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 EXPORT_SYMTAB
  24
  25 #include <linux/config.h>
  26 #include <linux/module.h>
  27 #include <linux/kernel.h>
  28
  29 #define DEBUG_SUBSYSTEM S_RPC
  30
  31 #include <linux/obd_support.h>
  32 #include <linux/obd_class.h>
  33 #include <linux/lustre_net.h>
  34
  35 extern int server_request_callback(ptl_event_t *ev, void *data);
  36
  37 static int ptlrpc_check_event(struct ptlrpc_service *svc)
  38 {
  39         int rc = 0;
  40
  41         spin_lock(&svc->srv_lock);
  42         if (sigismember(&(current->pending.signal), SIGKILL) ||
  43             sigismember(&(current->pending.signal), SIGSTOP) ||
  44             sigismember(&(current->pending.signal), SIGCONT) ||
  45             sigismember(&(current->pending.signal), SIGINT)) {
  46                 svc->srv_flags |= SVC_KILLED;
  47                 EXIT;
  48                 rc = 1;
  49                 goto out;
  50         }
  51
  52         if ( svc->srv_flags & SVC_STOPPING ) {
  53                 EXIT;
  54                 rc = 1;
  55                 goto out;
  56         }
  57
  58         if (svc->srv_flags & SVC_EVENT)
  59                 BUG();
  60
  61         if ( svc->srv_eq_h ) {
  62                 int err;
  63                 err = PtlEQGet(svc->srv_eq_h, &svc->srv_ev);
  64
  65                 if (err == PTL_OK) {
  66                         svc->srv_flags |= SVC_EVENT;
  67                         EXIT;
  68                         rc = 1;
  69                         goto out;
  70                 }
  71
  72                 if (err != PTL_EQ_EMPTY) {
  73                         CDEBUG(D_NET, "BUG: PtlEQGet returned %d\n", rc);
  74                         BUG();
  75                 }
  76
  77                 EXIT;
  78                 rc = 0;
  79                 goto out;
  80         }
  81
  82         if (!list_empty(&svc->srv_reqs)) {
  83                 svc->srv_flags |= SVC_LIST;
  84                 EXIT;
  85                 rc = 1;
  86                 goto out;
  87         }
  88
  89         EXIT;
  90  out:
  91         spin_unlock(&svc->srv_lock);
  92         return rc;
  93 }
  94
  95 struct ptlrpc_service *
  96 ptlrpc_init_svc(__u32 bufsize, int req_portal, int rep_portal, char *uuid,
  97                 req_unpack_t unpack, rep_pack_t pack, svc_handler_t handler)
  98 {
  99         int err;
 100         struct ptlrpc_service *svc;
 101
 102         OBD_ALLOC(svc, sizeof(*svc));
 103         if ( !svc ) {
 104                 CERROR("no memory\n");
 105                 return NULL;
 106         }
 107
 108         memset(svc, 0, sizeof(*svc));
 109
 110         spin_lock_init(&svc->srv_lock);
 111         INIT_LIST_HEAD(&svc->srv_reqs);
 112         init_waitqueue_head(&svc->srv_ctl_waitq);
 113         init_waitqueue_head(&svc->srv_waitq);
 114
 115         svc->srv_thread = NULL;
 116         svc->srv_flags = 0;
 117
 118         svc->srv_buf_size = bufsize;
 119         svc->srv_rep_portal = rep_portal;
 120         svc->srv_req_portal = req_portal;
 121         svc->srv_req_unpack = unpack;
 122         svc->srv_rep_pack = pack;
 123         svc->srv_handler = handler;
 124         err = kportal_uuid_to_peer(uuid, &svc->srv_self);
 125         if (err) {
 126                 CERROR("cannot get peer for uuid %s", uuid);
 127                 OBD_FREE(svc, sizeof(*svc));
 128                 return NULL;
 129         }
 130         return svc;
 131 }
 132
 133 static int ptlrpc_main(void *arg)
 134 {
 135         int rc;
 136         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 137         struct obd_device *obddev = data->dev;
 138         struct ptlrpc_service *svc = data->svc;
 139
 140         ENTRY;
 141
 142         lock_kernel();
 143         daemonize();
 144         spin_lock_irq(&current->sigmask_lock);
 145         sigfillset(&current->blocked);
 146         recalc_sigpending(current);
 147         spin_unlock_irq(&current->sigmask_lock);
 148
 149         sprintf(current->comm, data->name);
 150
 151         /* Record that the  thread is running */
 152         svc->srv_thread = current;
 153         svc->srv_flags = SVC_RUNNING;
 154         wake_up(&svc->srv_ctl_waitq);
 155
 156         /* XXX maintain a list of all managed devices: insert here */
