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                 INIT_LIST_HEAD(&bulk->b_page_list);
  97         }
  98
  99         return bulk;
 100 }
 101
 102 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
 103 {
 104         struct ptlrpc_bulk_page *page;
 105
 106         OBD_ALLOC(page, sizeof(*page));
 107         if (page != NULL) {
 108                 page->b_desc = desc;
 109                 ptl_set_inv_handle(&page->b_md_h);
 110                 ptl_set_inv_handle(&page->b_me_h);
 111                 list_add(&page->b_link, &desc->b_page_list);
 112                 desc->b_page_count++;
 113         }
 114         return page;
 115 }
 116
 117 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk)
 118 {
 119         struct list_head *tmp, *next;
 120         ENTRY;
 121         if (bulk == NULL) {
 122                 EXIT;
 123                 return;
 124         }
 125
 126         list_for_each_safe(tmp, next, &bulk->b_page_list) {
 127                 struct ptlrpc_bulk_page *page;
 128                 page = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 129                 ptlrpc_free_bulk_page(page);
 130         }
 131
 132         ptlrpc_put_connection(bulk->b_connection);
 133
 134         OBD_FREE(bulk, sizeof(*bulk));
 135         EXIT;
 136 }
 137
 138 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *page)
 139 {
 140         ENTRY;
 141         if (page == NULL) {
 142                 EXIT;
 143                 return;
 144         }
 145
 146         list_del(&page->b_link);
 147         page->b_desc->b_page_count--;
 148         OBD_FREE(page, sizeof(*page));
 149         EXIT;
 150 }
 151
 152 struct ptlrpc_request *ptlrpc_prep_req(struct ptlrpc_client *cl,
 153                                        struct ptlrpc_connection *conn,
 154                                        int opcode, int count, int *lengths,
 155                                        char **bufs)
 156 {
 157         struct ptlrpc_request *request;
 158         int rc;
 159         ENTRY;
 160
 161         OBD_ALLOC(request, sizeof(*request));
 162         if (!request) {
 163                 CERROR("request allocation out of memory\n");
 164                 RETURN(NULL);
 165         }
 166
 167         rc = lustre_pack_msg(count, lengths, bufs,
 168                              &request->rq_reqlen, &request->rq_reqmsg);
 169         if (rc) {
 170                 CERROR("cannot pack request %d\n", rc);
 171                 RETURN(NULL);
 172         }
 173
 174         request->rq_type = PTL_RPC_TYPE_REQUEST;
 175         request->rq_connection = ptlrpc_connection_addref(conn);
 176
 177         request->rq_reqmsg->conn = (__u64)(unsigned long)conn->c_remote_conn;
 178         request->rq_reqmsg->token = conn->c_remote_token;
 179         request->rq_reqmsg->opc = HTON__u32(opcode);
 180         request->rq_reqmsg->type = HTON__u32(PTL_RPC_MSG_REQUEST);
 181         INIT_LIST_HEAD(&request->rq_list);
 182
 183         /* this will be dec()d once in req_finished, once in free_committed */
 184         atomic_set(&request->rq_refcount, 2);
 185
 186         spin_lock(&conn->c_lock);
 187         request->rq_reqmsg->xid = HTON__u32(++conn->c_xid_out);
 188         request->rq_xid = conn->c_xid_out;
 189         spin_unlock(&conn->c_lock);
 190
 191         request->rq_client = cl;
 192
 193         RETURN(request);
 194 }
 195
 196 void ptlrpc_req_finished(struct ptlrpc_request *request)
 197 {
 198         if (request == NULL)
 199                 return;
 200
 201         if (request->rq_repmsg != NULL) {
 202                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 203                 request->rq_repmsg = NULL;
 204                 request->rq_reply_md.start = NULL;
 205         }
 206
 207         if (atomic_dec_and_test(&request->rq_refcount))
 208                 ptlrpc_free_req(request);
 209 }
 210
 211 void ptlrpc_free_req(struct ptlrpc_request *request)
 212 {
 213         if (request == NULL)
 214                 return;
 215
 216         if (request->rq_repmsg != NULL)
 217                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 218         if (request->rq_reqmsg != NULL)
 219                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 220
 221         if (request->rq_client) {
 222                 spin_lock(&request->rq_client->cli_lock);
 223                 list_del_init(&request->rq_list);
 224                 spin_unlock(&request->rq_client->cli_lock);
 225         }
