lustre/ost/ost_handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author: Peter J. Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  *
  23  *  Storage Target Handling functions
  24  *  Lustre Object Server Module (OST)
  25  *
  26  *  This server is single threaded at present (but can easily be multi
  27  *  threaded). For testing and management it is treated as an
  28  *  obd_device, although it does not export a full OBD method table
  29  *  (the requests are coming in over the wire, so object target
  30  *  modules do not have a full method table.)
  31  */
  32
  33 #ifndef EXPORT_SYMTAB
  34 # define EXPORT_SYMTAB
  35 #endif
  36 #define DEBUG_SUBSYSTEM S_OST
  37
  38 #include <linux/module.h>
  39 #include <linux/obd_ost.h>
  40 #include <linux/lustre_net.h>
  41 #include <linux/lustre_dlm.h>
  42 #include <linux/lustre_export.h>
  43 #include <linux/init.h>
  44 #include <linux/lprocfs_status.h>
  45 #include <linux/lustre_commit_confd.h>
  46 #include <libcfs/list.h>
  47 #include <linux/lustre_sec.h>
  48
  49 void oti_init(struct obd_trans_info *oti, struct ptlrpc_request *req)
  50 {
  51         if (oti == NULL)
  52                 return;
  53         memset(oti, 0, sizeof *oti);
  54
  55         if (req->rq_repmsg && req->rq_reqmsg != 0)
  56                 oti->oti_transno = req->rq_repmsg->transno;
  57 }
  58
  59 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
  60 {
  61         struct oti_req_ack_lock *ack_lock;
  62         int i;
  63
  64         if (oti == NULL)
  65                 return;
  66
  67         if (req->rq_repmsg)
  68                 req->rq_repmsg->transno = oti->oti_transno;
  69
  70         /* XXX 4 == entries in oti_ack_locks??? */
  71         for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
  72                 if (!ack_lock->mode)
  73                         break;
  74                 /* XXX not even calling target_send_reply in some cases... */
  75                 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
  76         }
  77 }
  78
  79 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
  80                        struct obd_trans_info *oti)
  81 {
  82         struct ost_body *body, *repbody;
  83         int rc, size = sizeof(*body);
  84         ENTRY;
  85
  86         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
  87         if (body == NULL)
  88                 RETURN(-EFAULT);
  89
  90         rc = lustre_pack_reply(req, 1, &size, NULL);
  91         if (rc)
  92                 RETURN(rc);
  93
  94         if (body->oa.o_valid & OBD_MD_FLCOOKIE)
  95                 oti->oti_logcookies = obdo_logcookie(&body->oa);
  96         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
  97         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
  98         req->rq_status = obd_destroy(exp, &body->oa, NULL, oti);
  99         RETURN(0);
 100 }
 101
 102 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
 103 {
 104         struct ost_body *body, *repbody;
 105         int rc, size = sizeof(*body);
 106         ENTRY;
 107
 108         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 109         if (body == NULL)
 110                 RETURN(-EFAULT);
 111
 112         rc = lustre_pack_reply(req, 1, &size, NULL);
 113         if (rc)
 114                 RETURN(rc);
 115
 116         repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody));
 117         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 118         req->rq_status = obd_getattr(exp, &repbody->oa, NULL);
 119         RETURN(0);
 120 }
 121
 122 static int ost_statfs(struct ptlrpc_request *req)
 123 {
 124         struct obd_statfs *osfs;
 125         int rc, size = sizeof(*osfs);
 126         ENTRY;
 127
 128         rc = lustre_pack_reply(req, 1, &size, NULL);
 129         if (rc)
 130                 RETURN(rc);
 131
 132         osfs = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*osfs));
 133
 134         req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, jiffies-HZ);
 135         if (req->rq_status != 0)
 136                 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
 137
 138         RETURN(0);
 139 }
 140
 141 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
 142                       struct obd_trans_info *oti)
 143 {
 144         struct ost_body *body, *repbody;
 145         int rc, size = sizeof(*repbody);
 146         ENTRY;
 147
 148         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 149         if (body == NULL)
 150                 RETURN(-EFAULT);
 151
 152         rc = lustre_pack_reply(req, 1, &size, NULL);
 153         if (rc)
 154                 RETURN(rc);
 155
 156         repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody));
 157         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 158         oti->oti_logcookies = obdo_logcookie(&repbody->oa);
 159         req->rq_status = obd_create(exp, &repbody->oa, NULL, 0, NULL, oti);
 160         //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
 161         RETURN(0);
 162 }
 163
 164 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
 165                      struct obd_trans_info *oti)
 166 {
 167         struct ost_body *body, *repbody;
 168         int rc, size = sizeof(*repbody);
 169         ENTRY;
 170
 171         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 172         if (body == NULL)
 173                 RETURN(-EFAULT);
 174
 175         if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
 176             (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
 177                 RETURN(-EINVAL);
 178
 179         rc = lustre_pack_reply(req, 1, &size, NULL);
 180         if (rc)
 181                 RETURN(rc);
 182
 183         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 184         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 185         req->rq_status = obd_punch(exp, &repbody->oa, NULL, repbody->oa.o_size,
 186                                    repbody->oa.o_blocks, oti);
 187         RETURN(0);
 188 }
 189
 190 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
 191 {
 192         struct ost_body *body, *repbody;
 193         int rc, size = sizeof(*repbody);
 194         ENTRY;
 195
 196         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 197         if (body == NULL)
 198                 RETURN(-EFAULT);
 199
 200         rc = lustre_pack_reply(req, 1, &size, NULL);
 201         if (rc)
 202                 RETURN(rc);
 203
 204         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 205         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 206         req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
 207                                   repbody->oa.o_blocks);
 208         RETURN(0);
 209 }
 210
 211 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
 212                        struct obd_trans_info *oti)
 213 {
 214         struct ost_body *body, *repbody;
 215         int rc, size = sizeof(*repbody);
 216         ENTRY;
 217
 218         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 219         if (body == NULL)
 220                 RETURN(-EFAULT);
 221
 222         rc = lustre_pack_reply(req, 1, &size, NULL);
 223         if (rc)
 224                 RETURN(rc);
 225
 226         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 227         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 228
 229         req->rq_status = obd_setattr(exp, &repbody->oa, NULL, oti);
 230         RETURN(0);
 231 }
 232
 233 static int ost_bulk_timeout(void *data)
 234 {
 235         ENTRY;
 236         /* We don't fail the connection here, because having the export
 237          * killed makes the (vital) call to commitrw very sad.
