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                        i_size_read(file->f_dentry->d_inode));
 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 (i_size_read(inode) + 1 != body->eadatasize)
 801                         CERROR("symlink size: %Lu, reply space: %d\n",
 802                                i_size_read(inode) + 1, body->eadatasize);
 803                 size[bufcount] = min_t(int, i_size_read(inode) + 1,
 804                                        body->eadatasize);
 805                 bufcount++;
 806                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 807                        i_size_read(inode) + 1, body->eadatasize);
 808         }
 809
 810 #ifdef CONFIG_FS_POSIX_ACL
 811         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
 812             (body->valid & OBD_MD_FLACL)) {
 813                 struct dentry de = { .d_inode = inode };
 814
 815                 size[bufcount] = 0;
 816                 if (inode->i_op && inode->i_op->getxattr) {
 817                         lock_24kernel();
 818                         rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
 819                                                    NULL, 0);
 820                         unlock_24kernel();
 821
 822                         if (rc < 0) {
 823                                 if (rc != -ENODATA) {
 824                                         CERROR("got acl size: %d\n", rc);
 825                                         RETURN(rc);
 826                                 }
 827                         } else
 828                                 size[bufcount] = rc;
 829                 }
 830                 bufcount++;
 831         }
 832 #endif
 833
 834         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 835                 CERROR("failed MDS_GETATTR_PACK test\n");
 836                 req->rq_status = -ENOMEM;
 837                 RETURN(-ENOMEM);
 838         }
 839
 840         rc = lustre_pack_reply(req, bufcount, size, NULL);
 841         if (rc) {
 842                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
 843                 req->rq_status = rc;
 844                 RETURN(rc);
 845         }
 846
 847         RETURN(0);
 848 }
 849
 850 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
 851                             int child_part, struct lustre_handle *child_lockh)
 852 {
 853         struct obd_device *obd = req->rq_export->exp_obd;
 854         struct mds_obd *mds = &obd->u.mds;
 855         struct ldlm_reply *rep = NULL;
 856         struct lvfs_run_ctxt saved;
 857         struct mds_body *body;
 858         struct dentry *dparent = NULL, *dchild = NULL;
 859         struct lvfs_ucred uc = {0,};
 860         struct lustre_handle parent_lockh;
 861         int namesize;
 862         int rc = 0, cleanup_phase = 0, resent_req = 0;
 863         char *name;
 864         ENTRY;
 865
 866         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 867
 868         /* Swab now, before anyone looks inside the request */
 869         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 870                                   lustre_swab_mds_body);
 871         if (body == NULL) {
 872                 CERROR("Can't swab mds_body\n");
 873                 RETURN(-EFAULT);
 874         }
 875
 876         LASSERT_REQSWAB(req, offset + 1);
 877         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 878         if (name == NULL) {
 879                 CERROR("Can't unpack name\n");
 880                 RETURN(-EFAULT);
 881         }
 882         namesize = lustre_msg_buflen(req->rq_reqmsg, offset + 1);
 883         /* namesize less than 2 means we have empty name, probably came from
 884            revalidate by cfid, so no point in having name to be set */
 885         if (namesize <= 1)
 886                 name = NULL;
 887
 888         rc = mds_init_ucred(&uc, req, offset);
 889         if (rc)
 890                 GOTO(cleanup, rc);
 891
 892         LASSERT(offset == REQ_REC_OFF || offset == DLM_INTENT_REC_OFF);
 893         /* if requests were at offset 2, the getattr reply goes back at 1 */
 894         if (offset == DLM_INTENT_REC_OFF) {
 895                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 896                                      sizeof(*rep));
 897                 offset = DLM_REPLY_REC_OFF;
 898         }
 899
 900         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 901         cleanup_phase = 1; /* kernel context */
 902         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 903
 904         /* FIXME: handle raw lookup */
 905 #if 0
 906         if (body->valid == OBD_MD_FLID) {
 907                 struct mds_body *mds_reply;
 908                 int size = sizeof(*mds_reply);
 909                 ino_t inum;
 910                 // The user requested ONLY the inode number, so do a raw lookup
 911                 rc = lustre_pack_reply(req, 1, &size, NULL);
 912                 if (rc) {
 913                         CERROR("out of memory\n");
 914                         GOTO(cleanup, rc);
 915                 }
 916
 917                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 918
 919                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 920                                            sizeof(*mds_reply));
 921                 mds_reply->fid1.id = inum;
 922                 mds_reply->valid = OBD_MD_FLID;
 923                 GOTO(cleanup, rc);
 924         }
 925 #endif
 926
 927         if (lustre_handle_is_used(child_lockh)) {
 928                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 929                 resent_req = 1;
 930         }
 931
 932         if (resent_req == 0) {
 933                 if (name) {
 934                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
 935                         rc = mds_get_parent_child_locked(obd, &obd->u.mds,
 936                                                          &body->fid1,
 937                                                          &parent_lockh,
 938                                                          &dparent, LCK_CR,
 939                                                          MDS_INODELOCK_UPDATE,
 940                                                          name, namesize,
 941                                                          child_lockh, &dchild,
 942                                                          LCK_CR, child_part);
 943                 } else {
 944                         /* For revalidate by fid we always take UPDATE lock */
 945                         dchild = mds_fid2locked_dentry(obd, &body->fid2, NULL,
 946                                                        LCK_CR, child_lockh,
 947                                                        child_part);
 948                         LASSERT(dchild);
 949                         if (IS_ERR(dchild))
 950                                 rc = PTR_ERR(dchild);
 951                 }
 952                 if (rc)
 953                         GOTO(cleanup, rc);
 954         } else {
 955                 struct ldlm_lock *granted_lock;
 956                 struct ll_fid child_fid;
 957                 struct ldlm_resource *res;
 958                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 959                 granted_lock = ldlm_handle2lock(child_lockh);
 960                 LASSERTF(granted_lock != NULL, LPU64"/%u lockh "LPX64"\n",
 961                          body->fid1.id, body->fid1.generation,
 962                          child_lockh->cookie);
 963
 964
 965                 res = granted_lock->l_resource;
 966                 child_fid.id = res->lr_name.name[0];
 967                 child_fid.generation = res->lr_name.name[1];
 968                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 969                 LASSERT(!IS_ERR(dchild));
 970                 LDLM_LOCK_PUT(granted_lock);
 971         }
 972
 973         cleanup_phase = 2; /* dchild, dparent, locks */
 974
 975         if (dchild->d_inode == NULL) {
 976                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 977                 /* in the intent case, the policy clears this error:
 978                    the disposition is enough */
 979                 GOTO(cleanup, rc = -ENOENT);
 980         } else {
 981                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 982         }
 983
 984         if (req->rq_repmsg == NULL) {
 985                 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 986                 if (rc != 0) {
 987                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 988                         GOTO (cleanup, rc);
 989                 }
 990         }
 991
 992         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 993         GOTO(cleanup, rc); /* returns the lock to the client */
 994
 995  cleanup:
 996         switch (cleanup_phase) {
 997         case 2:
 998                 if (resent_req == 0) {
 999                         if (rc && dchild->d_inode)
1000                                 ldlm_lock_decref(child_lockh, LCK_CR);
1001                         if (name) {
1002                                 ldlm_lock_decref(&parent_lockh, LCK_CR);
1003                                 l_dput(dparent);
1004                         }
1005                 }
1006                 l_dput(dchild);
1007         case 1:
1008                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1009         default:
1010                 mds_exit_ucred(&uc, mds);
1011                 if (req->rq_reply_state == NULL) {
1012                         req->rq_status = rc;
1013                         lustre_pack_reply(req, 1, NULL, NULL);
1014                 }
1015         }
1016         return rc;
1017 }
1018
1019 static int mds_getattr(struct ptlrpc_request *req, int offset)
1020 {
1021         struct mds_obd *mds = mds_req2mds(req);
1022         struct obd_device *obd = req->rq_export->exp_obd;
1023         struct lvfs_run_ctxt saved;
1024         struct dentry *de;
1025         struct mds_body *body;
1026         struct lvfs_ucred uc = {0,};
1027         int rc = 0;
1028         ENTRY;
1029
1030         OBD_COUNTER_INCREMENT(obd, getattr);
1031
1032         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1033                                   lustre_swab_mds_body);
1034         if (body == NULL)
1035                 RETURN(-EFAULT);
1036
1037         rc = mds_init_ucred(&uc, req, offset);
1038         if (rc)
1039                 GOTO(out_ucred, rc);
1040
1041         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1042         de = mds_fid2dentry(mds, &body->fid1, NULL);
1043         if (IS_ERR(de)) {
1044                 rc = req->rq_status = PTR_ERR(de);
1045                 GOTO(out_pop, rc);
1046         }
1047
1048         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1049         if (rc != 0) {
1050                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1051                 GOTO(out_pop, rc);
