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;
  35         ENTRY;
  36
  37         spin_lock(&svc->srv_lock);
  38
  39         if (thread->t_flags & SVC_STOPPING)
  40                 GOTO(out, rc = 1);
  41
  42         LASSERT ((thread->t_flags & SVC_EVENT) == 0);
  43         LASSERT (ptl_is_valid_handle (&svc->srv_eq_h));
  44
  45         rc = PtlEQGet(svc->srv_eq_h, event);
  46         switch (rc)
  47         {
  48         case PTL_OK:
  49                 thread->t_flags |= SVC_EVENT;
  50                 GOTO(out, rc = 1);
  51
  52         case PTL_EQ_EMPTY:
  53                 GOTO(out, rc = 0);
  54
  55         default:
  56                 CERROR("BUG: PtlEQGet returned %d\n", rc);
  57                 LBUG();
  58         }
  59  out:
  60         spin_unlock(&svc->srv_lock);
  61         return rc;
  62 }
  63
  64 struct ptlrpc_service *
  65 ptlrpc_init_svc(__u32 nevents, __u32 nbufs,
  66                 __u32 bufsize, __u32 max_req_size,
  67                 int req_portal, int rep_portal,
  68                 obd_uuid_t uuid, svc_handler_t handler, char *name)
  69 {
  70         int err;
  71         int rc, i;
  72         struct ptlrpc_service *service;
  73         ENTRY;
  74
  75         OBD_ALLOC(service, sizeof(*service));
  76         if (!service)
  77                 RETURN(NULL);
  78
  79         service->srv_name = name;
  80         spin_lock_init(&service->srv_lock);
  81         INIT_LIST_HEAD(&service->srv_threads);
  82         init_waitqueue_head(&service->srv_waitq);
  83
  84         service->srv_max_req_size = max_req_size;
  85         service->srv_buf_size = bufsize;
  86         INIT_LIST_HEAD(&service->srv_rqbds);
  87         service->srv_nrqbds = 0;
  88         atomic_set(&service->srv_nrqbds_receiving, 0);
  89
  90         service->srv_rep_portal = rep_portal;
  91         service->srv_req_portal = req_portal;
  92         service->srv_handler = handler;
  93
  94         err = kportal_uuid_to_peer(uuid, &service->srv_self);
  95         if (err) {
  96                 CERROR("%s: cannot get peer for uuid '%s'\n", name, uuid);
  97                 OBD_FREE(service, sizeof(*service));
  98                 RETURN(NULL);
  99         }
 100
 101         rc = PtlEQAlloc(service->srv_self.peer_ni, nevents,
 102                         request_in_callback, &(service->srv_eq_h));
 103
 104         if (rc != PTL_OK) {
 105                 CERROR("%s: PtlEQAlloc failed: %d\n", name, rc);
 106                 OBD_FREE(service, sizeof(*service));
 107                 RETURN(NULL);
 108         }
 109
 110         for (i = 0; i < nbufs; i++) {
 111                 struct ptlrpc_request_buffer_desc *rqbd;
 112
 113                 OBD_ALLOC(rqbd, sizeof(*rqbd));
 114                 if (rqbd == NULL)
 115                         GOTO(failed, NULL);
 116
 117                 rqbd->rqbd_service = service;
 118                 ptl_set_inv_handle(&rqbd->rqbd_me_h);
 119                 atomic_set(&rqbd->rqbd_refcount, 0);
 120                 OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 121                 if (rqbd->rqbd_buffer == NULL) {
 122                         OBD_FREE(rqbd, sizeof(*rqbd));
 123                         GOTO(failed, NULL);
 124                 }
 125                 list_add(&rqbd->rqbd_list, &service->srv_rqbds);
 126                 service->srv_nrqbds++;
 127
 128                 ptlrpc_link_svc_me(rqbd);
 129         }
 130
 131         CDEBUG(D_NET, "Starting service listening on portal %d (eq: %lu)\n",
 132                service->srv_req_portal, service->srv_eq_h.handle_idx);
 133
 134         RETURN(service);
 135 failed:
 136         ptlrpc_unregister_service(service);
 137         return NULL;
 138 }
 139
 140 static int handle_incoming_request(struct obd_device *obddev,
 141                                    struct ptlrpc_service *svc,
 142                                    ptl_event_t *event,
 143                                    struct ptlrpc_request *request)
 144 {
 145         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 146         int rc;
 147
 148         /* FIXME: If we move to an event-driven model, we should put the request
 149          * on the stack of mds_handle instead. */
 150
 151         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 152         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 153         LASSERT (rqbd->rqbd_service == svc);
 154         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 155         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 156
 157         memset(request, 0, sizeof(*request));
 158         request->rq_svc = svc;
 159         request->rq_obd = obddev;
 160         request->rq_xid = event->match_bits;
