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 extern int ptl_handled_rpc(struct ptlrpc_service *service, void *start);
  37
  38 static int ptlrpc_check_event(struct ptlrpc_service *svc)
  39 {
  40         int rc = 0;
  41
  42         spin_lock(&svc->srv_lock);
  43         if (sigismember(&(current->pending.signal), SIGKILL) ||
  44             sigismember(&(current->pending.signal), SIGSTOP) ||
  45             sigismember(&(current->pending.signal), SIGCONT) ||
  46             sigismember(&(current->pending.signal), SIGINT)) {
  47                 svc->srv_flags |= SVC_KILLED;
  48                 EXIT;
  49                 rc = 1;
  50                 goto out;
  51         }
  52
  53         if ( svc->srv_flags & SVC_STOPPING ) {
  54                 EXIT;
  55                 rc = 1;
  56                 goto out;
  57         }
  58
  59         if (svc->srv_flags & SVC_EVENT)
  60                 BUG();
  61
  62         if ( svc->srv_eq_h ) {
  63                 int err;
  64                 err = PtlEQGet(svc->srv_eq_h, &svc->srv_ev);
  65
  66                 if (err == PTL_OK) {
  67                         svc->srv_flags |= SVC_EVENT;
  68                         EXIT;
  69                         rc = 1;
  70                         goto out;
  71                 }
  72
  73                 if (err != PTL_EQ_EMPTY) {
  74                         CDEBUG(D_NET, "BUG: PtlEQGet returned %d\n", rc);
  75                         BUG();
  76                 }
  77
  78                 EXIT;
  79                 rc = 0;
  80                 goto out;
  81         }
  82
  83         if (!list_empty(&svc->srv_reqs)) {
  84                 svc->srv_flags |= SVC_LIST;
  85                 EXIT;
  86                 rc = 1;
  87                 goto out;
  88         }
  89
  90         EXIT;
  91  out:
  92         spin_unlock(&svc->srv_lock);
  93         return rc;
  94 }
  95
  96 struct ptlrpc_service *
  97 ptlrpc_init_svc(__u32 bufsize, int req_portal, int rep_portal, char *uuid,
  98                 req_unpack_t unpack, rep_pack_t pack, svc_handler_t handler)
  99 {
 100         int err;
 101         struct ptlrpc_service *svc;
 102
 103         OBD_ALLOC(svc, sizeof(*svc));
 104         if ( !svc ) {
 105                 CERROR("no memory\n");
 106                 return NULL;
 107         }
 108
 109         memset(svc, 0, sizeof(*svc));
 110
 111         spin_lock_init(&svc->srv_lock);
 112         INIT_LIST_HEAD(&svc->srv_reqs);
 113         init_waitqueue_head(&svc->srv_ctl_waitq);
 114         init_waitqueue_head(&svc->srv_waitq);
 115
 116         svc->srv_thread = NULL;
 117         svc->srv_flags = 0;
 118
 119         svc->srv_buf_size = bufsize;
 120         svc->srv_rep_portal = rep_portal;
 121         svc->srv_req_portal = req_portal;
 122         svc->srv_req_unpack = unpack;
 123         svc->srv_rep_pack = pack;
 124         svc->srv_handler = handler;
 125         err = kportal_uuid_to_peer(uuid, &svc->srv_self);
 126         if (err) {
 127                 CERROR("cannot get peer for uuid %s", uuid);
 128                 OBD_FREE(svc, sizeof(*svc));
 129                 return NULL;
 130         }
 131         return svc;
 132 }
 133
 134 static int ptlrpc_main(void *arg)
 135 {
 136         int rc;
 137         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 138         struct obd_device *obddev = data->dev;
 139         struct ptlrpc_service *svc = data->svc;
 140
 141         ENTRY;
 142
 143         lock_kernel();
 144         daemonize();
 145         spin_lock_irq(&current->sigmask_lock);
 146         sigfillset(&current->blocked);
 147         recalc_sigpending(current);
 148         spin_unlock_irq(&current->sigmask_lock);
 149
 150         sprintf(current->comm, data->name);
 151
 152         /* Record that the  thread is running */
 153         svc->srv_thread = current;
 154         svc->srv_flags = SVC_RUNNING;
 155         wake_up(&svc->srv_ctl_waitq);
 156
 157         /* XXX maintain a list of all managed devices: insert here */
 158
 159         /* And now, loop forever on requests */
 160         while (1) {
 161                 wait_event(svc->srv_waitq, ptlrpc_check_event(svc));
 162
 163                 spin_lock(&svc->srv_lock);
