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 DEBUG_SUBSYSTEM S_RPC
  24
  25 #include <linux/lustre_ha.h>
  26
  27 void ptlrpc_init_client(struct recovd_obd *recovd,
  28                         int (*recover)(struct ptlrpc_client *recover),
  29                         int req_portal,
  30                         int rep_portal, struct ptlrpc_client *cl)
  31 {
  32         memset(cl, 0, sizeof(*cl));
  33         cl->cli_recovd = recovd;
  34         cl->cli_recover = recover;
  35         if (recovd)
  36                 recovd_cli_manage(recovd, cl);
  37         cl->cli_obd = NULL;
  38         cl->cli_request_portal = req_portal;
  39         cl->cli_reply_portal = rep_portal;
  40         INIT_LIST_HEAD(&cl->cli_delayed_head);
  41         INIT_LIST_HEAD(&cl->cli_sending_head);
  42         INIT_LIST_HEAD(&cl->cli_dying_head);
  43         spin_lock_init(&cl->cli_lock);
  44         sema_init(&cl->cli_rpc_sem, 32);
  45 }
  46
  47 __u8 *ptlrpc_req_to_uuid(struct ptlrpc_request *req)
  48 {
  49         return req->rq_connection->c_remote_uuid;
  50 }
  51
  52 struct ptlrpc_connection *ptlrpc_uuid_to_connection(char *uuid)
  53 {
  54         struct ptlrpc_connection *c;
  55         struct lustre_peer peer;
  56         int err;
  57
  58         err = kportal_uuid_to_peer(uuid, &peer);
  59         if (err != 0) {
  60                 CERROR("cannot find peer %s!\n", uuid);
  61                 return NULL;
  62         }
  63
  64         c = ptlrpc_get_connection(&peer);
  65         if (c)
  66                 c->c_epoch++;
  67
  68         return c;
  69 }
  70
  71 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn, char *uuid)
  72 {
  73         struct lustre_peer peer;
  74         int err;
  75
  76         err = kportal_uuid_to_peer(uuid, &peer);
  77         if (err != 0) {
  78                 CERROR("cannot find peer %s!\n", uuid);
  79                 return;
  80         }
  81
  82         memcpy(&conn->c_peer, &peer, sizeof(peer));
  83         return;
  84 }
  85
  86 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk(struct ptlrpc_connection *conn)
  87 {
  88         struct ptlrpc_bulk_desc *bulk;
  89
  90         OBD_ALLOC(bulk, sizeof(*bulk));
  91         if (bulk != NULL) {
  92                 bulk->b_connection = ptlrpc_connection_addref(conn);
  93                 init_waitqueue_head(&bulk->b_waitq);
  94                 INIT_LIST_HEAD(&bulk->b_page_list);
  95         }
  96
  97         return bulk;
  98 }
  99
 100 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
 101 {
 102         struct ptlrpc_bulk_page *bulk;
 103
 104         OBD_ALLOC(bulk, sizeof(*bulk));
 105         if (bulk != NULL) {
 106                 bulk->b_desc = desc;
 107                 ptl_set_inv_handle(&bulk->b_md_h);
 108                 ptl_set_inv_handle(&bulk->b_me_h);
 109                 list_add_tail(&bulk->b_link, &desc->b_page_list);
 110                 desc->b_page_count++;
 111         }
 112         return bulk;
 113 }
 114
 115 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *bulk)
 116 {
 117         struct list_head *tmp, *next;
 118         ENTRY;
 119         if (bulk == NULL) {
 120                 EXIT;
 121                 return;
 122         }
 123
 124         list_for_each_safe(tmp, next, &bulk->b_page_list) {
 125                 struct ptlrpc_bulk_page *bulk;
 126                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 127                 ptlrpc_free_bulk_page(bulk);
 128         }
 129
 130         ptlrpc_put_connection(bulk->b_connection);
 131
 132         OBD_FREE(bulk, sizeof(*bulk));
 133         EXIT;
 134 }
 135
 136 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
 137 {
 138         ENTRY;
 139         if (bulk == NULL) {
 140                 EXIT;
 141                 return;
 142         }
 143
 144         list_del(&bulk->b_link);
 145         bulk->b_desc->b_page_count--;
 146         OBD_FREE(bulk, sizeof(*bulk));
 147         EXIT;
 148 }
 149
 150 struct ptlrpc_request *ptlrpc_prep_req(struct ptlrpc_client *cl,
