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 #include <linux/lustre_audit.h>
  49
  50 void oti_init(struct obd_trans_info *oti, struct ptlrpc_request *req)
  51 {
  52         if (oti == NULL)
  53                 return;
  54         memset(oti, 0, sizeof *oti);
  55
  56         if (req->rq_repmsg && req->rq_reqmsg != 0)
  57                 oti->oti_transno = req->rq_repmsg->transno;
  58 }
  59
  60 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
  61 {
  62         struct oti_req_ack_lock *ack_lock;
  63         int i;
  64
  65         if (oti == NULL)
  66                 return;
  67
  68         if (req->rq_repmsg)
  69                 req->rq_repmsg->transno = oti->oti_transno;
  70
  71         /* XXX 4 == entries in oti_ack_locks??? */
  72         for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
  73                 if (!ack_lock->mode)
  74                         break;
  75                 /* XXX not even calling target_send_reply in some cases... */
  76                 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
  77         }
  78 }
  79
  80 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
  81                        struct obd_trans_info *oti)
  82 {
  83         struct ost_body *body, *repbody;
  84         int rc, size = sizeof(*body);
  85         ENTRY;
  86
  87         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
  88         if (body == NULL)
  89                 RETURN(-EFAULT);
  90
  91         rc = lustre_pack_reply(req, 1, &size, NULL);
  92         if (rc)
  93                 RETURN(rc);
  94
  95         if (body->oa.o_valid & OBD_MD_FLCOOKIE)
  96                 oti->oti_logcookies = obdo_logcookie(&body->oa);
  97         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
  98         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
  99         req->rq_status = obd_destroy(exp, &body->oa, NULL, oti);
 100         RETURN(0);
 101 }
 102
 103 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
 104 {
 105         struct ost_body *body, *repbody;
 106         int rc, size = sizeof(*body);
 107         ENTRY;
 108
 109         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 110         if (body == NULL)
 111                 RETURN(-EFAULT);
 112
 113         rc = lustre_pack_reply(req, 1, &size, NULL);
 114         if (rc)
 115                 RETURN(rc);
 116
 117         repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody));
 118         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 119         req->rq_status = obd_getattr(exp, &repbody->oa, NULL);
 120         RETURN(0);
 121 }
 122
 123 static int ost_statfs(struct ptlrpc_request *req)
 124 {
 125         struct obd_statfs *osfs;
 126         int rc, size = sizeof(*osfs);
 127         ENTRY;
 128
 129         rc = lustre_pack_reply(req, 1, &size, NULL);
 130         if (rc)
 131                 RETURN(rc);
 132
 133         osfs = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*osfs));
 134
 135         req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, jiffies-HZ);
 136         if (req->rq_status != 0)
 137                 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
 138
 139         RETURN(0);
 140 }
 141
 142 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
 143                       struct obd_trans_info *oti)
 144 {
 145         struct ost_body *body, *repbody;
 146         int rc, size = sizeof(*repbody);
 147         ENTRY;
 148
 149         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 150         if (body == NULL)
 151                 RETURN(-EFAULT);
 152
 153         rc = lustre_pack_reply(req, 1, &size, NULL);
 154         if (rc)
 155                 RETURN(rc);
 156
 157         repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody));
 158         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 159         oti->oti_logcookies = obdo_logcookie(&repbody->oa);
 160         req->rq_status = obd_create(exp, &repbody->oa, NULL, 0, NULL, oti);
 161         //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
 162         RETURN(0);
 163 }
 164
 165 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
 166                      struct obd_trans_info *oti)
 167 {
 168         struct ost_body *body, *repbody;
 169         int rc, size = sizeof(*repbody);
 170         ENTRY;
 171
 172         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 173         if (body == NULL)
 174                 RETURN(-EFAULT);
 175
 176         if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
 177             (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
 178                 RETURN(-EINVAL);
 179
 180         rc = lustre_pack_reply(req, 1, &size, NULL);
 181         if (rc)
 182                 RETURN(rc);
 183
 184         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 185         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 186         req->rq_status = obd_punch(exp, &repbody->oa, NULL, repbody->oa.o_size,
 187                                    repbody->oa.o_blocks, oti);
 188         RETURN(0);
 189 }
 190
 191 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
 192 {
 193         struct ost_body *body, *repbody;
 194         int rc, size = sizeof(*repbody);
 195         ENTRY;
 196
 197         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 198         if (body == NULL)
 199                 RETURN(-EFAULT);
 200
 201         rc = lustre_pack_reply(req, 1, &size, NULL);
 202         if (rc)
 203                 RETURN(rc);
 204
 205         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 206         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 207         req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
 208                                   repbody->oa.o_blocks);
 209         RETURN(0);
 210 }
 211
 212 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
 213                        struct obd_trans_info *oti)
 214 {
 215         struct ost_body *body, *repbody;
 216         int rc, size = sizeof(*repbody);
 217         ENTRY;
 218
 219         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 220         if (body == NULL)
 221                 RETURN(-EFAULT);
 222
 223         rc = lustre_pack_reply(req, 1, &size, NULL);
 224         if (rc)
 225                 RETURN(rc);
 226
 227         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 228         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 229
 230         req->rq_status = obd_setattr(exp, &repbody->oa, NULL, oti);
 231         RETURN(0);
 232 }
 233
 234 static int ost_bulk_timeout(void *data)
 235 {
 236         ENTRY;
 237         /* We don't fail the connection here, because having the export
 238          * killed makes the (vital) call to commitrw very sad.