 238          */
 239         RETURN(1);
 240 }
 241
 242 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
 243                                 struct niobuf_remote *rnb, int nrnb,
 244                                 struct niobuf_remote **pp_rnbp)
 245 {
 246         /* Copy a remote niobuf, splitting it into page-sized chunks
 247          * and setting ioo[i].ioo_bufcnt accordingly */
 248         struct niobuf_remote *pp_rnb;
 249         int   i;
 250         int   j;
 251         int   page;
 252         int   rnbidx = 0;
 253         int   npages = 0;
 254
 255         /* first count and check the number of pages required */
 256         for (i = 0; i < nioo; i++)
 257                 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
 258                         obd_off offset = rnb[rnbidx].offset;
 259                         obd_off p0 = offset >> PAGE_SHIFT;
 260                         obd_off pn = (offset + rnb[rnbidx].len - 1)>>PAGE_SHIFT;
 261
 262                         LASSERT(rnbidx < nrnb);
 263
 264                         npages += (pn + 1 - p0);
 265
 266                         if (rnb[rnbidx].len == 0) {
 267                                 CERROR("zero len BRW: obj %d objid "LPX64
 268                                        " buf %u\n", i, ioo[i].ioo_id, j);
 269                                 return -EINVAL;
 270                         }
 271                         if (j > 0 &&
 272                             rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
 273                                 CERROR("unordered BRW: obj %d objid "LPX64
 274                                        " buf %u offset "LPX64" <= "LPX64"\n",
 275                                        i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
 276                                        rnb[rnbidx].offset);
 277                                 return -EINVAL;
 278                         }
 279                 }
 280
 281         LASSERT(rnbidx == nrnb);
 282
 283         if (npages == nrnb) {       /* all niobufs are for single pages */
 284                 *pp_rnbp = rnb;
 285                 return npages;
 286         }
 287
 288         OBD_ALLOC(pp_rnb, sizeof(*pp_rnb) * npages);
 289         if (pp_rnb == NULL)
 290                 return -ENOMEM;
 291
 292         /* now do the actual split */
 293         page = rnbidx = 0;
 294         for (i = 0; i < nioo; i++) {
 295                 int  obj_pages = 0;
 296
 297                 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
 298                         obd_off off = rnb[rnbidx].offset;
 299                         int     nob = rnb[rnbidx].len;
 300
 301                         LASSERT(rnbidx < nrnb);
 302                         do {
 303                                 obd_off  poff = off & (PAGE_SIZE - 1);
 304                                 int      pnob = (poff + nob > PAGE_SIZE) ?
 305                                                 PAGE_SIZE - poff : nob;
 306
 307                                 LASSERT(page < npages);
 308                                 pp_rnb[page].len = pnob;
 309                                 pp_rnb[page].offset = off;
 310                                 pp_rnb[page].flags = rnb[rnbidx].flags;
 311
 312                                 CDEBUG(0, "   obj %d id "LPX64
 313                                        "page %d(%d) "LPX64" for %d, flg %x\n",
 314                                        i, ioo[i].ioo_id, obj_pages, page,
 315                                        pp_rnb[page].offset, pp_rnb[page].len,
 316                                        pp_rnb[page].flags);
 317                                 page++;
 318                                 obj_pages++;
 319
 320                                 off += pnob;
 321                                 nob -= pnob;
 322                         } while (nob > 0);
 323                         LASSERT(nob == 0);
 324                 }
 325                 ioo[i].ioo_bufcnt = obj_pages;
 326         }
 327         LASSERT(page == npages);
 328
 329         *pp_rnbp = pp_rnb;
 330         return npages;
 331 }
 332
 333 static void free_per_page_niobufs (int npages, struct niobuf_remote *pp_rnb,
 334                                    struct niobuf_remote *rnb)
 335 {
 336         if (pp_rnb == rnb)                      /* didn't allocate above */
 337                 return;
 338
 339         OBD_FREE(pp_rnb, sizeof(*pp_rnb) * npages);
 340 }
 341
 342 #if CHECKSUM_BULK
 343 obd_count ost_checksum_bulk(struct ptlrpc_bulk_desc *desc)
 344 {
 345         obd_count cksum = 0;
 346         struct ptlrpc_bulk_page *bp;
 347
 348         list_for_each_entry(bp, &desc->bd_page_list, bp_link) {
 349                 ost_checksum(&cksum, kmap(bp->bp_page) + bp->bp_pageoffset,
 350                              bp->bp_buflen);
 351                 kunmap(bp->bp_page);
 352         }
 353
 354         return cksum;
 355 }
 356 #endif
 357
 358 static void ost_stime_record(struct ptlrpc_request *req, struct timeval *start,
 359                              unsigned rw, unsigned phase)
 360 {
 361         struct obd_device *obd = req->rq_svc->srv_obddev;
 362         struct timeval stop;
 363         int ind = rw *3 + phase;
 364
 365         if (obd && obd->obd_type && obd->obd_type->typ_name) {
 366                 if (!strcmp(obd->obd_type->typ_name, OBD_OST_DEVICENAME)) {
 367                         struct ost_obd *ost = NULL;
 368
 369                         ost = &obd->u.ost;
 370                         if (ind >= (sizeof(ost->ost_stimes) /
 371                                     sizeof(ost->ost_stimes[0])))
 372                                return;
 373                         do_gettimeofday(&stop);
 374
 375                         spin_lock(&ost->ost_lock);
 376                         lprocfs_stime_record(&ost->ost_stimes[ind],&stop,start);
 377                         spin_unlock(&ost->ost_lock);
 378                         memcpy(start, &stop, sizeof(*start));
 379                 }
 380        }
 381 }
 382
 383 static int ost_brw_read(struct ptlrpc_request *req)
 384 {
 385         struct ptlrpc_bulk_desc *desc;
 386         struct niobuf_remote    *remote_nb;
 387         struct niobuf_remote    *pp_rnb;
 388         struct niobuf_local     *local_nb;
 389         struct obd_ioobj        *ioo;
 390         struct ost_body         *body, *repbody;
 391         struct l_wait_info       lwi;
 392         struct obd_trans_info    oti = { 0 };
