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         oti->oti_nid = req->rq_peer.peer_id.nid;
  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         struct lustre_capa *capa = NULL;
 170         int rc, size = sizeof(*repbody);
 171         ENTRY;
 172
 173         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 174         if (body == NULL)
 175                 RETURN(-EFAULT);
 176
 177         if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
 178             (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
 179                 RETURN(-EINVAL);
 180
 181         if (body->oa.o_valid & OBD_MD_CAPA) {
 182                 capa = lustre_swab_reqbuf(req, 1, sizeof(*capa),
 183                                           lustre_swab_lustre_capa);
 184                 if (capa == NULL) {
 185                         CERROR("Missing/short capa\n");
 186                         RETURN(-EFAULT);
 187                 }
 188         }
 189
 190         rc = lustre_pack_reply(req, 1, &size, NULL);
 191         if (rc)
 192                 RETURN(rc);
 193
 194         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 195         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 196         req->rq_status = obd_punch(exp, &repbody->oa, NULL, repbody->oa.o_size,
 197                                    repbody->oa.o_blocks, oti, capa);
 198         RETURN(0);
 199 }
 200
 201 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
 202 {
 203         struct ost_body *body, *repbody;
 204         int rc, size = sizeof(*repbody);
 205         ENTRY;
 206
 207         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 208         if (body == NULL)
 209                 RETURN(-EFAULT);
 210
 211         rc = lustre_pack_reply(req, 1, &size, NULL);
 212         if (rc)
 213                 RETURN(rc);
 214
 215         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 216         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 217         req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
 218                                   repbody->oa.o_blocks);
 219         RETURN(0);
 220 }
 221
 222 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
 223                        struct obd_trans_info *oti)
 224 {
 225         struct ost_body *body, *repbody;
 226         int rc, size = sizeof(*repbody);
 227         ENTRY;
 228
 229         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 230         if (body == NULL)
 231                 RETURN(-EFAULT);
 232
 233         rc = lustre_pack_reply(req, 1, &size, NULL);
 234         if (rc)
 235                 RETURN(rc);
 236
 237         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 238         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 239
 240         req->rq_status = obd_setattr(exp, &repbody->oa, NULL, oti, NULL);
 241         RETURN(0);
 242 }
 243
 244 static int ost_bulk_timeout(void *data)
 245 {
 246         ENTRY;
 247         /* We don't fail the connection here, because having the export
 248          * killed makes the (vital) call to commitrw very sad.
 249          */
 250         RETURN(1);
 251 }
 252
 253 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
 254                                 struct niobuf_remote *rnb, int nrnb,
 255                                 struct niobuf_remote **pp_rnbp)
 256 {
 257         /* Copy a remote niobuf, splitting it into page-sized chunks
 258          * and setting ioo[i].ioo_bufcnt accordingly */
 259         struct niobuf_remote *pp_rnb;
 260         int   i;
 261         int   j;
 262         int   page;
 263         int   rnbidx = 0;
 264         int   npages = 0;
 265
 266         /* first count and check the number of pages required */
 267         for (i = 0; i < nioo; i++)
 268                 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
 269                         obd_off offset = rnb[rnbidx].offset;
 270                         obd_off p0 = offset >> PAGE_SHIFT;
 271                         obd_off pn = (offset + rnb[rnbidx].len - 1)>>PAGE_SHIFT;
 272
 273                         LASSERT(rnbidx < nrnb);
 274
 275                         npages += (pn + 1 - p0);
 276
 277                         if (rnb[rnbidx].len == 0) {
 278                                 CERROR("zero len BRW: obj %d objid "LPX64
 279                                        " buf %u\n", i, ioo[i].ioo_id, j);
 280                                 return -EINVAL;
 281                         }
 282                         if (j > 0 &&
 283                             rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
 284                                 CERROR("unordered BRW: obj %d objid "LPX64
 285                                        " buf %u offset "LPX64" <= "LPX64"\n",
 286                                        i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
 287                                        rnb[rnbidx].offset);
 288                                 return -EINVAL;
 289                         }
 290                 }
 291
 292         LASSERT(rnbidx == nrnb);
 293
 294         if (npages == nrnb) {       /* all niobufs are for single pages */
 295                 *pp_rnbp = rnb;
 296                 return npages;
 297         }
 298
 299         OBD_ALLOC(pp_rnb, sizeof(*pp_rnb) * npages);
 300         if (pp_rnb == NULL)
 301                 return -ENOMEM;
 302
 303         /* now do the actual split */
 304         page = rnbidx = 0;
 305         for (i = 0; i < nioo; i++) {
 306                 int  obj_pages = 0;
 307
 308                 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
 309                         obd_off off = rnb[rnbidx].offset;
 310                         int     nob = rnb[rnbidx].len;
 311
 312                         LASSERT(rnbidx < nrnb);
 313                         do {
 314                                 obd_off  poff = off & (PAGE_SIZE - 1);
 315                                 int      pnob = (poff + nob > PAGE_SIZE) ?