 239          */
 240         RETURN(1);
 241 }
 242
 243 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
 244                                 struct niobuf_remote *rnb, int nrnb,
 245                                 struct niobuf_remote **pp_rnbp)
 246 {
 247         /* Copy a remote niobuf, splitting it into page-sized chunks
 248          * and setting ioo[i].ioo_bufcnt accordingly */
 249         struct niobuf_remote *pp_rnb;
 250         int   i;
 251         int   j;
 252         int   page;
 253         int   rnbidx = 0;
 254         int   npages = 0;
 255
 256         /* first count and check the number of pages required */
 257         for (i = 0; i < nioo; i++)
 258                 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
 259                         obd_off offset = rnb[rnbidx].offset;
 260                         obd_off p0 = offset >> PAGE_SHIFT;
 261                         obd_off pn = (offset + rnb[rnbidx].len - 1)>>PAGE_SHIFT;
 262
 263                         LASSERT(rnbidx < nrnb);
 264
 265                         npages += (pn + 1 - p0);
 266
 267                         if (rnb[rnbidx].len == 0) {
 268                                 CERROR("zero len BRW: obj %d objid "LPX64
 269                                        " buf %u\n", i, ioo[i].ioo_id, j);
 270                                 return -EINVAL;
 271                         }
 272                         if (j > 0 &&
 273                             rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
 274                                 CERROR("unordered BRW: obj %d objid "LPX64
 275                                        " buf %u offset "LPX64" <= "LPX64"\n",
 276                                        i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
 277                                        rnb[rnbidx].offset);
 278                                 return -EINVAL;
 279                         }
 280                 }
 281
 282         LASSERT(rnbidx == nrnb);
 283
 284         if (npages == nrnb) {       /* all niobufs are for single pages */
 285                 *pp_rnbp = rnb;
 286                 return npages;
 287         }
 288
 289         OBD_ALLOC(pp_rnb, sizeof(*pp_rnb) * npages);
 290         if (pp_rnb == NULL)
 291                 return -ENOMEM;
 292
 293         /* now do the actual split */
 294         page = rnbidx = 0;
 295         for (i = 0; i < nioo; i++) {
 296                 int  obj_pages = 0;
 297
 298                 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
 299                         obd_off off = rnb[rnbidx].offset;
 300                         int     nob = rnb[rnbidx].len;
 301
 302                         LASSERT(rnbidx < nrnb);
 303                         do {
 304                                 obd_off  poff = off & (PAGE_SIZE - 1);
 305                                 int      pnob = (poff + nob > PAGE_SIZE) ?
 306                                                 PAGE_SIZE - poff : nob;
 307
 308                                 LASSERT(page < npages);
 309                                 pp_rnb[page].len = pnob;
 310                                 pp_rnb[page].offset = off;
 311                                 pp_rnb[page].flags = rnb[rnbidx].flags;
 312
 313                                 CDEBUG(0, "   obj %d id "LPX64
 314                                        "page %d(%d) "LPX64" for %d, flg %x\n",
 315                                        i, ioo[i].ioo_id, obj_pages, page,
 316                                        pp_rnb[page].offset, pp_rnb[page].len,
 317                                        pp_rnb[page].flags);
 318                                 page++;
 319                                 obj_pages++;
 320
 321                                 off += pnob;
 322                                 nob -= pnob;
 323                         } while (nob > 0);
 324                         LASSERT(nob == 0);
 325                 }
 326                 ioo[i].ioo_bufcnt = obj_pages;
 327         }
 328         LASSERT(page == npages);
 329
 330         *pp_rnbp = pp_rnb;
 331         return npages;
 332 }
 333
 334 static void free_per_page_niobufs (int npages, struct niobuf_remote *pp_rnb,
 335                                    struct niobuf_remote *rnb)
 336 {
 337         if (pp_rnb == rnb)                      /* didn't allocate above */
 338                 return;
 339
 340         OBD_FREE(pp_rnb, sizeof(*pp_rnb) * npages);
 341 }
 342
 343 #if CHECKSUM_BULK
 344 obd_count ost_checksum_bulk(struct ptlrpc_bulk_desc *desc)
 345 {
 346         obd_count cksum = 0;
 347         struct ptlrpc_bulk_page *bp;
 348
 349         list_for_each_entry(bp, &desc->bd_page_list, bp_link) {
 350                 ost_checksum(&cksum, kmap(bp->bp_page) + bp->bp_pageoffset,
 351                              bp->bp_buflen);
 352                 kunmap(bp->bp_page);
 353         }
 354
 355         return cksum;
 356 }
 357 #endif
 358
 359 static void ost_stime_record(struct ptlrpc_request *req, struct timeval *start,
 360                              unsigned rw, unsigned phase)
 361 {
 362         struct obd_device *obd = req->rq_svc->srv_obddev;
 363         struct timeval stop;
 364         int ind = rw *3 + phase;
 365
 366         if (obd && obd->obd_type && obd->obd_type->typ_name) {
 367                 if (!strcmp(obd->obd_type->typ_name, OBD_OST_DEVICENAME)) {
 368                         struct ost_obd *ost = NULL;
 369
 370                         ost = &obd->u.ost;
 371                         if (ind >= (sizeof(ost->ost_stimes) /
 372                                     sizeof(ost->ost_stimes[0])))
 373                                return;
 374                         do_gettimeofday(&stop);
 375
 376                         spin_lock(&ost->ost_lock);
 377                         lprocfs_stime_record(&ost->ost_stimes[ind],&stop,start);
 378                         spin_unlock(&ost->ost_lock);
 379                         memcpy(start, &stop, sizeof(*start));
 380                 }
 381        }
 382 }
 383
 384 static int ost_brw_read(struct ptlrpc_request *req)
 385 {
 386         struct ptlrpc_bulk_desc *desc;
 387         struct niobuf_remote    *remote_nb;
 388         struct niobuf_remote    *pp_rnb;
 389         struct niobuf_local     *local_nb;
 390         struct obd_ioobj        *ioo;
 391         struct ost_body         *body, *repbody;
 392         struct l_wait_info       lwi;
 393         struct obd_trans_info    oti = { 0 };
 394         int                      size[1] = { sizeof(*body) };
 395         int                      comms_error = 0;
 396         int                      niocount;
 397         int                      npages;
 398         int                      nob = 0;
 399         int                      rc;
 400         int                      i;
 401         struct timeval           start;
 402         ENTRY;