1052         }
1053
1054         req->rq_status = mds_getattr_internal(obd, de, req, body,
1055                                               REPLY_REC_OFF);
1056
1057         l_dput(de);
1058         GOTO(out_pop, rc);
1059 out_pop:
1060         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1061 out_ucred:
1062         if (req->rq_reply_state == NULL) {
1063                 req->rq_status = rc;
1064                 lustre_pack_reply(req, 1, NULL, NULL);
1065         }
1066         mds_exit_ucred(&uc, mds);
1067         return rc;
1068 }
1069
1070 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1071                           __u64 max_age)
1072 {
1073         int rc;
1074
1075         spin_lock(&obd->obd_osfs_lock);
1076         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
1077         if (rc == 0)
1078                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1079         spin_unlock(&obd->obd_osfs_lock);
1080
1081         return rc;
1082 }
1083
1084 static int mds_statfs(struct ptlrpc_request *req)
1085 {
1086         struct obd_device *obd = req->rq_export->exp_obd;
1087         int rc, size[2] = { sizeof(struct ptlrpc_body),
1088                             sizeof(struct obd_statfs) };
1089         ENTRY;
1090
1091         /* This will trigger a watchdog timeout */
1092         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1093                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1094         OBD_COUNTER_INCREMENT(obd, statfs);
1095
1096         rc = lustre_pack_reply(req, 2, size, NULL);
1097         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1098                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1099                 GOTO(out, rc);
1100         }
1101
1102         /* We call this so that we can cache a bit - 1 jiffie worth */
1103         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1104                                                 size[REPLY_REC_OFF]),
1105                             cfs_time_current_64() - HZ);
1106         if (rc) {
1107                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1108                 GOTO(out, rc);
1109         }
1110
1111         EXIT;
1112 out:
1113         req->rq_status = rc;
1114         return 0;
1115 }
1116
1117 static int mds_sync(struct ptlrpc_request *req, int offset)
1118 {
1119         struct obd_device *obd = req->rq_export->exp_obd;
1120         struct mds_obd *mds = &obd->u.mds;
1121         struct mds_body *body;
1122         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
1123         ENTRY;
1124
1125         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1126                                   lustre_swab_mds_body);
1127         if (body == NULL)
1128                 GOTO(out, rc = -EFAULT);
1129
1130         rc = lustre_pack_reply(req, 2, size, NULL);
1131         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1132                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1133                 GOTO(out, rc);
1134         }
1135
1136         if (body->fid1.id == 0) {
1137                 /* a fid of zero is taken to mean "sync whole filesystem" */
1138                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
1139                 GOTO(out, rc);
1140         } else {
1141                 struct dentry *de;
1142
1143                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1144                 if (IS_ERR(de))
1145                         GOTO(out, rc = PTR_ERR(de));
1146
1147                 /* The file parameter isn't used for anything */
1148                 if (de->d_inode->i_fop && de->d_inode->i_fop->fsync)
1149                         rc = de->d_inode->i_fop->fsync(NULL, de, 1);
1150                 if (rc == 0) {
1151                         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1152                                               sizeof(*body));
1153                         mds_pack_inode2fid(&body->fid1, de->d_inode);
1154                         mds_pack_inode2body(body, de->d_inode);
1155                 }
1156
1157                 l_dput(de);
1158                 GOTO(out, rc);
1159         }
1160 out:
1161         req->rq_status = rc;
1162         return 0;
1163 }
1164
1165 /* mds_readpage does not take a DLM lock on the inode, because the client must
1166  * already have a PR lock.
1167  *
1168  * If we were to take another one here, a deadlock will result, if another
1169  * thread is already waiting for a PW lock. */
1170 static int mds_readpage(struct ptlrpc_request *req, int offset)
1171 {
1172         struct obd_device *obd = req->rq_export->exp_obd;
1173         struct mds_obd *mds = &obd->u.mds;
1174         struct vfsmount *mnt;
1175         struct dentry *de;
1176         struct file *file;
1177         struct mds_body *body, *repbody;
1178         struct lvfs_run_ctxt saved;
1179         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
1180         struct lvfs_ucred uc = {0,};
1181         ENTRY;
1182
1183         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1184                 RETURN(-ENOMEM);
1185
1186         rc = lustre_pack_reply(req, 2, size, NULL);
1187         if (rc) {
1188                 CERROR("error packing readpage reply: rc %d\n", rc);
1189                 GOTO(out, rc);
1190         }
1191
1192         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1193                                   lustre_swab_mds_body);
1194         if (body == NULL)
1195                 GOTO (out, rc = -EFAULT);
1196
1197         rc = mds_init_ucred(&uc, req, offset);
1198         if (rc)
1199                 GOTO(out, rc);
1200
1201         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1202         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1203         if (IS_ERR(de))
1204                 GOTO(out_pop, rc = PTR_ERR(de));
1205
1206         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1207
1208         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1209         /* note: in case of an error, dentry_open puts dentry */
1210         if (IS_ERR(file))
1211                 GOTO(out_pop, rc = PTR_ERR(file));
1212
1213         /* body->size is actually the offset -eeb */
1214         if ((body->size & (de->d_inode->i_sb->s_blocksize - 1)) != 0) {
1215                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1216                        body->size, de->d_inode->i_sb->s_blocksize);
1217                 GOTO(out_file, rc = -EFAULT);
1218         }
1219
1220         /* body->nlink is actually the #bytes to read -eeb */
1221         if (body->nlink & (de->d_inode->i_sb->s_blocksize - 1)) {
1222                 CERROR("size %u is not multiple of blocksize %lu\n",
1223                        body->nlink, de->d_inode->i_sb->s_blocksize);
1224                 GOTO(out_file, rc = -EFAULT);
1225         }
1226
1227         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1228                                  sizeof(*repbody));
1229         repbody->size = i_size_read(file->f_dentry->d_inode);
1230         repbody->valid = OBD_MD_FLSIZE;
1231
1232         /* to make this asynchronous make sure that the handling function
1233            doesn't send a reply when this function completes. Instead a
1234            callback function would send the reply */
1235         /* body->size is actually the offset -eeb */
1236         rc = mds_sendpage(req, file, body->size, body->nlink);
1237
1238 out_file:
1239         filp_close(file, 0);
1240 out_pop:
1241         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1242 out:
1243         mds_exit_ucred(&uc, mds);
1244         req->rq_status = rc;
1245         RETURN(0);
1246 }
1247
1248 int mds_reint(struct ptlrpc_request *req, int offset,
1249               struct lustre_handle *lockh)
1250 {
1251         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1252         int rc;
1253
1254         OBD_ALLOC(rec, sizeof(*rec));
1255         if (rec == NULL)
1256                 RETURN(-ENOMEM);
1257
1258         rc = mds_update_unpack(req, offset, rec);
1259         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1260                 CERROR("invalid record\n");
1261                 GOTO(out, req->rq_status = -EINVAL);
1262         }
1263
1264         /* rc will be used to interrupt a for loop over multiple records */
1265         rc = mds_reint_rec(rec, offset, req, lockh);
1266  out:
1267         OBD_FREE(rec, sizeof(*rec));
1268         return rc;
1269 }
1270
1271 int mds_filter_recovery_request(struct ptlrpc_request *req,
1272                                 struct obd_device *obd, int *process)
1273 {
1274         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1275         case MDS_CONNECT: /* This will never get here, but for completeness. */
1276         case OST_CONNECT: /* This will never get here, but for completeness. */
1277         case MDS_DISCONNECT:
1278         case OST_DISCONNECT:
1279                *process = 1;
1280                RETURN(0);
1281
1282         case MDS_CLOSE:
1283         case MDS_DONE_WRITING:
1284         case MDS_SYNC: /* used in unmounting */
1285         case OBD_PING:
1286         case MDS_REINT:
1287         case SEQ_QUERY:
1288         case FLD_QUERY:
1289         case LDLM_ENQUEUE:
1290                 *process = target_queue_recovery_request(req, obd);
1291                 RETURN(0);
1292
1293         default:
1294                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1295                 *process = -EAGAIN;
1296                 RETURN(0);
1297         }
1298 }
1299 EXPORT_SYMBOL(mds_filter_recovery_request);
1300
1301 static char *reint_names[] = {
1302         [REINT_SETATTR] "setattr",
1303         [REINT_CREATE]  "create",
1304         [REINT_LINK]    "link",
1305         [REINT_UNLINK]  "unlink",
1306         [REINT_RENAME]  "rename",
1307         [REINT_OPEN]    "open",
1308 };
1309
1310 static int mds_set_info_rpc(struct obd_export *exp, struct ptlrpc_request *req)
1311 {
1312         char *key;
1313         __u32 *val;
1314         int keylen, rc = 0;
1315         ENTRY;
1316
1317         key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1318         if (key == NULL) {
1319                 DEBUG_REQ(D_HA, req, "no set_info key");
1320                 RETURN(-EFAULT);
1321         }
1322         keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1323
1324         val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, sizeof(*val));
1325         if (val == NULL) {
1326                 DEBUG_REQ(D_HA, req, "no set_info val");
1327                 RETURN(-EFAULT);
1328         }
1329
1330         rc = lustre_pack_reply(req, 1, NULL, NULL);
1331         if (rc)
1332                 RETURN(rc);
1333         lustre_msg_set_status(req->rq_repmsg, 0);
1334
1335         if (keylen < strlen("read-only") ||
1336             memcmp(key, "read-only", keylen) != 0)
1337                 RETURN(-EINVAL);
1338
1339         if (*val)