 164                 if (svc->srv_flags & SVC_SIGNAL) {
 165                         EXIT;
 166                         spin_unlock(&svc->srv_lock);
 167                         break;
 168                 }
 169
 170                 if (svc->srv_flags & SVC_STOPPING) {
 171                         EXIT;
 172                         spin_unlock(&svc->srv_lock);
 173                         break;
 174                 }
 175
 176                 if (svc->srv_flags & SVC_EVENT) {
 177                         struct ptlrpc_request request;
 178                         void *start;
 179                         svc->srv_flags = SVC_RUNNING;
 180
 181                         /* FIXME: If we move to an event-driven model,
 182                          * we should put the request on the stack of
 183                          * mds_handle instead. */
 184                         start = svc->srv_ev.mem_desc.start;
 185                         memset(&request, 0, sizeof(request));
 186                         request.rq_obd = obddev;
 187                         request.rq_reqbuf = (svc->srv_ev.mem_desc.start +
 188                                              svc->srv_ev.offset);
 189                         request.rq_reqlen = svc->srv_ev.mem_desc.length;
 190                         request.rq_xid = svc->srv_ev.match_bits;
 191                         CDEBUG(D_NET, "got req %d\n", request.rq_xid);
 192
 193                         request.rq_peer.peer_nid = svc->srv_ev.initiator.nid;
 194                         /* FIXME: this NI should be the incoming NI.
 195                          * We don't know how to find that from here. */
 196                         request.rq_peer.peer_ni = svc->srv_self.peer_ni;
 197                         svc->srv_flags &= ~SVC_EVENT;
 198
 199                         spin_unlock(&svc->srv_lock);
 200                         rc = svc->srv_handler(obddev, svc, &request);
 201                         ptl_handled_rpc(svc, start);
 202                         continue;
 203                 }
 204
 205                 if (svc->srv_flags & SVC_LIST) {
 206                         struct ptlrpc_request *request;
 207                         svc->srv_flags = SVC_RUNNING;
 208
 209                         request = list_entry(svc->srv_reqs.next,
 210                                              struct ptlrpc_request,
 211                                              rq_list);
 212                         list_del(&request->rq_list);
 213                         spin_unlock(&svc->srv_lock);
 214                         rc = svc->srv_handler(obddev, svc, request);
 215                         continue;
 216                 }
 217                 CERROR("unknown break in service");
 218                 spin_unlock(&svc->srv_lock);
 219                 break;
 220         }
 221
 222         svc->srv_thread = NULL;
 223         svc->srv_flags = SVC_STOPPED;
 224         wake_up(&svc->srv_ctl_waitq);
 225         CDEBUG(D_NET, "svc exiting process %d\n", current->pid);
 226         return 0;
 227 }
 228
 229 void ptlrpc_stop_thread(struct ptlrpc_service *svc)
 230 {
 231         svc->srv_flags = SVC_STOPPING;
 232
 233         wake_up(&svc->srv_waitq);
 234         wait_event_interruptible(svc->srv_ctl_waitq,
 235                                  (svc->srv_flags & SVC_STOPPED));
 236 }
 237
 238 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 239                                 char *name)
 240 {
 241         struct ptlrpc_svc_data d;
 242         int rc;
 243         ENTRY;
 244
 245         d.dev = dev;
 246         d.svc = svc;
 247         d.name = name;
 248
 249         init_waitqueue_head(&svc->srv_waitq);
 250
 251         init_waitqueue_head(&svc->srv_ctl_waitq);
 252         rc = kernel_thread(ptlrpc_main, (void *) &d,
 253                            CLONE_VM | CLONE_FS | CLONE_FILES);
 254         if (rc < 0) {
 255                 CERROR("cannot start thread\n");
 256                 return -EINVAL;
 257         }
 258         wait_event(svc->srv_ctl_waitq, svc->srv_flags & SVC_RUNNING);
 259
 260         EXIT;
 261         return 0;
 262 }
 263
 264
 265 int rpc_register_service(struct ptlrpc_service *service, char *uuid)
 266 {