 403
 404         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
 405                 GOTO(out, rc = -EIO);
 406
 407         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 408                          (obd_timeout + 1) / 4);
 409
 410         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 411         if (body == NULL) {
 412                 CERROR("Missing/short ost_body\n");
 413                 GOTO(out, rc = -EFAULT);
 414         }
 415
 416         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 417         if (ioo == NULL) {
 418                 CERROR("Missing/short ioobj\n");
 419                 GOTO(out, rc = -EFAULT);
 420         }
 421
 422         niocount = ioo->ioo_bufcnt;
 423         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 424                                        lustre_swab_niobuf_remote);
 425         if (remote_nb == NULL) {
 426                 CERROR("Missing/short niobuf\n");
 427                 GOTO(out, rc = -EFAULT);
 428         }
 429         if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
 430                 for (i = 1; i < niocount; i++)
 431                         lustre_swab_niobuf_remote (&remote_nb[i]);
 432         }
 433
 434         rc = lustre_pack_reply(req, 1, size, NULL);
 435         if (rc)
 436                 GOTO(out, rc);
 437
 438         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 439         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 440         npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
 441         if (npages < 0)
 442                 GOTO(out, rc = npages);
 443
 444         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 445         if (local_nb == NULL)
 446                 GOTO(out_pp_rnb, rc = -ENOMEM);
 447
 448         desc = ptlrpc_prep_bulk_exp (req, npages,
 449                                      BULK_PUT_SOURCE, OST_BULK_PORTAL);
 450         if (desc == NULL)
 451                 GOTO(out_local, rc = -ENOMEM);
 452
 453         do_gettimeofday(&start);
 454         rc = obd_preprw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 455                         ioo, npages, pp_rnb, local_nb, &oti);
 456         ost_stime_record(req, &start, 0, 0);
 457         if (rc != 0)
 458                 GOTO(out_bulk, rc);
 459
 460         /* We're finishing using body->oa as an input variable */
 461         body->oa.o_valid = 0;
 462
 463         nob = 0;
 464         for (i = 0; i < npages; i++) {
 465                 int page_rc = local_nb[i].rc;
 466
 467                 if (page_rc < 0) {              /* error */
 468                         rc = page_rc;
 469                         break;
 470                 }
 471
 472                 LASSERT(page_rc <= pp_rnb[i].len);
 473                 nob += page_rc;
 474                 if (page_rc != 0) {             /* some data! */
 475                         LASSERT (local_nb[i].page != NULL);
 476                         ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 477                                               pp_rnb[i].offset & (PAGE_SIZE-1),
 478                                               page_rc);
 479                 }
 480
 481                 if (page_rc != pp_rnb[i].len) { /* short read */
 482                         /* All subsequent pages should be 0 */
 483                         while(++i < npages)
 484                                 LASSERT(local_nb[i].rc == 0);
 485                         break;
 486                 }
 487         }
 488
 489         if (rc == 0) {
 490                 rc = ptlrpc_start_bulk_transfer(desc);
 491                 if (rc == 0) {
 492                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 493                                           ost_bulk_timeout, desc);
 494                         rc = l_wait_event(desc->bd_waitq,
 495                                           !ptlrpc_bulk_active(desc), &lwi);
 496                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 497                         if (rc == -ETIMEDOUT) {
 498                                 DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT");
 499                                 ptlrpc_abort_bulk(desc);
 500                         } else if (!desc->bd_success ||
 501                                    desc->bd_nob_transferred != desc->bd_nob) {
 502                                 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
 503                                           desc->bd_success ?
 504                                           "truncated" : "network error on",
 505                                           desc->bd_nob_transferred,
 506                                           desc->bd_nob);
 507                                 /* XXX should this be a different errno? */
 508                                 rc = -ETIMEDOUT;
 509                         }
 510                 } else {
 511                         DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d\n", rc);
 512                 }
 513                 comms_error = rc != 0;
 514         }
 515
 516         ost_stime_record(req, &start, 0, 1);
 517         /* Must commit after prep above in all cases */
 518         rc = obd_commitrw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 519                           ioo, npages, local_nb, &oti, rc);
 520         ost_stime_record(req, &start, 0, 2);
 521
 522         if (rc == 0) {
 523                 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 524                 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 525
 526 #if CHECKSUM_BULK
 527                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 528                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 529 #endif
 530         }
 531
 532  out_bulk:
 533         ptlrpc_free_bulk(desc);
 534  out_local:
 535         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 536  out_pp_rnb:
 537         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 538  out:
 539         LASSERT(rc <= 0);
 540         if (rc == 0) {
 541                 req->rq_status = nob;
 542                 ptlrpc_reply(req);
 543         } else if (!comms_error) {
 544                 /* only reply if comms OK */
 545                 req->rq_status = rc;
 546                 ptlrpc_error(req);
 547         } else {
 548                 if (req->rq_reply_state != NULL) {
 549                         /* reply out callback would free */
 550                         lustre_free_reply_state (req->rq_reply_state);
 551                 }
 552                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 553                         CERROR("bulk IO comms error: "
 554                                "evicting %s@%s id %s\n",
 555                                req->rq_export->exp_client_uuid.uuid,
