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) 2001-2003 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 Lustre, http://www.lustre.org.
  14  *
  15  *   Lustre is free software; you can redistribute it and/or
  16  *   modify it under the terms of version 2 of the GNU General Public
  17  *   License as published by the Free Software Foundation.
  18  *
  19  *   Lustre is distributed in the hope that it will be useful,
  20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   GNU General Public License for more details.
  23  *
  24  *   You should have received a copy of the GNU General Public License
  25  *   along with Lustre; if not, write to the Free Software
  26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/lustre_mds.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/init.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/random.h>
  40 #include <linux/fs.h>
  41 #include <linux/jbd.h>
  42 #include <linux/namei.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 #include <linux/obd_lov.h>
  53 #include <linux/obd_ost.h>
  54 #include <linux/lustre_mds.h>
  55 #include <linux/lustre_fsfilt.h>
  56 #include <linux/lprocfs_status.h>
  57 #include <linux/lustre_commit_confd.h>
  58 #include <linux/lustre_acl.h>
  59 #include "mds_internal.h"
  60 #include <linux/lustre_sec.h>
  61
  62 static int mds_intent_policy(struct ldlm_namespace *ns,
  63                              struct ldlm_lock **lockp, void *req_cookie,
  64                              ldlm_mode_t mode, int flags, void *data);
  65 static int mds_postsetup(struct obd_device *obd);
  66 static int mds_cleanup(struct obd_device *obd, int flags);
  67
  68
  69 /* Assumes caller has already pushed into the kernel filesystem context */
  70 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  71                         loff_t offset, int count)
  72 {
  73         struct ptlrpc_bulk_desc *desc;
  74         struct l_wait_info lwi;
  75         struct page **pages;
  76         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  77         ENTRY;
  78
  79         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  80
  81         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  82         OBD_ALLOC(pages, sizeof(*pages) * npages);
  83         if (!pages)
  84                 GOTO(out, rc = -ENOMEM);
  85
  86         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  87                                     MDS_BULK_PORTAL);
  88         if (desc == NULL)
  89                 GOTO(out_free, rc = -ENOMEM);
  90
  91         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  92                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  93
  94                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  95                 if (pages[i] == NULL)
  96                         GOTO(cleanup_buf, rc = -ENOMEM);
  97
  98                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  99         }
 100
 101         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 102                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 103                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 104                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 105                        file->f_dentry->d_inode->i_size);
 106
 107                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 108                                      kmap(pages[i]), tmpsize, &offset);
 109                 kunmap(pages[i]);
 110
 111                 if (rc != tmpsize)
 112                         GOTO(cleanup_buf, rc = -EIO);
 113         }
 114
 115         LASSERT(desc->bd_nob == count);
 116
 117         rc = ptlrpc_start_bulk_transfer(desc);
 118         if (rc)
 119                 GOTO(cleanup_buf, rc);
 120
 121         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 122                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 123                        OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
 124                 GOTO(abort_bulk, rc);
 125         }
 126
 127         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 128         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 129         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 130
 131         if (rc == 0) {
 132                 if (desc->bd_success &&
 133                     desc->bd_nob_transferred == count)
 134                         GOTO(cleanup_buf, rc);
 135
 136                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 137         }
 138
 139         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 140                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 141                   desc->bd_nob_transferred, count,
 142                   req->rq_export->exp_client_uuid.uuid,
 143                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 144
 145         ptlrpc_fail_export(req->rq_export);
 146
 147         EXIT;
 148  abort_bulk:
 149         ptlrpc_abort_bulk (desc);
 150  cleanup_buf:
 151         for (i = 0; i < npages; i++)
 152                 if (pages[i])
 153                         __free_pages(pages[i], 0);
 154
 155         ptlrpc_free_bulk(desc);
 156  out_free:
 157         OBD_FREE(pages, sizeof(*pages) * npages);
 158  out:
 159         return rc;
 160 }
 161
 162 extern char *ldlm_lockname[];
 163
 164 int mds_lock_mode_for_dir(struct obd_device *obd,
 165                           struct dentry *dentry, int mode)
 166 {
 167         int ret_mode = 0, split;
 168
 169         /* any dir access needs couple locks:
 170          * 1) on part of dir we gonna lookup/modify in
 171          * 2) on a whole dir to protect it from concurrent splitting
 172          *    and to flush client's cache for readdir()
 173          * so, for a given mode and dentry this routine decides what
 174          * lock mode to use for lock #2:
 175          * 1) if caller's gonna lookup in dir then we need to protect
 176          *    dir from being splitted only - LCK_CR
 177          * 2) if caller's gonna modify dir then we need to protect
 178          *    dir from being splitted and to flush cache - LCK_CW
 179          * 3) if caller's gonna modify dir and that dir seems ready
 180          *    for splitting then we need to protect it from any
 181          *    type of access (lookup/modify/split) - LCK_EX -bzzz */
 182
 183         split = mds_splitting_expected(obd, dentry);
 184
 185         /*
 186          * it is important to check here only for MDS_NO_SPLITTABLE. The reason
 187          * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
 188          * and another thread will not split it on the quiet. But if we have
 189          * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
 190          * but not now (for current thread) and we should consider that it can
 191          * happen soon and go that branch which can yield LCK_EX to protect from
 192          * possible splitting.
 193          */
 194         if (split == MDS_NO_SPLITTABLE) {
 195                 /*
 196                  * this inode won't be splitted. so we need not to protect from
 197                  * just flush client's cache on modification.
 198                  */
 199                 if (mode == LCK_PW)
 200                         ret_mode = LCK_CW;
 201                 else
 202                         ret_mode = 0;
 203         } else {
 204                 if (mode == LCK_EX) {
 205                         ret_mode = LCK_EX;
 206                 } else if (mode == LCK_PR) {
 207                         ret_mode = LCK_CR;
 208                 } else if (mode == LCK_PW) {
 209                         /*
 210                          * caller gonna modify directory. We use concurrent
 211                          * write lock here to retract client's cache for
 212                          * readdir.
 213                          */
 214                         if (split == MDS_EXPECT_SPLIT) {
 215                                 /*
 216                                  * splitting possible. serialize any access the
 217                                  * idea is that first one seen dir is splittable
 218                                  * is given exclusive lock and split
 219                                  * directory. caller passes lock mode to
 220                                  * mds_try_to_split_dir() and splitting would be
 221                                  * done with exclusive lock only -bzzz.
 222                                  */
 223                                 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
 224                                        obd->obd_name,
 225                                        (unsigned long)dentry->d_inode->i_ino,
 226                                        (unsigned long)dentry->d_inode->i_generation);
 227                                 ret_mode = LCK_EX;
 228                         } else {
 229                                 ret_mode = LCK_CW;
 230                         }
 231                 }
 232         }
 233
 234         return ret_mode;
 235 }
 236
 237 /* only valid locked dentries or errors should be returned */
 238 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
 239                                     struct vfsmount **mnt, int lock_mode,
 240                                     struct lustre_handle *lockh, int *mode,
 241                                     char *name, int namelen, __u64 lockpart)
 242 {
 243         struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
 244         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 245         struct ldlm_res_id res_id = { .name = {0} };
 246         int flags = 0, rc;
 247         ENTRY;
 248
 249         if (IS_ERR(de))
 250                 RETURN(de);
 251
 252         lockh[1].cookie = 0;
 253         res_id.name[0] = id_fid(id);
 254         res_id.name[1] = id_group(id);
 255
 256 #ifdef S_PDIROPS
 257         if (name && IS_PDIROPS(de->d_inode)) {
 258                 ldlm_policy_data_t cpolicy =
 259                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 260                 LASSERT(mode != NULL);
 261                 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
 262                 if (*mode) {
 263                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 264                                               res_id, LDLM_IBITS,
 265                                               &cpolicy, *mode, &flags,
 266                                               mds_blocking_ast,
 267                                               ldlm_completion_ast, NULL, NULL,
 268                                               NULL, 0, NULL, lockh + 1);
 269                         if (rc != ELDLM_OK) {
 270                                 l_dput(de);
 271                                 RETURN(ERR_PTR(-ENOLCK));
 272                         }
 273                 }
 274                 flags = 0;
 275
 276                 res_id.name[2] = full_name_hash((unsigned char *)name, namelen);
 277
 278                 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
 279                        OLID4(id), res_id.name[2]);
 280         }
 281 #else
 282 #warning "No PDIROPS support in the kernel"
 283 #endif
 284         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 285                               LDLM_IBITS, &policy, lock_mode, &flags,
 286                               mds_blocking_ast, ldlm_completion_ast,
 287                               NULL, NULL, NULL, 0, NULL, lockh);
 288         if (rc != ELDLM_OK) {
 289                 l_dput(de);
 290                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 291 #ifdef S_PDIROPS
 292                 if (lockh[1].cookie)
 293                         ldlm_lock_decref(lockh + 1, *mode);
 294 #endif
 295         }
 296
 297         RETURN(retval);
 298 }
 299
 300 #ifndef DCACHE_DISCONNECTED
 301 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 302 #endif
 303
 304 /* Look up an entry by inode number. This function ONLY returns valid dget'd
 305  * dentries with an initialized inode or errors */
 306 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
 307                              struct vfsmount **mnt)
 308 {
 309         unsigned long ino = (unsigned long)id_ino(id);
 310         __u32 generation = (__u32)id_gen(id);
 311         struct mds_obd *mds = &obd->u.mds;
 312         struct dentry *result;
 313         struct inode *inode;
 314         char idname[32];
 315
 316         if (ino == 0)
 317                 RETURN(ERR_PTR(-ESTALE));
 318
 319         snprintf(idname, sizeof(idname), "0x%lx", ino);
 320
 321         CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
 322                ino, generation, mds->mds_sb);
 323
 324         /* under ext3 this is neither supposed to return bad inodes nor NULL
 325            inodes. */
 326         result = ll_lookup_one_len(idname, mds->mds_id_de,
 327                                    strlen(idname));
 328         if (IS_ERR(result))
 329                 RETURN(result);
 330
 331         inode = result->d_inode;
 332         if (!inode)
 333                 RETURN(ERR_PTR(-ENOENT));
 334
 335         if (is_bad_inode(inode)) {
 336                 CERROR("bad inode returned %lu/%u\n",
 337                        inode->i_ino, inode->i_generation);
 338                 dput(result);
 339                 RETURN(ERR_PTR(-ENOENT));
 340         }
 341
 342         /* here we disabled generation check, as root inode i_generation
 343          * of cache mds and real mds are different. */
 344         if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
 345             inode->i_generation != generation) {
 346                 /* we didn't find the right inode.. */
 347                 CERROR("bad inode %lu, link: %lu, ct: %d, generation %u/%u\n",
 348                        inode->i_ino, (unsigned long)inode->i_nlink,
 349                        atomic_read(&inode->i_count), inode->i_generation,
 350                        generation);
 351                 dput(result);
 352                 RETURN(ERR_PTR(-ENOENT));
 353         }
 354
 355         if (mnt) {
 356                 *mnt = mds->mds_vfsmnt;
 357                 mntget(*mnt);
 358         }
 359
 360         RETURN(result);
 361 }
 362
 363 static
 364 int mds_req_add_idmapping(struct ptlrpc_request *req,
 365                           struct mds_export_data *med)
 366 {
 367         struct mds_req_sec_desc *rsd;
 368         struct lustre_sec_desc  *lsd;
 369         int rc;
 370
 371         if (!med->med_remote)
 372                 return 0;
 373
 374         /* maybe we should do it more completely: invalidate the gss ctxt? */
 375         if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
 376                 CWARN("didn't find mapped uid\n");
 377                 return -EPERM;
 378         }
 379
 380         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 381         if (!rsd) {
 382                 CERROR("Can't unpack security desc\n");
 383                 return -EPROTO;
 384         }
 385
 386         lsd = mds_get_lsd(req->rq_mapped_uid);
 387         if (!lsd) {
 388                 CERROR("can't get LSD(%u), no mapping added\n",
 389                        req->rq_mapped_uid);
 390                 return -EPERM;
 391         }
 392
 393         rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 394                            rsd->rsd_gid, lsd->lsd_gid);
 395         mds_put_lsd(lsd);
 396         return rc;
 397 }
 398
 399 static
 400 int mds_req_del_idmapping(struct ptlrpc_request *req,
 401                           struct mds_export_data *med)
 402 {
 403         struct mds_req_sec_desc *rsd;
 404         struct lustre_sec_desc  *lsd;
 405         int rc;
 406
 407         if (!med->med_remote)
 408                 return 0;
 409
 410         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 411         if (!rsd) {
 412                 CERROR("Can't unpack security desc\n");
 413                 return -EPROTO;
 414         }
 415
 416         LASSERT(req->rq_mapped_uid != -1);
 417         lsd = mds_get_lsd(req->rq_mapped_uid);
 418         if (!lsd) {
 419                 CERROR("can't get LSD(%u), no idmapping deleted\n",
 420                        req->rq_mapped_uid);
 421                 return -EPERM;
 422         }
 423
 424         rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 425                            rsd->rsd_gid, lsd->lsd_gid);
 426         mds_put_lsd(lsd);
 427         return rc;
 428 }
 429
 430 static int mds_init_export_data(struct ptlrpc_request *req,
 431                                 struct mds_export_data *med)
 432 {
 433         struct obd_connect_data *data, *reply;
 434         int ask_remote, ask_local;
 435         ENTRY;
 436
 437         data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
 438         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
 439         LASSERT(data && reply);
 440
 441         if (med->med_initialized) {
 442                 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
 443                 goto reply;
 444         }
 445
 446         ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
 447         ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
 448
 449         /* currently the policy is simple: satisfy client as possible
 450          * as we can.
 451          */
 452         if (req->rq_auth_uid == -1) {
 453                 if (ask_remote)
 454                         CWARN("null sec is used, force to be local\n");
 455                 med->med_remote = 0;
 456         } else {
 457                 if (ask_remote) {
 458                         if (!req->rq_remote_realm)
 459                                 CWARN("local realm asked to be remote\n");
 460                         med->med_remote = 1;
 461                 } else if (ask_local) {
 462                         if (req->rq_remote_realm)
 463                                 CWARN("remote realm asked to be local\n");
 464                         med->med_remote = 0;
 465                 } else
 466                         med->med_remote = (req->rq_remote_realm != 0);
 467         }
 468
 469         med->med_nllu = data->ocd_nllu[0];
 470         med->med_nllg = data->ocd_nllu[1];
 471
 472         med->med_initialized = 1;
 473 reply:
 474         reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
 475         if (med->med_remote) {
 476                 if (!med->med_idmap)
 477                         med->med_idmap = mds_idmap_alloc();
 478
 479                 if (!med->med_idmap)
 480                         CERROR("Failed to alloc idmap, following request from "
 481                                "this client will be refused\n");
 482
 483                 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
 484                 CDEBUG(D_SEC, "set client as remote\n");
 485         } else {
 486                 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
 487                 CDEBUG(D_SEC, "set client as local\n");
 488         }
 489
 490         RETURN(0);
 491 }
 492
 493 static void mds_free_export_data(struct mds_export_data *med)
 494 {
 495         if (!med->med_idmap)
 496                 return;
 497
 498         LASSERT(med->med_remote);
 499         mds_idmap_free(med->med_idmap);
 500         med->med_idmap = NULL;
 501 }
 502
 503 /* Establish a connection to the MDS.
 504  *
 505  * This will set up an export structure for the client to hold state data about
 506  * that client, like open files, the last operation number it did on the server,
 507  * etc.
 508  */
 509 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 510                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 511                        unsigned long flags)
 512 {
 513         struct mds_export_data *med;
 514         struct mds_client_data *mcd;
 515         struct obd_export *exp;
 516         int rc;
 517         ENTRY;
 518
 519         if (!conn || !obd || !cluuid)
 520                 RETURN(-EINVAL);
 521
 522         /* XXX There is a small race between checking the list and adding a new
 523          * connection for the same UUID, but the real threat (list corruption
 524          * when multiple different clients connect) is solved.
 525          *
 526          * There is a second race between adding the export to the list, and
 527          * filling in the client data below.  Hence skipping the case of NULL
 528          * mcd above.  We should already be controlling multiple connects at the
 529          * client, and we can't hold the spinlock over memory allocations
 530          * without risk of deadlocking.
