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 extern int ptl_handled_rpc(struct ptlrpc_service *service, void *start);
  31
  32 static int ptlrpc_check_event(struct ptlrpc_service *svc,
  33                               struct ptlrpc_thread *thread, ptl_event_t *event)
  34 {
  35         int rc = 0;
  36         ENTRY;
  37
  38         spin_lock(&svc->srv_lock);
  39         if (thread->t_flags & SVC_STOPPING)
  40                 GOTO(out, rc = 1);
  41
  42         if (ptl_is_valid_handle(&svc->srv_eq_h)) {
  43                 int err;
  44                 err = PtlEQGet(svc->srv_eq_h, event);
  45
  46                 if (err == PTL_OK) {
  47                         thread->t_flags |= SVC_EVENT;
  48                         GOTO(out, rc = 1);
  49                 }
  50
  51                 if (err != PTL_EQ_EMPTY) {
  52                         CERROR("BUG: PtlEQGet returned %d\n", rc);
  53                         LBUG();
  54                 }
  55
  56                 GOTO(out, rc = 0);
  57         }
  58
  59         EXIT;
  60  out:
  61         spin_unlock(&svc->srv_lock);
  62         return rc;
  63 }
  64
  65 struct ptlrpc_service *
  66 ptlrpc_init_svc(__u32 bufsize, int req_portal, int rep_portal, char *uuid,
  67                 svc_handler_t handler, char *name)
  68 {
  69         int err;
  70         int rc, i;
  71         struct ptlrpc_service *service;
  72         ENTRY;
  73
  74         OBD_ALLOC(service, sizeof(*service));
  75         if (!service) {
  76                 LBUG();
  77                 RETURN(NULL);
  78         }
  79
  80         service->srv_name = name;
  81         spin_lock_init(&service->srv_lock);
  82         INIT_LIST_HEAD(&service->srv_reqs);
  83         INIT_LIST_HEAD(&service->srv_threads);
  84         init_waitqueue_head(&service->srv_waitq);
  85
  86         service->srv_buf_size = bufsize;
  87         service->srv_rep_portal = rep_portal;
  88         service->srv_req_portal = req_portal;
  89         service->srv_handler = handler;
  90
  91         err = kportal_uuid_to_peer(uuid, &service->srv_self);
  92         if (err) {
  93                 CERROR("cannot get peer for uuid '%s'\n", uuid);
  94                 OBD_FREE(service, sizeof(*service));
  95                 RETURN(NULL);
  96         }
  97
  98         service->srv_ring_length = RPC_RING_LENGTH;
  99
 100         rc = PtlEQAlloc(service->srv_self.peer_ni, 1024, request_in_callback,
 101                         &(service->srv_eq_h));
 102
 103         if (rc != PTL_OK) {
 104                 CERROR("PtlEQAlloc failed: %d\n", rc);
 105                 LBUG();
 106                 OBD_FREE(service, sizeof(*service));
 107                 RETURN(NULL);
 108         }
 109
 110         for (i = 0; i < service->srv_ring_length; i++) {
 111                 OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
 112                 if (service->srv_buf[i] == NULL) {
 113                         CERROR("no memory\n");
 114                         LBUG();
 115                         GOTO(err_ring, NULL);
 116                 }
 117                 service->srv_ref_count[i] = 0;
 118                 ptlrpc_link_svc_me(service, i);
 119         }
 120
 121         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 122                service->srv_req_portal);
 123
 124         RETURN(service);
 125 err_ring:
 126         service->srv_ring_length = i;
 127         ptlrpc_unregister_service(service);
 128         return NULL;
 129 }
 130
 131 static int handle_incoming_request(struct obd_device *obddev,
 132                                    struct ptlrpc_service *svc,
 133                                    ptl_event_t *event)
 134 {
 135         struct ptlrpc_request request;
 136         struct lustre_peer peer;
 137         void *start;
 138         int rc;
 139
 140         /* FIXME: If we move to an event-driven model, we should put the request
 141          * on the stack of mds_handle instead. */
 142         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 143         start = event->mem_desc.start;
 144
 145         memset(&request, 0, sizeof(request));
 146         request.rq_svc = svc;
 147         request.rq_obd = obddev;
