lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001-2005 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *   Author: Mike Shaver <shaver@clusterfs.com>
  12  *
  13  *   This file is part of the Lustre file system, http://www.lustre.org
  14  *   Lustre is a trademark of Cluster File Systems, Inc.
  15  *
  16  *   You may have signed or agreed to another license before downloading
  17  *   this software.  If so, you are bound by the terms and conditions
  18  *   of that agreement, and the following does not apply to you.  See the
  19  *   LICENSE file included with this distribution for more information.
  20  *
  21  *   If you did not agree to a different license, then this copy of Lustre
  22  *   is open source software; you can redistribute it and/or modify it
  23  *   under the terms of version 2 of the GNU General Public License as
  24  *   published by the Free Software Foundation.
  25  *
  26  *   In either case, Lustre is distributed in the hope that it will be
  27  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  28  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  29  *   license text for more details.
  30  */
  31
  32 #ifndef EXPORT_SYMTAB
  33 # define EXPORT_SYMTAB
  34 #endif
  35 #define DEBUG_SUBSYSTEM S_MDS
  36
  37 #include <lustre_mds.h>
  38 #include <linux/module.h>
  39 #include <linux/init.h>
  40 #include <linux/random.h>
  41 #include <linux/fs.h>
  42 #include <linux/jbd.h>
  43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  44 # include <linux/smp_lock.h>
  45 # include <linux/buffer_head.h>
  46 # include <linux/workqueue.h>
  47 # include <linux/mount.h>
  48 #else
  49 # include <linux/locks.h>
  50 #endif
  51
  52 #include <linux/lustre_acl.h>
  53 #include <obd_class.h>
  54 #include <lustre_dlm.h>
  55 #include <obd_lov.h>
  56 #include <lustre_fsfilt.h>
  57 #include <lprocfs_status.h>
  58 #include <lustre_commit_confd.h>
  59 #include <lustre_quota.h>
  60 #include <lustre_disk.h>
  61 #include <lustre_param.h>
  62
  63 #include "mds_internal.h"
  64
  65 int mds_num_threads;
  66 CFS_MODULE_PARM(mds_num_threads, "i", int, 0444,
  67                 "number of MDS service threads to start");
  68
  69 static int mds_intent_policy(struct ldlm_namespace *ns,
  70                              struct ldlm_lock **lockp, void *req_cookie,
  71                              ldlm_mode_t mode, int flags, void *data);
  72 static int mds_postsetup(struct obd_device *obd);
  73 static int mds_cleanup(struct obd_device *obd);
  74
  75 /* Assumes caller has already pushed into the kernel filesystem context */
  76 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  77                         loff_t offset, int count)
  78 {
  79         struct ptlrpc_bulk_desc *desc;
  80         struct l_wait_info lwi;
  81         struct page **pages;
  82         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  83         ENTRY;
  84
  85         LASSERT((offset & ~CFS_PAGE_MASK) == 0); /* I'm dubious about this */
  86
  87         npages = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
  88         OBD_ALLOC(pages, sizeof(*pages) * npages);
  89         if (!pages)
  90                 GOTO(out, rc = -ENOMEM);
  91
  92         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  93                                     MDS_BULK_PORTAL);
  94         if (desc == NULL)
  95                 GOTO(out_free, rc = -ENOMEM);
  96
  97         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  98                 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
  99
 100                 pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
 101                 if (pages[i] == NULL)
 102                         GOTO(cleanup_buf, rc = -ENOMEM);
 103
 104                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 105         }
 106
 107         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 108                 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
 109                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 110                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 111                        file->f_dentry->d_inode->i_size);
 112
 113                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 114                                      kmap(pages[i]), tmpsize, &offset);
 115                 kunmap(pages[i]);
 116
 117                 if (rc != tmpsize)
 118                         GOTO(cleanup_buf, rc = -EIO);
 119         }
 120
 121         LASSERT(desc->bd_nob == count);
 122
 123         rc = ptlrpc_start_bulk_transfer(desc);
 124         if (rc)
 125                 GOTO(cleanup_buf, rc);
 126
 127         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 128                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 129                        OBD_FAIL_MDS_SENDPAGE, rc);
 130                 GOTO(abort_bulk, rc);
 131         }
 132
 133         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 134         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 135         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 136
 137         if (rc == 0) {
 138                 if (desc->bd_success &&
 139                     desc->bd_nob_transferred == count)
 140                         GOTO(cleanup_buf, rc);
 141
 142                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 143         }
 144
 145         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s",
 146                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 147                   desc->bd_nob_transferred, count,
 148                   req->rq_export->exp_client_uuid.uuid,
 149                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 150
 151         class_fail_export(req->rq_export);
 152
 153         EXIT;
 154  abort_bulk:
 155         ptlrpc_abort_bulk (desc);
 156  cleanup_buf:
 157         for (i = 0; i < npages; i++)
 158                 if (pages[i])
 159                         __cfs_free_page(pages[i]);
 160
 161         ptlrpc_free_bulk(desc);
 162  out_free:
 163         OBD_FREE(pages, sizeof(*pages) * npages);
 164  out:
 165         return rc;
 166 }
 167
 168 /* only valid locked dentries or errors should be returned */
 169 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 170                                      struct vfsmount **mnt, int lock_mode,
 171                                      struct lustre_handle *lockh,
 172                                      __u64 lockpart)
 173 {
 174         struct mds_obd *mds = &obd->u.mds;
 175         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 176         struct ldlm_res_id res_id = { .name = {0} };
 177         int flags = LDLM_FL_ATOMIC_CB, rc;
 178         ldlm_policy_data_t policy = { .l_inodebits = { lockpart} };
 179         ENTRY;
 180
 181         if (IS_ERR(de))
 182                 RETURN(de);
 183
 184         res_id.name[0] = de->d_inode->i_ino;
 185         res_id.name[1] = de->d_inode->i_generation;
 186         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id,
 187                                     LDLM_IBITS, &policy, lock_mode, &flags,
 188                                     ldlm_blocking_ast, ldlm_completion_ast,
 189                                     NULL, NULL, 0, NULL, lockh);
 190         if (rc != ELDLM_OK) {
 191                 l_dput(de);
 192                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 193         }
 194
 195         RETURN(retval);
 196 }
 197
 198 /* Look up an entry by inode number. */
 199 /* this function ONLY returns valid dget'd dentries with an initialized inode
 200    or errors */
 201 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 202                               struct vfsmount **mnt)
 203 {
 204         char fid_name[32];
 205         unsigned long ino = fid->id;
 206         __u32 generation = fid->generation;
 207         struct inode *inode;
 208         struct dentry *result;
 209
 210         if (ino == 0)
 211                 RETURN(ERR_PTR(-ESTALE));
 212
 213         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 214
 215         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 216                ino, generation, mds->mds_obt.obt_sb);
 217
 218         /* under ext3 this is neither supposed to return bad inodes
 219            nor NULL inodes. */
 220         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 221         if (IS_ERR(result))
 222                 RETURN(result);
 223
 224         inode = result->d_inode;
 225         if (!inode)
 226                 RETURN(ERR_PTR(-ENOENT));
 227
 228         if (inode->i_generation == 0 || inode->i_nlink == 0) {
 229                 LCONSOLE_WARN("Found inode with zero generation or link -- this"
 230                               " may indicate disk corruption (inode: %lu/%u, "
 231                               "link %lu, count %d)\n", inode->i_ino,
 232                               inode->i_generation,(unsigned long)inode->i_nlink,
 233                               atomic_read(&inode->i_count));
 234                 dput(result);
 235                 RETURN(ERR_PTR(-ENOENT));
 236         }
 237
 238         if (generation && inode->i_generation != generation) {
 239                 /* we didn't find the right inode.. */
 240                 CDEBUG(D_INODE, "found wrong generation: inode %lu, link: %lu, "
 241                        "count: %d, generation %u/%u\n", inode->i_ino,
 242                        (unsigned long)inode->i_nlink,
 243                        atomic_read(&inode->i_count), inode->i_generation,
 244                        generation);
 245                 dput(result);
 246                 RETURN(ERR_PTR(-ENOENT));
 247         }
 248
 249         if (mnt) {
 250                 *mnt = mds->mds_vfsmnt;
 251                 mntget(*mnt);
 252         }
 253
 254         RETURN(result);
 255 }
 256
 257 static int mds_connect_internal(struct obd_export *exp,
 258                                 struct obd_connect_data *data)
 259 {
 260         struct obd_device *obd = exp->exp_obd;
 261         if (data != NULL) {
 262                 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
 263                 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
 264
 265                 /* If no known bits (which should not happen, probably,
 266                    as everybody should support LOOKUP and UPDATE bits at least)
 267                    revert to compat mode with plain locks. */
 268                 if (!data->ocd_ibits_known &&
 269                     data->ocd_connect_flags & OBD_CONNECT_IBITS)
 270                         data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
 271
 272                 if (!obd->u.mds.mds_fl_acl)
 273                         data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
 274
 275                 if (!obd->u.mds.mds_fl_user_xattr)
 276                         data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
 277
 278                 exp->exp_connect_flags = data->ocd_connect_flags;
 279                 data->ocd_version = LUSTRE_VERSION_CODE;
 280                 exp->exp_mds_data.med_ibits_known = data->ocd_ibits_known;
 281         }
 282
 283         if (obd->u.mds.mds_fl_acl &&
 284             ((exp->exp_connect_flags & OBD_CONNECT_ACL) == 0)) {
 285                 CWARN("%s: MDS requires ACL support but client does not\n",
 286                       obd->obd_name);
 287                 return -EBADE;
 288         }
 289         return 0;
 290 }
 291
 292 static int mds_reconnect(struct obd_export *exp, struct obd_device *obd,
 293                          struct obd_uuid *cluuid,
 294                          struct obd_connect_data *data)
 295 {
 296         int rc;
 297         ENTRY;
 298
 299         if (exp == NULL || obd == NULL || cluuid == NULL)
 300                 RETURN(-EINVAL);
 301
 302         rc = mds_connect_internal(exp, data);
 303
 304         RETURN(rc);
 305 }
 306
 307 /* Establish a connection to the MDS.
 308  *
 309  * This will set up an export structure for the client to hold state data
 310  * about that client, like open files, the last operation number it did
 311  * on the server, etc.
 312  */
 313 static int mds_connect(const struct lu_env *env,
 314                        struct lustre_handle *conn, struct obd_device *obd,
 315                        struct obd_uuid *cluuid, struct obd_connect_data *data)
 316 {
 317         struct obd_export *exp;
 318         struct mds_export_data *med;
 319         struct mds_client_data *mcd = NULL;
 320         int rc;
 321         ENTRY;
 322
 323         if (!conn || !obd || !cluuid)
 324                 RETURN(-EINVAL);
 325
 326         /* XXX There is a small race between checking the list and adding a
 327          * new connection for the same UUID, but the real threat (list
 328          * corruption when multiple different clients connect) is solved.
 329          *
 330          * There is a second race between adding the export to the list,
 331          * and filling in the client data below.  Hence skipping the case
 332          * of NULL mcd above.  We should already be controlling multiple
 333          * connects at the client, and we can't hold the spinlock over
 334          * memory allocations without risk of deadlocking.
