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                 struct obd_uuid *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->uuid, &service->srv_self);
  95         if (err) {
  96                 CERROR("%s: cannot get peer for uuid '%s'\n", name,
  97                        uuid->uuid);
  98                 OBD_FREE(service, sizeof(*service));
  99                 RETURN(NULL);
 100         }
 101
 102         rc = PtlEQAlloc(service->srv_self.peer_ni, nevents,
 103                         request_in_callback, &(service->srv_eq_h));
 104
 105         if (rc != PTL_OK) {
 106                 CERROR("%s: PtlEQAlloc failed: %d\n", name, rc);
 107                 OBD_FREE(service, sizeof(*service));
 108                 RETURN(NULL);
 109         }
 110
 111         for (i = 0; i < nbufs; i++) {
 112                 struct ptlrpc_request_buffer_desc *rqbd;
 113
 114                 OBD_ALLOC(rqbd, sizeof(*rqbd));
 115                 if (rqbd == NULL)
 116                         GOTO(failed, NULL);
 117
 118                 rqbd->rqbd_service = service;
 119                 ptl_set_inv_handle(&rqbd->rqbd_me_h);
 120                 atomic_set(&rqbd->rqbd_refcount, 0);
 121                 OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
 122                 if (rqbd->rqbd_buffer == NULL) {
 123                         OBD_FREE(rqbd, sizeof(*rqbd));
 124                         GOTO(failed, NULL);
 125                 }
 126                 list_add(&rqbd->rqbd_list, &service->srv_rqbds);
 127                 service->srv_nrqbds++;
 128
 129                 ptlrpc_link_svc_me(rqbd);
 130         }
 131
 132         CDEBUG(D_NET, "Starting service listening on portal %d (eq: %lu)\n",
 133                service->srv_req_portal, service->srv_eq_h.handle_idx);
 134
 135         RETURN(service);
 136 failed:
 137         ptlrpc_unregister_service(service);
 138         return NULL;
 139 }
 140
 141 static int handle_incoming_request(struct obd_device *obddev,
 142                                    struct ptlrpc_service *svc,
 143                                    ptl_event_t *event,
 144                                    struct ptlrpc_request *request)
 145 {
 146         struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
 147         int rc;
 148
 149         /* FIXME: If we move to an event-driven model, we should put the request
 150          * on the stack of mds_handle instead. */
 151
 152         LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 153         LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
 154         LASSERT (rqbd->rqbd_service == svc);
 155         LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
 156         LASSERT (event->offset + event->mlength <= svc->srv_buf_size);
 157
 158         memset(request, 0, sizeof(*request));
 159         request->rq_svc = svc;
 160         request->rq_obd = obddev;
 161         request->rq_xid = event->match_bits;
 162         request->rq_reqmsg = event->mem_desc.start + event->offset;
 163         request->rq_reqlen = event->mlength;
 164
 165         rc = -EINVAL;
 166
 167         if (request->rq_reqlen < sizeof(struct lustre_msg)) {
 168                 CERROR("incomplete request (%d): ptl %d from "LPX64" xid "
 169                        LPU64"\n",
 170                        request->rq_reqlen, svc->srv_req_portal,
 171                        event->initiator.nid, request->rq_xid);
 172                 goto out;
 173         }
 174
 175         CDEBUG(D_RPCTRACE, "Handling RPC pid:xid:nid:opc %d:"LPU64":"LPX64":%d\n",
 176                NTOH__u32(request->rq_reqmsg->status),
 177                request->rq_xid,
 178                event->initiator.nid,
 179                NTOH__u32(request->rq_reqmsg->opc));
 180
 181         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 182                 CERROR("wrong packet type received (type=%u)\n",
 183                        request->rq_reqmsg->type);
 184                 goto out;
 185         }
 186
 187         if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
 188                 CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "
 189                        LPD64"\n",
 190                        request->rq_reqmsg->magic, svc->srv_req_portal,
 191                        event->initiator.nid, request->rq_xid);
 192                 goto out;
 193         }
 194
 195         if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
 196                 CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "
 197                        LPD64"\n",
 198                        request->rq_reqmsg->version, svc->srv_req_portal,
 199                        event->initiator.nid, request->rq_xid);
 200                 goto out;
 201         }
 202
 203         CDEBUG(D_NET, "got req "LPD64" (md: %p + %d)\n", request->rq_xid,
 204                event->mem_desc.start, event->offset);
 205
 206         request->rq_peer.peer_nid = event->initiator.nid;
 207         /* FIXME: this NI should be the incoming NI.