 157
 158         /* And now, loop forever on requests */
 159         while (1) {
 160                 wait_event(svc->srv_waitq, ptlrpc_check_event(svc));
 161
 162                 spin_lock(&svc->srv_lock);
 163                 if (svc->srv_flags & SVC_SIGNAL) {
 164                         EXIT;
 165                         spin_unlock(&svc->srv_lock);
 166                         break;
 167                 }
 168
 169                 if (svc->srv_flags & SVC_STOPPING) {
 170                         EXIT;
 171                         spin_unlock(&svc->srv_lock);
 172                         break;
 173                 }
 174
 175                 if (svc->srv_flags & SVC_EVENT) {
 176                         struct ptlrpc_request request;
 177                         svc->srv_flags = SVC_RUNNING;
 178
 179                         /* FIXME: If we move to an event-driven model,
 180                          * we should put the request on the stack of
 181                          * mds_handle instead. */
 182                         memset(&request, 0, sizeof(request));
 183                         request.rq_obd = obddev;
 184                         request.rq_reqbuf = (svc->srv_ev.mem_desc.start +
 185                                              svc->srv_ev.offset);
 186                         request.rq_reqlen = svc->srv_ev.mem_desc.length;
 187                         request.rq_xid = svc->srv_ev.match_bits;
 188                         CDEBUG(D_NET, "got req %d\n", request.rq_xid);
 189
 190                         request.rq_peer.peer_nid = svc->srv_ev.initiator.nid;
 191                         /* FIXME: this NI should be the incoming NI.
 192                          * We don't know how to find that from here. */
 193                         request.rq_peer.peer_ni = svc->srv_self.peer_ni;
 194                         svc->srv_flags &= ~SVC_EVENT;
 195
 196                         spin_unlock(&svc->srv_lock);
 197                         rc = svc->srv_handler(obddev, svc, &request);
 198                         ptl_received_rpc(svc);
 199                         continue;
 200                 }
 201
 202                 if (svc->srv_flags & SVC_LIST) {
 203                         struct ptlrpc_request *request;
 204                         svc->srv_flags = SVC_RUNNING;
 205
 206                         request = list_entry(svc->srv_reqs.next,
 207                                              struct ptlrpc_request,
 208                                              rq_list);
 209                         list_del(&request->rq_list);
 210                         spin_unlock(&svc->srv_lock);
 211                         rc = svc->srv_handler(obddev, svc, request);
 212                         continue;
 213                 }
 214                 CERROR("unknown break in service");
 215                 spin_unlock(&svc->srv_lock);
 216                 break;
 217         }
 218
 219         svc->srv_thread = NULL;
 220         svc->srv_flags = SVC_STOPPED;
 221         wake_up(&svc->srv_ctl_waitq);
 222         CDEBUG(D_NET, "svc exiting process %d\n", current->pid);
 223         return 0;
 224 }
 225
 226 void ptlrpc_stop_thread(struct ptlrpc_service *svc)
 227 {
 228         svc->srv_flags = SVC_STOPPING;
 229
 230         wake_up(&svc->srv_waitq);
 231         wait_event_interruptible(svc->srv_ctl_waitq,
 232                                  (svc->srv_flags & SVC_STOPPED));
 233 }
 234
 235 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 236                                 char *name)
 237 {
 238         struct ptlrpc_svc_data d;
 239         int rc;
 240         ENTRY;
 241
 242         d.dev = dev;
 243         d.svc = svc;
 244         d.name = name;
 245
 246         init_waitqueue_head(&svc->srv_waitq);
 247
 248         init_waitqueue_head(&svc->srv_ctl_waitq);
 249         rc = kernel_thread(ptlrpc_main, (void *) &d,
 250                            CLONE_VM | CLONE_FS | CLONE_FILES);
 251         if (rc < 0) {
 252                 CERROR("cannot start thread\n");
 253                 return -EINVAL;
 254         }
 255         wait_event(svc->srv_ctl_waitq, svc->srv_flags & SVC_RUNNING);
 256
 257         EXIT;
 258         return 0;
 259 }
 260
 261
 262 int rpc_register_service(struct ptlrpc_service *service, char *uuid)