 335          */
 336         rc = class_connect(conn, obd, cluuid);
 337         if (rc)
 338                 RETURN(rc);
 339         exp = class_conn2export(conn);
 340         LASSERT(exp);
 341         med = &exp->exp_mds_data;
 342
 343         rc = mds_connect_internal(exp, data);
 344         if (rc)
 345                 GOTO(out, rc);
 346
 347         OBD_ALLOC(mcd, sizeof(*mcd));
 348         if (!mcd)
 349                 GOTO(out, rc = -ENOMEM);
 350
 351         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 352         med->med_mcd = mcd;
 353
 354         rc = mds_client_add(obd, exp, -1);
 355         GOTO(out, rc);
 356
 357 out:
 358         if (rc) {
 359                 if (mcd) {
 360                         OBD_FREE(mcd, sizeof(*mcd));
 361                         med->med_mcd = NULL;
 362                 }
 363                 class_disconnect(exp);
 364         } else {
 365                 class_export_put(exp);
 366         }
 367
 368         RETURN(rc);
 369 }
 370
 371 int mds_init_export(struct obd_export *exp)
 372 {
 373         struct mds_export_data *med = &exp->exp_mds_data;
 374
 375         INIT_LIST_HEAD(&med->med_open_head);
 376         spin_lock_init(&med->med_open_lock);
 377
 378         spin_lock(&exp->exp_lock);
 379         exp->exp_connecting = 1;
 380         spin_unlock(&exp->exp_lock);
 381
 382         RETURN(0);
 383 }
 384
 385 static int mds_destroy_export(struct obd_export *export)
 386 {
 387         struct mds_export_data *med;
 388         struct obd_device *obd = export->exp_obd;
 389         struct mds_obd *mds = &obd->u.mds;
 390         struct lvfs_run_ctxt saved;
 391         struct lov_mds_md *lmm;
 392         struct llog_cookie *logcookies;
 393         int rc = 0;
 394         ENTRY;
 395
 396         med = &export->exp_mds_data;
 397         target_destroy_export(export);
 398
 399         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 400                 RETURN(0);
 401
 402         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 403         /* Close any open files (which may also cause orphan unlinking). */
 404
 405         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 406         if (lmm == NULL) {
 407                 CWARN("%s: allocation failure during cleanup; can not force "
 408                       "close file handles on this service.\n", obd->obd_name);
 409                 GOTO(out, rc = -ENOMEM);
 410         }
 411
 412         OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
 413         if (logcookies == NULL) {
 414                 CWARN("%s: allocation failure during cleanup; can not force "
 415                       "close file handles on this service.\n", obd->obd_name);
 416                 OBD_FREE(lmm, mds->mds_max_mdsize);
 417                 GOTO(out_lmm, rc = -ENOMEM);
 418         }
 419
 420         spin_lock(&med->med_open_lock);
 421         while (!list_empty(&med->med_open_head)) {
 422                 struct list_head *tmp = med->med_open_head.next;
 423                 struct mds_file_data *mfd =
 424                         list_entry(tmp, struct mds_file_data, mfd_list);
 425                 int lmm_size = mds->mds_max_mdsize;
 426                 umode_t mode = mfd->mfd_dentry->d_inode->i_mode;
 427                 __u64 valid = 0;
 428
 429                 /* Remove mfd handle so it can't be found again.
 430                  * We are consuming the mfd_list reference here. */
 431                 mds_mfd_unlink(mfd, 0);
 432                 spin_unlock(&med->med_open_lock);
 433
 434                 /* If you change this message, be sure to update
 435                  * replay_single:test_46 */
 436                 CDEBUG(D_INODE|D_IOCTL, "%s: force closing file handle for "
 437                        "%.*s (ino %lu)\n", obd->obd_name,
 438                        mfd->mfd_dentry->d_name.len,mfd->mfd_dentry->d_name.name,
 439                        mfd->mfd_dentry->d_inode->i_ino);
 440
 441                 rc = mds_get_md(obd, mfd->mfd_dentry->d_inode, lmm, &lmm_size, 1);
 442                 if (rc < 0)
 443                         CWARN("mds_get_md failure, rc=%d\n", rc);
 444                 else
 445                         valid |= OBD_MD_FLEASIZE;
 446
 447                 /* child orphan sem protects orphan_dec_test and
 448                  * is_orphan race, mds_mfd_close drops it */
 449                 MDS_DOWN_WRITE_ORPHAN_SEM(mfd->mfd_dentry->d_inode);
 450                 rc = mds_mfd_close(NULL, REQ_REC_OFF, obd, mfd,
 451                                    !(export->exp_flags & OBD_OPT_FAILOVER),
 452                                    lmm, lmm_size, logcookies,
 453                                    mds->mds_max_cookiesize,
 454                                    &valid);
 455
 456                 if (rc)
 457                         CDEBUG(D_INODE|D_IOCTL, "Error closing file: %d\n", rc);
 458
 459                 if (valid & OBD_MD_FLCOOKIE) {
 460                         rc = mds_osc_destroy_orphan(obd, mode, lmm,
 461                                                     lmm_size, logcookies, 1);
 462                         if (rc < 0) {
 463                                 CDEBUG(D_INODE, "%s: destroy of orphan failed,"
 464                                        " rc = %d\n", obd->obd_name, rc);
 465                                 rc = 0;
 466                         }
 467                         valid &= ~OBD_MD_FLCOOKIE;
 468                 }
 469
 470                 spin_lock(&med->med_open_lock);
 471         }
 472         spin_unlock(&med->med_open_lock);
 473
 474         OBD_FREE(logcookies, mds->mds_max_cookiesize);
 475 out_lmm:
 476         OBD_FREE(lmm, mds->mds_max_mdsize);
 477 out:
 478         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 479         mds_client_free(export);
 480         RETURN(rc);
 481 }
 482
 483 static int mds_disconnect(struct obd_export *exp)
 484 {
 485         int rc;
 486         ENTRY;
 487
 488         LASSERT(exp);
 489         class_export_get(exp);
 490
 491         /* Disconnect early so that clients can't keep using export */
 492         rc = class_disconnect(exp);
 493         if (exp->exp_obd->obd_namespace != NULL)
 494                 ldlm_cancel_locks_for_export(exp);
 495
 496         /* complete all outstanding replies */
 497         spin_lock(&exp->exp_lock);
 498         while (!list_empty(&exp->exp_outstanding_replies)) {
 499                 struct ptlrpc_reply_state *rs =
 500                         list_entry(exp->exp_outstanding_replies.next,
 501                                    struct ptlrpc_reply_state, rs_exp_list);
 502                 struct ptlrpc_service *svc = rs->rs_service;
 503
 504                 spin_lock(&svc->srv_lock);
 505                 list_del_init(&rs->rs_exp_list);
 506                 ptlrpc_schedule_difficult_reply(rs);
 507                 spin_unlock(&svc->srv_lock);
 508         }
 509         spin_unlock(&exp->exp_lock);
 510
 511         class_export_put(exp);
 512         RETURN(rc);
 513 }
 514
 515 static int mds_getstatus(struct ptlrpc_request *req)
 516 {
 517         struct mds_obd *mds = mds_req2mds(req);
 518         struct mds_body *body;
 519         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 520         ENTRY;
 521
 522         rc = lustre_pack_reply(req, 2, size, NULL);
 523         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 524                 CERROR("mds: out of memory for message\n");
 525                 req->rq_status = -ENOMEM;       /* superfluous? */
 526                 RETURN(-ENOMEM);
 527         }
 528
 529         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 530         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 531
 532         /* the last_committed and last_xid fields are filled in for all
 533          * replies already - no need to do so here also.
 534          */
 535         RETURN(0);
 536 }
 537
 538 /* get the LOV EA from @inode and store it into @md.  It can be at most
 539  * @size bytes, and @size is updated with the actual EA size.
 540  * The EA size is also returned on success, and -ve errno on failure.
 541  * If there is no EA then 0 is returned. */
 542 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
 543                int *size, int lock)
 544 {
 545         int rc = 0;
 546         int lmm_size;
 547
 548         if (lock)
 549                 LOCK_INODE_MUTEX(inode);
 550         rc = fsfilt_get_md(obd, inode, md, *size, "lov");
 551
 552         if (rc < 0) {
 553                 CERROR("Error %d reading eadata for ino %lu\n",
 554                        rc, inode->i_ino);
 555         } else if (rc > 0) {
 556                 lmm_size = rc;
 557                 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
 558
 559                 if (rc == 0) {
 560                         *size = lmm_size;
 561                         rc = lmm_size;
 562                 } else if (rc > 0) {
 563                         *size = rc;
 564                 }
 565         } else {
 566                 *size = 0;
 567         }
 568         if (lock)
 569                 UNLOCK_INODE_MUTEX(inode);
 570
 571         RETURN (rc);
 572 }
 573
 574
 575 /* Call with lock=1 if you want mds_pack_md to take the i_mutex.
 576  * Call with lock=0 if the caller has already taken the i_mutex. */
 577 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 578                 struct mds_body *body, struct inode *inode, int lock)
 579 {
 580         struct mds_obd *mds = &obd->u.mds;
 581         void *lmm;
 582         int lmm_size;
 583         int rc;
 584         ENTRY;
 585
 586         lmm = lustre_msg_buf(msg, offset, 0);
 587         if (lmm == NULL) {
 588                 /* Some problem with getting eadata when I sized the reply
 589                  * buffer... */
 590                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 591                        inode->i_ino);
 592                 RETURN(0);
 593         }
 594         lmm_size = lustre_msg_buflen(msg, offset);
 595
 596         /* I don't really like this, but it is a sanity check on the client
 597          * MD request.  However, if the client doesn't know how much space
 598          * to reserve for the MD, it shouldn't be bad to have too much space.
 599          */
 600         if (lmm_size > mds->mds_max_mdsize) {
 601                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 602                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 603                 // RETURN(-EINVAL);
 604         }
 605
 606         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
 607         if (rc > 0) {
 608                 if (S_ISDIR(inode->i_mode))
 609                         body->valid |= OBD_MD_FLDIREA;
 610                 else
 611                         body->valid |= OBD_MD_FLEASIZE;
 612                 body->eadatasize = lmm_size;
 613                 rc = 0;
 614         }
 615
 616         RETURN(rc);
 617 }
 618
 619 #ifdef CONFIG_FS_POSIX_ACL
 620 static
 621 int mds_pack_posix_acl(struct inode *inode, struct lustre_msg *repmsg,
 622                        struct mds_body *repbody, int repoff)
 623 {
 624         struct dentry de = { .d_inode = inode };
 625         int buflen, rc;
 626         ENTRY;
 627
 628         LASSERT(repbody->aclsize == 0);
 629         LASSERT(lustre_msg_bufcount(repmsg) > repoff);
 630
 631         buflen = lustre_msg_buflen(repmsg, repoff);
 632         if (!buflen)
 633                 GOTO(out, 0);
 634
 635         if (!inode->i_op || !inode->i_op->getxattr)
 636                 GOTO(out, 0);
 637
 638         lock_24kernel();
 639         rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
 640                                    lustre_msg_buf(repmsg, repoff, buflen),
 641                                    buflen);
 642         unlock_24kernel();
 643
 644         if (rc >= 0)
 645                 repbody->aclsize = rc;
 646         else if (rc != -ENODATA) {
 647                 CERROR("buflen %d, get acl: %d\n", buflen, rc);
 648                 RETURN(rc);
 649         }
 650         EXIT;
 651 out:
 652         repbody->valid |= OBD_MD_FLACL;
 653         return 0;
 654 }
 655 #else
 656 #define mds_pack_posix_acl(inode, repmsg, repbody, repoff) 0
 657 #endif
 658
 659 int mds_pack_acl(struct mds_export_data *med, struct inode *inode,
 660                  struct lustre_msg *repmsg, struct mds_body *repbody,
 661                  int repoff)
 662 {
 663         return mds_pack_posix_acl(inode, repmsg, repbody, repoff);
 664 }
 665
 666 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 667                                 struct ptlrpc_request *req,
 668                                 struct mds_body *reqbody, int reply_off)
 669 {
 670         struct mds_body *body;
 671         struct inode *inode = dentry->d_inode;
 672         int rc = 0;
 673         ENTRY;
 674
 675         if (inode == NULL)
 676                 RETURN(-ENOENT);
 677
 678         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 679         LASSERT(body != NULL);                 /* caller prepped reply */
 680
 681         mds_pack_inode2fid(&body->fid1, inode);
 682         body->flags = reqbody->flags; /* copy MDS_BFLAG_EXT_FLAGS if present */
 683         mds_pack_inode2body(body, inode);
 684         reply_off++;
 685
 686         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
 687             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
 688                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off, body,
 689                                  inode, 1);
 690
 691                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 692                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 693                     !(body->valid & OBD_MD_FLDIREA))
 694                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 695                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 696
 697                 lustre_shrink_reply(req, reply_off, body->eadatasize, 0);
 698                 if (body->eadatasize)
 699                         reply_off++;
 700         } else if (S_ISLNK(inode->i_mode) &&
 701                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 702                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off, 0);
 703                 int len;
 704
 705                 LASSERT (symname != NULL);       /* caller prepped reply */
 706                 len = lustre_msg_buflen(req->rq_repmsg, reply_off);
 707
 708                 rc = inode->i_op->readlink(dentry, symname, len);
 709                 if (rc < 0) {
 710                         CERROR("readlink failed: %d\n", rc);
 711                 } else if (rc != len - 1) {
 712                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 713                                 rc, len - 1);
 714                         rc = -EINVAL;
 715                 } else {
 716                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 717                         body->valid |= OBD_MD_LINKNAME;
 718                         body->eadatasize = rc + 1;
 719                         symname[rc] = 0;        /* NULL terminate */
 720                         rc = 0;
 721                 }
 722                 reply_off++;
 723         } else if (reqbody->valid == OBD_MD_FLFLAGS &&
 724                    reqbody->flags & MDS_BFLAG_EXT_FLAGS) {
 725                 int flags;
 726
 727                 /* We only return the full set of flags on ioctl, otherwise we
 728                  * get enough flags from the inode in mds_pack_inode2body(). */
 729                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_GETFLAGS,
 730                                       (long)&flags);
 731                 if (rc == 0)
 732                         body->flags = flags | MDS_BFLAG_EXT_FLAGS;
 733         }
 734
 735         if (reqbody->valid & OBD_MD_FLMODEASIZE) {
 736                 struct mds_obd *mds = mds_req2mds(req);
 737                 body->max_cookiesize = mds->mds_max_cookiesize;
 738                 body->max_mdsize = mds->mds_max_mdsize;
 739                 body->valid |= OBD_MD_FLMODEASIZE;
 740         }
 741
 742         if (rc)
 743                 RETURN(rc);
 744
 745 #ifdef CONFIG_FS_POSIX_ACL
 746         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
 747             (reqbody->valid & OBD_MD_FLACL)) {
 748                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 749                                   inode, req->rq_repmsg,
 750                                   body, reply_off);
 751
 752                 lustre_shrink_reply(req, reply_off, body->aclsize, 0);
 753                 if (body->aclsize)
 754                         reply_off++;
 755         }
 756 #endif
 757
 758         RETURN(rc);
 759 }
 760
 761 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 762                                 int offset)