1340                 exp->exp_connect_flags |= OBD_CONNECT_RDONLY;
1341         else
1342                 exp->exp_connect_flags &= ~OBD_CONNECT_RDONLY;
1343
1344         RETURN(0);
1345 }
1346
1347 static int mds_handle_quotacheck(struct ptlrpc_request *req)
1348 {
1349         struct obd_quotactl *oqctl;
1350         int rc;
1351         ENTRY;
1352
1353         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1354                                    lustre_swab_obd_quotactl);
1355         if (oqctl == NULL)
1356                 RETURN(-EPROTO);
1357
1358         rc = lustre_pack_reply(req, 1, NULL, NULL);
1359         if (rc) {
1360                 CERROR("mds: out of memory while packing quotacheck reply\n");
1361                 RETURN(rc);
1362         }
1363
1364         req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1365         RETURN(0);
1366 }
1367
1368 static int mds_handle_quotactl(struct ptlrpc_request *req)
1369 {
1370         struct obd_quotactl *oqctl, *repoqc;
1371         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1372         ENTRY;
1373
1374         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1375                                    lustre_swab_obd_quotactl);
1376         if (oqctl == NULL)
1377                 RETURN(-EPROTO);
1378
1379         rc = lustre_pack_reply(req, 2, size, NULL);
1380         if (rc)
1381                 RETURN(rc);
1382
1383         repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1384
1385         req->rq_status = obd_quotactl(req->rq_export, oqctl);
1386         *repoqc = *oqctl;
1387         RETURN(0);
1388 }
1389
1390 int mds_msg_check_version(struct lustre_msg *msg)
1391 {
1392         int rc;
1393
1394         switch (lustre_msg_get_opc(msg)) {
1395         case MDS_CONNECT:
1396         case MDS_DISCONNECT:
1397         case OBD_PING:
1398         case SEC_CTX_INIT:
1399         case SEC_CTX_INIT_CONT:
1400         case SEC_CTX_FINI:
1401                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1402                 if (rc)
1403                         CERROR("bad opc %u version %08x, expecting %08x\n",
1404                                lustre_msg_get_opc(msg),
1405                                lustre_msg_get_version(msg),
1406                                LUSTRE_OBD_VERSION);
1407                 break;
1408         case MDS_GETSTATUS:
1409         case MDS_GETATTR:
1410         case MDS_GETATTR_NAME:
1411         case MDS_STATFS:
1412         case MDS_READPAGE:
1413         case MDS_WRITEPAGE:
1414         case MDS_IS_SUBDIR:
1415         case MDS_REINT:
1416         case MDS_CLOSE:
1417         case MDS_DONE_WRITING:
1418         case MDS_PIN:
1419         case MDS_SYNC:
1420         case MDS_GETXATTR:
1421         case MDS_SETXATTR:
1422         case MDS_SET_INFO:
1423         case MDS_QUOTACHECK:
1424         case MDS_QUOTACTL:
1425         case QUOTA_DQACQ:
1426         case QUOTA_DQREL:
1427         case SEQ_QUERY:
1428         case FLD_QUERY:
1429                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1430                 if (rc)
1431                         CERROR("bad opc %u version %08x, expecting %08x\n",
1432                                lustre_msg_get_opc(msg),
1433                                lustre_msg_get_version(msg),
1434                                LUSTRE_MDS_VERSION);
1435                 break;
1436         case LDLM_ENQUEUE:
1437         case LDLM_CONVERT:
1438         case LDLM_BL_CALLBACK:
1439         case LDLM_CP_CALLBACK:
1440                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1441                 if (rc)
1442                         CERROR("bad opc %u version %08x, expecting %08x\n",
1443                                lustre_msg_get_opc(msg),
1444                                lustre_msg_get_version(msg),
1445                                LUSTRE_DLM_VERSION);
1446                 break;
1447         case OBD_LOG_CANCEL:
1448         case LLOG_ORIGIN_HANDLE_CREATE:
1449         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1450         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1451         case LLOG_ORIGIN_HANDLE_CLOSE:
1452         case LLOG_ORIGIN_HANDLE_DESTROY:
1453         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1454         case LLOG_CATINFO:
1455                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1456                 if (rc)
1457                         CERROR("bad opc %u version %08x, expecting %08x\n",
1458                                lustre_msg_get_opc(msg),
1459                                lustre_msg_get_version(msg),
1460                                LUSTRE_LOG_VERSION);
1461                 break;
1462         default:
1463                 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
1464                 rc = -ENOTSUPP;
1465         }
1466         return rc;
1467 }
1468 EXPORT_SYMBOL(mds_msg_check_version);
1469
1470 int mds_handle(struct ptlrpc_request *req)
1471 {
1472         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1473         int rc;
1474         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1475         struct obd_device *obd = NULL;
1476         ENTRY;
1477
1478         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1479
1480         LASSERT(current->journal_info == NULL);
1481
1482         rc = mds_msg_check_version(req->rq_reqmsg);
1483         if (rc) {
1484                 CERROR("MDS drop mal-formed request\n");
1485                 RETURN(rc);
1486         }
1487
1488         /* XXX identical to OST */
1489         if (lustre_msg_get_opc(req->rq_reqmsg) != MDS_CONNECT) {
1490                 struct mds_export_data *med;
1491                 int recovering;
1492
1493                 if (req->rq_export == NULL) {
1494                         CERROR("operation %d on unconnected MDS from %s\n",
1495                                lustre_msg_get_opc(req->rq_reqmsg),
1496                                libcfs_id2str(req->rq_peer));
1497                         req->rq_status = -ENOTCONN;
1498                         GOTO(out, rc = -ENOTCONN);
1499                 }
1500
1501                 med = &req->rq_export->exp_mds_data;
1502                 obd = req->rq_export->exp_obd;
1503                 mds = mds_req2mds(req);
1504
1505                 /* sanity check: if the xid matches, the request must
1506                  * be marked as a resent or replayed */
1507                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
1508                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid))
1509                         if (!(lustre_msg_get_flags(req->rq_reqmsg) &
1510                                  (MSG_RESENT | MSG_REPLAY))) {
1511                                 CERROR("rq_xid "LPU64" matches last_xid, "
1512                                        "expected RESENT flag\n",
1513                                         req->rq_xid);
1514                                 req->rq_status = -ENOTCONN;
1515                                 GOTO(out, rc = -EFAULT);
1516                         }
1517                 /* else: note the opposite is not always true; a
1518                  * RESENT req after a failover will usually not match
1519                  * the last_xid, since it was likely never
1520                  * committed. A REPLAYed request will almost never
1521                  * match the last xid, however it could for a
1522                  * committed, but still retained, open. */
1523
1524                 /* Check for aborted recovery. */
1525                 spin_lock_bh(&obd->obd_processing_task_lock);
1526                 recovering = obd->obd_recovering;
1527                 spin_unlock_bh(&obd->obd_processing_task_lock);
1528                 if (recovering) {
1529                         rc = mds_filter_recovery_request(req, obd,
1530                                                          &should_process);
1531                         if (rc || !should_process)
1532                                 RETURN(rc);
1533                         else if (should_process < 0) {
1534                                 req->rq_status = should_process;
1535                                 rc = ptlrpc_error(req);
1536                                 RETURN(rc);
1537                         }
1538                 }
1539         }
1540
1541         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1542         case MDS_CONNECT:
1543                 DEBUG_REQ(D_INODE, req, "connect");
1544                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1545                 rc = target_handle_connect(req);
1546                 if (!rc) {
1547                         /* Now that we have an export, set mds. */
1548                         /*
1549                          * XXX nikita: these assignments are useless: mds is
1550                          * never used below, and obd is only used for
1551                          * MSG_LAST_REPLAY case, which never happens for
1552                          * MDS_CONNECT.
1553                          */
1554                         obd = req->rq_export->exp_obd;
1555                         mds = mds_req2mds(req);
1556                 }
1557                 break;
1558
1559         case MDS_DISCONNECT:
1560                 DEBUG_REQ(D_INODE, req, "disconnect");
1561                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1562                 rc = target_handle_disconnect(req);
1563                 req->rq_status = rc;            /* superfluous? */
1564                 break;
1565
1566         case MDS_GETSTATUS:
1567                 DEBUG_REQ(D_INODE, req, "getstatus");
1568                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1569                 rc = mds_getstatus(req);
1570                 break;
1571
1572         case MDS_GETATTR:
1573                 DEBUG_REQ(D_INODE, req, "getattr");
1574                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1575                 rc = mds_getattr(req, REQ_REC_OFF);
1576                 break;
1577
1578         case MDS_SETXATTR:
1579                 DEBUG_REQ(D_INODE, req, "setxattr");
1580                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SETXATTR_NET, 0);
1581                 rc = mds_setxattr(req);
1582                 break;
1583
1584         case MDS_GETXATTR:
1585                 DEBUG_REQ(D_INODE, req, "getxattr");
1586                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETXATTR_NET, 0);
1587                 rc = mds_getxattr(req);
1588                 break;
1589
1590         case MDS_GETATTR_NAME: {
1591                 struct lustre_handle lockh = { 0 };
1592                 DEBUG_REQ(D_INODE, req, "getattr_name");
1593                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1594
1595                 /* If this request gets a reconstructed reply, we won't be
1596                  * acquiring any new locks in mds_getattr_lock, so we don't
1597                  * want to cancel.