 556                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 557                                req->rq_peerstr);
 558                         ptlrpc_fail_export(req->rq_export);
 559                 } else {
 560                         CERROR("ignoring bulk IO comms error: "
 561                                "client reconnected %s@%s id %s\n",
 562                                req->rq_export->exp_client_uuid.uuid,
 563                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 564                                req->rq_peerstr);
 565                 }
 566         }
 567
 568         RETURN(rc);
 569 }
 570
 571 int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
 572 {
 573         struct ptlrpc_bulk_desc *desc;
 574         struct niobuf_remote    *remote_nb;
 575         struct niobuf_remote    *pp_rnb;
 576         struct niobuf_local     *local_nb;
 577         struct obd_ioobj        *ioo;
 578         struct ost_body         *body, *repbody;
 579         struct l_wait_info       lwi;
 580         __u32                   *rcs;
 581         int                      size[2] = { sizeof(*body) };
 582         int                      objcount, niocount, npages;
 583         int                      comms_error = 0;
 584         int                      rc, swab, i, j;
 585         struct timeval           start;
 586         ENTRY;
 587
 588         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
 589                 GOTO(out, rc = -EIO);
 590
 591         /* pause before transaction has been started */
 592         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 593                          (obd_timeout + 1) / 4);
 594
 595         swab = lustre_msg_swabbed(req->rq_reqmsg);
 596         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 597         if (body == NULL) {
 598                 CERROR("Missing/short ost_body\n");
 599                 GOTO(out, rc = -EFAULT);
 600         }
 601
 602         LASSERT_REQSWAB(req, 1);
 603         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 604         if (objcount == 0) {
 605                 CERROR("Missing/short ioobj\n");
 606                 GOTO(out, rc = -EFAULT);
 607         }
 608         ioo = lustre_msg_buf (req->rq_reqmsg, 1, objcount * sizeof(*ioo));
 609         LASSERT (ioo != NULL);
 610         for (niocount = i = 0; i < objcount; i++) {
 611                 if (swab)
 612                         lustre_swab_obd_ioobj (&ioo[i]);
 613                 if (ioo[i].ioo_bufcnt == 0) {
 614                         CERROR("ioo[%d] has zero bufcnt\n", i);
 615                         GOTO(out, rc = -EFAULT);
 616                 }
 617                 niocount += ioo[i].ioo_bufcnt;
 618         }
 619
 620         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 621                                        lustre_swab_niobuf_remote);
 622         if (remote_nb == NULL) {
 623                 CERROR("Missing/short niobuf\n");
 624                 GOTO(out, rc = -EFAULT);
 625         }
 626         if (swab) {                             /* swab the remaining niobufs */
 627                 for (i = 1; i < niocount; i++)
 628                         lustre_swab_niobuf_remote (&remote_nb[i]);
 629         }
 630
 631         size[1] = niocount * sizeof(*rcs);
 632         rc = lustre_pack_reply(req, 2, size, NULL);
 633         if (rc != 0)
 634                 GOTO(out, rc);
 635         rcs = lustre_msg_buf(req->rq_repmsg, 1, niocount * sizeof(*rcs));
 636
 637 #if 0
 638         /* Do snap options here*/
 639         rc = obd_do_cow(req->rq_export, ioo, objcount, remote_nb);
 640         if (rc)
 641                 GOTO(out, rc);
 642 #endif
 643
 644         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 645         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 646         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 647         if (npages < 0)
 648                 GOTO(out, rc = npages);
 649
 650         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 651         if (local_nb == NULL)
 652                 GOTO(out_pp_rnb, rc = -ENOMEM);
 653
 654         desc = ptlrpc_prep_bulk_exp (req, npages,
 655                                      BULK_GET_SINK, OST_BULK_PORTAL);
 656         if (desc == NULL)
 657                 GOTO(out_local, rc = -ENOMEM);
 658
 659         do_gettimeofday(&start);
 660         rc = obd_preprw(OBD_BRW_WRITE, req->rq_export, &body->oa, objcount,
 661                         ioo, npages, pp_rnb, local_nb, oti);
 662         ost_stime_record(req, &start, 1, 0);
 663         if (rc != 0)
 664                 GOTO(out_bulk, rc);
 665
 666         /* NB Having prepped, we must commit... */
 667
 668         for (i = 0; i < npages; i++)
 669                 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 670                                       pp_rnb[i].offset & (PAGE_SIZE - 1),
 671                                       pp_rnb[i].len);
 672
 673         rc = ptlrpc_start_bulk_transfer (desc);
 674         if (rc == 0) {
 675                 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 676                                   ost_bulk_timeout, desc);
 677                 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc),
 678                                   &lwi);
 679                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 680                 if (rc == -ETIMEDOUT) {
 681                         DEBUG_REQ(D_ERROR, req, "timeout on bulk GET");
 682                         ptlrpc_abort_bulk(desc);
 683                 } else if (!desc->bd_success ||
 684                            desc->bd_nob_transferred != desc->bd_nob) {
 685                         DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
 686                                   desc->bd_success ?
 687                                   "truncated" : "network error on",
 688                                   desc->bd_nob_transferred, desc->bd_nob);
 689                         /* XXX should this be a different errno? */
 690                         rc = -ETIMEDOUT;
 691                 }
 692         } else {
 693                 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d\n", rc);
 694         }
 695         comms_error = rc != 0;
 696
 697         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 698         memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 699
 700 #if CHECKSUM_BULK
 701         if (rc == 0 && (body->oa.o_valid & OBD_MD_FLCKSUM) != 0) {
 702                 static int cksum_counter;
 703                 obd_count client_cksum = body->oa.o_cksum;
 704                 obd_count cksum = ost_checksum_bulk(desc);
 705
 706                 if (client_cksum != cksum) {
 707                         CERROR("Bad checksum: client %x, server %x id %s\n",
 708                                client_cksum, cksum,
 709                                req->rq_peerstr);