 531          */
 532         rc = class_connect(conn, obd, cluuid);
 533         if (rc)
 534                 RETURN(rc);
 535         exp = class_conn2export(conn);
 536
 537         LASSERT(exp != NULL);
 538         med = &exp->exp_mds_data;
 539
 540         OBD_ALLOC(mcd, sizeof(*mcd));
 541         if (!mcd) {
 542                 CERROR("%s: out of memory for client data.\n",
 543                         obd->obd_name);
 544                 GOTO(out, rc = -ENOMEM);
 545         }
 546
 547         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 548         med->med_mcd = mcd;
 549
 550         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 551         if (rc)
 552                 GOTO(out, rc);
 553
 554         EXIT;
 555 out:
 556         if (rc) {
 557                 if (mcd)
 558                         OBD_FREE(mcd, sizeof(*mcd));
 559                 class_disconnect(exp, 0);
 560         } else {
 561                 class_export_put(exp);
 562         }
 563         return rc;
 564 }
 565
 566 static int mds_connect_post(struct obd_export *exp, unsigned initial,
 567                             unsigned long flags)
 568 {
 569         struct obd_device *obd = exp->exp_obd;
 570         struct mds_obd *mds = &obd->u.mds;
 571         struct mds_export_data *med;
 572         struct mds_client_data *mcd;
 573         int rc = 0;
 574         ENTRY;
 575
 576         med = &exp->exp_mds_data;
 577         mcd = med->med_mcd;
 578
 579         if (initial) {
 580                 /* some one reconnect initially, we have to reset
 581                  * data existing export can have. bug 6102 */
 582                 if (mcd->mcd_last_xid != 0)
 583                         CDEBUG(D_HA, "initial reconnect to existing export\n");
 584                 mcd->mcd_last_transno = 0;
 585                 mcd->mcd_last_xid = 0;
 586                 mcd->mcd_last_result = 0;
 587                 mcd->mcd_last_data = 0;
 588         }
 589
 590         if (!(flags & OBD_OPT_MDS_CONNECTION)) {
 591                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 592                         atomic_inc(&mds->mds_real_clients);
 593                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 594                                obd->obd_name, exp->exp_client_uuid.uuid,
 595                                atomic_read(&mds->mds_real_clients));
 596                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 597                 }
 598                 if (mds->mds_md_name)
 599                         rc = mds_md_connect(obd, mds->mds_md_name);
 600         }
 601         RETURN(rc);
 602 }
 603
 604 static int mds_init_export(struct obd_export *exp)
 605 {
 606         struct mds_export_data *med = &exp->exp_mds_data;
 607
 608         INIT_LIST_HEAD(&med->med_open_head);
 609         spin_lock_init(&med->med_open_lock);
 610         return 0;
 611 }
 612
 613 static int mds_destroy_export(struct obd_export *export)
 614 {
 615         struct obd_device *obd = export->exp_obd;
 616         struct mds_export_data *med = &export->exp_mds_data;
 617         struct lvfs_run_ctxt saved;
 618         int rc = 0;
 619         ENTRY;
 620
 621         mds_free_export_data(med);
 622         target_destroy_export(export);
 623
 624         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 625                 GOTO(out, 0);
 626
 627         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 628
 629         /* Close any open files (which may also cause orphan unlinking). */
 630         spin_lock(&med->med_open_lock);
 631         while (!list_empty(&med->med_open_head)) {
 632                 struct list_head *tmp = med->med_open_head.next;
 633                 struct mds_file_data *mfd =
 634                         list_entry(tmp, struct mds_file_data, mfd_list);
 635                 struct lustre_id sid;
 636
 637                 BDEVNAME_DECLARE_STORAGE(btmp);
 638
 639                 /* bug 1579: fix force-closing for 2.5 */
 640                 struct dentry *dentry = mfd->mfd_dentry;
 641
 642                 list_del(&mfd->mfd_list);
 643                 spin_unlock(&med->med_open_lock);
 644
 645                 down(&dentry->d_inode->i_sem);
 646                 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
 647                 up(&dentry->d_inode->i_sem);
 648                 if (rc) {
 649                         CERROR("Can't read inode self id, inode %lu, "
 650                                "rc %d\n", dentry->d_inode->i_ino, rc);
 651                         memset(&sid, 0, sizeof(sid));
 652                 }
 653
 654                 /* If you change this message, be sure to update
 655                  * replay_single:test_46 */
 656                 CERROR("force closing client file handle for %.*s (%s:"
 657                        DLID4")\n", dentry->d_name.len, dentry->d_name.name,
 658                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 659                        OLID4(&sid));
 660
 661                 /* child inode->i_alloc_sem protects orphan_dec_test and
 662                  * is_orphan race, mds_mfd_close drops it */
 663                 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 664                 rc = mds_mfd_close(NULL, 0, obd, mfd,
 665                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 666                 if (rc)
 667                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 668                 spin_lock(&med->med_open_lock);
 669         }
 670         spin_unlock(&med->med_open_lock);
 671         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 672
 673         EXIT;
 674 out:
 675         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 676         return rc;
 677 }
 678
 679 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
 680 {
 681         unsigned long irqflags;
 682         struct obd_device *obd;
 683         struct mds_obd *mds;
 684         int rc;
 685         ENTRY;
 686
 687         LASSERT(exp != NULL);
 688         obd = class_exp2obd(exp);
 689         if (obd == NULL) {
 690                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 691                        exp->exp_handle.h_cookie);
 692                 RETURN(-EINVAL);
 693         }
 694         mds = &obd->u.mds;
 695
 696         /*
 697          * suppress any inter-mds requests durring disconnecting lmv if this is
 698          * detected --force mode. This is needed to avoid endless recovery.
 699          */
 700         if (atomic_read(&mds->mds_real_clients) > 0 &&
 701             !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
 702                 flags |= OBD_OPT_FORCE;
 703
 704         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 705             && !atomic_read(&mds->mds_real_clients)) {
 706                 /* there was no client at all */
 707                 mds_md_disconnect(obd, flags);
 708         }
 709
 710         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 711             && atomic_dec_and_test(&mds->mds_real_clients)) {
 712                 /* time to drop LMV connections */
 713                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 714                        "Disconnnect from LMV now\n",
 715                        obd->obd_name, exp->exp_client_uuid.uuid);
 716                 mds_md_disconnect(obd, flags);
 717         }
 718
 719         spin_lock_irqsave(&exp->exp_lock, irqflags);
 720         exp->exp_flags = flags;
 721         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 722
 723         /* disconnect early so that clients can't keep using export */
 724         rc = class_disconnect(exp, flags);
 725         ldlm_cancel_locks_for_export(exp);
 726
 727         /* complete all outstanding replies */
 728         spin_lock_irqsave(&exp->exp_lock, irqflags);
 729         while (!list_empty(&exp->exp_outstanding_replies)) {
 730                 struct ptlrpc_reply_state *rs =
 731                         list_entry(exp->exp_outstanding_replies.next,
 732                                    struct ptlrpc_reply_state, rs_exp_list);
 733                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 734
 735                 spin_lock(&svc->srv_lock);
 736                 list_del_init(&rs->rs_exp_list);
 737                 ptlrpc_schedule_difficult_reply(rs);
 738                 spin_unlock(&svc->srv_lock);
 739         }
 740         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 741         RETURN(rc);
 742 }
 743
 744 static int mds_getstatus(struct ptlrpc_request *req)
 745 {
 746         struct mds_obd *mds = mds_req2mds(req);
 747         struct mds_body *body;
 748         int rc, size;
 749         ENTRY;
 750
 751         size = sizeof(*body);
 752
 753         rc = lustre_pack_reply(req, 1, &size, NULL);
 754         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 755                 CERROR("mds: out of memory for message: size=%d\n", size);
 756                 req->rq_status = -ENOMEM;       /* superfluous? */
 757                 RETURN(-ENOMEM);
 758         }
 759
 760         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 761         body->valid |= OBD_MD_FID;
 762
 763         memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
 764
 765         /*
 766          * the last_committed and last_xid fields are filled in for all replies
 767          * already - no need to do so here also.
 768          */
 769         RETURN(0);
 770 }
 771
 772 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 773                      void *data, int flag)
 774 {
 775         int do_ast;
 776         ENTRY;
 777
 778         if (flag == LDLM_CB_CANCELING) {
 779                 /* Don't need to do anything here. */
 780                 RETURN(0);
 781         }
 782
 783         /* XXX layering violation!  -phil */
 784         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 785
 786         /*
 787          * get this: if mds_blocking_ast is racing with mds_intent_policy, such
 788          * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
 789          * then by the time we get the lock, we might not be the correct
 790          * blocking function anymore.  So check, and return early, if so.
 791          */
 792         if (lock->l_blocking_ast != mds_blocking_ast) {
 793                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 794                 RETURN(0);
 795         }
 796
 797         lock->l_flags |= LDLM_FL_CBPENDING;
 798         do_ast = (!lock->l_readers && !lock->l_writers);
 799         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 800
 801         if (do_ast) {
 802                 struct lustre_handle lockh;
 803                 int rc;
 804
 805                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 806                 ldlm_lock2handle(lock, &lockh);
 807                 rc = ldlm_cli_cancel(&lockh);
 808                 if (rc < 0)
 809                         CERROR("ldlm_cli_cancel: %d\n", rc);
 810         } else {
 811                 LDLM_DEBUG(lock, "Lock still has references, will be "
 812                            "cancelled later");
 813         }
 814         RETURN(0);
 815 }
 816
 817 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
 818                           void *md, int size, int mea)
 819 {
 820         int rc = size;
 821
 822         if (S_ISREG(inode->i_mode)) {
 823                 rc = mds_convert_lov_ea(obd, inode, md, size);
 824         } else if (S_ISDIR(inode->i_mode)) {
 825                 if (mea) {
 826                         rc = mds_convert_mea_ea(obd, inode, md, size);
 827                 } else {
 828                         rc = mds_convert_lov_ea(obd, inode, md, size);
 829                 }
 830                 if (rc == -EINVAL) {
 831                         CERROR("Invalid EA format (nor LOV or MEA) "
 832                                "is detected. Inode %lu/%u\n",
 833                                inode->i_ino, inode->i_generation);
 834                 }
 835         }
 836         return rc;
 837 }
 838
 839 int mds_get_md(struct obd_device *obd, struct inode *inode,
 840                void *md, int *size, int lock, int mea)
 841 {
 842         int lmm_size;
 843         int rc = 0;
 844         ENTRY;
 845
 846         if (lock)
 847                 down(&inode->i_sem);
 848
 849         rc = fsfilt_get_md(obd, inode, md, *size,
 850                            (mea ? EA_MEA : EA_LOV));
 851         if (rc < 0) {
 852                 CERROR("Error %d reading eadata for ino %lu\n",
 853                        rc, inode->i_ino);
 854         } else if (rc > 0) {
 855                 lmm_size = rc;
 856                 rc = mds_convert_md(obd, inode, md,
 857                                     lmm_size, mea);
 858                 if (rc == 0) {
 859                         *size = lmm_size;
 860                         rc = lmm_size;
 861                 } else if (rc > 0) {
 862                         *size = rc;
 863                 }
 864         }
 865         if (lock)
 866                 up(&inode->i_sem);
 867
 868         RETURN(rc);
 869 }
 870
 871
 872 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 873  * Call with lock=0 if the caller has already taken the i_sem. */
 874 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 875                 struct mds_body *body, struct inode *inode, int lock, int mea)
 876 {
 877         struct mds_obd *mds = &obd->u.mds;
 878         void *lmm;
 879         int lmm_size;
 880         int rc;
 881         ENTRY;
 882
 883         lmm = lustre_msg_buf(msg, offset, 0);
 884         if (lmm == NULL) {
 885                 /* Some problem with getting eadata when I sized the reply
 886                  * buffer... */
 887                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 888                        inode->i_ino);
 889                 RETURN(0);
 890         }
 891         lmm_size = msg->buflens[offset];
 892
 893         /* I don't really like this, but it is a sanity check on the client
 894          * MD request.  However, if the client doesn't know how much space
 895          * to reserve for the MD, it shouldn't be bad to have too much space.
 896          */
 897         if (lmm_size > mds->mds_max_mdsize) {
 898                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 899                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 900                 // RETURN(-EINVAL);
 901         }
 902
 903         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
 904         if (rc > 0) {
 905                 if (S_ISDIR(inode->i_mode))
 906                         body->valid |= OBD_MD_FLDIREA;
 907                 else
 908                         body->valid |= OBD_MD_FLEASIZE;
 909
 910                 if (mea)
 911                         body->valid |= OBD_MD_MEA;
 912
 913                 body->eadatasize = lmm_size;
 914                 rc = 0;
 915         }
 916
 917         RETURN(rc);
 918 }
 919 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
 920                   struct mds_body *repbody, int reply_off)
 921 {
 922         struct inode *inode = dentry->d_inode;
 923         char *symname;
 924         int len, rc;
 925         ENTRY;
 926
 927         symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 928         LASSERT(symname != NULL);
 929         len = req->rq_repmsg->buflens[reply_off + 1];
 930
 931         rc = inode->i_op->readlink(dentry, symname, len);
 932         if (rc < 0) {
 933                 CERROR("readlink failed: %d\n", rc);
 934         } else if (rc != len - 1) {
 935                 CERROR ("Unexpected readlink rc %d: expecting %d\n",
 936                         rc, len - 1);
 937                 rc = -EINVAL;
 938         } else {
 939                 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 940                 repbody->valid |= OBD_MD_LINKNAME;
 941                 repbody->eadatasize = rc + 1;
 942                 symname[rc] = 0;        /* NULL terminate */
 943                 rc = 0;
 944         }
 945
 946         RETURN(rc);
 947 }
 948
 949 int mds_pack_ea(struct dentry *dentry, struct ptlrpc_request *req,
 950                 struct mds_body *repbody, int req_off, int reply_off)
 951 {
 952         struct inode *inode = dentry->d_inode;
 953         char *ea_name;
 954         void *value = NULL;
 955         int len, rc;
 956         ENTRY;
 957
 958         ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
 959         len = req->rq_repmsg->buflens[reply_off + 1];
 960         if (len != 0)
 961                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 962
 963         rc = -EOPNOTSUPP;
 964         if (inode->i_op && inode->i_op->getxattr)
 965                 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
 966
 967         if (rc < 0) {
 968                 if (rc != -ENODATA && rc != -EOPNOTSUPP)
 969                         CERROR("getxattr failed: %d", rc);
 970         } else {
 971                 repbody->valid |= OBD_MD_FLEA;
 972                 repbody->eadatasize = rc;
 973                 rc = 0;
 974         }
 975
 976         RETURN(rc);
 977 }
 978
 979 int mds_pack_ealist(struct dentry *dentry, struct ptlrpc_request *req,
 980                     struct mds_body *repbody, int reply_off)
 981 {
 982         struct inode *inode = dentry->d_inode;
 983         void *value = NULL;
 984         int len, rc;
 985         ENTRY;
 986
 987         len = req->rq_repmsg->buflens[reply_off + 1];
 988         if (len != 0)
 989                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 990
 991         rc = -EOPNOTSUPP;
 992         if (inode->i_op && inode->i_op->getxattr)
 993                 rc = inode->i_op->listxattr(dentry, value, len);
 994
 995         if (rc < 0) {
 996                 CERROR("listxattr failed: %d", rc);
 997         } else {
 998                 repbody->valid |= OBD_MD_FLEALIST;
 999                 repbody->eadatasize = rc;
1000                 rc = 0;
1001         }
1002         RETURN(rc);
1003 }
1004
1005 int mds_pack_acl(struct obd_device *obd, struct lustre_msg *repmsg, int offset,
1006                  struct mds_body *body, struct inode *inode)
1007 {
1008         struct dentry de = { .d_inode = inode };
1009         __u32 buflen, *sizep;
1010         void *buf;
1011         int size;
1012         ENTRY;
1013
1014         if (!inode->i_op->getxattr)
1015                 RETURN(0);
1016
1017         buflen = repmsg->buflens[offset + 1];
1018         buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1019
1020         size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1021         if (size == -ENODATA || size == -EOPNOTSUPP)
1022                 RETURN(0);
1023         if (size < 0)
1024                 RETURN(size);
1025         LASSERT(size);
1026
1027         sizep = lustre_msg_buf(repmsg, offset, 4);
1028         if (!sizep) {
1029                 CERROR("can't locate returned acl size buf\n");
1030                 RETURN(-EPROTO);
1031         }
1032
1033         *sizep = cpu_to_le32(size);
1034         body->valid |= OBD_MD_FLACL_ACCESS;
1035
1036         RETURN(0);
1037 }
1038
1039 /*
1040  * here we take simple rule: once uid/fsuid is root, we also squash
1041  * the gid/fsgid, don't care setuid/setgid attributes.
