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 = LDLM_FL_ATOMIC_CB, 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 = LDLM_FL_ATOMIC_CB;
 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         } else if (de->d_inode && de->d_inode->i_nlink == 0) {
 296                 /* as sometimes we lookup inode by ino/generation through
 297                    iopen mechanism, it's possible to find already unlinked
 298                    inode with nlink == 0. let's interpretate the case as
 299                    ENOENT -bzzz */
 300                 CWARN("found already unlinked inode %lu/%u\n",
 301                       de->d_inode->i_ino, de->d_inode->i_generation);
 302                 l_dput(de);
 303                 retval = ERR_PTR(-ENOENT);
 304                 ldlm_lock_decref(lockh, lock_mode);
 305 #ifdef S_PDIROPS
 306                 if (lockh[1].cookie)
 307                         ldlm_lock_decref(lockh + 1, *mode);
 308 #endif
 309         }
 310
 311         RETURN(retval);
 312 }
 313
 314 #ifndef DCACHE_DISCONNECTED
 315 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 316 #endif
 317
 318 /* Look up an entry by inode number. This function ONLY returns valid dget'd
 319  * dentries with an initialized inode or errors */
 320 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
 321                              struct vfsmount **mnt)
 322 {
 323         unsigned long ino = (unsigned long)id_ino(id);
 324         __u32 generation = (__u32)id_gen(id);
 325         struct mds_obd *mds = &obd->u.mds;
 326         struct dentry *result;
 327         struct inode *inode;
 328         char idname[32];
 329
 330         if (ino == 0)
 331                 RETURN(ERR_PTR(-ESTALE));
 332
 333         snprintf(idname, sizeof(idname), "0x%lx", ino);
 334
 335         CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
 336                ino, generation, mds->mds_sb);
 337
 338         /* under ext3 this is neither supposed to return bad inodes nor NULL
 339            inodes. */
 340         result = ll_lookup_one_len(idname, mds->mds_id_de,
 341                                    strlen(idname));
 342         if (IS_ERR(result))
 343                 RETURN(result);
 344
 345         inode = result->d_inode;
 346         if (!inode)
 347                 RETURN(ERR_PTR(-ENOENT));
 348
 349         if (is_bad_inode(inode)) {
 350                 CERROR("bad inode returned %lu/%u\n",
 351                        inode->i_ino, inode->i_generation);
 352                 dput(result);
 353                 RETURN(ERR_PTR(-ENOENT));
 354         }
 355
 356         /* here we disabled generation check, as root inode i_generation
 357          * of cache mds and real mds are different. */
 358         if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
 359             inode->i_generation != generation) {
 360                 /* we didn't find the right inode.. */
 361                 if (id_group(id) != mds->mds_num) {
 362                         CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
 363                                "%u != %u, mds %u != %u, request to wrong MDS?\n",
 364                                inode->i_ino, (unsigned long)inode->i_nlink,
 365                                atomic_read(&inode->i_count), inode->i_generation,
 366                                generation, mds->mds_num, (unsigned)id_group(id));
 367                 } else {
 368                         CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
 369                                "%u != %u, inode is recreated while request handled?\n",
 370                                inode->i_ino, (unsigned long)inode->i_nlink,
 371                                atomic_read(&inode->i_count), inode->i_generation,
 372                                generation);
 373                 }
 374                 dput(result);
 375                 RETURN(ERR_PTR(-ENOENT));
 376         }
 377
 378         if (mnt) {
 379                 *mnt = mds->mds_vfsmnt;
 380                 mntget(*mnt);
 381         }
 382
 383         RETURN(result);
 384 }
 385
 386 static
 387 int mds_req_add_idmapping(struct ptlrpc_request *req,
 388                           struct mds_export_data *med)
 389 {
 390         struct mds_req_sec_desc *rsd;
 391         struct lustre_sec_desc  *lsd;
 392         int rc;
 393
 394         if (!med->med_remote)
 395                 return 0;
 396
 397         /* maybe we should do it more completely: invalidate the gss ctxt? */
 398         if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
 399                 CWARN("didn't find mapped uid\n");
 400                 return -EPERM;
 401         }
 402
 403         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 404         if (!rsd) {
 405                 CERROR("Can't unpack security desc\n");
 406                 return -EPROTO;
 407         }
 408
 409         lsd = mds_get_lsd(req->rq_mapped_uid);
 410         if (!lsd) {
 411                 CERROR("can't get LSD(%u), no mapping added\n",
 412                        req->rq_mapped_uid);
 413                 return -EPERM;
 414         }
 415
 416         rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 417                            rsd->rsd_gid, lsd->lsd_gid);
 418         mds_put_lsd(lsd);
 419         return rc;
 420 }
 421
 422 static
 423 int mds_req_del_idmapping(struct ptlrpc_request *req,
 424                           struct mds_export_data *med)
 425 {
 426         struct mds_req_sec_desc *rsd;
 427         struct lustre_sec_desc  *lsd;
 428         int rc;
 429
 430         if (!med->med_remote)
 431                 return 0;
 432
 433         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 434         if (!rsd) {
 435                 CERROR("Can't unpack security desc\n");
 436                 return -EPROTO;
 437         }
 438
 439         LASSERT(req->rq_mapped_uid != -1);
 440         lsd = mds_get_lsd(req->rq_mapped_uid);
 441         if (!lsd) {
 442                 CERROR("can't get LSD(%u), no idmapping deleted\n",
 443                        req->rq_mapped_uid);
 444                 return -EPERM;
 445         }
 446
 447         rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 448                            rsd->rsd_gid, lsd->lsd_gid);
 449         mds_put_lsd(lsd);
 450         return rc;
 451 }
 452
 453 static int mds_init_export_data(struct ptlrpc_request *req,
 454                                 struct mds_export_data *med)
 455 {
 456         struct obd_connect_data *data, *reply;
 457         int ask_remote, ask_local;
 458         ENTRY;
 459
 460         data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
 461         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
 462         LASSERT(data && reply);
 463
 464         if (med->med_initialized) {
 465                 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
 466                 goto reply;
 467         }
 468
 469         ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
 470         ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
 471
 472         /* currently the policy is simple: satisfy client as possible
 473          * as we can.
 474          */
 475         if (req->rq_auth_uid == -1) {
 476                 if (ask_remote)
 477                         CWARN("null sec is used, force to be local\n");
 478                 med->med_remote = 0;
 479         } else {
 480                 if (ask_remote) {
 481                         if (!req->rq_remote_realm)
 482                                 CWARN("local realm asked to be remote\n");
 483                         med->med_remote = 1;
 484                 } else if (ask_local) {
 485                         if (req->rq_remote_realm)
 486                                 CWARN("remote realm asked to be local\n");
 487                         med->med_remote = 0;
 488                 } else
 489                         med->med_remote = (req->rq_remote_realm != 0);
 490         }
 491
 492         med->med_nllu = data->ocd_nllu[0];
 493         med->med_nllg = data->ocd_nllu[1];
 494
 495         med->med_initialized = 1;
 496 reply:
 497         reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
 498         if (med->med_remote) {
 499                 if (!med->med_idmap)
 500                         med->med_idmap = mds_idmap_alloc();
 501
 502                 if (!med->med_idmap)
 503                         CERROR("Failed to alloc idmap, following request from "
 504                                "this client will be refused\n");
 505
 506                 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
 507                 CDEBUG(D_SEC, "set client as remote\n");
 508         } else {
 509                 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
 510                 CDEBUG(D_SEC, "set client as local\n");
 511         }
 512
 513         RETURN(0);
 514 }
 515
 516 static void mds_free_export_data(struct mds_export_data *med)
 517 {
 518         if (!med->med_idmap)
 519                 return;
 520
 521         LASSERT(med->med_remote);
 522         mds_idmap_free(med->med_idmap);
 523         med->med_idmap = NULL;
 524 }
 525
 526 /* Establish a connection to the MDS.