 393         int                      size[1] = { sizeof(*body) };
 394         int                      comms_error = 0;
 395         int                      niocount;
 396         int                      npages;
 397         int                      nob = 0;
 398         int                      rc;
 399         int                      i;
 400         struct timeval           start;
 401         ENTRY;
 402
 403         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
 404                 GOTO(out, rc = -EIO);
 405
 406         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 407                          (obd_timeout + 1) / 4);
 408
 409         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 410         if (body == NULL) {
 411                 CERROR("Missing/short ost_body\n");
 412                 GOTO(out, rc = -EFAULT);
 413         }
 414
 415         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 416         if (ioo == NULL) {
 417                 CERROR("Missing/short ioobj\n");
 418                 GOTO(out, rc = -EFAULT);
 419         }
 420
 421         niocount = ioo->ioo_bufcnt;
 422         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 423                                        lustre_swab_niobuf_remote);
 424         if (remote_nb == NULL) {
 425                 CERROR("Missing/short niobuf\n");
 426                 GOTO(out, rc = -EFAULT);
 427         }
 428         if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
 429                 for (i = 1; i < niocount; i++)
 430                         lustre_swab_niobuf_remote (&remote_nb[i]);
 431         }
 432
 433         rc = lustre_pack_reply(req, 1, size, NULL);
 434         if (rc)
 435                 GOTO(out, rc);
 436
 437         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 438         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 439         npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
 440         if (npages < 0)
 441                 GOTO(out, rc = npages);
 442
 443         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 444         if (local_nb == NULL)
 445                 GOTO(out_pp_rnb, rc = -ENOMEM);
 446
 447         desc = ptlrpc_prep_bulk_exp (req, npages,
 448                                      BULK_PUT_SOURCE, OST_BULK_PORTAL);
 449         if (desc == NULL)
 450                 GOTO(out_local, rc = -ENOMEM);
 451
 452         do_gettimeofday(&start);
 453         rc = obd_preprw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 454                         ioo, npages, pp_rnb, local_nb, &oti);
 455         ost_stime_record(req, &start, 0, 0);
 456         if (rc != 0)
 457                 GOTO(out_bulk, rc);
 458
 459         /* We're finishing using body->oa as an input variable */
 460         body->oa.o_valid = 0;
 461
 462         nob = 0;
 463         for (i = 0; i < npages; i++) {
 464                 int page_rc = local_nb[i].rc;
 465
 466                 if (page_rc < 0) {              /* error */
 467                         rc = page_rc;
 468                         break;
 469                 }
 470
 471                 LASSERT(page_rc <= pp_rnb[i].len);
 472                 nob += page_rc;
 473                 if (page_rc != 0) {             /* some data! */
 474                         LASSERT (local_nb[i].page != NULL);
 475                         ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 476                                               pp_rnb[i].offset & (PAGE_SIZE-1),
 477                                               page_rc);
 478                 }
 479
 480                 if (page_rc != pp_rnb[i].len) { /* short read */
 481                         /* All subsequent pages should be 0 */
 482                         while(++i < npages)
 483                                 LASSERT(local_nb[i].rc == 0);
 484                         break;
 485                 }
 486         }
 487
 488         if (rc == 0) {
 489                 rc = ptlrpc_start_bulk_transfer(desc);
 490                 if (rc == 0) {
 491                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 492                                           ost_bulk_timeout, desc);
 493                         rc = l_wait_event(desc->bd_waitq,
 494                                           !ptlrpc_bulk_active(desc), &lwi);
 495                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 496                         if (rc == -ETIMEDOUT) {
 497                                 DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT");
 498                                 ptlrpc_abort_bulk(desc);
 499                         } else if (!desc->bd_success ||
 500                                    desc->bd_nob_transferred != desc->bd_nob) {
 501                                 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
 502                                           desc->bd_success ?
 503                                           "truncated" : "network error on",
 504                                           desc->bd_nob_transferred,
 505                                           desc->bd_nob);
 506                                 /* XXX should this be a different errno? */
 507                                 rc = -ETIMEDOUT;
 508                         }
 509                 } else {
 510                         DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d\n", rc);
 511                 }
 512                 comms_error = rc != 0;
 513         }
 514
 515         ost_stime_record(req, &start, 0, 1);
 516         /* Must commit after prep above in all cases */
 517         rc = obd_commitrw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 518                           ioo, npages, local_nb, &oti, rc);
 519         ost_stime_record(req, &start, 0, 2);
 520
 521         if (rc == 0) {
 522                 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 523                 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 524
 525 #if CHECKSUM_BULK
 526                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 527                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 528 #endif
 529         }
 530
 531  out_bulk:
 532         ptlrpc_free_bulk(desc);
 533  out_local:
 534         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 535  out_pp_rnb:
 536         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 537  out:
 538         LASSERT(rc <= 0);
 539         if (rc == 0) {
 540                 req->rq_status = nob;
 541                 ptlrpc_reply(req);