 763 {
 764         struct mds_obd *mds = mds_req2mds(req);
 765         struct mds_body *body;
 766         int rc, bufcount = 2;
 767         int size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 768         ENTRY;
 769
 770         LASSERT(offset == REQ_REC_OFF); /* non-intent */
 771
 772         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
 773         LASSERT(body != NULL);                 /* checked by caller */
 774         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 775
 776         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 777             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
 778                 LOCK_INODE_MUTEX(inode);
 779                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
 780                                    "lov");
 781                 UNLOCK_INODE_MUTEX(inode);
 782                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 783                        rc, inode->i_ino);
 784                 if (rc < 0) {
 785                         if (rc != -ENODATA) {
 786                                 CERROR("error getting inode %lu MD: rc = %d\n",
 787                                        inode->i_ino, rc);
 788                                 RETURN(rc);
 789                         }
 790                         size[bufcount] = 0;
 791                 } else if (rc > mds->mds_max_mdsize) {
 792                         size[bufcount] = 0;
 793                         CERROR("MD size %d larger than maximum possible %u\n",
 794                                rc, mds->mds_max_mdsize);
 795                 } else {
 796                         size[bufcount] = rc;
 797                 }
 798                 bufcount++;
 799         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 800                 if (inode->i_size + 1 != body->eadatasize)
 801                         CERROR("symlink size: %Lu, reply space: %d\n",
 802                                inode->i_size + 1, body->eadatasize);
 803                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 804                 bufcount++;
 805                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 806                        inode->i_size + 1, body->eadatasize);
 807         }
 808
 809 #ifdef CONFIG_FS_POSIX_ACL
 810         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
 811             (body->valid & OBD_MD_FLACL)) {
 812                 struct dentry de = { .d_inode = inode };
 813
 814                 size[bufcount] = 0;
 815                 if (inode->i_op && inode->i_op->getxattr) {
 816                         lock_24kernel();
 817                         rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
 818                                                    NULL, 0);
 819                         unlock_24kernel();
 820
 821                         if (rc < 0) {
 822                                 if (rc != -ENODATA) {
 823                                         CERROR("got acl size: %d\n", rc);
 824                                         RETURN(rc);
 825                                 }
 826                         } else
 827                                 size[bufcount] = rc;
 828                 }
 829                 bufcount++;
 830         }
 831 #endif
 832
 833         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 834                 CERROR("failed MDS_GETATTR_PACK test\n");
 835                 req->rq_status = -ENOMEM;
 836                 RETURN(-ENOMEM);
 837         }
 838
 839         rc = lustre_pack_reply(req, bufcount, size, NULL);
 840         if (rc) {
 841                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
 842                 req->rq_status = rc;
 843                 RETURN(rc);
 844         }
 845
 846         RETURN(0);
 847 }
 848
 849 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
 850                             int child_part, struct lustre_handle *child_lockh)
 851 {
 852         struct obd_device *obd = req->rq_export->exp_obd;
 853         struct mds_obd *mds = &obd->u.mds;
 854         struct ldlm_reply *rep = NULL;
 855         struct lvfs_run_ctxt saved;
 856         struct mds_body *body;
 857         struct dentry *dparent = NULL, *dchild = NULL;
 858         struct lvfs_ucred uc = {0,};
 859         struct lustre_handle parent_lockh;
 860         int namesize;
 861         int rc = 0, cleanup_phase = 0, resent_req = 0;
 862         char *name;
 863         ENTRY;
 864
 865         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 866
 867         /* Swab now, before anyone looks inside the request */
 868         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 869                                   lustre_swab_mds_body);
 870         if (body == NULL) {
 871                 CERROR("Can't swab mds_body\n");
 872                 RETURN(-EFAULT);
 873         }
 874
 875         LASSERT_REQSWAB(req, offset + 1);
 876         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 877         if (name == NULL) {
 878                 CERROR("Can't unpack name\n");
 879                 RETURN(-EFAULT);
 880         }
 881         namesize = lustre_msg_buflen(req->rq_reqmsg, offset + 1);
 882         /* namesize less than 2 means we have empty name, probably came from
 883            revalidate by cfid, so no point in having name to be set */
 884         if (namesize <= 1)
 885                 name = NULL;
 886
 887         rc = mds_init_ucred(&uc, req, offset);
 888         if (rc)
 889                 GOTO(cleanup, rc);
 890
 891         LASSERT(offset == REQ_REC_OFF || offset == DLM_INTENT_REC_OFF);
 892         /* if requests were at offset 2, the getattr reply goes back at 1 */
 893         if (offset == DLM_INTENT_REC_OFF) {
 894                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 895                                      sizeof(*rep));
 896                 offset = DLM_REPLY_REC_OFF;
 897         }
 898
 899         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 900         cleanup_phase = 1; /* kernel context */
 901         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 902
 903         /* FIXME: handle raw lookup */
 904 #if 0
 905         if (body->valid == OBD_MD_FLID) {
 906                 struct mds_body *mds_reply;
 907                 int size = sizeof(*mds_reply);
 908                 ino_t inum;
 909                 // The user requested ONLY the inode number, so do a raw lookup
 910                 rc = lustre_pack_reply(req, 1, &size, NULL);
 911                 if (rc) {
 912                         CERROR("out of memory\n");
 913                         GOTO(cleanup, rc);
 914                 }
 915
 916                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 917
 918                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 919                                            sizeof(*mds_reply));
 920                 mds_reply->fid1.id = inum;
 921                 mds_reply->valid = OBD_MD_FLID;
 922                 GOTO(cleanup, rc);
 923         }
 924 #endif
 925
 926         if (lustre_handle_is_used(child_lockh)) {
 927                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 928                 resent_req = 1;
 929         }
 930
 931         if (resent_req == 0) {
 932                 if (name) {
 933                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
 934                         rc = mds_get_parent_child_locked(obd, &obd->u.mds,
 935                                                          &body->fid1,
 936                                                          &parent_lockh,
 937                                                          &dparent, LCK_CR,
 938                                                          MDS_INODELOCK_UPDATE,
 939                                                          name, namesize,
 940                                                          child_lockh, &dchild,
 941                                                          LCK_CR, child_part);
 942                 } else {
 943                         /* For revalidate by fid we always take UPDATE lock */
 944                         dchild = mds_fid2locked_dentry(obd, &body->fid2, NULL,
 945                                                        LCK_CR, child_lockh,
 946                                                        child_part);
 947                         LASSERT(dchild);
 948                         if (IS_ERR(dchild))
 949                                 rc = PTR_ERR(dchild);
 950                 }
 951                 if (rc)
 952                         GOTO(cleanup, rc);
 953         } else {
 954                 struct ldlm_lock *granted_lock;
 955                 struct ll_fid child_fid;
 956                 struct ldlm_resource *res;
 957                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 958                 granted_lock = ldlm_handle2lock(child_lockh);
 959                 LASSERTF(granted_lock != NULL, LPU64"/%u lockh "LPX64"\n",
 960                          body->fid1.id, body->fid1.generation,
 961                          child_lockh->cookie);
 962
 963
 964                 res = granted_lock->l_resource;
 965                 child_fid.id = res->lr_name.name[0];
 966                 child_fid.generation = res->lr_name.name[1];
 967                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 968                 LASSERT(!IS_ERR(dchild));
 969                 LDLM_LOCK_PUT(granted_lock);
 970         }
 971
 972         cleanup_phase = 2; /* dchild, dparent, locks */
 973
 974         if (dchild->d_inode == NULL) {
 975                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 976                 /* in the intent case, the policy clears this error:
 977                    the disposition is enough */
 978                 GOTO(cleanup, rc = -ENOENT);
 979         } else {
 980                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 981         }
 982
 983         if (req->rq_repmsg == NULL) {
 984                 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 985                 if (rc != 0) {
 986                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 987                         GOTO (cleanup, rc);
 988                 }
 989         }
 990
 991         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 992         GOTO(cleanup, rc); /* returns the lock to the client */
 993
 994  cleanup:
 995         switch (cleanup_phase) {
 996         case 2:
 997                 if (resent_req == 0) {
 998                         if (rc && dchild->d_inode)
 999                                 ldlm_lock_decref(child_lockh, LCK_CR);
1000                         if (name) {
1001                                 ldlm_lock_decref(&parent_lockh, LCK_CR);
1002                                 l_dput(dparent);
1003                         }
1004                 }
1005                 l_dput(dchild);
1006         case 1:
1007                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1008         default:
1009                 mds_exit_ucred(&uc, mds);
1010                 if (req->rq_reply_state == NULL) {
1011                         req->rq_status = rc;
1012                         lustre_pack_reply(req, 1, NULL, NULL);
1013                 }
1014         }
1015         return rc;
1016 }
1017
1018 static int mds_getattr(struct ptlrpc_request *req, int offset)
1019 {
1020         struct mds_obd *mds = mds_req2mds(req);
1021         struct obd_device *obd = req->rq_export->exp_obd;
1022         struct lvfs_run_ctxt saved;
1023         struct dentry *de;
1024         struct mds_body *body;
1025         struct lvfs_ucred uc = {0,};
1026         int rc = 0;
1027         ENTRY;
1028
1029         OBD_COUNTER_INCREMENT(obd, getattr);
1030
1031         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1032                                   lustre_swab_mds_body);
1033         if (body == NULL)
1034                 RETURN(-EFAULT);
1035
1036         rc = mds_init_ucred(&uc, req, offset);
1037         if (rc)
1038                 GOTO(out_ucred, rc);
1039
1040         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1041         de = mds_fid2dentry(mds, &body->fid1, NULL);
1042         if (IS_ERR(de)) {
1043                 rc = req->rq_status = PTR_ERR(de);
1044                 GOTO(out_pop, rc);
1045         }
1046
1047         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1048         if (rc != 0) {
1049                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1050                 GOTO(out_pop, rc);
1051         }
1052
1053         req->rq_status = mds_getattr_internal(obd, de, req, body,
1054                                               REPLY_REC_OFF);
1055
1056         l_dput(de);
1057         GOTO(out_pop, rc);
1058 out_pop:
1059         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1060 out_ucred:
1061         if (req->rq_reply_state == NULL) {
1062                 req->rq_status = rc;
1063                 lustre_pack_reply(req, 1, NULL, NULL);
1064         }
1065         mds_exit_ucred(&uc, mds);
1066         return rc;
1067 }
1068
1069 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1070                           __u64 max_age)
1071 {
1072         int rc;
1073
1074         spin_lock(&obd->obd_osfs_lock);
1075         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
1076         if (rc == 0)
1077                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1078         spin_unlock(&obd->obd_osfs_lock);
1079
1080         return rc;
1081 }
1082
1083 static int mds_statfs(struct ptlrpc_request *req)
1084 {
1085         struct obd_device *obd = req->rq_export->exp_obd;
1086         int rc, size[2] = { sizeof(struct ptlrpc_body),
1087                             sizeof(struct obd_statfs) };
1088         ENTRY;
1089
1090         /* This will trigger a watchdog timeout */
1091         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1092                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1093         OBD_COUNTER_INCREMENT(obd, statfs);
1094
1095         rc = lustre_pack_reply(req, 2, size, NULL);
1096         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1097                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1098                 GOTO(out, rc);
1099         }
1100
1101         /* We call this so that we can cache a bit - 1 jiffie worth */
1102         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1103                                                 size[REPLY_REC_OFF]),
1104                             cfs_time_current_64() - HZ);
1105         if (rc) {
1106                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1107                 GOTO(out, rc);
1108         }
1109
1110         EXIT;
1111 out:
1112         req->rq_status = rc;
1113         return 0;
1114 }
1115
1116 static int mds_sync(struct ptlrpc_request *req, int offset)
1117 {
1118         struct obd_device *obd = req->rq_export->exp_obd;
1119         struct mds_obd *mds = &obd->u.mds;
1120         struct mds_body *body;
1121         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
1122         ENTRY;
1123
1124         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1125                                   lustre_swab_mds_body);
1126         if (body == NULL)
1127                 GOTO(out, rc = -EFAULT);
1128
1129         rc = lustre_pack_reply(req, 2, size, NULL);
1130         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1131                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1132                 GOTO(out, rc);
1133         }
1134
1135         if (body->fid1.id == 0) {
1136                 /* a fid of zero is taken to mean "sync whole filesystem" */
1137                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
1138                 GOTO(out, rc);
1139         } else {
1140                 struct dentry *de;
1141
1142                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1143                 if (IS_ERR(de))
1144                         GOTO(out, rc = PTR_ERR(de));
1145
1146                 /* The file parameter isn't used for anything */
1147                 if (de->d_inode->i_fop && de->d_inode->i_fop->fsync)
1148                         rc = de->d_inode->i_fop->fsync(NULL, de, 1);
1149                 if (rc == 0) {
1150                         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1151                                               sizeof(*body));
1152                         mds_pack_inode2fid(&body->fid1, de->d_inode);
1153                         mds_pack_inode2body(body, de->d_inode);
1154                 }
1155
1156                 l_dput(de);
1157                 GOTO(out, rc);
1158         }
1159 out:
1160         req->rq_status = rc;
1161         return 0;
1162 }
1163
1164 /* mds_readpage does not take a DLM lock on the inode, because the client must
1165  * already have a PR lock.