 316                                                 PAGE_SIZE - poff : nob;
 317
 318                                 LASSERT(page < npages);
 319                                 pp_rnb[page].len = pnob;
 320                                 pp_rnb[page].offset = off;
 321                                 pp_rnb[page].flags = rnb[rnbidx].flags;
 322
 323                                 CDEBUG(0, "   obj %d id "LPX64
 324                                        "page %d(%d) "LPX64" for %d, flg %x\n",
 325                                        i, ioo[i].ioo_id, obj_pages, page,
 326                                        pp_rnb[page].offset, pp_rnb[page].len,
 327                                        pp_rnb[page].flags);
 328                                 page++;
 329                                 obj_pages++;
 330
 331                                 off += pnob;
 332                                 nob -= pnob;
 333                         } while (nob > 0);
 334                         LASSERT(nob == 0);
 335                 }
 336                 ioo[i].ioo_bufcnt = obj_pages;
 337         }
 338         LASSERT(page == npages);
 339
 340         *pp_rnbp = pp_rnb;
 341         return npages;
 342 }
 343
 344 static void free_per_page_niobufs (int npages, struct niobuf_remote *pp_rnb,
 345                                    struct niobuf_remote *rnb)
 346 {
 347         if (pp_rnb == rnb)                      /* didn't allocate above */
 348                 return;
 349
 350         OBD_FREE(pp_rnb, sizeof(*pp_rnb) * npages);
 351 }
 352
 353 #if CHECKSUM_BULK
 354 obd_count ost_checksum_bulk(struct ptlrpc_bulk_desc *desc)
 355 {
 356         obd_count cksum = 0;
 357         struct ptlrpc_bulk_page *bp;
 358
 359         list_for_each_entry(bp, &desc->bd_page_list, bp_link) {
 360                 ost_checksum(&cksum, kmap(bp->bp_page) + bp->bp_pageoffset,
 361                              bp->bp_buflen);
 362                 kunmap(bp->bp_page);
 363         }
 364
 365         return cksum;
 366 }
 367 #endif
 368
 369 static void ost_stime_record(struct ptlrpc_request *req, struct timeval *start,
 370                              unsigned rw, unsigned phase)
 371 {
 372         struct obd_device *obd = req->rq_svc->srv_obddev;
 373         struct timeval stop;
 374         int ind = rw *3 + phase;
 375
 376         if (obd && obd->obd_type && obd->obd_type->typ_name) {
 377                 if (!strcmp(obd->obd_type->typ_name, OBD_OST_DEVICENAME)) {
 378                         struct ost_obd *ost = NULL;
 379
 380                         ost = &obd->u.ost;
 381                         if (ind >= (sizeof(ost->ost_stimes) /
 382                                     sizeof(ost->ost_stimes[0])))
 383                                return;
 384                         do_gettimeofday(&stop);
 385
 386                         spin_lock(&ost->ost_lock);
 387                         lprocfs_stime_record(&ost->ost_stimes[ind],&stop,start);
 388                         spin_unlock(&ost->ost_lock);
 389                         memcpy(start, &stop, sizeof(*start));
 390                 }
 391        }
 392 }
 393
 394 static int ost_brw_read(struct ptlrpc_request *req)
 395 {
 396         struct ptlrpc_bulk_desc *desc;
 397         struct niobuf_remote    *remote_nb;
 398         struct niobuf_remote    *pp_rnb;
 399         struct niobuf_local     *local_nb;
 400         struct obd_ioobj        *ioo;
 401         struct ost_body         *body, *repbody;
 402         struct lustre_capa      *capa = NULL;
 403         struct l_wait_info       lwi;
 404         struct obd_trans_info    oti = { 0 };
 405         int                      size[1] = { sizeof(*body) };
 406         int                      comms_error = 0;
 407         int                      niocount;
 408         int                      npages;
 409         int                      nob = 0;
 410         int                      rc;
 411         int                      i, bufcnt = 0;
 412         struct timeval           start;
 413         ENTRY;
 414
 415         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
 416                 GOTO(out, rc = -EIO);
 417
 418         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 419                          (obd_timeout + 1) / 4);
 420
 421         body = lustre_swab_reqbuf(req, bufcnt++, sizeof(*body),
 422                                   lustre_swab_ost_body);
 423         if (body == NULL) {
 424                 CERROR("Missing/short ost_body\n");
 425                 GOTO(out, rc = -EFAULT);
 426         }
 427
 428         ioo = lustre_swab_reqbuf(req, bufcnt++, sizeof(*ioo),
 429                                  lustre_swab_obd_ioobj);
 430         if (ioo == NULL) {
 431                 CERROR("Missing/short ioobj\n");
 432                 GOTO(out, rc = -EFAULT);
 433         }
 434
 435         if (body->oa.o_valid & OBD_MD_CAPA) {
 436                 capa = lustre_swab_reqbuf(req, bufcnt++, sizeof(*capa),
 437                                           lustre_swab_lustre_capa);
 438                 if (capa == NULL) {
 439                         CERROR("Missing/short capa\n");
 440                         GOTO(out, rc = -EFAULT);
 441                 }
 442         }
 443
 444         niocount = ioo->ioo_bufcnt;
 445         remote_nb = lustre_swab_reqbuf(req, bufcnt++,
 446                                        niocount * sizeof(*remote_nb),
 447                                        lustre_swab_niobuf_remote);
 448         if (remote_nb == NULL) {
 449                 CERROR("Missing/short niobuf\n");
 450                 GOTO(out, rc = -EFAULT);
 451         }
 452         if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
 453                 for (i = 1; i < niocount; i++)
 454                         lustre_swab_niobuf_remote (&remote_nb[i]);
 455         }
 456
 457         rc = lustre_pack_reply(req, 1, size, NULL);
 458         if (rc)
 459                 GOTO(out, rc);
 460
 461         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 462         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 463         npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
 464         if (npages < 0)
 465                 GOTO(out, rc = npages);
 466
 467         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 468         if (local_nb == NULL)
 469                 GOTO(out_pp_rnb, rc = -ENOMEM);
 470
 471         desc = ptlrpc_prep_bulk_exp (req, npages,
 472                                      BULK_PUT_SOURCE, OST_BULK_PORTAL);
 473         if (desc == NULL)
 474                 GOTO(out_local, rc = -ENOMEM);
 475
 476         do_gettimeofday(&start);
 477         rc = obd_preprw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 478                         ioo, npages, pp_rnb, local_nb, &oti, capa);
 479         ost_stime_record(req, &start, 0, 0);
 480         if (rc != 0)
 481                 GOTO(out_bulk, rc);
 482
 483         /* We're finishing using body->oa as an input variable */
 484         body->oa.o_valid = 0;
 485
 486         nob = 0;
 487         for (i = 0; i < npages; i++) {
 488                 int page_rc = local_nb[i].rc;
 489
 490                 if (page_rc < 0) {              /* error */
 491                         rc = page_rc;
 492                         break;
 493                 }
 494
 495                 LASSERT(page_rc <= pp_rnb[i].len);
 496                 nob += page_rc;
 497                 if (page_rc != 0) {             /* some data! */
 498                         LASSERT (local_nb[i].page != NULL);
 499                         ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 500                                               pp_rnb[i].offset & (PAGE_SIZE-1),
 501                                               page_rc);
 502                 }
 503
 504                 if (page_rc != pp_rnb[i].len) { /* short read */
 505                         /* All subsequent pages should be 0 */
 506                         while(++i < npages)
 507                                 LASSERT(local_nb[i].rc == 0);
 508                         break;
 509                 }
 510         }
 511
 512         if (rc == 0) {
 513                 rc = ptlrpc_start_bulk_transfer(desc);
 514                 if (rc == 0) {
 515                         struct timeval tstart, now;
 516                         do_gettimeofday(&tstart);
 517                         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 518                                           ost_bulk_timeout, desc);
 519                         rc = l_wait_event(desc->bd_waitq,
 520                                           !ptlrpc_bulk_active(desc), &lwi);
 521                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 522                         do_gettimeofday(&now);
 523                         if (rc == -ETIMEDOUT) {
 524                                 DEBUG_REQ(D_ERROR, req, "timeout on bulk PUT"
 525                                           ", exp_conn_cnt = %u, real wait %us"
 526                                           ", arrived %u.%u, served %u.%u",
 527                                           req->rq_export->exp_conn_cnt,
 528                                           (unsigned) (now.tv_sec - tstart.tv_sec),
 529                                           (unsigned) req->rq_arrival_time.tv_sec,
 530                                           (unsigned) req->rq_arrival_time.tv_usec,
 531                                           (unsigned) req->rq_rpcd_start.tv_sec,
 532                                           (unsigned) req->rq_rpcd_start.tv_usec);
 533                                 ptlrpc_abort_bulk(desc);
 534                         } else if (!desc->bd_success ||
 535                                    desc->bd_nob_transferred != desc->bd_nob) {
 536                                 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
 537                                           desc->bd_success ?