 542         } else if (!comms_error) {
 543                 /* only reply if comms OK */
 544                 req->rq_status = rc;
 545                 ptlrpc_error(req);
 546         } else {
 547                 if (req->rq_reply_state != NULL) {
 548                         /* reply out callback would free */
 549                         lustre_free_reply_state (req->rq_reply_state);
 550                 }
 551                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 552                         CERROR("bulk IO comms error: "
 553                                "evicting %s@%s id %s\n",
 554                                req->rq_export->exp_client_uuid.uuid,
 555                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 556                                req->rq_peerstr);
 557                         ptlrpc_fail_export(req->rq_export);
 558                 } else {
 559                         CERROR("ignoring bulk IO comms error: "
 560                                "client reconnected %s@%s id %s\n",
 561                                req->rq_export->exp_client_uuid.uuid,
 562                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 563                                req->rq_peerstr);
 564                 }
 565         }
 566
 567         RETURN(rc);
 568 }
 569
 570 int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
 571 {
 572         struct ptlrpc_bulk_desc *desc;
 573         struct niobuf_remote    *remote_nb;
 574         struct niobuf_remote    *pp_rnb;
 575         struct niobuf_local     *local_nb;
 576         struct obd_ioobj        *ioo;
 577         struct ost_body         *body, *repbody;
 578         struct l_wait_info       lwi;
 579         __u32                   *rcs;
 580         int                      size[2] = { sizeof(*body) };
 581         int                      objcount, niocount, npages;
 582         int                      comms_error = 0;
 583         int                      rc, swab, i, j;
 584         struct timeval           start;
 585         ENTRY;
 586
 587         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
 588                 GOTO(out, rc = -EIO);
 589
 590         /* pause before transaction has been started */
 591         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 592                          (obd_timeout + 1) / 4);
 593
 594         swab = lustre_msg_swabbed(req->rq_reqmsg);
 595         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 596         if (body == NULL) {
 597                 CERROR("Missing/short ost_body\n");
 598                 GOTO(out, rc = -EFAULT);
 599         }
 600
 601         LASSERT_REQSWAB(req, 1);
 602         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 603         if (objcount == 0) {
 604                 CERROR("Missing/short ioobj\n");
 605                 GOTO(out, rc = -EFAULT);
 606         }
 607         ioo = lustre_msg_buf (req->rq_reqmsg, 1, objcount * sizeof(*ioo));
 608         LASSERT (ioo != NULL);
 609         for (niocount = i = 0; i < objcount; i++) {
 610                 if (swab)
 611                         lustre_swab_obd_ioobj (&ioo[i]);
 612                 if (ioo[i].ioo_bufcnt == 0) {
 613                         CERROR("ioo[%d] has zero bufcnt\n", i);
 614                         GOTO(out, rc = -EFAULT);
 615                 }
 616                 niocount += ioo[i].ioo_bufcnt;
 617         }
 618
 619         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 620                                        lustre_swab_niobuf_remote);
 621         if (remote_nb == NULL) {
 622                 CERROR("Missing/short niobuf\n");
 623                 GOTO(out, rc = -EFAULT);
 624         }
 625         if (swab) {                             /* swab the remaining niobufs */
 626                 for (i = 1; i < niocount; i++)
 627                         lustre_swab_niobuf_remote (&remote_nb[i]);
 628         }
 629
 630         size[1] = niocount * sizeof(*rcs);
 631         rc = lustre_pack_reply(req, 2, size, NULL);
 632         if (rc != 0)
 633                 GOTO(out, rc);
 634         rcs = lustre_msg_buf(req->rq_repmsg, 1, niocount * sizeof(*rcs));
 635
 636 #if 0
 637         /* Do snap options here*/
 638         rc = obd_do_cow(req->rq_export, ioo, objcount, remote_nb);
 639         if (rc)
 640                 GOTO(out, rc);
 641 #endif
 642
 643         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 644         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 645         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 646         if (npages < 0)
 647                 GOTO(out, rc = npages);
 648
 649         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 650         if (local_nb == NULL)
 651                 GOTO(out_pp_rnb, rc = -ENOMEM);
 652
 653         desc = ptlrpc_prep_bulk_exp (req, npages,
 654                                      BULK_GET_SINK, OST_BULK_PORTAL);
 655         if (desc == NULL)
 656                 GOTO(out_local, rc = -ENOMEM);
 657
 658         do_gettimeofday(&start);
 659         rc = obd_preprw(OBD_BRW_WRITE, req->rq_export, &body->oa, objcount,
 660                         ioo, npages, pp_rnb, local_nb, oti);
 661         ost_stime_record(req, &start, 1, 0);
 662         if (rc != 0)
 663                 GOTO(out_bulk, rc);
 664
 665         /* NB Having prepped, we must commit... */
 666
 667         for (i = 0; i < npages; i++)
 668                 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 669                                       pp_rnb[i].offset & (PAGE_SIZE - 1),
 670                                       pp_rnb[i].len);
 671
 672         rc = ptlrpc_start_bulk_transfer (desc);
 673         if (rc == 0) {
 674                 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 675                                   ost_bulk_timeout, desc);
 676                 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc),
 677                                   &lwi);
 678                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 679                 if (rc == -ETIMEDOUT) {
 680                         DEBUG_REQ(D_ERROR, req, "timeout on bulk GET");
 681                         ptlrpc_abort_bulk(desc);
 682                 } else if (!desc->bd_success ||
 683                            desc->bd_nob_transferred != desc->bd_nob) {
 684                         DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
 685                                   desc->bd_success ?