1042  */
1043 int mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
1044                     ptl_nid_t *peernid)
1045 {
1046         if (!mds->mds_squash_uid || *peernid == mds->mds_nosquash_nid)
1047                 return 0;
1048
1049         if (rsd->rsd_uid && rsd->rsd_fsuid)
1050                 return 0;
1051
1052         CDEBUG(D_SEC, "squash req from "LPX64":"
1053                "(%u:%u-%u:%u/%x)=>(%u:%u-%u:%u/%x)\n", *peernid,
1054                 rsd->rsd_uid, rsd->rsd_gid,
1055                 rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
1056                 rsd->rsd_uid ? rsd->rsd_uid : mds->mds_squash_uid,
1057                 rsd->rsd_uid ? rsd->rsd_gid : mds->mds_squash_gid,
1058                 rsd->rsd_fsuid ? rsd->rsd_fsuid : mds->mds_squash_uid,
1059                 rsd->rsd_fsuid ? rsd->rsd_fsgid : mds->mds_squash_gid,
1060                 rsd->rsd_cap & ~CAP_FS_MASK);
1061
1062         if (rsd->rsd_uid == 0) {
1063                 rsd->rsd_uid = mds->mds_squash_uid;
1064                 rsd->rsd_gid = mds->mds_squash_gid;
1065         }
1066         if (rsd->rsd_fsuid == 0) {
1067                 rsd->rsd_fsuid = mds->mds_squash_uid;
1068                 rsd->rsd_fsgid = mds->mds_squash_gid;
1069         }
1070         rsd->rsd_cap &= ~CAP_FS_MASK;
1071
1072         return 1;
1073 }
1074
1075 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1076                                 struct ptlrpc_request *req, int req_off,
1077                                 struct mds_body *reqbody, int reply_off)
1078 {
1079         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1080         struct inode *inode = dentry->d_inode;
1081         struct mds_body *body;
1082         int rc = 0;
1083         ENTRY;
1084
1085         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1086                 RETURN(-ENOENT);
1087
1088         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1089         LASSERT(body != NULL);                 /* caller prepped reply */
1090
1091         if (dentry->d_flags & DCACHE_CROSS_REF) {
1092                 mds_pack_dentry2body(obd, body, dentry,
1093                                      (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1094                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1095                        OLID4(&body->id1));
1096                 RETURN(0);
1097         }
1098
1099         mds_pack_inode2body(obd, body, inode,
1100                             (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1101
1102         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1103             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1104
1105                 /* guessing what kind og attribute do we need. */
1106                 int is_mea = (S_ISDIR(inode->i_mode) &&
1107                     (reqbody->valid & OBD_MD_MEA) != 0);
1108
1109                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1110                                  body, inode, 1, is_mea);
1111
1112                 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1113                 if (!(body->valid & OBD_MD_FLEASIZE) &&
1114                     !(body->valid & OBD_MD_FLDIREA))
1115                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1116                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1117         } else if (S_ISLNK(inode->i_mode) &&
1118                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1119                 rc = mds_pack_link(dentry, req, body, reply_off);
1120         } else if (reqbody->valid & OBD_MD_FLEA) {
1121                 rc = mds_pack_ea(dentry, req, body, req_off, reply_off);
1122         } else if (reqbody->valid & OBD_MD_FLEALIST) {
1123                 rc = mds_pack_ealist(dentry, req, body, reply_off);
1124         }
1125
1126         if (reqbody->valid & OBD_MD_FLACL_ACCESS) {
1127                 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1128                 rc = mds_pack_acl(obd, req->rq_repmsg, reply_off + inc,
1129                                   body, inode);
1130         }
1131
1132         if (rc == 0)
1133                 mds_body_do_reverse_map(med, body);
1134
1135         RETURN(rc);
1136 }
1137
1138 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1139                                    struct dentry *dentry,
1140                                    int offset)
1141 {
1142         int rc = 0, size[1] = {sizeof(struct mds_body)};
1143         ENTRY;
1144
1145         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1146                 CERROR("failed MDS_GETATTR_PACK test\n");
1147                 req->rq_status = -ENOMEM;
1148                 RETURN(-ENOMEM);
1149         }
1150
1151         rc = lustre_pack_reply(req, 1, size, NULL);
1152         if (rc) {
1153                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1154                 GOTO(out, req->rq_status = rc);
1155         }
1156
1157         EXIT;
1158 out:
1159         return rc;
1160 }
1161
1162 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1163                                 int offset)
1164 {
1165         struct inode *inode = de->d_inode;
1166         struct mds_obd *mds = mds_req2mds(req);
1167         struct mds_body *body;
1168         int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1169         ENTRY;
1170
1171         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1172         LASSERT(body != NULL);                 /* checked by caller */
1173         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
1174
1175         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1176             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1177                 int rc;
1178
1179                 down(&inode->i_sem);
1180                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1181                                    ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1182                 up(&inode->i_sem);
1183                 if (rc < 0) {
1184                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1185                                 CERROR("error getting inode %lu MD: rc = %d\n",
1186                                        inode->i_ino, rc);
1187                         size[bufcount] = 0;
1188                 } else if (rc > mds->mds_max_mdsize) {
1189                         size[bufcount] = 0;
1190                         CERROR("MD size %d larger than maximum possible %u\n",
1191                                rc, mds->mds_max_mdsize);
1192                 } else {
1193                         size[bufcount] = rc;
1194                 }
1195                 bufcount++;
1196         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1197                 if (inode->i_size + 1 != body->eadatasize)
1198                         CERROR("symlink size: %Lu, reply space: %d\n",
1199                                inode->i_size + 1, body->eadatasize);
1200                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1201                 bufcount++;
1202                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1203                        inode->i_size + 1, body->eadatasize);
1204         } else if ((body->valid & OBD_MD_FLEA)) {
1205                 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1206                                                   offset + 1, 0);
1207                 rc = -EOPNOTSUPP;
1208                 if (inode->i_op && inode->i_op->getxattr)
1209                         rc = inode->i_op->getxattr(de, ea_name, NULL, 0);
1210
1211                 if (rc < 0) {
1212                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1213                                 CERROR("error getting inode %lu EA: rc = %d\n",
1214                                        inode->i_ino, rc);
1215                         size[bufcount] = 0;
1216                 } else {
1217                         size[bufcount] = min_t(int, body->eadatasize, rc);
1218                 }
1219                 bufcount++;
1220         } else if (body->valid & OBD_MD_FLEALIST) {
1221                 rc = -EOPNOTSUPP;
1222                 if (inode->i_op && inode->i_op->getxattr)
1223                         rc = inode->i_op->listxattr(de, NULL, 0);
1224
1225                 if (rc < 0) {
1226                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1227                                 CERROR("error getting inode %lu EA: rc = %d\n",
1228                                        inode->i_ino, rc);
1229                         size[bufcount] = 0;
1230                 } else {
1231                         size[bufcount] = min_t(int, body->eadatasize, rc);
1232                 }
1233                 bufcount++;
1234         }
1235
1236         /* may co-exist with OBD_MD_FLEASIZE */
1237         if (body->valid & OBD_MD_FLACL_ACCESS) {
1238                 size[bufcount++] = 4;
1239                 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1240         }
1241
1242         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1243                 CERROR("failed MDS_GETATTR_PACK test\n");
1244                 req->rq_status = -ENOMEM;
1245                 GOTO(out, rc = -ENOMEM);
1246         }
1247
1248         rc = lustre_pack_reply(req, bufcount, size, NULL);
1249         if (rc) {
1250                 CERROR("out of memory\n");
1251                 GOTO(out, req->rq_status = rc);
1252         }
1253
1254         EXIT;
1255  out:
1256         return rc;
1257 }
1258
1259 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1260                       char *name, int namelen)
1261 {
1262         struct mea *mea = NULL;
1263         int mea_size, rc = 0;
1264         ENTRY;
1265
1266         rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1267         if (rc)
1268                 RETURN(rc);
1269         if (mea != NULL) {
1270                 /*
1271                  * dir is already splitted, check if requested filename should
1272                  * live at this MDS or at another one.
1273                  */
1274                 int i = mea_name2idx(mea, name, namelen - 1);
1275                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1276                         CDEBUG(D_OTHER,
1277                                "inapropriate MDS(%d) for %s. should be "
1278                                "%lu(%d)\n", mea->mea_master, name,
1279                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1280                         rc = -ERESTART;
1281                 }
1282         }
1283
1284         if (mea)
1285                 OBD_FREE(mea, mea_size);
1286         RETURN(rc);
1287 }
1288
1289 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1290                             struct lustre_handle *child_lockh, int child_part)
1291 {
1292         struct obd_device *obd = req->rq_export->exp_obd;
1293         struct mds_obd *mds = &obd->u.mds;
1294         struct ldlm_reply *rep = NULL;
1295         struct lvfs_run_ctxt saved;
1296         struct mds_req_sec_desc *rsd;
1297         struct mds_body *body;
1298         struct dentry *dparent = NULL, *dchild = NULL;
1299         struct lvfs_ucred uc = {NULL, NULL,};
1300         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1301         unsigned int namesize;
1302         int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1303         char *name = NULL;
1304         ENTRY;
1305
1306         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
1307         MD_COUNTER_INCREMENT(obd, getattr_lock);
1308
1309         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1310         if (!rsd) {
1311                 CERROR("Can't unpack security desc\n");
1312                 RETURN(-EFAULT);
1313         }
1314
1315         /* swab now, before anyone looks inside the request. */
1316         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1317                                   lustre_swab_mds_body);
1318         if (body == NULL) {
1319                 CERROR("Can't swab mds_body\n");
1320                 GOTO(cleanup, rc = -EFAULT);
1321         }
1322
1323         LASSERT_REQSWAB(req, offset + 1);
1324         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1325         if (name == NULL) {
1326                 CERROR("Can't unpack name\n");
1327                 GOTO(cleanup, rc = -EFAULT);
1328         }
1329         namesize = req->rq_reqmsg->buflens[offset + 1];
1330
1331         /* namesize less than 2 means we have empty name, probably came from
1332            revalidate by cfid, so no point in having name to be set */
1333         if (namesize <= 1)
1334                 name = NULL;
1335
1336         LASSERT (offset == 1 || offset == 3);
1337         if (offset == 3) {
1338                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1339                 reply_offset = 1;
1340         } else {
1341                 reply_offset = 0;
1342         }
1343
1344         rc = mds_init_ucred(&uc, req, rsd);
1345         if (rc) {
1346                 GOTO(cleanup, rc);
1347         }
1348
1349         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1350         cleanup_phase = 1; /* kernel context */
1351         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1352
1353         LASSERT(namesize > 0);
1354         if (child_lockh->cookie != 0) {
1355                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1356                 resent_req = 1;
1357         }
1358 #if HAVE_LOOKUP_RAW
1359         if (body->valid == OBD_MD_FLID) {
1360                 struct mds_body *mds_reply;
1361                 int size = sizeof(*mds_reply);
1362                 struct inode *dir;
1363                 ino_t inum;
1364
1365                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1366                 if (IS_ERR(dparent)) {
1367                         rc = PTR_ERR(dparent);
1368                         GOTO(cleanup, rc);
1369                 }
1370                 /*
1371                  * the user requested ONLY the inode number, so do a raw lookup.
1372                  */
1373                 rc = lustre_pack_reply(req, 1, &size, NULL);
1374                 if (rc) {
1375                         CERROR("out of memory\n");
1376                         l_dput(dparent);
1377                         GOTO(cleanup, rc);
1378                 }
1379                 dir  = dparent->d_inode;
1380                 LASSERT(dir->i_op->lookup_raw != NULL);
1381                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1382                 l_dput(dparent);
1383                 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1384                                            sizeof(*mds_reply));
1385
1386                 id_ino(&mds_reply->id1) = inum;
1387                 mds_reply->valid = OBD_MD_FLID;
1388                 GOTO(cleanup, rc);
1389         }
1390 #endif
1391         if (resent_req == 0) {
1392                 LASSERT(id_fid(&body->id1) != 0);
1393                 if (name) {
1394                         rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1395                                                          parent_lockh, &dparent,
1396                                                          LCK_PR,
1397                                                          MDS_INODELOCK_UPDATE,
1398                                                          &update_mode,
1399                                                          name, namesize,
1400                                                          child_lockh, &dchild,
1401                                                          LCK_PR, child_part);
1402                         if (rc)
1403                                 GOTO(cleanup, rc);
1404
1405                         cleanup_phase = 2; /* dchild, dparent, locks */
1406
1407                         /*
1408                          * let's make sure this name should leave on this mds
1409                          * node.
1410                          */
1411                         rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1412                         if (rc)
1413                                 GOTO(cleanup, rc);
1414                 } else {
1415                         /* we have no dentry here, drop LOOKUP bit */
1416                         /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1417                         child_part &= ~MDS_INODELOCK_LOOKUP;
1418                         CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1419                                obd->obd_name, OLID4(&body->id1));
1420
1421                         dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1422                                                       LCK_PR, parent_lockh,
1423                                                       &update_mode,
1424                                                       NULL, 0,
1425                                                       MDS_INODELOCK_UPDATE);
1426                         if (IS_ERR(dchild)) {
1427                                 CERROR("can't find inode with id "DLID4", err = %d\n",
1428                                        OLID4(&body->id1), (int)PTR_ERR(dchild));
1429                                 GOTO(cleanup, rc = PTR_ERR(dchild));
1430                         }
1431                         memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1432                 }
1433         } else {
1434                 struct ldlm_lock *granted_lock;
1435
1436                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1437                 granted_lock = ldlm_handle2lock(child_lockh);
1438
1439                 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1440                          id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1441                          child_lockh->cookie);
1442
1443                 if (name) {
1444                         /* usual named request */
1445                         dparent = mds_id2dentry(obd, &body->id1, NULL);
1446                         LASSERT(!IS_ERR(dparent));
1447                         dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1448                         LASSERT(!IS_ERR(dchild));
1449                 } else {
1450                         /* client wants to get attr. by id */
1451                         dchild = mds_id2dentry(obd, &body->id1, NULL);
1452                         LASSERT(!IS_ERR(dchild));
1453                 }
1454                 LDLM_LOCK_PUT(granted_lock);
1455         }
1456
1457         cleanup_phase = 2; /* dchild, dparent, locks */
1458
1459         if (!DENTRY_VALID(dchild)) {
1460                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1461                 /*
1462                  * in the intent case, the policy clears this error: the
1463                  * disposition is enough.
1464                  */
1465                 rc = -ENOENT;
1466                 GOTO(cleanup, rc);
1467         } else {
1468                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1469         }
1470
1471         if (req->rq_repmsg == NULL) {
1472                 if (dchild->d_flags & DCACHE_CROSS_REF)
1473                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1474                 else
1475                         rc = mds_getattr_pack_msg(req, dchild, offset);
1476                 if (rc != 0) {
1477                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1478                         GOTO (cleanup, rc);
1479                 }
1480         }
1481
1482         rc = mds_getattr_internal(obd, dchild, req, offset, body, reply_offset);
1483         GOTO(cleanup, rc); /* returns the lock to the client */
1484
1485  cleanup:
1486         switch (cleanup_phase) {
1487         case 2:
1488                 if (resent_req == 0) {
1489                         if (rc && DENTRY_VALID(dchild))
1490                                 ldlm_lock_decref(child_lockh, LCK_PR);
1491                         if (name)
1492                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1493 #ifdef S_PDIROPS
1494                         if (parent_lockh[1].cookie != 0)
1495                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1496 #endif
1497                         if (dparent)
1498                                 l_dput(dparent);
1499                 }
1500                 l_dput(dchild);
1501         case 1:
1502                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1503         default:
1504                 mds_exit_ucred(&uc);
1505         }
1506         return rc;
1507 }
1508
1509 static int mds_getattr(struct ptlrpc_request *req, int offset)
1510 {
1511         struct obd_device *obd = req->rq_export->exp_obd;
1512         struct lvfs_run_ctxt saved;
1513         struct dentry *de;
1514         struct mds_req_sec_desc *rsd;
1515         struct mds_body *body;
1516         struct lvfs_ucred uc = {NULL, NULL,};
1517         int rc = 0;
1518         ENTRY;
1519
1520         MD_COUNTER_INCREMENT(obd, getattr);
1521
1522         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1523         if (!rsd) {
1524                 CERROR("Can't unpack security desc\n");
1525                 RETURN(-EFAULT);
1526         }
1527
1528         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1529                                   lustre_swab_mds_body);
1530         if (body == NULL) {
1531                 CERROR ("Can't unpack body\n");
1532                 RETURN (-EFAULT);
1533         }
1534
1535         rc = mds_init_ucred(&uc, req, rsd);
1536         if (rc) {
1537                 mds_exit_ucred(&uc);
1538                 RETURN(rc);
1539         }
1540
1541         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1542         de = mds_id2dentry(obd, &body->id1, NULL);
1543         if (IS_ERR(de)) {
1544                 rc = req->rq_status = PTR_ERR(de);
1545                 GOTO(out_pop, rc);
1546         }
1547
1548         rc = mds_getattr_pack_msg(req, de, offset);
1549         if (rc != 0) {
1550                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1551                 GOTO(out_pop, rc);
1552         }
1553
1554         req->rq_status = mds_getattr_internal(obd, de, req, offset, body, 0);
1555         l_dput(de);
1556
1557         EXIT;
1558 out_pop:
1559         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1560         mds_exit_ucred(&uc);
1561         return rc;
1562 }
1563
1564 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1565                           unsigned long max_age)
1566 {
1567         int rc;
1568         ENTRY;
1569
1570         spin_lock(&obd->obd_osfs_lock);
1571         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1572         if (rc == 0)
1573                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1574         spin_unlock(&obd->obd_osfs_lock);
1575
1576         RETURN(rc);
1577 }
1578
1579 static int mds_statfs(struct ptlrpc_request *req)
1580 {
1581         struct obd_device *obd = req->rq_export->exp_obd;
1582         int rc, size = sizeof(struct obd_statfs);
1583         ENTRY;
1584
1585         /* This will trigger a watchdog timeout */
1586         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1587                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1588
1589         rc = lustre_pack_reply(req, 1, &size, NULL);
1590         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1591                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1592                 GOTO(out, rc);
1593         }
1594
1595         OBD_COUNTER_INCREMENT(obd, statfs);
1596
1597         /* We call this so that we can cache a bit - 1 jiffie worth */
1598         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1599                             jiffies - HZ);
1600         if (rc) {
1601                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1602                 GOTO(out, rc);
1603         }
1604
1605         EXIT;
1606 out:
1607         req->rq_status = rc;
1608         return rc;
1609 }
1610
1611 static int mds_sync(struct ptlrpc_request *req, int offset)
1612 {
1613         struct obd_device *obd = req->rq_export->exp_obd;
1614         struct mds_obd *mds = &obd->u.mds;
1615         struct mds_body *body;
1616         int rc, size = sizeof(*body);
1617         ENTRY;
1618
1619         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1620                                   lustre_swab_mds_body);
1621         if (body == NULL)
1622                 GOTO(out, rc = -EPROTO);
1623
1624         rc = lustre_pack_reply(req, 1, &size, NULL);
1625         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1626                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1627                 GOTO(out, rc);
1628         }
1629
1630         if (id_ino(&body->id1) == 0) {
1631                 /* an id of zero is taken to mean "sync whole filesystem" */
1632                 rc = fsfilt_sync(obd, mds->mds_sb);
1633                 if (rc)
1634                         GOTO(out, rc);
1635         } else {
1636                 /* just any file to grab fsync method - "file" arg unused */
1637                 struct file *file = mds->mds_rcvd_filp;
1638                 struct mds_body *rep_body;
1639                 struct dentry *de;
1640
1641                 de = mds_id2dentry(obd, &body->id1, NULL);
1642                 if (IS_ERR(de))
1643                         GOTO(out, rc = PTR_ERR(de));
1644
1645                 rc = file->f_op->fsync(NULL, de, 1);
1646                 if (rc)
1647                         GOTO(out, rc);
1648
1649                 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1650                 mds_pack_inode2body(obd, rep_body, de->d_inode,
1651                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1652                 l_dput(de);
1653         }
1654
1655         EXIT;
1656 out:
1657         req->rq_status = rc;
1658         return rc;
1659 }
1660
1661 /* mds_readpage does not take a DLM lock on the inode, because the client must
1662  * already have a PR lock.