 538                                           "truncated" : "network error on",
 539                                           desc->bd_nob_transferred,
 540                                           desc->bd_nob);
 541                                 /* XXX should this be a different errno? */
 542                                 rc = -ETIMEDOUT;
 543                         }
 544                 } else {
 545                         DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d\n", rc);
 546                 }
 547                 comms_error = rc != 0;
 548         }
 549
 550         ost_stime_record(req, &start, 0, 1);
 551         /* Must commit after prep above in all cases */
 552         rc = obd_commitrw(OBD_BRW_READ, req->rq_export, &body->oa, 1,
 553                           ioo, npages, local_nb, &oti, rc);
 554         ost_stime_record(req, &start, 0, 2);
 555
 556         if (rc == 0) {
 557                 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 558                 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 559
 560 #if CHECKSUM_BULK
 561                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 562                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 563 #endif
 564         }
 565
 566  out_bulk:
 567         ptlrpc_free_bulk(desc);
 568  out_local:
 569         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 570  out_pp_rnb:
 571         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 572  out:
 573         LASSERT(rc <= 0);
 574         if (rc == 0) {
 575                 req->rq_status = nob;
 576                 ptlrpc_reply(req);
 577         } else if (!comms_error) {
 578                 /* only reply if comms OK */
 579                 req->rq_status = rc;
 580                 ptlrpc_error(req);
 581         } else {
 582                 if (req->rq_reply_state != NULL) {
 583                         /* reply out callback would free */
 584                         lustre_free_reply_state (req->rq_reply_state);
 585                 }
 586                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 587                         CERROR("bulk IO comms error: "
 588                                "evicting %s@%s id %s\n",
 589                                req->rq_export->exp_client_uuid.uuid,
 590                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 591                                req->rq_peerstr);
 592                         ptlrpc_fail_export(req->rq_export);
 593                 } else {
 594                         CERROR("ignoring bulk IO comms error: "
 595                                "client reconnected %s@%s id %s\n",
 596                                req->rq_export->exp_client_uuid.uuid,
 597                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 598                                req->rq_peerstr);
 599                 }
 600         }
 601
 602         RETURN(rc);
 603 }
 604
 605 int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
 606 {
 607         struct ptlrpc_bulk_desc *desc;
 608         struct niobuf_remote    *remote_nb;
 609         struct niobuf_remote    *pp_rnb;
 610         struct niobuf_local     *local_nb;
 611         struct obd_ioobj        *ioo;
 612         struct lustre_capa      *capa = NULL;
 613         struct ost_body         *body, *repbody;
 614         struct l_wait_info       lwi;
 615         __u32                   *rcs;
 616         int                      size[2] = { sizeof(*body) };
 617         int                      objcount, niocount, npages;
 618         int                      comms_error = 0;
 619         int                      rc, swab, i, j, bufcnt = 0;
 620         struct timeval           start;
 621         ENTRY;
 622
 623         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
 624                 GOTO(out, rc = -EIO);
 625
 626         /* pause before transaction has been started */
 627         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK | OBD_FAIL_ONCE,
 628                          (obd_timeout + 1) / 4);
 629
 630         swab = lustre_msg_swabbed(req->rq_reqmsg);
 631         body = lustre_swab_reqbuf(req, bufcnt++, sizeof(*body),
 632                                   lustre_swab_ost_body);
 633         if (body == NULL) {
 634                 CERROR("Missing/short ost_body\n");
 635                 GOTO(out, rc = -EFAULT);
 636         }
 637
 638         LASSERT_REQSWAB(req, 1);
 639         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 640         if (objcount == 0) {
 641                 CERROR("Missing/short ioobj\n");
 642                 GOTO(out, rc = -EFAULT);
 643         }
 644         ioo = lustre_msg_buf(req->rq_reqmsg, bufcnt++,
 645                              objcount * sizeof(*ioo));
 646         LASSERT (ioo != NULL);
 647         for (niocount = i = 0; i < objcount; i++) {
 648                 if (swab)
 649                         lustre_swab_obd_ioobj (&ioo[i]);
 650                 if (ioo[i].ioo_bufcnt == 0) {
 651                         CERROR("ioo[%d] has zero bufcnt\n", i);
 652                         GOTO(out, rc = -EFAULT);
 653                 }
 654                 niocount += ioo[i].ioo_bufcnt;
 655         }
 656
 657         if (body->oa.o_valid & OBD_MD_CAPA) {
 658                 capa = lustre_swab_reqbuf(req, bufcnt++, sizeof(*capa),
 659                                           lustre_swab_lustre_capa);
 660                 if (capa == NULL) {
 661                         CERROR("Missing/short capa\n");
 662                         GOTO(out, rc = -EFAULT);
 663                 }
 664         }
 665
 666         remote_nb = lustre_swab_reqbuf(req, bufcnt++,
 667                                        niocount * sizeof(*remote_nb),
 668                                        lustre_swab_niobuf_remote);
 669         if (remote_nb == NULL) {
 670                 CERROR("Missing/short niobuf\n");
 671                 GOTO(out, rc = -EFAULT);
 672         }
 673         if (swab) {                             /* swab the remaining niobufs */
 674                 for (i = 1; i < niocount; i++)
 675                         lustre_swab_niobuf_remote (&remote_nb[i]);
 676         }
 677
 678         size[1] = niocount * sizeof(*rcs);
 679         rc = lustre_pack_reply(req, 2, size, NULL);
 680         if (rc != 0)
 681                 GOTO(out, rc);
 682         rcs = lustre_msg_buf(req->rq_repmsg, 1, niocount * sizeof(*rcs));
 683
 684 #if 0
 685         /* Do snap options here*/
