lustre/ptlrpc/client.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 #define DEBUG_SUBSYSTEM S_RPC
  26
  27 #include <linux/lustre_ha.h>
  28
  29 void ptlrpc_init_client(struct recovd_obd *recovd,
  30                         int (*recover)(struct ptlrpc_client *recover),
  31                         int req_portal,
  32                         int rep_portal, struct ptlrpc_client *cl)
  33 {
  34         memset(cl, 0, sizeof(*cl));
  35         cl->cli_recovd = recovd;
  36         cl->cli_recover = recover;
  37         if (recovd)
  38                 recovd_cli_manage(recovd, cl);
  39         cl->cli_obd = NULL;
  40         cl->cli_request_portal = req_portal;
  41         cl->cli_reply_portal = rep_portal;
  42         INIT_LIST_HEAD(&cl->cli_delayed_head);
  43         INIT_LIST_HEAD(&cl->cli_sending_head);
  44         INIT_LIST_HEAD(&cl->cli_dying_head);
  45         spin_lock_init(&cl->cli_lock);
  46         sema_init(&cl->cli_rpc_sem, 32);
  47 }
  48
  49 __u8 *ptlrpc_req_to_uuid(struct ptlrpc_request *req)
  50 {
  51         return req->rq_connection->c_remote_uuid;
  52 }
  53
  54 struct ptlrpc_connection *ptlrpc_uuid_to_connection(char *uuid)
  55 {
  56         struct ptlrpc_connection *c;
  57         struct lustre_peer peer;
  58         int err;
  59
  60         err = kportal_uuid_to_peer(uuid, &peer);
  61         if (err != 0) {
  62                 CERROR("cannot find peer %s!\n", uuid);
  63                 return NULL;
  64         }
  65
  66         c = ptlrpc_get_connection(&peer);
  67         if (c)
  68                 c->c_epoch++;
  69
  70         return c;
  71 }
  72
  73 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn, char *uuid)
  74 {
  75         struct lustre_peer peer;
  76         int err;
  77
  78         err = kportal_uuid_to_peer(uuid, &peer);
  79         if (err != 0) {
  80                 CERROR("cannot find peer %s!\n", uuid);
  81                 return;
  82         }
  83
  84         memcpy(&conn->c_peer, &peer, sizeof(peer));
  85         return;
  86 }
  87
  88 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk(struct ptlrpc_connection *conn)
  89 {
  90         struct ptlrpc_bulk_desc *bulk;
  91
  92         OBD_ALLOC(bulk, sizeof(*bulk));
  93         if (bulk != NULL) {
  94                 bulk->b_connection = ptlrpc_connection_addref(conn);
  95                 init_waitqueue_head(&bulk->b_waitq);
  96         }
  97
  98         return bulk;
  99 }
 100
 101 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk)
 102 {
 103         ENTRY;
 104         if (bulk == NULL) {
 105                 EXIT;
 106                 return;
 107         }
 108
 109         ptlrpc_put_connection(bulk->b_connection);
 110
 111         OBD_FREE(bulk, sizeof(*bulk));
 112         EXIT;
 113 }
 114
 115 struct ptlrpc_request *ptlrpc_prep_req(struct ptlrpc_client *cl,
 116                                        struct ptlrpc_connection *conn,
 117                                        int opcode, int count, int *lengths,
 118                                        char **bufs)
 119 {
 120         struct ptlrpc_request *request;
 121         int rc;
 122         ENTRY;
 123
 124         OBD_ALLOC(request, sizeof(*request));
 125         if (!request) {
 126                 CERROR("request allocation out of memory\n");
 127                 RETURN(NULL);
 128         }
 129
 130         rc = lustre_pack_msg(count, lengths, bufs,
 131                              &request->rq_reqlen, &request->rq_reqmsg);
 132         if (rc) {
 133                 CERROR("cannot pack request %d\n", rc);
 134                 RETURN(NULL);
 135         }
 136
 137         request->rq_type = PTL_RPC_TYPE_REQUEST;
 138         request->rq_connection = ptlrpc_connection_addref(conn);
 139
 140         request->rq_reqmsg->conn = (__u64)(unsigned long)conn->c_remote_conn;
 141         request->rq_reqmsg->token = conn->c_remote_token;
 142         request->rq_reqmsg->opc = HTON__u32(opcode);
 143         request->rq_reqmsg->type = HTON__u32(PTL_RPC_MSG_REQUEST);
 144         INIT_LIST_HEAD(&request->rq_list);
 145
 146         /* this will be dec()d once in req_finished, once in free_committed */
 147         atomic_set(&request->rq_refcount, 2);
 148
 149         spin_lock(&conn->c_lock);
 150         request->rq_reqmsg->xid = HTON__u32(++conn->c_xid_out);
 151         request->rq_xid = conn->c_xid_out;
 152         spin_unlock(&conn->c_lock);
 153
 154         request->rq_client = cl;
 155
 156         RETURN(request);
 157 }
 158
 159 void ptlrpc_req_finished(struct ptlrpc_request *request)
 160 {
 161         if (request == NULL)
 162                 return;
 163
 164         if (request->rq_repmsg != NULL) {