1166  *
1167  * If we were to take another one here, a deadlock will result, if another
1168  * thread is already waiting for a PW lock. */
1169 static int mds_readpage(struct ptlrpc_request *req, int offset)
1170 {
1171         struct obd_device *obd = req->rq_export->exp_obd;
1172         struct mds_obd *mds = &obd->u.mds;
1173         struct vfsmount *mnt;
1174         struct dentry *de;
1175         struct file *file;
1176         struct mds_body *body, *repbody;
1177         struct lvfs_run_ctxt saved;
1178         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
1179         struct lvfs_ucred uc = {0,};
1180         ENTRY;
1181
1182         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1183                 RETURN(-ENOMEM);
1184
1185         rc = lustre_pack_reply(req, 2, size, NULL);
1186         if (rc) {
1187                 CERROR("error packing readpage reply: rc %d\n", rc);
1188                 GOTO(out, rc);
1189         }
1190
1191         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1192                                   lustre_swab_mds_body);
1193         if (body == NULL)
1194                 GOTO (out, rc = -EFAULT);
1195
1196         rc = mds_init_ucred(&uc, req, offset);
1197         if (rc)
1198                 GOTO(out, rc);
1199
1200         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1201         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1202         if (IS_ERR(de))
1203                 GOTO(out_pop, rc = PTR_ERR(de));
1204
1205         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1206
1207         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1208         /* note: in case of an error, dentry_open puts dentry */
1209         if (IS_ERR(file))
1210                 GOTO(out_pop, rc = PTR_ERR(file));
1211
1212         /* body->size is actually the offset -eeb */
1213         if ((body->size & (de->d_inode->i_sb->s_blocksize - 1)) != 0) {
1214                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1215                        body->size, de->d_inode->i_sb->s_blocksize);
1216                 GOTO(out_file, rc = -EFAULT);
1217         }
1218
1219         /* body->nlink is actually the #bytes to read -eeb */
1220         if (body->nlink & (de->d_inode->i_sb->s_blocksize - 1)) {
1221                 CERROR("size %u is not multiple of blocksize %lu\n",
1222                        body->nlink, de->d_inode->i_sb->s_blocksize);
1223                 GOTO(out_file, rc = -EFAULT);
1224         }
1225
1226         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1227                                  sizeof(*repbody));
1228         repbody->size = file->f_dentry->d_inode->i_size;
1229         repbody->valid = OBD_MD_FLSIZE;
1230
1231         /* to make this asynchronous make sure that the handling function
1232            doesn't send a reply when this function completes. Instead a
1233            callback function would send the reply */
1234         /* body->size is actually the offset -eeb */
1235         rc = mds_sendpage(req, file, body->size, body->nlink);
1236
1237 out_file:
1238         filp_close(file, 0);
1239 out_pop:
1240         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1241 out:
1242         mds_exit_ucred(&uc, mds);
1243         req->rq_status = rc;
1244         RETURN(0);
1245 }
1246
1247 int mds_reint(struct ptlrpc_request *req, int offset,
1248               struct lustre_handle *lockh)
1249 {
1250         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1251         int rc;
1252
1253         OBD_ALLOC(rec, sizeof(*rec));
1254         if (rec == NULL)
1255                 RETURN(-ENOMEM);
1256
1257         rc = mds_update_unpack(req, offset, rec);
1258         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1259                 CERROR("invalid record\n");
1260                 GOTO(out, req->rq_status = -EINVAL);
1261         }
1262
1263         /* rc will be used to interrupt a for loop over multiple records */
1264         rc = mds_reint_rec(rec, offset, req, lockh);
1265  out:
1266         OBD_FREE(rec, sizeof(*rec));
1267         return rc;
1268 }
1269
1270 int mds_filter_recovery_request(struct ptlrpc_request *req,
1271                                 struct obd_device *obd, int *process)
1272 {
1273         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1274         case MDS_CONNECT: /* This will never get here, but for completeness. */
1275         case OST_CONNECT: /* This will never get here, but for completeness. */
1276         case MDS_DISCONNECT:
1277         case OST_DISCONNECT:
1278                *process = 1;
1279                RETURN(0);
1280
1281         case MDS_CLOSE:
1282         case MDS_DONE_WRITING:
1283         case MDS_SYNC: /* used in unmounting */
1284         case OBD_PING:
1285         case MDS_REINT:
1286         case SEQ_QUERY:
1287         case FLD_QUERY:
1288         case LDLM_ENQUEUE:
1289                 *process = target_queue_recovery_request(req, obd);
1290                 RETURN(0);
1291
1292         default:
1293                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1294                 *process = -EAGAIN;
1295                 RETURN(0);
1296         }
1297 }
1298 EXPORT_SYMBOL(mds_filter_recovery_request);
1299
1300 static char *reint_names[] = {
1301         [REINT_SETATTR] "setattr",
1302         [REINT_CREATE]  "create",
1303         [REINT_LINK]    "link",
1304         [REINT_UNLINK]  "unlink",
1305         [REINT_RENAME]  "rename",
1306         [REINT_OPEN]    "open",
1307 };
1308
1309 static int mds_set_info_rpc(struct obd_export *exp, struct ptlrpc_request *req)
1310 {
1311         char *key;
1312         __u32 *val;
1313         int keylen, rc = 0;
1314         ENTRY;
1315
1316         key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1317         if (key == NULL) {
1318                 DEBUG_REQ(D_HA, req, "no set_info key");
1319                 RETURN(-EFAULT);
1320         }
1321         keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1322
1323         val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, sizeof(*val));
1324         if (val == NULL) {
1325                 DEBUG_REQ(D_HA, req, "no set_info val");
1326                 RETURN(-EFAULT);
1327         }
1328
1329         rc = lustre_pack_reply(req, 1, NULL, NULL);
1330         if (rc)
1331                 RETURN(rc);
1332         lustre_msg_set_status(req->rq_repmsg, 0);
1333
1334         if (keylen < strlen("read-only") ||
1335             memcmp(key, "read-only", keylen) != 0)
1336                 RETURN(-EINVAL);
1337
1338         if (*val)
1339                 exp->exp_connect_flags |= OBD_CONNECT_RDONLY;
1340         else
1341                 exp->exp_connect_flags &= ~OBD_CONNECT_RDONLY;
1342
1343         RETURN(0);
1344 }
1345
1346 static int mds_handle_quotacheck(struct ptlrpc_request *req)
1347 {
1348         struct obd_quotactl *oqctl;
1349         int rc;
1350         ENTRY;
1351
1352         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1353                                    lustre_swab_obd_quotactl);
1354         if (oqctl == NULL)
1355                 RETURN(-EPROTO);
1356
1357         rc = lustre_pack_reply(req, 1, NULL, NULL);
1358         if (rc) {
1359                 CERROR("mds: out of memory while packing quotacheck reply\n");
1360                 RETURN(rc);
1361         }
1362
1363         req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1364         RETURN(0);
1365 }
1366
1367 static int mds_handle_quotactl(struct ptlrpc_request *req)
1368 {
1369         struct obd_quotactl *oqctl, *repoqc;
1370         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1371         ENTRY;
1372
1373         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1374                                    lustre_swab_obd_quotactl);
1375         if (oqctl == NULL)
1376                 RETURN(-EPROTO);
1377
1378         rc = lustre_pack_reply(req, 2, size, NULL);
1379         if (rc)
1380                 RETURN(rc);
1381
1382         repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1383
1384         req->rq_status = obd_quotactl(req->rq_export, oqctl);
1385         *repoqc = *oqctl;
1386         RETURN(0);
1387 }
1388
1389 int mds_msg_check_version(struct lustre_msg *msg)
1390 {
1391         int rc;
1392
1393         switch (lustre_msg_get_opc(msg)) {
1394         case MDS_CONNECT:
1395         case MDS_DISCONNECT:
1396         case OBD_PING:
1397         case SEC_CTX_INIT:
1398         case SEC_CTX_INIT_CONT:
1399         case SEC_CTX_FINI:
1400                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1401                 if (rc)
1402                         CERROR("bad opc %u version %08x, expecting %08x\n",
1403                                lustre_msg_get_opc(msg),
1404                                lustre_msg_get_version(msg),
1405                                LUSTRE_OBD_VERSION);
1406                 break;
1407         case MDS_GETSTATUS:
1408         case MDS_GETATTR:
1409         case MDS_GETATTR_NAME:
1410         case MDS_STATFS:
1411         case MDS_READPAGE:
1412         case MDS_WRITEPAGE:
1413         case MDS_IS_SUBDIR:
1414         case MDS_REINT:
1415         case MDS_CLOSE:
1416         case MDS_DONE_WRITING:
1417         case MDS_PIN:
1418         case MDS_SYNC:
1419         case MDS_GETXATTR:
1420         case MDS_SETXATTR:
1421         case MDS_SET_INFO:
1422         case MDS_QUOTACHECK:
1423         case MDS_QUOTACTL:
1424         case QUOTA_DQACQ:
1425         case QUOTA_DQREL:
1426         case SEQ_QUERY:
1427         case FLD_QUERY:
1428                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1429                 if (rc)
1430                         CERROR("bad opc %u version %08x, expecting %08x\n",
1431                                lustre_msg_get_opc(msg),
1432                                lustre_msg_get_version(msg),
1433                                LUSTRE_MDS_VERSION);
1434                 break;
1435         case LDLM_ENQUEUE:
1436         case LDLM_CONVERT:
1437         case LDLM_BL_CALLBACK:
1438         case LDLM_CP_CALLBACK:
1439                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1440                 if (rc)
1441                         CERROR("bad opc %u version %08x, expecting %08x\n",
1442                                lustre_msg_get_opc(msg),
1443                                lustre_msg_get_version(msg),
1444                                LUSTRE_DLM_VERSION);
1445                 break;
1446         case OBD_LOG_CANCEL:
1447         case LLOG_ORIGIN_HANDLE_CREATE:
1448         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1449         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1450         case LLOG_ORIGIN_HANDLE_CLOSE:
1451         case LLOG_ORIGIN_HANDLE_DESTROY:
1452         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1453         case LLOG_CATINFO:
1454                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1455                 if (rc)
1456                         CERROR("bad opc %u version %08x, expecting %08x\n",
1457                                lustre_msg_get_opc(msg),
1458                                lustre_msg_get_version(msg),
1459                                LUSTRE_LOG_VERSION);
1460                 break;
1461         default:
1462                 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
1463                 rc = -ENOTSUPP;
1464         }
1465         return rc;
1466 }
1467 EXPORT_SYMBOL(mds_msg_check_version);
1468
1469 int mds_handle(struct ptlrpc_request *req)
1470 {
1471         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1472         int rc;
1473         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1474         struct obd_device *obd = NULL;
1475         ENTRY;
1476
1477         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1478
1479         LASSERT(current->journal_info == NULL);
1480
1481         rc = mds_msg_check_version(req->rq_reqmsg);
1482         if (rc) {
1483                 CERROR("MDS drop mal-formed request\n");
1484                 RETURN(rc);
1485         }
1486
1487         /* XXX identical to OST */
1488         if (lustre_msg_get_opc(req->rq_reqmsg) != MDS_CONNECT) {
1489                 struct mds_export_data *med;
1490                 int recovering;
1491
1492                 if (req->rq_export == NULL) {
1493                         CERROR("operation %d on unconnected MDS from %s\n",
1494                                lustre_msg_get_opc(req->rq_reqmsg),
1495                                libcfs_id2str(req->rq_peer));
1496                         req->rq_status = -ENOTCONN;
1497                         GOTO(out, rc = -ENOTCONN);
1498                 }
1499
1500                 med = &req->rq_export->exp_mds_data;
1501                 obd = req->rq_export->exp_obd;
1502                 mds = mds_req2mds(req);
1503
1504                 /* sanity check: if the xid matches, the request must
1505                  * be marked as a resent or replayed */
1506                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
1507                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid))
1508                         if (!(lustre_msg_get_flags(req->rq_reqmsg) &
1509                                  (MSG_RESENT | MSG_REPLAY))) {
1510                                 CERROR("rq_xid "LPU64" matches last_xid, "
1511                                        "expected RESENT flag\n",
1512                                         req->rq_xid);
1513                                 req->rq_status = -ENOTCONN;
1514                                 GOTO(out, rc = -EFAULT);
1515                         }
1516                 /* else: note the opposite is not always true; a
1517                  * RESENT req after a failover will usually not match
1518                  * the last_xid, since it was likely never
1519                  * committed. A REPLAYed request will almost never
1520                  * match the last xid, however it could for a
1521                  * committed, but still retained, open. */
1522
1523                 /* Check for aborted recovery. */
1524                 spin_lock_bh(&obd->obd_processing_task_lock);
1525                 recovering = obd->obd_recovering;
1526                 spin_unlock_bh(&obd->obd_processing_task_lock);
1527                 if (recovering) {
1528                         rc = mds_filter_recovery_request(req, obd,
1529                                                          &should_process);
1530                         if (rc || !should_process)
1531                                 RETURN(rc);
1532                         else if (should_process < 0) {
1533                                 req->rq_status = should_process;
1534                                 rc = ptlrpc_error(req);
1535                                 RETURN(rc);
1536                         }
1537                 }
1538         }
1539
1540         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1541         case MDS_CONNECT:
1542                 DEBUG_REQ(D_INODE, req, "connect");
1543                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1544                 rc = target_handle_connect(req);
1545                 if (!rc) {
1546                         /* Now that we have an export, set mds. */
1547                         /*
1548                          * XXX nikita: these assignments are useless: mds is
1549                          * never used below, and obd is only used for
1550                          * MSG_LAST_REPLAY case, which never happens for
1551                          * MDS_CONNECT.