 710                         cksum_counter = 1;
 711                         repbody->oa.o_cksum = cksum;
 712                 } else {
 713                         cksum_counter++;
 714                         if ((cksum_counter & (-cksum_counter)) == cksum_counter)
 715                                 CWARN("Checksum %u from NID %s: %x OK\n",
 716                                       cksum_counter, req->rq_peerstr, cksum);
 717                 }
 718         }
 719 #endif
 720         ost_stime_record(req, &start, 1, 1);
 721         /* Must commit after prep above in all cases */
 722         rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa,
 723                           objcount, ioo, npages, local_nb, oti, rc);
 724
 725         ost_stime_record(req, &start, 1, 2);
 726         if (rc == 0) {
 727 #if CHECKSUM_BULK
 728                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 729                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 730 #endif
 731                 /* set per-requested niobuf return codes */
 732                 for (i = j = 0; i < niocount; i++) {
 733                         int nob = remote_nb[i].len;
 734
 735                         rcs[i] = 0;
 736                         do {
 737                                 LASSERT(j < npages);
 738                                 if (local_nb[j].rc < 0)
 739                                         rcs[i] = local_nb[j].rc;
 740                                 nob -= pp_rnb[j].len;
 741                                 j++;
 742                         } while (nob > 0);
 743                         LASSERT(nob == 0);
 744                 }
 745                 LASSERT(j == npages);
 746         }
 747         /*XXX This write extents only for write-back cache extents*/
 748         rc = obd_write_extents(req->rq_export, ioo, objcount, niocount,
 749                                local_nb, rc);
 750  out_bulk:
 751         ptlrpc_free_bulk(desc);
 752  out_local:
 753         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 754  out_pp_rnb:
 755         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 756  out:
 757         if (rc == 0) {
 758                 oti_to_request(oti, req);
 759                 rc = ptlrpc_reply(req);
 760         } else if (!comms_error) {
 761                 /* Only reply if there was no comms problem with bulk */
 762                 req->rq_status = rc;
 763                 ptlrpc_error(req);
 764         } else {
 765                 if (req->rq_reply_state != NULL) {
 766                         /* reply out callback would free */
 767                         lustre_free_reply_state (req->rq_reply_state);
 768                 }
 769                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 770                         CERROR("%s: bulk IO comm error evicting %s@%s id %s\n",
 771                                req->rq_export->exp_obd->obd_name,
 772                                req->rq_export->exp_client_uuid.uuid,
 773                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 774                                req->rq_peerstr);
 775                         ptlrpc_fail_export(req->rq_export);
 776                 } else {
 777                         CERROR("ignoring bulk IO comms error: "
 778                                "client reconnected %s@%s id %s\n",
 779                                req->rq_export->exp_client_uuid.uuid,
 780                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 781                                req->rq_peerstr);
 782                 }
 783         }
 784         RETURN(rc);
 785 }
 786 EXPORT_SYMBOL(ost_brw_write);
 787
 788 static int ost_san_brw(struct ptlrpc_request *req, int cmd)
 789 {
 790         struct niobuf_remote *remote_nb, *res_nb, *pp_rnb;
 791         struct obd_ioobj *ioo;
 792         struct ost_body *body, *repbody;
 793         int rc, i, objcount, niocount, size[2] = {sizeof(*body)}, npages;
 794         int swab;
 795         ENTRY;
 796
 797         /* XXX not set to use latest protocol */
 798
 799         swab = lustre_msg_swabbed(req->rq_reqmsg);
 800         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 801         if (body == NULL) {
 802                 CERROR("Missing/short ost_body\n");
 803                 GOTO(out, rc = -EFAULT);
 804         }
 805
 806         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 807         if (ioo == NULL) {
 808                 CERROR("Missing/short ioobj\n");
 809                 GOTO(out, rc = -EFAULT);
 810         }
 811         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 812         niocount = ioo[0].ioo_bufcnt;
 813         for (i = 1; i < objcount; i++) {
 814                 if (swab)
 815                         lustre_swab_obd_ioobj (&ioo[i]);
 816                 niocount += ioo[i].ioo_bufcnt;
 817         }
 818
 819         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 820                                        lustre_swab_niobuf_remote);
 821         if (remote_nb == NULL) {
 822                 CERROR("Missing/short niobuf\n");
 823                 GOTO(out, rc = -EFAULT);
 824         }
 825         if (swab) {                             /* swab the remaining niobufs */
 826                 for (i = 1; i < niocount; i++)
 827                         lustre_swab_niobuf_remote (&remote_nb[i]);
 828         }
 829
 830         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 831         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 832         if (npages < 0)
 833                 GOTO (out, rc = npages);
 834
 835         size[1] = npages * sizeof(*pp_rnb);
 836         rc = lustre_pack_reply(req, 2, size, NULL);
 837         if (rc)
 838                 GOTO(out_pp_rnb, rc);
 839
 840         req->rq_status = obd_san_preprw(cmd, req->rq_export, &body->oa,
 841                                         objcount, ioo, npages, pp_rnb);
 842
 843         if (req->rq_status)
 844                 GOTO(out_pp_rnb, rc = 0);
 845
 846         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 847         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 848
 849         res_nb = lustre_msg_buf(req->rq_repmsg, 1, size[1]);
 850         memcpy(res_nb, remote_nb, size[1]);
 851         rc = 0;
 852 out_pp_rnb:
 853         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 854 out:
 855         if (rc) {
 856                 req->rq_status = rc;
 857                 ptlrpc_error(req);
 858         } else
 859                 ptlrpc_reply(req);
 860
 861         return rc;
 862 }
 863
 864 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
 865 {
 866         char *key, *val;
 867         int keylen, rc = 0;
 868         ENTRY;
 869
 870         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 871         if (key == NULL) {
 872                 DEBUG_REQ(D_HA, req, "no set_info key");