 148         request.rq_xid = event->match_bits;
 149         request.rq_reqmsg = event->mem_desc.start + event->offset;
 150         request.rq_reqlen = event->mem_desc.length;
 151
 152         if (request.rq_reqlen < sizeof(struct lustre_msg)) {
 153                 CERROR("incomplete request (%d): ptl %d from %Lx xid %Ld\n",
 154                        request.rq_reqlen, svc->srv_req_portal,
 155                        event->initiator.nid, request.rq_xid);
 156                 spin_unlock(&svc->srv_lock);
 157                 RETURN(-EINVAL);
 158         }
 159
 160         if (request.rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 161                 CERROR("wrong lustre_msg magic %d: ptl %d from %Lx xid %Ld\n",
 162                        request.rq_reqmsg->magic, svc->srv_req_portal,
 163                        event->initiator.nid, request.rq_xid);
 164                 spin_unlock(&svc->srv_lock);
 165                 RETURN(-EINVAL);
 166         }
 167
 168         if (request.rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 169                 CERROR("wrong lustre_msg version %d: ptl %d from %Lx xid %Ld\n",
 170                        request.rq_reqmsg->version, svc->srv_req_portal,
 171                        event->initiator.nid, request.rq_xid);
 172                 spin_unlock(&svc->srv_lock);
 173                 RETURN(-EINVAL);
 174         }
 175
 176         CDEBUG(D_NET, "got req %Ld\n", request.rq_xid);
 177
 178         peer.peer_nid = event->initiator.nid;
 179         /* FIXME: this NI should be the incoming NI.
 180          * We don't know how to find that from here. */
 181         peer.peer_ni = svc->srv_self.peer_ni;
 182
 183         request.rq_export = class_conn2export((struct lustre_handle *) request.rq_reqmsg);
 184
 185         if (request.rq_export) {
 186                 request.rq_connection = request.rq_export->exp_connection;
 187                 ptlrpc_connection_addref(request.rq_connection);
 188         } else {
 189                 request.rq_connection = ptlrpc_get_connection(&peer);
 190         }
 191
 192         spin_unlock(&svc->srv_lock);
 193         rc = svc->srv_handler(&request);
 194         ptlrpc_put_connection(request.rq_connection);
 195         ptl_handled_rpc(svc, start);
 196         return rc;
 197 }
 198
 199 void ptlrpc_rotate_reqbufs(struct ptlrpc_service *service,
 200                             ptl_event_t *ev)
 201 {
 202         int index;
 203
 204         LASSERT ((ev->mem_desc.options & PTL_MD_IOV) == 0);
 205
 206         for (index = 0; index < service->srv_ring_length; index++)
 207                 if (service->srv_buf[index] == ev->mem_desc.start)
 208                         break;
 209
 210         if (index == service->srv_ring_length)
 211                 LBUG();
 212
 213         service->srv_ref_count[index]++;
 214
 215         if (ptl_is_valid_handle(&ev->unlinked_me)) {
 216                 int idx;
 217
 218                 for (idx = 0; idx < service->srv_ring_length; idx++)
 219                         if (service->srv_me_h[idx].handle_idx ==
 220                             ev->unlinked_me.handle_idx)
 221                                 break;
 222                 if (idx == service->srv_ring_length)
 223                         LBUG();
 224
 225                 CDEBUG(D_NET, "unlinked %d\n", idx);
 226                 ptl_set_inv_handle(&(service->srv_me_h[idx]));
 227
 228                 if (service->srv_ref_count[idx] == 0)
 229                         ptlrpc_link_svc_me(service, idx);
 230         }
 231 }
 232
 233 static int ptlrpc_main(void *arg)
 234 {
 235         int rc;
 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
 241         ENTRY;
 242
 243         lock_kernel();
 244         daemonize();
 245         spin_lock_irq(&current->sigmask_lock);
 246         sigfillset(&current->blocked);
 247         recalc_sigpending(current);
 248         spin_unlock_irq(&current->sigmask_lock);
 249
 250         sprintf(current->comm, data->name);
 251         unlock_kernel();
 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
 268                 if (thread->t_flags & SVC_STOPPING) {
 269                         thread->t_flags &= ~SVC_STOPPING;
 270                         spin_unlock(&svc->srv_lock);