 226
 227         ptlrpc_put_connection(request->rq_connection);
 228
 229         OBD_FREE(request, sizeof(*request));
 230 }
 231
 232 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 233 {
 234         int rc = 0;
 235
 236         if (req->rq_repmsg != NULL) {
 237                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 238                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 239                 GOTO(out, rc = 1);
 240         }
 241
 242         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 243                 CERROR("-- RESEND --\n");
 244                 GOTO(out, rc = 1);
 245         }
 246
 247         if (req->rq_flags & PTL_RPC_FL_RECOVERY) {
 248                 CERROR("-- RESTART --\n");
 249                 GOTO(out, rc = 1);
 250         }
 251
 252
 253         if (CURRENT_TIME - req->rq_time >= req->rq_timeout) {
 254                 CERROR("-- REQ TIMEOUT --\n");
 255                 /* clear the timeout */
 256                 req->rq_timeout = 0;
 257                 req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
 258                 req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 259                 if (req->rq_client && req->rq_client->cli_recovd)
 260                         recovd_cli_fail(req->rq_client);
 261                 if (req->rq_level < LUSTRE_CONN_FULL)
 262                         rc = 1;
 263                 else
 264                         rc = 0;
 265                 GOTO(out, rc);
 266         }
 267
 268         if (req->rq_timeout) {
 269                 schedule_timeout(req->rq_timeout * HZ);
 270         }
 271
 272         if (sigismember(&(current->pending.signal), SIGKILL) ||
 273             sigismember(&(current->pending.signal), SIGTERM) ||
 274             sigismember(&(current->pending.signal), SIGINT)) {
 275                 req->rq_flags |= PTL_RPC_FL_INTR;
 276                 GOTO(out, rc = 1);
 277         }
 278
 279  out:
 280         return rc;
 281 }
 282
 283 int ptlrpc_check_status(struct ptlrpc_request *req, int err)
 284 {
 285         ENTRY;
 286
 287         if (err != 0) {
 288                 CERROR("err is %d\n", err);
 289                 RETURN(err);
 290         }
 291
 292         if (req == NULL) {
 293                 CERROR("req == NULL\n");
 294                 RETURN(-ENOMEM);
 295         }
 296
 297         if (req->rq_repmsg == NULL) {
 298                 CERROR("req->rq_repmsg == NULL\n");
 299                 RETURN(-ENOMEM);
 300         }
 301
 302         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 303                 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
 304                 RETURN(-EINVAL);
 305         }
 306
 307         if (req->rq_repmsg->status != 0) {
 308                 CERROR("req->rq_repmsg->status is %d\n",
 309                        req->rq_repmsg->status);
 310                 /* XXX: translate this error from net to host */
 311                 RETURN(req->rq_repmsg->status);
 312         }
 313
 314         RETURN(0);
 315 }
 316
 317 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 318 {
 319         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 320         request->rq_reqmsg = NULL;
 321         request->rq_reqlen = 0;
 322 }
 323
 324 /* Abort this request and cleanup any resources associated with it. */
 325 static int ptlrpc_abort(struct ptlrpc_request *request)
 326 {
 327         /* First remove the ME for the reply; in theory, this means
 328          * that we can tear down the buffer safely. */
 329         PtlMEUnlink(request->rq_reply_me_h);
 330         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 331         request->rq_repmsg = NULL;
 332         request->rq_replen = 0;
 333         return 0;
 334 }
 335
 336 /* caller must lock cli */
 337 void ptlrpc_free_committed(struct ptlrpc_client *cli)
 338 {
 339         struct list_head *tmp, *saved;
 340         struct ptlrpc_request *req;
 341
 342         list_for_each_safe(tmp, saved, &cli->cli_sending_head) {
 343                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 344
 345                 if ( (req->rq_flags & PTL_RPC_FL_REPLAY) ) {
 346                         CDEBUG(D_INFO, "Retaining request %Ld for replay\n",
 347                                req->rq_xid);
 348                         continue;
 349                 }
 350
 351                 /* not yet committed */
 352                 if (!req->rq_transno ||