1663  *
1664  * If we were to take another one here, a deadlock will result, if another
1665  * thread is already waiting for a PW lock. */
1666 static int mds_readpage(struct ptlrpc_request *req, int offset)
1667 {
1668         struct obd_device *obd = req->rq_export->exp_obd;
1669         struct vfsmount *mnt;
1670         struct dentry *de;
1671         struct file *file;
1672         struct mds_req_sec_desc *rsd;
1673         struct mds_body *body, *repbody;
1674         struct lvfs_run_ctxt saved;
1675         int rc, size = sizeof(*repbody);
1676         struct lvfs_ucred uc = {NULL, NULL,};
1677         ENTRY;
1678
1679         rc = lustre_pack_reply(req, 1, &size, NULL);
1680         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1681                 CERROR("mds: out of memory\n");
1682                 GOTO(out, rc = -ENOMEM);
1683         }
1684
1685         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1686         if (!rsd) {
1687                 CERROR("Can't unpack security desc\n");
1688                 GOTO (out, rc = -EFAULT);
1689         }
1690
1691         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1692                                   lustre_swab_mds_body);
1693         if (body == NULL) {
1694                 CERROR("Can't unpack body\n");
1695                 GOTO (out, rc = -EFAULT);
1696         }
1697
1698         rc = mds_init_ucred(&uc, req, rsd);
1699         if (rc) {
1700                 GOTO(out, rc);
1701         }
1702
1703         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1704         de = mds_id2dentry(obd, &body->id1, &mnt);
1705         if (IS_ERR(de))
1706                 GOTO(out_pop, rc = PTR_ERR(de));
1707
1708         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1709
1710         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1711         /* note: in case of an error, dentry_open puts dentry */
1712         if (IS_ERR(file))
1713                 GOTO(out_pop, rc = PTR_ERR(file));
1714
1715         /* body->size is actually the offset -eeb */
1716         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1717                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1718                        body->size, de->d_inode->i_blksize);
1719                 GOTO(out_file, rc = -EFAULT);
1720         }
1721
1722         /* body->nlink is actually the #bytes to read -eeb */
1723         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1724                 CERROR("size %u is not multiple of blocksize %lu\n",
1725                        body->nlink, de->d_inode->i_blksize);
1726                 GOTO(out_file, rc = -EFAULT);
1727         }
1728
1729         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1730         repbody->size = file->f_dentry->d_inode->i_size;
1731         repbody->valid = OBD_MD_FLSIZE;
1732
1733         /* to make this asynchronous make sure that the handling function
1734            doesn't send a reply when this function completes. Instead a
1735            callback function would send the reply */
1736         /* body->size is actually the offset -eeb */
1737         rc = mds_sendpage(req, file, body->size, body->nlink);
1738
1739         EXIT;
1740 out_file:
1741         filp_close(file, 0);
1742 out_pop:
1743         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1744 out:
1745         mds_exit_ucred(&uc);
1746         req->rq_status = rc;
1747         return 0;
1748 }
1749
1750 /* update master MDS ID, which is stored in local inode EA. */
1751 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1752                    void *data, int data_len)
1753 {
1754         struct mds_obd *mds = &obd->u.mds;
1755         struct dentry *dentry;
1756         void *handle;
1757         int rc = 0;
1758         ENTRY;
1759
1760         LASSERT(id);
1761         LASSERT(obd);
1762
1763         dentry = mds_id2dentry(obd, id, NULL);
1764         if (IS_ERR(dentry))
1765                 GOTO(out, rc = PTR_ERR(dentry));
1766
1767         if (!dentry->d_inode) {
1768                 CERROR("Can't find object "DLID4".\n",
1769                        OLID4(id));
1770                 GOTO(out_dentry, rc = -EINVAL);
1771         }
1772
1773         handle = fsfilt_start(obd, dentry->d_inode,
1774                               FSFILT_OP_SETATTR, NULL);
1775         if (IS_ERR(handle))
1776                 GOTO(out_dentry, rc = PTR_ERR(handle));
1777
1778         rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1779                                   (struct lustre_id *)data);
1780         if (rc) {
1781                 CERROR("Can't update inode "DLID4" master id, "
1782                        "error = %d.\n", OLID4(id), rc);
1783                 GOTO(out_commit, rc);
1784         }
1785
1786         EXIT;
1787 out_commit:
1788         fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1789                       handle, 0);
1790 out_dentry:
1791         l_dput(dentry);
1792 out:
1793         return rc;
1794 }
1795 EXPORT_SYMBOL(mds_update_mid);
1796
1797 /* read master MDS ID, which is stored in local inode EA. */
1798 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1799                  void *data, int data_len)
1800 {
1801         struct dentry *dentry;
1802         int rc = 0;
1803         ENTRY;
1804
1805         LASSERT(id);
1806         LASSERT(obd);
1807
1808         dentry = mds_id2dentry(obd, id, NULL);
1809         if (IS_ERR(dentry))
1810                 GOTO(out, rc = PTR_ERR(dentry));
1811
1812         if (!dentry->d_inode) {
1813                 CERROR("Can't find object "DLID4".\n",
1814                        OLID4(id));
1815                 GOTO(out_dentry, rc = -EINVAL);
1816         }
1817
1818         down(&dentry->d_inode->i_sem);
1819         rc = mds_read_inode_mid(obd, dentry->d_inode,
1820                                 (struct lustre_id *)data);
1821         up(&dentry->d_inode->i_sem);
1822         if (rc) {
1823                 CERROR("Can't read inode "DLID4" master id, "
1824                        "error = %d.\n", OLID4(id), rc);
1825                 GOTO(out_dentry, rc);
1826         }
1827
1828         EXIT;
1829 out_dentry:
1830         l_dput(dentry);
1831 out:
1832         return rc;
1833 }
1834 EXPORT_SYMBOL(mds_read_mid);
1835
1836 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
1837                 char **data, int *datalen)
1838 {
1839         struct dentry *dentry;
1840         struct mds_obd *mds = &obd->u.mds;
1841         int rc = 0, mea = 0;
1842         char *ea;
1843         ENTRY;
1844
1845         LASSERT(id);
1846         LASSERT(obd);
1847
1848         dentry = mds_id2dentry(obd, id, NULL);
1849         if (IS_ERR(dentry))
1850                 GOTO(out, rc = PTR_ERR(dentry));
1851
1852         if (!dentry->d_inode) {
1853                 CERROR("Can't find object "DLID4".\n",
1854                        OLID4(id));
1855                 GOTO(out_dentry, rc = -EINVAL);
1856         }
1857         if (S_ISDIR(dentry->d_inode->i_mode)) {
1858                 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
1859                 mea = 1;
1860         } else {
1861                 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
1862                 mea = 0;
1863         }
1864         OBD_ALLOC(ea, *datalen);
1865         if (!ea) {
1866                 *datalen = 0;
1867                 GOTO(out_dentry, rc = PTR_ERR(dentry));
1868         }
1869         *data = ea;
1870         down(&dentry->d_inode->i_sem);
1871         rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
1872                            (mea ? EA_MEA : EA_LOV));
1873         up(&dentry->d_inode->i_sem);
1874
1875         if (rc < 0)
1876                 CERROR("Error %d reading eadata for ino %lu\n",
1877                         rc, dentry->d_inode->i_ino);
1878 out_dentry:
1879         l_dput(dentry);
1880 out:
1881         RETURN(rc);
1882 }
1883 EXPORT_SYMBOL(mds_read_md);
1884
1885 int mds_reint(struct ptlrpc_request *req, int offset,
1886               struct lustre_handle *lockh)
1887 {
1888         struct mds_update_record *rec;
1889         struct mds_req_sec_desc *rsd;
1890         int rc;
1891         ENTRY;
1892
1893         OBD_ALLOC(rec, sizeof(*rec));
1894         if (rec == NULL)
1895                 RETURN(-ENOMEM);
1896
1897         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1898         if (!rsd) {
1899                 CERROR("Can't unpack security desc\n");
1900                 GOTO(out, rc = -EFAULT);
1901         }
1902
1903         rc = mds_update_unpack(req, offset, rec);
1904         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1905                 CERROR("invalid record\n");
1906                 GOTO(out, req->rq_status = -EINVAL);
1907         }
1908
1909         rc = mds_init_ucred(&rec->ur_uc, req, rsd);
1910         if (rc) {
1911                 GOTO(out, rc);
1912         }
1913
1914         /* rc will be used to interrupt a for loop over multiple records */
1915         rc = mds_reint_rec(rec, offset, req, lockh);
1916
1917  out:
1918         mds_exit_ucred(&rec->ur_uc);
1919         OBD_FREE(rec, sizeof(*rec));
1920         RETURN(rc);
1921 }
1922
1923 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1924                                        struct obd_device *obd, int *process)
1925 {
1926         switch (req->rq_reqmsg->opc) {
1927         case MDS_CONNECT: /* This will never get here, but for completeness. */
1928         case OST_CONNECT: /* This will never get here, but for completeness. */
1929         case MDS_DISCONNECT:
1930         case OST_DISCONNECT:
1931                *process = 1;
1932                RETURN(0);
1933
1934         case MDS_CLOSE:
1935         case MDS_SYNC: /* used in unmounting */
1936         case OBD_PING:
1937         case MDS_REINT:
1938         case LDLM_ENQUEUE:
1939         case OST_CREATE:
1940                 *process = target_queue_recovery_request(req, obd);
1941                 RETURN(0);
1942
1943         default:
1944                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1945                 *process = 0;
1946                 /* XXX what should we set rq_status to here? */
1947                 req->rq_status = -EAGAIN;
1948                 RETURN(ptlrpc_error(req));
1949         }
1950 }
1951
1952 static char *reint_names[] = {
1953         [REINT_SETATTR] "setattr",
1954         [REINT_CREATE]  "create",
1955         [REINT_LINK]    "link",
1956         [REINT_UNLINK]  "unlink",
1957         [REINT_RENAME]  "rename",
1958         [REINT_OPEN]    "open",
1959 };
1960
1961 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
1962                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1963                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1964                             OBD_MD_FLID)
1965
1966 static void reconstruct_create(struct ptlrpc_request *req)
1967 {
1968         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1969         struct mds_client_data *mcd = med->med_mcd;
1970         struct dentry *dentry;
1971         struct ost_body *body;
1972         struct lustre_id id;
1973         int rc;
1974         ENTRY;
1975
1976         /* copy rc, transno and disp; steal locks */
1977         mds_req_from_mcd(req, mcd);
1978         if (req->rq_status) {
1979                 EXIT;
1980                 return;
1981         }
1982
1983         id_gen(&id) = 0;
1984         id_group(&id) = 0;
1985
1986         id_ino(&id) = mcd->mcd_last_data;
1987         LASSERT(id_ino(&id) != 0);
1988
1989         dentry = mds_id2dentry(req2obd(req), &id, NULL);
1990         if (IS_ERR(dentry)) {
1991                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1992                 req->rq_status = PTR_ERR(dentry);
1993                 EXIT;
1994                 return;
1995         }
1996
1997         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1998               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1999               dentry->d_inode->i_generation);
2000
2001         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2002         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2003         body->oa.o_id = dentry->d_inode->i_ino;
2004         body->oa.o_generation = dentry->d_inode->i_generation;
2005         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2006
2007         down(&dentry->d_inode->i_sem);
2008         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2009         up(&dentry->d_inode->i_sem);
2010         if (rc) {
2011                 CERROR("Can't read inode self id, inode %lu, "
2012                        "rc %d\n", dentry->d_inode->i_ino, rc);
2013                 id_fid(&id) = 0;
2014         }
2015
2016         body->oa.o_fid = id_fid(&id);
2017         body->oa.o_mds = id_group(&id);
2018         l_dput(dentry);
2019
2020         EXIT;
2021 }
2022
2023 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2024                               struct dentry *de, void *xattr, int xattr_size)
2025 {
2026         struct inode *inode = de->d_inode;
2027         struct posix_acl *acl;
2028         mode_t mode;
2029         int rc = 0;
2030
2031         LASSERT(handle);
2032         LASSERT(inode);
2033         LASSERT(xattr);
2034         LASSERT(xattr_size > 0);
2035
2036         if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2037                 CERROR("backend fs dosen't support xattr\n");
2038                 return -EOPNOTSUPP;
2039         }
2040
2041         /* set default acl */
2042         if (S_ISDIR(inode->i_mode)) {
2043                 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2044                                            xattr, xattr_size, 0);
2045                 if (rc) {
2046                         CERROR("set default acl err: %d\n", rc);
2047                         return rc;
2048                 }
2049         }
2050
2051         /* set access acl */
2052         acl = posix_acl_from_xattr(xattr, xattr_size);
2053         if (acl == NULL || IS_ERR(acl)) {
2054                 CERROR("insane attr data\n");
2055                 return PTR_ERR(acl);
2056         }
2057
2058         if (posix_acl_valid(acl)) {
2059                 CERROR("default acl not valid: %d\n", rc);
2060                 rc = -EFAULT;
2061                 goto out;
2062         }
2063
2064         mode = inode->i_mode;
2065         rc = posix_acl_create_masq(acl, &mode);
2066         if (rc < 0) {
2067                 CERROR("create masq err %d\n", rc);
2068                 goto out;
2069         }
2070
2071         if (inode->i_mode != mode) {
2072                 struct iattr iattr = { .ia_valid = ATTR_MODE,
2073                                        .ia_mode = mode };
2074                 int rc2;
2075
2076                 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2077                 if (rc2) {
2078                         CERROR("setattr mode err: %d\n", rc2);
2079                         rc = rc2;
2080                         goto out;
2081                 }
2082         }
2083
2084         if (rc > 0) {
2085                 /* we didn't change acl except mode bits of some
2086                  * entries, so should be fit into original size.