1598                  */
1599                 rc = mds_getattr_lock(req, REQ_REC_OFF, MDS_INODELOCK_UPDATE,
1600                                       &lockh);
1601                 /* this non-intent call (from an ioctl) is special */
1602                 req->rq_status = rc;
1603                 if (rc == 0 && lustre_handle_is_used(&lockh))
1604                         ldlm_lock_decref(&lockh, LCK_CR);
1605                 break;
1606         }
1607         case MDS_STATFS:
1608                 DEBUG_REQ(D_INODE, req, "statfs");
1609                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1610                 rc = mds_statfs(req);
1611                 break;
1612
1613         case MDS_READPAGE:
1614                 DEBUG_REQ(D_INODE, req, "readpage");
1615                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1616                 rc = mds_readpage(req, REQ_REC_OFF);
1617
1618                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1619                         RETURN(0);
1620                 }
1621
1622                 break;
1623
1624         case MDS_REINT: {
1625                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
1626                                              sizeof(*opcp));
1627                 __u32  opc;
1628                 int size[4] = { sizeof(struct ptlrpc_body),
1629                                 sizeof(struct mds_body),
1630                                 mds->mds_max_mdsize,
1631                                 mds->mds_max_cookiesize };
1632                 int bufcount;
1633
1634                 /* NB only peek inside req now; mds_reint() will swab it */
1635                 if (opcp == NULL) {
1636                         CERROR ("Can't inspect opcode\n");
1637                         rc = -EINVAL;
1638                         break;
1639                 }
1640                 opc = *opcp;
1641                 if (lustre_msg_swabbed(req->rq_reqmsg))
1642                         __swab32s(&opc);
1643
1644                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1645                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1646                            reint_names[opc] == NULL) ? reint_names[opc] :
1647                                                        "unknown opcode");
1648
1649                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1650
1651                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1652                         bufcount = 4;
1653                 else if (opc == REINT_OPEN)
1654                         bufcount = 3;
1655                 else
1656                         bufcount = 2;
1657
1658                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1659                 if (rc)
1660                         break;
1661
1662                 rc = mds_reint(req, REQ_REC_OFF, NULL);
1663                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1664                 break;
1665         }
1666
1667         case MDS_CLOSE:
1668                 DEBUG_REQ(D_INODE, req, "close");
1669                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1670                 rc = mds_close(req, REQ_REC_OFF);
1671                 break;
1672
1673         case MDS_DONE_WRITING:
1674                 DEBUG_REQ(D_INODE, req, "done_writing");
1675                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1676                 rc = mds_done_writing(req, REQ_REC_OFF);
1677                 break;
1678
1679         case MDS_PIN:
1680                 DEBUG_REQ(D_INODE, req, "pin");
1681                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1682                 rc = mds_pin(req, REQ_REC_OFF);
1683                 break;
1684
1685         case MDS_SYNC:
1686                 DEBUG_REQ(D_INODE, req, "sync");
1687                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1688                 rc = mds_sync(req, REQ_REC_OFF);
1689                 break;
1690
1691         case MDS_SET_INFO:
1692                 DEBUG_REQ(D_INODE, req, "set_info");
1693                 rc = mds_set_info_rpc(req->rq_export, req);
1694                 break;
1695
1696         case MDS_QUOTACHECK:
1697                 DEBUG_REQ(D_INODE, req, "quotacheck");
1698                 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACHECK_NET, 0);
1699                 rc = mds_handle_quotacheck(req);
1700                 break;
1701
1702         case MDS_QUOTACTL:
1703                 DEBUG_REQ(D_INODE, req, "quotactl");
1704                 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACTL_NET, 0);
1705                 rc = mds_handle_quotactl(req);
1706                 break;
1707
1708         case OBD_PING:
1709                 DEBUG_REQ(D_INODE, req, "ping");
1710                 rc = target_handle_ping(req);
1711                 break;
1712
1713         case OBD_LOG_CANCEL:
1714                 CDEBUG(D_INODE, "log cancel\n");
1715                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1716                 rc = -ENOTSUPP; /* la la la */
1717                 break;
1718
1719         case LDLM_ENQUEUE:
1720                 DEBUG_REQ(D_INODE, req, "enqueue");
1721                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1722                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1723                                          ldlm_server_blocking_ast, NULL);
1724                 fail = OBD_FAIL_LDLM_REPLY;
1725                 break;
1726         case LDLM_CONVERT:
1727                 DEBUG_REQ(D_INODE, req, "convert");
1728                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1729                 rc = ldlm_handle_convert(req);
1730                 break;
1731         case LDLM_BL_CALLBACK:
1732         case LDLM_CP_CALLBACK:
1733                 DEBUG_REQ(D_INODE, req, "callback");
1734                 CERROR("callbacks should not happen on MDS\n");
1735                 LBUG();
1736                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1737                 break;
1738         case LLOG_ORIGIN_HANDLE_CREATE:
1739                 DEBUG_REQ(D_INODE, req, "llog_init");
1740                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1741                 rc = llog_origin_handle_create(req);
1742                 break;
1743         case LLOG_ORIGIN_HANDLE_DESTROY:
1744                 DEBUG_REQ(D_INODE, req, "llog_init");
1745                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1746                 rc = llog_origin_handle_destroy(req);
1747                 break;
1748         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1749                 DEBUG_REQ(D_INODE, req, "llog next block");
1750                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1751                 rc = llog_origin_handle_next_block(req);
1752                 break;
1753         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1754                 DEBUG_REQ(D_INODE, req, "llog prev block");
1755                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1756                 rc = llog_origin_handle_prev_block(req);
1757                 break;
1758         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1759                 DEBUG_REQ(D_INODE, req, "llog read header");
1760                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1761                 rc = llog_origin_handle_read_header(req);
1762                 break;
1763         case LLOG_ORIGIN_HANDLE_CLOSE:
1764                 DEBUG_REQ(D_INODE, req, "llog close");
1765                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1766                 rc = llog_origin_handle_close(req);
1767                 break;
1768         case LLOG_CATINFO:
1769                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1770                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1771                 rc = llog_catinfo(req);
1772                 break;
1773         default:
1774                 req->rq_status = -ENOTSUPP;
1775                 rc = ptlrpc_error(req);
1776                 RETURN(rc);
1777         }
1778
1779         LASSERT(current->journal_info == NULL);
1780
1781         /* If we're DISCONNECTing, the mds_export_data is already freed */
1782         if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != MDS_DISCONNECT) {
1783                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1784
1785                 /* I don't think last_xid is used for anyway, so I'm not sure
1786                    if we need to care about last_close_xid here.*/
1787                 lustre_msg_set_last_xid(req->rq_repmsg,
1788                                        le64_to_cpu(med->med_mcd->mcd_last_xid));
1789
1790                 target_committed_to_req(req);
1791         }
1792
1793         EXIT;
1794  out:
1795
1796         target_send_reply(req, rc, fail);
1797         return 0;
1798 }
1799
1800 /* Update the server data on disk.  This stores the new mount_count and
1801  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1802  * then the server last_rcvd value may be less than that of the clients.
1803  * This will alert us that we may need to do client recovery.
1804  *
1805  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1806  */
1807 int mds_update_server_data(struct obd_device *obd, int force_sync)
1808 {
1809         struct mds_obd *mds = &obd->u.mds;
1810         struct lr_server_data *lsd = mds->mds_server_data;
1811         struct file *filp = mds->mds_rcvd_filp;
1812         struct lvfs_run_ctxt saved;
1813         loff_t off = 0;
1814         int rc;
1815         ENTRY;
1816
1817         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1818                mds->mds_mount_count, mds->mds_last_transno);
1819
1820         spin_lock(&mds->mds_transno_lock);
1821         lsd->lsd_last_transno = cpu_to_le64(mds->mds_last_transno);
1822         spin_unlock(&mds->mds_transno_lock);
1823
1824         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1825         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off,force_sync);
1826         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1827         if (rc)
1828                 CERROR("error writing MDS server data: rc = %d\n", rc);
1829         RETURN(rc);
1830 }
1831
1832 static void fsoptions_to_mds_flags(struct mds_obd *mds, char *options)
1833 {
1834         char *p = options;
1835
1836         if (!options)
1837                 return;
1838
1839         while (*options) {
1840                 int len;
1841
1842                 while (*p && *p != ',')
1843                         p++;
1844
1845                 len = p - options;
1846                 if (len == sizeof("user_xattr") - 1 &&
1847                     memcmp(options, "user_xattr", len) == 0) {
1848                         mds->mds_fl_user_xattr = 1;
1849                         LCONSOLE_INFO("Enabling user_xattr\n");
1850                 } else if (len == sizeof("nouser_xattr") - 1 &&
1851                            memcmp(options, "nouser_xattr", len) == 0) {
1852                         mds->mds_fl_user_xattr = 0;
1853                         LCONSOLE_INFO("Disabling user_xattr\n");
1854                 } else if (len == sizeof("acl") - 1 &&
1855                            memcmp(options, "acl", len) == 0) {
1856 #ifdef CONFIG_FS_POSIX_ACL
1857                         mds->mds_fl_acl = 1;
1858                         LCONSOLE_INFO("Enabling ACL\n");
1859 #else
1860                         CWARN("ignoring unsupported acl mount option\n");
1861 #endif
1862                 } else if (len == sizeof("noacl") - 1 &&
1863                            memcmp(options, "noacl", len) == 0) {
1864 #ifdef CONFIG_FS_POSIX_ACL
1865                         mds->mds_fl_acl = 0;
1866                         LCONSOLE_INFO("Disabling ACL\n");
1867 #endif
1868                 }
1869
1870                 options = ++p;
1871         }
1872 }
1873 static int mds_lov_presetup (struct mds_obd *mds, struct lustre_cfg *lcfg)
1874 {
1875         int rc;
1876         ENTRY;
1877
1878         rc = llog_start_commit_thread();
1879         if (rc < 0)
1880                 RETURN(rc);
1881
1882         if (lcfg->lcfg_bufcount >= 4 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1883                 class_uuid_t uuid;
1884
1885                 ll_generate_random_uuid(uuid);
1886                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1887
1888                 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
1889                 if (mds->mds_profile == NULL)
1890                         RETURN(-ENOMEM);
1891
1892                 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
1893                         LUSTRE_CFG_BUFLEN(lcfg, 3));
1894         }
1895         RETURN(rc);
1896 }
1897
1898 /* mount the file system (secretly).  lustre_cfg parameters are:
1899  * 1 = device
1900  * 2 = fstype
1901  * 3 = config name
1902  * 4 = mount options
1903  */
1904 static int mds_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1905 {
1906         struct lprocfs_static_vars lvars;
1907         struct mds_obd *mds = &obd->u.mds;
1908         struct lustre_mount_info *lmi;
1909         struct vfsmount *mnt;
1910         struct lustre_sb_info *lsi;
1911         struct obd_uuid uuid;
1912         __u8 *uuid_ptr;
1913         char *str, *label;
1914         char ns_name[48];
1915         int rc = 0;
1916         ENTRY;
1917
1918         /* setup 1:/dev/loop/0 2:ext3 3:mdsA 4:errors=remount-ro,iopen_nopriv */
1919
1920         CLASSERT(offsetof(struct obd_device, u.obt) ==
1921                  offsetof(struct obd_device, u.mds.mds_obt));
1922
1923         if (lcfg->lcfg_bufcount < 3)
1924                 RETURN(-EINVAL);
1925
1926         if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
1927                 RETURN(-EINVAL);
1928
1929         lmi = server_get_mount(obd->obd_name);
1930         if (!lmi) {
1931                 CERROR("Not mounted in lustre_fill_super?\n");
1932                 RETURN(-EINVAL);
1933         }
1934
1935         /* We mounted in lustre_fill_super.