 353                     req->rq_transno > cli->cli_last_committed)
 354                         break;
 355
 356                 CDEBUG(D_INFO, "Marking request %Ld as committed ("
 357                        "transno=%Lu, last_committed=%Lu\n",
 358                        req->rq_xid, req->rq_transno,
 359                        cli->cli_last_committed);
 360                 if (atomic_dec_and_test(&req->rq_refcount)) {
 361                         /* we do this to prevent free_req deadlock */
 362                         list_del_init(&req->rq_list);
 363                         req->rq_client = NULL;
 364                         ptlrpc_free_req(req);
 365                 } else {
 366                         list_del_init(&req->rq_list);
 367                         list_add(&req->rq_list, &cli->cli_dying_head);
 368                 }
 369         }
 370
 371         EXIT;
 372         return;
 373 }
 374
 375 void ptlrpc_cleanup_client(struct ptlrpc_client *cli)
 376 {
 377         struct list_head *tmp, *saved;
 378         struct ptlrpc_request *req;
 379         ENTRY;
 380
 381         spin_lock(&cli->cli_lock);
 382         list_for_each_safe(tmp, saved, &cli->cli_sending_head) {
 383                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 384                 CDEBUG(D_INFO, "Cleaning req %p from sending list.\n", req);
 385                 list_del_init(&req->rq_list);
 386                 req->rq_client = NULL;
 387                 ptlrpc_free_req(req);
 388         }
 389         list_for_each_safe(tmp, saved, &cli->cli_dying_head) {
 390                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 391                 CERROR("Request %p is on the dying list at cleanup!\n", req);
 392                 list_del_init(&req->rq_list);
 393                 req->rq_client = NULL;
 394                 ptlrpc_free_req(req);
 395         }
 396         spin_unlock(&cli->cli_lock);
 397
 398         EXIT;
 399         return;
 400 }
 401
 402 void ptlrpc_continue_req(struct ptlrpc_request *req)
 403 {
 404         ENTRY;
 405         CDEBUG(D_INODE, "continue delayed request %Ld opc %d\n",
 406                req->rq_xid, req->rq_reqmsg->opc);
 407         wake_up_interruptible(&req->rq_wait_for_rep);
 408         EXIT;
 409 }
 410
 411 void ptlrpc_resend_req(struct ptlrpc_request *req)
 412 {
 413         ENTRY;
 414         CDEBUG(D_INODE, "resend request %Ld, opc %d\n",
 415                req->rq_xid, req->rq_reqmsg->opc);
 416         req->rq_status = -EAGAIN;
 417         req->rq_level = LUSTRE_CONN_RECOVD;
 418         req->rq_flags |= PTL_RPC_FL_RESEND;
 419         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 420         wake_up_interruptible(&req->rq_wait_for_rep);
 421         EXIT;
 422 }
 423
 424 void ptlrpc_restart_req(struct ptlrpc_request *req)
 425 {
 426         ENTRY;
 427         CDEBUG(D_INODE, "restart completed request %Ld, opc %d\n",
 428                req->rq_xid, req->rq_reqmsg->opc);
 429         req->rq_status = -ERESTARTSYS;
 430         req->rq_flags |= PTL_RPC_FL_RECOVERY;
 431         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 432         wake_up_interruptible(&req->rq_wait_for_rep);
 433         EXIT;
 434 }
 435
 436 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 437 {
 438         int rc = 0;
 439         struct ptlrpc_client *cli = req->rq_client;
 440         ENTRY;
 441
 442         init_waitqueue_head(&req->rq_wait_for_rep);
 443         CDEBUG(D_NET, "subsys: %s req %Ld opc %d level %d, conn level %d\n",
 444                cli->cli_name, req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 445                req->rq_connection->c_level);
 446
 447         /* XXX probably both an import and connection level are needed */
 448         if (req->rq_level > req->rq_connection->c_level) {
 449                 CERROR("process %d waiting for recovery\n", current->pid);
 450                 spin_lock(&cli->cli_lock);
 451                 list_del_init(&req->rq_list);
 452                 list_add(&req->rq_list, cli->cli_delayed_head.prev);
 453                 spin_unlock(&cli->cli_lock);
 454                 wait_event_interruptible
 455                         (req->rq_wait_for_rep,
 456                          req->rq_level <= req->rq_connection->c_level);
 457                 spin_lock(&cli->cli_lock);
 458                 list_del_init(&req->rq_list);
 459                 spin_unlock(&cli->cli_lock);
 460                 CERROR("process %d resumed\n", current->pid);
 461         }
