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 #include <linux/ext3_fs.h>
  44 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  45 # include <linux/smp_lock.h>
  46 # include <linux/buffer_head.h>
  47 # include <linux/workqueue.h>
  48 # include <linux/mount.h>
  49 #else
  50 # include <linux/locks.h>
  51 #endif
  52
  53 #include <linux/lustre_acl.h>
  54 #include <obd_class.h>
  55 #include <lustre_dlm.h>
  56 #include <obd_lov.h>
  57 #include <lustre_fsfilt.h>
  58 #include <lprocfs_status.h>
  59 #include <lustre_commit_confd.h>
  60 #include <lustre_quota.h>
  61 #include <lustre_disk.h>
  62 #include <lustre_param.h>
  63
  64 #include "mds_internal.h"
  65
  66 int mds_num_threads;
  67 CFS_MODULE_PARM(mds_num_threads, "i", int, 0444,
  68                 "number of MDS service threads to start");
  69
  70 static int mds_intent_policy(struct ldlm_namespace *ns,
  71                              struct ldlm_lock **lockp, void *req_cookie,
  72                              ldlm_mode_t mode, int flags, void *data);
  73 static int mds_postsetup(struct obd_device *obd);
  74 static int mds_cleanup(struct obd_device *obd);
  75
  76 /* Assumes caller has already pushed into the kernel filesystem context */
  77 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  78                         loff_t offset, int count)
  79 {
  80         struct ptlrpc_bulk_desc *desc;
  81         struct l_wait_info lwi;
  82         struct page **pages;
  83         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  84         ENTRY;
  85
  86         LASSERT((offset & ~CFS_PAGE_MASK) == 0); /* I'm dubious about this */
  87
  88         npages = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
  89         OBD_ALLOC(pages, sizeof(*pages) * npages);
  90         if (!pages)
  91                 GOTO(out, rc = -ENOMEM);
  92
  93         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  94                                     MDS_BULK_PORTAL);
  95         if (desc == NULL)
  96                 GOTO(out_free, rc = -ENOMEM);
  97
  98         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  99                 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
 100
 101                 pages[i] = alloc_pages(GFP_KERNEL, 0);
 102                 if (pages[i] == NULL)
 103                         GOTO(cleanup_buf, rc = -ENOMEM);
 104
 105                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 106         }
 107
 108         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 109                 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
 110                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 111                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 112                        file->f_dentry->d_inode->i_size);
 113
 114                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 115                                      kmap(pages[i]), tmpsize, &offset);
 116                 kunmap(pages[i]);
 117
 118                 if (rc != tmpsize)
 119                         GOTO(cleanup_buf, rc = -EIO);
 120         }
 121
 122         LASSERT(desc->bd_nob == count);
 123
 124         rc = ptlrpc_start_bulk_transfer(desc);
 125         if (rc)
 126                 GOTO(cleanup_buf, rc);
 127
 128         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 129                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 130                        OBD_FAIL_MDS_SENDPAGE, rc);
 131                 GOTO(abort_bulk, rc);
 132         }
 133
 134         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 135         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 136         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 137
 138         if (rc == 0) {
 139                 if (desc->bd_success &&
 140                     desc->bd_nob_transferred == count)
 141                         GOTO(cleanup_buf, rc);
 142
 143                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 144         }
 145
 146         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s",
 147                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 148                   desc->bd_nob_transferred, count,
 149                   req->rq_export->exp_client_uuid.uuid,
 150                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 151
 152         class_fail_export(req->rq_export);
 153
 154         EXIT;
 155  abort_bulk:
 156         ptlrpc_abort_bulk (desc);
 157  cleanup_buf:
 158         for (i = 0; i < npages; i++)
 159                 if (pages[i])
 160                         __free_pages(pages[i], 0);
 161
 162         ptlrpc_free_bulk(desc);
 163  out_free:
 164         OBD_FREE(pages, sizeof(*pages) * npages);
 165  out:
 166         return rc;
 167 }
 168
 169 /* only valid locked dentries or errors should be returned */
 170 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 171                                      struct vfsmount **mnt, int lock_mode,
 172                                      struct lustre_handle *lockh,
 173                                      __u64 lockpart)
 174 {
 175         struct mds_obd *mds = &obd->u.mds;
 176         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 177         struct ldlm_res_id res_id = { .name = {0} };
 178         int flags = LDLM_FL_ATOMIC_CB, rc;
 179         ldlm_policy_data_t policy = { .l_inodebits = { lockpart} };
 180         ENTRY;
 181
 182         if (IS_ERR(de))
 183                 RETURN(de);
 184
 185         res_id.name[0] = de->d_inode->i_ino;
 186         res_id.name[1] = de->d_inode->i_generation;
 187         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id,
 188                                     LDLM_IBITS, &policy, lock_mode, &flags,
 189                                     ldlm_blocking_ast, ldlm_completion_ast,
 190                                     NULL, NULL, 0, NULL, lockh);
 191         if (rc != ELDLM_OK) {
 192                 l_dput(de);
 193                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 194         }
 195
 196         RETURN(retval);
 197 }
 198
 199 /* Look up an entry by inode number. */
 200 /* this function ONLY returns valid dget'd dentries with an initialized inode
 201    or errors */
 202 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 203                               struct vfsmount **mnt)
 204 {
 205         char fid_name[32];
 206         unsigned long ino = fid->id;
 207         __u32 generation = fid->generation;
 208         struct inode *inode;
 209         struct dentry *result;
 210
 211         if (ino == 0)
 212                 RETURN(ERR_PTR(-ESTALE));
 213
 214         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 215
 216         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 217                ino, generation, mds->mds_obt.obt_sb);
 218
 219         /* under ext3 this is neither supposed to return bad inodes
 220            nor NULL inodes. */
 221         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 222         if (IS_ERR(result))
 223                 RETURN(result);
 224
 225         inode = result->d_inode;
 226         if (!inode)
 227                 RETURN(ERR_PTR(-ENOENT));
 228
 229         if (inode->i_generation == 0 || inode->i_nlink == 0) {
 230                 LCONSOLE_WARN("Found inode with zero generation or link -- this"
 231                               " may indicate disk corruption (inode: %lu/%u, "
 232                               "link %lu, count %d)\n", inode->i_ino,
 233                               inode->i_generation,(unsigned long)inode->i_nlink,
 234                               atomic_read(&inode->i_count));
 235                 dput(result);
 236                 RETURN(ERR_PTR(-ENOENT));
 237         }
 238
 239         if (generation && inode->i_generation != generation) {
 240                 /* we didn't find the right inode.. */
 241                 CDEBUG(D_INODE, "found wrong generation: inode %lu, link: %lu, "
 242                        "count: %d, generation %u/%u\n", inode->i_ino,
 243                        (unsigned long)inode->i_nlink,
 244                        atomic_read(&inode->i_count), inode->i_generation,
 245                        generation);
 246                 dput(result);
 247                 RETURN(ERR_PTR(-ENOENT));
 248         }
 249
 250         if (mnt) {
 251                 *mnt = mds->mds_vfsmnt;
 252                 mntget(*mnt);
 253         }
 254
 255         RETURN(result);
 256 }
 257
 258 static int mds_connect_internal(struct obd_export *exp,
 259                                 struct obd_connect_data *data)
 260 {
 261         struct obd_device *obd = exp->exp_obd;
 262         if (data != NULL) {
 263                 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
 264                 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
 265
 266                 /* If no known bits (which should not happen, probably,
 267                    as everybody should support LOOKUP and UPDATE bits at least)
 268                    revert to compat mode with plain locks. */
 269                 if (!data->ocd_ibits_known &&
 270                     data->ocd_connect_flags & OBD_CONNECT_IBITS)
 271                         data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
 272
 273                 if (!obd->u.mds.mds_fl_acl)
 274                         data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
 275
 276                 if (!obd->u.mds.mds_fl_user_xattr)
 277                         data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
 278
 279                 exp->exp_connect_flags = data->ocd_connect_flags;
 280                 data->ocd_version = LUSTRE_VERSION_CODE;
 281                 exp->exp_mds_data.med_ibits_known = data->ocd_ibits_known;
 282         }
 283
 284         if (obd->u.mds.mds_fl_acl &&
 285             ((exp->exp_connect_flags & OBD_CONNECT_ACL) == 0)) {
 286                 CWARN("%s: MDS requires ACL support but client does not\n",
 287                       obd->obd_name);
 288                 return -EBADE;
 289         }
 290         return 0;
 291 }
 292
 293 static int mds_reconnect(struct obd_export *exp, struct obd_device *obd,
 294                          struct obd_uuid *cluuid,
 295                          struct obd_connect_data *data)
 296 {
 297         int rc;
 298         ENTRY;
 299
 300         if (exp == NULL || obd == NULL || cluuid == NULL)
 301                 RETURN(-EINVAL);
 302
 303         rc = mds_connect_internal(exp, data);
 304
 305         RETURN(rc);
 306 }
 307
 308 /* Establish a connection to the MDS.