 165                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 166                 request->rq_repmsg = NULL;
 167                 request->rq_reply_md.start = NULL;
 168         }
 169
 170         if (atomic_dec_and_test(&request->rq_refcount))
 171                 ptlrpc_free_req(request);
 172 }
 173
 174 void ptlrpc_free_req(struct ptlrpc_request *request)
 175 {
 176         if (request == NULL)
 177                 return;
 178
 179         if (request->rq_repmsg != NULL)
 180                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 181         if (request->rq_reqmsg != NULL)
 182                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 183
 184         if (request->rq_client) {
 185                 spin_lock(&request->rq_client->cli_lock);
 186                 list_del_init(&request->rq_list);
 187                 spin_unlock(&request->rq_client->cli_lock);
 188         }
 189
 190         ptlrpc_put_connection(request->rq_connection);
 191
 192         OBD_FREE(request, sizeof(*request));
 193 }
 194
 195 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 196 {
 197         int rc = 0;
 198
 199         if (req->rq_repmsg != NULL) {
 200                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 201                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 202                 GOTO(out, rc = 1);
 203         }
 204
 205         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 206                 CERROR("-- RESEND --\n");
 207                 GOTO(out, rc = 1);
 208         }
 209
 210         if (req->rq_flags & PTL_RPC_FL_RECOVERY) {
 211                 CERROR("-- RESTART --\n");
 212                 GOTO(out, rc = 1);
 213         }
 214
 215
 216         if (CURRENT_TIME - req->rq_time >= req->rq_timeout) {
 217                 CERROR("-- REQ TIMEOUT --\n");
 218                 /* clear the timeout */
 219                 req->rq_timeout = 0;
 220                 req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
 221                 req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 222                 if (req->rq_client && req->rq_client->cli_recovd)
 223                         recovd_cli_fail(req->rq_client);
 224                 if (req->rq_level < LUSTRE_CONN_FULL)
 225                         rc = -ETIMEDOUT;
 226                 else
 227                         rc = 0;
 228
 229                 GOTO(out, rc);
 230         }
 231
 232         if (req->rq_timeout) {
 233                 schedule_timeout(req->rq_timeout * HZ);
 234         }
 235
 236         if (sigismember(&(current->pending.signal), SIGKILL) ||
 237             sigismember(&(current->pending.signal), SIGTERM) ||
 238             sigismember(&(current->pending.signal), SIGINT)) {
 239                 req->rq_flags |= PTL_RPC_FL_INTR;
 240                 GOTO(out, rc = 1);
 241         }
 242
 243  out:
 244         return rc;
 245 }
 246
 247 int ptlrpc_check_status(struct ptlrpc_request *req, int err)
 248 {
 249         ENTRY;
 250
 251         if (err != 0) {
 252                 CERROR("err is %d\n", err);
 253                 RETURN(err);
 254         }
 255
 256         if (req == NULL) {
 257                 CERROR("req == NULL\n");
 258                 RETURN(-ENOMEM);
 259         }
 260
 261         if (req->rq_repmsg == NULL) {
 262                 CERROR("req->rq_repmsg == NULL\n");
 263                 RETURN(-ENOMEM);
 264         }
 265
 266         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 267                 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
 268                 RETURN(-EINVAL);
 269         }
 270
 271         if (req->rq_repmsg->status != 0) {
 272                 CERROR("req->rq_repmsg->status is %d\n",
 273                        req->rq_repmsg->status);
 274                 /* XXX: translate this error from net to host */
 275                 RETURN(req->rq_repmsg->status);
 276         }
 277
 278         RETURN(0);
 279 }
 280
 281 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 282 {
 283         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 284         request->rq_reqmsg = NULL;
 285         request->rq_reqlen = 0;
 286 }
 287
 288 /* Abort this request and cleanup any resources associated with it. */
 289 static int ptlrpc_abort(struct ptlrpc_request *request)
 290 {
 291         /* First remove the ME for the reply; in theory, this means
 292          * that we can tear down the buffer safely. */
 293         PtlMEUnlink(request->rq_reply_me_h);
 294         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 295         request->rq_repmsg = NULL;
 296         request->rq_replen = 0;
 297         return 0;
 298 }
 299
 300 /* caller must lock cli */
 301 void ptlrpc_free_committed(struct ptlrpc_client *cli)