 151                                        struct ptlrpc_connection *conn,
 152                                        int opcode, int count, int *lengths,
 153                                        char **bufs)
 154 {
 155         struct ptlrpc_request *request;
 156         int rc;
 157         ENTRY;
 158
 159         OBD_ALLOC(request, sizeof(*request));
 160         if (!request) {
 161                 CERROR("request allocation out of memory\n");
 162                 RETURN(NULL);
 163         }
 164
 165         rc = lustre_pack_msg(count, lengths, bufs,
 166                              &request->rq_reqlen, &request->rq_reqmsg);
 167         if (rc) {
 168                 CERROR("cannot pack request %d\n", rc);
 169                 RETURN(NULL);
 170         }
 171
 172         request->rq_type = PTL_RPC_TYPE_REQUEST;
 173         request->rq_connection = ptlrpc_connection_addref(conn);
 174
 175         request->rq_reqmsg->conn = (__u64)(unsigned long)conn->c_remote_conn;
 176         request->rq_reqmsg->token = conn->c_remote_token;
 177         request->rq_reqmsg->opc = HTON__u32(opcode);
 178         request->rq_reqmsg->type = HTON__u32(PTL_RPC_MSG_REQUEST);
 179         INIT_LIST_HEAD(&request->rq_list);
 180
 181         /* this will be dec()d once in req_finished, once in free_committed */
 182         atomic_set(&request->rq_refcount, 2);
 183
 184         spin_lock(&conn->c_lock);
 185         request->rq_reqmsg->xid = HTON__u32(++conn->c_xid_out);
 186         request->rq_xid = conn->c_xid_out;
 187         spin_unlock(&conn->c_lock);
 188
 189         request->rq_client = cl;
 190
 191         RETURN(request);
 192 }
 193
 194 void ptlrpc_req_finished(struct ptlrpc_request *request)
 195 {
 196         if (request == NULL)
 197                 return;
 198
 199         if (request->rq_repmsg != NULL) {
 200                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 201                 request->rq_repmsg = NULL;
 202                 request->rq_reply_md.start = NULL;
 203         }
 204
 205         if (atomic_dec_and_test(&request->rq_refcount))
 206                 ptlrpc_free_req(request);
 207 }
 208
 209 void ptlrpc_free_req(struct ptlrpc_request *request)
 210 {
 211         if (request == NULL)
 212                 return;
 213
 214         if (request->rq_repmsg != NULL)
 215                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 216         if (request->rq_reqmsg != NULL)
 217                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 218
 219         if (request->rq_client) {
 220                 spin_lock(&request->rq_client->cli_lock);
 221                 list_del_init(&request->rq_list);
 222                 spin_unlock(&request->rq_client->cli_lock);
 223         }
 224
 225         ptlrpc_put_connection(request->rq_connection);
 226
 227         OBD_FREE(request, sizeof(*request));
 228 }
 229
 230 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 231 {
 232         int rc = 0;
 233
 234         if (req->rq_repmsg != NULL) {
 235                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 236                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 237                 GOTO(out, rc = 1);
 238         }
 239
 240         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 241                 CERROR("-- RESEND --\n");
 242                 GOTO(out, rc = 1);
 243         }
 244
 245         if (req->rq_flags & PTL_RPC_FL_RECOVERY) {
 246                 CERROR("-- RESTART --\n");
 247                 GOTO(out, rc = 1);
 248         }
 249
 250
 251         if (CURRENT_TIME - req->rq_time >= req->rq_timeout) {
 252                 CERROR("-- REQ TIMEOUT --\n");
 253                 /* clear the timeout */
 254                 req->rq_timeout = 0;
 255                 req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
 256                 req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 257                 if (req->rq_client && req->rq_client->cli_recovd)
 258                         recovd_cli_fail(req->rq_client);