 267         struct lustre_peer peer;
 268         int rc, i;
 269
 270         rc = kportal_uuid_to_peer(uuid, &peer);
 271         if (rc != 0) {
 272                 CERROR("Invalid uuid \"%s\"\n", uuid);
 273                 return -EINVAL;
 274         }
 275
 276         service->srv_ring_length = RPC_RING_LENGTH;
 277         service->srv_me_active = 0;
 278         service->srv_md_active = 0;
 279
 280         service->srv_id.addr_kind = PTL_ADDR_GID;
 281         service->srv_id.gid = PTL_ID_ANY;
 282         service->srv_id.rid = PTL_ID_ANY;
 283
 284         rc = PtlEQAlloc(peer.peer_ni, 128, server_request_callback,
 285                         service, &(service->srv_eq_h));
 286
 287         if (rc != PTL_OK) {
 288                 CERROR("PtlEQAlloc failed: %d\n", rc);
 289                 return rc;
 290         }
 291
 292         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 293                service->srv_req_portal);
 294
 295         /* Attach the leading ME on which we build the ring */
 296         rc = PtlMEAttach(peer.peer_ni, service->srv_req_portal,
 297                          service->srv_id, 0, ~0, PTL_RETAIN,
 298                          &(service->srv_me_h[0]));
 299
 300         if (rc != PTL_OK) {
 301                 CERROR("PtlMEAttach failed: %d\n", rc);
 302                 return rc;
 303         }
 304
 305         for (i = 0; i < service->srv_ring_length; i++) {
 306                 OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
 307
 308                 if (service->srv_buf[i] == NULL) {
 309                         CERROR("no memory\n");
 310                         return -ENOMEM;
 311                 }
 312
 313
 314                 /* Insert additional ME's to the ring */
 315                 if (i > 0) {
 316                         rc = PtlMEInsert(service->srv_me_h[i-1],
 317                                          service->srv_id, 0, ~0, PTL_RETAIN,
 318                                          PTL_INS_AFTER,&(service->srv_me_h[i]));
 319                         service->srv_me_tail = i;
 320
 321                         if (rc != PTL_OK) {
 322                                 CERROR("PtlMEInsert failed: %d\n", rc);
 323                                 return rc;
 324                         }
 325                 }
 326
 327                 service->srv_ref_count[i] = 0;
 328                 service->srv_md[i].start         = service->srv_buf[i];
 329                 service->srv_md[i].length        = service->srv_buf_size;
 330                 service->srv_md[i].max_offset    = service->srv_buf_size;
 331                 service->srv_md[i].threshold     = PTL_MD_THRESH_INF;
 332                 service->srv_md[i].options       = PTL_MD_OP_PUT;
 333                 service->srv_md[i].user_ptr      = service;
 334                 service->srv_md[i].eventq        = service->srv_eq_h;
 335                 service->srv_md[i].max_offset    = service->srv_buf_size;
 336
 337                 rc = PtlMDAttach(service->srv_me_h[i], service->srv_md[i],
 338                                  PTL_RETAIN, &(service->srv_md_h[i]));
 339
 340                 if (rc != PTL_OK) {
 341                         /* cleanup */
 342                         CERROR("PtlMDAttach failed: %d\n", rc);
 343                         return rc;
 344                 }
 345         }
 346
 347         return 0;
 348 }
 349
 350 int rpc_unregister_service(struct ptlrpc_service *service)
 351 {
 352         int rc, i;
 353
 354         for (i = 0; i < service->srv_ring_length; i++) {
 355                 rc = PtlMDUnlink(service->srv_md_h[i]);
 356                 if (rc)
 357                         CERROR("PtlMDUnlink failed: %d\n", rc);
 358
 359                 rc = PtlMEUnlink(service->srv_me_h[i]);
 360                 if (rc)
 361                         CERROR("PtlMEUnlink failed: %d\n", rc);
 362
 363                 if (service->srv_buf[i] != NULL)
 364                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 365                 service->srv_buf[i] = NULL;
 366         }
 367
 368         rc = PtlEQFree(service->srv_eq_h);
 369         if (rc)
 370                 CERROR("PtlEQFree failed: %d\n", rc);
 371
 372         return 0;
 373 }