 161         request->rq_reqmsg = event->mem_desc.start + event->offset;
 162         request->rq_reqlen = event->mlength;
 163
 164         rc = -EINVAL;
 165
 166         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 167                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "
 168                        LPD64"\n",
 169                        request->rq_reqlen, svc->srv_req_portal,
 170                        event->initiator.nid, request->rq_xid);
 171                 goto out;
 172         }
 173
 174         CDEBUG(D_RPCTRACE, "Handling RPC pid:xid:nid:opc %d:"LPX64":"LPX64":%d\n",
 175                NTOH__u32(request->rq_reqmsg->status),
 176                request->rq_xid,
 177                event->initiator.nid,
 178                NTOH__u32(request->rq_reqmsg->opc));
 179
 180         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 181                 CERROR("wrong packet type received (type=%u)\n",
 182                        request->rq_reqmsg->type);
 183                 goto out;
 184         }
 185
 186         if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 187                 CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "
 188                        LPD64"\n",
 189                        request->rq_reqmsg->magic, svc->srv_req_portal,
 190                        event->initiator.nid, request->rq_xid);
 191                 goto out;
 192         }
 193
 194         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 195                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "
 196                        LPD64"\n",
 197                        request->rq_reqmsg->version, svc->srv_req_portal,
 198                        event->initiator.nid, request->rq_xid);
 199                 goto out;
 200         }
 201
 202         CDEBUG(D_NET, "got req "LPD64" (md: %p + %d)\n", request->rq_xid,
 203                event->mem_desc.start, event->offset);
 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 /* Don't use daemonize, it removes fs struct from new thread  (bug 418) */
 235 static void ptlrpc_daemonize(void)
 236 {
 237         exit_mm(current);
 238
 239         current->session = 1;
 240         current->pgrp = 1;
 241         current->tty = NULL;
 242
 243         exit_files(current);
 244         reparent_to_init();
 245 }
 246
 247 static int ptlrpc_main(void *arg)
 248 {
 249         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 250         struct obd_device *obddev = data->dev;
 251         struct ptlrpc_service *svc = data->svc;
 252         struct ptlrpc_thread *thread = data->thread;
 253         struct ptlrpc_request *request;
 254         ptl_event_t *event;
 255         int rc = 0;
 256         unsigned long flags;
 257         ENTRY;
 258
 259         lock_kernel();
 260         ptlrpc_daemonize();
 261
 262 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
 263         sigfillset(&current->blocked);
 264         recalc_sigpending();
 265 #else
 266         spin_lock_irqsave(&current->sigmask_lock, flags);
 267         sigfillset(&current->blocked);
 268         recalc_sigpending(current);
 269         spin_unlock_irqrestore(&current->sigmask_lock, flags);
 270 #endif
 271
 272 #ifdef __arch_um__
 273         sprintf(current->comm, "%s|%d", data->name,current->thread.extern_pid);
 274 #else
 275         strcpy(current->comm, data->name);
 276 #endif
 277         unlock_kernel();
 278
 279         OBD_ALLOC(event, sizeof(*event));
 280         if (!event)
 281                 GOTO(out, rc = -ENOMEM);
 282         OBD_ALLOC(request, sizeof(*request));
 283         if (!request)
 284                 GOTO(out_event, rc = -ENOMEM);
 285
 286         /* Record that the thread is running */
 287         thread->t_flags = SVC_RUNNING;
 288         wake_up(&thread->t_ctl_waitq);
 289
 290         /* XXX maintain a list of all managed devices: insert here */
 291
 292         /* And now, loop forever on requests */
 293         while (1) {
 294                 wait_event(svc->srv_waitq,
 295                            ptlrpc_check_event(svc, thread, event));
 296
 297                 if (thread->t_flags & SVC_STOPPING) {
 298                         spin_lock(&svc->srv_lock);
 299                         thread->t_flags &= ~SVC_STOPPING;
 300                         spin_unlock(&svc->srv_lock);
 301
 302                         EXIT;
 303                         break;
 304                 }
 305
 306                 if (thread->t_flags & SVC_EVENT) {
 307                         spin_lock(&svc->srv_lock);
 308                         thread->t_flags &= ~SVC_EVENT;
 309                         spin_unlock(&svc->srv_lock);
 310
 311                         rc = handle_incoming_request(obddev, svc, event,