2087                  */
2088                 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2089                 LASSERT(rc > 0);
2090
2091                 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2092                                            xattr, xattr_size, 0);
2093                 if (rc)
2094                         CERROR("set access acl err: %d\n", rc);
2095         }
2096 out:
2097         posix_acl_release(acl);
2098         return rc;
2099 }
2100
2101 static int mdt_obj_create(struct ptlrpc_request *req)
2102 {
2103         struct obd_device *obd = req->rq_export->exp_obd;
2104         struct mds_obd *mds = &obd->u.mds;
2105         struct ost_body *body, *repbody;
2106         void *acl = NULL;
2107         int acl_size;
2108         char idname[LL_ID_NAMELEN];
2109         int size = sizeof(*repbody);
2110         struct inode *parent_inode;
2111         struct lvfs_run_ctxt saved;
2112         int rc, cleanup_phase = 0;
2113         struct dentry *new = NULL;
2114         struct dentry_params dp;
2115         int mealen, flags = 0;
2116         struct lvfs_ucred uc;
2117         struct lustre_id id;
2118         struct mea *mea;
2119         void *handle = NULL;
2120         unsigned long cr_inum = 0;
2121         ENTRY;
2122
2123         DEBUG_REQ(D_HA, req, "create remote object");
2124         parent_inode = mds->mds_unnamed_dir->d_inode;
2125
2126         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2127                                   lustre_swab_ost_body);
2128         if (body == NULL)
2129                 RETURN(-EFAULT);
2130
2131         /* acl data is packed transparently, no swab here */
2132         LASSERT(req->rq_reqmsg->bufcount >= 2);
2133         acl_size = req->rq_reqmsg->buflens[1];
2134         if (acl_size) {
2135                 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2136                 if (!acl) {
2137                         CERROR("No default acl buf?\n");
2138                         RETURN(-EFAULT);
2139                 }
2140         }
2141
2142         rc = lustre_pack_reply(req, 1, &size, NULL);
2143         if (rc)
2144                 RETURN(rc);
2145
2146         MDS_CHECK_RESENT(req, reconstruct_create(req));
2147
2148         uc.luc_lsd = NULL;
2149         uc.luc_ginfo = NULL;
2150         uc.luc_uid = body->oa.o_uid;
2151         uc.luc_gid = body->oa.o_gid;
2152         uc.luc_fsuid = body->oa.o_uid;
2153         uc.luc_fsgid = body->oa.o_gid;
2154
2155         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2156         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2157
2158         /* in REPLAY case inum should be given (client or other MDS fills it) */
2159         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2160             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2161                 /*
2162                  * this is re-create request from MDS holding directory name.
2163                  * we have to lookup given ino/gen first. if it exists (good
2164                  * case) then there is nothing to do. if it does not then we
2165                  * have to recreate it.
2166                  */
2167                 id_ino(&id) = body->oa.o_id;
2168                 id_gen(&id) = body->oa.o_generation;
2169
2170                 new = mds_id2dentry(obd, &id, NULL);
2171                 if (!IS_ERR(new) && new->d_inode) {
2172                         struct lustre_id sid;
2173
2174                         CWARN("mkdir() repairing is on its way: %lu/%lu\n",
2175                               (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
2176
2177                         obdo_from_inode(&repbody->oa, new->d_inode,
2178                                         FILTER_VALID_FLAGS);
2179
2180                         repbody->oa.o_id = new->d_inode->i_ino;
2181                         repbody->oa.o_generation = new->d_inode->i_generation;
2182                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2183                         cleanup_phase = 1;
2184
2185                         down(&new->d_inode->i_sem);
2186                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2187                         up(&new->d_inode->i_sem);
2188                         if (rc) {
2189                                 CERROR("Can't read inode self id "
2190                                        "inode %lu, rc %d.\n",
2191                                        new->d_inode->i_ino, rc);
2192                                 GOTO(cleanup, rc);
2193                         }
2194
2195                         repbody->oa.o_fid = id_fid(&sid);
2196                         repbody->oa.o_mds = id_group(&sid);
2197                         LASSERT(id_fid(&sid) != 0);
2198
2199                         /*
2200                          * here we could use fid passed in body->oa.o_fid and
2201                          * thus avoid mds_read_inode_sid().
2202                          */
2203                         cr_inum = new->d_inode->i_ino;
2204                         GOTO(cleanup, rc = 0);
2205                 }
2206         }
2207
2208         down(&parent_inode->i_sem);
2209         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2210         if (IS_ERR(handle)) {
2211                 up(&parent_inode->i_sem);
2212                 CERROR("fsfilt_start() failed, rc = %d\n",
2213                        (int)PTR_ERR(handle));
2214                 GOTO(cleanup, rc = PTR_ERR(handle));
2215         }
2216         cleanup_phase = 1; /* transaction */
2217
2218 repeat:
2219         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2220         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2221         if (IS_ERR(new)) {
2222                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2223                        obd->obd_name, idname, (int) PTR_ERR(new));
2224                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2225                 up(&parent_inode->i_sem);
2226                 RETURN(PTR_ERR(new));
2227         } else if (new->d_inode) {
2228                 CERROR("%s: name exists. repeat\n", obd->obd_name);
2229                 goto repeat;
2230         }
2231
2232         new->d_fsdata = (void *)&dp;
2233         dp.p_inum = 0;
2234         dp.p_ptr = req;
2235
2236         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2237             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2238                 LASSERT(body->oa.o_id != 0);
2239                 dp.p_inum = body->oa.o_id;
2240                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2241                           (unsigned long)body->oa.o_id,
2242                           (unsigned long)body->oa.o_generation);
2243         }
2244
2245         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2246         if (rc == 0) {
2247                 if (acl) {
2248                         rc = mds_inode_init_acl(obd, handle, new,
2249                                                 acl, acl_size);
2250                         if (rc) {
2251                                 up(&parent_inode->i_sem);
2252                                 GOTO(cleanup, rc);
2253                         }
2254                 }
2255                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2256                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2257                         new->d_inode->i_generation = body->oa.o_generation;
2258                         mark_inode_dirty(new->d_inode);
2259
2260                         /*
2261                          * avoiding asserts in cache flush case, as
2262                          * @body->oa.o_id should be zero.
2263                          */
2264                         if (body->oa.o_id) {
2265                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2266                                          "BUG 3550: failed to recreate obj "
2267                                          LPU64" -> %lu\n", body->oa.o_id,
2268                                          new->d_inode->i_ino);
2269
2270                                 LASSERTF(body->oa.o_generation ==
2271                                          new->d_inode->i_generation,
2272                                          "BUG 3550: failed to recreate obj/gen "
2273                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2274                                          body->oa.o_generation,
2275                                          new->d_inode->i_ino,
2276                                          new->d_inode->i_generation);
2277                         }
2278                 }
2279
2280                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2281                 repbody->oa.o_id = new->d_inode->i_ino;
2282                 repbody->oa.o_generation = new->d_inode->i_generation;
2283                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2284
2285                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2286                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2287                         id_group(&id) = mds->mds_num;
2288
2289                         LASSERT(body->oa.o_fid != 0);
2290                         id_fid(&id) = body->oa.o_fid;
2291
2292                         LASSERT(body->oa.o_id != 0);
2293                         id_ino(&id) = repbody->oa.o_id;
2294                         id_gen(&id) = repbody->oa.o_generation;
2295
2296                         down(&new->d_inode->i_sem);
2297                         rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2298                         up(&new->d_inode->i_sem);
2299
2300                         /*
2301                          * make sure, that fid is up-to-date.
2302                          */
2303                         mds_set_last_fid(obd, id_fid(&id));
2304                 } else {
2305                         /*
2306                          * allocate new sid, as object is created from scratch
2307                          * and this is not replay.
2308                          */
2309                         down(&new->d_inode->i_sem);
2310                         rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2311                         up(&new->d_inode->i_sem);
2312                 }
2313                 if (rc) {
2314                         CERROR("Can't update lustre ID for inode %lu, "
2315                                "error = %d\n", new->d_inode->i_ino, rc);
2316                         GOTO(cleanup, rc);
2317                 }
2318
2319                 /* initializing o_fid after it is allocated. */
2320                 repbody->oa.o_fid = id_fid(&id);
2321                 repbody->oa.o_mds = id_group(&id);
2322
2323                 rc = fsfilt_del_dir_entry(obd, new);
2324                 up(&parent_inode->i_sem);
2325                 if (rc) {
2326                         CERROR("can't remove name for object: %d\n", rc);
2327                         GOTO(cleanup, rc);
2328                 }
2329
2330                 cleanup_phase = 2; /* created directory object */
2331
2332                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2333                        (unsigned long)new->d_inode->i_ino,
2334                        (unsigned long)new->d_inode->i_generation,
2335                        (unsigned)new->d_inode->i_mode);
2336                 cr_inum = new->d_inode->i_ino;
2337         } else {
2338                 up(&parent_inode->i_sem);
2339                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2340                 GOTO(cleanup, rc);
2341         }
2342
2343         if (body->oa.o_valid & OBD_MD_FLID) {
2344                 /* this is new object for splitted dir. We have to prevent
2345                  * recursive splitting on it -bzzz */
2346                 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2347
2348                 OBD_ALLOC(mea, mealen);
2349                 if (mea == NULL)
2350                         GOTO(cleanup, rc = -ENOMEM);
2351
2352                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2353                 mea->mea_master = 0;
2354                 mea->mea_count = 0;
2355
2356                 down(&new->d_inode->i_sem);
2357                 rc = fsfilt_set_md(obd, new->d_inode, handle,
2358                                    mea, mealen, EA_MEA);
2359                 up(&new->d_inode->i_sem);
2360                 if (rc)
2361                         CERROR("fsfilt_set_md() failed, "
2362                                "rc = %d\n", rc);
2363
2364                 OBD_FREE(mea, mealen);
2365
2366                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2367                        obd->obd_name, new->d_inode->i_ino,
2368                        new->d_inode->i_generation, flags);
2369         } else if (body->oa.o_easize) {
2370                 /* we pass LCK_EX to split routine to signal that we have
2371                  * exclusive access to the directory. simple because nobody
2372                  * knows it already exists -bzzz */
2373                 rc = mds_try_to_split_dir(obd, new, NULL,
2374                                           body->oa.o_easize, LCK_EX);
2375                 if (rc < 0) {
2376                         CERROR("Can't split directory %lu, error = %d.\n",
2377                                new->d_inode->i_ino, rc);
2378                 } else {
2379                         rc = 0;
2380                 }
2381         }
2382
2383         EXIT;
2384 cleanup:
2385         switch (cleanup_phase) {
2386         case 2: /* object has been created, but we'll may want to replay it later */
2387                 if (rc == 0)
2388                         ptlrpc_require_repack(req);
2389         case 1: /* transaction */
2390                 rc = mds_finish_transno(mds, parent_inode, handle,
2391                                         req, rc, cr_inum);
2392         }
2393
2394         l_dput(new);
2395         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2396         return rc;
2397 }
2398
2399 static int mdt_get_info(struct ptlrpc_request *req)
2400 {
2401         struct obd_export *exp = req->rq_export;
2402         int keylen, rc = 0;
2403         char *key;
2404         ENTRY;
2405
2406         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2407         if (key == NULL) {
2408                 DEBUG_REQ(D_HA, req, "no get_info key");
2409                 RETURN(-EFAULT);
2410         }
2411         keylen = req->rq_reqmsg->buflens[0];
2412
2413         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2414             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2415             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2416                 RETURN(-EPROTO);
2417
2418         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2419                 struct lustre_id *reply;
2420                 int size = sizeof(*reply);
2421
2422                 rc = lustre_pack_reply(req, 1, &size, NULL);
2423                 if (rc)
2424                         RETURN(rc);
2425
2426                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2427                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2428         } else {
2429                 obd_id *reply;
2430                 int size = sizeof(*reply);
2431
2432                 rc = lustre_pack_reply(req, 1, &size, NULL);
2433                 if (rc)
2434                         RETURN(rc);
2435
2436                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2437                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2438         }
2439
2440         req->rq_repmsg->status = 0;
2441         RETURN(rc);
2442 }
2443
2444 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2445                         void *key, __u32 vallen, void *val)
2446 {
2447         struct obd_device *obd;
2448         struct mds_obd *mds;
2449         int rc = 0;
2450         ENTRY;
2451
2452         obd = class_exp2obd(exp);
2453         if (obd == NULL) {
2454                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2455                        exp->exp_handle.h_cookie);
2456                 RETURN(-EINVAL);
2457         }
2458
2459         mds = &obd->u.mds;
2460         if (keylen >= strlen("mds_type") &&
2461              memcmp(key, "mds_type", keylen) == 0) {
2462                 int valsize;
2463                 __u32 group;
2464
2465                 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2466
2467                 mds->mds_obd_type = *(int*)val;
2468                 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2469                 valsize = sizeof(group);
2470
2471                 /* mds number has been changed, so the corresponding obdfilter
2472                  * exp need to be changed too. */
2473                 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2474                                   "mds_conn", valsize, &group);
2475                 RETURN(rc);
2476         }
2477         CDEBUG(D_IOCTL, "invalid key\n");
2478         RETURN(-EINVAL);
2479 }
2480
2481 static int mdt_set_info(struct ptlrpc_request *req)
2482 {
2483         char *key, *val;
2484         struct obd_export *exp = req->rq_export;
2485         int keylen, rc = 0, vallen;
2486         ENTRY;
2487
2488         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2489         if (key == NULL) {
2490                 DEBUG_REQ(D_HA, req, "no set_info key");
2491                 RETURN(-EFAULT);
2492         }
2493         keylen = req->rq_reqmsg->buflens[0];
2494
2495         if (keylen == strlen("mds_type") &&
2496             memcmp(key, "mds_type", keylen) == 0) {
2497                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2498                 if (rc)
2499                         RETURN(rc);
2500
2501                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2502                 vallen = req->rq_reqmsg->buflens[1];
2503
2504                 rc = obd_set_info(exp, keylen, key, vallen, val);
2505                 req->rq_repmsg->status = 0;
2506                 RETURN(rc);
2507         }
2508         CDEBUG(D_IOCTL, "invalid key\n");
2509         RETURN(-EINVAL);
2510 }
2511
2512 static void mds_revoke_export_locks(struct obd_export *exp)
2513 {
2514         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
2515         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2516         struct ldlm_lock *lock, *next;
2517         struct ldlm_lock_desc desc;
2518
2519         if (!exp->u.eu_mds_data.med_remote)
2520                 return;
2521
2522         ENTRY;
2523         l_lock(&ns->ns_lock);
2524         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2525                 if (lock->l_req_mode != lock->l_granted_mode)
2526                         continue;
2527
2528                 LASSERT(lock->l_resource);
2529                 if (lock->l_resource->lr_type != LDLM_IBITS &&
2530                     lock->l_resource->lr_type != LDLM_PLAIN)
2531                         continue;
2532
2533                 if (lock->l_flags & LDLM_FL_AST_SENT)
2534                         continue;
2535
2536                 lock->l_flags |= LDLM_FL_AST_SENT;
2537
2538                 /* the desc just pretend to exclusive */
2539                 ldlm_lock2desc(lock, &desc);
2540                 desc.l_req_mode = LCK_EX;
2541                 desc.l_granted_mode = 0;
2542
2543                 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2544         }
2545         l_unlock(&ns->ns_lock);
2546         EXIT;
2547 }
2548
2549 static int mds_msg_check_version(struct lustre_msg *msg)
2550 {
2551         int rc;
2552
2553         switch (msg->opc) {
2554         case MDS_CONNECT:
2555         case MDS_DISCONNECT:
2556         case OBD_PING:
2557                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2558                 if (rc)
2559                         CERROR("bad opc %u version %08x, expecting %08x\n",
2560                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2561                 break;
2562         case MDS_STATFS:
2563         case MDS_GETSTATUS:
2564         case MDS_GETATTR:
2565         case MDS_GETATTR_LOCK:
2566         case MDS_READPAGE:
2567         case MDS_REINT:
2568         case MDS_CLOSE:
2569         case MDS_DONE_WRITING:
2570         case MDS_PIN:
2571         case MDS_SYNC:
2572                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2573                 if (rc)
2574                         CERROR("bad opc %u version %08x, expecting %08x\n",
2575                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2576                 break;
2577         case LDLM_ENQUEUE:
2578         case LDLM_CONVERT:
2579         case LDLM_BL_CALLBACK:
2580         case LDLM_CP_CALLBACK:
2581                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2582                 if (rc)
2583                         CERROR("bad opc %u version %08x, expecting %08x\n",
2584                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2585                 break;
2586         case OBD_LOG_CANCEL:
2587         case LLOG_ORIGIN_HANDLE_OPEN:
2588         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2589         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2590         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2591         case LLOG_ORIGIN_HANDLE_CLOSE:
2592         case LLOG_CATINFO:
2593                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2594                 if (rc)
2595                         CERROR("bad opc %u version %08x, expecting %08x\n",
2596                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2597                 break;
2598         case OST_CREATE:
2599         case OST_WRITE:
2600         case OST_GET_INFO:
2601         case OST_SET_INFO:
2602                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2603                 if (rc)
2604                         CERROR("bad opc %u version %08x, expecting %08x\n",
2605                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2606                 break;
2607         case SEC_INIT:
2608         case SEC_INIT_CONTINUE:
2609         case SEC_FINI:
2610                 rc = 0;
2611                 break;
2612         default:
2613                 CERROR("MDS unknown opcode %d\n", msg->opc);
2614                 rc = -ENOTSUPP;
2615                 break;
2616         }
2617
2618         return rc;
2619 }
2620
2621 int mds_handle(struct ptlrpc_request *req)
2622 {
2623         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2624         struct obd_device *obd = NULL;
2625         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2626         int rc = 0;
2627         ENTRY;
2628
2629         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2630
2631         rc = mds_msg_check_version(req->rq_reqmsg);
2632         if (rc) {
2633                 CERROR("MDS drop mal-formed request\n");
2634                 RETURN(rc);
2635         }
2636
2637         /* Security opc should NOT trigger any recovery events */
2638         if (req->rq_reqmsg->opc == SEC_INIT ||
2639             req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2640                 if (req->rq_export) {
2641                         mds_req_add_idmapping(req,
2642                                               &req->rq_export->exp_mds_data);
2643                         mds_revoke_export_locks(req->rq_export);
2644                 }
2645                 GOTO(out, rc = 0);
2646         } else if (req->rq_reqmsg->opc == SEC_FINI) {
2647                 if (req->rq_export) {
2648                         mds_req_del_idmapping(req,
2649                                               &req->rq_export->exp_mds_data);
2650                         mds_revoke_export_locks(req->rq_export);
2651                 }
2652                 GOTO(out, rc = 0);
2653         }
2654
2655         LASSERT(current->journal_info == NULL);
2656         /* XXX identical to OST */
2657         if (req->rq_reqmsg->opc != MDS_CONNECT) {
2658                 struct mds_export_data *med;
2659                 int recovering;
2660
2661                 if (req->rq_export == NULL) {
2662                         CERROR("operation %d on unconnected MDS from %s\n",
2663                                req->rq_reqmsg->opc,
2664                                req->rq_peerstr);
2665                         req->rq_status = -ENOTCONN;
2666                         GOTO(out, rc = -ENOTCONN);
2667                 }
2668
2669                 med = &req->rq_export->exp_mds_data;
2670                 obd = req->rq_export->exp_obd;
2671                 mds = &obd->u.mds;
2672
2673                 /* sanity check: if the xid matches, the request must
2674                  * be marked as a resent or replayed */
2675                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2676                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2677                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2678                                  (MSG_RESENT | MSG_REPLAY),
2679                                  "rq_xid "LPU64" matches last_xid, "
2680                                  "expected RESENT flag\n",
2681                                  req->rq_xid);
2682                 }
2683                 /* else: note the opposite is not always true; a
2684                  * RESENT req after a failover will usually not match
2685                  * the last_xid, since it was likely never
2686                  * committed. A REPLAYed request will almost never
2687                  * match the last xid, however it could for a
2688                  * committed, but still retained, open. */
2689
2690                 spin_lock_bh(&obd->obd_processing_task_lock);
2691                 recovering = obd->obd_recovering;
2692                 spin_unlock_bh(&obd->obd_processing_task_lock);
2693                 if (recovering) {
2694                         rc = mds_filter_recovery_request(req, obd,
2695                                                          &should_process);
2696                         if (rc || should_process == 0) {
2697                                 RETURN(rc);
2698                         } else if (should_process < 0) {
2699                                 req->rq_status = should_process;
2700                                 rc = ptlrpc_error(req);
2701                                 RETURN(rc);
2702                         }
2703                 }
2704         }
2705
2706         switch (req->rq_reqmsg->opc) {
2707         case MDS_CONNECT:
2708                 DEBUG_REQ(D_INODE, req, "connect");
2709                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2710                 rc = target_handle_connect(req);
2711                 if (!rc) {
2712                         struct mds_export_data *med;
2713
2714                         LASSERT(req->rq_export);
2715                         med = &req->rq_export->u.eu_mds_data;
2716                         mds_init_export_data(req, med);
2717                         mds_req_add_idmapping(req, med);
2718
2719                         /* Now that we have an export, set mds. */
2720                         obd = req->rq_export->exp_obd;
2721                         mds = mds_req2mds(req);
2722                 }
2723                 break;
2724
2725         case MDS_DISCONNECT:
2726                 DEBUG_REQ(D_INODE, req, "disconnect");
2727                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2728                 rc = target_handle_disconnect(req);
2729                 req->rq_status = rc;            /* superfluous? */
2730                 break;
2731
2732         case MDS_GETSTATUS:
2733                 DEBUG_REQ(D_INODE, req, "getstatus");
2734                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2735                 rc = mds_getstatus(req);
2736                 break;
2737
2738         case MDS_GETATTR:
2739                 DEBUG_REQ(D_INODE, req, "getattr");
2740                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2741                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2742                 break;
2743
2744         case MDS_GETATTR_LOCK: {
2745                 struct lustre_handle lockh;
2746                 DEBUG_REQ(D_INODE, req, "getattr_lock");
2747                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2748
2749                 /* If this request gets a reconstructed reply, we won't be
2750                  * acquiring any new locks in mds_getattr_lock, so we don't
2751                  * want to cancel.
