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 <portals/list.h>
  47
  48 void oti_init(struct obd_trans_info *oti, struct ptlrpc_request *req)
  49 {
  50         if (oti == NULL)
  51                 return;
  52         memset(oti, 0, sizeof *oti);
  53
  54         if (req->rq_repmsg && req->rq_reqmsg != 0)
  55                 oti->oti_transno = req->rq_repmsg->transno;
  56 }
  57
  58 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
  59 {
  60         struct oti_req_ack_lock *ack_lock;
  61         int i;
  62
  63         if (oti == NULL)
  64                 return;
  65
  66         if (req->rq_repmsg)
  67                 req->rq_repmsg->transno = oti->oti_transno;
  68
  69         /* XXX 4 == entries in oti_ack_locks??? */
  70         for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
  71                 if (!ack_lock->mode)
  72                         break;
  73                 /* XXX not even calling target_send_reply in some cases... */
  74                 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
  75         }
  76 }
  77
  78 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
  79                        struct obd_trans_info *oti)
  80 {
  81         struct ost_body *body, *repbody;
  82         int rc, size = sizeof(*body);
  83         ENTRY;
  84
  85         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
  86         if (body == NULL)
  87                 RETURN(-EFAULT);
  88
  89         rc = lustre_pack_reply(req, 1, &size, NULL);
  90         if (rc)
  91                 RETURN(rc);
  92
  93         if (body->oa.o_valid & OBD_MD_FLCOOKIE)
  94                 oti->oti_logcookies = obdo_logcookie(&body->oa);
  95         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
  96         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
  97         req->rq_status = obd_destroy(exp, &body->oa, NULL, oti);
  98         RETURN(0);
  99 }
 100
 101 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
 102 {
 103         struct ost_body *body, *repbody;
 104         int rc, size = sizeof(*body);
 105         ENTRY;
 106
 107         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 108         if (body == NULL)
 109                 RETURN(-EFAULT);
 110
 111         rc = lustre_pack_reply(req, 1, &size, NULL);
 112         if (rc)
 113                 RETURN(rc);
 114
 115         repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody));
 116         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 117         req->rq_status = obd_getattr(exp, &repbody->oa, NULL);
 118         RETURN(0);
 119 }
 120
 121 static int ost_statfs(struct ptlrpc_request *req)
 122 {
 123         struct obd_statfs *osfs;
 124         int rc, size = sizeof(*osfs);
 125         ENTRY;
 126
 127         rc = lustre_pack_reply(req, 1, &size, NULL);
 128         if (rc)
 129                 RETURN(rc);
 130
 131         osfs = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*osfs));
 132
 133         req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, jiffies-HZ);
 134         if (req->rq_status != 0)
 135                 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
 136
 137         RETURN(0);
 138 }
 139
 140 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
 141                       struct obd_trans_info *oti)
 142 {
 143         struct ost_body *body, *repbody;
 144         int rc, size = sizeof(*repbody);
 145         ENTRY;
 146
 147         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 148         if (body == NULL)
 149                 RETURN(-EFAULT);
 150
 151         rc = lustre_pack_reply(req, 1, &size, NULL);
 152         if (rc)
 153                 RETURN(rc);
 154
 155         repbody = lustre_msg_buf (req->rq_repmsg, 0, sizeof(*repbody));
 156         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 157         oti->oti_logcookies = obdo_logcookie(&repbody->oa);
 158         req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
 159         //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
 160         RETURN(0);
 161 }
 162
 163 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
 164                      struct obd_trans_info *oti)
 165 {
 166         struct ost_body *body, *repbody;
 167         int rc, size = sizeof(*repbody);
 168         ENTRY;
 169
 170         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 171         if (body == NULL)
 172                 RETURN(-EFAULT);
 173
 174         if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
 175             (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
 176                 RETURN(-EINVAL);
 177
 178         rc = lustre_pack_reply(req, 1, &size, NULL);
 179         if (rc)
 180                 RETURN(rc);
 181
 182         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 183         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 184         req->rq_status = obd_punch(exp, &repbody->oa, NULL, repbody->oa.o_size,
 185                                    repbody->oa.o_blocks, oti);
 186         RETURN(0);
 187 }
 188
 189 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
 190 {
 191         struct ost_body *body, *repbody;
 192         int rc, size = sizeof(*repbody);
 193         ENTRY;
 194
 195         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 196         if (body == NULL)
 197                 RETURN(-EFAULT);
 198
 199         rc = lustre_pack_reply(req, 1, &size, NULL);
 200         if (rc)
 201                 RETURN(rc);
 202
 203         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 204         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 205         req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
 206                                   repbody->oa.o_blocks);
 207         RETURN(0);
 208 }
 209
 210 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
 211                        struct obd_trans_info *oti)
 212 {
 213         struct ost_body *body, *repbody;
 214         int rc, size = sizeof(*repbody);
 215         ENTRY;
 216
 217         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 218         if (body == NULL)
 219                 RETURN(-EFAULT);
 220
 221         rc = lustre_pack_reply(req, 1, &size, NULL);
 222         if (rc)
 223                 RETURN(rc);
 224
 225         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 226         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 227
 228         req->rq_status = obd_setattr(exp, &repbody->oa, NULL, oti);
 229         RETURN(0);
 230 }
 231
 232 static int ost_bulk_timeout(void *data)
 233 {
 234         ENTRY;
 235         /* We don't fail the connection here, because having the export
 236          * killed makes the (vital) call to commitrw very sad.