 271                         EXIT;
 272                         break;
 273                 }
 274
 275                 if (thread->t_flags & SVC_EVENT) {
 276                         thread->t_flags &= ~SVC_EVENT;
 277                         ptlrpc_rotate_reqbufs(svc, &event);
 278
 279                         rc = handle_incoming_request(obddev, svc, &event);
 280                         thread->t_flags &= ~SVC_EVENT;
 281                         continue;
 282                 }
 283
 284                 CERROR("unknown break in service");
 285                 spin_unlock(&svc->srv_lock);
 286                 EXIT;
 287                 break;
 288         }
 289
 290         thread->t_flags = SVC_STOPPED;
 291         wake_up(&thread->t_ctl_waitq);
 292         CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
 293         return 0;
 294 }
 295
 296 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 297                                struct ptlrpc_thread *thread)
 298 {
 299         spin_lock(&svc->srv_lock);
 300         thread->t_flags = SVC_STOPPING;
 301         spin_unlock(&svc->srv_lock);
 302
 303         wake_up(&svc->srv_waitq);
 304         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 305 }
 306
 307 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 308 {
 309         spin_lock(&svc->srv_lock);
 310         while (!list_empty(&svc->srv_threads)) {
 311                 struct ptlrpc_thread *thread;
 312                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 313                                     t_link);
 314                 spin_unlock(&svc->srv_lock);
 315                 ptlrpc_stop_thread(svc, thread);
 316                 spin_lock(&svc->srv_lock);
 317                 list_del(&thread->t_link);
 318                 OBD_FREE(thread, sizeof(*thread));
 319         }
 320         spin_unlock(&svc->srv_lock);
 321 }
 322
 323 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 324                         char *name)
 325 {
 326         struct ptlrpc_svc_data d;
 327         struct ptlrpc_thread *thread;
 328         int rc;
 329         ENTRY;
 330
 331         OBD_ALLOC(thread, sizeof(*thread));
 332         if (thread == NULL) {
 333                 LBUG();
 334                 RETURN(-ENOMEM);
 335         }
 336         init_waitqueue_head(&thread->t_ctl_waitq);
 337
 338         d.dev = dev;
 339         d.svc = svc;
 340         d.name = name;
 341         d.thread = thread;
 342
 343         spin_lock(&svc->srv_lock);
 344         list_add(&thread->t_link, &svc->srv_threads);
 345         spin_unlock(&svc->srv_lock);
 346
 347         rc = kernel_thread(ptlrpc_main, (void *) &d,
 348                            CLONE_VM | CLONE_FS | CLONE_FILES);
 349         if (rc < 0) {
 350                 CERROR("cannot start thread\n");
 351                 OBD_FREE(thread, sizeof(*thread));
 352                 RETURN(-EINVAL);
 353         }
 354         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 355
 356         RETURN(0);
 357 }
 358
 359 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 360 {
 361         int rc, i;
 362
 363         for (i = 0; i < service->srv_ring_length; i++) {
 364                 if (ptl_is_valid_handle(&(service->srv_me_h[i]))) {
 365                         rc = PtlMEUnlink(service->srv_me_h[i]);
 366                         if (rc)
 367                                 CERROR("PtlMEUnlink failed: %d\n", rc);
 368                         ptl_set_inv_handle(&(service->srv_me_h[i]));
 369                 }
 370
 371                 if (service->srv_buf[i] != NULL)
 372                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 373                 service->srv_buf[i] = NULL;
 374         }
 375
 376         rc = PtlEQFree(service->srv_eq_h);
 377         if (rc)
 378                 CERROR("PtlEQFree failed: %d\n", rc);
 379
 380         if (!list_empty(&service->srv_reqs)) {
 381                 // XXX reply with errors and clean up
 382                 CERROR("Request list not empty!\n");
 383                 rc = -EBUSY;
 384         }
 385
 386         OBD_FREE(service, sizeof(*service));
 387         if (rc)
 388                 LBUG();
 389         return rc;
 390 }