 527  *
 528  * This will set up an export structure for the client to hold state data about
 529  * that client, like open files, the last operation number it did on the server,
 530  * etc.
 531  */
 532 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 533                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 534                        unsigned long flags)
 535 {
 536         struct mds_export_data *med;
 537         struct mds_client_data *mcd;
 538         struct obd_export *exp;
 539         int rc;
 540         ENTRY;
 541
 542         if (!conn || !obd || !cluuid)
 543                 RETURN(-EINVAL);
 544
 545         /* XXX There is a small race between checking the list and adding a new
 546          * connection for the same UUID, but the real threat (list corruption
 547          * when multiple different clients connect) is solved.
 548          *
 549          * There is a second race between adding the export to the list, and
 550          * filling in the client data below.  Hence skipping the case of NULL
 551          * mcd above.  We should already be controlling multiple connects at the
 552          * client, and we can't hold the spinlock over memory allocations
 553          * without risk of deadlocking.
 554          */
 555         rc = class_connect(conn, obd, cluuid);
 556         if (rc)
 557                 RETURN(rc);
 558         exp = class_conn2export(conn);
 559
 560         LASSERT(exp != NULL);
 561         med = &exp->exp_mds_data;
 562
 563         OBD_ALLOC(mcd, sizeof(*mcd));
 564         if (!mcd) {
 565                 CERROR("%s: out of memory for client data.\n",
 566                         obd->obd_name);
 567                 GOTO(out, rc = -ENOMEM);
 568         }
 569
 570         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 571         med->med_mcd = mcd;
 572
 573         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 574         if (rc)
 575                 GOTO(out, rc);
 576
 577         EXIT;
 578 out:
 579         if (rc) {
 580                 if (mcd)
 581                         OBD_FREE(mcd, sizeof(*mcd));
 582                 class_disconnect(exp, 0);
 583         } else {
 584                 class_export_put(exp);
 585         }
 586         return rc;
 587 }
 588
 589 static int mds_connect_post(struct obd_export *exp, unsigned initial,
 590                             unsigned long flags)
 591 {
 592         struct obd_device *obd = exp->exp_obd;
 593         struct mds_obd *mds = &obd->u.mds;
 594         struct mds_export_data *med;
 595         struct mds_client_data *mcd;
 596         int rc = 0;
 597         ENTRY;
 598
 599         med = &exp->exp_mds_data;
 600         mcd = med->med_mcd;
 601
 602         if (initial) {
 603                 /* some one reconnect initially, we have to reset
 604                  * data existing export can have. bug 6102 */
 605                 if (mcd->mcd_last_xid != 0)
 606                         CDEBUG(D_HA, "initial reconnect to existing export\n");
 607                 mcd->mcd_last_transno = 0;
 608                 mcd->mcd_last_xid = 0;
 609                 mcd->mcd_last_result = 0;
 610                 mcd->mcd_last_data = 0;
 611         }
 612
 613         if (!(flags & OBD_OPT_MDS_CONNECTION)) {
 614                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 615                         atomic_inc(&mds->mds_real_clients);
 616                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 617                                obd->obd_name, exp->exp_client_uuid.uuid,
 618                                atomic_read(&mds->mds_real_clients));
 619                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 620                 }
 621                 if (mds->mds_md_name)
 622                         rc = mds_md_connect(obd, mds->mds_md_name);
 623         }
 624         RETURN(rc);
 625 }
 626
 627 static int mds_init_export(struct obd_export *exp)
 628 {
 629         struct mds_export_data *med = &exp->exp_mds_data;
 630
 631         INIT_LIST_HEAD(&med->med_open_head);
 632         spin_lock_init(&med->med_open_lock);
 633         return 0;
 634 }
 635
 636 static int mds_destroy_export(struct obd_export *export)
 637 {
 638         struct obd_device *obd = export->exp_obd;
 639         struct mds_export_data *med = &export->exp_mds_data;
 640         struct lvfs_run_ctxt saved;
 641         int rc = 0;
 642         ENTRY;
 643
 644         mds_free_export_data(med);
 645         target_destroy_export(export);
 646
 647         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 648                 GOTO(out, 0);
 649
 650         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 651
 652         /* Close any open files (which may also cause orphan unlinking). */
 653         spin_lock(&med->med_open_lock);
 654         while (!list_empty(&med->med_open_head)) {
 655                 struct list_head *tmp = med->med_open_head.next;
 656                 struct mds_file_data *mfd =
 657                         list_entry(tmp, struct mds_file_data, mfd_list);
 658                 struct lustre_id sid;
 659
 660                 BDEVNAME_DECLARE_STORAGE(btmp);
 661
 662                 /* bug 1579: fix force-closing for 2.5 */
 663                 struct dentry *dentry = mfd->mfd_dentry;
 664
 665                 list_del(&mfd->mfd_list);
 666                 spin_unlock(&med->med_open_lock);
 667
 668                 down(&dentry->d_inode->i_sem);
 669                 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
 670                 up(&dentry->d_inode->i_sem);
 671                 if (rc) {
 672                         CERROR("Can't read inode self id, inode %lu, "
 673                                "rc %d\n", dentry->d_inode->i_ino, rc);
 674                         memset(&sid, 0, sizeof(sid));
 675                 }
 676
 677                 /* If you change this message, be sure to update
 678                  * replay_single:test_46 */
 679                 CERROR("force closing client file handle for %.*s (%s:"
 680                        DLID4")\n", dentry->d_name.len, dentry->d_name.name,
 681                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 682                        OLID4(&sid));
 683
 684                 /* child inode->i_alloc_sem protects orphan_dec_test and
 685                  * is_orphan race, mds_mfd_close drops it */
 686                 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 687                 rc = mds_mfd_close(NULL, 0, obd, mfd,
 688                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 689                 if (rc)
 690                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 691                 spin_lock(&med->med_open_lock);
 692         }
 693         spin_unlock(&med->med_open_lock);
 694         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 695
 696         EXIT;
 697 out:
 698         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 699         return rc;
 700 }
 701
 702 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
 703 {
 704         unsigned long irqflags;
 705         struct obd_device *obd;
 706         struct mds_obd *mds;
 707         int rc;
 708         ENTRY;
 709
 710         LASSERT(exp != NULL);
 711         obd = class_exp2obd(exp);
 712         if (obd == NULL) {
 713                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 714                        exp->exp_handle.h_cookie);
 715                 RETURN(-EINVAL);
 716         }
 717         mds = &obd->u.mds;
 718
 719         /*
 720          * suppress any inter-mds requests durring disconnecting lmv if this is
 721          * detected --force mode. This is needed to avoid endless recovery.
 722          */
 723         if (atomic_read(&mds->mds_real_clients) > 0 &&
 724             !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
 725                 flags |= OBD_OPT_FORCE;
 726
 727         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 728             && !atomic_read(&mds->mds_real_clients)) {
 729                 /* there was no client at all */
 730                 mds_md_disconnect(obd, flags);
 731         }
 732
 733         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 734             && atomic_dec_and_test(&mds->mds_real_clients)) {
 735                 /* time to drop LMV connections */
 736                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 737                        "Disconnnect from LMV now\n",
 738                        obd->obd_name, exp->exp_client_uuid.uuid);
 739                 mds_md_disconnect(obd, flags);
 740         }
 741
 742         spin_lock_irqsave(&exp->exp_lock, irqflags);
 743         exp->exp_flags = flags;
 744         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 745
 746         /* disconnect early so that clients can't keep using export */
 747         rc = class_disconnect(exp, flags);
 748         ldlm_cancel_locks_for_export(exp);
 749
 750         /* complete all outstanding replies */
 751         spin_lock_irqsave(&exp->exp_lock, irqflags);
 752         while (!list_empty(&exp->exp_outstanding_replies)) {
 753                 struct ptlrpc_reply_state *rs =
 754                         list_entry(exp->exp_outstanding_replies.next,
 755                                    struct ptlrpc_reply_state, rs_exp_list);
 756                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 757
 758                 spin_lock(&svc->srv_lock);
 759                 list_del_init(&rs->rs_exp_list);
 760                 ptlrpc_schedule_difficult_reply(rs);
 761                 spin_unlock(&svc->srv_lock);
 762         }
 763         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 764         RETURN(rc);
 765 }
 766
 767 static int mds_getstatus(struct ptlrpc_request *req)
 768 {
 769         struct mds_obd *mds = mds_req2mds(req);
 770         struct mds_body *body;
 771         int rc, size;
 772         ENTRY;
 773
 774         size = sizeof(*body);
 775
 776         rc = lustre_pack_reply(req, 1, &size, NULL);
 777         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 778                 CERROR("mds: out of memory for message: size=%d\n", size);
 779                 req->rq_status = -ENOMEM;       /* superfluous? */
 780                 RETURN(-ENOMEM);
 781         }
 782
 783         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 784         body->valid |= OBD_MD_FID;
 785
 786         memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
 787
 788         /*
 789          * the last_committed and last_xid fields are filled in for all replies
 790          * already - no need to do so here also.