 686         rc = obd_do_cow(req->rq_export, ioo, objcount, remote_nb);
 687         if (rc)
 688                 GOTO(out, rc);
 689 #endif
 690
 691         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 692         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 693         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 694         if (npages < 0)
 695                 GOTO(out, rc = npages);
 696
 697         OBD_ALLOC(local_nb, sizeof(*local_nb) * npages);
 698         if (local_nb == NULL)
 699                 GOTO(out_pp_rnb, rc = -ENOMEM);
 700
 701         desc = ptlrpc_prep_bulk_exp (req, npages,
 702                                      BULK_GET_SINK, OST_BULK_PORTAL);
 703         if (desc == NULL)
 704                 GOTO(out_local, rc = -ENOMEM);
 705
 706         do_gettimeofday(&start);
 707         rc = obd_preprw(OBD_BRW_WRITE, req->rq_export, &body->oa, objcount,
 708                         ioo, npages, pp_rnb, local_nb, oti, capa);
 709         ost_stime_record(req, &start, 1, 0);
 710         if (rc != 0)
 711                 GOTO(out_bulk, rc);
 712
 713         /* NB Having prepped, we must commit... */
 714
 715         for (i = 0; i < npages; i++)
 716                 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 717                                       pp_rnb[i].offset & (PAGE_SIZE - 1),
 718                                       pp_rnb[i].len);
 719
 720         rc = ptlrpc_start_bulk_transfer (desc);
 721         if (rc == 0) {
 722                 struct timeval tstart, now;
 723                 do_gettimeofday(&tstart);
 724                 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4,
 725                                   ost_bulk_timeout, desc);
 726                 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc),
 727                                   &lwi);
 728                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 729                 do_gettimeofday(&now);
 730                 if (rc == -ETIMEDOUT) {
 731                         DEBUG_REQ(D_ERROR, req, "timeout on bulk GET, "
 732                                   "exp_conn_cnt = %u, real wait %us\n",
 733                                   req->rq_export->exp_conn_cnt,
 734                                   (unsigned) (now.tv_sec - tstart.tv_sec));
 735                         ptlrpc_abort_bulk(desc);
 736                 } else if (!desc->bd_success ||
 737                            desc->bd_nob_transferred != desc->bd_nob) {
 738                         DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
 739                                   desc->bd_success ?
 740                                   "truncated" : "network error on",
 741                                   desc->bd_nob_transferred, desc->bd_nob);
 742                         /* XXX should this be a different errno? */
 743                         rc = -ETIMEDOUT;
 744                 }
 745         } else {
 746                 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d\n", rc);
 747         }
 748         comms_error = rc != 0;
 749
 750         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 751         memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 752
 753 #if CHECKSUM_BULK
 754         if (rc == 0 && (body->oa.o_valid & OBD_MD_FLCKSUM) != 0) {
 755                 static int cksum_counter;
 756                 obd_count client_cksum = body->oa.o_cksum;
 757                 obd_count cksum = ost_checksum_bulk(desc);
 758
 759                 if (client_cksum != cksum) {
 760                         CERROR("Bad checksum: client %x, server %x id %s\n",
 761                                client_cksum, cksum,
 762                                req->rq_peerstr);
 763                         cksum_counter = 1;
 764                         repbody->oa.o_cksum = cksum;
 765                 } else {
 766                         cksum_counter++;
 767                         if ((cksum_counter & (-cksum_counter)) == cksum_counter)
 768                                 CWARN("Checksum %u from NID %s: %x OK\n",
 769                                       cksum_counter, req->rq_peerstr, cksum);
 770                 }
 771         }
 772 #endif
 773         ost_stime_record(req, &start, 1, 1);
 774         /* Must commit after prep above in all cases */
 775         rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa,
 776                           objcount, ioo, npages, local_nb, oti, rc);
 777
 778         ost_stime_record(req, &start, 1, 2);
 779         if (rc == 0) {
 780 #if CHECKSUM_BULK
 781                 repbody->oa.o_cksum = ost_checksum_bulk(desc);
 782                 repbody->oa.o_valid |= OBD_MD_FLCKSUM;
 783 #endif
 784                 /* set per-requested niobuf return codes */
 785                 for (i = j = 0; i < niocount; i++) {
 786                         int nob = remote_nb[i].len;
 787
 788                         rcs[i] = 0;
 789                         do {
 790                                 LASSERT(j < npages);
 791                                 if (local_nb[j].rc < 0)
 792                                         rcs[i] = local_nb[j].rc;
 793                                 nob -= pp_rnb[j].len;
 794                                 j++;
 795                         } while (nob > 0);
 796                         LASSERT(nob == 0);
 797                 }
 798                 LASSERT(j == npages);
 799         }
 800         /*XXX This write extents only for write-back cache extents*/
 801         rc = obd_write_extents(req->rq_export, ioo, objcount, niocount,
 802                                local_nb, rc);
 803  out_bulk:
 804         ptlrpc_free_bulk(desc);
 805  out_local:
 806         OBD_FREE(local_nb, sizeof(*local_nb) * npages);
 807  out_pp_rnb:
 808         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 809  out:
 810         if (rc == 0) {
 811                 oti_to_request(oti, req);
 812                 rc = ptlrpc_reply(req);
 813         } else if (!comms_error) {
 814                 /* Only reply if there was no comms problem with bulk */
 815                 req->rq_status = rc;
 816                 ptlrpc_error(req);
 817         } else {
 818                 if (req->rq_reply_state != NULL) {
 819                         /* reply out callback would free */
 820                         lustre_free_reply_state (req->rq_reply_state);
 821                 }
 822                 if (req->rq_reqmsg->conn_cnt == req->rq_export->exp_conn_cnt) {