 873                 RETURN(-EFAULT);
 874         }
 875         keylen = req->rq_reqmsg->buflens[0];
 876
 877         rc = lustre_pack_reply(req, 0, NULL, NULL);
 878         if (rc)
 879                 RETURN(rc);
 880
 881         val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
 882
 883         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
 884                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
 885                                    lustre_swab_audit_msg);
 886         }
 887         else if (keylen == 5 && strcmp(key, "audit") == 0) {
 888                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
 889                                    lustre_swab_audit_attr);
 890         }
 891         else if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
 892                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
 893                                    lustre_swab_obdo);
 894         }
 895
 896         rc = obd_set_info(exp, keylen, key, req->rq_reqmsg->buflens[1], val);
 897         req->rq_repmsg->status = 0;
 898         RETURN(rc);
 899 }
 900
 901 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
 902 {
 903         char *key;
 904         int keylen, rc = 0, size = sizeof(obd_id);
 905         obd_id *reply;
 906         ENTRY;
 907
 908         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 909         if (key == NULL) {
 910                 DEBUG_REQ(D_HA, req, "no get_info key");
 911                 RETURN(-EFAULT);
 912         }
 913         keylen = req->rq_reqmsg->buflens[0];
 914
 915         if (keylen < strlen("last_id") || memcmp(key, "last_id", 7) != 0)
 916                 RETURN(-EPROTO);
 917
 918         rc = lustre_pack_reply(req, 1, &size, NULL);
 919         if (rc)
 920                 RETURN(rc);
 921
 922         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
 923         rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
 924         req->rq_repmsg->status = 0;
 925         RETURN(rc);
 926 }
 927
 928 static int ost_llog_handle_connect(struct obd_export *exp,
 929                                    struct ptlrpc_request *req)
 930 {
 931         struct llogd_conn_body *body;
 932         int rc;
 933         ENTRY;
 934
 935         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 936         rc = obd_llog_connect(exp, body);
 937         RETURN(rc);
 938 }
 939
 940 static int ost_filter_recovery_request(struct ptlrpc_request *req,
 941                                        struct obd_device *obd, int *process)
 942 {
 943         switch (req->rq_reqmsg->opc) {
 944         case OST_CONNECT: /* This will never get here, but for completeness. */
 945         case OST_DISCONNECT:
 946                *process = 1;
 947                RETURN(0);
 948
 949         case OBD_PING:
 950         case OST_CREATE:
 951         case OST_DESTROY:
 952         case OST_PUNCH:
 953         case OST_SETATTR:
 954         case OST_SYNC:
 955         case OST_WRITE:
 956         case OBD_LOG_CANCEL:
 957         case LDLM_ENQUEUE:
 958                 *process = target_queue_recovery_request(req, obd);
 959                 RETURN(0);
 960
 961         default:
 962                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 963                 *process = 0;
 964                 /* XXX what should we set rq_status to here? */
 965                 req->rq_status = -EAGAIN;
 966                 RETURN(ptlrpc_error(req));
 967         }
 968 }
 969
 970 int ost_msg_check_version(struct lustre_msg *msg)
 971 {
 972         int rc;
 973
 974         switch(msg->opc) {
 975         case OST_CONNECT:
 976         case OST_DISCONNECT:
 977         case OBD_PING:
 978         case OST_CREATE:
 979         case OST_DESTROY:
 980         case OST_GETATTR:
 981         case OST_SETATTR:
 982         case OST_WRITE:
 983         case OST_READ:
 984         case OST_SAN_READ:
 985         case OST_SAN_WRITE:
 986         case OST_PUNCH:
 987         case OST_STATFS:
 988         case OST_SYNC:
 989         case OST_SET_INFO:
 990         case OST_GET_INFO:
 991                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
 992                 if (rc)
 993                         CERROR("bad opc %u version %08x, expecting %08x\n",
 994                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
 995                 break;
 996         case LDLM_ENQUEUE:
 997         case LDLM_CONVERT:
 998         case LDLM_CANCEL:
 999         case LDLM_BL_CALLBACK:
1000         case LDLM_CP_CALLBACK:
1001                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1002                 if (rc)
1003                         CERROR("bad opc %u version %08x, expecting %08x\n",
1004                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1005                 break;
1006         case OBD_LOG_CANCEL:
1007         case LLOG_ORIGIN_CONNECT:
1008                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1009                 if (rc)
1010                         CERROR("bad opc %u version %08x, expecting %08x\n",
1011                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1012                 break;
1013         case SEC_INIT:
1014         case SEC_INIT_CONTINUE:
1015         case SEC_FINI:
1016                 rc = 0;
1017                 break;
1018         default:
1019                 CERROR("OST unexpected opcode %d\n", msg->opc);
1020                 rc = -ENOTSUPP;
1021                 break;
1022         }
1023         return rc;
1024 }
1025
1026 int ost_handle(struct ptlrpc_request *req)
1027 {
1028         int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1029         struct obd_trans_info *oti = NULL;
1030         struct obd_device *obd = NULL;
1031         ENTRY;
1032
1033         LASSERT(current->journal_info == NULL);
1034
1035         rc = ost_msg_check_version(req->rq_reqmsg);
1036         if (rc) {
1037                 CERROR("OST drop mal-formed request\n");
1038                 RETURN(rc);
1039         }
1040
1041         /* Security opc should NOT trigger any recovery events */
1042         if (req->rq_reqmsg->opc == SEC_INIT ||
1043             req->rq_reqmsg->opc == SEC_INIT_CONTINUE ||
1044             req->rq_reqmsg->opc == SEC_FINI) {
1045                 GOTO(out_check_req, rc = 0);
1046         }
1047
1048         /* XXX identical to MDS */
1049         if (req->rq_reqmsg->opc != OST_CONNECT) {
1050                 int recovering;
1051
1052                 if (req->rq_export == NULL) {
1053                         CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1054                                req->rq_reqmsg->opc,
1055                                req->rq_peerstr);
1056                         req->rq_status = -ENOTCONN;
1057                         GOTO(out_check_req, rc = -ENOTCONN);
1058                 }
1059