 791          */
 792         RETURN(0);
 793 }
 794
 795 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 796                      void *data, int flag)
 797 {
 798         int do_ast;
 799         ENTRY;
 800
 801         if (flag == LDLM_CB_CANCELING) {
 802                 /* Don't need to do anything here. */
 803                 RETURN(0);
 804         }
 805
 806         /* XXX layering violation!  -phil */
 807         lock_res(lock->l_resource);
 808
 809         /*
 810          * get this: if mds_blocking_ast is racing with mds_intent_policy, such
 811          * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
 812          * then by the time we get the lock, we might not be the correct
 813          * blocking function anymore.  So check, and return early, if so.
 814          */
 815         if (lock->l_blocking_ast != mds_blocking_ast) {
 816                 unlock_res(lock->l_resource);
 817                 RETURN(0);
 818         }
 819
 820         lock->l_flags |= LDLM_FL_CBPENDING;
 821         do_ast = (!lock->l_readers && !lock->l_writers);
 822         unlock_res(lock->l_resource);
 823
 824         if (do_ast) {
 825                 struct lustre_handle lockh;
 826                 int rc;
 827
 828                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 829                 ldlm_lock2handle(lock, &lockh);
 830                 rc = ldlm_cli_cancel(&lockh);
 831                 if (rc < 0)
 832                         CERROR("ldlm_cli_cancel: %d\n", rc);
 833         } else {
 834                 LDLM_DEBUG(lock, "Lock still has references, will be "
 835                            "cancelled later");
 836         }
 837         RETURN(0);
 838 }
 839
 840 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
 841                           void *md, int size, int mea)
 842 {
 843         int rc = size;
 844
 845         if (S_ISREG(inode->i_mode)) {
 846                 rc = mds_convert_lov_ea(obd, inode, md, size);
 847         } else if (S_ISDIR(inode->i_mode)) {
 848                 if (mea) {
 849                         rc = mds_convert_mea_ea(obd, inode, md, size);
 850                 } else {
 851                         rc = mds_convert_lov_ea(obd, inode, md, size);
 852                 }
 853                 if (rc == -EINVAL) {
 854                         CERROR("Invalid EA format (nor LOV or MEA) "
 855                                "is detected. Inode %lu/%u\n",
 856                                inode->i_ino, inode->i_generation);
 857                 }
 858         }
 859         return rc;
 860 }
 861
 862 int mds_get_md(struct obd_device *obd, struct inode *inode,
 863                void *md, int *size, int lock, int mea)
 864 {
 865         int lmm_size;
 866         int rc = 0;
 867         ENTRY;
 868
 869         if (lock)
 870                 down(&inode->i_sem);
 871
 872         rc = fsfilt_get_md(obd, inode, md, *size,
 873                            (mea ? EA_MEA : EA_LOV));
 874         if (rc < 0) {
 875                 CERROR("Error %d reading eadata for ino %lu\n",
 876                        rc, inode->i_ino);
 877         } else if (rc > 0) {
 878                 lmm_size = rc;
 879                 rc = mds_convert_md(obd, inode, md,
 880                                     lmm_size, mea);
 881                 if (rc == 0) {
 882                         *size = lmm_size;
 883                         rc = lmm_size;
 884                 } else if (rc > 0) {
 885                         *size = rc;
 886                 }
 887         }
 888         if (lock)
 889                 up(&inode->i_sem);
 890
 891         RETURN(rc);
 892 }
 893
 894 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 895  * Call with lock=0 if the caller has already taken the i_sem. */
 896 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 897                 struct mds_body *body, struct inode *inode, int lock, int mea)
 898 {
 899         struct mds_obd *mds = &obd->u.mds;
 900         int rc, lmm_size;
 901         void *lmm;
 902         ENTRY;
 903
 904         lmm = lustre_msg_buf(msg, offset, 0);
 905         if (lmm == NULL) {
 906                 /* Some problem with getting eadata when I sized the reply
 907                  * buffer... */
 908                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 909                        inode->i_ino);
 910                 RETURN(0);
 911         }
 912         lmm_size = msg->buflens[offset];
 913
 914         /* I don't really like this, but it is a sanity check on the client
 915          * MD request.  However, if the client doesn't know how much space
 916          * to reserve for the MD, it shouldn't be bad to have too much space.
 917          */
 918         if (lmm_size > mds->mds_max_mdsize) {
 919                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 920                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 921                 // RETURN(-EINVAL);
 922         }
 923
 924         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
 925         if (rc > 0) {
 926                 body->valid |= S_ISDIR(inode->i_mode) ?
 927                         OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
 928
 929                 if (mea)
 930                         body->valid |= OBD_MD_MEA;
 931
 932                 body->eadatasize = lmm_size;
 933                 rc = 0;
 934         }
 935
 936         RETURN(rc);
 937 }
 938
 939 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
 940                   struct mds_body *repbody, int reply_off)
 941 {
 942         struct inode *inode = dentry->d_inode;
 943         char *symname;
 944         int len, rc;
 945         ENTRY;
 946
 947         symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 948         LASSERT(symname != NULL);
 949         len = req->rq_repmsg->buflens[reply_off + 1];
 950
 951         rc = inode->i_op->readlink(dentry, symname, len);
 952         if (rc < 0) {
 953                 CERROR("readlink failed: %d\n", rc);
 954         } else if (rc != len - 1) {
 955                 CERROR ("Unexpected readlink rc %d: expecting %d\n",
 956                         rc, len - 1);
 957                 rc = -EINVAL;
 958         } else {
 959                 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 960                 repbody->valid |= OBD_MD_LINKNAME;
 961                 repbody->eadatasize = rc + 1;
 962                 symname[rc] = 0;        /* NULL terminate */
 963                 rc = 0;
 964         }
 965
 966         RETURN(rc);
 967 }
 968
 969 int mds_pack_ea(struct dentry *dentry, struct ptlrpc_request *req,
 970                 struct mds_body *repbody, int req_off, int reply_off)
 971 {
 972         struct inode *inode = dentry->d_inode;
 973         char *ea_name;
 974         void *value = NULL;
 975         int len, rc;
 976         ENTRY;
 977
 978         ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
 979         len = req->rq_repmsg->buflens[reply_off + 1];
 980         if (len != 0)
 981                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 982
 983         rc = -EOPNOTSUPP;
 984         if (inode->i_op && inode->i_op->getxattr)
 985                 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
 986
 987         if (rc < 0) {
 988                 if (rc != -ENODATA && rc != -EOPNOTSUPP)
 989                         CERROR("getxattr failed: %d", rc);
 990         } else {
 991                 repbody->valid |= OBD_MD_FLEA;
 992                 repbody->eadatasize = rc;
 993                 rc = 0;
 994         }
 995
 996         RETURN(rc);
 997 }
 998
 999 int mds_pack_ealist(struct dentry *dentry, struct ptlrpc_request *req,
1000                     struct mds_body *repbody, int reply_off)
1001 {
1002         struct inode *inode = dentry->d_inode;
1003         void *value = NULL;
1004         int len, rc;
1005         ENTRY;
1006
1007         len = req->rq_repmsg->buflens[reply_off + 1];
1008         if (len != 0)
1009                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1010
1011         rc = -EOPNOTSUPP;
1012         if (inode->i_op && inode->i_op->getxattr)
1013                 rc = inode->i_op->listxattr(dentry, value, len);
1014
1015         if (rc < 0) {
1016                 CERROR("listxattr failed: %d", rc);
1017         } else {
1018                 repbody->valid |= OBD_MD_FLEALIST;
1019                 repbody->eadatasize = rc;
1020                 rc = 0;
1021         }
1022         RETURN(rc);
1023 }
1024
1025 int mds_pack_acl(struct obd_device *obd, struct lustre_msg *repmsg, int offset,
1026                  struct mds_body *body, struct inode *inode)
1027 {
1028         struct dentry de = { .d_inode = inode };
1029         __u32 buflen, *sizep;
1030         void *buf;
1031         int size;
1032         ENTRY;
1033
1034         if (!inode->i_op->getxattr)
1035                 RETURN(0);
1036
1037         buflen = repmsg->buflens[offset + 1];
1038         buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1039
1040         size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1041         if (size == -ENODATA || size == -EOPNOTSUPP)
1042                 RETURN(0);
1043         if (size < 0)
1044                 RETURN(size);
1045         LASSERT(size);
1046
1047         sizep = lustre_msg_buf(repmsg, offset, 4);
1048         if (!sizep) {
1049                 CERROR("can't locate returned acl size buf\n");
1050                 RETURN(-EPROTO);
1051         }
1052
1053         *sizep = cpu_to_le32(size);
1054         body->valid |= OBD_MD_FLACL_ACCESS;
1055
1056         RETURN(0);
1057 }
1058
1059 /*
1060  * here we take simple rule: once uid/fsuid is root, we also squash
1061  * the gid/fsgid, don't care setuid/setgid attributes.