 263 {
 264         struct lustre_peer peer;
 265         int rc, i;
 266
 267         rc = kportal_uuid_to_peer(uuid, &peer);
 268         if (rc != 0) {
 269                 CERROR("Invalid uuid \"%s\"\n", uuid);
 270                 return -EINVAL;
 271         }
 272
 273         service->srv_ring_length = RPC_RING_LENGTH;
 274         service->srv_me_active = 0;
 275         service->srv_md_active = 0;
 276
 277         service->srv_id.addr_kind = PTL_ADDR_GID;
 278         service->srv_id.gid = PTL_ID_ANY;
 279         service->srv_id.rid = PTL_ID_ANY;
 280
 281         rc = PtlEQAlloc(peer.peer_ni, 128, server_request_callback,
 282                         service, &(service->srv_eq_h));
 283
 284         if (rc != PTL_OK) {
 285                 CERROR("PtlEQAlloc failed: %d\n", rc);
 286                 return rc;
 287         }
 288
 289         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 290                service->srv_req_portal);
 291
 292         /* Attach the leading ME on which we build the ring */
 293         rc = PtlMEAttach(peer.peer_ni, service->srv_req_portal,
 294                          service->srv_id, 0, ~0, PTL_RETAIN,
 295                          &(service->srv_me_h[0]));
 296
 297         if (rc != PTL_OK) {
 298                 CERROR("PtlMEAttach failed: %d\n", rc);
 299                 return rc;
 300         }
 301
 302         for (i = 0; i < service->srv_ring_length; i++) {
 303                 OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
 304
 305                 if (service->srv_buf[i] == NULL) {
 306                         CERROR("no memory\n");
 307                         return -ENOMEM;
 308                 }
 309
 310
 311                 /* Insert additional ME's to the ring */
 312                 if (i > 0) {
 313                         rc = PtlMEInsert(service->srv_me_h[i-1],
 314                                          service->srv_id, 0, ~0, PTL_RETAIN,
 315                                          PTL_INS_AFTER,&(service->srv_me_h[i]));
 316                         service->srv_me_tail = i;
 317
 318                         if (rc != PTL_OK) {
 319                                 CERROR("PtlMEInsert failed: %d\n", rc);
 320                                 return rc;
 321                         }
 322                 }
 323
 324                 service->srv_ref_count[i] = 0;
 325                 service->srv_md[i].start         = service->srv_buf[i];
 326                 service->srv_md[i].length        = service->srv_buf_size;
 327                 service->srv_md[i].max_offset    = service->srv_buf_size;
 328                 service->srv_md[i].threshold     = PTL_MD_THRESH_INF;
 329                 service->srv_md[i].options       = PTL_MD_OP_PUT;
 330                 service->srv_md[i].user_ptr      = service;
 331                 service->srv_md[i].eventq        = service->srv_eq_h;
 332                 service->srv_md[i].max_offset    = service->srv_buf_size;
 333
 334                 rc = PtlMDAttach(service->srv_me_h[i], service->srv_md[i],
 335                                  PTL_RETAIN, &(service->srv_md_h[i]));
 336
 337                 if (rc != PTL_OK) {
 338                         /* cleanup */
 339                         CERROR("PtlMDAttach failed: %d\n", rc);
 340                         return rc;
 341                 }
 342         }
 343
 344         return 0;
 345 }
 346
 347 int rpc_unregister_service(struct ptlrpc_service *service)
 348 {
 349         int rc, i;
 350
 351         for (i = 0; i < service->srv_ring_length; i++) {
 352                 rc = PtlMDUnlink(service->srv_md_h[i]);
 353                 if (rc)
 354                         CERROR("PtlMDUnlink failed: %d\n", rc);
 355
 356                 rc = PtlMEUnlink(service->srv_me_h[i]);
 357                 if (rc)
 358                         CERROR("PtlMEUnlink failed: %d\n", rc);
 359
 360                 if (service->srv_buf[i] != NULL)
 361                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 362                 service->srv_buf[i] = NULL;
 363         }
 364
 365         rc = PtlEQFree(service->srv_eq_h);
 366         if (rc)
 367                 CERROR("PtlEQFree failed: %d\n", rc);
 368
 369         return 0;
 370 }