 686                                   "truncated" : "network error on",
 687                                   desc->bd_nob_transferred, desc->bd_nob);
 688                         /* XXX should this be a different errno? */
 689                         rc = -ETIMEDOUT;
 690                 }
 691         } else {
 692                 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d\n", rc);
 693         }
 694         comms_error = rc != 0;
 695
 696         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 697         memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 698
 699 #if CHECKSUM_BULK
 700         if (rc == 0 && (body->oa.o_valid & OBD_MD_FLCKSUM) != 0) {
 701                 static int cksum_counter;
 702                 obd_count client_cksum = body->oa.o_cksum;
 703                 obd_count cksum = ost_checksum_bulk(desc);
 704
 705                 if (client_cksum != cksum) {
 706                         CERROR("Bad checksum: client %x, server %x id %s\n",
 707                                client_cksum, cksum,
 708                                req->rq_peerstr);
 709                         cksum_counter = 1;
 710                         repbody->oa.o_cksum = cksum;
 711                 } else {
 712                         cksum_counter++;
 713                         if ((cksum_counter & (-cksum_counter)) == cksum_counter)
 714                                 CWARN("Checksum %u from NID %s: %x OK\n",
 715                                       cksum_counter, req->rq_peerstr, cksum);
 716                 }
 717         }
 718 #endif
 719         ost_stime_record(req, &start, 1, 1);
 720         /* Must commit after prep above in all cases */
 721         rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa,
 722                           objcount, ioo, npages, local_nb, oti, rc);
 723
 724         ost_stime_record(req, &start, 1, 2);
 725         if (rc == 0) {
 726 #if CHECKSUM_BULK
 727                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 728                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 729 #endif
 730                 /* set per-requested niobuf return codes */
 731                 for (i = j = 0; i < niocount; i++) {
 732                         int nob = remote_nb[i].len;
 733
 734                         rcs[i] = 0;
 735                         do {
 736                                 LASSERT(j < npages);
 737                                 if (local_nb[j].rc < 0)
 738                                         rcs[i] = local_nb[j].rc;
 739                                 nob -= pp_rnb[j].len;
 740                                 j++;
 741                         } while (nob > 0);
 742                         LASSERT(nob == 0);
 743                 }
 744                 LASSERT(j == npages);
 745         }
 746         /*XXX This write extents only for write-back cache extents*/
 747         rc = obd_write_extents(req->rq_export, ioo, objcount, niocount,
 748                                local_nb, rc);
 749  out_bulk:
 750         ptlrpc_free_bulk(desc);
 751  out_local:
 752         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 753  out_pp_rnb:
 754         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 755  out:
 756         if (rc == 0) {
 757                 oti_to_request(oti, req);
 758                 rc = ptlrpc_reply(req);
 759         } else if (!comms_error) {
 760                 /* Only reply if there was no comms problem with bulk */
 761                 req->rq_status = rc;
 762                 ptlrpc_error(req);
 763         } else {
 764                 if (req->rq_reply_state != NULL) {
 765                         /* reply out callback would free */
 766                         lustre_free_reply_state (req->rq_reply_state);
 767                 }
 768                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 769                         CERROR("%s: bulk IO comm error evicting %s@%s id %s\n",
 770                                req->rq_export->exp_obd->obd_name,
 771                                req->rq_export->exp_client_uuid.uuid,
 772                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 773                                req->rq_peerstr);
 774                         ptlrpc_fail_export(req->rq_export);
 775                 } else {
 776                         CERROR("ignoring bulk IO comms error: "
 777                                "client reconnected %s@%s id %s\n",
 778                                req->rq_export->exp_client_uuid.uuid,
 779                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 780                                req->rq_peerstr);
 781                 }
 782         }
 783         RETURN(rc);
 784 }
 785 EXPORT_SYMBOL(ost_brw_write);
 786
 787 static int ost_san_brw(struct ptlrpc_request *req, int cmd)
 788 {
 789         struct niobuf_remote *remote_nb, *res_nb, *pp_rnb;
 790         struct obd_ioobj *ioo;
 791         struct ost_body *body, *repbody;
 792         int rc, i, objcount, niocount, size[2] = {sizeof(*body)}, npages;
 793         int swab;
 794         ENTRY;
 795
 796         /* XXX not set to use latest protocol */
 797
 798         swab = lustre_msg_swabbed(req->rq_reqmsg);
 799         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 800         if (body == NULL) {
 801                 CERROR("Missing/short ost_body\n");
 802                 GOTO(out, rc = -EFAULT);
 803         }
 804
 805         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 806         if (ioo == NULL) {
 807                 CERROR("Missing/short ioobj\n");
 808                 GOTO(out, rc = -EFAULT);
 809         }
 810         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 811         niocount = ioo[0].ioo_bufcnt;
 812         for (i = 1; i < objcount; i++) {
 813                 if (swab)
 814                         lustre_swab_obd_ioobj (&ioo[i]);
 815                 niocount += ioo[i].ioo_bufcnt;
 816         }
 817
 818         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 819                                        lustre_swab_niobuf_remote);
 820         if (remote_nb == NULL) {
 821                 CERROR("Missing/short niobuf\n");
 822                 GOTO(out, rc = -EFAULT);
 823         }
 824         if (swab) {                             /* swab the remaining niobufs */
 825                 for (i = 1; i < niocount; i++)
 826                         lustre_swab_niobuf_remote (&remote_nb[i]);
 827         }
 828
 829         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 830         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 831         if (npages < 0)
 832                 GOTO (out, rc = npages);
 833
 834         size[1] = npages * sizeof(*pp_rnb);
 835         rc = lustre_pack_reply(req, 2, size, NULL);
 836         if (rc)
 837                 GOTO(out_pp_rnb, rc);
 838
 839         req->rq_status = obd_san_preprw(cmd, req->rq_export, &body->oa,