 237          */
 238         RETURN(1);
 239 }
 240
 241 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
 242                                 struct niobuf_remote *rnb, int nrnb,
 243                                 struct niobuf_remote **pp_rnbp)
 244 {
 245         /* Copy a remote niobuf, splitting it into page-sized chunks
 246          * and setting ioo[i].ioo_bufcnt accordingly */
 247         struct niobuf_remote *pp_rnb;
 248         int   i;
 249         int   j;
 250         int   page;
 251         int   rnbidx = 0;
 252         int   npages = 0;
 253
 254         /* first count and check the number of pages required */
 255         for (i = 0; i < nioo; i++)
 256                 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
 257                         obd_off offset = rnb[rnbidx].offset;
 258                         obd_off p0 = offset >> PAGE_SHIFT;
 259                         obd_off pn = (offset + rnb[rnbidx].len - 1)>>PAGE_SHIFT;
 260
 261                         LASSERT(rnbidx < nrnb);
 262
 263                         npages += (pn + 1 - p0);
 264
 265                         if (rnb[rnbidx].len == 0) {
 266                                 CERROR("zero len BRW: obj %d objid "LPX64
 267                                        " buf %u\n", i, ioo[i].ioo_id, j);
 268                                 return -EINVAL;
 269                         }
 270                         if (j > 0 &&
 271                             rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
 272                                 CERROR("unordered BRW: obj %d objid "LPX64
 273                                        " buf %u offset "LPX64" <= "LPX64"\n",
 274                                        i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
 275                                        rnb[rnbidx].offset);
 276                                 return -EINVAL;
 277                         }
 278                 }
 279
 280         LASSERT(rnbidx == nrnb);
 281
 282         if (npages == nrnb) {       /* all niobufs are for single pages */
 283                 *pp_rnbp = rnb;
 284                 return npages;
 285         }
 286
 287         OBD_ALLOC(pp_rnb, sizeof(*pp_rnb) * npages);
 288         if (pp_rnb == NULL)
 289                 return -ENOMEM;
 290
 291         /* now do the actual split */
 292         page = rnbidx = 0;
 293         for (i = 0; i < nioo; i++) {
 294                 int  obj_pages = 0;
 295
 296                 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
 297                         obd_off off = rnb[rnbidx].offset;
 298                         int     nob = rnb[rnbidx].len;
 299
 300                         LASSERT(rnbidx < nrnb);
 301                         do {
 302                                 obd_off  poff = off & (PAGE_SIZE - 1);
 303                                 int      pnob = (poff + nob > PAGE_SIZE) ?
 304                                                 PAGE_SIZE - poff : nob;
 305
 306                                 LASSERT(page < npages);
 307                                 pp_rnb[page].len = pnob;
 308                                 pp_rnb[page].offset = off;
 309                                 pp_rnb[page].flags = rnb[rnbidx].flags;
 310
 311                                 CDEBUG(0, "   obj %d id "LPX64
 312                                        "page %d(%d) "LPX64" for %d, flg %x\n",
 313                                        i, ioo[i].ioo_id, obj_pages, page,
 314                                        pp_rnb[page].offset, pp_rnb[page].len,
 315                                        pp_rnb[page].flags);
 316                                 page++;
 317                                 obj_pages++;
 318
 319                                 off += pnob;
 320                                 nob -= pnob;
 321                         } while (nob > 0);
 322                         LASSERT(nob == 0);
 323                 }
 324                 ioo[i].ioo_bufcnt = obj_pages;
 325         }
 326         LASSERT(page == npages);
 327
 328         *pp_rnbp = pp_rnb;
 329         return npages;
 330 }
 331
 332 static void free_per_page_niobufs (int npages, struct niobuf_remote *pp_rnb,
 333                                    struct niobuf_remote *rnb)
 334 {
 335         if (pp_rnb == rnb)                      /* didn't allocate above */
 336                 return;
 337
 338         OBD_FREE(pp_rnb, sizeof(*pp_rnb) * npages);
 339 }
 340
 341 #if CHECKSUM_BULK
 342 obd_count ost_checksum_bulk(struct ptlrpc_bulk_desc *desc)
 343 {
 344         obd_count cksum = 0;
 345         struct ptlrpc_bulk_page *bp;
 346
 347         list_for_each_entry(bp, &desc->bd_page_list, bp_link) {
 348                 ost_checksum(&cksum, kmap(bp->bp_page) + bp->bp_pageoffset,
 349                              bp->bp_buflen);
 350                 kunmap(bp->bp_page);
 351         }
 352
 353         return cksum;
 354 }
 355 #endif
 356
 357 static void ost_stime_record(struct ptlrpc_request *req, struct timeval *start,
 358                              unsigned rw, unsigned phase)
 359 {
 360         struct obd_device *obd = req->rq_svc->srv_obddev;
 361         struct timeval stop;
 362         int ind = rw *3 + phase;
 363
 364         if (obd && obd->obd_type && obd->obd_type->typ_name) {
 365                 if (!strcmp(obd->obd_type->typ_name, LUSTRE_OST_NAME)) {
 366                         struct ost_obd *ost = NULL;
 367
 368                         ost = &obd->u.ost;
 369                         if (ind >= (sizeof(ost->ost_stimes) /
 370                                     sizeof(ost->ost_stimes[0])))
 371                                return;
 372                         do_gettimeofday(&stop);
 373
 374                         spin_lock(&ost->ost_lock);
 375                         lprocfs_stime_record(&ost->ost_stimes[ind],&stop,start);
 376                         spin_unlock(&ost->ost_lock);
 377                         memcpy(start, &stop, sizeof(*start));
 378                 }
 379        }
 380 }
 381
 382 static char str[PTL_NALFMT_SIZE];
 383
 384
 385 static int ost_brw_read(struct ptlrpc_request *req)
 386 {
 387         struct ptlrpc_bulk_desc *desc;
 388         struct niobuf_remote    *remote_nb;
 389         struct niobuf_remote    *pp_rnb;
 390         struct niobuf_local     *local_nb;
 391         struct obd_ioobj        *ioo;
 392         struct ost_body         *body, *repbody;
 393         struct l_wait_info       lwi;
 394         struct obd_trans_info    oti = { 0 };
 395         int                      size[1] = { sizeof(*body) };
 396         int                      comms_error = 0;
 397         int                      niocount;
 398         int                      npages;
 399         int                      nob = 0;
 400         int                      rc;