 309  *
 310  * This will set up an export structure for the client to hold state data
 311  * about that client, like open files, the last operation number it did
 312  * on the server, etc.
 313  */
 314 static int mds_connect(const struct lu_env *env,
 315                        struct lustre_handle *conn, struct obd_device *obd,
 316                        struct obd_uuid *cluuid, struct obd_connect_data *data)
 317 {
 318         struct obd_export *exp;
 319         struct mds_export_data *med;
 320         struct mds_client_data *mcd = NULL;
 321         int rc;
 322         ENTRY;
 323
 324         if (!conn || !obd || !cluuid)
 325                 RETURN(-EINVAL);
 326
 327         /* XXX There is a small race between checking the list and adding a
 328          * new connection for the same UUID, but the real threat (list
 329          * corruption when multiple different clients connect) is solved.
 330          *
 331          * There is a second race between adding the export to the list,
 332          * and filling in the client data below.  Hence skipping the case
 333          * of NULL mcd above.  We should already be controlling multiple
 334          * connects at the client, and we can't hold the spinlock over
 335          * memory allocations without risk of deadlocking.
 336          */
 337         rc = class_connect(conn, obd, cluuid);
 338         if (rc)
 339                 RETURN(rc);
 340         exp = class_conn2export(conn);
 341         LASSERT(exp);
 342         med = &exp->exp_mds_data;
 343
 344         rc = mds_connect_internal(exp, data);
 345         if (rc)
 346                 GOTO(out, rc);
 347
 348         OBD_ALLOC(mcd, sizeof(*mcd));
 349         if (!mcd)
 350                 GOTO(out, rc = -ENOMEM);
 351
 352         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 353         med->med_mcd = mcd;
 354
 355         rc = mds_client_add(obd, exp, -1);
 356         GOTO(out, rc);
 357
 358 out:
 359         if (rc) {
 360                 if (mcd) {
 361                         OBD_FREE(mcd, sizeof(*mcd));
 362                         med->med_mcd = NULL;
 363                 }
 364                 class_disconnect(exp);
 365         } else {
 366                 class_export_put(exp);
 367         }
 368
 369         RETURN(rc);
 370 }
 371
 372 int mds_init_export(struct obd_export *exp)
 373 {
 374         struct mds_export_data *med = &exp->exp_mds_data;
 375
 376         INIT_LIST_HEAD(&med->med_open_head);
 377         spin_lock_init(&med->med_open_lock);
 378
 379         spin_lock(&exp->exp_lock);
 380         exp->exp_connecting = 1;
 381         spin_unlock(&exp->exp_lock);
 382
 383         RETURN(0);
 384 }
 385
 386 static int mds_destroy_export(struct obd_export *export)
 387 {
 388         struct mds_export_data *med;
 389         struct obd_device *obd = export->exp_obd;
 390         struct mds_obd *mds = &obd->u.mds;
 391         struct lvfs_run_ctxt saved;
 392         struct lov_mds_md *lmm;
 393         struct llog_cookie *logcookies;
 394         int rc = 0;
 395         ENTRY;
 396
 397         med = &export->exp_mds_data;
 398         target_destroy_export(export);
 399
 400         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 401                 RETURN(0);
 402
 403         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 404         /* Close any open files (which may also cause orphan unlinking). */
 405
 406         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 407         if (lmm == NULL) {
 408                 CWARN("%s: allocation failure during cleanup; can not force "
 409                       "close file handles on this service.\n", obd->obd_name);
 410                 GOTO(out, rc = -ENOMEM);
 411         }
 412
 413         OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
 414         if (logcookies == NULL) {
 415                 CWARN("%s: allocation failure during cleanup; can not force "
 416                       "close file handles on this service.\n", obd->obd_name);
 417                 OBD_FREE(lmm, mds->mds_max_mdsize);
 418                 GOTO(out_lmm, rc = -ENOMEM);
 419         }
 420
 421         spin_lock(&med->med_open_lock);
 422         while (!list_empty(&med->med_open_head)) {
 423                 struct list_head *tmp = med->med_open_head.next;
 424                 struct mds_file_data *mfd =
 425                         list_entry(tmp, struct mds_file_data, mfd_list);
 426                 int lmm_size = mds->mds_max_mdsize;
 427                 umode_t mode = mfd->mfd_dentry->d_inode->i_mode;
 428                 __u64 valid = 0;
 429
 430                 /* Remove mfd handle so it can't be found again.
 431                  * We are consuming the mfd_list reference here. */
 432                 mds_mfd_unlink(mfd, 0);
 433                 spin_unlock(&med->med_open_lock);
 434
 435                 /* If you change this message, be sure to update
 436                  * replay_single:test_46 */
 437                 CDEBUG(D_INODE|D_IOCTL, "%s: force closing file handle for "
 438                        "%.*s (ino %lu)\n", obd->obd_name,
 439                        mfd->mfd_dentry->d_name.len,mfd->mfd_dentry->d_name.name,
 440                        mfd->mfd_dentry->d_inode->i_ino);
 441
 442                 rc = mds_get_md(obd, mfd->mfd_dentry->d_inode, lmm, &lmm_size, 1);
 443                 if (rc < 0)
 444                         CWARN("mds_get_md failure, rc=%d\n", rc);
 445                 else
 446                         valid |= OBD_MD_FLEASIZE;
 447
 448                 /* child orphan sem protects orphan_dec_test and
 449                  * is_orphan race, mds_mfd_close drops it */
 450                 MDS_DOWN_WRITE_ORPHAN_SEM(mfd->mfd_dentry->d_inode);
 451                 rc = mds_mfd_close(NULL, REQ_REC_OFF, obd, mfd,
 452                                    !(export->exp_flags & OBD_OPT_FAILOVER),
 453                                    lmm, lmm_size, logcookies,
 454                                    mds->mds_max_cookiesize,
 455                                    &valid);
 456
 457                 if (rc)
 458                         CDEBUG(D_INODE|D_IOCTL, "Error closing file: %d\n", rc);
 459
 460                 if (valid & OBD_MD_FLCOOKIE) {
 461                         rc = mds_osc_destroy_orphan(obd, mode, lmm,
 462                                                     lmm_size, logcookies, 1);
 463                         if (rc < 0) {
 464                                 CDEBUG(D_INODE, "%s: destroy of orphan failed,"
 465                                        " rc = %d\n", obd->obd_name, rc);
 466                                 rc = 0;
 467                         }
 468                         valid &= ~OBD_MD_FLCOOKIE;
 469                 }
 470
 471                 spin_lock(&med->med_open_lock);
 472         }
 473         spin_unlock(&med->med_open_lock);
 474
 475         OBD_FREE(logcookies, mds->mds_max_cookiesize);
 476 out_lmm:
 477         OBD_FREE(lmm, mds->mds_max_mdsize);
 478 out:
 479         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 480         mds_client_free(export);
 481         RETURN(rc);
 482 }
 483
 484 static int mds_disconnect(struct obd_export *exp)
 485 {
 486         int rc;
 487         ENTRY;
 488
 489         LASSERT(exp);
 490         class_export_get(exp);
 491
 492         /* Disconnect early so that clients can't keep using export */
 493         rc = class_disconnect(exp);
 494         if (exp->exp_obd->obd_namespace != NULL)
 495                 ldlm_cancel_locks_for_export(exp);
 496
 497         /* complete all outstanding replies */
 498         spin_lock(&exp->exp_lock);
 499         while (!list_empty(&exp->exp_outstanding_replies)) {
 500                 struct ptlrpc_reply_state *rs =
 501                         list_entry(exp->exp_outstanding_replies.next,
 502                                    struct ptlrpc_reply_state, rs_exp_list);
 503                 struct ptlrpc_service *svc = rs->rs_service;
 504
 505                 spin_lock(&svc->srv_lock);
 506                 list_del_init(&rs->rs_exp_list);
 507                 ptlrpc_schedule_difficult_reply(rs);
 508                 spin_unlock(&svc->srv_lock);
 509         }
 510         spin_unlock(&exp->exp_lock);
 511
 512         class_export_put(exp);
 513         RETURN(rc);
 514 }
 515
 516 static int mds_getstatus(struct ptlrpc_request *req)
 517 {
 518         struct mds_obd *mds = mds_req2mds(req);
 519         struct mds_body *body;
 520         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 521         ENTRY;
 522
 523         rc = lustre_pack_reply(req, 2, size, NULL);
 524         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 525                 CERROR("mds: out of memory for message\n");
 526                 req->rq_status = -ENOMEM;       /* superfluous? */
 527                 RETURN(-ENOMEM);
 528         }
 529
 530         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 531         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 532
 533         /* the last_committed and last_xid fields are filled in for all
 534          * replies already - no need to do so here also.