 259                 if (req->rq_level < LUSTRE_CONN_FULL)
 260                         rc = 1;
 261                 else
 262                         rc = 0;
 263                 GOTO(out, rc);
 264         }
 265
 266         if (req->rq_timeout) {
 267                 schedule_timeout(req->rq_timeout * HZ);
 268         }
 269
 270         if (sigismember(&(current->pending.signal), SIGKILL) ||
 271             sigismember(&(current->pending.signal), SIGTERM) ||
 272             sigismember(&(current->pending.signal), SIGINT)) {
 273                 req->rq_flags |= PTL_RPC_FL_INTR;
 274                 GOTO(out, rc = 1);
 275         }
 276
 277  out:
 278         return rc;
 279 }
 280
 281 int ptlrpc_check_status(struct ptlrpc_request *req, int err)
 282 {
 283         ENTRY;
 284
 285         if (err != 0) {
 286                 CERROR("err is %d\n", err);
 287                 RETURN(err);
 288         }
 289
 290         if (req == NULL) {
 291                 CERROR("req == NULL\n");
 292                 RETURN(-ENOMEM);
 293         }
 294
 295         if (req->rq_repmsg == NULL) {
 296                 CERROR("req->rq_repmsg == NULL\n");
 297                 RETURN(-ENOMEM);
 298         }
 299
 300         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 301                 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
 302                 RETURN(-EINVAL);
 303         }
 304
 305         if (req->rq_repmsg->status != 0) {
 306                 CERROR("req->rq_repmsg->status is %d\n",
 307                        req->rq_repmsg->status);
 308                 /* XXX: translate this error from net to host */
 309                 RETURN(req->rq_repmsg->status);
 310         }
 311
 312         RETURN(0);
 313 }
 314
 315 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 316 {
 317         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 318         request->rq_reqmsg = NULL;
 319         request->rq_reqlen = 0;
 320 }
 321
 322 /* Abort this request and cleanup any resources associated with it. */
 323 static int ptlrpc_abort(struct ptlrpc_request *request)
 324 {
 325         /* First remove the ME for the reply; in theory, this means
 326          * that we can tear down the buffer safely. */
 327         PtlMEUnlink(request->rq_reply_me_h);
 328         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 329         request->rq_repmsg = NULL;
 330         request->rq_replen = 0;
 331         return 0;
 332 }
 333
 334 /* caller must lock cli */
 335 void ptlrpc_free_committed(struct ptlrpc_client *cli)
 336 {
 337         struct list_head *tmp, *saved;
 338         struct ptlrpc_request *req;
 339
 340         list_for_each_safe(tmp, saved, &cli->cli_sending_head) {
 341                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 342
 343                 if ( (req->rq_flags & PTL_RPC_FL_REPLAY) ) {
 344                         CDEBUG(D_INFO, "Retaining request %Ld for replay\n",
 345                                req->rq_xid);
 346                         continue;
 347                 }
 348
 349                 /* not yet committed */
 350                 if (!req->rq_transno ||
 351                     req->rq_transno > cli->cli_last_committed)
 352                         break;
 353
 354                 CDEBUG(D_INFO, "Marking request %Ld as committed ("
 355                        "transno=%Lu, last_committed=%Lu\n",
 356                        req->rq_xid, req->rq_transno,
 357                        cli->cli_last_committed);
 358                 if (atomic_dec_and_test(&req->rq_refcount)) {
 359                         /* we do this to prevent free_req deadlock */
 360                         list_del_init(&req->rq_list);
 361                         req->rq_client = NULL;
 362                         ptlrpc_free_req(req);
 363                 } else {
 364                         list_del_init(&req->rq_list);
 365                         list_add(&req->rq_list, &cli->cli_dying_head);
 366                 }
 367         }
 368
 369         EXIT;