1062  */
1063 int mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
1064                     ptl_nid_t *peernid)
1065 {
1066         if (!mds->mds_squash_uid || *peernid == mds->mds_nosquash_nid)
1067                 return 0;
1068
1069         if (rsd->rsd_uid && rsd->rsd_fsuid)
1070                 return 0;
1071
1072         CDEBUG(D_SEC, "squash req from "LPX64":"
1073                "(%u:%u-%u:%u/%x)=>(%u:%u-%u:%u/%x)\n", *peernid,
1074                 rsd->rsd_uid, rsd->rsd_gid,
1075                 rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
1076                 rsd->rsd_uid ? rsd->rsd_uid : mds->mds_squash_uid,
1077                 rsd->rsd_uid ? rsd->rsd_gid : mds->mds_squash_gid,
1078                 rsd->rsd_fsuid ? rsd->rsd_fsuid : mds->mds_squash_uid,
1079                 rsd->rsd_fsuid ? rsd->rsd_fsgid : mds->mds_squash_gid,
1080                 rsd->rsd_cap & ~CAP_FS_MASK);
1081
1082         if (rsd->rsd_uid == 0) {
1083                 rsd->rsd_uid = mds->mds_squash_uid;
1084                 rsd->rsd_gid = mds->mds_squash_gid;
1085         }
1086         if (rsd->rsd_fsuid == 0) {
1087                 rsd->rsd_fsuid = mds->mds_squash_uid;
1088                 rsd->rsd_fsgid = mds->mds_squash_gid;
1089         }
1090         rsd->rsd_cap &= ~CAP_FS_MASK;
1091
1092         return 1;
1093 }
1094
1095 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1096                                 struct ptlrpc_request *req, int req_off,
1097                                 struct mds_body *reqbody, int reply_off)
1098 {
1099         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1100         struct inode *inode = dentry->d_inode;
1101         struct mds_body *body;
1102         int rc = 0;
1103         ENTRY;
1104
1105         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1106                 RETURN(-ENOENT);
1107
1108         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1109         LASSERT(body != NULL);                 /* caller prepped reply */
1110
1111         if (dentry->d_flags & DCACHE_CROSS_REF) {
1112                 mds_pack_dentry2body(obd, body, dentry,
1113                                      (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1114                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1115                        OLID4(&body->id1));
1116                 RETURN(0);
1117         }
1118
1119         mds_pack_inode2body(obd, body, inode,
1120                             (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1121
1122         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1123             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1124
1125                 /* guessing what kind og attribute do we need. */
1126                 int is_mea = (S_ISDIR(inode->i_mode) &&
1127                     (reqbody->valid & OBD_MD_MEA) != 0);
1128
1129                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1130                                  body, inode, 1, is_mea);
1131
1132                 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1133                 if (!(body->valid & OBD_MD_FLEASIZE) &&
1134                     !(body->valid & OBD_MD_FLDIREA))
1135                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1136                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1137         } else if (S_ISLNK(inode->i_mode) &&
1138                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1139                 rc = mds_pack_link(dentry, req, body, reply_off);
1140         } else if (reqbody->valid & OBD_MD_FLEA) {
1141                 rc = mds_pack_ea(dentry, req, body, req_off, reply_off);
1142         } else if (reqbody->valid & OBD_MD_FLEALIST) {
1143                 rc = mds_pack_ealist(dentry, req, body, reply_off);
1144         }
1145
1146         if (reqbody->valid & OBD_MD_FLACL_ACCESS) {
1147                 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1148                 rc = mds_pack_acl(obd, req->rq_repmsg, reply_off + inc,
1149                                   body, inode);
1150         }
1151
1152         if (rc == 0)
1153                 mds_body_do_reverse_map(med, body);
1154
1155         RETURN(rc);
1156 }
1157
1158 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1159                                    struct dentry *dentry,
1160                                    int offset)
1161 {
1162         int rc = 0, size[1] = {sizeof(struct mds_body)};
1163         ENTRY;
1164
1165         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1166                 CERROR("failed MDS_GETATTR_PACK test\n");
1167                 req->rq_status = -ENOMEM;
1168                 RETURN(-ENOMEM);
1169         }
1170
1171         rc = lustre_pack_reply(req, 1, size, NULL);
1172         if (rc) {
1173                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1174                 GOTO(out, req->rq_status = rc);
1175         }
1176
1177         EXIT;
1178 out:
1179         return rc;
1180 }
1181
1182 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1183                                 int offset)
1184 {
1185         struct inode *inode = de->d_inode;
1186         struct mds_obd *mds = mds_req2mds(req);
1187         struct mds_body *body;
1188         int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1189         ENTRY;
1190
1191         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1192         LASSERT(body != NULL);                 /* checked by caller */
1193         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
1194
1195         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1196             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1197                 int rc;
1198
1199                 down(&inode->i_sem);
1200                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1201                                    ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1202                 up(&inode->i_sem);
1203                 if (rc < 0) {
1204                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1205                                 CERROR("error getting inode %lu MD: rc = %d\n",
1206                                        inode->i_ino, rc);
1207                         size[bufcount] = 0;
1208                 } else if (rc > mds->mds_max_mdsize) {
1209                         size[bufcount] = 0;
1210                         CERROR("MD size %d larger than maximum possible %u\n",
1211                                rc, mds->mds_max_mdsize);
1212                 } else {
1213                         size[bufcount] = rc;
1214                 }
1215                 bufcount++;
1216         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1217                 if (inode->i_size + 1 != body->eadatasize)
1218                         CERROR("symlink size: %Lu, reply space: %d\n",
1219                                inode->i_size + 1, body->eadatasize);
1220                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1221                 bufcount++;
1222                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1223                        inode->i_size + 1, body->eadatasize);
1224         } else if ((body->valid & OBD_MD_FLEA)) {
1225                 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1226                                                   offset + 1, 0);
1227                 rc = -EOPNOTSUPP;
1228                 if (inode->i_op && inode->i_op->getxattr)
1229                         rc = inode->i_op->getxattr(de, ea_name, NULL, 0);
1230
1231                 if (rc < 0) {
1232                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1233                                 CERROR("error getting inode %lu EA: rc = %d\n",
1234                                        inode->i_ino, rc);
1235                         size[bufcount] = 0;
1236                 } else {
1237                         size[bufcount] = min_t(int, body->eadatasize, rc);
1238                 }
1239                 bufcount++;
1240         } else if (body->valid & OBD_MD_FLEALIST) {
1241                 rc = -EOPNOTSUPP;
1242                 if (inode->i_op && inode->i_op->getxattr)
1243                         rc = inode->i_op->listxattr(de, NULL, 0);
1244
1245                 if (rc < 0) {
1246                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1247                                 CERROR("error getting inode %lu EA: rc = %d\n",
1248                                        inode->i_ino, rc);
1249                         size[bufcount] = 0;
1250                 } else {
1251                         size[bufcount] = min_t(int, body->eadatasize, rc);
1252                 }
1253                 bufcount++;
1254         }
1255
1256         /* may co-exist with OBD_MD_FLEASIZE */
1257         if (body->valid & OBD_MD_FLACL_ACCESS) {
1258                 size[bufcount++] = 4;
1259                 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1260         }
1261
1262         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1263                 CERROR("failed MDS_GETATTR_PACK test\n");
1264                 req->rq_status = -ENOMEM;
1265                 GOTO(out, rc = -ENOMEM);
1266         }
1267
1268         rc = lustre_pack_reply(req, bufcount, size, NULL);
1269         if (rc) {
1270                 CERROR("out of memory\n");
1271                 GOTO(out, req->rq_status = rc);
1272         }
1273
1274         EXIT;
1275  out:
1276         return rc;
1277 }
1278
1279 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1280                       char *name, int namelen)
1281 {
1282         struct mea *mea = NULL;
1283         int mea_size, rc = 0;
1284         ENTRY;
1285
1286         rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1287         if (rc)
1288                 RETURN(rc);
1289         if (mea != NULL) {
1290                 /*
1291                  * dir is already splitted, check if requested filename should
1292                  * live at this MDS or at another one.
