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                 LBUG();
  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                         LBUG();
  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                 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
 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                         void *start;
 178                         svc->srv_flags = SVC_RUNNING;
 179
 180                         /* FIXME: If we move to an event-driven model,
 181                          * we should put the request on the stack of
 182                          * mds_handle instead. */
 183                         start = svc->srv_ev.mem_desc.start;
 184                         memset(&request, 0, sizeof(request));
 185                         request.rq_obd = obddev;
 186                         request.rq_reqbuf = (svc->srv_ev.mem_desc.start +
 187                                              svc->srv_ev.offset);
 188                         request.rq_reqlen = svc->srv_ev.mem_desc.length;
 189                         request.rq_xid = svc->srv_ev.match_bits;
 190                         CDEBUG(D_NET, "got req %d\n", request.rq_xid);
 191
 192                         request.rq_peer.peer_nid = svc->srv_ev.initiator.nid;
 193                         /* FIXME: this NI should be the incoming NI.
 194                          * We don't know how to find that from here. */
 195                         request.rq_peer.peer_ni = svc->srv_self.peer_ni;
 196                         svc->srv_flags &= ~SVC_EVENT;
 197
 198                         spin_unlock(&svc->srv_lock);
 199                         rc = svc->srv_handler(obddev, svc, &request);
 200                         ptl_handled_rpc(svc, start);
 201                         continue;
 202                 }
 203
 204                 if (svc->srv_flags & SVC_LIST) {
 205                         struct ptlrpc_request *request;
 206                         svc->srv_flags = SVC_RUNNING;
 207
 208                         request = list_entry(svc->srv_reqs.next,
 209                                              struct ptlrpc_request,
 210                                              rq_list);
 211                         list_del(&request->rq_list);
 212                         spin_unlock(&svc->srv_lock);
 213                         rc = svc->srv_handler(obddev, svc, request);
 214                         continue;
 215                 }
 216                 CERROR("unknown break in service");
 217                 spin_unlock(&svc->srv_lock);
 218                 break;
 219         }
 220
 221         svc->srv_thread = NULL;
 222         svc->srv_flags = SVC_STOPPED;
 223         wake_up(&svc->srv_ctl_waitq);
 224         CDEBUG(D_NET, "svc exiting process %d\n", current->pid);
 225         return 0;
 226 }
 227
 228 void ptlrpc_stop_thread(struct ptlrpc_service *svc)
 229 {
 230         svc->srv_flags = SVC_STOPPING;
 231
 232         wake_up(&svc->srv_waitq);
 233         wait_event_interruptible(svc->srv_ctl_waitq,
 234                                  (svc->srv_flags & SVC_STOPPED));
 235 }
 236
 237 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
 238                                 char *name)
 239 {
 240         struct ptlrpc_svc_data d;
 241         int rc;
 242         ENTRY;
 243
 244         d.dev = dev;
 245         d.svc = svc;
 246         d.name = name;
 247
 248         init_waitqueue_head(&svc->srv_waitq);
 249
 250         init_waitqueue_head(&svc->srv_ctl_waitq);
 251         rc = kernel_thread(ptlrpc_main, (void *) &d,
 252                            CLONE_VM | CLONE_FS | CLONE_FILES);
 253         if (rc < 0) {
 254                 CERROR("cannot start thread\n");
 255                 return -EINVAL;
 256         }
 257         wait_event(svc->srv_ctl_waitq, svc->srv_flags & SVC_RUNNING);
 258
 259         EXIT;
 260         return 0;
 261 }
 262
 263
 264 int rpc_register_service(struct ptlrpc_service *service, char *uuid)
 265 {
 266         struct lustre_peer peer;