2752                  */
2753                 lockh.cookie = 0;
2754                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2755                                       MDS_INODELOCK_UPDATE);
2756                 /* this non-intent call (from an ioctl) is special */
2757                 req->rq_status = rc;
2758                 if (rc == 0 && lockh.cookie)
2759                         ldlm_lock_decref(&lockh, LCK_PR);
2760                 break;
2761         }
2762         case MDS_STATFS:
2763                 DEBUG_REQ(D_INODE, req, "statfs");
2764                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2765                 rc = mds_statfs(req);
2766                 break;
2767
2768         case MDS_READPAGE:
2769                 DEBUG_REQ(D_INODE, req, "readpage");
2770                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2771                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2772
2773                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2774                         if (req->rq_reply_state) {
2775                                 lustre_free_reply_state (req->rq_reply_state);
2776                                 req->rq_reply_state = NULL;
2777                         }
2778                         RETURN(0);
2779                 }
2780
2781                 break;
2782         case MDS_REINT: {
2783                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2784                                              sizeof (*opcp));
2785                 __u32  opc;
2786                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2787                                mds->mds_max_cookiesize};
2788                 int bufcount;
2789
2790                 /* NB only peek inside req now; mds_reint() will swab it */
2791                 if (opcp == NULL) {
2792                         CERROR ("Can't inspect opcode\n");
2793                         rc = -EINVAL;
2794                         break;
2795                 }
2796                 opc = *opcp;
2797                 if (lustre_msg_swabbed (req->rq_reqmsg))
2798                         __swab32s(&opc);
2799
2800                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2801                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2802                            reint_names[opc] == NULL) ? reint_names[opc] :
2803                                                        "unknown opcode");
2804
2805                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2806
2807                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2808                         bufcount = 3;
2809                 else if (opc == REINT_OPEN)
2810                         bufcount = 2;
2811                 else
2812                         bufcount = 1;
2813
2814                 rc = lustre_pack_reply(req, bufcount, size, NULL);
2815                 if (rc)
2816                         break;
2817
2818                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2819                 fail = OBD_FAIL_MDS_REINT_NET_REP;
2820                 break;
2821         }
2822
2823         case MDS_CLOSE:
2824                 DEBUG_REQ(D_INODE, req, "close");
2825                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2826                 rc = mds_close(req, MDS_REQ_REC_OFF);
2827                 break;
2828
2829         case MDS_DONE_WRITING:
2830                 DEBUG_REQ(D_INODE, req, "done_writing");
2831                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2832                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2833                 break;
2834
2835         case MDS_PIN:
2836                 DEBUG_REQ(D_INODE, req, "pin");
2837                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2838                 rc = mds_pin(req, MDS_REQ_REC_OFF);
2839                 break;
2840
2841         case MDS_SYNC:
2842                 DEBUG_REQ(D_INODE, req, "sync");
2843                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2844                 rc = mds_sync(req, MDS_REQ_REC_OFF);
2845                 break;
2846
2847         case OBD_PING:
2848                 DEBUG_REQ(D_INODE, req, "ping");
2849                 rc = target_handle_ping(req);
2850                 break;
2851
2852         case OBD_LOG_CANCEL:
2853                 CDEBUG(D_INODE, "log cancel\n");
2854                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2855                 rc = -ENOTSUPP; /* la la la */
2856                 break;
2857
2858         case LDLM_ENQUEUE:
2859                 DEBUG_REQ(D_INODE, req, "enqueue");
2860                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2861                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2862                                          ldlm_server_blocking_ast, NULL);
2863                 fail = OBD_FAIL_LDLM_REPLY;
2864                 break;
2865         case LDLM_CONVERT:
2866                 DEBUG_REQ(D_INODE, req, "convert");
2867                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2868                 rc = ldlm_handle_convert(req);
2869                 break;
2870         case LDLM_BL_CALLBACK:
2871         case LDLM_CP_CALLBACK:
2872                 DEBUG_REQ(D_INODE, req, "callback");
2873                 CERROR("callbacks should not happen on MDS\n");
2874                 LBUG();
2875                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2876                 break;
2877         case LLOG_ORIGIN_HANDLE_OPEN:
2878                 DEBUG_REQ(D_INODE, req, "llog_init");
2879                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2880                 rc = llog_origin_handle_open(req);
2881                 break;
2882         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2883                 DEBUG_REQ(D_INODE, req, "llog next block");
2884                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2885                 rc = llog_origin_handle_next_block(req);
2886                 break;
2887         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2888                 DEBUG_REQ(D_INODE, req, "llog prev block");
2889                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2890                 rc = llog_origin_handle_prev_block(req);
2891                 break;
2892         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2893                 DEBUG_REQ(D_INODE, req, "llog read header");
2894                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2895                 rc = llog_origin_handle_read_header(req);
2896                 break;
2897         case LLOG_ORIGIN_HANDLE_CLOSE:
2898                 DEBUG_REQ(D_INODE, req, "llog close");
2899                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2900                 rc = llog_origin_handle_close(req);
2901                 break;
2902         case OST_CREATE:
2903                 DEBUG_REQ(D_INODE, req, "ost_create");
2904                 rc = mdt_obj_create(req);
2905                 break;
2906         case OST_GET_INFO:
2907                 DEBUG_REQ(D_INODE, req, "get_info");
2908                 rc = mdt_get_info(req);
2909                 break;
2910         case OST_SET_INFO:
2911                 DEBUG_REQ(D_INODE, req, "set_info");
2912                 rc = mdt_set_info(req);
2913                 break;
2914         case OST_WRITE:
2915                 CDEBUG(D_INODE, "write\n");
2916                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2917                 rc = ost_brw_write(req, NULL);
2918                 LASSERT(current->journal_info == NULL);
2919                 /* mdt_brw sends its own replies */
2920                 RETURN(rc);
2921                 break;
2922         case LLOG_CATINFO:
2923                 DEBUG_REQ(D_INODE, req, "llog catinfo");
2924                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2925                 rc = llog_catinfo(req);
2926                 break;
2927         default:
2928                 req->rq_status = -ENOTSUPP;
2929                 rc = ptlrpc_error(req);
2930                 RETURN(rc);
2931         }
2932
2933         LASSERT(current->journal_info == NULL);
2934
2935         EXIT;
2936
2937         /* If we're DISCONNECTing, the mds_export_data is already freed */
2938         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2939                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2940                 struct obd_device *obd = list_entry(mds, struct obd_device,
2941                                                     u.mds);
2942                 req->rq_repmsg->last_xid =
2943                         le64_to_cpu(med->med_mcd->mcd_last_xid);
2944
2945                 if (!obd->obd_no_transno) {
2946                         req->rq_repmsg->last_committed =
2947                                 obd->obd_last_committed;
2948                 } else {
2949                         DEBUG_REQ(D_IOCTL, req,
2950                                   "not sending last_committed update");
2951                 }
2952                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2953                        ", xid "LPU64"\n",
2954                        mds->mds_last_transno, obd->obd_last_committed,
2955                        req->rq_xid);
2956         }
2957  out:
2958
2959
2960         target_send_reply(req, rc, fail);
2961         return 0;
2962 }
2963
2964 /* Update the server data on disk.  This stores the new mount_count and also the
2965  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
2966  * last_rcvd value may be less than that of the clients.  This will alert us
2967  * that we may need to do client recovery.
2968  *
2969  * Also assumes for mds_last_transno that we are not modifying it (no locking).
2970  */
2971 int mds_update_server_data(struct obd_device *obd, int force_sync)
2972 {
2973         struct mds_obd *mds = &obd->u.mds;
2974         struct mds_server_data *msd = mds->mds_server_data;
2975         struct file *filp = mds->mds_rcvd_filp;
2976         struct lvfs_run_ctxt saved;
2977         loff_t off = 0;
2978         int rc;
2979         ENTRY;
2980
2981         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2982         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2983
2984         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2985                mds->mds_mount_count, mds->mds_last_transno);
2986         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2987         if (rc)
2988                 CERROR("error writing MDS server data: rc = %d\n", rc);
2989         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2990
2991         RETURN(rc);
2992 }
2993
2994 /* saves last allocated fid counter to file. */
2995 int mds_update_last_fid(struct obd_device *obd, void *handle,
2996                         int force_sync)
2997 {
2998         struct mds_obd *mds = &obd->u.mds;
2999         struct file *filp = mds->mds_fid_filp;
3000         struct lvfs_run_ctxt saved;
3001         loff_t off = 0;
3002         __u64 last_fid;
3003         int rc = 0;
3004         ENTRY;
3005
3006         spin_lock(&mds->mds_last_fid_lock);
3007         last_fid = mds->mds_last_fid;
3008         spin_unlock(&mds->mds_last_fid_lock);
3009
3010         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3011                last_fid);
3012
3013         if (handle) {
3014                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3015                                       handle, mds_commit_last_fid_cb,
3016                                       NULL);
3017         }
3018
3019         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3020         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3021                                  &off, force_sync);
3022         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3023
3024         if (rc) {
3025                 CERROR("error writing MDS last_fid #"LPU64
3026                        ", err = %d\n", last_fid, rc);
3027                 RETURN(rc);
3028         }
3029
3030         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3031                "%llu: err = %d\n", last_fid, off, rc);
3032
3033         RETURN(rc);
3034 }
3035
3036 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3037 {
3038         struct mds_obd *mds = &obd->u.mds;
3039
3040         spin_lock(&mds->mds_last_fid_lock);
3041         if (fid > mds->mds_last_fid)
3042                 mds->mds_last_fid = fid;
3043         spin_unlock(&mds->mds_last_fid_lock);
3044 }
3045
3046 void mds_commit_last_transno_cb(struct obd_device *obd,
3047                                 __u64 transno, void *data,
3048                                 int error)
3049 {
3050         obd_transno_commit_cb(obd, transno, error);
3051 }
3052
3053 void mds_commit_last_fid_cb(struct obd_device *obd,
3054                             __u64 fid, void *data,
3055                             int error)
3056 {
3057         if (error) {
3058                 CERROR("%s: fid "LPD64" commit error: %d\n",
3059                        obd->obd_name, fid, error);
3060                 return;
3061         }
3062
3063         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3064                obd->obd_name, fid);
3065 }
3066
3067 __u64 mds_alloc_fid(struct obd_device *obd)
3068 {
3069         struct mds_obd *mds = &obd->u.mds;
3070         __u64 fid;
3071
3072         spin_lock(&mds->mds_last_fid_lock);
3073         fid = ++mds->mds_last_fid;
3074         spin_unlock(&mds->mds_last_fid_lock);
3075
3076         return fid;
3077 }
3078
3079 /*
3080  * allocates new lustre_id on passed @inode and saves it to inode EA.
3081  */
3082 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
3083                         void *handle, struct lustre_id *id)
3084 {
3085         struct mds_obd *mds = &obd->u.mds;
3086         int alloc = 0, rc = 0;
3087         ENTRY;
3088
3089         LASSERT(obd != NULL);
3090         LASSERT(inode != NULL);
3091
3092         if (id == NULL) {
3093                 OBD_ALLOC(id, sizeof(*id));
3094                 if (id == NULL)
3095                         RETURN(-ENOMEM);
3096                 alloc = 1;
3097         }
3098
3099         id_group(id) = mds->mds_num;
3100         id_fid(id) = mds_alloc_fid(obd);
3101         id_ino(id) = inode->i_ino;
3102         id_gen(id) = inode->i_generation;
3103         id_type(id) = (S_IFMT & inode->i_mode);
3104
3105         rc = mds_update_inode_sid(obd, inode, handle, id);
3106         if (rc) {
3107                 CERROR("Can't update inode FID EA, "
3108                        "rc = %d\n", rc);
3109         }
3110
3111         if (alloc)
3112                 OBD_FREE(id, sizeof(*id));
3113         RETURN(rc);
3114 }
3115
3116 /*
3117  * reads inode self id from inode EA. Probably later this should be replaced by
3118  * caching inode self id to avoid raeding it every time it is needed.
3119  */
3120 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3121                        struct lustre_id *id)
3122 {
3123         int rc;
3124         ENTRY;
3125
3126         LASSERT(id != NULL);
3127         LASSERT(obd != NULL);
3128         LASSERT(inode != NULL);
3129
3130         rc = fsfilt_get_md(obd, inode, &id->li_fid,
3131                            sizeof(id->li_fid), EA_SID);
3132         if (rc < 0) {
3133                 CERROR("fsfilt_get_md() failed, "
3134                        "rc = %d\n", rc);
3135                 RETURN(rc);
3136         } else if (!rc) {
3137                 rc = -ENODATA;
3138                 RETURN(rc);
3139         } else {
3140                 rc = 0;
3141         }
3142
3143         RETURN(rc);
3144 }
3145
3146 /* updates inode self id in EA. */
3147 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3148                          void *handle, struct lustre_id *id)
3149 {
3150         int rc = 0;
3151         ENTRY;
3152
3153         LASSERT(id != NULL);
3154         LASSERT(obd != NULL);
3155         LASSERT(inode != NULL);
3156
3157         rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3158                            sizeof(id->li_fid), EA_SID);
3159         if (rc) {
3160                 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3161                 RETURN(rc);
3162         }
3163
3164         RETURN(rc);
3165 }
3166
3167 /*
3168  * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3169  * to master MDS by correct store cookie, needed to find inode on master MDS
3170  * quickly.