 823                         CERROR("%s: bulk IO comm error evicting %s@%s id %s\n",
 824                                req->rq_export->exp_obd->obd_name,
 825                                req->rq_export->exp_client_uuid.uuid,
 826                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 827                                req->rq_peerstr);
 828                         ptlrpc_fail_export(req->rq_export);
 829                 } else {
 830                         CERROR("ignoring bulk IO comms error: "
 831                                "client reconnected %s@%s id %s\n",
 832                                req->rq_export->exp_client_uuid.uuid,
 833                                req->rq_export->exp_connection->c_remote_uuid.uuid,
 834                                req->rq_peerstr);
 835                 }
 836         }
 837         RETURN(rc);
 838 }
 839 EXPORT_SYMBOL(ost_brw_write);
 840
 841 static int ost_san_brw(struct ptlrpc_request *req, int cmd)
 842 {
 843         struct niobuf_remote *remote_nb, *res_nb, *pp_rnb;
 844         struct obd_ioobj *ioo;
 845         struct ost_body *body, *repbody;
 846         int rc, i, objcount, niocount, size[2] = {sizeof(*body)}, npages;
 847         int swab;
 848         ENTRY;
 849
 850         /* XXX not set to use latest protocol */
 851
 852         swab = lustre_msg_swabbed(req->rq_reqmsg);
 853         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 854         if (body == NULL) {
 855                 CERROR("Missing/short ost_body\n");
 856                 GOTO(out, rc = -EFAULT);
 857         }
 858
 859         ioo = lustre_swab_reqbuf(req, 1, sizeof(*ioo), lustre_swab_obd_ioobj);
 860         if (ioo == NULL) {
 861                 CERROR("Missing/short ioobj\n");
 862                 GOTO(out, rc = -EFAULT);
 863         }
 864         objcount = req->rq_reqmsg->buflens[1] / sizeof(*ioo);
 865         niocount = ioo[0].ioo_bufcnt;
 866         for (i = 1; i < objcount; i++) {
 867                 if (swab)
 868                         lustre_swab_obd_ioobj (&ioo[i]);
 869                 niocount += ioo[i].ioo_bufcnt;
 870         }
 871
 872         remote_nb = lustre_swab_reqbuf(req, 2, niocount * sizeof(*remote_nb),
 873                                        lustre_swab_niobuf_remote);
 874         if (remote_nb == NULL) {
 875                 CERROR("Missing/short niobuf\n");
 876                 GOTO(out, rc = -EFAULT);
 877         }
 878         if (swab) {                             /* swab the remaining niobufs */
 879                 for (i = 1; i < niocount; i++)
 880                         lustre_swab_niobuf_remote (&remote_nb[i]);
 881         }
 882
 883         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 884         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
 885         if (npages < 0)
 886                 GOTO (out, rc = npages);
 887
 888         size[1] = npages * sizeof(*pp_rnb);
 889         rc = lustre_pack_reply(req, 2, size, NULL);
 890         if (rc)
 891                 GOTO(out_pp_rnb, rc);
 892
 893         req->rq_status = obd_san_preprw(cmd, req->rq_export, &body->oa,
 894                                         objcount, ioo, npages, pp_rnb);
 895
 896         if (req->rq_status)
 897                 GOTO(out_pp_rnb, rc = 0);
 898
 899         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
 900         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 901
 902         res_nb = lustre_msg_buf(req->rq_repmsg, 1, size[1]);
 903         memcpy(res_nb, remote_nb, size[1]);
 904         rc = 0;
 905 out_pp_rnb:
 906         free_per_page_niobufs(npages, pp_rnb, remote_nb);
 907 out:
 908         if (rc) {
 909                 req->rq_status = rc;
 910                 ptlrpc_error(req);
 911         } else
 912                 ptlrpc_reply(req);
 913
 914         return rc;
 915 }
 916
 917 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
 918 {
 919         char *key, *val;
 920         int keylen, rc = 0;
 921         ENTRY;
 922
 923         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 924         if (key == NULL) {
 925                 DEBUG_REQ(D_HA, req, "no set_info key");
 926                 RETURN(-EFAULT);
 927         }
 928         keylen = req->rq_reqmsg->buflens[0];
 929
 930         rc = lustre_pack_reply(req, 0, NULL, NULL);
 931         if (rc)
 932                 RETURN(rc);
 933
 934         val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
 935         if (val == NULL) {
 936                 CERROR("val for setinfo can't be NULL\n");
 937                 RETURN(-EFAULT);
 938         }
 939
 940         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
 941                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
 942                                    lustre_swab_audit_msg);
 943         } else if (keylen == 5 && strcmp(key, "audit") == 0) {
 944                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
 945                                    lustre_swab_audit_attr);
 946         } else if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
 947                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
 948                                    lustre_swab_obdo);
 949         } else if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
 950                 lustre_swab_reqbuf(req, 1, sizeof(struct lustre_capa_key),
 951                                    lustre_swab_lustre_capa_key);
 952         }
 953
 954         rc = obd_set_info(exp, keylen, key, req->rq_reqmsg->buflens[1], val);
 955         req->rq_repmsg->status = 0;
 956         RETURN(rc);
 957 }
 958
 959 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
 960 {
 961         char *key;
 962         int keylen, rc = 0, size = sizeof(obd_id);
 963         obd_id *reply;
 964         ENTRY;
 965
 966         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
 967         if (key == NULL) {
 968                 DEBUG_REQ(D_HA, req, "no get_info key");
 969                 RETURN(-EFAULT);
 970         }
 971         keylen = req->rq_reqmsg->buflens[0];
 972
 973         if (keylen < strlen("last_id") || memcmp(key, "last_id", 7) != 0)
 974                 RETURN(-EPROTO);
 975
 976         rc = lustre_pack_reply(req, 1, &size, NULL);
 977         if (rc)
 978                 RETURN(rc);
 979
 980         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
 981         rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
 982         req->rq_repmsg->status = 0;
 983         RETURN(rc);