 535          */
 536         RETURN(0);
 537 }
 538
 539 /* get the LOV EA from @inode and store it into @md.  It can be at most
 540  * @size bytes, and @size is updated with the actual EA size.
 541  * The EA size is also returned on success, and -ve errno on failure.
 542  * If there is no EA then 0 is returned. */
 543 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
 544                int *size, int lock)
 545 {
 546         int rc = 0;
 547         int lmm_size;
 548
 549         if (lock)
 550                 LOCK_INODE_MUTEX(inode);
 551         rc = fsfilt_get_md(obd, inode, md, *size, "lov");
 552
 553         if (rc < 0) {
 554                 CERROR("Error %d reading eadata for ino %lu\n",
 555                        rc, inode->i_ino);
 556         } else if (rc > 0) {
 557                 lmm_size = rc;
 558                 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
 559
 560                 if (rc == 0) {
 561                         *size = lmm_size;
 562                         rc = lmm_size;
 563                 } else if (rc > 0) {
 564                         *size = rc;
 565                 }
 566         } else {
 567                 *size = 0;
 568         }
 569         if (lock)
 570                 UNLOCK_INODE_MUTEX(inode);
 571
 572         RETURN (rc);
 573 }
 574
 575
 576 /* Call with lock=1 if you want mds_pack_md to take the i_mutex.
 577  * Call with lock=0 if the caller has already taken the i_mutex. */
 578 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 579                 struct mds_body *body, struct inode *inode, int lock)
 580 {
 581         struct mds_obd *mds = &obd->u.mds;
 582         void *lmm;
 583         int lmm_size;
 584         int rc;
 585         ENTRY;
 586
 587         lmm = lustre_msg_buf(msg, offset, 0);
 588         if (lmm == NULL) {
 589                 /* Some problem with getting eadata when I sized the reply
 590                  * buffer... */
 591                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 592                        inode->i_ino);
 593                 RETURN(0);
 594         }
 595         lmm_size = lustre_msg_buflen(msg, offset);
 596
 597         /* I don't really like this, but it is a sanity check on the client
 598          * MD request.  However, if the client doesn't know how much space
 599          * to reserve for the MD, it shouldn't be bad to have too much space.
 600          */
 601         if (lmm_size > mds->mds_max_mdsize) {
 602                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 603                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 604                 // RETURN(-EINVAL);
 605         }
 606
 607         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
 608         if (rc > 0) {
 609                 if (S_ISDIR(inode->i_mode))
 610                         body->valid |= OBD_MD_FLDIREA;
 611                 else
 612                         body->valid |= OBD_MD_FLEASIZE;
 613                 body->eadatasize = lmm_size;
 614                 rc = 0;
 615         }
 616
 617         RETURN(rc);
 618 }
 619
 620 #ifdef CONFIG_FS_POSIX_ACL
 621 static
 622 int mds_pack_posix_acl(struct inode *inode, struct lustre_msg *repmsg,
 623                        struct mds_body *repbody, int repoff)
 624 {
 625         struct dentry de = { .d_inode = inode };
 626         int buflen, rc;
 627         ENTRY;
 628
 629         LASSERT(repbody->aclsize == 0);
 630         LASSERT(lustre_msg_bufcount(repmsg) > repoff);
 631
 632         buflen = lustre_msg_buflen(repmsg, repoff);
 633         if (!buflen)
 634                 GOTO(out, 0);
 635
 636         if (!inode->i_op || !inode->i_op->getxattr)
 637                 GOTO(out, 0);
 638
 639         lock_24kernel();
 640         rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
 641                                    lustre_msg_buf(repmsg, repoff, buflen),
 642                                    buflen);
 643         unlock_24kernel();
 644
 645         if (rc >= 0)
 646                 repbody->aclsize = rc;
 647         else if (rc != -ENODATA) {
 648                 CERROR("buflen %d, get acl: %d\n", buflen, rc);
 649                 RETURN(rc);
 650         }
 651         EXIT;
 652 out:
 653         repbody->valid |= OBD_MD_FLACL;
 654         return 0;
 655 }
 656 #else
 657 #define mds_pack_posix_acl(inode, repmsg, repbody, repoff) 0
 658 #endif
 659
 660 int mds_pack_acl(struct mds_export_data *med, struct inode *inode,
 661                  struct lustre_msg *repmsg, struct mds_body *repbody,
 662                  int repoff)
 663 {
 664         return mds_pack_posix_acl(inode, repmsg, repbody, repoff);
 665 }
 666
 667 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 668                                 struct ptlrpc_request *req,
 669                                 struct mds_body *reqbody, int reply_off)
 670 {
 671         struct mds_body *body;
 672         struct inode *inode = dentry->d_inode;
 673         int rc = 0;
 674         ENTRY;
 675
 676         if (inode == NULL)
 677                 RETURN(-ENOENT);
 678
 679         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 680         LASSERT(body != NULL);                 /* caller prepped reply */
 681
 682         mds_pack_inode2fid(&body->fid1, inode);
 683         body->flags = reqbody->flags; /* copy MDS_BFLAG_EXT_FLAGS if present */
 684         mds_pack_inode2body(body, inode);
 685         reply_off++;
 686
 687         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
 688             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
 689                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off, body,
 690                                  inode, 1);
 691
 692                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 693                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 694                     !(body->valid & OBD_MD_FLDIREA))
 695                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 696                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 697
 698                 lustre_shrink_reply(req, reply_off, body->eadatasize, 0);
 699                 if (body->eadatasize)
 700                         reply_off++;
 701         } else if (S_ISLNK(inode->i_mode) &&
 702                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 703                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off, 0);
 704                 int len;
 705
 706                 LASSERT (symname != NULL);       /* caller prepped reply */
 707                 len = lustre_msg_buflen(req->rq_repmsg, reply_off);
 708
 709                 rc = inode->i_op->readlink(dentry, symname, len);
 710                 if (rc < 0) {
 711                         CERROR("readlink failed: %d\n", rc);
 712                 } else if (rc != len - 1) {
 713                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 714                                 rc, len - 1);
 715                         rc = -EINVAL;
 716                 } else {
 717                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 718                         body->valid |= OBD_MD_LINKNAME;
 719                         body->eadatasize = rc + 1;
 720                         symname[rc] = 0;        /* NULL terminate */
 721                         rc = 0;
 722                 }
 723                 reply_off++;
 724         } else if (reqbody->valid == OBD_MD_FLFLAGS &&
 725                    reqbody->flags & MDS_BFLAG_EXT_FLAGS) {
 726                 int flags;
 727