3171  */
3172 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3173                        struct lustre_id *id)
3174 {
3175         int rc;
3176         ENTRY;
3177
3178         LASSERT(id != NULL);
3179         LASSERT(obd != NULL);
3180         LASSERT(inode != NULL);
3181
3182         rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3183         if (rc < 0) {
3184                 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3185                 RETURN(rc);
3186         } else if (!rc) {
3187                 rc = -ENODATA;
3188                 RETURN(rc);
3189         } else {
3190                 rc = 0;
3191         }
3192
3193         RETURN(rc);
3194 }
3195
3196 /*
3197  * updates master inode id. Usualy this is done by CMOBD after an inode is
3198  * created and relationship between cache MDS and master one should be
3199  * established.
3200  */
3201 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3202                          void *handle, struct lustre_id *id)
3203 {
3204         int rc = 0;
3205         ENTRY;
3206
3207         LASSERT(id != NULL);
3208         LASSERT(obd != NULL);
3209         LASSERT(inode != NULL);
3210
3211         rc = fsfilt_set_md(obd, inode, handle, id,
3212                            sizeof(*id), EA_MID);
3213         if (rc) {
3214                 CERROR("fsfilt_set_md() failed, "
3215                        "rc = %d\n", rc);
3216                 RETURN(rc);
3217         }
3218
3219         RETURN(rc);
3220 }
3221
3222 /* mount the file system (secretly) */
3223 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3224 {
3225         struct lustre_cfg* lcfg = buf;
3226         struct mds_obd *mds = &obd->u.mds;
3227         struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3228         char *options = NULL;
3229         struct vfsmount *mnt;
3230         char ns_name[48];
3231         unsigned long page;
3232         int rc = 0;
3233         ENTRY;
3234
3235         if (lcfg->lcfg_bufcount < 3)
3236                 RETURN(rc = -EINVAL);
3237
3238         if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3239                 RETURN(rc = -EINVAL);
3240
3241         obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3242         if (IS_ERR(obd->obd_fsops))
3243                 RETURN(rc = PTR_ERR(obd->obd_fsops));
3244
3245         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3246
3247         page = __get_free_page(GFP_KERNEL);
3248         if (!page)
3249                 RETURN(-ENOMEM);
3250
3251         options = (char *)page;
3252         memset(options, 0, PAGE_SIZE);
3253
3254         /*
3255          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3256          * and the rest of options are passed by mount options. Probably this
3257          * should be moved to somewhere else like startup scripts or lconf. */
3258         sprintf(options, "iopen_nopriv");
3259
3260         if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3261                 sprintf(options + strlen(options), ",%s",
3262                         lustre_cfg_string(lcfg, 4));
3263
3264         /* we have to know mdsnum before touching underlying fs -bzzz */
3265         atomic_set(&mds->mds_open_count, 0);
3266         sema_init(&mds->mds_md_sem, 1);
3267         mds->mds_md_connected = 0;
3268         mds->mds_md_name = NULL;
3269
3270         if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3271             strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3272                 class_uuid_t uuid;
3273
3274                 generate_random_uuid(uuid);
3275                 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3276
3277                 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3278                 if (mds->mds_md_name == NULL)
3279                         RETURN(rc = -ENOMEM);
3280
3281                 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3282                        LUSTRE_CFG_BUFLEN(lcfg, 5));
3283
3284                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3285                        obd->obd_name, mds->mds_md_name);
3286
3287                 rc = mds_md_connect(obd, mds->mds_md_name);
3288                 if (rc) {
3289                         OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3290                         GOTO(err_ops, rc);
3291                 }
3292         }
3293
3294         mds->mds_obd_type = MDS_MASTER_OBD;
3295
3296         if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3297             strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3298                     LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3299                 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3300                             strlen("master"))) {
3301                         mds->mds_obd_type = MDS_MASTER_OBD;
3302                 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3303                                    strlen("cache"))) {
3304                         mds->mds_obd_type = MDS_CACHE_OBD;
3305                 }
3306         }
3307
3308         rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3309                            lustre_cfg_string(lcfg, 2),
3310                            options, 0, &lvfs_ctxt);
3311
3312         free_page(page);
3313
3314         if (rc || !lvfs_ctxt) {
3315                 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3316                 GOTO(err_ops, rc);
3317         }
3318
3319         mnt = lvfs_ctxt->loc_mnt;
3320         mds->mds_lvfs_ctxt = lvfs_ctxt;
3321         ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3322
3323         CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3324
3325         sema_init(&mds->mds_epoch_sem, 1);
3326         atomic_set(&mds->mds_real_clients, 0);
3327         spin_lock_init(&mds->mds_transno_lock);
3328         spin_lock_init(&mds->mds_last_fid_lock);
3329         sema_init(&mds->mds_orphan_recovery_sem, 1);
3330         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3331
3332         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3333         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3334
3335         if (obd->obd_namespace == NULL) {
3336                 mds_cleanup(obd, 0);
3337                 GOTO(err_put, rc = -ENOMEM);
3338         }
3339         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3340
3341         rc = mds_fs_setup(obd, mnt);
3342         if (rc) {
3343                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3344                        obd->obd_name, rc);
3345                 GOTO(err_ns, rc);
3346         }
3347
3348         rc = llog_start_commit_thread();
3349         if (rc < 0)
3350
3351                 GOTO(err_fs, rc);
3352
3353
3354         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3355             strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3356                     LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3357                 class_uuid_t uuid;
3358
3359                 generate_random_uuid(uuid);
3360                 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3361
3362                 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3363                 if (mds->mds_profile == NULL)
3364                         GOTO(err_fs, rc = -ENOMEM);
3365
3366                 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3367                         LUSTRE_CFG_BUFLEN(lcfg, 3));
3368         }
3369
3370         /*
3371          * setup root dir and files ID dir if lmv already connected, or there is
3372          * not lmv at all.
3373          */
3374         if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3375                                 lustre_cfg_buf(lcfg, 3) &&
3376                                 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3377                                         LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3378                 rc = mds_fs_setup_rootid(obd);
3379                 if (rc)
3380                         GOTO(err_fs, rc);
3381
3382                 rc = mds_fs_setup_virtid(obd);
3383                 if (rc)
3384                         GOTO(err_fs, rc);
3385         }
3386
3387         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3388                            "mds_ldlm_client", &obd->obd_ldlm_client);
3389         obd->obd_replayable = 1;
3390
3391         rc = mds_postsetup(obd);
3392         if (rc)
3393                 GOTO(err_fs, rc);
3394
3395         RETURN(0);
3396
3397 err_fs:
3398         /* No extra cleanup needed for llog_init_commit_thread() */
3399         mds_fs_cleanup(obd, 0);
3400 err_ns:
3401         ldlm_namespace_free(obd->obd_namespace, 0);
3402         obd->obd_namespace = NULL;
3403 err_put:
3404         unlock_kernel();
3405         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3406         mds->mds_sb = 0;
3407         lock_kernel();
3408 err_ops:
3409         fsfilt_put_ops(obd->obd_fsops);
3410         return rc;
3411 }
3412
3413 static int mds_postsetup(struct obd_device *obd)
3414 {
3415         struct mds_obd *mds = &obd->u.mds;
3416         int rc = 0;
3417         ENTRY;
3418
3419         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3420                             obd, 0, NULL, &llog_lvfs_ops);
3421         if (rc)
3422                 RETURN(rc);
3423
3424         if (mds->mds_profile) {
3425                 struct llog_ctxt *lgctxt;
3426                 struct lvfs_run_ctxt saved;
3427                 struct lustre_profile *lprof;
3428                 struct config_llog_instance cfg;
3429
3430                 cfg.cfg_instance = NULL;
3431                 cfg.cfg_uuid = mds->mds_dt_uuid;
3432                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3433
3434                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3435                 if (!lgctxt) {
3436                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3437                         GOTO(err_llog, rc = -EINVAL);
3438                 }
3439
3440                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3441                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3442
3443                 if (rc)
3444                         GOTO(err_llog, rc);
3445
3446                 lprof = class_get_profile(mds->mds_profile);
3447                 if (lprof == NULL) {
3448                         CERROR("No profile found: %s\n", mds->mds_profile);
3449                         GOTO(err_cleanup, rc = -ENOENT);
3450                 }
3451                 rc = mds_dt_connect(obd, lprof->lp_lov);
3452                 if (rc)
3453                         GOTO(err_cleanup, rc);
3454
3455                 rc = mds_md_postsetup(obd);
3456                 if (rc)
3457                         GOTO(err_cleanup, rc);
3458         }
3459
3460         RETURN(rc);
3461 err_cleanup:
3462         mds_dt_clean(obd);
3463 err_llog:
3464         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3465                                           LLOG_CONFIG_ORIG_CTXT));
3466         return rc;
3467 }
3468
3469 int mds_postrecov_common(struct obd_device *obd)
3470 {
3471         struct mds_obd *mds = &obd->u.mds;
3472         struct llog_ctxt *ctxt;
3473         int rc, item = 0, valsize;
3474          __u32 group;
3475         ENTRY;
3476
3477         LASSERT(!obd->obd_recovering);
3478         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3479         LASSERT(ctxt != NULL);
3480
3481         /* set nextid first, so we are sure it happens */
3482         rc = mds_dt_set_nextid(obd);
3483         if (rc) {
3484                 CERROR("%s: mds_dt_set_nextid() failed\n", obd->obd_name);
3485                 GOTO(out, rc);
3486         }
3487
3488         /* clean PENDING dir */
3489         rc = mds_cleanup_orphans(obd);
3490         if (rc < 0)
3491                 GOTO(out, rc);
3492         item = rc;
3493
3494         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3495         valsize = sizeof(group);
3496         rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"), "mds_conn",
3497                           valsize, &group);
3498         if (rc)
3499                 GOTO(out, rc);
3500
3501         rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3502                           NULL, NULL, NULL);
3503         if (rc) {
3504                 CERROR("%s: failed at llog_origin_connect: %d\n",
3505                        obd->obd_name, rc);
3506                 GOTO(out, rc);
3507         }
3508
3509         /* remove the orphaned precreated objects */
3510         rc = mds_dt_clearorphans(mds, NULL /* all OSTs */);
3511         if (rc)
3512                 GOTO(err_llog, rc);
3513
3514 out:
3515         RETURN(rc < 0 ? rc : item);
3516
3517 err_llog:
3518         /* cleanup all llogging subsystems */
3519         rc = obd_llog_finish(obd, &obd->obd_llogs,
3520                              mds->mds_dt_desc.ld_tgt_count);
3521         if (rc)
3522                 CERROR("%s: failed to cleanup llogging subsystems\n",
3523                         obd->obd_name);
3524         goto out;
3525 }
3526
3527 int mds_postrecov(struct obd_device *obd)
3528 {
3529         int rc;
3530         ENTRY;
3531         rc = mds_postrecov_common(obd);
3532         if (rc == 0)
3533                 rc = mds_md_reconnect(obd);
3534         RETURN(rc);
3535 }
3536
3537 int mds_dt_clean(struct obd_device *obd)
3538 {
3539         struct mds_obd *mds = &obd->u.mds;
3540         ENTRY;
3541
3542         if (mds->mds_profile) {
3543                 char * cln_prof;
3544                 struct llog_ctxt *llctx;
3545                 struct lvfs_run_ctxt saved;
3546                 struct config_llog_instance cfg;
3547                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3548
3549                 OBD_ALLOC(cln_prof, len);
3550                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3551
3552                 cfg.cfg_instance = NULL;
3553                 cfg.cfg_uuid = mds->mds_dt_uuid;
3554
3555                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3556                 llctx = llog_get_context(&obd->obd_llogs,
3557                                          LLOG_CONFIG_ORIG_CTXT);
3558                 class_config_process_llog(llctx, cln_prof, &cfg);
3559                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3560
3561                 OBD_FREE(cln_prof, len);
3562                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3563                 mds->mds_profile = NULL;
3564         }
3565         RETURN(0);
3566 }
3567
3568 int mds_md_clean(struct obd_device *obd)
3569 {
3570         struct mds_obd *mds = &obd->u.mds;
3571         ENTRY;
3572
3573         if (mds->mds_md_name) {
3574                 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3575                 mds->mds_md_name = NULL;
3576         }
3577         RETURN(0);
3578 }
3579
3580 static int mds_precleanup(struct obd_device *obd, int flags)
3581 {
3582         int rc = 0;
3583         ENTRY;
3584
3585         mds_md_clean(obd);
3586         mds_dt_disconnect(obd, flags);
3587         mds_dt_clean(obd);
3588         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3589         RETURN(rc);
3590 }
3591
3592 extern void lgss_svc_cache_purge_all(void);
3593 static int mds_cleanup(struct obd_device *obd, int flags)
3594 {
3595         struct mds_obd *mds = &obd->u.mds;
3596         ENTRY;
3597
3598         if (mds->mds_sb == NULL)
3599                 RETURN(0);
3600
3601         mds_update_server_data(obd, 1);
3602         mds_update_last_fid(obd, NULL, 1);
3603
3604         if (mds->mds_dt_objids != NULL) {
3605                 int size = mds->mds_dt_desc.ld_tgt_count *
3606                         sizeof(obd_id);
3607                 OBD_FREE(mds->mds_dt_objids, size);
3608         }
3609         mds_fs_cleanup(obd, flags);
3610
3611         unlock_kernel();
3612
3613         /* 2 seems normal on mds, (may_umount() also expects 2
3614           fwiw), but we only see 1 at this point in obdfilter. */
3615         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3616
3617         mds->mds_sb = 0;
3618
3619         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3620
3621         spin_lock_bh(&obd->obd_processing_task_lock);
3622         if (obd->obd_recovering) {
3623                 target_cancel_recovery_timer(obd);
3624                 obd->obd_recovering = 0;
3625         }
3626         spin_unlock_bh(&obd->obd_processing_task_lock);
3627
3628         lock_kernel();
3629         fsfilt_put_ops(obd->obd_fsops);
3630
3631 #ifdef ENABLE_GSS
3632         /* XXX */
3633         lgss_svc_cache_purge_all();
3634 #endif
3635
3636         spin_lock(&mds->mds_denylist_lock);
3637         while (!list_empty( &mds->mds_denylist ) ) {
3638                 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3639                                                     deny_sec_t, list);
3640                 list_del(&p_deny_sec->list);
3641                 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3642         }
3643         spin_unlock(&mds->mds_denylist_lock);
3644         if(mds->mds_mds_sec)
3645                 OBD_FREE(mds->mds_mds_sec, strlen(mds->mds_mds_sec) + 1);
3646         if(mds->mds_ost_sec)
3647                 OBD_FREE(mds->mds_ost_sec, strlen(mds->mds_ost_sec) + 1);
3648
3649         RETURN(0);
3650 }
3651
3652 static int set_security(const char *value, char **sec)
3653 {
3654         int rc = 0;
3655
3656         if (!strcmp(value, "null") ||
3657             !strcmp(value, "krb5i") ||
3658             !strcmp(value, "krb5p")) {
3659                 OBD_ALLOC(*sec, strlen(value) + 1);
3660                 if(!*sec)
3661                         RETURN(-ENOMEM);
3662                 memcpy(*sec, value, strlen(value) + 1);
3663         } else {
3664                 CERROR("Unrecognized value, force use NULL\n");
3665                 rc = -EINVAL;
3666         }
3667
3668         return rc;
3669 }
3670
3671 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3672 {
3673         struct lustre_cfg *lcfg = buf;
3674         struct mds_obd *mds = &obd->u.mds;
3675         int rc = 0;
3676         ENTRY;
3677
3678         switch(lcfg->lcfg_command) {
3679         case LCFG_SET_SECURITY: {
3680                 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
3681                     (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
3682                         GOTO(out, rc = -EINVAL);
3683
3684                 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
3685                         rc = set_security(lustre_cfg_string(lcfg, 2),
3686                                           &mds->mds_mds_sec);
3687                 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
3688                         rc = set_security(lustre_cfg_string(lcfg, 2),
3689                                           &mds->mds_ost_sec);
3690                 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
3691                         spin_lock(&mds->mds_denylist_lock);
3692                         rc = add_deny_security(lustre_cfg_string(lcfg, 2),
3693                                                &mds->mds_denylist);
3694                         spin_unlock(&mds->mds_denylist_lock);
3695                 } else {
3696                         CERROR("Unrecognized key\n");
3697                         rc = -EINVAL;
3698                 }
3699                 break;
3700         }
3701         default: {
3702                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3703                 GOTO(out, rc = -EINVAL);
3704
3705         }
3706         }
3707 out:
3708         RETURN(rc);
3709 }
3710
3711 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3712                                         int offset,
3713                                         struct ldlm_lock *new_lock,
3714                                         struct ldlm_lock **old_lock,
3715                                         struct lustre_handle *lockh)
3716 {
3717         struct obd_export *exp = req->rq_export;
3718         struct obd_device *obd = exp->exp_obd;
3719         struct ldlm_request *dlmreq =
3720                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3721         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3722         struct list_head *iter;
3723
3724         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3725                 return;
3726
3727         l_lock(&obd->obd_namespace->ns_lock);
3728         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3729                 struct ldlm_lock *lock;
3730                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3731                 if (lock == new_lock)
3732                         continue;
3733                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3734                         lockh->cookie = lock->l_handle.h_cookie;
3735                         LDLM_DEBUG(lock, "restoring lock cookie");
3736                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3737                                   lockh->cookie);
3738                         if (old_lock)
3739                                 *old_lock = LDLM_LOCK_GET(lock);
3740                         l_unlock(&obd->obd_namespace->ns_lock);
3741                         return;
3742                 }
3743         }
3744         l_unlock(&obd->obd_namespace->ns_lock);
3745
3746         /* If the xid matches, then we know this is a resent request,
3747          * and allow it. (It's probably an OPEN, for which we don't
3748          * send a lock */
3749         if (req->rq_xid ==
3750             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3751                 return;
3752
3753         if (req->rq_xid ==
3754             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3755                 return;
3756
3757         /* This remote handle isn't enqueued, so we never received or
3758          * processed this request.  Clear MSG_RESENT, because it can
3759          * be handled like any normal request now. */
3760
3761         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3762
3763         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3764                   remote_hdl.cookie);
3765 }
3766
3767 int intent_disposition(struct ldlm_reply *rep, int flag)
3768 {
3769         if (!rep)
3770                 return 0;
3771         return (rep->lock_policy_res1 & flag);
3772 }
3773
3774 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3775 {
3776         if (!rep)
3777                 return;
3778         rep->lock_policy_res1 |= flag;
3779 }
3780
3781 static int mds_intent_policy(struct ldlm_namespace *ns,
3782                              struct ldlm_lock **lockp, void *req_cookie,
3783                              ldlm_mode_t mode, int flags, void *data)
3784 {
3785         struct ptlrpc_request *req = req_cookie;
3786         struct ldlm_lock *lock = *lockp;
3787         struct ldlm_intent *it;
3788         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3789         struct ldlm_reply *rep;
3790         struct lustre_handle lockh[2] = {{0}, {0}};
3791         struct ldlm_lock *new_lock = NULL;
3792         int getattr_part = MDS_INODELOCK_UPDATE;
3793         int rc, reply_buffers;
3794         int repsize[5] = {sizeof(struct ldlm_reply),
3795                           sizeof(struct mds_body),
3796                           mds->mds_max_mdsize};
3797
3798         int offset = MDS_REQ_INTENT_REC_OFF;
3799         ENTRY;
3800
3801         LASSERT(req != NULL);
3802         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3803
3804         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3805                 /* No intent was provided */
3806                 int size = sizeof(struct ldlm_reply);
3807                 rc = lustre_pack_reply(req, 1, &size, NULL);
3808                 LASSERT(rc == 0);
3809                 RETURN(0);
3810         }
3811
3812         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3813                                 lustre_swab_ldlm_intent);
3814         if (it == NULL) {
3815                 CERROR("Intent missing\n");
3816                 RETURN(req->rq_status = -EFAULT);
3817         }
3818
3819         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3820
3821         reply_buffers = 3;
3822         if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3823                 reply_buffers = 5;
3824                 repsize[3] = 4;
3825                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
3826         }
3827
3828         rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
3829         if (rc)
3830                 RETURN(req->rq_status = rc);
3831
3832         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3833         LASSERT(rep != NULL);
3834
3835         intent_set_disposition(rep, DISP_IT_EXECD);
3836
3837         /* execute policy */
3838         switch ((long)it->opc) {
3839         case IT_OPEN:
3840         case IT_CREAT|IT_OPEN:
3841                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3842                                             lock, NULL, lockh);
3843                 /* XXX swab here to assert that an mds_open reint
3844                  * packet is following */
3845                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3846                                             lock, NULL, lockh);
3847                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3848
3849                 if (rep->lock_policy_res2) {
3850                         /*
3851                          * mds_open() returns ENOLCK where it should return
3852                          * zero, but it has no lock to return.