 302 {
 303         struct list_head *tmp, *saved;
 304         struct ptlrpc_request *req;
 305
 306         list_for_each_safe(tmp, saved, &cli->cli_sending_head) {
 307                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 308
 309                 if ( (req->rq_flags & PTL_RPC_FL_REPLAY) ) {
 310                         CDEBUG(D_INFO, "Retaining request %Ld for replay\n",
 311                                req->rq_xid);
 312                         continue;
 313                 }
 314
 315                 /* not yet committed */
 316                 if (!req->rq_transno ||
 317                     req->rq_transno > cli->cli_last_committed)
 318                         break;
 319
 320                 CDEBUG(D_INFO, "Marking request %Ld as committed ("
 321                        "transno=%Lu, last_committed=%Lu\n",
 322                        req->rq_xid, req->rq_transno,
 323                        cli->cli_last_committed);
 324                 if (atomic_dec_and_test(&req->rq_refcount)) {
 325                         /* we do this to prevent free_req deadlock */
 326                         list_del_init(&req->rq_list);
 327                         req->rq_client = NULL;
 328                         ptlrpc_free_req(req);
 329                 } else {
 330                         list_del_init(&req->rq_list);
 331                         list_add(&req->rq_list, &cli->cli_dying_head);
 332                 }
 333         }
 334
 335         EXIT;
 336         return;
 337 }
 338
 339 void ptlrpc_cleanup_client(struct ptlrpc_client *cli)
 340 {
 341         struct list_head *tmp, *saved;
 342         struct ptlrpc_request *req;
 343         ENTRY;
 344
 345         spin_lock(&cli->cli_lock);
 346         list_for_each_safe(tmp, saved, &cli->cli_sending_head) {
 347                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 348                 CDEBUG(D_INFO, "Cleaning req %p from sending list.\n", req);
 349                 list_del_init(&req->rq_list);
 350                 req->rq_client = NULL;
 351                 ptlrpc_free_req(req);
 352         }
 353         list_for_each_safe(tmp, saved, &cli->cli_dying_head) {
 354                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 355                 CERROR("Request %p is on the dying list at cleanup!\n", req);
 356                 list_del_init(&req->rq_list);
 357                 req->rq_client = NULL;
 358                 ptlrpc_free_req(req);
 359         }
 360         spin_unlock(&cli->cli_lock);
 361
 362         EXIT;
 363         return;
 364 }
 365
 366 void ptlrpc_continue_req(struct ptlrpc_request *req)
 367 {
 368         ENTRY;
 369         CDEBUG(D_INODE, "continue delayed request %Ld opc %d\n",
 370                req->rq_xid, req->rq_reqmsg->opc);
 371         wake_up_interruptible(&req->rq_wait_for_rep);
 372         EXIT;
 373 }
 374
 375 void ptlrpc_resend_req(struct ptlrpc_request *req)
 376 {
 377         ENTRY;
 378         CDEBUG(D_INODE, "resend request %Ld, opc %d\n",
 379                req->rq_xid, req->rq_reqmsg->opc);
 380         req->rq_status = -EAGAIN;
 381         req->rq_level = LUSTRE_CONN_RECOVD;
 382         req->rq_flags |= PTL_RPC_FL_RESEND;
 383         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 384         wake_up_interruptible(&req->rq_wait_for_rep);
 385         EXIT;
 386 }
 387
 388 void ptlrpc_restart_req(struct ptlrpc_request *req)
 389 {
 390         ENTRY;
 391         CDEBUG(D_INODE, "restart completed request %Ld, opc %d\n",
 392                req->rq_xid, req->rq_reqmsg->opc);
 393         req->rq_status = -ERESTARTSYS;
 394         req->rq_flags |= PTL_RPC_FL_RECOVERY;
 395         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 396         wake_up_interruptible(&req->rq_wait_for_rep);
 397         EXIT;
 398 }
 399
 400 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 401 {
 402         int rc = 0;
 403         struct ptlrpc_client *cli = req->rq_client;
 404         ENTRY;
 405
 406         init_waitqueue_head(&req->rq_wait_for_rep);
 407         CERROR("subsys: %s req %Ld opc %d level %d, conn level %d\n",
 408                cli->cli_name, req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 409                req->rq_connection->c_level);
 410
 411         /* XXX probably both an import and connection level are needed */
 412         if (req->rq_level > req->rq_connection->c_level) {
 413                 CERROR("process %d waiting for recovery\n", current->pid);
 414                 spin_lock(&cli->cli_lock);
 415                 list_del_init(&req->rq_list);
 416                 list_add(&req->rq_list, cli->cli_delayed_head.prev);
 417                 spin_unlock(&cli->cli_lock);
 418                 wait_event_interruptible