 401         int                      i;
 402         struct timeval           start;
 403         ENTRY;
 404
 405         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
 406                 GOTO(out, rc = -EIO);
 407
 408         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 409                          (obd_timeout + 1) / 4);
 410
 411         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 412         if (body == NULL) {
 413                 CERROR("Missing/short ost_body\n");
 414                 GOTO(out, rc = -EFAULT);
 415         }
 416
 417         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 418         if (ioo == NULL) {
 419                 CERROR("Missing/short ioobj\n");
 420                 GOTO(out, rc = -EFAULT);
 421         }
 422
 423         niocount = ioo->ioo_bufcnt;
 424         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 425                                        lustre_swab_niobuf_remote);
 426         if (remote_nb == NULL) {
 427                 CERROR("Missing/short niobuf\n");
 428                 GOTO(out, rc = -EFAULT);
 429         }
 430         if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
 431                 for (i = 1; i < niocount; i++)
 432                         lustre_swab_niobuf_remote (&remote_nb[i]);
 433         }
 434
 435         rc = lustre_pack_reply(req, 1, size, NULL);
 436         if (rc)
 437                 GOTO(out, rc);
 438
 439         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 440         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 441         npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
 442         if (npages < 0)
 443                 GOTO(out, rc = npages);
 444
 445         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 446         if (local_nb == NULL)
 447                 GOTO(out_pp_rnb, rc = -ENOMEM);
 448
 449         desc = ptlrpc_prep_bulk_exp (req, npages,
 450                                      BULK_PUT_SOURCE, OST_BULK_PORTAL);
 451         if (desc == NULL)
 452                 GOTO(out_local, rc = -ENOMEM);
 453
 454         do_gettimeofday(&start);
 455         rc = obd_preprw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 456                         ioo, npages, pp_rnb, local_nb, &oti);
 457         ost_stime_record(req, &start, 0, 0);
 458         if (rc != 0)
 459                 GOTO(out_bulk, rc);
 460
 461         nob = 0;
 462         for (i = 0; i < npages; i++) {
 463                 int page_rc = local_nb[i].rc;
 464
 465                 if (page_rc < 0) {              /* error */
 466                         rc = page_rc;
 467                         break;
 468                 }
 469
 470                 LASSERT(page_rc <= pp_rnb[i].len);
 471                 nob += page_rc;
 472                 if (page_rc != 0) {             /* some data! */
 473                         LASSERT (local_nb[i].page != NULL);
 474                         ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 475                                               pp_rnb[i].offset & (PAGE_SIZE-1),
 476                                               page_rc);
 477                 }
 478
 479                 if (page_rc != pp_rnb[i].len) { /* short read */
 480                         /* All subsequent pages should be 0 */
 481                         while(++i < npages)
 482                                 LASSERT(local_nb[i].rc == 0);
 483                         break;
 484                 }
 485         }
 486
 487         if (rc == 0) {
 488                 rc = ptlrpc_start_bulk_transfer(desc);
 489                 if (rc == 0) {
 490                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 491                                           ost_bulk_timeout, desc);
 492                         rc = l_wait_event(desc->bd_waitq,
 493                                           !ptlrpc_bulk_active(desc), &lwi);
 494                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 495                         if (rc == -ETIMEDOUT) {
 496                                 DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT");
 497                                 ptlrpc_abort_bulk(desc);
 498                         } else if (!desc->bd_success ||
 499                                    desc->bd_nob_transferred != desc->bd_nob) {
 500                                 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
 501                                           desc->bd_success ?
 502                                           "truncated" : "network error on",
 503                                           desc->bd_nob_transferred,
 504                                           desc->bd_nob);
 505                                 /* XXX should this be a different errno? */
 506                                 rc = -ETIMEDOUT;
 507                         }
 508                 } else {
 509                         DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d\n", rc);
 510                 }
 511                 comms_error = rc != 0;
 512         }
 513
 514         ost_stime_record(req, &start, 0, 1);
 515         /* Must commit after prep above in all cases */
 516         rc = obd_commitrw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 517                           ioo, npages, local_nb, &oti, rc);
 518         ost_stime_record(req, &start, 0, 2);
 519
 520         if (rc == 0) {
 521                 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 522                 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 523
 524 #if CHECKSUM_BULK
 525                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 526                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 527 #endif
 528         }
 529
 530  out_bulk:
 531         ptlrpc_free_bulk(desc);
 532  out_local:
 533         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 534  out_pp_rnb:
 535         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 536  out:
 537         LASSERT(rc <= 0);
 538         if (rc == 0) {
 539                 req->rq_status = nob;
 540                 ptlrpc_reply(req);
 541         } else if (!comms_error) {
 542                 /* only reply if comms OK */
 543                 req->rq_status = rc;
 544                 ptlrpc_error(req);
 545         } else {
 546                 if (req->rq_reply_state != NULL) {
 547                         /* reply out callback would free */
 548                         lustre_free_reply_state (req->rq_reply_state);
 549                 }
 550                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 551                         CERROR("bulk IO comms error: "
 552                                "evicting %s@%s nid %s\n",
 553                                req->rq_export->exp_client_uuid.uuid,