 728                 /* We only return the full set of flags on ioctl, otherwise we
 729                  * get enough flags from the inode in mds_pack_inode2body(). */
 730                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_GETFLAGS,
 731                                       (long)&flags);
 732                 if (rc == 0)
 733                         body->flags = flags | MDS_BFLAG_EXT_FLAGS;
 734         }
 735
 736         if (reqbody->valid & OBD_MD_FLMODEASIZE) {
 737                 struct mds_obd *mds = mds_req2mds(req);
 738                 body->max_cookiesize = mds->mds_max_cookiesize;
 739                 body->max_mdsize = mds->mds_max_mdsize;
 740                 body->valid |= OBD_MD_FLMODEASIZE;
 741         }
 742
 743         if (rc)
 744                 RETURN(rc);
 745
 746 #ifdef CONFIG_FS_POSIX_ACL
 747         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
 748             (reqbody->valid & OBD_MD_FLACL)) {
 749                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 750                                   inode, req->rq_repmsg,
 751                                   body, reply_off);
 752
 753                 lustre_shrink_reply(req, reply_off, body->aclsize, 0);
 754                 if (body->aclsize)
 755                         reply_off++;
 756         }
 757 #endif
 758
 759         RETURN(rc);
 760 }
 761
 762 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 763                                 int offset)
 764 {
 765         struct mds_obd *mds = mds_req2mds(req);
 766         struct mds_body *body;
 767         int rc, bufcount = 2;
 768         int size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 769         ENTRY;
 770
 771         LASSERT(offset == REQ_REC_OFF); /* non-intent */
 772
 773         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
 774         LASSERT(body != NULL);                 /* checked by caller */
 775         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 776
 777         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 778             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
 779                 LOCK_INODE_MUTEX(inode);
 780                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
 781                                    "lov");
 782                 UNLOCK_INODE_MUTEX(inode);
 783                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 784                        rc, inode->i_ino);
 785                 if (rc < 0) {
 786                         if (rc != -ENODATA) {
 787                                 CERROR("error getting inode %lu MD: rc = %d\n",
 788                                        inode->i_ino, rc);
 789                                 RETURN(rc);
 790                         }
 791                         size[bufcount] = 0;
 792                 } else if (rc > mds->mds_max_mdsize) {
 793                         size[bufcount] = 0;
 794                         CERROR("MD size %d larger than maximum possible %u\n",
 795                                rc, mds->mds_max_mdsize);
 796                 } else {
 797                         size[bufcount] = rc;
 798                 }
 799                 bufcount++;
 800         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 801                 if (inode->i_size + 1 != body->eadatasize)
 802                         CERROR("symlink size: %Lu, reply space: %d\n",
 803                                inode->i_size + 1, body->eadatasize);
 804                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 805                 bufcount++;
 806                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 807                        inode->i_size + 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_blksize - 1)) != 0) {
1215                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1216                        body->size, de->d_inode->i_blksize);
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_blksize - 1)) {
1222                 CERROR("size %u is not multiple of blocksize %lu\n",
1223                        body->nlink, de->d_inode->i_blksize);
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 = file->f_dentry->d_inode->i_size;
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
1956         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1957         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
1958         if (obd->obd_namespace == NULL) {
1959                 mds_cleanup(obd);
1960                 GOTO(err_ops, rc = -ENOMEM);
1961         }
1962         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1963
1964         lprocfs_init_vars(mds, &lvars);
1965         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1966             lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
1967                 /* Init private stats here */
1968                 mds_stats_counter_init(obd->obd_stats);
1969                 obd->obd_proc_exports = proc_mkdir("exports",
1970                                                    obd->obd_proc_entry);
1971         }
1972
1973         rc = mds_fs_setup(obd, mnt);
1974         if (rc) {
1975                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
1976                        obd->obd_name, rc);
1977                 GOTO(err_ns, rc);
1978         }
1979
1980         rc = mds_lov_presetup(mds, lcfg);
1981         if (rc < 0)
1982                 GOTO(err_fs, rc);
1983
1984         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1985                            "mds_ldlm_client", &obd->obd_ldlm_client);
1986         obd->obd_replayable = 1;
1987
1988         rc = lquota_setup(mds_quota_interface_ref, obd);
1989         if (rc)
1990                 GOTO(err_fs, rc);
1991
1992 #if 0
1993         mds->mds_group_hash = upcall_cache_init(obd->obd_name);
1994         if (IS_ERR(mds->mds_group_hash)) {
1995                 rc = PTR_ERR(mds->mds_group_hash);
1996                 mds->mds_group_hash = NULL;
1997                 GOTO(err_qctxt, rc);
1998         }
1999 #endif
2000
2001         /* Don't wait for mds_postrecov trying to clear orphans */
2002         obd->obd_async_recov = 1;
2003         rc = mds_postsetup(obd);
2004         /* Bug 11557 - allow async abort_recov start
2005            FIXME can remove most of this obd_async_recov plumbing
2006         obd->obd_async_recov = 0;
2007         */
2008         if (rc)
2009                 GOTO(err_qctxt, rc);
2010
2011         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2012         if (uuid_ptr != NULL) {
2013                 class_uuid_unparse(uuid_ptr, &uuid);
2014                 str = uuid.uuid;
2015         } else {
2016                 str = "no UUID";
2017         }
2018
2019         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2020         if (obd->obd_recovering) {
2021                 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2022                               "recovery until %d %s reconnect, or if no clients"
2023                               " reconnect for %d:%.02d; during that time new "
2024                               "clients will not be allowed to connect. "
2025                               "Recovery progress can be monitored by watching "
2026                               "/proc/fs/lustre/mds/%s/recovery_status.\n",
2027                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2028                               label ?: "", label ? "/" : "", str,
2029                               obd->obd_max_recoverable_clients,
2030                               (obd->obd_max_recoverable_clients == 1) ?