1060                 obd = req->rq_export->exp_obd;
1061
1062                 /* Check for aborted recovery. */
1063                 spin_lock_bh(&obd->obd_processing_task_lock);
1064                 recovering = obd->obd_recovering;
1065                 spin_unlock_bh(&obd->obd_processing_task_lock);
1066                 if (recovering) {
1067                         rc = ost_filter_recovery_request(req, obd,
1068                                                          &should_process);
1069                         if (rc || !should_process)
1070                                 RETURN(rc);
1071                         if (should_process < 0) {
1072                                 req->rq_status = should_process;
1073                                 rc = ptlrpc_error(req);
1074                                 RETURN(rc);
1075                         }
1076                 }
1077         }
1078
1079         OBD_ALLOC(oti, sizeof(*oti));
1080         if (oti == NULL)
1081                 RETURN(-ENOMEM);
1082
1083         oti_init(oti, req);
1084
1085         switch (req->rq_reqmsg->opc) {
1086         case OST_CONNECT: {
1087                 CDEBUG(D_INODE, "connect\n");
1088                 OBD_FAIL_GOTO(OBD_FAIL_OST_CONNECT_NET, out_free_oti, rc = 0);
1089                 rc = target_handle_connect(req);
1090                 if (!rc)
1091                         obd = req->rq_export->exp_obd;
1092                 break;
1093         }
1094         case OST_DISCONNECT:
1095                 CDEBUG(D_INODE, "disconnect\n");
1096                 OBD_FAIL_GOTO(OBD_FAIL_OST_DISCONNECT_NET, out_free_oti, rc = 0);
1097                 rc = target_handle_disconnect(req);
1098                 break;
1099         case OST_CREATE:
1100                 CDEBUG(D_INODE, "create\n");
1101                 OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC);
1102                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1103                 OBD_FAIL_GOTO(OBD_FAIL_OST_CREATE_NET, out_free_oti, rc = 0);
1104                 rc = ost_create(req->rq_export, req, oti);
1105                 break;
1106         case OST_DESTROY:
1107                 CDEBUG(D_INODE, "destroy\n");
1108                 OBD_FAIL_GOTO(OBD_FAIL_OST_DESTROY_NET, out_free_oti, rc = 0);
1109                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1110                 rc = ost_destroy(req->rq_export, req, oti);
1111                 break;
1112         case OST_GETATTR:
1113                 CDEBUG(D_INODE, "getattr\n");
1114                 OBD_FAIL_GOTO(OBD_FAIL_OST_GETATTR_NET, out_free_oti, rc = 0);
1115                 rc = ost_getattr(req->rq_export, req);
1116                 break;
1117         case OST_SETATTR:
1118                 CDEBUG(D_INODE, "setattr\n");
1119                 OBD_FAIL_GOTO(OBD_FAIL_OST_SETATTR_NET, out_free_oti, rc = 0);
1120                 rc = ost_setattr(req->rq_export, req, oti);
1121                 break;
1122         case OST_WRITE:
1123                 CDEBUG(D_INODE, "write\n");
1124                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1125                 OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC);
1126                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1127                 rc = ost_brw_write(req, oti);
1128                 LASSERT(current->journal_info == NULL);
1129                 /* ost_brw sends its own replies */
1130                 GOTO(out_free_oti, rc);
1131         case OST_READ:
1132                 CDEBUG(D_INODE, "read\n");
1133                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1134                 rc = ost_brw_read(req);
1135                 LASSERT(current->journal_info == NULL);
1136                 /* ost_brw sends its own replies */
1137                 GOTO(out_free_oti, rc);
1138         case OST_SAN_READ:
1139                 CDEBUG(D_INODE, "san read\n");
1140                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1141                 rc = ost_san_brw(req, OBD_BRW_READ);
1142                 /* ost_san_brw sends its own replies */
1143                 GOTO(out_free_oti, rc);
1144         case OST_SAN_WRITE:
1145                 CDEBUG(D_INODE, "san write\n");
1146                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1147                 rc = ost_san_brw(req, OBD_BRW_WRITE);
1148                 /* ost_san_brw sends its own replies */
1149                 GOTO(out_free_oti, rc);
1150         case OST_PUNCH:
1151                 CDEBUG(D_INODE, "punch\n");
1152                 OBD_FAIL_GOTO(OBD_FAIL_OST_PUNCH_NET, out_free_oti, rc = 0);
1153                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1154                 rc = ost_punch(req->rq_export, req, oti);
1155                 break;
1156         case OST_STATFS:
1157                 CDEBUG(D_INODE, "statfs\n");
1158                 OBD_FAIL_GOTO(OBD_FAIL_OST_STATFS_NET, out_free_oti, rc = 0);
1159                 rc = ost_statfs(req);
1160                 break;
1161         case OST_SYNC:
1162                 CDEBUG(D_INODE, "sync\n");
1163                 OBD_FAIL_GOTO(OBD_FAIL_OST_SYNC_NET, out_free_oti, rc = 0);
1164                 rc = ost_sync(req->rq_export, req);
1165                 break;
1166         case OST_SET_INFO:
1167                 DEBUG_REQ(D_INODE, req, "set_info");
1168                 rc = ost_set_info(req->rq_export, req);
1169                 break;
1170         case OST_GET_INFO:
1171                 DEBUG_REQ(D_INODE, req, "get_info");
1172                 rc = ost_get_info(req->rq_export, req);
1173                 break;
1174         case OBD_PING:
1175                 DEBUG_REQ(D_INODE, req, "ping");
1176                 rc = target_handle_ping(req);
1177                 break;
1178         /* FIXME - just reply status */
1179         case LLOG_ORIGIN_CONNECT:
1180                 DEBUG_REQ(D_INODE, req, "log connect\n");
1181                 rc = ost_llog_handle_connect(req->rq_export, req);
1182                 req->rq_status = rc;
1183                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1184                 if (rc)
1185                         GOTO(out_free_oti, rc);
1186                 GOTO(out_free_oti, rc = ptlrpc_reply(req));
1187         case OBD_LOG_CANCEL:
1188                 CDEBUG(D_INODE, "log cancel\n");
1189                 OBD_FAIL_GOTO(OBD_FAIL_OBD_LOG_CANCEL_NET, out_free_oti, rc = 0);
1190                 rc = llog_origin_handle_cancel(req);
1191                 req->rq_status = rc;
1192                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1193                 if (rc)
1194                         GOTO(out_free_oti, rc);
1195                 GOTO(out_free_oti, rc = ptlrpc_reply(req));
1196         case LDLM_ENQUEUE:
1197                 CDEBUG(D_INODE, "enqueue\n");
1198                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_ENQUEUE, out_free_oti, rc = 0);