3853                          */
3854                         if (rep->lock_policy_res2 == ENOLCK)
3855                                 rep->lock_policy_res2 = 0;
3856
3857                         RETURN(ELDLM_LOCK_ABORTED);
3858                 }
3859
3860                 /*
3861                  * IT_OPEN may return lock on cross-node dentry that we want to
3862                  * hold during attr retrival -bzzz
3863                  */
3864                 if (lockh[0].cookie == 0)
3865                         RETURN(ELDLM_LOCK_ABORTED);
3866
3867                 break;
3868         case IT_LOOKUP:
3869                 getattr_part = MDS_INODELOCK_LOOKUP;
3870         case IT_CHDIR:
3871         case IT_GETATTR:
3872                 getattr_part |= MDS_INODELOCK_LOOKUP;
3873         case IT_READDIR:
3874                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3875                                             lock, &new_lock, lockh);
3876                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3877                                                          getattr_part);
3878                 /* FIXME: LDLM can set req->rq_status. MDS sets
3879                    policy_res{1,2} with disposition and status.
3880                    - replay: returns 0 & req->status is old status
3881                    - otherwise: returns req->status */
3882                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3883                         rep->lock_policy_res2 = 0;
3884                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3885                     rep->lock_policy_res2)
3886                         RETURN(ELDLM_LOCK_ABORTED);
3887                 if (req->rq_status != 0) {
3888                         LBUG();
3889                         rep->lock_policy_res2 = req->rq_status;
3890                         RETURN(ELDLM_LOCK_ABORTED);
3891                 }
3892                 break;
3893         case IT_UNLINK:
3894                 rc = mds_lock_and_check_slave(offset, req, lockh);
3895                 if ((rep->lock_policy_res2 = rc)) {
3896                         if (rc == ENOLCK)
3897                                 rep->lock_policy_res2 = 0;
3898                         RETURN(ELDLM_LOCK_ABORTED);
3899                 }
3900                 break;
3901         default:
3902                 CERROR("Unhandled intent "LPD64"\n", it->opc);
3903                 LBUG();
3904         }
3905
3906         /* By this point, whatever function we called above must have either
3907          * filled in 'lockh', been an intent replay, or returned an error.  We
3908          * want to allow replayed RPCs to not get a lock, since we would just
3909          * drop it below anyways because lock replay is done separately by the
3910          * client afterwards.  For regular RPCs we want to give the new lock to
3911          * the client instead of whatever lock it was about to get. */
3912         if (new_lock == NULL)
3913                 new_lock = ldlm_handle2lock(&lockh[0]);
3914         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3915                 RETURN(0);
3916
3917         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3918                  it->opc, lockh[0].cookie);
3919
3920         /* If we've already given this lock to a client once, then we should
3921          * have no readers or writers.  Otherwise, we should have one reader
3922          * _or_ writer ref (which will be zeroed below) before returning the
3923          * lock to a client. */
3924         if (new_lock->l_export == req->rq_export) {
3925                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3926         } else {
3927                 LASSERT(new_lock->l_export == NULL);
3928                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3929         }
3930
3931         *lockp = new_lock;
3932
3933         if (new_lock->l_export == req->rq_export) {
3934                 /* Already gave this to the client, which means that we
3935                  * reconstructed a reply. */
3936                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3937                         MSG_RESENT);
3938                 RETURN(ELDLM_LOCK_REPLACED);
3939         }
3940
3941         /* Fixup the lock to be given to the client */
3942         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3943         new_lock->l_readers = 0;
3944         new_lock->l_writers = 0;
3945
3946         new_lock->l_export = class_export_get(req->rq_export);
3947         list_add(&new_lock->l_export_chain,
3948                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
3949
3950         new_lock->l_blocking_ast = lock->l_blocking_ast;
3951         new_lock->l_completion_ast = lock->l_completion_ast;
3952
3953         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3954                sizeof(lock->l_remote_handle));
3955
3956         new_lock->l_flags &= ~LDLM_FL_LOCAL;
3957
3958         LDLM_LOCK_PUT(new_lock);
3959         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3960
3961         RETURN(ELDLM_LOCK_REPLACED);
3962 }
3963
3964 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3965 {
3966         struct lprocfs_static_vars lvars;
3967         int rc = 0;
3968         struct mds_obd *mds = &dev->u.mds;
3969
3970         spin_lock_init(&mds->mds_denylist_lock);
3971         INIT_LIST_HEAD(&mds->mds_denylist);
3972
3973         lprocfs_init_multi_vars(0, &lvars);
3974
3975         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3976         if (rc)
3977                 return rc;
3978
3979         return lprocfs_alloc_md_stats(dev, 0);
3980 }
3981
3982 int mds_detach(struct obd_device *dev)
3983 {
3984         lprocfs_free_md_stats(dev);
3985         return lprocfs_obd_detach(dev);
3986 }
3987
3988 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3989 {
3990         struct lprocfs_static_vars lvars;
3991
3992         lprocfs_init_multi_vars(1, &lvars);
3993         return lprocfs_obd_attach(dev, lvars.obd_vars);
3994 }
3995
3996 int mdt_detach(struct obd_device *dev)
3997 {
3998         return lprocfs_obd_detach(dev);
3999 }
4000
4001 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4002 {
4003         struct mds_obd *mds = &obd->u.mds;
4004         int rc = 0;
4005         ENTRY;
4006
4007         mds->mds_service =
4008                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4009                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4010                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4011                                 mds_handle, "mds", obd->obd_proc_entry);
4012
4013         if (!mds->mds_service) {
4014                 CERROR("failed to start service\n");
4015                 RETURN(-ENOMEM);
4016         }
4017
4018         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4019                                     "ll_mdt");
4020         if (rc)
4021                 GOTO(err_thread, rc);
4022
4023         mds->mds_setattr_service =
4024                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4025                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4026                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4027                                 mds_handle, "mds_setattr",
4028                                 obd->obd_proc_entry);
4029         if (!mds->mds_setattr_service) {
4030                 CERROR("failed to start getattr service\n");
4031                 GOTO(err_thread, rc = -ENOMEM);
4032         }
4033
4034         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4035                                     MDT_NUM_THREADS, "ll_mdt_attr");
4036         if (rc)
4037                 GOTO(err_thread2, rc);
4038
4039         mds->mds_readpage_service =
4040                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4041                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4042                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4043                                 mds_handle, "mds_readpage",
4044                                 obd->obd_proc_entry);
4045         if (!mds->mds_readpage_service) {
4046                 CERROR("failed to start readpage service\n");
4047                 GOTO(err_thread2, rc = -ENOMEM);
4048         }
4049
4050         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4051                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
4052
4053         if (rc)
4054                 GOTO(err_thread3, rc);
4055
4056         RETURN(0);
4057
4058 err_thread3:
4059         ptlrpc_unregister_service(mds->mds_readpage_service);
4060 err_thread2:
4061         ptlrpc_unregister_service(mds->mds_setattr_service);
4062 err_thread:
4063         ptlrpc_unregister_service(mds->mds_service);
4064         return rc;
4065 }
4066
4067 static int mdt_cleanup(struct obd_device *obd, int flags)
4068 {
4069         struct mds_obd *mds = &obd->u.mds;
4070         ENTRY;
4071
4072         ptlrpc_stop_all_threads(mds->mds_readpage_service);
4073         ptlrpc_unregister_service(mds->mds_readpage_service);
4074
4075         ptlrpc_stop_all_threads(mds->mds_setattr_service);
4076         ptlrpc_unregister_service(mds->mds_setattr_service);
4077
4078         ptlrpc_stop_all_threads(mds->mds_service);
4079         ptlrpc_unregister_service(mds->mds_service);
4080
4081         RETURN(0);
4082 }
4083
4084 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4085                                          __u64 gr, void *data)
4086 {
4087         struct lustre_id id;
4088         struct obd_device *obd = data;
4089
4090         id_ino(&id) = ino;
4091         id_gen(&id) = gen;
4092         return mds_id2dentry(obd, &id, NULL);
4093 }
4094
4095 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4096                         void *key, __u32 *valsize, void *val)
4097 {
4098         struct obd_device *obd;
4099         struct mds_obd *mds;
4100         ENTRY;
4101
4102         obd = class_exp2obd(exp);
4103         mds = &obd->u.mds;
4104
4105         if (obd == NULL) {
4106                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4107                        exp->exp_handle.h_cookie);
4108                 RETURN(-EINVAL);
4109         }
4110
4111         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4112                 /* get log_context handle. */
4113                 unsigned long *llh_handle = val;
4114                 *valsize = sizeof(unsigned long);
4115                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
4116                 RETURN(0);
4117         }
4118         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4119                 /* get log_context handle. */
4120                 unsigned long *sb = val;
4121                 *valsize = sizeof(unsigned long);
4122                 *sb = (unsigned long)obd->u.mds.mds_sb;
4123                 RETURN(0);
4124         }
4125
4126         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4127                 __u32 *mdsize = val;
4128                 *valsize = sizeof(*mdsize);
4129                 *mdsize = mds->mds_max_mdsize;
4130                 RETURN(0);
4131         }
4132
4133         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4134                 __u32 *mdsnum = val;
4135                 *valsize = sizeof(*mdsnum);
4136                 *mdsnum = mds->mds_num;
4137                 RETURN(0);
4138         }
4139
4140         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4141                 struct lustre_id *rootid = val;
4142                 *valsize = sizeof(struct lustre_id);
4143                 *rootid = mds->mds_rootid;
4144                 RETURN(0);
4145         }
4146
4147         CDEBUG(D_IOCTL, "invalid key\n");
4148         RETURN(-EINVAL);
4149
4150 }
4151 struct lvfs_callback_ops mds_lvfs_ops = {
4152         l_id2dentry:     mds_lvfs_id2dentry,
4153 };
4154
4155 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4156                 int objcount, struct obd_ioobj *obj,
4157                 int niocount, struct niobuf_remote *nb,
4158                 struct niobuf_local *res,
4159                 struct obd_trans_info *oti);
4160
4161 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4162                  int objcount, struct obd_ioobj *obj, int niocount,
4163                  struct niobuf_local *res, struct obd_trans_info *oti,
4164                  int rc);
4165
4166 /* use obd ops to offer management infrastructure */
4167 static struct obd_ops mds_obd_ops = {
4168         .o_owner           = THIS_MODULE,
4169         .o_attach          = mds_attach,
4170         .o_detach          = mds_detach,
4171         .o_connect         = mds_connect,
4172         .o_connect_post    = mds_connect_post,
4173         .o_init_export     = mds_init_export,
4174         .o_destroy_export  = mds_destroy_export,
4175         .o_disconnect      = mds_disconnect,
4176         .o_setup           = mds_setup,
4177         .o_precleanup      = mds_precleanup,
4178         .o_cleanup         = mds_cleanup,
4179         .o_process_config  = mds_process_config,
4180         .o_postrecov       = mds_postrecov,
4181         .o_statfs          = mds_obd_statfs,
4182         .o_iocontrol       = mds_iocontrol,
4183         .o_create          = mds_obd_create,
4184         .o_destroy         = mds_obd_destroy,
4185         .o_llog_init       = mds_llog_init,
4186         .o_llog_finish     = mds_llog_finish,
4187         .o_notify          = mds_notify,
4188         .o_get_info        = mds_get_info,
4189         .o_set_info        = mds_set_info,
4190         .o_preprw          = mds_preprw,
4191         .o_commitrw        = mds_commitrw,
4192 };
4193
4194 static struct obd_ops mdt_obd_ops = {
4195         .o_owner           = THIS_MODULE,
4196         .o_attach          = mdt_attach,
4197         .o_detach          = mdt_detach,
4198         .o_setup           = mdt_setup,
4199         .o_cleanup         = mdt_cleanup,
4200 };
4201
4202 static int __init mds_init(void)
4203 {
4204         struct lprocfs_static_vars lvars;
4205
4206         mds_init_lsd_cache();
4207
4208         lprocfs_init_multi_vars(0, &lvars);
4209         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4210                             LUSTRE_MDS_NAME);
4211         lprocfs_init_multi_vars(1, &lvars);
4212         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4213                             LUSTRE_MDT_NAME);
4214
4215         return 0;
4216 }
4217
4218 static void /*__exit*/ mds_exit(void)
4219 {
4220         mds_cleanup_lsd_cache();
4221
4222         class_unregister_type(LUSTRE_MDS_NAME);
4223         class_unregister_type(LUSTRE_MDT_NAME);
4224 }
4225
4226 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4227 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4228 MODULE_LICENSE("GPL");
4229
4230 module_init(mds_init);
4231 module_exit(mds_exit);