2031                               "client" : "clients",
2032                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
2033                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
2034                               obd->obd_name);
2035         } else {
2036                 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2037                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2038                               label ?: "", label ? "/" : "", str,
2039                               obd->obd_replayable ? "enabled" : "disabled");
2040         }
2041
2042         if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
2043                 ldlm_timeout = 6;
2044
2045         RETURN(0);
2046
2047 err_qctxt:
2048         lquota_cleanup(mds_quota_interface_ref, obd);
2049 err_fs:
2050         /* No extra cleanup needed for llog_init_commit_thread() */
2051         mds_fs_cleanup(obd);
2052 #if 0
2053         upcall_cache_cleanup(mds->mds_group_hash);
2054         mds->mds_group_hash = NULL;
2055 #endif
2056 err_ns:
2057         lprocfs_obd_cleanup(obd);
2058         lprocfs_free_obd_stats(obd);
2059         ldlm_namespace_free(obd->obd_namespace, 0);
2060         obd->obd_namespace = NULL;
2061 err_ops:
2062         fsfilt_put_ops(obd->obd_fsops);
2063 err_put:
2064         server_put_mount(obd->obd_name, mnt);
2065         obd->u.obt.obt_sb = NULL;
2066         return rc;
2067 }
2068
2069 static int mds_lov_clean(struct obd_device *obd)
2070 {
2071         struct mds_obd *mds = &obd->u.mds;
2072         struct obd_device *osc = mds->mds_osc_obd;
2073         ENTRY;
2074
2075         if (mds->mds_profile) {
2076                 class_del_profile(mds->mds_profile);
2077                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2078                 mds->mds_profile = NULL;
2079         }
2080
2081         /* There better be a lov */
2082         if (!osc)
2083                 RETURN(0);
2084         if (IS_ERR(osc))
2085                 RETURN(PTR_ERR(osc));
2086
2087         obd_register_observer(osc, NULL);
2088
2089         /* Give lov our same shutdown flags */
2090         osc->obd_force = obd->obd_force;
2091         osc->obd_fail = obd->obd_fail;
2092
2093         /* Cleanup the lov */
2094         obd_disconnect(mds->mds_osc_exp);
2095         class_manual_cleanup(osc);
2096         mds->mds_osc_exp = NULL;
2097
2098         RETURN(0);
2099 }
2100
2101 static int mds_postsetup(struct obd_device *obd)
2102 {
2103         struct mds_obd *mds = &obd->u.mds;
2104         int rc = 0;
2105         ENTRY;
2106
2107         rc = llog_setup(obd, NULL, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2108                         &llog_lvfs_ops);
2109         if (rc)
2110                 RETURN(rc);
2111
2112         rc = llog_setup(obd, NULL, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2113                         &llog_lvfs_ops);
2114         if (rc)
2115                 RETURN(rc);
2116
2117         if (mds->mds_profile) {
2118                 struct lustre_profile *lprof;
2119                 /* The profile defines which osc and mdc to connect to, for a
2120                    client.  We reuse that here to figure out the name of the
2121                    lov to use (and ignore lprof->lp_md).
2122                    The profile was set in the config log with
2123                    LCFG_MOUNTOPT profilenm oscnm mdcnm */
2124                 lprof = class_get_profile(mds->mds_profile);
2125                 if (lprof == NULL) {
2126                         CERROR("No profile found: %s\n", mds->mds_profile);
2127                         GOTO(err_cleanup, rc = -ENOENT);
2128                 }
2129                 rc = mds_lov_connect(obd, lprof->lp_dt);
2130                 if (rc)
2131                         GOTO(err_cleanup, rc);
2132         }
2133
2134         RETURN(rc);
2135
2136 err_cleanup:
2137         mds_lov_clean(obd);
2138         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2139         llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2140         RETURN(rc);
2141 }
2142
2143 int mds_postrecov(struct obd_device *obd)
2144 {
2145         int rc;
2146         ENTRY;
2147
2148         if (obd->obd_fail)
2149                 RETURN(0);
2150
2151         LASSERT(!obd->obd_recovering);
2152         LASSERT(llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT) != NULL);
2153
2154         /* FIXME why not put this in the synchronize? */
2155         /* set nextid first, so we are sure it happens */
2156         rc = mds_lov_set_nextid(obd);
2157         if (rc) {
2158                 CERROR("%s: mds_lov_set_nextid failed %d\n",
2159                        obd->obd_name, rc);
2160                 GOTO(out, rc);
2161         }
2162
2163         /* clean PENDING dir */
2164         if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2165                 rc = mds_cleanup_pending(obd);
2166                 if (rc < 0)
2167                         GOTO(out, rc);
2168
2169         /* FIXME Does target_finish_recovery really need this to block? */
2170         /* Notify the LOV, which will in turn call mds_notify for each tgt */
2171         /* This means that we have to hack obd_notify to think we're obd_set_up
2172            during mds_lov_connect. */
2173         obd_notify(obd->u.mds.mds_osc_obd, NULL,
2174                    obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
2175                    OBD_NOTIFY_SYNC, NULL);
2176
2177         /* quota recovery */
2178         lquota_recovery(mds_quota_interface_ref, obd);
2179
2180 out:
2181         RETURN(rc);
2182 }
2183
2184 /* We need to be able to stop an mds_lov_synchronize */
2185 static int mds_lov_early_clean(struct obd_device *obd)
2186 {
2187         struct mds_obd *mds = &obd->u.mds;
2188         struct obd_device *osc = mds->mds_osc_obd;
2189
2190         if (!osc || (!obd->obd_force && !obd->obd_fail))
2191                 return(0);
2192
2193         CDEBUG(D_HA, "abort inflight\n");
2194         return (obd_precleanup(osc, OBD_CLEANUP_EARLY));
2195 }
2196
2197 static int mds_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2198 {
2199         int rc = 0;
2200         ENTRY;
2201
2202         switch (stage) {
2203         case OBD_CLEANUP_EARLY:
2204                 break;
2205         case OBD_CLEANUP_EXPORTS:
2206                 /*XXX Use this for mdd mds cleanup, so comment out
2207                  *this target_cleanup_recovery for this tmp MDD MDS
2208                  *Wangdi*/
2209                 if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2210                         target_cleanup_recovery(obd);
2211                 mds_lov_early_clean(obd);
2212                 break;
2213         case OBD_CLEANUP_SELF_EXP:
2214                 mds_lov_disconnect(obd);
2215                 mds_lov_clean(obd);
2216                 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2217                 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2218                 rc = obd_llog_finish(obd, 0);
2219                 break;
2220         case OBD_CLEANUP_OBD:
2221                 break;
2222         }
2223         RETURN(rc);
2224 }
2225
2226 static int mds_cleanup(struct obd_device *obd)
2227 {
2228         struct mds_obd *mds = &obd->u.mds;
2229         lvfs_sbdev_type save_dev;
2230         ENTRY;
2231
2232         if (obd->u.obt.obt_sb == NULL)
2233                 RETURN(0);
2234         save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2235
2236         if (mds->mds_osc_exp)
2237                 /* lov export was disconnected by mds_lov_clean;
2238                    we just need to drop our ref */
2239                 class_export_put(mds->mds_osc_exp);
2240
2241         lprocfs_obd_cleanup(obd);
2242         lprocfs_free_obd_stats(obd);
2243
2244         lquota_cleanup(mds_quota_interface_ref, obd);
2245
2246         mds_update_server_data(obd, 1);
2247         if (mds->mds_lov_objids != NULL)
2248                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2249         mds_fs_cleanup(obd);
2250
2251 #if 0
2252         upcall_cache_cleanup(mds->mds_group_hash);
2253         mds->mds_group_hash = NULL;
2254 #endif
2255
2256         server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2257         obd->u.obt.obt_sb = NULL;
2258
2259         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2260
2261         spin_lock_bh(&obd->obd_processing_task_lock);
2262         if (obd->obd_recovering) {
2263                 target_cancel_recovery_timer(obd);
2264                 obd->obd_recovering = 0;
2265         }
2266         spin_unlock_bh(&obd->obd_processing_task_lock);
2267
2268         fsfilt_put_ops(obd->obd_fsops);
2269
2270         LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2271
2272         RETURN(0);
2273 }
2274
2275 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2276                                         struct ldlm_lock *new_lock,
2277                                         struct ldlm_lock **old_lock,
2278                                         struct lustre_handle *lockh)
2279 {
2280         struct obd_export *exp = req->rq_export;
2281         struct ldlm_request *dlmreq =