 840                                         objcount, ioo, npages, pp_rnb);
 841
 842         if (req->rq_status)
 843                 GOTO(out_pp_rnb, rc = 0);
 844
 845         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 846         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 847
 848         res_nb = lustre_msg_buf(req->rq_repmsg, 1, size[1]);
 849         memcpy(res_nb, remote_nb, size[1]);
 850         rc = 0;
 851 out_pp_rnb:
 852         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 853 out:
 854         if (rc) {
 855                 req->rq_status = rc;
 856                 ptlrpc_error(req);
 857         } else
 858                 ptlrpc_reply(req);
 859
 860         return rc;
 861 }
 862
 863 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
 864 {
 865         char *key, *val;
 866         int keylen, rc = 0;
 867         ENTRY;
 868
 869         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 870         if (key == NULL) {
 871                 DEBUG_REQ(D_HA, req, "no set_info key");
 872                 RETURN(-EFAULT);
 873         }
 874         keylen = req->rq_reqmsg->buflens[0];
 875
 876         rc = lustre_pack_reply(req, 0, NULL, NULL);
 877         if (rc)
 878                 RETURN(rc);
 879
 880         val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
 881
 882         rc = obd_set_info(exp, keylen, key, req->rq_reqmsg->buflens[1], val);
 883         req->rq_repmsg->status = 0;
 884         RETURN(rc);
 885 }
 886
 887 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
 888 {
 889         char *key;
 890         int keylen, rc = 0, size = sizeof(obd_id);
 891         obd_id *reply;
 892         ENTRY;
 893
 894         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 895         if (key == NULL) {
 896                 DEBUG_REQ(D_HA, req, "no get_info key");
 897                 RETURN(-EFAULT);
 898         }
 899         keylen = req->rq_reqmsg->buflens[0];
 900
 901         if (keylen < strlen("last_id") || memcmp(key, "last_id", 7) != 0)
 902                 RETURN(-EPROTO);
 903
 904         rc = lustre_pack_reply(req, 1, &size, NULL);
 905         if (rc)
 906                 RETURN(rc);
 907
 908         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
 909         rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
 910         req->rq_repmsg->status = 0;
 911         RETURN(rc);
 912 }
 913
 914 static int ost_llog_handle_connect(struct obd_export *exp,
 915                                    struct ptlrpc_request *req)
 916 {
 917         struct llogd_conn_body *body;
 918         int rc;
 919         ENTRY;
 920
 921         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 922         rc = obd_llog_connect(exp, body);
 923         RETURN(rc);
 924 }
 925
 926 static int ost_filter_recovery_request(struct ptlrpc_request *req,
 927                                        struct obd_device *obd, int *process)
 928 {
 929         switch (req->rq_reqmsg->opc) {
 930         case OST_CONNECT: /* This will never get here, but for completeness. */
 931         case OST_DISCONNECT:
 932                *process = 1;
 933                RETURN(0);
 934
 935         case OBD_PING:
 936         case OST_CREATE:
 937         case OST_DESTROY:
 938         case OST_PUNCH:
 939         case OST_SETATTR:
 940         case OST_SYNC:
 941         case OST_WRITE:
 942         case OBD_LOG_CANCEL:
 943         case LDLM_ENQUEUE:
 944                 *process = target_queue_recovery_request(req, obd);
 945                 RETURN(0);
 946
 947         default:
 948                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 949                 *process = 0;
 950                 /* XXX what should we set rq_status to here? */
 951                 req->rq_status = -EAGAIN;
 952                 RETURN(ptlrpc_error(req));
 953         }
 954 }
 955
 956 int ost_msg_check_version(struct lustre_msg *msg)
 957 {
 958         int rc;
 959
 960         switch(msg->opc) {
 961         case OST_CONNECT:
 962         case OST_DISCONNECT:
 963         case OBD_PING:
 964         case OST_CREATE:
 965         case OST_DESTROY:
 966         case OST_GETATTR:
 967         case OST_SETATTR:
 968         case OST_WRITE:
 969         case OST_READ:
 970         case OST_SAN_READ:
 971         case OST_SAN_WRITE:
 972         case OST_PUNCH:
 973         case OST_STATFS:
 974         case OST_SYNC:
 975         case OST_SET_INFO:
 976         case OST_GET_INFO:
 977                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
 978                 if (rc)
 979                         CERROR("bad opc %u version %08x, expecting %08x\n",
 980                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
 981                 break;
 982         case LDLM_ENQUEUE:
 983         case LDLM_CONVERT:
 984         case LDLM_CANCEL:
 985         case LDLM_BL_CALLBACK:
 986         case LDLM_CP_CALLBACK:
 987                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
 988                 if (rc)
 989                         CERROR("bad opc %u version %08x, expecting %08x\n",
 990                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
 991                 break;
 992         case OBD_LOG_CANCEL:
 993         case LLOG_ORIGIN_CONNECT:
 994                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
 995                 if (rc)
 996                         CERROR("bad opc %u version %08x, expecting %08x\n",
 997                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
 998                 break;
 999         case SEC_INIT:
1000         case SEC_INIT_CONTINUE:
1001         case SEC_FINI:
1002                 rc = 0;
1003                 break;
1004         default:
1005                 CERROR("OST unexpected opcode %d\n", msg->opc);
1006                 rc = -ENOTSUPP;
1007                 break;
1008         }
1009         return rc;
1010 }
1011
1012 int ost_handle(struct ptlrpc_request *req)
1013 {
1014         int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1015         struct obd_trans_info *oti = NULL;
1016         struct obd_device *obd = NULL;
1017         ENTRY;
1018
1019         LASSERT(current->journal_info == NULL);
1020
1021         rc = ost_msg_check_version(req->rq_reqmsg);
1022         if (rc) {
1023                 CERROR("OST drop mal-formed request\n");
1024                 RETURN(rc);
1025         }
1026
1027         /* Security opc should NOT trigger any recovery events */
1028         if (req->rq_reqmsg->opc == SEC_INIT ||
1029             req->rq_reqmsg->opc == SEC_INIT_CONTINUE ||
1030             req->rq_reqmsg->opc == SEC_FINI) {
1031                 GOTO(out_check_req, rc = 0);
1032         }
1033
1034         /* XXX identical to MDS */
1035         if (req->rq_reqmsg->opc != OST_CONNECT) {
1036                 int recovering;
1037
1038                 if (req->rq_export == NULL) {
1039                         CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1040                                req->rq_reqmsg->opc,