1552                          */
1553                         obd = req->rq_export->exp_obd;
1554                         mds = mds_req2mds(req);
1555                 }
1556                 break;
1557
1558         case MDS_DISCONNECT:
1559                 DEBUG_REQ(D_INODE, req, "disconnect");
1560                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1561                 rc = target_handle_disconnect(req);
1562                 req->rq_status = rc;            /* superfluous? */
1563                 break;
1564
1565         case MDS_GETSTATUS:
1566                 DEBUG_REQ(D_INODE, req, "getstatus");
1567                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1568                 rc = mds_getstatus(req);
1569                 break;
1570
1571         case MDS_GETATTR:
1572                 DEBUG_REQ(D_INODE, req, "getattr");
1573                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1574                 rc = mds_getattr(req, REQ_REC_OFF);
1575                 break;
1576
1577         case MDS_SETXATTR:
1578                 DEBUG_REQ(D_INODE, req, "setxattr");
1579                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SETXATTR_NET, 0);
1580                 rc = mds_setxattr(req);
1581                 break;
1582
1583         case MDS_GETXATTR:
1584                 DEBUG_REQ(D_INODE, req, "getxattr");
1585                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETXATTR_NET, 0);
1586                 rc = mds_getxattr(req);
1587                 break;
1588
1589         case MDS_GETATTR_NAME: {
1590                 struct lustre_handle lockh = { 0 };
1591                 DEBUG_REQ(D_INODE, req, "getattr_name");
1592                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1593
1594                 /* If this request gets a reconstructed reply, we won't be
1595                  * acquiring any new locks in mds_getattr_lock, so we don't
1596                  * want to cancel.
1597                  */
1598                 rc = mds_getattr_lock(req, REQ_REC_OFF, MDS_INODELOCK_UPDATE,
1599                                       &lockh);
1600                 /* this non-intent call (from an ioctl) is special */
1601                 req->rq_status = rc;
1602                 if (rc == 0 && lustre_handle_is_used(&lockh))
1603                         ldlm_lock_decref(&lockh, LCK_CR);
1604                 break;
1605         }
1606         case MDS_STATFS:
1607                 DEBUG_REQ(D_INODE, req, "statfs");
1608                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1609                 rc = mds_statfs(req);
1610                 break;
1611
1612         case MDS_READPAGE:
1613                 DEBUG_REQ(D_INODE, req, "readpage");
1614                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1615                 rc = mds_readpage(req, REQ_REC_OFF);
1616
1617                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1618                         RETURN(0);
1619                 }
1620
1621                 break;
1622
1623         case MDS_REINT: {
1624                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
1625                                              sizeof(*opcp));
1626                 __u32  opc;
1627                 int size[4] = { sizeof(struct ptlrpc_body),
1628                                 sizeof(struct mds_body),
1629                                 mds->mds_max_mdsize,
1630                                 mds->mds_max_cookiesize };
1631                 int bufcount;
1632
1633                 /* NB only peek inside req now; mds_reint() will swab it */
1634                 if (opcp == NULL) {
1635                         CERROR ("Can't inspect opcode\n");
1636                         rc = -EINVAL;
1637                         break;
1638                 }
1639                 opc = *opcp;
1640                 if (lustre_msg_swabbed(req->rq_reqmsg))
1641                         __swab32s(&opc);
1642
1643                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1644                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1645                            reint_names[opc] == NULL) ? reint_names[opc] :
1646                                                        "unknown opcode");
1647
1648                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1649
1650                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1651                         bufcount = 4;
1652                 else if (opc == REINT_OPEN)
1653                         bufcount = 3;
1654                 else
1655                         bufcount = 2;
1656
1657                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1658                 if (rc)
1659                         break;
1660
1661                 rc = mds_reint(req, REQ_REC_OFF, NULL);
1662                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1663                 break;
1664         }
1665
1666         case MDS_CLOSE:
1667                 DEBUG_REQ(D_INODE, req, "close");
1668                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1669                 rc = mds_close(req, REQ_REC_OFF);
1670                 break;
1671
1672         case MDS_DONE_WRITING:
1673                 DEBUG_REQ(D_INODE, req, "done_writing");
1674                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1675                 rc = mds_done_writing(req, REQ_REC_OFF);
1676                 break;
1677
1678         case MDS_PIN:
1679                 DEBUG_REQ(D_INODE, req, "pin");
1680                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1681                 rc = mds_pin(req, REQ_REC_OFF);
1682                 break;
1683
1684         case MDS_SYNC:
1685                 DEBUG_REQ(D_INODE, req, "sync");
1686                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1687                 rc = mds_sync(req, REQ_REC_OFF);
1688                 break;
1689
1690         case MDS_SET_INFO:
1691                 DEBUG_REQ(D_INODE, req, "set_info");
1692                 rc = mds_set_info_rpc(req->rq_export, req);
1693                 break;
1694
1695         case MDS_QUOTACHECK:
1696                 DEBUG_REQ(D_INODE, req, "quotacheck");
1697                 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACHECK_NET, 0);
1698                 rc = mds_handle_quotacheck(req);
1699                 break;
1700
1701         case MDS_QUOTACTL:
1702                 DEBUG_REQ(D_INODE, req, "quotactl");
1703                 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACTL_NET, 0);
1704                 rc = mds_handle_quotactl(req);
1705                 break;
1706
1707         case OBD_PING:
1708                 DEBUG_REQ(D_INODE, req, "ping");
1709                 rc = target_handle_ping(req);
1710                 break;
1711
1712         case OBD_LOG_CANCEL:
1713                 CDEBUG(D_INODE, "log cancel\n");
1714                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1715                 rc = -ENOTSUPP; /* la la la */
1716                 break;
1717
1718         case LDLM_ENQUEUE:
1719                 DEBUG_REQ(D_INODE, req, "enqueue");
1720                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1721                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1722                                          ldlm_server_blocking_ast, NULL);
1723                 fail = OBD_FAIL_LDLM_REPLY;
1724                 break;
1725         case LDLM_CONVERT:
1726                 DEBUG_REQ(D_INODE, req, "convert");
1727                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1728                 rc = ldlm_handle_convert(req);
1729                 break;
1730         case LDLM_BL_CALLBACK:
1731         case LDLM_CP_CALLBACK:
1732                 DEBUG_REQ(D_INODE, req, "callback");
1733                 CERROR("callbacks should not happen on MDS\n");
1734                 LBUG();
1735                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1736                 break;
1737         case LLOG_ORIGIN_HANDLE_CREATE:
1738                 DEBUG_REQ(D_INODE, req, "llog_init");
1739                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1740                 rc = llog_origin_handle_create(req);
1741                 break;
1742         case LLOG_ORIGIN_HANDLE_DESTROY:
1743                 DEBUG_REQ(D_INODE, req, "llog_init");
1744                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1745                 rc = llog_origin_handle_destroy(req);
1746                 break;
1747         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1748                 DEBUG_REQ(D_INODE, req, "llog next block");
1749                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1750                 rc = llog_origin_handle_next_block(req);
1751                 break;
1752         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1753                 DEBUG_REQ(D_INODE, req, "llog prev block");
1754                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1755                 rc = llog_origin_handle_prev_block(req);
1756                 break;
1757         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1758                 DEBUG_REQ(D_INODE, req, "llog read header");
1759                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1760                 rc = llog_origin_handle_read_header(req);
1761                 break;
1762         case LLOG_ORIGIN_HANDLE_CLOSE:
1763                 DEBUG_REQ(D_INODE, req, "llog close");
1764                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1765                 rc = llog_origin_handle_close(req);
1766                 break;
1767         case LLOG_CATINFO:
1768                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1769                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1770                 rc = llog_catinfo(req);
1771                 break;
1772         default:
1773                 req->rq_status = -ENOTSUPP;
1774                 rc = ptlrpc_error(req);
1775                 RETURN(rc);
1776         }
1777
1778         LASSERT(current->journal_info == NULL);
1779
1780         /* If we're DISCONNECTing, the mds_export_data is already freed */
1781         if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != MDS_DISCONNECT) {
1782                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1783
1784                 /* I don't think last_xid is used for anyway, so I'm not sure
1785                    if we need to care about last_close_xid here.*/
1786                 lustre_msg_set_last_xid(req->rq_repmsg,
1787                                        le64_to_cpu(med->med_mcd->mcd_last_xid));
1788
1789                 target_committed_to_req(req);
1790         }
1791
1792         EXIT;
1793  out:
1794
1795         target_send_reply(req, rc, fail);
1796         return 0;
1797 }
1798
1799 /* Update the server data on disk.  This stores the new mount_count and
1800  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1801  * then the server last_rcvd value may be less than that of the clients.
1802  * This will alert us that we may need to do client recovery.