1293                  */
1294                 int i = mea_name2idx(mea, name, namelen - 1);
1295                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1296                         CDEBUG(D_OTHER,
1297                                "inapropriate MDS(%d) for %s. should be "
1298                                "%lu(%d)\n", mea->mea_master, name,
1299                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1300                         rc = -ERESTART;
1301                 }
1302         }
1303
1304         if (mea)
1305                 OBD_FREE(mea, mea_size);
1306         RETURN(rc);
1307 }
1308
1309 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1310                      struct ptlrpc_request *req, struct mds_body *body)
1311 {
1312         struct inode *inode = dentry->d_inode;
1313         ENTRY;
1314
1315         LASSERT(body != NULL);
1316
1317         if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1318                 RETURN(0);
1319
1320         if (obd->obd_recovering) {
1321                 CDEBUG(D_ERROR, "size for "DLID4" is unknown yet (recovering)\n",
1322                        OLID4(&body->id1));
1323                 RETURN(0);
1324         }
1325
1326         if (atomic_read(&inode->i_writecount)) {
1327                 /* some one has opened the file for write.
1328                  * mds doesn't know actual size */
1329                 CDEBUG(D_OTHER, "MDS doesn't know actual size for "DLID4"\n",
1330                        OLID4(&body->id1));
1331                 RETURN(0);
1332         }
1333         CDEBUG(D_OTHER, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1334                body->size, body->blocks, OLID4(&body->id1));
1335         body->valid |= OBD_MD_FLSIZE;
1336         RETURN(0);
1337 }
1338
1339 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1340                             struct lustre_handle *child_lockh, int child_part)
1341 {
1342         struct obd_device *obd = req->rq_export->exp_obd;
1343         struct mds_obd *mds = &obd->u.mds;
1344         struct ldlm_reply *rep = NULL;
1345         struct lvfs_run_ctxt saved;
1346         struct mds_req_sec_desc *rsd;
1347         struct mds_body *body;
1348         struct dentry *dparent = NULL, *dchild = NULL;
1349         struct lvfs_ucred uc = {NULL, NULL,};
1350         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1351         unsigned int namesize;
1352         int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1353         char *name = NULL;
1354         ENTRY;
1355
1356         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
1357         MD_COUNTER_INCREMENT(obd, getattr_lock);
1358
1359         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1360         if (!rsd) {
1361                 CERROR("Can't unpack security desc\n");
1362                 RETURN(-EFAULT);
1363         }
1364
1365         /* swab now, before anyone looks inside the request. */
1366         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1367                                   lustre_swab_mds_body);
1368         if (body == NULL) {
1369                 CERROR("Can't swab mds_body\n");
1370                 GOTO(cleanup, rc = -EFAULT);
1371         }
1372
1373         LASSERT_REQSWAB(req, offset + 1);
1374         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1375         if (name == NULL) {
1376                 CERROR("Can't unpack name\n");
1377                 GOTO(cleanup, rc = -EFAULT);
1378         }
1379         namesize = req->rq_reqmsg->buflens[offset + 1];
1380
1381         /* namesize less than 2 means we have empty name, probably came from
1382            revalidate by cfid, so no point in having name to be set */
1383         if (namesize <= 1)
1384                 name = NULL;
1385
1386         LASSERT (offset == 1 || offset == 3);
1387         if (offset == 3) {
1388                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1389                 reply_offset = 1;
1390         } else {
1391                 reply_offset = 0;
1392         }
1393
1394         rc = mds_init_ucred(&uc, req, rsd);
1395         if (rc) {
1396                 GOTO(cleanup, rc);
1397         }
1398
1399         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1400         cleanup_phase = 1; /* kernel context */
1401         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1402
1403         LASSERT(namesize > 0);
1404         if (child_lockh->cookie != 0) {
1405                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1406                 resent_req = 1;
1407         }
1408 #if HAVE_LOOKUP_RAW
1409         if (body->valid == OBD_MD_FLID) {
1410                 struct mds_body *mds_reply;
1411                 int size = sizeof(*mds_reply);
1412                 struct inode *dir;
1413                 ino_t inum;
1414
1415                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1416                 if (IS_ERR(dparent)) {
1417                         rc = PTR_ERR(dparent);
1418                         GOTO(cleanup, rc);
1419                 }
1420                 /*
1421                  * the user requested ONLY the inode number, so do a raw lookup.
1422                  */
1423                 rc = lustre_pack_reply(req, 1, &size, NULL);
1424                 if (rc) {
1425                         CERROR("out of memory\n");
1426                         l_dput(dparent);
1427                         GOTO(cleanup, rc);
1428                 }
1429                 dir  = dparent->d_inode;
1430                 LASSERT(dir->i_op->lookup_raw != NULL);
1431                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1432                 l_dput(dparent);
1433                 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1434                                            sizeof(*mds_reply));
1435
1436                 id_ino(&mds_reply->id1) = inum;
1437                 mds_reply->valid = OBD_MD_FLID;
1438                 GOTO(cleanup, rc);
1439         }
1440 #endif
1441         if (resent_req == 0) {
1442                 LASSERT(id_fid(&body->id1) != 0);
1443                 if (name) {
1444                         rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1445                                                          parent_lockh, &dparent,
1446                                                          LCK_PR,
1447                                                          MDS_INODELOCK_UPDATE,
1448                                                          &update_mode,
1449                                                          name, namesize,
1450                                                          child_lockh, &dchild,
1451                                                          LCK_PR, child_part);
1452                         if (rc)
1453                                 GOTO(cleanup, rc);
1454
1455                         cleanup_phase = 2; /* dchild, dparent, locks */
1456
1457                         /*
1458                          * let's make sure this name should leave on this mds
1459                          * node.