1041                                req->rq_peerstr);
1042                         req->rq_status = -ENOTCONN;
1043                         GOTO(out_check_req, rc = -ENOTCONN);
1044                 }
1045
1046                 obd = req->rq_export->exp_obd;
1047
1048                 /* Check for aborted recovery. */
1049                 spin_lock_bh(&obd->obd_processing_task_lock);
1050                 recovering = obd->obd_recovering;
1051                 spin_unlock_bh(&obd->obd_processing_task_lock);
1052                 if (recovering) {
1053                         rc = ost_filter_recovery_request(req, obd,
1054                                                          &should_process);
1055                         if (rc || !should_process)
1056                                 RETURN(rc);
1057                         if (should_process < 0) {
1058                                 req->rq_status = should_process;
1059                                 rc = ptlrpc_error(req);
1060                                 RETURN(rc);
1061                         }
1062                 }
1063         }
1064
1065         OBD_ALLOC(oti, sizeof(*oti));
1066         if (oti == NULL)
1067                 RETURN(-ENOMEM);
1068
1069         oti_init(oti, req);
1070
1071         switch (req->rq_reqmsg->opc) {
1072         case OST_CONNECT: {
1073                 CDEBUG(D_INODE, "connect\n");
1074                 OBD_FAIL_GOTO(OBD_FAIL_OST_CONNECT_NET, out_free_oti, rc = 0);
1075                 rc = target_handle_connect(req);
1076                 if (!rc)
1077                         obd = req->rq_export->exp_obd;
1078                 break;
1079         }
1080         case OST_DISCONNECT:
1081                 CDEBUG(D_INODE, "disconnect\n");
1082                 OBD_FAIL_GOTO(OBD_FAIL_OST_DISCONNECT_NET, out_free_oti, rc = 0);
1083                 rc = target_handle_disconnect(req);
1084                 break;
1085         case OST_CREATE:
1086                 CDEBUG(D_INODE, "create\n");
1087                 OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC);
1088                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1089                 OBD_FAIL_GOTO(OBD_FAIL_OST_CREATE_NET, out_free_oti, rc = 0);
1090                 rc = ost_create(req->rq_export, req, oti);
1091                 break;
1092         case OST_DESTROY:
1093                 CDEBUG(D_INODE, "destroy\n");
1094                 OBD_FAIL_GOTO(OBD_FAIL_OST_DESTROY_NET, out_free_oti, rc = 0);
1095                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1096                 rc = ost_destroy(req->rq_export, req, oti);
1097                 break;
1098         case OST_GETATTR:
1099                 CDEBUG(D_INODE, "getattr\n");
1100                 OBD_FAIL_GOTO(OBD_FAIL_OST_GETATTR_NET, out_free_oti, rc = 0);
1101                 rc = ost_getattr(req->rq_export, req);
1102                 break;
1103         case OST_SETATTR:
1104                 CDEBUG(D_INODE, "setattr\n");
1105                 OBD_FAIL_GOTO(OBD_FAIL_OST_SETATTR_NET, out_free_oti, rc = 0);
1106                 rc = ost_setattr(req->rq_export, req, oti);
1107                 break;
1108         case OST_WRITE:
1109                 CDEBUG(D_INODE, "write\n");
1110                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1111                 OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC);
1112                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1113                 rc = ost_brw_write(req, oti);
1114                 LASSERT(current->journal_info == NULL);
1115                 /* ost_brw sends its own replies */
1116                 GOTO(out_free_oti, rc);
1117         case OST_READ:
1118                 CDEBUG(D_INODE, "read\n");
1119                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1120                 rc = ost_brw_read(req);
1121                 LASSERT(current->journal_info == NULL);
1122                 /* ost_brw sends its own replies */
1123                 GOTO(out_free_oti, rc);
1124         case OST_SAN_READ:
1125                 CDEBUG(D_INODE, "san read\n");
1126                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1127                 rc = ost_san_brw(req, OBD_BRW_READ);
1128                 /* ost_san_brw sends its own replies */
1129                 GOTO(out_free_oti, rc);
1130         case OST_SAN_WRITE:
1131                 CDEBUG(D_INODE, "san write\n");
1132                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1133                 rc = ost_san_brw(req, OBD_BRW_WRITE);
1134                 /* ost_san_brw sends its own replies */
1135                 GOTO(out_free_oti, rc);
1136         case OST_PUNCH:
1137                 CDEBUG(D_INODE, "punch\n");
1138                 OBD_FAIL_GOTO(OBD_FAIL_OST_PUNCH_NET, out_free_oti, rc = 0);
1139                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1140                 rc = ost_punch(req->rq_export, req, oti);
1141                 break;
1142         case OST_STATFS:
1143                 CDEBUG(D_INODE, "statfs\n");
1144                 OBD_FAIL_GOTO(OBD_FAIL_OST_STATFS_NET, out_free_oti, rc = 0);
1145                 rc = ost_statfs(req);
1146                 break;
1147         case OST_SYNC:
1148                 CDEBUG(D_INODE, "sync\n");
1149                 OBD_FAIL_GOTO(OBD_FAIL_OST_SYNC_NET, out_free_oti, rc = 0);
1150                 rc = ost_sync(req->rq_export, req);
1151                 break;
1152         case OST_SET_INFO:
1153                 DEBUG_REQ(D_INODE, req, "set_info");
1154                 rc = ost_set_info(req->rq_export, req);
1155                 break;
1156         case OST_GET_INFO:
1157                 DEBUG_REQ(D_INODE, req, "get_info");
1158                 rc = ost_get_info(req->rq_export, req);
1159                 break;
1160         case OBD_PING:
1161                 DEBUG_REQ(D_INODE, req, "ping");
1162                 rc = target_handle_ping(req);
1163                 break;
1164         /* FIXME - just reply status */
1165         case LLOG_ORIGIN_CONNECT:
1166                 DEBUG_REQ(D_INODE, req, "log connect\n");
1167                 rc = ost_llog_handle_connect(req->rq_export, req);
1168                 req->rq_status = rc;
1169                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1170                 if (rc)
1171                         GOTO(out_free_oti, rc);
1172                 GOTO(out_free_oti, rc = ptlrpc_reply(req));
1173         case OBD_LOG_CANCEL:
1174                 CDEBUG(D_INODE, "log cancel\n");
1175                 OBD_FAIL_GOTO(OBD_FAIL_OBD_LOG_CANCEL_NET, out_free_oti, rc = 0);
1176                 rc = llog_origin_handle_cancel(req);
1177                 req->rq_status = rc;
1178                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1179                 if (rc)
1180                         GOTO(out_free_oti, rc);
1181                 GOTO(out_free_oti, rc = ptlrpc_reply(req));
1182         case LDLM_ENQUEUE:
1183                 CDEBUG(D_INODE, "enqueue\n");
1184                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_ENQUEUE, out_free_oti, rc = 0);
1185                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1186                                          ldlm_server_blocking_ast,
1187                                          ldlm_server_glimpse_ast);
1188                 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1189                 break;
1190         case LDLM_CONVERT:
1191                 CDEBUG(D_INODE, "convert\n");
1192                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_CONVERT, out_free_oti, rc = 0);
1193                 rc = ldlm_handle_convert(req);
1194                 break;
1195         case LDLM_CANCEL:
1196                 CDEBUG(D_INODE, "cancel\n");
1197                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_CANCEL, out_free_oti, rc = 0);
1198                 rc = ldlm_handle_cancel(req);
1199                 break;
1200         case LDLM_BL_CALLBACK:
1201         case LDLM_CP_CALLBACK:
1202                 CDEBUG(D_INODE, "callback\n");
1203                 CERROR("callbacks should not happen on OST\n");
1204                 /* fall through */
1205         default:
1206                 CERROR("Unexpected opcode %d\n", req->rq_reqmsg->opc);
1207                 req->rq_status = -ENOTSUPP;
1208                 rc = ptlrpc_error(req);
1209                 GOTO(out_free_oti, rc);
1210         }
1211
1212         LASSERT(current->journal_info == NULL);
1213
1214         EXIT;
1215         /* If we're DISCONNECTing, the export_data is already freed */
1216         if (!rc && req->rq_reqmsg->opc != OST_DISCONNECT) {
1217                 if (!obd->obd_no_transno) {
1218                         req->rq_repmsg->last_committed =
1219                                 obd->obd_last_committed;
1220                 } else {
1221                         DEBUG_REQ(D_IOCTL, req,
1222                                   "not sending last_committed update");
1223                 }
1224                 CDEBUG(D_INFO, "last_committed "LPU64", xid "LPX64"\n",
1225                        obd->obd_last_committed, req->rq_xid);
1226         }
1227
1228 out_check_req:
1229
1230         if (!rc)
1231                 oti_to_request(oti, req);
1232         target_send_reply(req, rc, fail);
1233         rc = 0;
1234
1235 out_free_oti:
1236         if (oti)
1237                 OBD_FREE(oti, sizeof(*oti));
1238         return rc;
1239 }
1240 EXPORT_SYMBOL(ost_handle);
1241
1242 int ost_attach(struct obd_device *dev, obd_count len, void *data)
1243 {
1244         struct lprocfs_static_vars lvars;
1245
1246         lprocfs_init_vars(ost,&lvars);
1247         return lprocfs_obd_attach(dev, lvars.obd_vars);
1248 }
1249
1250 int ost_detach(struct obd_device *dev)
1251 {
1252         return lprocfs_obd_detach(dev);
1253 }
1254
1255 extern struct file_operations ost_stimes_fops;
1256
1257 static int ost_setup(struct obd_device *obd, obd_count len, void *buf)
1258 {
1259         struct ost_obd *ost = &obd->u.ost;
1260         int rc;
1261         ENTRY;
1262
1263         rc = cleanup_group_info();
1264         if (rc)
1265                 RETURN(rc);
1266
1267         rc = llog_start_commit_thread();
1268         if (rc < 0)
1269                 RETURN(rc);
1270
1271         lprocfs_obd_seq_create(obd, "service_times", 0444, &ost_stimes_fops,
1272                                obd);
1273
1274         ost->ost_service =
1275                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1276                                 OST_REQUEST_PORTAL, OSC_REPLY_PORTAL, 30000,
1277                                 ost_handle, "ost",
1278                                 obd->obd_proc_entry);
1279         if (ost->ost_service == NULL) {
1280                 CERROR("failed to start service\n");
1281                 RETURN(-ENOMEM);
1282         }
1283
1284         rc = ptlrpc_start_n_threads(obd, ost->ost_service, OST_NUM_THREADS,
1285                                     "ll_ost");
1286         if (rc)
1287                 GOTO(out_service, rc = -EINVAL);
1288
1289         ost->ost_create_service =
1290                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1291                                 OST_CREATE_PORTAL, OSC_REPLY_PORTAL, 30000,
1292                                 ost_handle, "ost_create",
1293                                 obd->obd_proc_entry);
1294         if (ost->ost_create_service == NULL) {
1295                 CERROR("failed to start OST create service\n");
1296                 GOTO(out_service, rc = -ENOMEM);
1297         }
1298
1299
1300         spin_lock_init(&ost->ost_lock);
1301         ost->ost_service->srv_obddev = obd;
1302
1303         rc = ptlrpc_start_n_threads(obd, ost->ost_create_service, 1,
1304                                     "ll_ost_creat");
1305         if (rc)
1306                 GOTO(out_create, rc = -EINVAL);
1307
1308         RETURN(0);
1309
1310 out_create:
1311         ptlrpc_unregister_service(ost->ost_create_service);
1312 out_service:
1313         ptlrpc_unregister_service(ost->ost_service);
1314         RETURN(rc);
1315 }
1316
1317 extern void lgss_svc_cache_purge_all(void);
1318 static int ost_cleanup(struct obd_device *obd, int flags)
1319 {
1320         struct ost_obd *ost = &obd->u.ost;
1321         int err = 0;
1322         ENTRY;
1323
1324         spin_lock_bh(&obd->obd_processing_task_lock);
1325         if (obd->obd_recovering) {
1326                 target_cancel_recovery_timer(obd);
1327                 obd->obd_recovering = 0;
1328         }
1329         spin_unlock_bh(&obd->obd_processing_task_lock);
1330
1331         ptlrpc_stop_all_threads(ost->ost_service);
1332         ptlrpc_unregister_service(ost->ost_service);
1333
1334         ptlrpc_stop_all_threads(ost->ost_create_service);
1335         ptlrpc_unregister_service(ost->ost_create_service);
1336
1337 #ifdef ENABLE_GSS
1338         /* XXX */
1339         lgss_svc_cache_purge_all();
1340 #endif
1341         RETURN(err);
1342 }
1343
1344 /* use obd ops to offer management infrastructure */
1345 static struct obd_ops ost_obd_ops = {
1346         .o_owner        = THIS_MODULE,
1347         .o_attach       = ost_attach,
1348         .o_detach       = ost_detach,
1349         .o_setup        = ost_setup,
1350         .o_cleanup      = ost_cleanup,
1351 };
1352
1353 static int __init ost_init(void)
1354 {
1355         struct lprocfs_static_vars lvars;
1356         ENTRY;
1357
1358         lprocfs_init_vars(ost,&lvars);
1359         RETURN(class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
1360                                    OBD_OST_DEVICENAME));
1361 }
1362
1363 static void /*__exit*/ ost_exit(void)
1364 {
1365         class_unregister_type(OBD_OST_DEVICENAME);
1366 }
1367
1368 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1369 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
1370 MODULE_LICENSE("GPL");
1371
1372 module_init(ost_init);
1373 module_exit(ost_exit);