2282                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2283         struct lustre_handle remote_hdl = dlmreq->lock_handle[0];
2284         struct list_head *iter;
2285
2286         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2287                 return;
2288
2289         spin_lock(&exp->exp_ldlm_data.led_lock);
2290         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2291                 struct ldlm_lock *lock;
2292                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2293                 if (lock == new_lock)
2294                         continue;
2295                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2296                         lockh->cookie = lock->l_handle.h_cookie;
2297                         LDLM_DEBUG(lock, "restoring lock cookie");
2298                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2299                                   lockh->cookie);
2300                         if (old_lock)
2301                                 *old_lock = LDLM_LOCK_GET(lock);
2302                         spin_unlock(&exp->exp_ldlm_data.led_lock);
2303                         return;
2304                 }
2305         }
2306         spin_unlock(&exp->exp_ldlm_data.led_lock);
2307
2308         /* If the xid matches, then we know this is a resent request,
2309          * and allow it. (It's probably an OPEN, for which we don't
2310          * send a lock */
2311         if (req->rq_xid ==
2312             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2313                 return;
2314
2315         if (req->rq_xid ==
2316             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2317                 return;
2318
2319         /* This remote handle isn't enqueued, so we never received or
2320          * processed this request.  Clear MSG_RESENT, because it can
2321          * be handled like any normal request now. */
2322
2323         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2324
2325         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2326                   remote_hdl.cookie);
2327 }
2328
2329 int intent_disposition(struct ldlm_reply *rep, int flag)
2330 {
2331         if (!rep)
2332                 return 0;
2333         return (rep->lock_policy_res1 & flag);
2334 }
2335
2336 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2337 {
2338         if (!rep)
2339                 return;
2340         rep->lock_policy_res1 |= flag;
2341 }
2342
2343 static int mds_intent_policy(struct ldlm_namespace *ns,
2344                              struct ldlm_lock **lockp, void *req_cookie,
2345                              ldlm_mode_t mode, int flags, void *data)
2346 {
2347         struct ptlrpc_request *req = req_cookie;
2348         struct ldlm_lock *lock = *lockp;
2349         struct ldlm_intent *it;
2350         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2351         struct ldlm_reply *rep;
2352         struct lustre_handle lockh = { 0 };
2353         struct ldlm_lock *new_lock = NULL;
2354         int getattr_part = MDS_INODELOCK_UPDATE;
2355         int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2356                            [DLM_LOCKREPLY_OFF]   = sizeof(struct ldlm_reply),
2357                            [DLM_REPLY_REC_OFF]   = sizeof(struct mds_body),
2358                            [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2359         int repbufcnt = 4, rc;
2360         ENTRY;
2361
2362         LASSERT(req != NULL);
2363
2364         if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2365                 /* No intent was provided */
2366                 rc = lustre_pack_reply(req, 2, repsize, NULL);
2367                 LASSERT(rc == 0);
2368                 RETURN(0);
2369         }
2370
2371         it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2372                                 lustre_swab_ldlm_intent);
2373         if (it == NULL) {
2374                 CERROR("Intent missing\n");
2375                 RETURN(req->rq_status = -EFAULT);
2376         }
2377
2378         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2379
2380         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2381             (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2382                 /* we should never allow OBD_CONNECT_ACL if not configured */
2383                 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2384         else if (it->opc & IT_UNLINK)
2385                 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2386
2387         rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2388         if (rc)
2389                 RETURN(req->rq_status = rc);
2390
2391         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2392         intent_set_disposition(rep, DISP_IT_EXECD);
2393
2394
2395         /* execute policy */
2396         switch ((long)it->opc) {
2397         case IT_OPEN:
2398         case IT_CREAT|IT_OPEN:
2399                 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2400                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2401                                             &lockh);
2402                 /* XXX swab here to assert that an mds_open reint
2403                  * packet is following */
2404                 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2405                                                   &lockh);
2406 #if 0
2407                 /* We abort the lock if the lookup was negative and
2408                  * we did not make it to the OPEN portion */
2409                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2410                         RETURN(ELDLM_LOCK_ABORTED);
2411                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2412                     !intent_disposition(rep, DISP_OPEN_OPEN))
2413 #endif
2414
2415                 /* If there was an error of some sort or if we are not
2416                  * returning any locks */
2417                 if (rep->lock_policy_res2 ||
2418                     !intent_disposition(rep, DISP_OPEN_LOCK))
2419                         RETURN(ELDLM_LOCK_ABORTED);
2420                 break;
2421         case IT_LOOKUP:
2422                         getattr_part = MDS_INODELOCK_LOOKUP;
2423         case IT_GETATTR:
2424                         getattr_part |= MDS_INODELOCK_LOOKUP;
2425                         OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2426         case IT_READDIR:
2427                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2428                                             &new_lock, &lockh);
2429
2430                 /* INODEBITS_INTEROP: if this lock was converted from a
2431                  * plain lock (client does not support inodebits), then
2432                  * child lock must be taken with both lookup and update
2433                  * bits set for all operations.
2434                  */
2435                 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2436                         getattr_part = MDS_INODELOCK_LOOKUP |
2437                                        MDS_INODELOCK_UPDATE;
2438
2439                 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2440                                                          getattr_part, &lockh);
2441                 /* FIXME: LDLM can set req->rq_status. MDS sets
2442                    policy_res{1,2} with disposition and status.
2443                    - replay: returns 0 & req->status is old status
2444                    - otherwise: returns req->status */
2445                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2446                         rep->lock_policy_res2 = 0;
2447                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2448                     rep->lock_policy_res2)
2449                         RETURN(ELDLM_LOCK_ABORTED);
2450                 if (req->rq_status != 0) {
2451                         LBUG();
2452                         rep->lock_policy_res2 = req->rq_status;
2453                         RETURN(ELDLM_LOCK_ABORTED);
2454                 }
2455                 break;
2456         default:
2457                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2458                 RETURN(-EFAULT);
2459         }
2460
2461         /* By this point, whatever function we called above must have either
2462          * filled in 'lockh', been an intent replay, or returned an error.  We
2463          * want to allow replayed RPCs to not get a lock, since we would just
2464          * drop it below anyways because lock replay is done separately by the
2465          * client afterwards.  For regular RPCs we want to give the new lock to
2466          * the client instead of whatever lock it was about to get. */
2467         if (new_lock == NULL)
2468                 new_lock = ldlm_handle2lock(&lockh);
2469         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2470                 RETURN(0);
2471
2472         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2473                  it->opc, lockh.cookie);
2474
2475         /* If we've already given this lock to a client once, then we should
2476          * have no readers or writers.  Otherwise, we should have one reader
2477          * _or_ writer ref (which will be zeroed below) before returning the