1936            lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1937
1938         lsi = s2lsi(lmi->lmi_sb);
1939         fsoptions_to_mds_flags(mds, lsi->lsi_ldd->ldd_mount_opts);
1940         fsoptions_to_mds_flags(mds, lsi->lsi_lmd->lmd_opts);
1941         mnt = lmi->lmi_mnt;
1942         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1943         if (IS_ERR(obd->obd_fsops))
1944                 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
1945
1946         CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
1947
1948         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1949
1950         sema_init(&mds->mds_epoch_sem, 1);
1951         spin_lock_init(&mds->mds_transno_lock);
1952         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1953         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1954         mds->mds_atime_diff = MAX_ATIME_DIFF;
1955         mds->mds_evict_ost_nids = 1;
1956
1957         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1958         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1959                                                 LDLM_NAMESPACE_GREEDY);
1960         if (obd->obd_namespace == NULL) {
1961                 mds_cleanup(obd);
1962                 GOTO(err_ops, rc = -ENOMEM);
1963         }
1964         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1965
1966         lprocfs_init_vars(mds, &lvars);
1967         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1968             lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
1969                 /* Init private stats here */
1970                 mds_stats_counter_init(obd->obd_stats);
1971                 obd->obd_proc_exports = proc_mkdir("exports",
1972                                                    obd->obd_proc_entry);
1973         }
1974
1975         rc = mds_fs_setup(obd, mnt);
1976         if (rc) {
1977                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
1978                        obd->obd_name, rc);
1979                 GOTO(err_ns, rc);
1980         }
1981
1982         rc = mds_lov_presetup(mds, lcfg);
1983         if (rc < 0)
1984                 GOTO(err_fs, rc);
1985
1986         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1987                            "mds_ldlm_client", &obd->obd_ldlm_client);
1988         obd->obd_replayable = 1;
1989
1990         rc = lquota_setup(mds_quota_interface_ref, obd);
1991         if (rc)
1992                 GOTO(err_fs, rc);
1993
1994 #if 0
1995         mds->mds_group_hash = upcall_cache_init(obd->obd_name);
1996         if (IS_ERR(mds->mds_group_hash)) {
1997                 rc = PTR_ERR(mds->mds_group_hash);
1998                 mds->mds_group_hash = NULL;
1999                 GOTO(err_qctxt, rc);
2000         }
2001 #endif
2002
2003         /* Don't wait for mds_postrecov trying to clear orphans */
2004         obd->obd_async_recov = 1;
2005         rc = mds_postsetup(obd);
2006         /* Bug 11557 - allow async abort_recov start
2007            FIXME can remove most of this obd_async_recov plumbing
2008         obd->obd_async_recov = 0;
2009         */
2010         if (rc)
2011                 GOTO(err_qctxt, rc);
2012
2013         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2014         if (uuid_ptr != NULL) {
2015                 class_uuid_unparse(uuid_ptr, &uuid);
2016                 str = uuid.uuid;
2017         } else {
2018                 str = "no UUID";
2019         }
2020
2021         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2022         if (obd->obd_recovering) {
2023                 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2024                               "recovery until %d %s reconnect, or if no clients"
2025                               " reconnect for %d:%.02d; during that time new "
2026                               "clients will not be allowed to connect. "
2027                               "Recovery progress can be monitored by watching "
2028                               "/proc/fs/lustre/mds/%s/recovery_status.\n",
2029                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2030                               label ?: "", label ? "/" : "", str,
2031                               obd->obd_max_recoverable_clients,
2032                               (obd->obd_max_recoverable_clients == 1) ?
2033                               "client" : "clients",
2034                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
2035                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
2036                               obd->obd_name);
2037         } else {
2038                 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2039                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2040                               label ?: "", label ? "/" : "", str,
2041                               obd->obd_replayable ? "enabled" : "disabled");
2042         }
2043
2044         if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
2045                 ldlm_timeout = 6;
2046
2047         RETURN(0);
2048
2049 err_qctxt:
2050         lquota_cleanup(mds_quota_interface_ref, obd);
2051 err_fs:
2052         /* No extra cleanup needed for llog_init_commit_thread() */
2053         mds_fs_cleanup(obd);
2054 #if 0
2055         upcall_cache_cleanup(mds->mds_group_hash);
2056         mds->mds_group_hash = NULL;
2057 #endif
2058 err_ns:
2059         lprocfs_obd_cleanup(obd);
2060         lprocfs_free_obd_stats(obd);
2061         ldlm_namespace_free(obd->obd_namespace, 0);
2062         obd->obd_namespace = NULL;
2063 err_ops:
2064         fsfilt_put_ops(obd->obd_fsops);
2065 err_put:
2066         server_put_mount(obd->obd_name, mnt);
2067         obd->u.obt.obt_sb = NULL;
2068         return rc;
2069 }
2070
2071 static int mds_lov_clean(struct obd_device *obd)
2072 {
2073         struct mds_obd *mds = &obd->u.mds;
2074         struct obd_device *osc = mds->mds_osc_obd;
2075         ENTRY;
2076
2077         if (mds->mds_profile) {
2078                 class_del_profile(mds->mds_profile);
2079                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2080                 mds->mds_profile = NULL;
2081         }
2082
2083         /* There better be a lov */
2084         if (!osc)
2085                 RETURN(0);
2086         if (IS_ERR(osc))
2087                 RETURN(PTR_ERR(osc));
2088
2089         obd_register_observer(osc, NULL);
2090
2091         /* Give lov our same shutdown flags */
2092         osc->obd_force = obd->obd_force;
2093         osc->obd_fail = obd->obd_fail;
2094
2095         /* Cleanup the lov */
2096         obd_disconnect(mds->mds_osc_exp);
2097         class_manual_cleanup(osc);
2098         mds->mds_osc_exp = NULL;
2099
2100         RETURN(0);
2101 }
2102
2103 static int mds_postsetup(struct obd_device *obd)
2104 {
2105         struct mds_obd *mds = &obd->u.mds;
2106         int rc = 0;
2107         ENTRY;
2108
2109         rc = llog_setup(obd, NULL, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2110                         &llog_lvfs_ops);
2111         if (rc)
2112                 RETURN(rc);
2113
2114         rc = llog_setup(obd, NULL, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2115                         &llog_lvfs_ops);
2116         if (rc)
2117                 RETURN(rc);
2118
2119         if (mds->mds_profile) {
2120                 struct lustre_profile *lprof;
2121                 /* The profile defines which osc and mdc to connect to, for a
2122                    client.  We reuse that here to figure out the name of the
2123                    lov to use (and ignore lprof->lp_md).