1803  *
1804  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1805  */
1806 int mds_update_server_data(struct obd_device *obd, int force_sync)
1807 {
1808         struct mds_obd *mds = &obd->u.mds;
1809         struct lr_server_data *lsd = mds->mds_server_data;
1810         struct file *filp = mds->mds_rcvd_filp;
1811         struct lvfs_run_ctxt saved;
1812         loff_t off = 0;
1813         int rc;
1814         ENTRY;
1815
1816         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1817                mds->mds_mount_count, mds->mds_last_transno);
1818
1819         spin_lock(&mds->mds_transno_lock);
1820         lsd->lsd_last_transno = cpu_to_le64(mds->mds_last_transno);
1821         spin_unlock(&mds->mds_transno_lock);
1822
1823         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1824         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off,force_sync);
1825         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1826         if (rc)
1827                 CERROR("error writing MDS server data: rc = %d\n", rc);
1828         RETURN(rc);
1829 }
1830
1831 static void fsoptions_to_mds_flags(struct mds_obd *mds, char *options)
1832 {
1833         char *p = options;
1834
1835         if (!options)
1836                 return;
1837
1838         while (*options) {
1839                 int len;
1840
1841                 while (*p && *p != ',')
1842                         p++;
1843
1844                 len = p - options;
1845                 if (len == sizeof("user_xattr") - 1 &&
1846                     memcmp(options, "user_xattr", len) == 0) {
1847                         mds->mds_fl_user_xattr = 1;
1848                         LCONSOLE_INFO("Enabling user_xattr\n");
1849                 } else if (len == sizeof("nouser_xattr") - 1 &&
1850                            memcmp(options, "nouser_xattr", len) == 0) {
1851                         mds->mds_fl_user_xattr = 0;
1852                         LCONSOLE_INFO("Disabling user_xattr\n");
1853                 } else if (len == sizeof("acl") - 1 &&
1854                            memcmp(options, "acl", len) == 0) {
1855 #ifdef CONFIG_FS_POSIX_ACL
1856                         mds->mds_fl_acl = 1;
1857                         LCONSOLE_INFO("Enabling ACL\n");
1858 #else
1859                         CWARN("ignoring unsupported acl mount option\n");
1860 #endif
1861                 } else if (len == sizeof("noacl") - 1 &&
1862                            memcmp(options, "noacl", len) == 0) {
1863 #ifdef CONFIG_FS_POSIX_ACL
1864                         mds->mds_fl_acl = 0;
1865                         LCONSOLE_INFO("Disabling ACL\n");
1866 #endif
1867                 }
1868
1869                 options = ++p;
1870         }
1871 }
1872 static int mds_lov_presetup (struct mds_obd *mds, struct lustre_cfg *lcfg)
1873 {
1874         int rc;
1875         ENTRY;
1876
1877         rc = llog_start_commit_thread();
1878         if (rc < 0)
1879                 RETURN(rc);
1880
1881         if (lcfg->lcfg_bufcount >= 4 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1882                 class_uuid_t uuid;
1883
1884                 ll_generate_random_uuid(uuid);
1885                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1886
1887                 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
1888                 if (mds->mds_profile == NULL)
1889                         RETURN(-ENOMEM);
1890
1891                 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
1892                         LUSTRE_CFG_BUFLEN(lcfg, 3));
1893         }
1894         RETURN(rc);
1895 }
1896
1897 /* mount the file system (secretly).  lustre_cfg parameters are:
1898  * 1 = device
1899  * 2 = fstype
1900  * 3 = config name
1901  * 4 = mount options
1902  */
1903 static int mds_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1904 {
1905         struct lprocfs_static_vars lvars;
1906         struct mds_obd *mds = &obd->u.mds;
1907         struct lustre_mount_info *lmi;
1908         struct vfsmount *mnt;
1909         struct lustre_sb_info *lsi;
1910         struct obd_uuid uuid;
1911         __u8 *uuid_ptr;
1912         char *str, *label;
1913         char ns_name[48];
1914         int rc = 0;
1915         ENTRY;
1916
1917         /* setup 1:/dev/loop/0 2:ext3 3:mdsA 4:errors=remount-ro,iopen_nopriv */
1918
1919         CLASSERT(offsetof(struct obd_device, u.obt) ==
1920                  offsetof(struct obd_device, u.mds.mds_obt));
1921
1922         if (lcfg->lcfg_bufcount < 3)
1923                 RETURN(-EINVAL);
1924
1925         if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
1926                 RETURN(-EINVAL);
1927
1928         lmi = server_get_mount(obd->obd_name);
1929         if (!lmi) {
1930                 CERROR("Not mounted in lustre_fill_super?\n");
1931                 RETURN(-EINVAL);
1932         }
1933
1934         /* We mounted in lustre_fill_super.
1935            lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1936
1937         lsi = s2lsi(lmi->lmi_sb);
1938         fsoptions_to_mds_flags(mds, lsi->lsi_ldd->ldd_mount_opts);
1939         fsoptions_to_mds_flags(mds, lsi->lsi_lmd->lmd_opts);
1940         mnt = lmi->lmi_mnt;
1941         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1942         if (IS_ERR(obd->obd_fsops))
1943                 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
1944
1945         CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
1946
1947         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1948
1949         sema_init(&mds->mds_epoch_sem, 1);
1950         spin_lock_init(&mds->mds_transno_lock);
1951         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1952         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1953         mds->mds_atime_diff = MAX_ATIME_DIFF;
1954
1955         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1956         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
1957         if (obd->obd_namespace == NULL) {
1958                 mds_cleanup(obd);
1959                 GOTO(err_ops, rc = -ENOMEM);
1960         }
1961         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1962
1963         lprocfs_init_vars(mds, &lvars);
1964         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1965             lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
1966                 /* Init private stats here */
1967                 mds_stats_counter_init(obd->obd_stats);
1968                 obd->obd_proc_exports = proc_mkdir("exports",
1969                                                    obd->obd_proc_entry);
1970         }
1971
1972         rc = mds_fs_setup(obd, mnt);
1973         if (rc) {
1974                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
1975                        obd->obd_name, rc);
1976                 GOTO(err_ns, rc);
1977         }
1978
1979         rc = mds_lov_presetup(mds, lcfg);
1980         if (rc < 0)
1981                 GOTO(err_fs, rc);
1982
1983         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1984                            "mds_ldlm_client", &obd->obd_ldlm_client);
1985         obd->obd_replayable = 1;
1986
1987         rc = lquota_setup(mds_quota_interface_ref, obd);
1988         if (rc)
1989                 GOTO(err_fs, rc);
1990
1991 #if 0
1992         mds->mds_group_hash = upcall_cache_init(obd->obd_name);
1993         if (IS_ERR(mds->mds_group_hash)) {
1994                 rc = PTR_ERR(mds->mds_group_hash);
1995                 mds->mds_group_hash = NULL;
1996                 GOTO(err_qctxt, rc);
1997         }
1998 #endif
1999
2000         /* Don't wait for mds_postrecov trying to clear orphans */
2001         obd->obd_async_recov = 1;
2002         rc = mds_postsetup(obd);
2003         /* Bug 11557 - allow async abort_recov start
2004            FIXME can remove most of this obd_async_recov plumbing
2005         obd->obd_async_recov = 0;
2006         */
2007         if (rc)
2008                 GOTO(err_qctxt, rc);
2009
2010         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2011         if (uuid_ptr != NULL) {
2012                 class_uuid_unparse(uuid_ptr, &uuid);
2013                 str = uuid.uuid;
2014         } else {
2015                 str = "no UUID";
2016         }
2017
2018         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2019         if (obd->obd_recovering) {
2020                 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2021                               "recovery until %d %s reconnect, or if no clients"
2022                               " reconnect for %d:%.02d; during that time new "
2023                               "clients will not be allowed to connect. "
2024                               "Recovery progress can be monitored by watching "
2025                               "/proc/fs/lustre/mds/%s/recovery_status.\n",
2026                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2027                               label ?: "", label ? "/" : "", str,
2028                               obd->obd_max_recoverable_clients,
2029                               (obd->obd_max_recoverable_clients == 1) ?
2030                               "client" : "clients",
2031                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
2032                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
2033                               obd->obd_name);
2034         } else {
2035                 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2036                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2037                               label ?: "", label ? "/" : "", str,
2038                               obd->obd_replayable ? "enabled" : "disabled");
2039         }
2040
2041         if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
2042                 ldlm_timeout = 6;
2043
2044         RETURN(0);
2045
2046 err_qctxt:
2047         lquota_cleanup(mds_quota_interface_ref, obd);
2048 err_fs:
2049         /* No extra cleanup needed for llog_init_commit_thread() */
2050         mds_fs_cleanup(obd);
2051 #if 0
2052         upcall_cache_cleanup(mds->mds_group_hash);
2053         mds->mds_group_hash = NULL;
2054 #endif
2055 err_ns:
2056         lprocfs_obd_cleanup(obd);
2057         lprocfs_free_obd_stats(obd);
2058         ldlm_namespace_free(obd->obd_namespace, 0);
2059         obd->obd_namespace = NULL;
2060 err_ops:
2061         fsfilt_put_ops(obd->obd_fsops);
2062 err_put:
2063         server_put_mount(obd->obd_name, mnt);
2064         obd->u.obt.obt_sb = NULL;
2065         return rc;
2066 }
2067
2068 static int mds_lov_clean(struct obd_device *obd)
2069 {
2070         struct mds_obd *mds = &obd->u.mds;
2071         struct obd_device *osc = mds->mds_osc_obd;
2072         ENTRY;
2073
2074         if (mds->mds_profile) {
2075                 class_del_profile(mds->mds_profile);
2076                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2077                 mds->mds_profile = NULL;
2078         }
2079
2080         /* There better be a lov */
2081         if (!osc)
2082                 RETURN(0);
2083         if (IS_ERR(osc))
2084                 RETURN(PTR_ERR(osc));
2085
2086         obd_register_observer(osc, NULL);
2087
2088         /* Give lov our same shutdown flags */
2089         osc->obd_force = obd->obd_force;
2090         osc->obd_fail = obd->obd_fail;
2091
2092         /* Cleanup the lov */
2093         obd_disconnect(mds->mds_osc_exp);
2094         class_manual_cleanup(osc);
2095         mds->mds_osc_exp = NULL;
2096
2097         RETURN(0);
2098 }
2099
2100 static int mds_postsetup(struct obd_device *obd)
2101 {
2102         struct mds_obd *mds = &obd->u.mds;
2103         int rc = 0;
2104         ENTRY;
2105
2106         rc = llog_setup(obd, NULL, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2107                         &llog_lvfs_ops);
2108         if (rc)
2109                 RETURN(rc);
2110
2111         rc = llog_setup(obd, NULL, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2112                         &llog_lvfs_ops);
2113         if (rc)
2114                 RETURN(rc);
2115
2116         if (mds->mds_profile) {
2117                 struct lustre_profile *lprof;
2118                 /* The profile defines which osc and mdc to connect to, for a
2119                    client.  We reuse that here to figure out the name of the
2120                    lov to use (and ignore lprof->lp_md).