 208          * We don't know how to find that from here. */
 209         request->rq_peer.peer_ni = svc->srv_self.peer_ni;
 210
 211         request->rq_export = class_conn2export((struct lustre_handle *)
 212                                                request->rq_reqmsg);
 213
 214         if (request->rq_export) {
 215                 request->rq_connection = request->rq_export->exp_connection;
 216                 ptlrpc_connection_addref(request->rq_connection);
 217         } else {
 218                 /* create a (hopefully temporary) connection that will be used
 219                  * to send the reply if this call doesn't create an export.
 220                  * XXX revisit this when we revamp ptlrpc */
 221                 request->rq_connection =
 222                         ptlrpc_get_connection(&request->rq_peer, NULL);
 223         }
 224
 225         rc = svc->srv_handler(request);
 226         ptlrpc_put_connection(request->rq_connection);
 227
 228  out:
 229         if (atomic_dec_and_test (&rqbd->rqbd_refcount)) /* last reference? */
 230                 ptlrpc_link_svc_me (rqbd);
 231
 232         return rc;
 233 }
 234
 235 /* Don't use daemonize, it removes fs struct from new thread  (bug 418) */
 236 static void ptlrpc_daemonize(void)
 237 {
 238         exit_mm(current);
 239
 240         current->session = 1;
 241         current->pgrp = 1;
 242         current->tty = NULL;
 243
 244         exit_files(current);
 245         reparent_to_init();
 246 }
 247
 248 static int ptlrpc_main(void *arg)
 249 {
 250         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 251         struct obd_device *obddev = data->dev;
 252         struct ptlrpc_service *svc = data->svc;
 253         struct ptlrpc_thread *thread = data->thread;
 254         struct ptlrpc_request *request;
 255         ptl_event_t *event;
 256         int rc = 0;
 257         unsigned long flags;
 258         ENTRY;
 259
 260         lock_kernel();
 261         ptlrpc_daemonize();
 262
 263 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)
 264         sigfillset(&current->blocked);
 265         recalc_sigpending();
 266 #else
 267         spin_lock_irqsave(&current->sigmask_lock, flags);
 268         sigfillset(&current->blocked);
 269         recalc_sigpending(current);
 270         spin_unlock_irqrestore(&current->sigmask_lock, flags);
 271 #endif
 272
 273 #ifdef __arch_um__
 274         sprintf(current->comm, "%s|%d", data->name,current->thread.extern_pid);
 275 #else
 276         strcpy(current->comm, data->name);
 277 #endif
 278         unlock_kernel();
 279
 280         OBD_ALLOC(event, sizeof(*event));
 281         if (!event)
 282                 GOTO(out, rc = -ENOMEM);
 283         OBD_ALLOC(request, sizeof(*request));
 284         if (!request)
 285                 GOTO(out_event, rc = -ENOMEM);
 286
 287         /* Record that the thread is running */
 288         thread->t_flags = SVC_RUNNING;
 289         wake_up(&thread->t_ctl_waitq);
 290
 291         /* XXX maintain a list of all managed devices: insert here */
 292
 293         /* And now, loop forever on requests */
 294         while (1) {
 295                 wait_event(svc->srv_waitq,
 296                            ptlrpc_check_event(svc, thread, event));
 297
 298                 if (thread->t_flags & SVC_STOPPING) {
 299                         spin_lock(&svc->srv_lock);
 300                         thread->t_flags &= ~SVC_STOPPING;
 301                         spin_unlock(&svc->srv_lock);
 302
 303                         EXIT;
 304                         break;
 305                 }
 306
 307                 if (thread->t_flags & SVC_EVENT) {
 308                         spin_lock(&svc->srv_lock);
 309                         thread->t_flags &= ~SVC_EVENT;
 310                         spin_unlock(&svc->srv_lock);
 311
 312                         rc = handle_incoming_request(obddev, svc, event,
 313                                                      request);
 314                         continue;
 315                 }
 316
 317                 CERROR("unknown break in service");
 318                 LBUG();
 319                 EXIT;
 320                 break;
 321         }
 322
 323         OBD_FREE(request, sizeof(*request));
 324 out_event:
 325         OBD_FREE(event, sizeof(*event));
 326 out:
 327         thread->t_flags = SVC_STOPPED;