 462  resend:
 463         req->rq_time = CURRENT_TIME;
 464         req->rq_timeout = 30;
 465         rc = ptl_send_rpc(req);
 466         if (rc) {
 467                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 468                 if ( rc > 0 )
 469                         rc = -rc;
 470                 ptlrpc_cleanup_request_buf(req);
 471                 up(&cli->cli_rpc_sem);
 472                 RETURN(-rc);
 473         }
 474
 475         spin_lock(&cli->cli_lock);
 476         list_del_init(&req->rq_list);
 477         list_add(&req->rq_list, cli->cli_sending_head.prev);
 478         spin_unlock(&cli->cli_lock);
 479
 480         CDEBUG(D_OTHER, "-- sleeping\n");
 481         wait_event_interruptible(req->rq_wait_for_rep,
 482                                  ptlrpc_check_reply(req));
 483         CDEBUG(D_OTHER, "-- done\n");
 484
 485         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 486                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 487                 goto resend;
 488         }
 489
 490         up(&cli->cli_rpc_sem);
 491         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 492                 GOTO(out, rc = -ETIMEDOUT);
 493
 494         if (req->rq_flags & PTL_RPC_FL_INTR) {
 495                 /* Clean up the dangling reply buffers */
 496                 ptlrpc_abort(req);
 497                 GOTO(out, rc = -EINTR);
 498         }
 499
 500         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 501                 GOTO(out, rc = req->rq_status);
 502
 503         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 504         if (rc) {
 505                 CERROR("unpack_rep failed: %d\n", rc);
 506                 GOTO(out, rc);
 507         }
 508         CDEBUG(D_NET, "got rep %d\n", req->rq_repmsg->xid);
 509         if (req->rq_repmsg->status == 0)
 510                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 511                        req->rq_replen, req->rq_repmsg->status);
 512
 513         spin_lock(&cli->cli_lock);
 514         cli->cli_last_rcvd = req->rq_repmsg->last_rcvd;
 515         cli->cli_last_committed = req->rq_repmsg->last_committed;
 516         ptlrpc_free_committed(cli);
 517         spin_unlock(&cli->cli_lock);
 518
 519         EXIT;
 520  out:
 521         return rc;
 522 }
 523
 524 int ptlrpc_replay_req(struct ptlrpc_request *req)
 525 {
 526         int rc = 0;
 527         struct ptlrpc_client *cli = req->rq_client;
 528         ENTRY;
 529
 530         init_waitqueue_head(&req->rq_wait_for_rep);
 531         CERROR("req %Ld opc %d level %d, conn level %d\n",
 532                req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 533                req->rq_connection->c_level);
 534
 535         req->rq_time = CURRENT_TIME;
 536         req->rq_timeout = 3;
 537         rc = ptl_send_rpc(req);
 538         if (rc) {
 539                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 540                 ptlrpc_cleanup_request_buf(req);
 541                 up(&cli->cli_rpc_sem);
 542                 RETURN(-rc);
 543         }
 544
 545         CDEBUG(D_OTHER, "-- sleeping\n");
 546         wait_event_interruptible(req->rq_wait_for_rep,
 547                                  ptlrpc_check_reply(req));
 548         CDEBUG(D_OTHER, "-- done\n");
 549
 550         up(&cli->cli_rpc_sem);
 551
 552         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 553                 CERROR("Unknown reason for wakeup\n");
 554                 /* XXX Phil - I end up here when I kill obdctl */
 555                 ptlrpc_abort(req);
 556                 GOTO(out, rc = -EINTR);
 557         }
 558
 559         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 560         if (rc) {
 561                 CERROR("unpack_rep failed: %d\n", rc);
 562                 GOTO(out, rc);
 563         }
 564
 565         CDEBUG(D_NET, "got rep %d\n", req->rq_repmsg->xid);
 566         if (req->rq_repmsg->status == 0)
 567                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 568                        req->rq_replen, req->rq_repmsg->status);
 569         else {
 570                 CERROR("recovery failed: ");
 571                 CERROR("req %Ld opc %d level %d, conn level %d\n",
 572                        req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 573                        req->rq_connection->c_level);
 574                 LBUG();
 575         }
 576
 577  out:
 578         RETURN(rc);
 579 }