2124                    The profile was set in the config log with
2125                    LCFG_MOUNTOPT profilenm oscnm mdcnm */
2126                 lprof = class_get_profile(mds->mds_profile);
2127                 if (lprof == NULL) {
2128                         CERROR("No profile found: %s\n", mds->mds_profile);
2129                         GOTO(err_cleanup, rc = -ENOENT);
2130                 }
2131                 rc = mds_lov_connect(obd, lprof->lp_dt);
2132                 if (rc)
2133                         GOTO(err_cleanup, rc);
2134         }
2135
2136         RETURN(rc);
2137
2138 err_cleanup:
2139         mds_lov_clean(obd);
2140         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2141         llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2142         RETURN(rc);
2143 }
2144
2145 int mds_postrecov(struct obd_device *obd)
2146 {
2147         int rc;
2148         ENTRY;
2149
2150         if (obd->obd_fail)
2151                 RETURN(0);
2152
2153         LASSERT(!obd->obd_recovering);
2154         LASSERT(llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT) != NULL);
2155
2156         /* FIXME why not put this in the synchronize? */
2157         /* set nextid first, so we are sure it happens */
2158         rc = mds_lov_set_nextid(obd);
2159         if (rc) {
2160                 CERROR("%s: mds_lov_set_nextid failed %d\n",
2161                        obd->obd_name, rc);
2162                 GOTO(out, rc);
2163         }
2164
2165         /* clean PENDING dir */
2166         if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2167                 rc = mds_cleanup_pending(obd);
2168                 if (rc < 0)
2169                         GOTO(out, rc);
2170
2171         /* FIXME Does target_finish_recovery really need this to block? */
2172         /* Notify the LOV, which will in turn call mds_notify for each tgt */
2173         /* This means that we have to hack obd_notify to think we're obd_set_up
2174            during mds_lov_connect. */
2175         obd_notify(obd->u.mds.mds_osc_obd, NULL,
2176                    obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
2177                    OBD_NOTIFY_SYNC, NULL);
2178
2179         /* quota recovery */
2180         lquota_recovery(mds_quota_interface_ref, obd);
2181
2182 out:
2183         RETURN(rc);
2184 }
2185
2186 /* We need to be able to stop an mds_lov_synchronize */
2187 static int mds_lov_early_clean(struct obd_device *obd)
2188 {
2189         struct mds_obd *mds = &obd->u.mds;
2190         struct obd_device *osc = mds->mds_osc_obd;
2191
2192         if (!osc || (!obd->obd_force && !obd->obd_fail))
2193                 return(0);
2194
2195         CDEBUG(D_HA, "abort inflight\n");
2196         return (obd_precleanup(osc, OBD_CLEANUP_EARLY));
2197 }
2198
2199 static int mds_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2200 {
2201         int rc = 0;
2202         ENTRY;
2203
2204         switch (stage) {
2205         case OBD_CLEANUP_EARLY:
2206                 break;
2207         case OBD_CLEANUP_EXPORTS:
2208                 /*XXX Use this for mdd mds cleanup, so comment out
2209                  *this target_cleanup_recovery for this tmp MDD MDS
2210                  *Wangdi*/
2211                 if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2212                         target_cleanup_recovery(obd);
2213                 mds_lov_early_clean(obd);
2214                 break;
2215         case OBD_CLEANUP_SELF_EXP:
2216                 mds_lov_disconnect(obd);
2217                 mds_lov_clean(obd);
2218                 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2219                 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2220                 rc = obd_llog_finish(obd, 0);
2221                 break;
2222         case OBD_CLEANUP_OBD:
2223                 break;
2224         }
2225         RETURN(rc);
2226 }
2227
2228 static int mds_cleanup(struct obd_device *obd)
2229 {
2230         struct mds_obd *mds = &obd->u.mds;
2231         lvfs_sbdev_type save_dev;
2232         ENTRY;
2233
2234         if (obd->u.obt.obt_sb == NULL)
2235                 RETURN(0);
2236         save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2237
2238         if (mds->mds_osc_exp)
2239                 /* lov export was disconnected by mds_lov_clean;
2240                    we just need to drop our ref */
2241                 class_export_put(mds->mds_osc_exp);
2242
2243         lprocfs_obd_cleanup(obd);
2244         lprocfs_free_obd_stats(obd);
2245
2246         lquota_cleanup(mds_quota_interface_ref, obd);
2247
2248         mds_update_server_data(obd, 1);
2249         if (mds->mds_lov_objids != NULL)
2250                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2251         mds_fs_cleanup(obd);
2252
2253 #if 0
2254         upcall_cache_cleanup(mds->mds_group_hash);
2255         mds->mds_group_hash = NULL;
2256 #endif
2257
2258         server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2259         obd->u.obt.obt_sb = NULL;
2260
2261         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2262
2263         spin_lock_bh(&obd->obd_processing_task_lock);
2264         if (obd->obd_recovering) {
2265                 target_cancel_recovery_timer(obd);
2266                 obd->obd_recovering = 0;
2267         }
2268         spin_unlock_bh(&obd->obd_processing_task_lock);
2269
2270         fsfilt_put_ops(obd->obd_fsops);
2271
2272         LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2273
2274         RETURN(0);
2275 }
2276
2277 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2278                                         struct ldlm_lock *new_lock,
2279                                         struct ldlm_lock **old_lock,
2280                                         struct lustre_handle *lockh)
2281 {
2282         struct obd_export *exp = req->rq_export;
2283         struct ldlm_request *dlmreq =
2284                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2285         struct lustre_handle remote_hdl = dlmreq->lock_handle[0];
2286         struct list_head *iter;
2287
2288         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2289                 return;
2290
2291         spin_lock(&exp->exp_ldlm_data.led_lock);
2292         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2293                 struct ldlm_lock *lock;
2294                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2295                 if (lock == new_lock)
2296                         continue;
2297                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2298                         lockh->cookie = lock->l_handle.h_cookie;
2299                         LDLM_DEBUG(lock, "restoring lock cookie");
2300                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2301                                   lockh->cookie);
2302                         if (old_lock)
2303                                 *old_lock = LDLM_LOCK_GET(lock);
2304                         spin_unlock(&exp->exp_ldlm_data.led_lock);
2305                         return;
2306                 }
2307         }
2308         spin_unlock(&exp->exp_ldlm_data.led_lock);
2309
2310         /* If the xid matches, then we know this is a resent request,
2311          * and allow it. (It's probably an OPEN, for which we don't
2312          * send a lock */
2313         if (req->rq_xid ==
2314             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2315                 return;
2316
2317         if (req->rq_xid ==
2318             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2319                 return;
2320
2321         /* This remote handle isn't enqueued, so we never received or
2322          * processed this request.  Clear MSG_RESENT, because it can
2323          * be handled like any normal request now. */
2324
2325         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2326
2327         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2328                   remote_hdl.cookie);
2329 }
2330
2331 int intent_disposition(struct ldlm_reply *rep, int flag)
2332 {
2333         if (!rep)
2334                 return 0;
2335         return (rep->lock_policy_res1 & flag);
2336 }
2337
2338 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2339 {
2340         if (!rep)
2341                 return;
2342         rep->lock_policy_res1 |= flag;
2343 }
2344
2345 static int mds_intent_policy(struct ldlm_namespace *ns,
2346                              struct ldlm_lock **lockp, void *req_cookie,
2347                              ldlm_mode_t mode, int flags, void *data)
2348 {
2349         struct ptlrpc_request *req = req_cookie;
2350         struct ldlm_lock *lock = *lockp;
2351         struct ldlm_intent *it;
2352         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2353         struct ldlm_reply *rep;
2354         struct lustre_handle lockh = { 0 };
2355         struct ldlm_lock *new_lock = NULL;
2356         int getattr_part = MDS_INODELOCK_UPDATE;
2357         int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2358                            [DLM_LOCKREPLY_OFF]   = sizeof(struct ldlm_reply),
2359                            [DLM_REPLY_REC_OFF]   = sizeof(struct mds_body),
2360                            [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2361         int repbufcnt = 4, rc;
2362         ENTRY;
2363
2364         LASSERT(req != NULL);
2365
2366         if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2367                 /* No intent was provided */
2368                 rc = lustre_pack_reply(req, 2, repsize, NULL);
2369                 LASSERT(rc == 0);
2370                 RETURN(0);
2371         }
2372
2373         it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2374                                 lustre_swab_ldlm_intent);
2375         if (it == NULL) {
2376                 CERROR("Intent missing\n");
2377                 RETURN(req->rq_status = -EFAULT);
2378         }
2379
2380         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2381
2382         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2383             (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2384                 /* we should never allow OBD_CONNECT_ACL if not configured */
2385                 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2386         else if (it->opc & IT_UNLINK)
2387                 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2388
2389         rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2390         if (rc)
2391                 RETURN(req->rq_status = rc);
2392
2393         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2394         intent_set_disposition(rep, DISP_IT_EXECD);
2395
2396
2397         /* execute policy */
2398         switch ((long)it->opc) {
2399         case IT_OPEN:
2400         case IT_CREAT|IT_OPEN:
2401                 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2402                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2403                                             &lockh);
2404                 /* XXX swab here to assert that an mds_open reint
2405                  * packet is following */
2406                 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2407                                                   &lockh);
2408 #if 0
2409                 /* We abort the lock if the lookup was negative and
2410                  * we did not make it to the OPEN portion */
2411                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2412                         RETURN(ELDLM_LOCK_ABORTED);
2413                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2414                     !intent_disposition(rep, DISP_OPEN_OPEN))
2415 #endif
2416
2417                 /* If there was an error of some sort or if we are not
2418                  * returning any locks */
2419                 if (rep->lock_policy_res2 ||
2420                     !intent_disposition(rep, DISP_OPEN_LOCK))
2421                         RETURN(ELDLM_LOCK_ABORTED);
2422                 break;
2423         case IT_LOOKUP:
2424                         getattr_part = MDS_INODELOCK_LOOKUP;
2425         case IT_GETATTR:
2426                         getattr_part |= MDS_INODELOCK_LOOKUP;
2427                         OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2428         case IT_READDIR:
2429                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2430                                             &new_lock, &lockh);
2431
2432                 /* INODEBITS_INTEROP: if this lock was converted from a
2433                  * plain lock (client does not support inodebits), then
2434                  * child lock must be taken with both lookup and update
2435                  * bits set for all operations.