1460                          */
1461                         rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1462                         if (rc)
1463                                 GOTO(cleanup, rc);
1464                 } else {
1465                         /* we have no dentry here, drop LOOKUP bit */
1466                         /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1467                         child_part &= ~MDS_INODELOCK_LOOKUP;
1468                         CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1469                                obd->obd_name, OLID4(&body->id1));
1470
1471                         dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1472                                                       LCK_PR, parent_lockh,
1473                                                       &update_mode,
1474                                                       NULL, 0,
1475                                                       MDS_INODELOCK_UPDATE);
1476                         if (IS_ERR(dchild)) {
1477                                 CERROR("can't find inode with id "DLID4", err = %d\n",
1478                                        OLID4(&body->id1), (int)PTR_ERR(dchild));
1479                                 GOTO(cleanup, rc = PTR_ERR(dchild));
1480                         }
1481                         memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1482                 }
1483         } else {
1484                 struct ldlm_lock *granted_lock;
1485
1486                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1487                 granted_lock = ldlm_handle2lock(child_lockh);
1488
1489                 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1490                          id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1491                          child_lockh->cookie);
1492
1493                 if (name) {
1494                         /* usual named request */
1495                         dparent = mds_id2dentry(obd, &body->id1, NULL);
1496                         LASSERT(!IS_ERR(dparent));
1497                         dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1498                         LASSERT(!IS_ERR(dchild));
1499                 } else {
1500                         /* client wants to get attr. by id */
1501                         dchild = mds_id2dentry(obd, &body->id1, NULL);
1502                         LASSERT(!IS_ERR(dchild));
1503                 }
1504                 LDLM_LOCK_PUT(granted_lock);
1505         }
1506
1507         cleanup_phase = 2; /* dchild, dparent, locks */
1508
1509         if (!DENTRY_VALID(dchild)) {
1510                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1511                 /*
1512                  * in the intent case, the policy clears this error: the
1513                  * disposition is enough.
1514                  */
1515                 rc = -ENOENT;
1516                 GOTO(cleanup, rc);
1517         } else {
1518                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1519         }
1520
1521         if (req->rq_repmsg == NULL) {
1522                 if (dchild->d_flags & DCACHE_CROSS_REF)
1523                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1524                 else
1525                         rc = mds_getattr_pack_msg(req, dchild, offset);
1526                 if (rc != 0) {
1527                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1528                         GOTO (cleanup, rc);
1529                 }
1530         }
1531
1532         rc = mds_getattr_internal(obd, dchild, req, offset, body, reply_offset);
1533         if (rc)
1534                 GOTO(cleanup, rc); /* returns the lock to the client */
1535
1536         /* probably MDS knows actual size? */
1537         body = lustre_msg_buf(req->rq_repmsg, reply_offset, sizeof(*body));
1538         LASSERT(body != NULL);
1539         mds_getattr_size(obd, dchild, req, body);
1540
1541         GOTO(cleanup, rc);
1542
1543  cleanup:
1544         switch (cleanup_phase) {
1545         case 2:
1546                 if (resent_req == 0) {
1547                         if (rc && DENTRY_VALID(dchild))
1548                                 ldlm_lock_decref(child_lockh, LCK_PR);
1549                         if (name)
1550                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1551 #ifdef S_PDIROPS
1552                         if (parent_lockh[1].cookie != 0)
1553                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1554 #endif
1555                         if (dparent)
1556                                 l_dput(dparent);
1557                 }
1558                 l_dput(dchild);
1559         case 1:
1560                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1561         default:
1562                 mds_exit_ucred(&uc);
1563         }
1564         return rc;
1565 }
1566
1567 static int mds_getattr(struct ptlrpc_request *req, int offset)
1568 {
1569         struct obd_device *obd = req->rq_export->exp_obd;
1570         struct lvfs_run_ctxt saved;
1571         struct dentry *de;
1572         struct mds_req_sec_desc *rsd;
1573         struct mds_body *body;
1574         struct lvfs_ucred uc = {NULL, NULL,};
1575         int rc = 0;
1576         ENTRY;
1577
1578         MD_COUNTER_INCREMENT(obd, getattr);
1579
1580         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1581         if (!rsd) {
1582                 CERROR("Can't unpack security desc\n");
1583                 RETURN(-EFAULT);
1584         }
1585
1586         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1587                                   lustre_swab_mds_body);
1588         if (body == NULL) {
1589                 CERROR ("Can't unpack body\n");
1590                 RETURN (-EFAULT);
1591         }
1592
1593         rc = mds_init_ucred(&uc, req, rsd);
1594         if (rc) {
1595                 mds_exit_ucred(&uc);
1596                 RETURN(rc);
1597         }
1598
1599         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1600         de = mds_id2dentry(obd, &body->id1, NULL);
1601         if (IS_ERR(de)) {
1602                 rc = req->rq_status = PTR_ERR(de);
1603                 GOTO(out_pop, rc);
1604         }
1605
1606         rc = mds_getattr_pack_msg(req, de, offset);
1607         if (rc != 0) {
1608                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1609                 GOTO(out_pop, rc);
1610         }
1611
1612         req->rq_status = mds_getattr_internal(obd, de, req, offset, body, 0);
1613         l_dput(de);
1614
1615         EXIT;
1616 out_pop:
1617         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1618         mds_exit_ucred(&uc);
1619         return rc;
1620 }
1621
1622 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1623                           unsigned long max_age)
1624 {
1625         int rc;
1626         ENTRY;
1627
1628         spin_lock(&obd->obd_osfs_lock);
1629         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1630         if (rc == 0)
1631                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1632         spin_unlock(&obd->obd_osfs_lock);
1633
1634         RETURN(rc);
1635 }
1636
1637 static int mds_statfs(struct ptlrpc_request *req)
1638 {
1639         struct obd_device *obd = req->rq_export->exp_obd;
1640         int rc, size = sizeof(struct obd_statfs);
1641         ENTRY;
1642
1643         /* This will trigger a watchdog timeout */
1644         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1645                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1646
1647         rc = lustre_pack_reply(req, 1, &size, NULL);
1648         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1649                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1650                 GOTO(out, rc);
1651         }
1652
1653         OBD_COUNTER_INCREMENT(obd, statfs);
1654
1655         /* We call this so that we can cache a bit - 1 jiffie worth */
1656         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1657                             jiffies - HZ);
1658         if (rc) {
1659                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1660                 GOTO(out, rc);
1661         }
1662
1663         EXIT;
1664 out:
1665         req->rq_status = rc;
1666         return rc;
1667 }
1668
1669 static int mds_sync(struct ptlrpc_request *req, int offset)
1670 {
1671         struct obd_device *obd = req->rq_export->exp_obd;
1672         struct mds_obd *mds = &obd->u.mds;
1673         struct mds_body *body;
1674         int rc, size = sizeof(*body);
1675         ENTRY;
1676
1677         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1678                                   lustre_swab_mds_body);
1679         if (body == NULL)
1680                 GOTO(out, rc = -EPROTO);
1681
1682         rc = lustre_pack_reply(req, 1, &size, NULL);
1683         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1684                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1685                 GOTO(out, rc);
1686         }
1687
1688         if (id_ino(&body->id1) == 0) {
1689                 /* an id of zero is taken to mean "sync whole filesystem" */
1690                 rc = fsfilt_sync(obd, mds->mds_sb);
1691                 if (rc)
1692                         GOTO(out, rc);
1693         } else {
1694                 /* just any file to grab fsync method - "file" arg unused */
1695                 struct file *file = mds->mds_rcvd_filp;
1696                 struct mds_body *rep_body;
1697                 struct dentry *de;
1698
1699                 de = mds_id2dentry(obd, &body->id1, NULL);
1700                 if (IS_ERR(de))
1701                         GOTO(out, rc = PTR_ERR(de));
1702
1703                 rc = file->f_op->fsync(NULL, de, 1);
1704                 if (rc)
1705                         GOTO(out, rc);
1706
1707                 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1708                 mds_pack_inode2body(obd, rep_body, de->d_inode,
1709                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1710                 l_dput(de);
1711         }
1712
1713         EXIT;
1714 out:
1715         req->rq_status = rc;
1716         return rc;
1717 }
1718
1719 /* mds_readpage does not take a DLM lock on the inode, because the client must
1720  * already have a PR lock.