 554                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 555                                ptlrpc_peernid2str(&req->rq_peer, str));
 556                         ptlrpc_fail_export(req->rq_export);
 557                 } else {
 558                         CERROR("ignoring bulk IO comms error: "
 559                                "client reconnected %s@%s nid %s\n",
 560                                req->rq_export->exp_client_uuid.uuid,
 561                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 562                                ptlrpc_peernid2str(&req->rq_peer, str));
 563                 }
 564         }
 565
 566         RETURN(rc);
 567 }
 568
 569 int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
 570 {
 571         struct ptlrpc_bulk_desc *desc;
 572         struct niobuf_remote    *remote_nb;
 573         struct niobuf_remote    *pp_rnb;
 574         struct niobuf_local     *local_nb;
 575         struct obd_ioobj        *ioo;
 576         struct ost_body         *body, *repbody;
 577         struct l_wait_info       lwi;
 578         __u32                   *rcs;
 579         int                      size[2] = { sizeof(*body) };
 580         int                      objcount, niocount, npages;
 581         int                      comms_error = 0;
 582         int                      rc, swab, i, j;
 583         struct timeval           start;
 584         ENTRY;
 585
 586         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
 587                 GOTO(out, rc = -EIO);
 588
 589         /* pause before transaction has been started */
 590         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 591                          (obd_timeout + 1) / 4);
 592
 593         swab = lustre_msg_swabbed(req->rq_reqmsg);
 594         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 595         if (body == NULL) {
 596                 CERROR("Missing/short ost_body\n");
 597                 GOTO(out, rc = -EFAULT);
 598         }
 599
 600         LASSERT_REQSWAB(req, 1);
 601         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 602         if (objcount == 0) {
 603                 CERROR("Missing/short ioobj\n");
 604                 GOTO(out, rc = -EFAULT);
 605         }
 606         ioo = lustre_msg_buf (req->rq_reqmsg, 1, objcount * sizeof(*ioo));
 607         LASSERT (ioo != NULL);
 608         for (niocount = i = 0; i < objcount; i++) {
 609                 if (swab)
 610                         lustre_swab_obd_ioobj (&ioo[i]);
 611                 if (ioo[i].ioo_bufcnt == 0) {
 612                         CERROR("ioo[%d] has zero bufcnt\n", i);
 613                         GOTO(out, rc = -EFAULT);
 614                 }
 615                 niocount += ioo[i].ioo_bufcnt;
 616         }
 617
 618         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 619                                        lustre_swab_niobuf_remote);
 620         if (remote_nb == NULL) {
 621                 CERROR("Missing/short niobuf\n");
 622                 GOTO(out, rc = -EFAULT);
 623         }
 624         if (swab) {                             /* swab the remaining niobufs */
 625                 for (i = 1; i < niocount; i++)
 626                         lustre_swab_niobuf_remote (&remote_nb[i]);
 627         }
 628
 629         size[1] = niocount * sizeof(*rcs);
 630         rc = lustre_pack_reply(req, 2, size, NULL);
 631         if (rc != 0)
 632                 GOTO(out, rc);
 633         rcs = lustre_msg_buf(req->rq_repmsg, 1, niocount * sizeof(*rcs));
 634
 635         /* Do snap options here*/
 636         rc = obd_do_cow(req->rq_export, ioo, objcount, remote_nb);
 637         if (rc)
 638                 GOTO(out, rc);
 639
 640         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 641         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 642         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 643         if (npages < 0)
 644                 GOTO(out, rc = npages);
 645
 646         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 647         if (local_nb == NULL)
 648                 GOTO(out_pp_rnb, rc = -ENOMEM);
 649
 650         desc = ptlrpc_prep_bulk_exp (req, npages,
 651                                      BULK_GET_SINK, OST_BULK_PORTAL);
 652         if (desc == NULL)
 653                 GOTO(out_local, rc = -ENOMEM);
 654
 655         do_gettimeofday(&start);
 656         rc = obd_preprw(OBD_BRW_WRITE, req->rq_export, &body->oa, objcount,
 657                         ioo, npages, pp_rnb, local_nb, oti);
 658         ost_stime_record(req, &start, 1, 0);
 659         if (rc != 0)
 660                 GOTO(out_bulk, rc);
 661
 662         /* NB Having prepped, we must commit... */
 663
 664         for (i = 0; i < npages; i++)
 665                 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 666                                       pp_rnb[i].offset & (PAGE_SIZE - 1),
 667                                       pp_rnb[i].len);
 668
 669         rc = ptlrpc_start_bulk_transfer (desc);
 670         if (rc == 0) {
 671                 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 672                                   ost_bulk_timeout, desc);
 673                 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc),
 674                                   &lwi);
 675                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 676                 if (rc == -ETIMEDOUT) {
 677                         DEBUG_REQ(D_ERROR, req, "timeout on bulk GET");
 678                         ptlrpc_abort_bulk(desc);
 679                 } else if (!desc->bd_success ||
 680                            desc->bd_nob_transferred != desc->bd_nob) {
 681                         DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
 682                                   desc->bd_success ?
 683                                   "truncated" : "network error on",
 684                                   desc->bd_nob_transferred, desc->bd_nob);
 685                         /* XXX should this be a different errno? */
 686                         rc = -ETIMEDOUT;
 687                 }
 688         } else {
 689                 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d\n", rc);
 690         }
 691         comms_error = rc != 0;
 692
 693         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 694         memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 695
 696 #if CHECKSUM_BULK
 697         if (rc == 0 && (body->oa.o_valid & OBD_MD_FLCKSUM) != 0) {
 698                 static int cksum_counter;
 699                 obd_count client_cksum = body->oa.o_cksum;
 700                 obd_count cksum = ost_checksum_bulk(desc);
 701
 702                 if (client_cksum != cksum) {
 703                         CERROR("Bad checksum: client %x, server %x NID %s\n",