2121                    The profile was set in the config log with
2122                    LCFG_MOUNTOPT profilenm oscnm mdcnm */
2123                 lprof = class_get_profile(mds->mds_profile);
2124                 if (lprof == NULL) {
2125                         CERROR("No profile found: %s\n", mds->mds_profile);
2126                         GOTO(err_cleanup, rc = -ENOENT);
2127                 }
2128                 rc = mds_lov_connect(obd, lprof->lp_dt);
2129                 if (rc)
2130                         GOTO(err_cleanup, rc);
2131         }
2132
2133         RETURN(rc);
2134
2135 err_cleanup:
2136         mds_lov_clean(obd);
2137         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2138         llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2139         RETURN(rc);
2140 }
2141
2142 int mds_postrecov(struct obd_device *obd)
2143 {
2144         int rc;
2145         ENTRY;
2146
2147         if (obd->obd_fail)
2148                 RETURN(0);
2149
2150         LASSERT(!obd->obd_recovering);
2151         LASSERT(llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT) != NULL);
2152
2153         /* FIXME why not put this in the synchronize? */
2154         /* set nextid first, so we are sure it happens */
2155         rc = mds_lov_set_nextid(obd);
2156         if (rc) {
2157                 CERROR("%s: mds_lov_set_nextid failed %d\n",
2158                        obd->obd_name, rc);
2159                 GOTO(out, rc);
2160         }
2161
2162         /* clean PENDING dir */
2163         if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2164                 rc = mds_cleanup_pending(obd);
2165                 if (rc < 0)
2166                         GOTO(out, rc);
2167
2168         /* FIXME Does target_finish_recovery really need this to block? */
2169         /* Notify the LOV, which will in turn call mds_notify for each tgt */
2170         /* This means that we have to hack obd_notify to think we're obd_set_up
2171            during mds_lov_connect. */
2172         obd_notify(obd->u.mds.mds_osc_obd, NULL,
2173                    obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
2174                    OBD_NOTIFY_SYNC, NULL);
2175
2176         /* quota recovery */
2177         lquota_recovery(mds_quota_interface_ref, obd);
2178
2179 out:
2180         RETURN(rc);
2181 }
2182
2183 /* We need to be able to stop an mds_lov_synchronize */
2184 static int mds_lov_early_clean(struct obd_device *obd)
2185 {
2186         struct mds_obd *mds = &obd->u.mds;
2187         struct obd_device *osc = mds->mds_osc_obd;
2188
2189         if (!osc || (!obd->obd_force && !obd->obd_fail))
2190                 return(0);
2191
2192         CDEBUG(D_HA, "abort inflight\n");
2193         return (obd_precleanup(osc, OBD_CLEANUP_EARLY));
2194 }
2195
2196 static int mds_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2197 {
2198         int rc = 0;
2199         ENTRY;
2200
2201         switch (stage) {
2202         case OBD_CLEANUP_EARLY:
2203                 break;
2204         case OBD_CLEANUP_EXPORTS:
2205                 /*XXX Use this for mdd mds cleanup, so comment out
2206                  *this target_cleanup_recovery for this tmp MDD MDS
2207                  *Wangdi*/
2208                 if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2209                         target_cleanup_recovery(obd);
2210                 mds_lov_early_clean(obd);
2211                 break;
2212         case OBD_CLEANUP_SELF_EXP:
2213                 mds_lov_disconnect(obd);
2214                 mds_lov_clean(obd);
2215                 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2216                 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2217                 rc = obd_llog_finish(obd, 0);
2218                 break;
2219         case OBD_CLEANUP_OBD:
2220                 break;
2221         }
2222         RETURN(rc);
2223 }
2224
2225 static int mds_cleanup(struct obd_device *obd)
2226 {
2227         struct mds_obd *mds = &obd->u.mds;
2228         lvfs_sbdev_type save_dev;
2229         ENTRY;
2230
2231         if (obd->u.obt.obt_sb == NULL)
2232                 RETURN(0);
2233         save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2234
2235         if (mds->mds_osc_exp)
2236                 /* lov export was disconnected by mds_lov_clean;
2237                    we just need to drop our ref */
2238                 class_export_put(mds->mds_osc_exp);
2239
2240         lprocfs_obd_cleanup(obd);
2241         lprocfs_free_obd_stats(obd);
2242
2243         lquota_cleanup(mds_quota_interface_ref, obd);
2244
2245         mds_update_server_data(obd, 1);
2246         if (mds->mds_lov_objids != NULL)
2247                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2248         mds_fs_cleanup(obd);
2249
2250 #if 0
2251         upcall_cache_cleanup(mds->mds_group_hash);
2252         mds->mds_group_hash = NULL;
2253 #endif
2254
2255         server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2256         obd->u.obt.obt_sb = NULL;
2257
2258         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2259
2260         spin_lock_bh(&obd->obd_processing_task_lock);
2261         if (obd->obd_recovering) {
2262                 target_cancel_recovery_timer(obd);
2263                 obd->obd_recovering = 0;
2264         }
2265         spin_unlock_bh(&obd->obd_processing_task_lock);
2266
2267         fsfilt_put_ops(obd->obd_fsops);
2268
2269         LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2270
2271         RETURN(0);
2272 }
2273
2274 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2275                                         struct ldlm_lock *new_lock,
2276                                         struct ldlm_lock **old_lock,
2277                                         struct lustre_handle *lockh)
2278 {
2279         struct obd_export *exp = req->rq_export;
2280         struct ldlm_request *dlmreq =
2281                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2282         struct lustre_handle remote_hdl = dlmreq->lock_handle[0];
2283         struct list_head *iter;
2284
2285         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2286                 return;
2287
2288         spin_lock(&exp->exp_ldlm_data.led_lock);
2289         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2290                 struct ldlm_lock *lock;
2291                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2292                 if (lock == new_lock)
2293                         continue;
2294                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2295                         lockh->cookie = lock->l_handle.h_cookie;
2296                         LDLM_DEBUG(lock, "restoring lock cookie");
2297                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2298                                   lockh->cookie);
2299                         if (old_lock)
2300                                 *old_lock = LDLM_LOCK_GET(lock);
2301                         spin_unlock(&exp->exp_ldlm_data.led_lock);
2302                         return;
2303                 }
2304         }
2305         spin_unlock(&exp->exp_ldlm_data.led_lock);
2306
2307         /* If the xid matches, then we know this is a resent request,
2308          * and allow it. (It's probably an OPEN, for which we don't
2309          * send a lock */
2310         if (req->rq_xid ==
2311             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2312                 return;
2313
2314         if (req->rq_xid ==
2315             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2316                 return;
2317
2318         /* This remote handle isn't enqueued, so we never received or
2319          * processed this request.  Clear MSG_RESENT, because it can
2320          * be handled like any normal request now. */
2321
2322         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2323
2324         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2325                   remote_hdl.cookie);
2326 }
2327
2328 int intent_disposition(struct ldlm_reply *rep, int flag)
2329 {
2330         if (!rep)
2331                 return 0;
2332         return (rep->lock_policy_res1 & flag);
2333 }
2334
2335 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2336 {
2337         if (!rep)
2338                 return;
2339         rep->lock_policy_res1 |= flag;
2340 }
2341
2342 static int mds_intent_policy(struct ldlm_namespace *ns,
2343                              struct ldlm_lock **lockp, void *req_cookie,
2344                              ldlm_mode_t mode, int flags, void *data)
2345 {
2346         struct ptlrpc_request *req = req_cookie;
2347         struct ldlm_lock *lock = *lockp;
2348         struct ldlm_intent *it;
2349         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2350         struct ldlm_reply *rep;
2351         struct lustre_handle lockh = { 0 };
2352         struct ldlm_lock *new_lock = NULL;
2353         int getattr_part = MDS_INODELOCK_UPDATE;
2354         int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2355                            [DLM_LOCKREPLY_OFF]   = sizeof(struct ldlm_reply),
2356                            [DLM_REPLY_REC_OFF]   = sizeof(struct mds_body),
2357                            [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2358         int repbufcnt = 4, rc;
2359         ENTRY;
2360
2361         LASSERT(req != NULL);
2362
2363         if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2364                 /* No intent was provided */
2365                 rc = lustre_pack_reply(req, 2, repsize, NULL);
2366                 LASSERT(rc == 0);
2367                 RETURN(0);
2368         }
2369
2370         it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2371                                 lustre_swab_ldlm_intent);
2372         if (it == NULL) {
2373                 CERROR("Intent missing\n");
2374                 RETURN(req->rq_status = -EFAULT);
2375         }
2376
2377         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2378
2379         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2380             (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2381                 /* we should never allow OBD_CONNECT_ACL if not configured */
2382                 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2383         else if (it->opc & IT_UNLINK)
2384                 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2385
2386         rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2387         if (rc)
2388                 RETURN(req->rq_status = rc);
2389
2390         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2391         intent_set_disposition(rep, DISP_IT_EXECD);
2392
2393
2394         /* execute policy */
2395         switch ((long)it->opc) {
2396         case IT_OPEN:
2397         case IT_CREAT|IT_OPEN:
2398                 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2399                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2400                                             &lockh);
2401                 /* XXX swab here to assert that an mds_open reint
2402                  * packet is following */
2403                 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2404                                                   &lockh);
2405 #if 0
2406                 /* We abort the lock if the lookup was negative and
2407                  * we did not make it to the OPEN portion */
2408                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2409                         RETURN(ELDLM_LOCK_ABORTED);
2410                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2411                     !intent_disposition(rep, DISP_OPEN_OPEN))
2412 #endif
2413
2414                 /* If there was an error of some sort or if we are not
2415                  * returning any locks */
2416                 if (rep->lock_policy_res2 ||
2417                     !intent_disposition(rep, DISP_OPEN_LOCK))
2418                         RETURN(ELDLM_LOCK_ABORTED);
2419                 break;
2420         case IT_LOOKUP:
2421                         getattr_part = MDS_INODELOCK_LOOKUP;
2422         case IT_GETATTR:
2423                         getattr_part |= MDS_INODELOCK_LOOKUP;
2424                         OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2425         case IT_READDIR:
2426                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2427                                             &new_lock, &lockh);
2428
2429                 /* INODEBITS_INTEROP: if this lock was converted from a
2430                  * plain lock (client does not support inodebits), then
2431                  * child lock must be taken with both lookup and update
2432                  * bits set for all operations.
2433                  */
2434                 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2435                         getattr_part = MDS_INODELOCK_LOOKUP |
2436                                        MDS_INODELOCK_UPDATE;
2437
2438                 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2439                                                          getattr_part, &lockh);
2440                 /* FIXME: LDLM can set req->rq_status. MDS sets
2441                    policy_res{1,2} with disposition and status.
2442                    - replay: returns 0 & req->status is old status
2443                    - otherwise: returns req->status */
2444                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2445                         rep->lock_policy_res2 = 0;
2446                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2447                     rep->lock_policy_res2)
2448                         RETURN(ELDLM_LOCK_ABORTED);
2449                 if (req->rq_status != 0) {
2450                         LBUG();
2451                         rep->lock_policy_res2 = req->rq_status;
2452                         RETURN(ELDLM_LOCK_ABORTED);
2453                 }
2454                 break;
2455         default:
2456                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2457                 RETURN(-EFAULT);
2458         }
2459
2460         /* By this point, whatever function we called above must have either
2461          * filled in 'lockh', been an intent replay, or returned an error.  We
2462          * want to allow replayed RPCs to not get a lock, since we would just
2463          * drop it below anyways because lock replay is done separately by the
2464          * client afterwards.  For regular RPCs we want to give the new lock to
2465          * the client instead of whatever lock it was about to get. */
2466         if (new_lock == NULL)
2467                 new_lock = ldlm_handle2lock(&lockh);
2468         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2469                 RETURN(0);
2470
2471         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2472                  it->opc, lockh.cookie);
2473
2474         /* If we've already given this lock to a client once, then we should
2475          * have no readers or writers.  Otherwise, we should have one reader
2476          * _or_ writer ref (which will be zeroed below) before returning the
2477          * lock to a client. */
2478         if (new_lock->l_export == req->rq_export) {
2479                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2480         } else {
2481                 LASSERT(new_lock->l_export == NULL);
2482                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2483         }
2484
2485         *lockp = new_lock;
2486
2487         if (new_lock->l_export == req->rq_export) {
2488                 /* Already gave this to the client, which means that we
2489                  * reconstructed a reply. */
2490                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2491                         MSG_RESENT);
2492                 RETURN(ELDLM_LOCK_REPLACED);
2493         }
2494
2495         /* Fixup the lock to be given to the client */
2496         lock_res_and_lock(new_lock);
2497         new_lock->l_readers = 0;
2498         new_lock->l_writers = 0;
2499
2500         new_lock->l_export = class_export_get(req->rq_export);
2501         spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2502         list_add(&new_lock->l_export_chain,
2503                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2504         spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2505
2506         new_lock->l_blocking_ast = lock->l_blocking_ast;
2507         new_lock->l_completion_ast = lock->l_completion_ast;
2508
2509         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2510                sizeof(lock->l_remote_handle));
2511
2512         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2513
2514         unlock_res_and_lock(new_lock);
2515         LDLM_LOCK_PUT(new_lock);
2516
2517         RETURN(ELDLM_LOCK_REPLACED);
2518 }
2519
2520 static int mdt_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2521 {
2522         struct mds_obd *mds = &obd->u.mds;
2523         struct lprocfs_static_vars lvars;
2524         int mds_min_threads;
2525         int mds_max_threads;
2526         int rc = 0;
2527         ENTRY;
2528
2529         lprocfs_init_vars(mdt, &lvars);
2530         lprocfs_obd_setup(obd, lvars.obd_vars);
2531
2532         sema_init(&mds->mds_health_sem, 1);
2533
2534         if (mds_num_threads) {
2535                 /* If mds_num_threads is set, it is the min and the max. */
2536                 if (mds_num_threads > MDS_THREADS_MAX)
2537                         mds_num_threads = MDS_THREADS_MAX;
2538                 if (mds_num_threads < MDS_THREADS_MIN)
2539                         mds_num_threads = MDS_THREADS_MIN;
2540                 mds_max_threads = mds_min_threads = mds_num_threads;
2541         } else {
2542                 /* Base min threads on memory and cpus */
2543                 mds_min_threads = smp_num_cpus * num_physpages >>
2544                         (27 - CFS_PAGE_SHIFT);
2545                 if (mds_min_threads < MDS_THREADS_MIN)
2546                         mds_min_threads = MDS_THREADS_MIN;
2547                 /* Largest auto threads start value */
2548                 if (mds_min_threads > 32)
2549                         mds_min_threads = 32;
2550                 mds_max_threads = min(MDS_THREADS_MAX, mds_min_threads * 4);
2551         }
2552
2553         mds->mds_service =
2554                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2555                                 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2556                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2557                                 mds_handle, LUSTRE_MDS_NAME,
2558                                 obd->obd_proc_entry, NULL,
2559                                 mds_min_threads, mds_max_threads, "ll_mdt", 0);
2560
2561         if (!mds->mds_service) {
2562                 CERROR("failed to start service\n");
2563                 GOTO(err_lprocfs, rc = -ENOMEM);
2564         }
2565
2566         rc = ptlrpc_start_threads(obd, mds->mds_service);
2567         if (rc)
2568                 GOTO(err_thread, rc);
2569
2570         mds->mds_setattr_service =
2571                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2572                                 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2573                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2574                                 mds_handle, "mds_setattr",
2575                                 obd->obd_proc_entry, NULL,
2576                                 mds_min_threads, mds_max_threads,
2577                                 "ll_mdt_attr", 0);
2578         if (!mds->mds_setattr_service) {
2579                 CERROR("failed to start getattr service\n");
2580                 GOTO(err_thread, rc = -ENOMEM);
2581         }
2582
2583         rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2584         if (rc)
2585                 GOTO(err_thread2, rc);
2586
2587         mds->mds_readpage_service =
2588                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2589                                 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2590                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2591                                 mds_handle, "mds_readpage",
2592                                 obd->obd_proc_entry, NULL,
2593                                 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2594                                 "ll_mdt_rdpg", 0);
2595         if (!mds->mds_readpage_service) {
2596                 CERROR("failed to start readpage service\n");
2597                 GOTO(err_thread2, rc = -ENOMEM);
2598         }
2599
2600         rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2601
2602         if (rc)
2603                 GOTO(err_thread3, rc);
2604
2605         ping_evictor_start();
2606
2607         RETURN(0);
2608
2609 err_thread3:
2610         ptlrpc_unregister_service(mds->mds_readpage_service);
2611         mds->mds_readpage_service = NULL;
2612 err_thread2:
2613         ptlrpc_unregister_service(mds->mds_setattr_service);
2614         mds->mds_setattr_service = NULL;
2615 err_thread:
2616         ptlrpc_unregister_service(mds->mds_service);
2617         mds->mds_service = NULL;
2618 err_lprocfs:
2619         lprocfs_obd_cleanup(obd);
2620         return rc;
2621 }
2622
2623 static int mdt_cleanup(struct obd_device *obd)
2624 {
2625         struct mds_obd *mds = &obd->u.mds;
2626         ENTRY;
2627
2628         ping_evictor_stop();
2629
2630         down(&mds->mds_health_sem);
2631         ptlrpc_unregister_service(mds->mds_readpage_service);
2632         ptlrpc_unregister_service(mds->mds_setattr_service);
2633         ptlrpc_unregister_service(mds->mds_service);
2634         mds->mds_readpage_service = NULL;
2635         mds->mds_setattr_service = NULL;
2636         mds->mds_service = NULL;
2637         up(&mds->mds_health_sem);
2638
2639         lprocfs_obd_cleanup(obd);
2640
2641         RETURN(0);
2642 }
2643
2644 static int mdt_health_check(struct obd_device *obd)
2645 {
2646         struct mds_obd *mds = &obd->u.mds;
2647         int rc = 0;
2648
2649         down(&mds->mds_health_sem);
2650         rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2651         rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2652         rc |= ptlrpc_service_health_check(mds->mds_service);
2653         up(&mds->mds_health_sem);
2654
2655         /*
2656          * health_check to return 0 on healthy
2657          * and 1 on unhealthy.