2436                  */
2437                 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2438                         getattr_part = MDS_INODELOCK_LOOKUP |
2439                                        MDS_INODELOCK_UPDATE;
2440
2441                 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2442                                                          getattr_part, &lockh);
2443                 /* FIXME: LDLM can set req->rq_status. MDS sets
2444                    policy_res{1,2} with disposition and status.
2445                    - replay: returns 0 & req->status is old status
2446                    - otherwise: returns req->status */
2447                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2448                         rep->lock_policy_res2 = 0;
2449                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2450                     rep->lock_policy_res2)
2451                         RETURN(ELDLM_LOCK_ABORTED);
2452                 if (req->rq_status != 0) {
2453                         LBUG();
2454                         rep->lock_policy_res2 = req->rq_status;
2455                         RETURN(ELDLM_LOCK_ABORTED);
2456                 }
2457                 break;
2458         default:
2459                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2460                 RETURN(-EFAULT);
2461         }
2462
2463         /* By this point, whatever function we called above must have either
2464          * filled in 'lockh', been an intent replay, or returned an error.  We
2465          * want to allow replayed RPCs to not get a lock, since we would just
2466          * drop it below anyways because lock replay is done separately by the
2467          * client afterwards.  For regular RPCs we want to give the new lock to
2468          * the client instead of whatever lock it was about to get. */
2469         if (new_lock == NULL)
2470                 new_lock = ldlm_handle2lock(&lockh);
2471         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2472                 RETURN(0);
2473
2474         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2475                  it->opc, lockh.cookie);
2476
2477         /* If we've already given this lock to a client once, then we should
2478          * have no readers or writers.  Otherwise, we should have one reader
2479          * _or_ writer ref (which will be zeroed below) before returning the
2480          * lock to a client. */
2481         if (new_lock->l_export == req->rq_export) {
2482                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2483         } else {
2484                 LASSERT(new_lock->l_export == NULL);
2485                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2486         }
2487
2488         *lockp = new_lock;
2489
2490         if (new_lock->l_export == req->rq_export) {
2491                 /* Already gave this to the client, which means that we
2492                  * reconstructed a reply. */
2493                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2494                         MSG_RESENT);
2495                 RETURN(ELDLM_LOCK_REPLACED);
2496         }
2497
2498         /* Fixup the lock to be given to the client */
2499         lock_res_and_lock(new_lock);
2500         new_lock->l_readers = 0;
2501         new_lock->l_writers = 0;
2502
2503         new_lock->l_export = class_export_get(req->rq_export);
2504         spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2505         list_add(&new_lock->l_export_chain,
2506                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2507         spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2508
2509         new_lock->l_blocking_ast = lock->l_blocking_ast;
2510         new_lock->l_completion_ast = lock->l_completion_ast;
2511
2512         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2513                sizeof(lock->l_remote_handle));
2514
2515         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2516
2517         unlock_res_and_lock(new_lock);
2518         LDLM_LOCK_PUT(new_lock);
2519
2520         RETURN(ELDLM_LOCK_REPLACED);
2521 }
2522
2523 static int mdt_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2524 {
2525         struct mds_obd *mds = &obd->u.mds;
2526         struct lprocfs_static_vars lvars;
2527         int mds_min_threads;
2528         int mds_max_threads;
2529         int rc = 0;
2530         ENTRY;
2531
2532         lprocfs_init_vars(mdt, &lvars);
2533         lprocfs_obd_setup(obd, lvars.obd_vars);
2534
2535         sema_init(&mds->mds_health_sem, 1);
2536
2537         if (mds_num_threads) {
2538                 /* If mds_num_threads is set, it is the min and the max. */
2539                 if (mds_num_threads > MDS_THREADS_MAX)
2540                         mds_num_threads = MDS_THREADS_MAX;
2541                 if (mds_num_threads < MDS_THREADS_MIN)
2542                         mds_num_threads = MDS_THREADS_MIN;
2543                 mds_max_threads = mds_min_threads = mds_num_threads;
2544         } else {
2545                 /* Base min threads on memory and cpus */
2546                 mds_min_threads = smp_num_cpus * num_physpages >>
2547                         (27 - CFS_PAGE_SHIFT);
2548                 if (mds_min_threads < MDS_THREADS_MIN)
2549                         mds_min_threads = MDS_THREADS_MIN;
2550                 /* Largest auto threads start value */
2551                 if (mds_min_threads > 32)
2552                         mds_min_threads = 32;
2553                 mds_max_threads = min(MDS_THREADS_MAX, mds_min_threads * 4);
2554         }
2555
2556         mds->mds_service =
2557                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2558                                 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2559                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2560                                 mds_handle, LUSTRE_MDS_NAME,
2561                                 obd->obd_proc_entry, NULL,
2562                                 mds_min_threads, mds_max_threads, "ll_mdt", 0);
2563
2564         if (!mds->mds_service) {
2565                 CERROR("failed to start service\n");
2566                 GOTO(err_lprocfs, rc = -ENOMEM);
2567         }
2568
2569         rc = ptlrpc_start_threads(obd, mds->mds_service);
2570         if (rc)
2571                 GOTO(err_thread, rc);
2572
2573         mds->mds_setattr_service =
2574                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2575                                 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2576                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2577                                 mds_handle, "mds_setattr",
2578                                 obd->obd_proc_entry, NULL,
2579                                 mds_min_threads, mds_max_threads,
2580                                 "ll_mdt_attr", 0);
2581         if (!mds->mds_setattr_service) {
2582                 CERROR("failed to start getattr service\n");
2583                 GOTO(err_thread, rc = -ENOMEM);
2584         }
2585
2586         rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2587         if (rc)
2588                 GOTO(err_thread2, rc);
2589
2590         mds->mds_readpage_service =
2591                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2592                                 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2593                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2594                                 mds_handle, "mds_readpage",
2595                                 obd->obd_proc_entry, NULL,
2596                                 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2597                                 "ll_mdt_rdpg", 0);
2598         if (!mds->mds_readpage_service) {
2599                 CERROR("failed to start readpage service\n");
2600                 GOTO(err_thread2, rc = -ENOMEM);
2601         }
2602
2603         rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2604
2605         if (rc)
2606                 GOTO(err_thread3, rc);
2607
2608         ping_evictor_start();
2609
2610         RETURN(0);
2611
2612 err_thread3:
2613         ptlrpc_unregister_service(mds->mds_readpage_service);
2614         mds->mds_readpage_service = NULL;
2615 err_thread2:
2616         ptlrpc_unregister_service(mds->mds_setattr_service);
2617         mds->mds_setattr_service = NULL;
2618 err_thread:
2619         ptlrpc_unregister_service(mds->mds_service);
2620         mds->mds_service = NULL;
2621 err_lprocfs:
2622         lprocfs_obd_cleanup(obd);
2623         return rc;
2624 }
2625
2626 static int mdt_cleanup(struct obd_device *obd)
2627 {
2628         struct mds_obd *mds = &obd->u.mds;
2629         ENTRY;
2630
2631         ping_evictor_stop();
2632
2633         down(&mds->mds_health_sem);
2634         ptlrpc_unregister_service(mds->mds_readpage_service);
2635         ptlrpc_unregister_service(mds->mds_setattr_service);
2636         ptlrpc_unregister_service(mds->mds_service);
2637         mds->mds_readpage_service = NULL;
2638         mds->mds_setattr_service = NULL;
2639         mds->mds_service = NULL;
2640         up(&mds->mds_health_sem);
2641
2642         lprocfs_obd_cleanup(obd);
2643
2644         RETURN(0);
2645 }
2646
2647 static int mdt_health_check(struct obd_device *obd)
2648 {
2649         struct mds_obd *mds = &obd->u.mds;
2650         int rc = 0;
2651
2652         down(&mds->mds_health_sem);
2653         rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2654         rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2655         rc |= ptlrpc_service_health_check(mds->mds_service);
2656         up(&mds->mds_health_sem);
2657
2658         /*
2659          * health_check to return 0 on healthy
2660          * and 1 on unhealthy.