 312                                                      request);
 313                         continue;
 314                 }
 315
 316                 CERROR("unknown break in service");
 317                 LBUG();
 318                 EXIT;
 319                 break;
 320         }
 321
 322         OBD_FREE(request, sizeof(*request));
 323 out_event:
 324         OBD_FREE(event, sizeof(*event));
 325 out:
 326         thread->t_flags = SVC_STOPPED;
 327         wake_up(&thread->t_ctl_waitq);
 328
 329         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 330                current->pid, rc);
 331         return rc;
 332 }
 333
 334 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 335                                struct ptlrpc_thread *thread)
 336 {
 337         spin_lock(&svc->srv_lock);
 338         thread->t_flags = SVC_STOPPING;
 339         spin_unlock(&svc->srv_lock);
 340
 341         wake_up(&svc->srv_waitq);
 342         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 343 }
 344
 345 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 346 {
 347         spin_lock(&svc->srv_lock);
 348         while (!list_empty(&svc->srv_threads)) {
 349                 struct ptlrpc_thread *thread;
 350                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 351                                     t_link);
 352                 spin_unlock(&svc->srv_lock);
 353                 ptlrpc_stop_thread(svc, thread);
 354                 spin_lock(&svc->srv_lock);
 355                 list_del(&thread->t_link);
 356                 OBD_FREE(thread, sizeof(*thread));
 357         }
 358         spin_unlock(&svc->srv_lock);
 359 }
 360
 361 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 362                         char *name)
 363 {
 364         struct ptlrpc_svc_data d;
 365         struct ptlrpc_thread *thread;
 366         int rc;
 367         ENTRY;
 368
 369         OBD_ALLOC(thread, sizeof(*thread));
 370         if (thread == NULL) {
 371                 LBUG();
 372                 RETURN(-ENOMEM);
 373         }
 374         init_waitqueue_head(&thread->t_ctl_waitq);
 375
 376         d.dev = dev;
 377         d.svc = svc;
 378         d.name = name;
 379         d.thread = thread;
 380
 381         spin_lock(&svc->srv_lock);
 382         list_add(&thread->t_link, &svc->srv_threads);
 383         spin_unlock(&svc->srv_lock);
 384
 385         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 386          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 387          */
 388         rc = kernel_thread(ptlrpc_main, (void *) &d, CLONE_VM | CLONE_FILES);
 389         if (rc < 0) {
 390                 CERROR("cannot start thread\n");
 391                 OBD_FREE(thread, sizeof(*thread));
 392                 RETURN(rc);
 393         }
 394         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 395
 396         RETURN(0);
 397 }
 398
 399 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 400 {
 401         int rc;
 402
 403         LASSERT (list_empty (&service->srv_threads));
 404
 405         /* XXX We could reply (with failure) to all buffered requests
 406          * _after_ unlinking _all_ the request buffers, but _before_
 407          * freeing them.
 408          */
 409
 410         while (!list_empty (&service->srv_rqbds)) {
 411                 struct ptlrpc_request_buffer_desc *rqbd =
 412                         list_entry (service->srv_rqbds.next,
 413                                     struct ptlrpc_request_buffer_desc,
 414                                     rqbd_list);
 415
 416                 list_del (&rqbd->rqbd_list);
 417
 418                 LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 419                 /* refcount could be anything; it's possible for the
 420                  * buffers to continued to get filled after all the server
 421                  * threads exited.  But we know they _have_ exited.
 422                  */
 423
 424                 (void) PtlMEUnlink(rqbd->rqbd_me_h);
 425                 /* The callback handler could have unlinked this ME already
 426                  * (we're racing with her) but it's safe to ensure it _has_
 427                  * been unlinked.
 428                  */
 429
 430                 OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 431                 OBD_FREE (rqbd, sizeof (*rqbd));
 432                 service->srv_nrqbds--;
 433         }
 434
 435         LASSERT (service->srv_nrqbds == 0);
 436
 437         rc = PtlEQFree(service->srv_eq_h);
 438         if (rc)
 439                 CERROR("PtlEQFree failed: %d\n", rc);
 440
 441         OBD_FREE(service, sizeof(*service));
 442         if (rc)
 443                 LBUG();
 444         return rc;
 445 }