2478          * lock to a client. */
2479         if (new_lock->l_export == req->rq_export) {
2480                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2481         } else {
2482                 LASSERT(new_lock->l_export == NULL);
2483                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2484         }
2485
2486         *lockp = new_lock;
2487
2488         if (new_lock->l_export == req->rq_export) {
2489                 /* Already gave this to the client, which means that we
2490                  * reconstructed a reply. */
2491                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2492                         MSG_RESENT);
2493                 RETURN(ELDLM_LOCK_REPLACED);
2494         }
2495
2496         /* Fixup the lock to be given to the client */
2497         lock_res_and_lock(new_lock);
2498         new_lock->l_readers = 0;
2499         new_lock->l_writers = 0;
2500
2501         new_lock->l_export = class_export_get(req->rq_export);
2502         spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2503         list_add(&new_lock->l_export_chain,
2504                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2505         spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2506
2507         new_lock->l_blocking_ast = lock->l_blocking_ast;
2508         new_lock->l_completion_ast = lock->l_completion_ast;
2509
2510         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2511                sizeof(lock->l_remote_handle));
2512
2513         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2514
2515         unlock_res_and_lock(new_lock);
2516         LDLM_LOCK_PUT(new_lock);
2517
2518         RETURN(ELDLM_LOCK_REPLACED);
2519 }
2520
2521 static int mdt_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2522 {
2523         struct mds_obd *mds = &obd->u.mds;
2524         struct lprocfs_static_vars lvars;
2525         int mds_min_threads;
2526         int mds_max_threads;
2527         int rc = 0;
2528         ENTRY;
2529
2530         lprocfs_init_vars(mdt, &lvars);
2531         lprocfs_obd_setup(obd, lvars.obd_vars);
2532
2533         sema_init(&mds->mds_health_sem, 1);
2534
2535         if (mds_num_threads) {
2536                 /* If mds_num_threads is set, it is the min and the max. */
2537                 if (mds_num_threads > MDS_THREADS_MAX)
2538                         mds_num_threads = MDS_THREADS_MAX;
2539                 if (mds_num_threads < MDS_THREADS_MIN)
2540                         mds_num_threads = MDS_THREADS_MIN;
2541                 mds_max_threads = mds_min_threads = mds_num_threads;
2542         } else {
2543                 /* Base min threads on memory and cpus */
2544                 mds_min_threads = smp_num_cpus * num_physpages >>
2545                         (27 - CFS_PAGE_SHIFT);
2546                 if (mds_min_threads < MDS_THREADS_MIN)
2547                         mds_min_threads = MDS_THREADS_MIN;
2548                 /* Largest auto threads start value */
2549                 if (mds_min_threads > 32)
2550                         mds_min_threads = 32;
2551                 mds_max_threads = min(MDS_THREADS_MAX, mds_min_threads * 4);
2552         }
2553
2554         mds->mds_service =
2555                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2556                                 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2557                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2558                                 mds_handle, LUSTRE_MDS_NAME,
2559                                 obd->obd_proc_entry, NULL,
2560                                 mds_min_threads, mds_max_threads, "ll_mdt", 0);
2561
2562         if (!mds->mds_service) {
2563                 CERROR("failed to start service\n");
2564                 GOTO(err_lprocfs, rc = -ENOMEM);
2565         }
2566
2567         rc = ptlrpc_start_threads(obd, mds->mds_service);
2568         if (rc)
2569                 GOTO(err_thread, rc);
2570
2571         mds->mds_setattr_service =
2572                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2573                                 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2574                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2575                                 mds_handle, "mds_setattr",
2576                                 obd->obd_proc_entry, NULL,
2577                                 mds_min_threads, mds_max_threads,
2578                                 "ll_mdt_attr", 0);
2579         if (!mds->mds_setattr_service) {
2580                 CERROR("failed to start getattr service\n");
2581                 GOTO(err_thread, rc = -ENOMEM);
2582         }
2583
2584         rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2585         if (rc)
2586                 GOTO(err_thread2, rc);
2587
2588         mds->mds_readpage_service =
2589                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2590                                 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2591                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2592                                 mds_handle, "mds_readpage",
2593                                 obd->obd_proc_entry, NULL,
2594                                 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2595                                 "ll_mdt_rdpg", 0);
2596         if (!mds->mds_readpage_service) {
2597                 CERROR("failed to start readpage service\n");
2598                 GOTO(err_thread2, rc = -ENOMEM);
2599         }
2600
2601         rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2602
2603         if (rc)
2604                 GOTO(err_thread3, rc);
2605
2606         ping_evictor_start();
2607
2608         RETURN(0);
2609
2610 err_thread3:
2611         ptlrpc_unregister_service(mds->mds_readpage_service);
2612         mds->mds_readpage_service = NULL;
2613 err_thread2:
2614         ptlrpc_unregister_service(mds->mds_setattr_service);
2615         mds->mds_setattr_service = NULL;
2616 err_thread:
2617         ptlrpc_unregister_service(mds->mds_service);
2618         mds->mds_service = NULL;
2619 err_lprocfs:
2620         lprocfs_obd_cleanup(obd);
2621         return rc;
2622 }
2623
2624 static int mdt_cleanup(struct obd_device *obd)
2625 {
2626         struct mds_obd *mds = &obd->u.mds;
2627         ENTRY;
2628
2629         ping_evictor_stop();
2630
2631         down(&mds->mds_health_sem);
2632         ptlrpc_unregister_service(mds->mds_readpage_service);
2633         ptlrpc_unregister_service(mds->mds_setattr_service);
2634         ptlrpc_unregister_service(mds->mds_service);
2635         mds->mds_readpage_service = NULL;
2636         mds->mds_setattr_service = NULL;
2637         mds->mds_service = NULL;
2638         up(&mds->mds_health_sem);
2639
2640         lprocfs_obd_cleanup(obd);
2641
2642         RETURN(0);
2643 }
2644
2645 static int mdt_health_check(struct obd_device *obd)
2646 {
2647         struct mds_obd *mds = &obd->u.mds;
2648         int rc = 0;
2649
2650         down(&mds->mds_health_sem);
2651         rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2652         rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2653         rc |= ptlrpc_service_health_check(mds->mds_service);
2654         up(&mds->mds_health_sem);
2655
2656         /*
2657          * health_check to return 0 on healthy
2658          * and 1 on unhealthy.
2659          */
2660         if(rc != 0)
2661                 rc = 1;
2662
2663         return rc;
2664 }
2665
2666 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2667                                           void *data)
2668 {
2669         struct obd_device *obd = data;
2670         struct ll_fid fid;
2671         fid.id = id;
2672         fid.generation = gen;
2673         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2674 }
2675
2676 static int mds_health_check(struct obd_device *obd)
2677 {
2678         struct obd_device_target *odt = &obd->u.obt;
2679 #ifdef USE_HEALTH_CHECK_WRITE
2680         struct mds_obd *mds = &obd->u.mds;
2681 #endif
2682         int rc = 0;
2683
2684         if (odt->obt_sb->s_flags & MS_RDONLY)
2685                 rc = 1;
2686
2687 #ifdef USE_HEALTH_CHECK_WRITE
2688         LASSERT(mds->mds_health_check_filp != NULL);
2689         rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2690 #endif
2691         return rc;
2692 }
2693
2694 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2695 {
2696         struct lustre_cfg *lcfg = buf;
2697         struct lprocfs_static_vars lvars;
2698         int rc;
2699
2700         lprocfs_init_vars(mds, &lvars);
2701
2702         rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2703         return(rc);
2704 }
2705
2706 struct lvfs_callback_ops mds_lvfs_ops = {
2707         l_fid2dentry:     mds_lvfs_fid2dentry,
2708 };
2709
2710 /* use obd ops to offer management infrastructure */
2711 static struct obd_ops mds_obd_ops = {
2712         .o_owner           = THIS_MODULE,
2713         .o_connect         = mds_connect,
2714         .o_reconnect       = mds_reconnect,
2715         .o_init_export     = mds_init_export,
2716         .o_destroy_export  = mds_destroy_export,
2717         .o_disconnect      = mds_disconnect,
2718         .o_setup           = mds_setup,