 984 }
 985
 986 static int ost_llog_handle_connect(struct obd_export *exp,
 987                                    struct ptlrpc_request *req)
 988 {
 989         struct llogd_conn_body *body;
 990         int rc;
 991         ENTRY;
 992
 993         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 994         rc = obd_llog_connect(exp, body);
 995         RETURN(rc);
 996 }
 997
 998 static int ost_filter_recovery_request(struct ptlrpc_request *req,
 999                                        struct obd_device *obd, int *process)
1000 {
1001         switch (req->rq_reqmsg->opc) {
1002         case OST_CONNECT: /* This will never get here, but for completeness. */
1003         case OST_DISCONNECT:
1004                *process = 1;
1005                RETURN(0);
1006
1007         case OBD_PING:
1008         case OST_CREATE:
1009         case OST_DESTROY:
1010         case OST_PUNCH:
1011         case OST_SETATTR:
1012         case OST_SYNC:
1013         case OST_WRITE:
1014         case OBD_LOG_CANCEL:
1015         case LDLM_ENQUEUE:
1016                 *process = target_queue_recovery_request(req, obd);
1017                 RETURN(0);
1018
1019         default:
1020                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1021                 *process = 0;
1022                 /* XXX what should we set rq_status to here? */
1023                 req->rq_status = -EAGAIN;
1024                 RETURN(ptlrpc_error(req));
1025         }
1026 }
1027
1028 int ost_msg_check_version(struct lustre_msg *msg)
1029 {
1030         int rc;
1031
1032         switch(msg->opc) {
1033         case OST_CONNECT:
1034         case OST_DISCONNECT:
1035         case OBD_PING:
1036         case OST_CREATE:
1037         case OST_DESTROY:
1038         case OST_GETATTR:
1039         case OST_SETATTR:
1040         case OST_WRITE:
1041         case OST_READ:
1042         case OST_SAN_READ:
1043         case OST_SAN_WRITE:
1044         case OST_PUNCH:
1045         case OST_STATFS:
1046         case OST_SYNC:
1047         case OST_SET_INFO:
1048         case OST_GET_INFO:
1049                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1050                 if (rc)
1051                         CERROR("bad opc %u version %08x, expecting %08x\n",
1052                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
1053                 break;
1054         case LDLM_ENQUEUE:
1055         case LDLM_CONVERT:
1056         case LDLM_CANCEL:
1057         case LDLM_BL_CALLBACK:
1058         case LDLM_CP_CALLBACK:
1059                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1060                 if (rc)
1061                         CERROR("bad opc %u version %08x, expecting %08x\n",
1062                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1063                 break;
1064         case OBD_LOG_CANCEL:
1065         case LLOG_ORIGIN_CONNECT:
1066                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1067                 if (rc)
1068                         CERROR("bad opc %u version %08x, expecting %08x\n",
1069                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1070                 break;
1071         case SEC_INIT:
1072         case SEC_INIT_CONTINUE:
1073         case SEC_FINI:
1074                 rc = 0;
1075                 break;
1076         default:
1077                 CERROR("OST unexpected opcode %d\n", msg->opc);
1078                 rc = -ENOTSUPP;
1079                 break;
1080         }
1081         return rc;
1082 }
1083
1084 int ost_handle(struct ptlrpc_request *req)
1085 {
1086         int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1087         struct obd_trans_info *oti = NULL;
1088         struct obd_device *obd = NULL;
1089         ENTRY;
1090
1091         LASSERT(current->journal_info == NULL);
1092
1093         rc = ost_msg_check_version(req->rq_reqmsg);
1094         if (rc) {
1095                 CERROR("OST drop mal-formed request\n");
1096                 RETURN(rc);
1097         }
1098
1099         /* Security opc should NOT trigger any recovery events */
1100         if (req->rq_reqmsg->opc == SEC_INIT ||
1101             req->rq_reqmsg->opc == SEC_INIT_CONTINUE ||
1102             req->rq_reqmsg->opc == SEC_FINI) {
1103                 GOTO(out_check_req, rc = 0);
1104         }
1105
1106         /* XXX identical to MDS */
1107         if (req->rq_reqmsg->opc != OST_CONNECT) {
1108                 int recovering;
1109
1110                 if (req->rq_export == NULL) {
1111                         CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1112                                req->rq_reqmsg->opc,
1113                                req->rq_peerstr);
1114                         req->rq_status = -ENOTCONN;
1115                         GOTO(out_check_req, rc = -ENOTCONN);
1116                 }
1117
1118                 obd = req->rq_export->exp_obd;
1119
1120                 /* Check for aborted recovery. */
1121                 spin_lock_bh(&obd->obd_processing_task_lock);
1122                 recovering = obd->obd_recovering;
1123                 spin_unlock_bh(&obd->obd_processing_task_lock);
1124                 if (recovering) {
1125                         rc = ost_filter_recovery_request(req, obd,
1126                                                          &should_process);
1127                         if (rc || !should_process)
1128                                 RETURN(rc);
1129                         if (should_process < 0) {
1130                                 req->rq_status = should_process;
1131                                 rc = ptlrpc_error(req);
1132                                 RETURN(rc);
1133                         }
1134                 }
1135         }
1136
1137         OBD_ALLOC(oti, sizeof(*oti));
1138         if (oti == NULL)
1139                 RETURN(-ENOMEM);
1140
1141         oti_init(oti, req);
1142
1143         switch (req->rq_reqmsg->opc) {
1144         case OST_CONNECT: {
1145                 CDEBUG(D_INODE, "connect\n");
1146                 OBD_FAIL_GOTO(OBD_FAIL_OST_CONNECT_NET, out_free_oti, rc = 0);
1147                 rc = target_handle_connect(req);
1148                 if (!rc)
1149                         obd = req->rq_export->exp_obd;
1150                 break;
1151         }
1152         case OST_DISCONNECT:
1153                 CDEBUG(D_INODE, "disconnect\n");
1154                 OBD_FAIL_GOTO(OBD_FAIL_OST_DISCONNECT_NET, out_free_oti, rc = 0);
1155                 rc = target_handle_disconnect(req);