 328         wake_up(&thread->t_ctl_waitq);
 329
 330         CDEBUG(D_NET, "service thread exiting, process %d: rc = %d\n",
 331                current->pid, rc);
 332         return rc;
 333 }
 334
 335 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
 336                                struct ptlrpc_thread *thread)
 337 {
 338         spin_lock(&svc->srv_lock);
 339         thread->t_flags = SVC_STOPPING;
 340         spin_unlock(&svc->srv_lock);
 341
 342         wake_up(&svc->srv_waitq);
 343         wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
 344 }
 345
 346 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
 347 {
 348         spin_lock(&svc->srv_lock);
 349         while (!list_empty(&svc->srv_threads)) {
 350                 struct ptlrpc_thread *thread;
 351                 thread = list_entry(svc->srv_threads.next, struct ptlrpc_thread,
 352                                     t_link);
 353                 spin_unlock(&svc->srv_lock);
 354                 ptlrpc_stop_thread(svc, thread);
 355                 spin_lock(&svc->srv_lock);
 356                 list_del(&thread->t_link);
 357                 OBD_FREE(thread, sizeof(*thread));
 358         }
 359         spin_unlock(&svc->srv_lock);
 360 }
 361
 362 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 363                         char *name)
 364 {
 365         struct ptlrpc_svc_data d;
 366         struct ptlrpc_thread *thread;
 367         int rc;
 368         ENTRY;
 369
 370         OBD_ALLOC(thread, sizeof(*thread));
 371         if (thread == NULL) {
 372                 LBUG();
 373                 RETURN(-ENOMEM);
 374         }
 375         init_waitqueue_head(&thread->t_ctl_waitq);
 376
 377         d.dev = dev;
 378         d.svc = svc;
 379         d.name = name;
 380         d.thread = thread;
 381
 382         spin_lock(&svc->srv_lock);
 383         list_add(&thread->t_link, &svc->srv_threads);
 384         spin_unlock(&svc->srv_lock);
 385
 386         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 387          * just drop the VM and FILES in ptlrpc_daemonize() right away.
 388          */
 389         rc = kernel_thread(ptlrpc_main, (void *) &d, CLONE_VM | CLONE_FILES);
 390         if (rc < 0) {
 391                 CERROR("cannot start thread\n");
 392                 OBD_FREE(thread, sizeof(*thread));
 393                 RETURN(rc);
 394         }
 395         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 396
 397         RETURN(0);
 398 }
 399
 400 int ptlrpc_unregister_service(struct ptlrpc_service *service)
 401 {
 402         int rc;
 403
 404         LASSERT (list_empty (&service->srv_threads));
 405
 406         /* XXX We could reply (with failure) to all buffered requests
 407          * _after_ unlinking _all_ the request buffers, but _before_
 408          * freeing them.
 409          */
 410
 411         while (!list_empty (&service->srv_rqbds)) {
 412                 struct ptlrpc_request_buffer_desc *rqbd =
 413                         list_entry (service->srv_rqbds.next,
 414                                     struct ptlrpc_request_buffer_desc,
 415                                     rqbd_list);
 416
 417                 list_del (&rqbd->rqbd_list);
 418
 419                 LASSERT (atomic_read (&rqbd->rqbd_refcount) > 0);
 420                 /* refcount could be anything; it's possible for the
 421                  * buffers to continued to get filled after all the server
 422                  * threads exited.  But we know they _have_ exited.
 423                  */
 424
 425                 (void) PtlMEUnlink(rqbd->rqbd_me_h);
 426                 /* The callback handler could have unlinked this ME already
 427                  * (we're racing with her) but it's safe to ensure it _has_
 428                  * been unlinked.
 429                  */
 430
 431                 OBD_FREE (rqbd->rqbd_buffer, service->srv_buf_size);
 432                 OBD_FREE (rqbd, sizeof (*rqbd));
 433                 service->srv_nrqbds--;
 434         }
 435
 436         LASSERT (service->srv_nrqbds == 0);
 437
 438         rc = PtlEQFree(service->srv_eq_h);
 439         if (rc)
 440                 CERROR("PtlEQFree failed: %d\n", rc);
 441
 442         OBD_FREE(service, sizeof(*service));
 443         if (rc)
 444                 LBUG();
 445         return rc;
 446 }