 704                                client_cksum, cksum,
 705                                ptlrpc_peernid2str(&req->rq_peer, str));
 706                         cksum_counter = 1;
 707                         repbody->oa.o_cksum = cksum;
 708                 } else {
 709                         cksum_counter++;
 710                         if ((cksum_counter & (-cksum_counter)) == cksum_counter)
 711                                 CWARN("Checksum %u from NID %s: %x OK\n",
 712                                       cksum_counter,
 713                                       ptlrpc_peernid2str(&req->rq_peer, str),
 714                                       cksum);
 715                 }
 716         }
 717 #endif
 718         ost_stime_record(req, &start, 1, 1);
 719         /* Must commit after prep above in all cases */
 720         rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa,
 721                           objcount, ioo, npages, local_nb, oti, rc);
 722
 723         ost_stime_record(req, &start, 1, 2);
 724         if (rc == 0) {
 725                 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 726                 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 727
 728 #if CHECKSUM_BULK
 729                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 730                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 731 #endif
 732                 /* set per-requested niobuf return codes */
 733                 for (i = j = 0; i < niocount; i++) {
 734                         int nob = remote_nb[i].len;
 735
 736                         rcs[i] = 0;
 737                         do {
 738                                 LASSERT(j < npages);
 739                                 if (local_nb[j].rc < 0)
 740                                         rcs[i] = local_nb[j].rc;
 741                                 nob -= pp_rnb[j].len;
 742                                 j++;
 743                         } while (nob > 0);
 744                         LASSERT(nob == 0);
 745                 }
 746                 LASSERT(j == npages);
 747         }
 748         rc = obd_write_extents(req->rq_export, ioo, objcount, niocount,
 749                                local_nb, rc);
 750         if (rc) {
 751                 CERROR("write extents error of id "LPU64" rc=%d\n",
 752                         ioo->ioo_id, rc);
 753         }
 754  out_bulk:
 755         ptlrpc_free_bulk(desc);
 756  out_local:
 757         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 758  out_pp_rnb:
 759         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 760  out:
 761         if (rc == 0) {
 762                 oti_to_request(oti, req);
 763                 rc = ptlrpc_reply(req);
 764         } else if (!comms_error) {
 765                 /* Only reply if there was no comms problem with bulk */
 766                 req->rq_status = rc;
 767                 ptlrpc_error(req);
 768         } else {
 769                 if (req->rq_reply_state != NULL) {
 770                         /* reply out callback would free */
 771                         lustre_free_reply_state (req->rq_reply_state);
 772                 }
 773                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 774                         CERROR("%s: bulk IO comm error evicting %s@%s NID %s\n",
 775                                req->rq_export->exp_obd->obd_name,
 776                                req->rq_export->exp_client_uuid.uuid,
 777                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 778                                ptlrpc_peernid2str(&req->rq_peer, str));
 779                         ptlrpc_fail_export(req->rq_export);
 780                 } else {
 781                         CERROR("ignoring bulk IO comms error: "
 782                                "client reconnected %s@%s nid %s\n",
 783                                req->rq_export->exp_client_uuid.uuid,
 784                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 785                                ptlrpc_peernid2str(&req->rq_peer, str));
 786                 }
 787         }
 788         RETURN(rc);
 789 }
 790 EXPORT_SYMBOL(ost_brw_write);
 791
 792 static int ost_san_brw(struct ptlrpc_request *req, int cmd)
 793 {
 794         struct niobuf_remote *remote_nb, *res_nb, *pp_rnb;
 795         struct obd_ioobj *ioo;
 796         struct ost_body *body, *repbody;
 797         int rc, i, objcount, niocount, size[2] = {sizeof(*body)}, npages;
 798         int swab;
 799         ENTRY;
 800
 801         /* XXX not set to use latest protocol */
 802
 803         swab = lustre_msg_swabbed(req->rq_reqmsg);
 804         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 805         if (body == NULL) {
 806                 CERROR("Missing/short ost_body\n");
 807                 GOTO(out, rc = -EFAULT);
 808         }
 809
 810         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 811         if (ioo == NULL) {
 812                 CERROR("Missing/short ioobj\n");
 813                 GOTO(out, rc = -EFAULT);
 814         }
 815         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 816         niocount = ioo[0].ioo_bufcnt;
 817         for (i = 1; i < objcount; i++) {
 818                 if (swab)
 819                         lustre_swab_obd_ioobj (&ioo[i]);
 820                 niocount += ioo[i].ioo_bufcnt;
 821         }
 822
 823         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 824                                        lustre_swab_niobuf_remote);
 825         if (remote_nb == NULL) {
 826                 CERROR("Missing/short niobuf\n");
 827                 GOTO(out, rc = -EFAULT);
 828         }
 829         if (swab) {                             /* swab the remaining niobufs */
 830                 for (i = 1; i < niocount; i++)
 831                         lustre_swab_niobuf_remote (&remote_nb[i]);
 832         }
 833
 834         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 835         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 836         if (npages < 0)
 837                 GOTO (out, rc = npages);
 838
 839         size[1] = npages * sizeof(*pp_rnb);
 840         rc = lustre_pack_reply(req, 2, size, NULL);
 841         if (rc)
 842                 GOTO(out_pp_rnb, rc);
 843
 844         req->rq_status = obd_san_preprw(cmd, req->rq_export, &body->oa,
 845                                         objcount, ioo, npages, pp_rnb);
 846
 847         if (req->rq_status)
 848                 GOTO(out_pp_rnb, rc = 0);
 849
 850         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 851         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 852
 853         res_nb = lustre_msg_buf(req->rq_repmsg, 1, size[1]);
 854         memcpy(res_nb, remote_nb, size[1]);
 855         rc = 0;
 856 out_pp_rnb:
 857         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 858 out:
 859         if (rc) {
 860                 req->rq_status = rc;
 861                 ptlrpc_error(req);
 862         } else
 863                 ptlrpc_reply(req);
 864
 865         return rc;
 866 }
 867
 868 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
 869 {
 870         char *key, *val;
 871         int keylen, rc = 0;
 872         ENTRY;
 873
 874         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 875         if (key == NULL) {
 876                 DEBUG_REQ(D_HA, req, "no set_info key");
 877                 RETURN(-EFAULT);
 878         }
 879         keylen = req->rq_reqmsg->buflens[0];
 880
 881         rc = lustre_pack_reply(req, 0, NULL, NULL);
 882         if (rc)
 883                 RETURN(rc);
 884
 885         val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
 886
 887         rc = obd_set_info(exp, keylen, key, req->rq_reqmsg->buflens[1], val);
 888         req->rq_repmsg->status = 0;
 889         RETURN(rc);
 890 }
 891
 892 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
 893 {
 894         char *key;
 895         int keylen, rc = 0, size = sizeof(obd_id);
 896         obd_id *reply;
 897         ENTRY;
 898
 899         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 900         if (key == NULL) {
 901                 DEBUG_REQ(D_HA, req, "no get_info key");
 902                 RETURN(-EFAULT);
 903         }
 904         keylen = req->rq_reqmsg->buflens[0];
 905
 906         if (keylen < strlen("last_id") || memcmp(key, "last_id", 7) != 0)
 907                 RETURN(-EPROTO);
 908
 909         rc = lustre_pack_reply(req, 1, &size, NULL);
 910         if (rc)
 911                 RETURN(rc);
 912
 913         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
 914         rc = obd_get_info(exp, keylen, key, &size, reply);
 915         req->rq_repmsg->status = 0;
 916         RETURN(rc);
 917 }
 918
 919 static int ost_llog_handle_connect(struct obd_export *exp,
 920                 struct ptlrpc_request *req)
 921 {
 922         struct llogd_conn_body *body;
 923         int rc;
 924         ENTRY;
 925
 926         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 927         rc = obd_llog_connect(exp, body);
 928         RETURN(rc);
 929 }
 930
 931 static int ost_filter_recovery_request(struct ptlrpc_request *req,
 932                                        struct obd_device *obd, int *process)
 933 {
 934         switch (req->rq_reqmsg->opc) {
 935         case OST_CONNECT: /* This will never get here, but for completeness. */
 936         case OST_DISCONNECT:
 937                *process = 1;
 938                RETURN(0);
 939
 940         case OBD_PING:
 941         case OST_CREATE:
 942         case OST_DESTROY:
 943         case OST_PUNCH:
 944         case OST_SETATTR:
 945         case OST_SYNC:
 946         case OST_WRITE:
 947         case OBD_LOG_CANCEL:
 948         case LDLM_ENQUEUE:
 949                 *process = target_queue_recovery_request(req, obd);
 950                 RETURN(0);
 951
 952         default:
 953                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 954                 *process = 0;
 955                 /* XXX what should we set rq_status to here? */
 956                 req->rq_status = -EAGAIN;
 957                 RETURN(ptlrpc_error(req));
 958         }
 959 }
 960
 961 int ost_msg_check_version(struct lustre_msg *msg)
 962 {
 963         int rc;
 964
 965         switch(msg->opc) {
 966         case OST_CONNECT:
 967         case OST_DISCONNECT:
 968         case OBD_PING:
 969                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
 970                 if (rc)
 971                         CERROR("bad opc %u version %08x, expecting %08x\n",
 972                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
 973                 break;
 974         case OST_CREATE:
 975         case OST_DESTROY:
 976         case OST_GETATTR:
 977         case OST_SETATTR:
 978         case OST_WRITE:
 979         case OST_READ:
 980         case OST_SAN_READ:
 981         case OST_SAN_WRITE:
 982         case OST_PUNCH:
 983         case OST_STATFS:
 984         case OST_SYNC:
 985         case OST_SET_INFO:
 986         case OST_GET_INFO:
 987                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
 988                 if (rc)
 989                         CERROR("bad opc %u version %08x, expecting %08x\n",
 990                                msg->opc, msg->version, LUSTRE_OST_VERSION);
 991                 break;
 992         case LDLM_ENQUEUE:
 993         case LDLM_CONVERT:
 994         case LDLM_CANCEL:
 995         case LDLM_BL_CALLBACK:
 996         case LDLM_CP_CALLBACK:
 997                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
 998                 if (rc)
 999                         CERROR("bad opc %u version %08x, expecting %08x\n",
1000                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1001                 break;
1002         case OBD_LOG_CANCEL:
1003         case LLOG_ORIGIN_CONNECT:
1004                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1005                 if (rc)
1006                         CERROR("bad opc %u version %08x, expecting %08x\n",
1007                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1008                 break;
1009         default:
1010                 CERROR("OST unexpected opcode %d\n", msg->opc);
1011                 rc = -ENOTSUPP;
1012                 break;
1013         }
1014         return rc;
1015 }
1016
1017 int ost_handle(struct ptlrpc_request *req)
1018 {
1019         struct obd_trans_info trans_info = { 0, };
1020         struct obd_trans_info *oti = &trans_info;
1021         int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1022         struct obd_export *exp = NULL;
1023         ENTRY;
1024
1025         LASSERT(current->journal_info == NULL);
1026
1027         rc = ost_msg_check_version(req->rq_reqmsg);
1028         if (rc) {
1029                 CERROR("OST drop mal-formed request\n");
1030                 RETURN(rc);
1031         }
1032
1033         /* XXX identical to MDS */
1034         if (req->rq_reqmsg->opc != OST_CONNECT) {
1035                 struct obd_device *obd;
1036                 int recovering;
1037
1038                 exp = req->rq_export;
1039
1040                 if (exp == NULL) {
1041                         CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1042                                req->rq_reqmsg->opc,
1043                                ptlrpc_peernid2str(&req->rq_peer, str));
1044                         req->rq_status = -ENOTCONN;
1045                         GOTO(out, rc = -ENOTCONN);
1046                 }
1047
1048                 obd = exp->exp_obd;
1049