 370         return;
 371 }
 372
 373 void ptlrpc_cleanup_client(struct ptlrpc_client *cli)
 374 {
 375         struct list_head *tmp, *saved;
 376         struct ptlrpc_request *req;
 377         ENTRY;
 378
 379         spin_lock(&cli->cli_lock);
 380         list_for_each_safe(tmp, saved, &cli->cli_sending_head) {
 381                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 382                 CDEBUG(D_INFO, "Cleaning req %p from sending list.\n", req);
 383                 list_del_init(&req->rq_list);
 384                 req->rq_client = NULL;
 385                 ptlrpc_free_req(req);
 386         }
 387         list_for_each_safe(tmp, saved, &cli->cli_dying_head) {
 388                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 389                 CERROR("Request %p is on the dying list at cleanup!\n", req);
 390                 list_del_init(&req->rq_list);
 391                 req->rq_client = NULL;
 392                 ptlrpc_free_req(req);
 393         }
 394         spin_unlock(&cli->cli_lock);
 395
 396         EXIT;
 397         return;
 398 }
 399
 400 void ptlrpc_continue_req(struct ptlrpc_request *req)
 401 {
 402         ENTRY;
 403         CDEBUG(D_INODE, "continue delayed request %Ld opc %d\n",
 404                req->rq_xid, req->rq_reqmsg->opc);
 405         wake_up_interruptible(&req->rq_wait_for_rep);
 406         EXIT;
 407 }
 408
 409 void ptlrpc_resend_req(struct ptlrpc_request *req)
 410 {
 411         ENTRY;
 412         CDEBUG(D_INODE, "resend request %Ld, opc %d\n",
 413                req->rq_xid, req->rq_reqmsg->opc);
 414         req->rq_status = -EAGAIN;
 415         req->rq_level = LUSTRE_CONN_RECOVD;
 416         req->rq_flags |= PTL_RPC_FL_RESEND;
 417         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 418         wake_up_interruptible(&req->rq_wait_for_rep);
 419         EXIT;
 420 }
 421
 422 void ptlrpc_restart_req(struct ptlrpc_request *req)
 423 {
 424         ENTRY;
 425         CDEBUG(D_INODE, "restart completed request %Ld, opc %d\n",
 426                req->rq_xid, req->rq_reqmsg->opc);
 427         req->rq_status = -ERESTARTSYS;
 428         req->rq_flags |= PTL_RPC_FL_RECOVERY;
 429         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 430         wake_up_interruptible(&req->rq_wait_for_rep);
 431         EXIT;
 432 }
 433
 434 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 435 {
 436         int rc = 0;
 437         struct ptlrpc_client *cli = req->rq_client;
 438         ENTRY;
 439
 440         init_waitqueue_head(&req->rq_wait_for_rep);
 441         CDEBUG(D_NET, "subsys: %s req %Ld opc %d level %d, conn level %d\n",
 442                cli->cli_name, req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 443                req->rq_connection->c_level);
 444
 445         /* XXX probably both an import and connection level are needed */
 446         if (req->rq_level > req->rq_connection->c_level) {
 447                 CERROR("process %d waiting for recovery\n", current->pid);
 448                 spin_lock(&cli->cli_lock);
 449                 list_del_init(&req->rq_list);
 450                 list_add(&req->rq_list, cli->cli_delayed_head.prev);
 451                 spin_unlock(&cli->cli_lock);
 452                 wait_event_interruptible
 453                         (req->rq_wait_for_rep,
 454                          req->rq_level <= req->rq_connection->c_level);
 455                 spin_lock(&cli->cli_lock);
 456                 list_del_init(&req->rq_list);
 457                 spin_unlock(&cli->cli_lock);
 458                 CERROR("process %d resumed\n", current->pid);
 459         }
 460  resend:
 461         req->rq_time = CURRENT_TIME;
 462         req->rq_timeout = 30;
 463         rc = ptl_send_rpc(req);
 464         if (rc) {
 465                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 466                 if ( rc > 0 )
 467                         rc = -rc;
 468                 ptlrpc_cleanup_request_buf(req);
 469                 up(&cli->cli_rpc_sem);
 470                 RETURN(-rc);