 267         int rc, i;
 268
 269         rc = kportal_uuid_to_peer(uuid, &peer);
 270         if (rc != 0) {
 271                 CERROR("Invalid uuid \"%s\"\n", uuid);
 272                 return -EINVAL;
 273         }
 274
 275         service->srv_ring_length = RPC_RING_LENGTH;
 276         service->srv_md_active = 0;
 277
 278         service->srv_id.addr_kind = PTL_ADDR_GID;
 279         service->srv_id.gid = PTL_ID_ANY;
 280         service->srv_id.rid = PTL_ID_ANY;
 281
 282         rc = PtlEQAlloc(peer.peer_ni, 128, server_request_callback,
 283                         service, &(service->srv_eq_h));
 284
 285         if (rc != PTL_OK) {
 286                 CERROR("PtlEQAlloc failed: %d\n", rc);
 287                 return rc;
 288         }
 289
 290         CDEBUG(D_NET, "Starting service listening on portal %d\n",
 291                service->srv_req_portal);
 292
 293         /* Attach the leading ME on which we build the ring */
 294         rc = PtlMEAttach(peer.peer_ni, service->srv_req_portal,
 295                          service->srv_id, 0, ~0, PTL_RETAIN,
 296                          &(service->srv_me_h[0]));
 297
 298         if (rc != PTL_OK) {
 299                 CERROR("PtlMEAttach failed: %d\n", rc);
 300                 return rc;
 301         }
 302
 303         for (i = 0; i < service->srv_ring_length; i++) {
 304                 OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
 305
 306                 if (service->srv_buf[i] == NULL) {
 307                         CERROR("no memory\n");
 308                         return -ENOMEM;
 309                 }
 310
 311
 312                 /* Insert additional ME's to the ring */
 313                 if (i > 0) {
 314                         rc = PtlMEInsert(service->srv_me_h[i-1],
 315                                          service->srv_id, 0, ~0, PTL_RETAIN,
 316                                          PTL_INS_AFTER,&(service->srv_me_h[i]));
 317                         service->srv_me_tail = i;
 318
 319                         if (rc != PTL_OK) {
 320                                 CERROR("PtlMEInsert failed: %d\n", rc);
 321                                 return rc;
 322                         }
 323                 }
 324
 325                 service->srv_ref_count[i] = 0;
 326                 service->srv_md[i].start         = service->srv_buf[i];
 327                 service->srv_md[i].length        = service->srv_buf_size;
 328                 service->srv_md[i].max_offset    = service->srv_buf_size;
 329                 service->srv_md[i].threshold     = PTL_MD_THRESH_INF;
 330                 service->srv_md[i].options       = PTL_MD_OP_PUT;
 331                 service->srv_md[i].user_ptr      = service;
 332                 service->srv_md[i].eventq        = service->srv_eq_h;
 333                 service->srv_md[i].max_offset    = service->srv_buf_size;
 334
 335                 rc = PtlMDAttach(service->srv_me_h[i], service->srv_md[i],
 336                                  PTL_RETAIN, &(service->srv_md_h[i]));
 337
 338                 if (rc != PTL_OK) {
 339                         /* cleanup */
 340                         CERROR("PtlMDAttach failed: %d\n", rc);
 341                         return rc;
 342                 }
 343         }
 344
 345         return 0;
 346 }
 347
 348 int rpc_unregister_service(struct ptlrpc_service *service)
 349 {
 350         int rc, i;
 351
 352         for (i = 0; i < service->srv_ring_length; i++) {
 353                 rc = PtlMEUnlink(service->srv_me_h[i]);
 354                 if (rc)
 355                         CERROR("PtlMEUnlink failed: %d\n", rc);
 356
 357                 if (service->srv_buf[i] != NULL)
 358                         OBD_FREE(service->srv_buf[i], service->srv_buf_size);
 359                 service->srv_buf[i] = NULL;
 360         }
 361
 362         rc = PtlEQFree(service->srv_eq_h);
 363         if (rc)
 364                 CERROR("PtlEQFree failed: %d\n", rc);
 365
 366         return 0;
 367 }