1050                 /* Check for aborted recovery. */
1051                 spin_lock_bh(&obd->obd_processing_task_lock);
1052                 recovering = obd->obd_recovering;
1053                 spin_unlock_bh(&obd->obd_processing_task_lock);
1054                 if (recovering) {
1055                         rc = ost_filter_recovery_request(req, obd,
1056                                                          &should_process);
1057                         if (rc || !should_process)
1058                                 RETURN(rc);
1059                         if (should_process < 0) {
1060                                 req->rq_status = should_process;
1061                                 rc = ptlrpc_error(req);
1062                                 RETURN(rc);
1063                         }
1064                 }
1065         }
1066
1067         oti_init(oti, req);
1068
1069         switch (req->rq_reqmsg->opc) {
1070         case OST_CONNECT: {
1071                 CDEBUG(D_INODE, "connect\n");
1072                 OBD_FAIL_RETURN(OBD_FAIL_OST_CONNECT_NET, 0);
1073                 rc = target_handle_connect(req);
1074                 break;
1075         }
1076         case OST_DISCONNECT:
1077                 CDEBUG(D_INODE, "disconnect\n");
1078                 OBD_FAIL_RETURN(OBD_FAIL_OST_DISCONNECT_NET, 0);
1079                 rc = target_handle_disconnect(req);
1080                 break;
1081         case OST_CREATE:
1082                 CDEBUG(D_INODE, "create\n");
1083                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_ENOSPC))
1084                         GOTO(out, rc = -ENOSPC);
1085                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS))
1086                         GOTO(out, rc = -EROFS);
1087                 OBD_FAIL_RETURN(OBD_FAIL_OST_CREATE_NET, 0);
1088                 rc = ost_create(exp, req, oti);
1089                 break;
1090         case OST_DESTROY:
1091                 CDEBUG(D_INODE, "destroy\n");
1092                 OBD_FAIL_RETURN(OBD_FAIL_OST_DESTROY_NET, 0);
1093                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS))
1094                         GOTO(out, rc = -EROFS);
1095                 rc = ost_destroy(exp, req, oti);
1096                 break;
1097         case OST_GETATTR:
1098                 CDEBUG(D_INODE, "getattr\n");
1099                 OBD_FAIL_RETURN(OBD_FAIL_OST_GETATTR_NET, 0);
1100                 rc = ost_getattr(exp, req);
1101                 break;
1102         case OST_SETATTR:
1103                 CDEBUG(D_INODE, "setattr\n");
1104                 OBD_FAIL_RETURN(OBD_FAIL_OST_SETATTR_NET, 0);
1105                 rc = ost_setattr(exp, req, oti);
1106                 break;
1107         case OST_WRITE:
1108                 CDEBUG(D_INODE, "write\n");
1109                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1110                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_ENOSPC))
1111                         GOTO(out, rc = -ENOSPC);
1112                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS))
1113                         GOTO(out, rc = -EROFS);
1114                 rc = ost_brw_write(req, oti);
1115                 LASSERT(current->journal_info == NULL);
1116                 /* ost_brw sends its own replies */
1117                 RETURN(rc);
1118         case OST_READ:
1119                 CDEBUG(D_INODE, "read\n");
1120                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1121                 rc = ost_brw_read(req);
1122                 LASSERT(current->journal_info == NULL);
1123                 /* ost_brw sends its own replies */
1124                 RETURN(rc);
1125         case OST_SAN_READ:
1126                 CDEBUG(D_INODE, "san read\n");
1127                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1128                 rc = ost_san_brw(req, OBD_BRW_READ);
1129                 /* ost_san_brw sends its own replies */
1130                 RETURN(rc);
1131         case OST_SAN_WRITE:
1132                 CDEBUG(D_INODE, "san write\n");
1133                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1134                 rc = ost_san_brw(req, OBD_BRW_WRITE);
1135                 /* ost_san_brw sends its own replies */
1136                 RETURN(rc);
1137         case OST_PUNCH:
1138                 CDEBUG(D_INODE, "punch\n");
1139                 OBD_FAIL_RETURN(OBD_FAIL_OST_PUNCH_NET, 0);
1140                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_OST_EROFS))
1141                         GOTO(out, rc = -EROFS);
1142                 rc = ost_punch(exp, req, oti);
1143                 break;
1144         case OST_STATFS:
1145                 CDEBUG(D_INODE, "statfs\n");
1146                 OBD_FAIL_RETURN(OBD_FAIL_OST_STATFS_NET, 0);
1147                 rc = ost_statfs(req);
1148                 break;
1149         case OST_SYNC:
1150                 CDEBUG(D_INODE, "sync\n");
1151                 OBD_FAIL_RETURN(OBD_FAIL_OST_SYNC_NET, 0);
1152                 rc = ost_sync(exp, req);
1153                 break;
1154         case OST_SET_INFO:
1155                 DEBUG_REQ(D_INODE, req, "set_info");
1156                 rc = ost_set_info(exp, req);
1157                 break;
1158         case OST_GET_INFO:
1159                 DEBUG_REQ(D_INODE, req, "get_info");
1160                 rc = ost_get_info(exp, req);
1161                 break;
1162         case OBD_PING:
1163                 DEBUG_REQ(D_INODE, req, "ping");
1164                 rc = target_handle_ping(req);
1165                 break;
1166         /* FIXME - just reply status */
1167         case LLOG_ORIGIN_CONNECT:
1168                 DEBUG_REQ(D_INODE, req, "log connect\n");
1169                 rc = ost_llog_handle_connect(exp, req);
1170                 req->rq_status = rc;
1171                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1172                 if (rc)
1173                         RETURN(rc);
1174                 RETURN(ptlrpc_reply(req));
1175         case OBD_LOG_CANCEL:
1176                 CDEBUG(D_INODE, "log cancel\n");
1177                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1178                 rc = llog_origin_handle_cancel(req);
1179                 req->rq_status = rc;
1180                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1181                 if (rc)
1182                         RETURN(rc);
1183                 RETURN(ptlrpc_reply(req));
1184         case LDLM_ENQUEUE:
1185                 CDEBUG(D_INODE, "enqueue\n");
1186                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1187                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1188                                          ldlm_server_blocking_ast,
1189                                          ldlm_server_glimpse_ast);