 419                         (req->rq_wait_for_rep,
 420                          req->rq_level <= req->rq_connection->c_level);
 421                 spin_lock(&cli->cli_lock);
 422                 list_del_init(&req->rq_list);
 423                 spin_unlock(&cli->cli_lock);
 424                 CERROR("process %d resumed\n", current->pid);
 425         }
 426  resend:
 427         req->rq_time = CURRENT_TIME;
 428         req->rq_timeout = 30;
 429         rc = ptl_send_rpc(req);
 430         if (rc) {
 431                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 432                 if ( rc > 0 )
 433                         rc = -rc;
 434                 ptlrpc_cleanup_request_buf(req);
 435                 up(&cli->cli_rpc_sem);
 436                 RETURN(-rc);
 437         }
 438
 439         spin_lock(&cli->cli_lock);
 440         list_del_init(&req->rq_list);
 441         list_add(&req->rq_list, cli->cli_sending_head.prev);
 442         spin_unlock(&cli->cli_lock);
 443
 444         CDEBUG(D_OTHER, "-- sleeping\n");
 445         wait_event_interruptible(req->rq_wait_for_rep,
 446                                  ptlrpc_check_reply(req));
 447         CDEBUG(D_OTHER, "-- done\n");
 448
 449         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 450                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 451                 goto resend;
 452         }
 453
 454         up(&cli->cli_rpc_sem);
 455         if (req->rq_flags & PTL_RPC_FL_INTR) {
 456                 /* Clean up the dangling reply buffers */
 457                 ptlrpc_abort(req);
 458                 GOTO(out, rc = -EINTR);
 459         }
 460
 461         if (! (req->rq_flags & PTL_RPC_FL_REPLIED)) {
 462                 GOTO(out, rc = req->rq_status);
 463         }
 464
 465         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 466         if (rc) {
 467                 CERROR("unpack_rep failed: %d\n", rc);
 468                 GOTO(out, rc);
 469         }
 470         CDEBUG(D_NET, "got rep %d\n", req->rq_repmsg->xid);
 471         if (req->rq_repmsg->status == 0)
 472                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 473                        req->rq_replen, req->rq_repmsg->status);
 474
 475         spin_lock(&cli->cli_lock);
 476         cli->cli_last_rcvd = req->rq_repmsg->last_rcvd;
 477         cli->cli_last_committed = req->rq_repmsg->last_committed;
 478         ptlrpc_free_committed(cli);
 479         spin_unlock(&cli->cli_lock);
 480
 481         EXIT;
 482  out:
 483         return rc;
 484 }
 485
 486 int ptlrpc_replay_req(struct ptlrpc_request *req)
 487 {
 488         int rc = 0;
 489         struct ptlrpc_client *cli = req->rq_client;
 490         ENTRY;
 491
 492         init_waitqueue_head(&req->rq_wait_for_rep);
 493         CERROR("req %Ld opc %d level %d, conn level %d\n",
 494                req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 495                req->rq_connection->c_level);
 496
 497         req->rq_time = CURRENT_TIME;
 498         req->rq_timeout = 3;
 499         rc = ptl_send_rpc(req);
 500         if (rc) {
 501                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 502                 ptlrpc_cleanup_request_buf(req);
 503                 up(&cli->cli_rpc_sem);
 504                 RETURN(-rc);
 505         }
 506
 507         CDEBUG(D_OTHER, "-- sleeping\n");
 508         wait_event_interruptible(req->rq_wait_for_rep,
 509                                  ptlrpc_check_reply(req));
 510         CDEBUG(D_OTHER, "-- done\n");
 511
 512         up(&cli->cli_rpc_sem);
 513
 514         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 515                 CERROR("Unknown reason for wakeup\n");
 516                 /* XXX Phil - I end up here when I kill obdctl */
 517                 ptlrpc_abort(req);
 518                 GOTO(out, rc = -EINTR);
 519         }
 520
 521         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 522         if (rc) {
 523                 CERROR("unpack_rep failed: %d\n", rc);
 524                 GOTO(out, rc);
 525         }
 526
 527         CDEBUG(D_NET, "got rep %d\n", req->rq_repmsg->xid);
 528         if (req->rq_repmsg->status == 0)
 529                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 530                        req->rq_replen, req->rq_repmsg->status);
 531         else {
 532                 CERROR("recovery failed: ");
 533                 CERROR("req %Ld opc %d level %d, conn level %d\n",
 534                        req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 535                        req->rq_connection->c_level);
 536                 LBUG();
 537         }
 538
 539  out:
 540         RETURN(rc);
 541 }