 471         }
 472
 473         spin_lock(&cli->cli_lock);
 474         list_del_init(&req->rq_list);
 475         list_add(&req->rq_list, cli->cli_sending_head.prev);
 476         spin_unlock(&cli->cli_lock);
 477
 478         CDEBUG(D_OTHER, "-- sleeping\n");
 479         wait_event_interruptible(req->rq_wait_for_rep,
 480                                  ptlrpc_check_reply(req));
 481         CDEBUG(D_OTHER, "-- done\n");
 482
 483         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 484                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 485                 goto resend;
 486         }
 487
 488         up(&cli->cli_rpc_sem);
 489         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 490                 GOTO(out, rc = -ETIMEDOUT);
 491
 492         if (req->rq_flags & PTL_RPC_FL_INTR) {
 493                 /* Clean up the dangling reply buffers */
 494                 ptlrpc_abort(req);
 495                 GOTO(out, rc = -EINTR);
 496         }
 497
 498         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 499                 GOTO(out, rc = req->rq_status);
 500
 501         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 502         if (rc) {
 503                 CERROR("unpack_rep failed: %d\n", rc);
 504                 GOTO(out, rc);
 505         }
 506         CDEBUG(D_NET, "got rep %d\n", req->rq_repmsg->xid);
 507         if (req->rq_repmsg->status == 0)
 508                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 509                        req->rq_replen, req->rq_repmsg->status);
 510
 511         spin_lock(&cli->cli_lock);
 512         cli->cli_last_rcvd = req->rq_repmsg->last_rcvd;
 513         cli->cli_last_committed = req->rq_repmsg->last_committed;
 514         ptlrpc_free_committed(cli);
 515         spin_unlock(&cli->cli_lock);
 516
 517         EXIT;
 518  out:
 519         return rc;
 520 }
 521
 522 int ptlrpc_replay_req(struct ptlrpc_request *req)
 523 {
 524         int rc = 0;
 525         struct ptlrpc_client *cli = req->rq_client;
 526         ENTRY;
 527
 528         init_waitqueue_head(&req->rq_wait_for_rep);
 529         CERROR("req %Ld opc %d level %d, conn level %d\n",
 530                req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 531                req->rq_connection->c_level);
 532
 533         req->rq_time = CURRENT_TIME;
 534         req->rq_timeout = 3;
 535         rc = ptl_send_rpc(req);
 536         if (rc) {
 537                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 538                 ptlrpc_cleanup_request_buf(req);
 539                 up(&cli->cli_rpc_sem);
 540                 RETURN(-rc);
 541         }
 542
 543         CDEBUG(D_OTHER, "-- sleeping\n");
 544         wait_event_interruptible(req->rq_wait_for_rep,
 545                                  ptlrpc_check_reply(req));
 546         CDEBUG(D_OTHER, "-- done\n");
 547
 548         up(&cli->cli_rpc_sem);
 549
 550         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 551                 CERROR("Unknown reason for wakeup\n");
 552                 /* XXX Phil - I end up here when I kill obdctl */
 553                 ptlrpc_abort(req);
 554                 GOTO(out, rc = -EINTR);
 555         }
 556
 557         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 558         if (rc) {
 559                 CERROR("unpack_rep failed: %d\n", rc);
 560                 GOTO(out, rc);
 561         }
 562
 563         CDEBUG(D_NET, "got rep %d\n", req->rq_repmsg->xid);
 564         if (req->rq_repmsg->status == 0)
 565                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 566                        req->rq_replen, req->rq_repmsg->status);
 567         else {
 568                 CERROR("recovery failed: ");
 569                 CERROR("req %Ld opc %d level %d, conn level %d\n",
 570                        req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
 571                        req->rq_connection->c_level);
 572                 LBUG();
 573         }
 574
 575  out:
 576         RETURN(rc);
 577 }