2719         .o_precleanup      = mds_precleanup,
2720         .o_cleanup         = mds_cleanup,
2721         .o_postrecov       = mds_postrecov,
2722         .o_statfs          = mds_obd_statfs,
2723         .o_iocontrol       = mds_iocontrol,
2724         .o_create          = mds_obd_create,
2725         .o_destroy         = mds_obd_destroy,
2726         .o_llog_init       = mds_llog_init,
2727         .o_llog_finish     = mds_llog_finish,
2728         .o_notify          = mds_notify,
2729         .o_health_check    = mds_health_check,
2730         .o_process_config  = mds_process_config,
2731 };
2732
2733 static struct obd_ops mdt_obd_ops = {
2734         .o_owner           = THIS_MODULE,
2735         .o_setup           = mdt_setup,
2736         .o_cleanup         = mdt_cleanup,
2737         .o_health_check    = mdt_health_check,
2738 };
2739
2740 quota_interface_t *mds_quota_interface_ref;
2741 extern quota_interface_t mds_quota_interface;
2742
2743 static __attribute__((unused)) int __init mds_init(void)
2744 {
2745         int rc;
2746         struct lprocfs_static_vars lvars;
2747
2748         request_module("lquota");
2749         mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2750         rc = lquota_init(mds_quota_interface_ref);
2751         if (rc) {
2752                 if (mds_quota_interface_ref)
2753                         PORTAL_SYMBOL_PUT(mds_quota_interface);
2754                 return rc;
2755         }
2756         init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2757
2758         lprocfs_init_vars(mds, &lvars);
2759         class_register_type(&mds_obd_ops, NULL,
2760                             lvars.module_vars, LUSTRE_MDS_NAME, NULL);
2761         lprocfs_init_vars(mdt, &lvars);
2762         mdt_obd_ops = mdt_obd_ops; //make compiler happy
2763 //        class_register_type(&mdt_obd_ops, NULL,
2764 //                            lvars.module_vars, LUSTRE_MDT_NAME, NULL);
2765
2766         return 0;
2767 }
2768
2769 static __attribute__((unused)) void /*__exit*/ mds_exit(void)
2770 {
2771         lquota_exit(mds_quota_interface_ref);
2772         if (mds_quota_interface_ref)
2773                 PORTAL_SYMBOL_PUT(mds_quota_interface);
2774
2775         class_unregister_type(LUSTRE_MDS_NAME);
2776 //        class_unregister_type(LUSTRE_MDT_NAME);
2777 }
2778 /*mds still need lov setup here*/
2779 static int mds_cmd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2780 {
2781         struct mds_obd *mds = &obd->u.mds;
2782         struct lvfs_run_ctxt saved;
2783         const char     *dev;
2784         struct vfsmount *mnt;
2785         struct lustre_sb_info *lsi;
2786         struct lustre_mount_info *lmi;
2787         struct dentry  *dentry;
2788         struct file *file;
2789         int rc = 0;
2790         ENTRY;
2791
2792         CDEBUG(D_INFO, "obd %s setup \n", obd->obd_name);
2793         if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2794                 RETURN(0);
2795
2796         if (lcfg->lcfg_bufcount < 5) {
2797                 CERROR("invalid arg for setup %s\n", MDD_OBD_NAME);
2798                 RETURN(-EINVAL);
2799         }
2800         dev = lustre_cfg_string(lcfg, 4);
2801         lmi = server_get_mount(dev);
2802         LASSERT(lmi != NULL);
2803
2804         lsi = s2lsi(lmi->lmi_sb);
2805         mnt = lmi->lmi_mnt;
2806         /* FIXME: MDD LOV initialize objects.
2807          * we need only lmi here but not get mount
2808          * OSD did mount already, so put mount back
2809          */
2810         atomic_dec(&lsi->lsi_mounts);
2811         mntput(mnt);
2812
2813         obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2814         mds_init_ctxt(obd, mnt);
2815
2816         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2817         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
2818         if (IS_ERR(dentry)) {
2819                 rc = PTR_ERR(dentry);
2820                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
2821                 GOTO(err_putfs, rc);
2822         }
2823         mds->mds_objects_dir = dentry;
2824
2825         dentry = lookup_one_len("__iopen__", current->fs->pwd,
2826                                 strlen("__iopen__"));
2827         if (IS_ERR(dentry)) {
2828                 rc = PTR_ERR(dentry);
2829                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
2830                 GOTO(err_objects, rc);
2831         }
2832
2833         mds->mds_fid_de = dentry;
2834         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
2835                 rc = -ENOENT;
2836                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
2837                 GOTO(err_fid, rc);
2838         }
2839
2840         /* open and test the lov objd file */
2841         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
2842         if (IS_ERR(file)) {
2843                 rc = PTR_ERR(file);
2844                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
2845                 GOTO(err_fid, rc = PTR_ERR(file));
2846         }
2847         mds->mds_lov_objid_filp = file;
2848         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
2849                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
2850                        file->f_dentry->d_inode->i_mode);
2851                 GOTO(err_lov_objid, rc = -ENOENT);
2852         }
2853
2854         rc = mds_lov_presetup(mds, lcfg);
2855         if (rc < 0)
2856                 GOTO(err_objects, rc);
2857
2858         /* Don't wait for mds_postrecov trying to clear orphans */
2859         obd->obd_async_recov = 1;
2860         rc = mds_postsetup(obd);
2861         /* Bug 11557 - allow async abort_recov start
2862            FIXME can remove most of this obd_async_recov plumbing
2863         obd->obd_async_recov = 0;
2864         */
2865
2866         if (rc)
2867                 GOTO(err_objects, rc);
2868
2869         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2870         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2871
2872 err_pop:
2873         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2874         RETURN(rc);
2875 err_lov_objid:
2876         if (mds->mds_lov_objid_filp &&
2877                 filp_close((struct file *)mds->mds_lov_objid_filp, 0))
2878                 CERROR("can't close %s after error\n", LOV_OBJID);
2879 err_fid:
2880         dput(mds->mds_fid_de);
2881 err_objects:
2882         dput(mds->mds_objects_dir);
2883 err_putfs:
2884         fsfilt_put_ops(obd->obd_fsops);
2885         goto err_pop;
2886 }
2887
2888 static int mds_cmd_cleanup(struct obd_device *obd)
2889 {
2890         struct mds_obd *mds = &obd->u.mds;
2891         struct lvfs_run_ctxt saved;
2892         int rc = 0;
2893         ENTRY;
2894
2895         if (obd->obd_fail)
2896                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2897                               "will be preserved.\n", obd->obd_name);
2898
2899         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2900         if (mds->mds_lov_objid_filp) {
2901                 rc = filp_close((struct file *)mds->mds_lov_objid_filp, 0);
2902                 mds->mds_lov_objid_filp = NULL;
2903                 if (rc)
2904                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
2905         }
2906         if (mds->mds_objects_dir != NULL) {
2907                 l_dput(mds->mds_objects_dir);
2908                 mds->mds_objects_dir = NULL;
2909         }
2910
2911         if (mds->mds_lov_objids != NULL)
2912                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2913
2914         shrink_dcache_parent(mds->mds_fid_de);
2915         dput(mds->mds_fid_de);
2916         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2917         fsfilt_put_ops(obd->obd_fsops);
2918
2919         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2920         RETURN(rc);
2921 }
2922
2923 #if 0
2924 static int mds_cmd_health_check(struct obd_device *obd)
2925 {
2926         return 0;
2927 }
2928 #endif
2929 static struct obd_ops mds_cmd_obd_ops = {
2930         .o_owner           = THIS_MODULE,
2931         .o_setup           = mds_cmd_setup,
2932         .o_cleanup         = mds_cmd_cleanup,
2933         .o_precleanup      = mds_precleanup,
2934         .o_create          = mds_obd_create,
2935         .o_destroy         = mds_obd_destroy,
2936         .o_llog_init       = mds_llog_init,
2937         .o_llog_finish     = mds_llog_finish,
2938         .o_notify          = mds_notify,
2939         .o_postrecov       = mds_postrecov,
2940         //   .o_health_check    = mds_cmd_health_check,
2941 };
2942
2943 static int __init mds_cmd_init(void)
2944 {
2945         struct lprocfs_static_vars lvars;
2946
2947         lprocfs_init_vars(mds, &lvars);
2948         class_register_type(&mds_cmd_obd_ops, NULL, lvars.module_vars,
2949                             LUSTRE_MDS_NAME, NULL);
2950
2951         return 0;
2952 }
2953
2954 static void /*__exit*/ mds_cmd_exit(void)
2955 {
2956         class_unregister_type(LUSTRE_MDS_NAME);
2957 }
2958
2959 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2960 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2961 MODULE_LICENSE("GPL");
2962
2963 module_init(mds_cmd_init);
2964 module_exit(mds_cmd_exit);