267                 if (thread->t_flags & SVC_SIGNAL) {
 268                         thread->t_flags &= ~SVC_SIGNAL;
 269                         spin_unlock(&svc->srv_lock);
 270                         EXIT;
 271                         break;
 272                 }
 273
 274                 if (thread->t_flags & SVC_STOPPING) {
 275                         thread->t_flags &= ~SVC_STOPPING;
 276                         spin_unlock(&svc->srv_lock);
 277                         EXIT;
 278                         break;
 279                 }
 280
 281                 if (thread->t_flags & SVC_EVENT) {
 282                         thread->t_flags &= ~SVC_EVENT;
 283                         ptlrpc_rotate_reqbufs(svc, &event);
 284
 285                         rc = handle_incoming_request(obddev, svc, &event);
 286                         thread->t_flags &= ~SVC_EVENT;
 287                         continue;
 288                 }
 289
 290                 CERROR("unknown break in service");
 291                 spin_unlock(&svc->srv_lock);
 292                 EXIT;
 293                 break;
 294         }
 295
 296         thread->t_flags = SVC_STOPPED;
 297         wake_up(&thread->t_ctl_waitq);
 298         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 299         return 0;
 300 }
 301
 302 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 303                                struct ptlrpc_thread *thread)
 304 {
 305         spin_lock(&svc->srv_lock);
 306         thread->t_flags = SVC_STOPPING;
 307         spin_unlock(&svc->srv_lock);
 308
 309         wake_up(&svc->srv_waitq);
 310         wait_event_interruptible(thread->t_ctl_waitq,
 311                                  (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         rc = kernel_thread(ptlrpc_main, (void *) &d,
 355                            CLONE_VM | CLONE_FS | CLONE_FILES);
 356         if (rc < 0) {
 357                 CERROR("cannot start thread\n");
 358                 RETURN(-EINVAL);
 359         }
 360         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 361
 362         RETURN(0);
 363 }
 364
 365 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 366 {
 367         int rc, i;
 368
 369         for (i = 0; i < service->srv_ring_length; i++) {
 370                 if (ptl_is_valid_handle(&(service->srv_me_h[i]))) {
 371                         rc = PtlMEUnlink(service->srv_me_h[i]);
 372                         if (rc)
 373                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 374                         ptl_set_inv_handle(&(service->srv_me_h[i]));
 375                 }
 376
 377                 if (service->srv_buf[i] != NULL)
 378                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 379                 service->srv_buf[i] = NULL;
 380         }
 381
 382         rc = PtlEQFree(service->srv_eq_h);
 383         if (rc)
 384                 CERROR("PtlEQFree failed: %d\n", rc);
 385
 386         if (!list_empty(&service->srv_reqs)) {
 387                 // XXX reply with errors and clean up
 388                 CERROR("Request list not empty!\n");
 389         }
 390
 391         OBD_FREE(service, sizeof(*service));
 392         return 0;
 393 }