1721  *
1722  * If we were to take another one here, a deadlock will result, if another
1723  * thread is already waiting for a PW lock. */
1724 static int mds_readpage(struct ptlrpc_request *req, int offset)
1725 {
1726         struct obd_device *obd = req->rq_export->exp_obd;
1727         struct vfsmount *mnt;
1728         struct dentry *de;
1729         struct file *file;
1730         struct mds_req_sec_desc *rsd;
1731         struct mds_body *body, *repbody;
1732         struct lvfs_run_ctxt saved;
1733         int rc, size = sizeof(*repbody);
1734         struct lvfs_ucred uc = {NULL, NULL,};
1735         ENTRY;
1736
1737         rc = lustre_pack_reply(req, 1, &size, NULL);
1738         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1739                 CERROR("mds: out of memory\n");
1740                 GOTO(out, rc = -ENOMEM);
1741         }
1742
1743         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1744         if (!rsd) {
1745                 CERROR("Can't unpack security desc\n");
1746                 GOTO (out, rc = -EFAULT);
1747         }
1748
1749         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1750                                   lustre_swab_mds_body);
1751         if (body == NULL) {
1752                 CERROR("Can't unpack body\n");
1753                 GOTO (out, rc = -EFAULT);
1754         }
1755
1756         rc = mds_init_ucred(&uc, req, rsd);
1757         if (rc) {
1758                 GOTO(out, rc);
1759         }
1760
1761         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1762         de = mds_id2dentry(obd, &body->id1, &mnt);
1763         if (IS_ERR(de))
1764                 GOTO(out_pop, rc = PTR_ERR(de));
1765
1766         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1767
1768         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1769         /* note: in case of an error, dentry_open puts dentry */
1770         if (IS_ERR(file))
1771                 GOTO(out_pop, rc = PTR_ERR(file));
1772
1773         /* body->size is actually the offset -eeb */
1774         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1775                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1776                        body->size, de->d_inode->i_blksize);
1777                 GOTO(out_file, rc = -EFAULT);
1778         }
1779
1780         /* body->nlink is actually the #bytes to read -eeb */
1781         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1782                 CERROR("size %u is not multiple of blocksize %lu\n",
1783                        body->nlink, de->d_inode->i_blksize);
1784                 GOTO(out_file, rc = -EFAULT);
1785         }
1786
1787         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1788         repbody->size = file->f_dentry->d_inode->i_size;
1789         repbody->valid = OBD_MD_FLSIZE;
1790
1791         /* to make this asynchronous make sure that the handling function
1792            doesn't send a reply when this function completes. Instead a
1793            callback function would send the reply */
1794         /* body->size is actually the offset -eeb */
1795         rc = mds_sendpage(req, file, body->size, body->nlink);
1796
1797         EXIT;
1798 out_file:
1799         filp_close(file, 0);
1800 out_pop:
1801         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1802 out:
1803         mds_exit_ucred(&uc);
1804         req->rq_status = rc;
1805         return 0;
1806 }
1807
1808 /* update master MDS ID, which is stored in local inode EA. */
1809 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1810                    void *data, int data_len)
1811 {
1812         struct mds_obd *mds = &obd->u.mds;
1813         struct dentry *dentry;
1814         void *handle;
1815         int rc = 0;
1816         ENTRY;
1817
1818         LASSERT(id);
1819         LASSERT(obd);
1820
1821         dentry = mds_id2dentry(obd, id, NULL);
1822         if (IS_ERR(dentry))
1823                 GOTO(out, rc = PTR_ERR(dentry));
1824
1825         if (!dentry->d_inode) {
1826                 CERROR("Can't find object "DLID4".\n",
1827                        OLID4(id));
1828                 GOTO(out_dentry, rc = -EINVAL);
1829         }
1830
1831         handle = fsfilt_start(obd, dentry->d_inode,
1832                               FSFILT_OP_SETATTR, NULL);
1833         if (IS_ERR(handle))
1834                 GOTO(out_dentry, rc = PTR_ERR(handle));
1835
1836         rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1837                                   (struct lustre_id *)data);
1838         if (rc) {
1839                 CERROR("Can't update inode "DLID4" master id, "
1840                        "error = %d.\n", OLID4(id), rc);
1841                 GOTO(out_commit, rc);
1842         }
1843
1844         EXIT;
1845 out_commit:
1846         fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1847                       handle, 0);
1848 out_dentry:
1849         l_dput(dentry);
1850 out:
1851         return rc;
1852 }
1853 EXPORT_SYMBOL(mds_update_mid);
1854
1855 /* read master MDS ID, which is stored in local inode EA. */
1856 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1857                  void *data, int data_len)
1858 {
1859         struct dentry *dentry;
1860         int rc = 0;
1861         ENTRY;
1862
1863         LASSERT(id);
1864         LASSERT(obd);
1865
1866         dentry = mds_id2dentry(obd, id, NULL);
1867         if (IS_ERR(dentry))
1868                 GOTO(out, rc = PTR_ERR(dentry));
1869
1870         if (!dentry->d_inode) {
1871                 CERROR("Can't find object "DLID4".\n",
1872                        OLID4(id));
1873                 GOTO(out_dentry, rc = -EINVAL);
1874         }
1875
1876         down(&dentry->d_inode->i_sem);
1877         rc = mds_read_inode_mid(obd, dentry->d_inode,
1878                                 (struct lustre_id *)data);
1879         up(&dentry->d_inode->i_sem);
1880         if (rc) {
1881                 CERROR("Can't read inode "DLID4" master id, "
1882                        "error = %d.\n", OLID4(id), rc);
1883                 GOTO(out_dentry, rc);
1884         }
1885
1886         EXIT;
1887 out_dentry:
1888         l_dput(dentry);
1889 out:
1890         return rc;
1891 }
1892 EXPORT_SYMBOL(mds_read_mid);
1893
1894 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
1895                 char **data, int *datalen)
1896 {
1897         struct dentry *dentry;
1898         struct mds_obd *mds = &obd->u.mds;
1899         int rc = 0, mea = 0;
1900         char *ea;
1901         ENTRY;
1902
1903         LASSERT(id);
1904         LASSERT(obd);
1905
1906         dentry = mds_id2dentry(obd, id, NULL);
1907         if (IS_ERR(dentry))
1908                 GOTO(out, rc = PTR_ERR(dentry));
1909
1910         if (!dentry->d_inode) {
1911                 CERROR("Can't find object "DLID4".\n",
1912                        OLID4(id));
1913                 GOTO(out_dentry, rc = -EINVAL);
1914         }
1915         if (S_ISDIR(dentry->d_inode->i_mode)) {
1916                 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
1917                 mea = 1;
1918         } else {
1919                 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
1920                 mea = 0;
1921         }
1922         OBD_ALLOC(ea, *datalen);
1923         if (!ea) {
1924                 *datalen = 0;
1925                 GOTO(out_dentry, rc = PTR_ERR(dentry));
1926         }
1927         *data = ea;
1928         down(&dentry->d_inode->i_sem);
1929         rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
1930                            (mea ? EA_MEA : EA_LOV));
1931         up(&dentry->d_inode->i_sem);
1932
1933         if (rc < 0)
1934                 CERROR("Error %d reading eadata for ino %lu\n",
1935                         rc, dentry->d_inode->i_ino);
1936 out_dentry:
1937         l_dput(dentry);
1938 out:
1939         RETURN(rc);
1940 }
1941 EXPORT_SYMBOL(mds_read_md);
1942
1943 int mds_reint(struct ptlrpc_request *req, int offset,
1944               struct lustre_handle *lockh)
1945 {
1946         struct mds_update_record *rec;
1947         struct mds_req_sec_desc *rsd;
1948         int rc;
1949         ENTRY;
1950
1951         OBD_ALLOC(rec, sizeof(*rec));
1952         if (rec == NULL)
1953                 RETURN(-ENOMEM);