2658          */
2659         if(rc != 0)
2660                 rc = 1;
2661
2662         return rc;
2663 }
2664
2665 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2666                                           void *data)
2667 {
2668         struct obd_device *obd = data;
2669         struct ll_fid fid;
2670         fid.id = id;
2671         fid.generation = gen;
2672         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2673 }
2674
2675 static int mds_health_check(struct obd_device *obd)
2676 {
2677         struct obd_device_target *odt = &obd->u.obt;
2678 #ifdef USE_HEALTH_CHECK_WRITE
2679         struct mds_obd *mds = &obd->u.mds;
2680 #endif
2681         int rc = 0;
2682
2683         if (odt->obt_sb->s_flags & MS_RDONLY)
2684                 rc = 1;
2685
2686 #ifdef USE_HEALTH_CHECK_WRITE
2687         LASSERT(mds->mds_health_check_filp != NULL);
2688         rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2689 #endif
2690         return rc;
2691 }
2692
2693 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2694 {
2695         struct lustre_cfg *lcfg = buf;
2696         struct lprocfs_static_vars lvars;
2697         int rc;
2698
2699         lprocfs_init_vars(mds, &lvars);
2700
2701         rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2702         return(rc);
2703 }
2704
2705 struct lvfs_callback_ops mds_lvfs_ops = {
2706         l_fid2dentry:     mds_lvfs_fid2dentry,
2707 };
2708
2709 /* use obd ops to offer management infrastructure */
2710 static struct obd_ops mds_obd_ops = {
2711         .o_owner           = THIS_MODULE,
2712         .o_connect         = mds_connect,
2713         .o_reconnect       = mds_reconnect,
2714         .o_init_export     = mds_init_export,
2715         .o_destroy_export  = mds_destroy_export,
2716         .o_disconnect      = mds_disconnect,
2717         .o_setup           = mds_setup,
2718         .o_precleanup      = mds_precleanup,
2719         .o_cleanup         = mds_cleanup,
2720         .o_postrecov       = mds_postrecov,
2721         .o_statfs          = mds_obd_statfs,
2722         .o_iocontrol       = mds_iocontrol,
2723         .o_create          = mds_obd_create,
2724         .o_destroy         = mds_obd_destroy,
2725         .o_llog_init       = mds_llog_init,
2726         .o_llog_finish     = mds_llog_finish,
2727         .o_notify          = mds_notify,
2728         .o_health_check    = mds_health_check,
2729         .o_process_config  = mds_process_config,
2730 };
2731
2732 static struct obd_ops mdt_obd_ops = {
2733         .o_owner           = THIS_MODULE,
2734         .o_setup           = mdt_setup,
2735         .o_cleanup         = mdt_cleanup,
2736         .o_health_check    = mdt_health_check,
2737 };
2738
2739 quota_interface_t *mds_quota_interface_ref;
2740 extern quota_interface_t mds_quota_interface;
2741
2742 static __attribute__((unused)) int __init mds_init(void)
2743 {
2744         int rc;
2745         struct lprocfs_static_vars lvars;
2746
2747         request_module("lquota");
2748         mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2749         rc = lquota_init(mds_quota_interface_ref);
2750         if (rc) {
2751                 if (mds_quota_interface_ref)
2752                         PORTAL_SYMBOL_PUT(mds_quota_interface);
2753                 return rc;
2754         }
2755         init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2756
2757         lprocfs_init_vars(mds, &lvars);
2758         class_register_type(&mds_obd_ops, NULL,
2759                             lvars.module_vars, LUSTRE_MDS_NAME, NULL);
2760         lprocfs_init_vars(mdt, &lvars);
2761         mdt_obd_ops = mdt_obd_ops; //make compiler happy
2762 //        class_register_type(&mdt_obd_ops, NULL,
2763 //                            lvars.module_vars, LUSTRE_MDT_NAME, NULL);
2764
2765         return 0;
2766 }
2767
2768 static __attribute__((unused)) void /*__exit*/ mds_exit(void)
2769 {
2770         lquota_exit(mds_quota_interface_ref);
2771         if (mds_quota_interface_ref)
2772                 PORTAL_SYMBOL_PUT(mds_quota_interface);
2773
2774         class_unregister_type(LUSTRE_MDS_NAME);
2775 //        class_unregister_type(LUSTRE_MDT_NAME);
2776 }
2777 /*mds still need lov setup here*/
2778 static int mds_cmd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2779 {
2780         struct mds_obd *mds = &obd->u.mds;
2781         struct lvfs_run_ctxt saved;
2782         const char     *dev;
2783         struct vfsmount *mnt;
2784         struct lustre_sb_info *lsi;
2785         struct lustre_mount_info *lmi;
2786         struct dentry  *dentry;
2787         struct file *file;
2788         int rc = 0;
2789         ENTRY;
2790
2791         CDEBUG(D_INFO, "obd %s setup \n", obd->obd_name);
2792         if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2793                 RETURN(0);
2794
2795         if (lcfg->lcfg_bufcount < 5) {
2796                 CERROR("invalid arg for setup %s\n", MDD_OBD_NAME);
2797                 RETURN(-EINVAL);
2798         }
2799         dev = lustre_cfg_string(lcfg, 4);
2800         lmi = server_get_mount(dev);
2801         LASSERT(lmi != NULL);
2802
2803         lsi = s2lsi(lmi->lmi_sb);
2804         mnt = lmi->lmi_mnt;
2805         /* FIXME: MDD LOV initialize objects.
2806          * we need only lmi here but not get mount
2807          * OSD did mount already, so put mount back
2808          */
2809         atomic_dec(&lsi->lsi_mounts);
2810         mntput(mnt);
2811
2812         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2813         mds_init_ctxt(obd, mnt);
2814
2815         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2816         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
2817         if (IS_ERR(dentry)) {
2818                 rc = PTR_ERR(dentry);
2819                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
2820                 GOTO(err_putfs, rc);
2821         }
2822         mds->mds_objects_dir = dentry;
2823
2824         dentry = lookup_one_len("__iopen__", current->fs->pwd,
2825                                 strlen("__iopen__"));
2826         if (IS_ERR(dentry)) {
2827                 rc = PTR_ERR(dentry);
2828                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
2829                 GOTO(err_objects, rc);
2830         }
2831
2832         mds->mds_fid_de = dentry;
2833         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
2834                 rc = -ENOENT;
2835                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
2836                 GOTO(err_fid, rc);
2837         }
2838
2839         /* open and test the lov objd file */
2840         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
2841         if (IS_ERR(file)) {
2842                 rc = PTR_ERR(file);
2843                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
2844                 GOTO(err_fid, rc = PTR_ERR(file));
2845         }
2846         mds->mds_lov_objid_filp = file;
2847         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
2848                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
2849                        file->f_dentry->d_inode->i_mode);
2850                 GOTO(err_lov_objid, rc = -ENOENT);
2851         }
2852
2853         rc = mds_lov_presetup(mds, lcfg);
2854         if (rc < 0)
2855                 GOTO(err_objects, rc);
2856
2857         /* Don't wait for mds_postrecov trying to clear orphans */
2858         obd->obd_async_recov = 1;
2859         rc = mds_postsetup(obd);
2860         /* Bug 11557 - allow async abort_recov start
2861            FIXME can remove most of this obd_async_recov plumbing
2862         obd->obd_async_recov = 0;
2863         */
2864
2865         if (rc)
2866                 GOTO(err_objects, rc);
2867
2868         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2869         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2870
2871 err_pop:
2872         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2873         RETURN(rc);
2874 err_lov_objid:
2875         if (mds->mds_lov_objid_filp &&
2876                 filp_close((struct file *)mds->mds_lov_objid_filp, 0))
2877                 CERROR("can't close %s after error\n", LOV_OBJID);
2878 err_fid:
2879         dput(mds->mds_fid_de);
2880 err_objects:
2881         dput(mds->mds_objects_dir);
2882 err_putfs:
2883         fsfilt_put_ops(obd->obd_fsops);
2884         goto err_pop;
2885 }
2886
2887 static int mds_cmd_cleanup(struct obd_device *obd)
2888 {
2889         struct mds_obd *mds = &obd->u.mds;
2890         struct lvfs_run_ctxt saved;
2891         int rc = 0;
2892         ENTRY;
2893
2894         if (obd->obd_fail)
2895                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2896                               "will be preserved.\n", obd->obd_name);
2897
2898         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2899         if (mds->mds_lov_objid_filp) {
2900                 rc = filp_close((struct file *)mds->mds_lov_objid_filp, 0);
2901                 mds->mds_lov_objid_filp = NULL;
2902                 if (rc)
2903                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
2904         }
2905         if (mds->mds_objects_dir != NULL) {
2906                 l_dput(mds->mds_objects_dir);
2907                 mds->mds_objects_dir = NULL;
2908         }
2909
2910         if (mds->mds_lov_objids != NULL)
2911                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2912
2913         shrink_dcache_parent(mds->mds_fid_de);
2914         dput(mds->mds_fid_de);
2915         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2916         fsfilt_put_ops(obd->obd_fsops);
2917
2918         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2919         RETURN(rc);
2920 }
2921
2922 #if 0
2923 static int mds_cmd_health_check(struct obd_device *obd)
2924 {
2925         return 0;
2926 }
2927 #endif
2928 static struct obd_ops mds_cmd_obd_ops = {
2929         .o_owner           = THIS_MODULE,
2930         .o_setup           = mds_cmd_setup,
2931         .o_cleanup         = mds_cmd_cleanup,
2932         .o_precleanup      = mds_precleanup,
2933         .o_create          = mds_obd_create,
2934         .o_destroy         = mds_obd_destroy,
2935         .o_llog_init       = mds_llog_init,
2936         .o_llog_finish     = mds_llog_finish,
2937         .o_notify          = mds_notify,
2938         .o_postrecov       = mds_postrecov,
2939         //   .o_health_check    = mds_cmd_health_check,
2940 };
2941
2942 static int __init mds_cmd_init(void)
2943 {
2944         struct lprocfs_static_vars lvars;
2945
2946         lprocfs_init_vars(mds, &lvars);
2947         class_register_type(&mds_cmd_obd_ops, NULL, lvars.module_vars,
2948                             LUSTRE_MDS_NAME, NULL);
2949
2950         return 0;
2951 }
2952
2953 static void /*__exit*/ mds_cmd_exit(void)
2954 {
2955         class_unregister_type(LUSTRE_MDS_NAME);
2956 }
2957
2958 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2959 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2960 MODULE_LICENSE("GPL");
2961
2962 module_init(mds_cmd_init);
2963 module_exit(mds_cmd_exit);