1199                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1200                                          ldlm_server_blocking_ast,
1201                                          ldlm_server_glimpse_ast);
1202                 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1203                 break;
1204         case LDLM_CONVERT:
1205                 CDEBUG(D_INODE, "convert\n");
1206                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_CONVERT, out_free_oti, rc = 0);
1207                 rc = ldlm_handle_convert(req);
1208                 break;
1209         case LDLM_CANCEL:
1210                 CDEBUG(D_INODE, "cancel\n");
1211                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_CANCEL, out_free_oti, rc = 0);
1212                 rc = ldlm_handle_cancel(req);
1213                 break;
1214         case LDLM_BL_CALLBACK:
1215         case LDLM_CP_CALLBACK:
1216                 CDEBUG(D_INODE, "callback\n");
1217                 CERROR("callbacks should not happen on OST\n");
1218                 /* fall through */
1219         default:
1220                 CERROR("Unexpected opcode %d\n", req->rq_reqmsg->opc);
1221                 req->rq_status = -ENOTSUPP;
1222                 rc = ptlrpc_error(req);
1223                 GOTO(out_free_oti, rc);
1224         }
1225
1226         LASSERT(current->journal_info == NULL);
1227
1228         EXIT;
1229         /* If we're DISCONNECTing, the export_data is already freed */
1230         if (!rc && req->rq_reqmsg->opc != OST_DISCONNECT) {
1231                 if (!obd->obd_no_transno) {
1232                         req->rq_repmsg->last_committed =
1233                                 obd->obd_last_committed;
1234                 } else {
1235                         DEBUG_REQ(D_IOCTL, req,
1236                                   "not sending last_committed update");
1237                 }
1238                 CDEBUG(D_INFO, "last_committed "LPU64", xid "LPX64"\n",
1239                        obd->obd_last_committed, req->rq_xid);
1240         }
1241
1242 out_check_req:
1243
1244         if (!rc)
1245                 oti_to_request(oti, req);
1246         target_send_reply(req, rc, fail);
1247         rc = 0;
1248
1249 out_free_oti:
1250         if (oti)
1251                 OBD_FREE(oti, sizeof(*oti));
1252         return rc;
1253 }
1254 EXPORT_SYMBOL(ost_handle);
1255
1256 int ost_attach(struct obd_device *dev, obd_count len, void *data)
1257 {
1258         struct lprocfs_static_vars lvars;
1259
1260         lprocfs_init_vars(ost,&lvars);
1261         return lprocfs_obd_attach(dev, lvars.obd_vars);
1262 }
1263
1264 int ost_detach(struct obd_device *dev)
1265 {
1266         return lprocfs_obd_detach(dev);
1267 }
1268
1269 extern struct file_operations ost_stimes_fops;
1270
1271 static int ost_setup(struct obd_device *obd, obd_count len, void *buf)
1272 {
1273         struct ost_obd *ost = &obd->u.ost;
1274         int rc;
1275         ENTRY;
1276
1277         rc = cleanup_group_info();
1278         if (rc)
1279                 RETURN(rc);
1280
1281         rc = llog_start_commit_thread();
1282         if (rc < 0)
1283                 RETURN(rc);
1284
1285         lprocfs_obd_seq_create(obd, "service_times", 0444, &ost_stimes_fops,
1286                                obd);
1287
1288         ost->ost_service =
1289                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1290                                 OST_REQUEST_PORTAL, OSC_REPLY_PORTAL, 30000,
1291                                 ost_handle, "ost",
1292                                 obd->obd_proc_entry);
1293         if (ost->ost_service == NULL) {
1294                 CERROR("failed to start service\n");
1295                 RETURN(-ENOMEM);
1296         }
1297
1298         rc = ptlrpc_start_n_threads(obd, ost->ost_service, OST_NUM_THREADS,
1299                                     "ll_ost");
1300         if (rc)
1301                 GOTO(out_service, rc = -EINVAL);
1302
1303         ost->ost_create_service =
1304                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1305                                 OST_CREATE_PORTAL, OSC_REPLY_PORTAL, 30000,
1306                                 ost_handle, "ost_create",
1307                                 obd->obd_proc_entry);
1308         if (ost->ost_create_service == NULL) {
1309                 CERROR("failed to start OST create service\n");
1310                 GOTO(out_service, rc = -ENOMEM);
1311         }
1312
1313
1314         spin_lock_init(&ost->ost_lock);
1315         ost->ost_service->srv_obddev = obd;
1316
1317         rc = ptlrpc_start_n_threads(obd, ost->ost_create_service, 1,
1318                                     "ll_ost_creat");
1319         if (rc)
1320                 GOTO(out_create, rc = -EINVAL);
1321
1322         RETURN(0);
1323
1324 out_create:
1325         ptlrpc_unregister_service(ost->ost_create_service);
1326 out_service:
1327         ptlrpc_unregister_service(ost->ost_service);
1328         RETURN(rc);
1329 }
1330
1331 extern void lgss_svc_cache_purge_all(void);
1332 static int ost_cleanup(struct obd_device *obd, int flags)
1333 {
1334         struct ost_obd *ost = &obd->u.ost;
1335         int err = 0;
1336         ENTRY;
1337
1338         spin_lock_bh(&obd->obd_processing_task_lock);
1339         if (obd->obd_recovering) {
1340                 target_cancel_recovery_timer(obd);
1341                 obd->obd_recovering = 0;
1342         }
1343         spin_unlock_bh(&obd->obd_processing_task_lock);
1344
1345         ptlrpc_stop_all_threads(ost->ost_service);
1346         ptlrpc_unregister_service(ost->ost_service);
1347
1348         ptlrpc_stop_all_threads(ost->ost_create_service);
1349         ptlrpc_unregister_service(ost->ost_create_service);
1350
1351 #ifdef ENABLE_GSS
1352         /* XXX */
1353         lgss_svc_cache_purge_all();
1354 #endif
1355         RETURN(err);
1356 }
1357
1358 /* use obd ops to offer management infrastructure */
1359 static struct obd_ops ost_obd_ops = {
1360         .o_owner        = THIS_MODULE,
1361         .o_attach       = ost_attach,
1362         .o_detach       = ost_detach,
1363         .o_setup        = ost_setup,
1364         .o_cleanup      = ost_cleanup,
1365 };
1366
1367 static int __init ost_init(void)
1368 {
1369         struct lprocfs_static_vars lvars;
1370         ENTRY;
1371
1372         lprocfs_init_vars(ost,&lvars);
1373         RETURN(class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
1374                                    OBD_OST_DEVICENAME));
1375 }
1376
1377 static void /*__exit*/ ost_exit(void)
1378 {
1379         class_unregister_type(OBD_OST_DEVICENAME);
1380 }
1381
1382 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1383 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
1384 MODULE_LICENSE("GPL");
1385
1386 module_init(ost_init);
1387 module_exit(ost_exit);