1190                 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1191                 break;
1192         case LDLM_CONVERT:
1193                 CDEBUG(D_INODE, "convert\n");
1194                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1195                 rc = ldlm_handle_convert(req);
1196                 break;
1197         case LDLM_CANCEL:
1198                 CDEBUG(D_INODE, "cancel\n");
1199                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1200                 rc = ldlm_handle_cancel(req);
1201                 break;
1202         case LDLM_BL_CALLBACK:
1203         case LDLM_CP_CALLBACK:
1204                 CDEBUG(D_INODE, "callback\n");
1205                 CERROR("callbacks should not happen on OST\n");
1206                 /* fall through */
1207         default:
1208                 CERROR("Unexpected opcode %d\n", req->rq_reqmsg->opc);
1209                 req->rq_status = -ENOTSUPP;
1210                 rc = ptlrpc_error(req);
1211                 RETURN(rc);
1212         }
1213
1214         LASSERT(current->journal_info == NULL);
1215
1216         EXIT;
1217         /* If we're DISCONNECTing, the export_data is already freed */
1218         if (!rc && req->rq_reqmsg->opc != OST_DISCONNECT) {
1219                 struct obd_device *obd  = req->rq_export->exp_obd;
1220                 if (!obd->obd_no_transno) {
1221                         req->rq_repmsg->last_committed =
1222                                 obd->obd_last_committed;
1223                 } else {
1224                         DEBUG_REQ(D_IOCTL, req,
1225                                   "not sending last_committed update");
1226                 }
1227                 CDEBUG(D_INFO, "last_committed "LPU64", xid "LPX64"\n",
1228                        obd->obd_last_committed, req->rq_xid);
1229         }
1230
1231 out:
1232         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1233                 struct obd_device *obd = req->rq_export->exp_obd;
1234
1235                 if (obd && obd->obd_recovering) {
1236                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1237                         return target_queue_final_reply(req, rc);
1238                 }
1239                 /* Lost a race with recovery; let the error path DTRT. */
1240                 rc = req->rq_status = -ENOTCONN;
1241         }
1242
1243         if (!rc)
1244                 oti_to_request(oti, req);
1245
1246         target_send_reply(req, rc, fail);
1247         return 0;
1248 }
1249 EXPORT_SYMBOL(ost_handle);
1250
1251 int ost_attach(struct obd_device *dev, obd_count len, void *data)
1252 {
1253         struct lprocfs_static_vars lvars;
1254
1255         lprocfs_init_vars(ost,&lvars);
1256         return lprocfs_obd_attach(dev, lvars.obd_vars);
1257 }
1258
1259 int ost_detach(struct obd_device *dev)
1260 {
1261         return lprocfs_obd_detach(dev);
1262 }
1263
1264 extern struct file_operations ost_stimes_fops;
1265
1266 static int ost_setup(struct obd_device *obd, obd_count len, void *buf)
1267 {
1268         struct ost_obd *ost = &obd->u.ost;
1269         int rc;
1270         ENTRY;
1271
1272         rc = cleanup_group_info();
1273         if (rc)
1274                 RETURN(rc);
1275
1276         rc = llog_start_commit_thread();
1277         if (rc < 0)
1278                 RETURN(rc);
1279
1280         lprocfs_obd_seq_create(obd, "service_times", 0444, &ost_stimes_fops,
1281                                obd);
1282
1283         ost->ost_service =
1284                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1285                                 OST_REQUEST_PORTAL, OSC_REPLY_PORTAL,
1286                                 ost_handle, "ost",
1287                                 obd->obd_proc_entry);
1288         if (ost->ost_service == NULL) {
1289                 CERROR("failed to start service\n");
1290                 RETURN(-ENOMEM);
1291         }
1292
1293         rc = ptlrpc_start_n_threads(obd, ost->ost_service, OST_NUM_THREADS,
1294                                     "ll_ost");
1295         if (rc)
1296                 GOTO(out_service, rc = -EINVAL);
1297
1298         ost->ost_create_service =
1299                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1300                                 OST_CREATE_PORTAL, OSC_REPLY_PORTAL,
1301                                 ost_handle, "ost_create",
1302                                 obd->obd_proc_entry);
1303         if (ost->ost_create_service == NULL) {
1304                 CERROR("failed to start OST create service\n");
1305                 GOTO(out_service, rc = -ENOMEM);
1306         }
1307
1308
1309         spin_lock_init(&ost->ost_lock);
1310         ost->ost_service->srv_obddev = obd;
1311
1312         rc = ptlrpc_start_n_threads(obd, ost->ost_create_service, 1,
1313                                     "ll_ost_creat");
1314         if (rc)
1315                 GOTO(out_create, rc = -EINVAL);
1316
1317         RETURN(0);
1318
1319 out_create:
1320         ptlrpc_unregister_service(ost->ost_create_service);
1321 out_service:
1322         ptlrpc_unregister_service(ost->ost_service);
1323         RETURN(rc);
1324 }
1325
1326 static int ost_cleanup(struct obd_device *obd, int flags)
1327 {
1328         struct ost_obd *ost = &obd->u.ost;
1329         int err = 0;
1330         ENTRY;
1331
1332         spin_lock_bh(&obd->obd_processing_task_lock);
1333         if (obd->obd_recovering) {
1334                 target_cancel_recovery_timer(obd);
1335                 obd->obd_recovering = 0;
1336         }
1337         spin_unlock_bh(&obd->obd_processing_task_lock);
1338
1339         ptlrpc_stop_all_threads(ost->ost_service);
1340         ptlrpc_unregister_service(ost->ost_service);
1341
1342         ptlrpc_stop_all_threads(ost->ost_create_service);
1343         ptlrpc_unregister_service(ost->ost_create_service);
1344
1345         RETURN(err);
1346 }
1347
1348 /* use obd ops to offer management infrastructure */
1349 static struct obd_ops ost_obd_ops = {
1350         .o_owner        = THIS_MODULE,
1351         .o_attach       = ost_attach,
1352         .o_detach       = ost_detach,
1353         .o_setup        = ost_setup,
1354         .o_cleanup      = ost_cleanup,
1355 };
1356
1357 static int __init ost_init(void)
1358 {
1359         struct lprocfs_static_vars lvars;
1360         ENTRY;
1361
1362         lprocfs_init_vars(ost,&lvars);
1363         RETURN(class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
1364                                    LUSTRE_OST_NAME));
1365 }
1366
1367 static void /*__exit*/ ost_exit(void)
1368 {
1369         class_unregister_type(LUSTRE_OST_NAME);
1370 }
1371
1372 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1373 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
1374 MODULE_LICENSE("GPL");
1375
1376 module_init(ost_init);
1377 module_exit(ost_exit);