1156                 break;
1157         case OST_CREATE:
1158                 CDEBUG(D_INODE, "create\n");
1159                 OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC);
1160                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1161                 OBD_FAIL_GOTO(OBD_FAIL_OST_CREATE_NET, out_free_oti, rc = 0);
1162                 rc = ost_create(req->rq_export, req, oti);
1163                 break;
1164         case OST_DESTROY:
1165                 CDEBUG(D_INODE, "destroy\n");
1166                 OBD_FAIL_GOTO(OBD_FAIL_OST_DESTROY_NET, out_free_oti, rc = 0);
1167                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1168                 rc = ost_destroy(req->rq_export, req, oti);
1169                 break;
1170         case OST_GETATTR:
1171                 CDEBUG(D_INODE, "getattr\n");
1172                 OBD_FAIL_GOTO(OBD_FAIL_OST_GETATTR_NET, out_free_oti, rc = 0);
1173                 rc = ost_getattr(req->rq_export, req);
1174                 break;
1175         case OST_SETATTR:
1176                 CDEBUG(D_INODE, "setattr\n");
1177                 OBD_FAIL_GOTO(OBD_FAIL_OST_SETATTR_NET, out_free_oti, rc = 0);
1178                 rc = ost_setattr(req->rq_export, req, oti);
1179                 break;
1180         case OST_WRITE:
1181                 CDEBUG(D_INODE, "write\n");
1182                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1183                 OBD_FAIL_GOTO(OBD_FAIL_OST_ENOSPC, out_check_req, rc = -ENOSPC);
1184                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1185                 rc = ost_brw_write(req, oti);
1186                 LASSERT(current->journal_info == NULL);
1187                 /* ost_brw sends its own replies */
1188                 GOTO(out_free_oti, rc);
1189         case OST_READ:
1190                 CDEBUG(D_INODE, "read\n");
1191                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1192                 rc = ost_brw_read(req);
1193                 LASSERT(current->journal_info == NULL);
1194                 /* ost_brw sends its own replies */
1195                 GOTO(out_free_oti, rc);
1196         case OST_SAN_READ:
1197                 CDEBUG(D_INODE, "san read\n");
1198                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1199                 rc = ost_san_brw(req, OBD_BRW_READ);
1200                 /* ost_san_brw sends its own replies */
1201                 GOTO(out_free_oti, rc);
1202         case OST_SAN_WRITE:
1203                 CDEBUG(D_INODE, "san write\n");
1204                 OBD_FAIL_GOTO(OBD_FAIL_OST_BRW_NET, out_free_oti, rc = 0);
1205                 rc = ost_san_brw(req, OBD_BRW_WRITE);
1206                 /* ost_san_brw sends its own replies */
1207                 GOTO(out_free_oti, rc);
1208         case OST_PUNCH:
1209                 CDEBUG(D_INODE, "punch\n");
1210                 OBD_FAIL_GOTO(OBD_FAIL_OST_PUNCH_NET, out_free_oti, rc = 0);
1211                 OBD_FAIL_GOTO(OBD_FAIL_OST_EROFS, out_check_req, rc = -EROFS);
1212                 rc = ost_punch(req->rq_export, req, oti);
1213                 break;
1214         case OST_STATFS:
1215                 CDEBUG(D_INODE, "statfs\n");
1216                 OBD_FAIL_GOTO(OBD_FAIL_OST_STATFS_NET, out_free_oti, rc = 0);
1217                 rc = ost_statfs(req);
1218                 break;
1219         case OST_SYNC:
1220                 CDEBUG(D_INODE, "sync\n");
1221                 OBD_FAIL_GOTO(OBD_FAIL_OST_SYNC_NET, out_free_oti, rc = 0);
1222                 rc = ost_sync(req->rq_export, req);
1223                 break;
1224         case OST_SET_INFO:
1225                 DEBUG_REQ(D_INODE, req, "set_info");
1226                 rc = ost_set_info(req->rq_export, req);
1227                 break;
1228         case OST_GET_INFO:
1229                 DEBUG_REQ(D_INODE, req, "get_info");
1230                 rc = ost_get_info(req->rq_export, req);
1231                 break;
1232         case OBD_PING:
1233                 DEBUG_REQ(D_INODE, req, "ping");
1234                 rc = target_handle_ping(req);
1235                 break;
1236         /* FIXME - just reply status */
1237         case LLOG_ORIGIN_CONNECT:
1238                 DEBUG_REQ(D_INODE, req, "log connect\n");
1239                 rc = ost_llog_handle_connect(req->rq_export, req);
1240                 req->rq_status = rc;
1241                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1242                 if (rc)
1243                         GOTO(out_free_oti, rc);
1244                 GOTO(out_free_oti, rc = ptlrpc_reply(req));
1245         case OBD_LOG_CANCEL:
1246                 CDEBUG(D_INODE, "log cancel\n");
1247                 OBD_FAIL_GOTO(OBD_FAIL_OBD_LOG_CANCEL_NET, out_free_oti, rc = 0);
1248                 rc = llog_origin_handle_cancel(req);
1249                 req->rq_status = rc;
1250                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1251                 if (rc)
1252                         GOTO(out_free_oti, rc);
1253                 GOTO(out_free_oti, rc = ptlrpc_reply(req));
1254         case LDLM_ENQUEUE:
1255                 CDEBUG(D_INODE, "enqueue\n");
1256                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_ENQUEUE, out_free_oti, rc = 0);
1257                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1258                                          ldlm_server_blocking_ast,
1259                                          ldlm_server_glimpse_ast);
1260                 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1261                 break;
1262         case LDLM_CONVERT:
1263                 CDEBUG(D_INODE, "convert\n");
1264                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_CONVERT, out_free_oti, rc = 0);
1265                 rc = ldlm_handle_convert(req);
1266                 break;
1267         case LDLM_CANCEL:
1268                 CDEBUG(D_INODE, "cancel\n");
1269                 OBD_FAIL_GOTO(OBD_FAIL_LDLM_CANCEL, out_free_oti, rc = 0);
1270                 rc = ldlm_handle_cancel(req);
1271                 break;
1272         case LDLM_BL_CALLBACK:
1273         case LDLM_CP_CALLBACK:
1274                 CDEBUG(D_INODE, "callback\n");
1275                 CERROR("callbacks should not happen on OST\n");
1276                 /* fall through */
1277         default:
1278                 CERROR("Unexpected opcode %d\n", req->rq_reqmsg->opc);
1279                 req->rq_status = -ENOTSUPP;
1280                 rc = ptlrpc_error(req);