2661          */
2662         if(rc != 0)
2663                 rc = 1;
2664
2665         return rc;
2666 }
2667
2668 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2669                                           void *data)
2670 {
2671         struct obd_device *obd = data;
2672         struct ll_fid fid;
2673         fid.id = id;
2674         fid.generation = gen;
2675         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2676 }
2677
2678 static int mds_health_check(struct obd_device *obd)
2679 {
2680         struct obd_device_target *odt = &obd->u.obt;
2681 #ifdef USE_HEALTH_CHECK_WRITE
2682         struct mds_obd *mds = &obd->u.mds;
2683 #endif
2684         int rc = 0;
2685
2686         if (odt->obt_sb->s_flags & MS_RDONLY)
2687                 rc = 1;
2688
2689 #ifdef USE_HEALTH_CHECK_WRITE
2690         LASSERT(mds->mds_health_check_filp != NULL);
2691         rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2692 #endif
2693         return rc;
2694 }
2695
2696 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2697 {
2698         struct lustre_cfg *lcfg = buf;
2699         struct lprocfs_static_vars lvars;
2700         int rc;
2701
2702         lprocfs_init_vars(mds, &lvars);
2703
2704         rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2705         return(rc);
2706 }
2707
2708 struct lvfs_callback_ops mds_lvfs_ops = {
2709         l_fid2dentry:     mds_lvfs_fid2dentry,
2710 };
2711
2712 /* use obd ops to offer management infrastructure */
2713 static struct obd_ops mds_obd_ops = {
2714         .o_owner           = THIS_MODULE,
2715         .o_connect         = mds_connect,
2716         .o_reconnect       = mds_reconnect,
2717         .o_init_export     = mds_init_export,
2718         .o_destroy_export  = mds_destroy_export,
2719         .o_disconnect      = mds_disconnect,
2720         .o_setup           = mds_setup,
2721         .o_precleanup      = mds_precleanup,
2722         .o_cleanup         = mds_cleanup,
2723         .o_postrecov       = mds_postrecov,
2724         .o_statfs          = mds_obd_statfs,
2725         .o_iocontrol       = mds_iocontrol,
2726         .o_create          = mds_obd_create,
2727         .o_destroy         = mds_obd_destroy,
2728         .o_llog_init       = mds_llog_init,
2729         .o_llog_finish     = mds_llog_finish,
2730         .o_notify          = mds_notify,
2731         .o_health_check    = mds_health_check,
2732         .o_process_config  = mds_process_config,
2733 };
2734
2735 static struct obd_ops mdt_obd_ops = {
2736         .o_owner           = THIS_MODULE,
2737         .o_setup           = mdt_setup,
2738         .o_cleanup         = mdt_cleanup,
2739         .o_health_check    = mdt_health_check,
2740 };
2741
2742 quota_interface_t *mds_quota_interface_ref;
2743 extern quota_interface_t mds_quota_interface;
2744
2745 static __attribute__((unused)) int __init mds_init(void)
2746 {
2747         int rc;
2748         struct lprocfs_static_vars lvars;
2749
2750         request_module("lquota");
2751         mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2752         rc = lquota_init(mds_quota_interface_ref);
2753         if (rc) {
2754                 if (mds_quota_interface_ref)
2755                         PORTAL_SYMBOL_PUT(mds_quota_interface);
2756                 return rc;
2757         }
2758         init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2759
2760         lprocfs_init_vars(mds, &lvars);
2761         class_register_type(&mds_obd_ops, NULL,
2762                             lvars.module_vars, LUSTRE_MDS_NAME, NULL);
2763         lprocfs_init_vars(mdt, &lvars);
2764         mdt_obd_ops = mdt_obd_ops; //make compiler happy
2765 //        class_register_type(&mdt_obd_ops, NULL,
2766 //                            lvars.module_vars, LUSTRE_MDT_NAME, NULL);
2767
2768         return 0;
2769 }
2770
2771 static __attribute__((unused)) void /*__exit*/ mds_exit(void)
2772 {
2773         lquota_exit(mds_quota_interface_ref);
2774         if (mds_quota_interface_ref)
2775                 PORTAL_SYMBOL_PUT(mds_quota_interface);
2776
2777         class_unregister_type(LUSTRE_MDS_NAME);
2778 //        class_unregister_type(LUSTRE_MDT_NAME);
2779 }
2780 /*mds still need lov setup here*/
2781 static int mds_cmd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2782 {
2783         struct mds_obd *mds = &obd->u.mds;
2784         struct lvfs_run_ctxt saved;
2785         const char     *dev;
2786         struct vfsmount *mnt;
2787         struct lustre_sb_info *lsi;
2788         struct lustre_mount_info *lmi;
2789         struct dentry  *dentry;
2790         struct file *file;
2791         int rc = 0;
2792         ENTRY;
2793
2794         CDEBUG(D_INFO, "obd %s setup \n", obd->obd_name);
2795         if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2796                 RETURN(0);
2797
2798         if (lcfg->lcfg_bufcount < 5) {
2799                 CERROR("invalid arg for setup %s\n", MDD_OBD_NAME);
2800                 RETURN(-EINVAL);
2801         }
2802         dev = lustre_cfg_string(lcfg, 4);
2803         lmi = server_get_mount(dev);
2804         LASSERT(lmi != NULL);
2805
2806         lsi = s2lsi(lmi->lmi_sb);
2807         mnt = lmi->lmi_mnt;
2808         /* FIXME: MDD LOV initialize objects.
2809          * we need only lmi here but not get mount
2810          * OSD did mount already, so put mount back
2811          */
2812         atomic_dec(&lsi->lsi_mounts);
2813         mntput(mnt);
2814
2815         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2816         mds_init_ctxt(obd, mnt);
2817
2818         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2819         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
2820         if (IS_ERR(dentry)) {
2821                 rc = PTR_ERR(dentry);
2822                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
2823                 GOTO(err_putfs, rc);
2824         }
2825         mds->mds_objects_dir = dentry;
2826
2827         dentry = lookup_one_len("__iopen__", current->fs->pwd,
2828                                 strlen("__iopen__"));
2829         if (IS_ERR(dentry)) {
2830                 rc = PTR_ERR(dentry);
2831                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
2832                 GOTO(err_objects, rc);
2833         }
2834
2835         mds->mds_fid_de = dentry;
2836         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
2837                 rc = -ENOENT;
2838                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
2839                 GOTO(err_fid, rc);
2840         }
2841
2842         /* open and test the lov objd file */
2843         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
2844         if (IS_ERR(file)) {
2845                 rc = PTR_ERR(file);
2846                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
2847                 GOTO(err_fid, rc = PTR_ERR(file));
2848         }
2849         mds->mds_lov_objid_filp = file;
2850         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
2851                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
2852                        file->f_dentry->d_inode->i_mode);
2853                 GOTO(err_lov_objid, rc = -ENOENT);
2854         }
2855
2856         rc = mds_lov_presetup(mds, lcfg);
2857         if (rc < 0)
2858                 GOTO(err_objects, rc);
2859
2860         /* Don't wait for mds_postrecov trying to clear orphans */
2861         obd->obd_async_recov = 1;
2862         rc = mds_postsetup(obd);
2863         /* Bug 11557 - allow async abort_recov start
2864            FIXME can remove most of this obd_async_recov plumbing
2865         obd->obd_async_recov = 0;
2866         */
2867
2868         if (rc)
2869                 GOTO(err_objects, rc);
2870
2871         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2872         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2873
2874 err_pop:
2875         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2876         RETURN(rc);
2877 err_lov_objid:
2878         if (mds->mds_lov_objid_filp &&
2879                 filp_close((struct file *)mds->mds_lov_objid_filp, 0))
2880                 CERROR("can't close %s after error\n", LOV_OBJID);
2881 err_fid:
2882         dput(mds->mds_fid_de);
2883 err_objects:
2884         dput(mds->mds_objects_dir);
2885 err_putfs:
2886         fsfilt_put_ops(obd->obd_fsops);
2887         goto err_pop;
2888 }
2889
2890 static int mds_cmd_cleanup(struct obd_device *obd)
2891 {
2892         struct mds_obd *mds = &obd->u.mds;
2893         struct lvfs_run_ctxt saved;
2894         int rc = 0;
2895         ENTRY;
2896
2897         if (obd->obd_fail)
2898                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2899                               "will be preserved.\n", obd->obd_name);
2900
2901         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2902         if (mds->mds_lov_objid_filp) {
2903                 rc = filp_close((struct file *)mds->mds_lov_objid_filp, 0);
2904                 mds->mds_lov_objid_filp = NULL;
2905                 if (rc)
2906                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
2907         }
2908         if (mds->mds_objects_dir != NULL) {
2909                 l_dput(mds->mds_objects_dir);
2910                 mds->mds_objects_dir = NULL;
2911         }
2912
2913         if (mds->mds_lov_objids != NULL)
2914                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2915
2916         shrink_dcache_parent(mds->mds_fid_de);
2917         dput(mds->mds_fid_de);
2918         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2919         fsfilt_put_ops(obd->obd_fsops);
2920
2921         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2922         RETURN(rc);
2923 }
2924
2925 #if 0
2926 static int mds_cmd_health_check(struct obd_device *obd)
2927 {
2928         return 0;
2929 }
2930 #endif
2931 static struct obd_ops mds_cmd_obd_ops = {
2932         .o_owner           = THIS_MODULE,
2933         .o_setup           = mds_cmd_setup,
2934         .o_cleanup         = mds_cmd_cleanup,
2935         .o_precleanup      = mds_precleanup,
2936         .o_create          = mds_obd_create,
2937         .o_destroy         = mds_obd_destroy,
2938         .o_llog_init       = mds_llog_init,
2939         .o_llog_finish     = mds_llog_finish,
2940         .o_notify          = mds_notify,
2941         .o_postrecov       = mds_postrecov,
2942         //   .o_health_check    = mds_cmd_health_check,
2943 };
2944
2945 static int __init mds_cmd_init(void)
2946 {
2947         struct lprocfs_static_vars lvars;
2948
2949         lprocfs_init_vars(mds, &lvars);
2950         class_register_type(&mds_cmd_obd_ops, NULL, lvars.module_vars,
2951                             LUSTRE_MDS_NAME, NULL);
2952
2953         return 0;
2954 }
2955
2956 static void /*__exit*/ mds_cmd_exit(void)
2957 {
2958         class_unregister_type(LUSTRE_MDS_NAME);
2959 }
2960
2961 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2962 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2963 MODULE_LICENSE("GPL");
2964
2965 module_init(mds_cmd_init);
2966 module_exit(mds_cmd_exit);