1281                 GOTO(out_free_oti, rc);
1282         }
1283
1284         LASSERT(current->journal_info == NULL);
1285
1286         EXIT;
1287         /* If we're DISCONNECTing, the export_data is already freed */
1288         if (!rc && req->rq_reqmsg->opc != OST_DISCONNECT) {
1289                 if (!obd->obd_no_transno) {
1290                         req->rq_repmsg->last_committed =
1291                                 obd->obd_last_committed;
1292                 } else {
1293                         DEBUG_REQ(D_IOCTL, req,
1294                                   "not sending last_committed update");
1295                 }
1296                 CDEBUG(D_INFO, "last_committed "LPU64", xid "LPX64"\n",
1297                        obd->obd_last_committed, req->rq_xid);
1298         }
1299
1300 out_check_req:
1301
1302         if (!rc)
1303                 oti_to_request(oti, req);
1304         target_send_reply(req, rc, fail);
1305         rc = 0;
1306
1307 out_free_oti:
1308         if (oti)
1309                 OBD_FREE(oti, sizeof(*oti));
1310         return rc;
1311 }
1312 EXPORT_SYMBOL(ost_handle);
1313
1314 int ost_attach(struct obd_device *dev, obd_count len, void *data)
1315 {
1316         struct lprocfs_static_vars lvars;
1317
1318         lprocfs_init_vars(ost,&lvars);
1319         return lprocfs_obd_attach(dev, lvars.obd_vars);
1320 }
1321
1322 int ost_detach(struct obd_device *dev)
1323 {
1324         return lprocfs_obd_detach(dev);
1325 }
1326
1327 extern struct file_operations ost_stimes_fops;
1328
1329 static int ost_setup(struct obd_device *obd, obd_count len, void *buf)
1330 {
1331         struct ost_obd *ost = &obd->u.ost;
1332         int rc;
1333         ENTRY;
1334
1335         rc = cleanup_group_info();
1336         if (rc)
1337                 RETURN(rc);
1338
1339         rc = llog_start_commit_thread();
1340         if (rc < 0)
1341                 RETURN(rc);
1342
1343         lprocfs_obd_seq_create(obd, "service_times", 0444, &ost_stimes_fops,
1344                                obd);
1345
1346         ost->ost_service =
1347                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1348                                 OST_REQUEST_PORTAL, OSC_REPLY_PORTAL, 30000,
1349                                 ost_handle, "ost",
1350                                 obd->obd_proc_entry);
1351         if (ost->ost_service == NULL) {
1352                 CERROR("failed to start service\n");
1353                 RETURN(-ENOMEM);
1354         }
1355
1356         rc = ptlrpc_start_n_threads(obd, ost->ost_service, OST_NUM_THREADS,
1357                                     "ll_ost");
1358         if (rc)
1359                 GOTO(out_service, rc = -EINVAL);
1360
1361         ost->ost_create_service =
1362                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1363                                 OST_CREATE_PORTAL, OSC_REPLY_PORTAL, 30000,
1364                                 ost_handle, "ost_create",
1365                                 obd->obd_proc_entry);
1366         if (ost->ost_create_service == NULL) {
1367                 CERROR("failed to start OST create service\n");
1368                 GOTO(out_service, rc = -ENOMEM);
1369         }
1370
1371
1372         spin_lock_init(&ost->ost_lock);
1373         ost->ost_service->srv_obddev = obd;
1374
1375         rc = ptlrpc_start_n_threads(obd, ost->ost_create_service, 1,
1376                                     "ll_ost_creat");
1377         if (rc)
1378                 GOTO(out_create, rc = -EINVAL);
1379
1380         ost->ost_destroy_service =
1381                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1382                                 OST_DESTROY_PORTAL, OSC_REPLY_PORTAL, 30000,
1383                                 ost_handle, "ost_destroy",
1384                                 obd->obd_proc_entry);
1385         if (ost->ost_destroy_service == NULL) {
1386                 CERROR("failed to start service\n");
1387                 GOTO(out_create, rc = -ENOMEM);
1388         }
1389
1390         rc = ptlrpc_start_n_threads(obd, ost->ost_destroy_service,
1391                                     OST_NUM_THREADS, "ll_dstr_ost");
1392         if (rc)
1393                 GOTO(out_destroy, rc = -EINVAL);
1394
1395         RETURN(0);
1396
1397 out_destroy:
1398         ptlrpc_unregister_service(ost->ost_destroy_service);
1399 out_create:
1400         ptlrpc_unregister_service(ost->ost_create_service);
1401 out_service:
1402         ptlrpc_unregister_service(ost->ost_service);
1403         RETURN(rc);
1404 }
1405
1406 extern void lgss_svc_cache_purge_all(void);
1407 static int ost_cleanup(struct obd_device *obd, int flags)
1408 {
1409         struct ost_obd *ost = &obd->u.ost;
1410         int err = 0;
1411         ENTRY;
1412
1413         spin_lock_bh(&obd->obd_processing_task_lock);
1414         if (obd->obd_recovering) {
1415                 target_cancel_recovery_timer(obd);
1416                 obd->obd_recovering = 0;
1417         }
1418         spin_unlock_bh(&obd->obd_processing_task_lock);
1419
1420         ptlrpc_stop_all_threads(ost->ost_service);
1421         ptlrpc_unregister_service(ost->ost_service);
1422
1423         ptlrpc_stop_all_threads(ost->ost_create_service);
1424         ptlrpc_unregister_service(ost->ost_create_service);
1425
1426         ptlrpc_stop_all_threads(ost->ost_destroy_service);
1427         ptlrpc_unregister_service(ost->ost_destroy_service);
1428
1429 #ifdef ENABLE_GSS
1430         /* XXX */
1431         lgss_svc_cache_purge_all();
1432 #endif
1433         RETURN(err);
1434 }
1435
1436 /* use obd ops to offer management infrastructure */
1437 static struct obd_ops ost_obd_ops = {
1438         .o_owner        = THIS_MODULE,
1439         .o_attach       = ost_attach,
1440         .o_detach       = ost_detach,
1441         .o_setup        = ost_setup,
1442         .o_cleanup      = ost_cleanup,
1443 };
1444
1445 static int __init ost_init(void)
1446 {
1447         struct lprocfs_static_vars lvars;
1448         ENTRY;
1449
1450         lprocfs_init_vars(ost,&lvars);
1451         RETURN(class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
1452                                    OBD_OST_DEVICENAME));
1453 }
1454
1455 static void /*__exit*/ ost_exit(void)
1456 {
1457         class_unregister_type(OBD_OST_DEVICENAME);
1458 }
1459
1460 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1461 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
1462 MODULE_LICENSE("GPL");
1463
1464 module_init(ost_init);
1465 module_exit(ost_exit);