1954
1955         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1956         if (!rsd) {
1957                 CERROR("Can't unpack security desc\n");
1958                 GOTO(out, rc = -EFAULT);
1959         }
1960
1961         rc = mds_update_unpack(req, offset, rec);
1962         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1963                 CERROR("invalid record\n");
1964                 GOTO(out, req->rq_status = -EINVAL);
1965         }
1966
1967         rc = mds_init_ucred(&rec->ur_uc, req, rsd);
1968         if (rc) {
1969                 GOTO(out, rc);
1970         }
1971
1972         /* rc will be used to interrupt a for loop over multiple records */
1973         rc = mds_reint_rec(rec, offset, req, lockh);
1974
1975  out:
1976         mds_exit_ucred(&rec->ur_uc);
1977         OBD_FREE(rec, sizeof(*rec));
1978         RETURN(rc);
1979 }
1980
1981 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1982                                        struct obd_device *obd, int *process)
1983 {
1984         switch (req->rq_reqmsg->opc) {
1985         case MDS_CONNECT: /* This will never get here, but for completeness. */
1986         case OST_CONNECT: /* This will never get here, but for completeness. */
1987         case MDS_DISCONNECT:
1988         case OST_DISCONNECT:
1989                *process = 1;
1990                RETURN(0);
1991
1992         case MDS_CLOSE:
1993         case MDS_SYNC: /* used in unmounting */
1994         case OBD_PING:
1995         case MDS_REINT:
1996         case LDLM_ENQUEUE:
1997         case OST_CREATE:
1998                 *process = target_queue_recovery_request(req, obd);
1999                 RETURN(0);
2000
2001         default:
2002                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2003                 *process = 0;
2004                 /* XXX what should we set rq_status to here? */
2005                 req->rq_status = -EAGAIN;
2006                 RETURN(ptlrpc_error(req));
2007         }
2008 }
2009
2010 static char *reint_names[] = {
2011         [REINT_SETATTR] "setattr",
2012         [REINT_CREATE]  "create",
2013         [REINT_LINK]    "link",
2014         [REINT_UNLINK]  "unlink",
2015         [REINT_RENAME]  "rename",
2016         [REINT_OPEN]    "open",
2017 };
2018
2019 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
2020                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2021                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2022                             OBD_MD_FLID)
2023
2024 static void reconstruct_create(struct ptlrpc_request *req)
2025 {
2026         struct mds_export_data *med = &req->rq_export->exp_mds_data;
2027         struct mds_client_data *mcd = med->med_mcd;
2028         struct dentry *dentry;
2029         struct ost_body *body;
2030         struct lustre_id id;
2031         int rc;
2032         ENTRY;
2033
2034         /* copy rc, transno and disp; steal locks */
2035         mds_req_from_mcd(req, mcd);
2036         if (req->rq_status) {
2037                 EXIT;
2038                 return;
2039         }
2040
2041         id_gen(&id) = 0;
2042         id_group(&id) = 0;
2043
2044         id_ino(&id) = mcd->mcd_last_data;
2045         LASSERT(id_ino(&id) != 0);
2046
2047         dentry = mds_id2dentry(req2obd(req), &id, NULL);
2048         if (IS_ERR(dentry)) {
2049                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2050                 req->rq_status = PTR_ERR(dentry);
2051                 EXIT;
2052                 return;
2053         }
2054
2055         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2056               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2057               dentry->d_inode->i_generation);
2058
2059         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2060         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2061         body->oa.o_id = dentry->d_inode->i_ino;
2062         body->oa.o_generation = dentry->d_inode->i_generation;
2063         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2064
2065         down(&dentry->d_inode->i_sem);
2066         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2067         up(&dentry->d_inode->i_sem);
2068         if (rc) {
2069                 CERROR("Can't read inode self id, inode %lu, "
2070                        "rc %d\n", dentry->d_inode->i_ino, rc);
2071                 id_fid(&id) = 0;
2072         }
2073
2074         body->oa.o_fid = id_fid(&id);
2075         body->oa.o_mds = id_group(&id);
2076         l_dput(dentry);
2077
2078         EXIT;
2079 }
2080
2081 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2082                               struct dentry *de, void *xattr, int xattr_size)
2083 {
2084         struct inode *inode = de->d_inode;
2085         struct posix_acl *acl;
2086         mode_t mode;
2087         int rc = 0;
2088
2089         LASSERT(handle);
2090         LASSERT(inode);
2091         LASSERT(xattr);
2092         LASSERT(xattr_size > 0);
2093
2094         if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2095                 CERROR("backend fs dosen't support xattr\n");
2096                 return -EOPNOTSUPP;
2097         }
2098
2099         /* set default acl */
2100         if (S_ISDIR(inode->i_mode)) {
2101                 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2102                                            xattr, xattr_size, 0);
2103                 if (rc) {
2104                         CERROR("set default acl err: %d\n", rc);
2105                         return rc;
2106                 }
2107         }
2108
2109         /* set access acl */
2110         acl = posix_acl_from_xattr(xattr, xattr_size);
2111         if (acl == NULL || IS_ERR(acl)) {
2112                 CERROR("insane attr data\n");
2113                 return PTR_ERR(acl);
2114         }
2115
2116         if (posix_acl_valid(acl)) {
2117                 CERROR("default acl not valid: %d\n", rc);
2118                 rc = -EFAULT;
2119                 goto out;
2120         }
2121
2122         mode = inode->i_mode;
2123         rc = posix_acl_create_masq(acl, &mode);
2124         if (rc < 0) {
2125                 CERROR("create masq err %d\n", rc);
2126                 goto out;
2127         }
2128
2129         if (inode->i_mode != mode) {
2130                 struct iattr iattr = { .ia_valid = ATTR_MODE,
2131                                        .ia_mode = mode };
2132                 int rc2;
2133
2134                 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2135                 if (rc2) {
2136                         CERROR("setattr mode err: %d\n", rc2);
2137                         rc = rc2;
2138                         goto out;
2139                 }
2140         }
2141
2142         if (rc > 0) {
2143                 /* we didn't change acl except mode bits of some
2144                  * entries, so should be fit into original size.
2145                  */
2146                 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2147                 LASSERT(rc > 0);
2148
2149                 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2150                                            xattr, xattr_size, 0);
2151                 if (rc)
2152                         CERROR("set access acl err: %d\n", rc);
2153         }
2154 out:
2155         posix_acl_release(acl);
2156         return rc;
2157 }
2158
2159 static int mdt_obj_create(struct ptlrpc_request *req)
2160 {
2161         struct obd_device *obd = req->rq_export->exp_obd;
2162         struct mds_obd *mds = &obd->u.mds;
2163         struct ost_body *body, *repbody;
2164         void *acl = NULL;
2165         int acl_size;
2166         char idname[LL_ID_NAMELEN];
2167         int size = sizeof(*repbody);
2168         struct inode *parent_inode;
2169         struct lvfs_run_ctxt saved;
2170         int rc, cleanup_phase = 0;
2171         struct dentry *new = NULL;
2172         struct dentry_params dp;
2173         int mealen, flags = 0;
2174         struct lvfs_ucred uc;
2175         struct lustre_id id;
2176         struct mea *mea;
2177         void *handle = NULL;
2178         unsigned long cr_inum = 0;
2179         ENTRY;
2180
2181         DEBUG_REQ(D_HA, req, "create remote object");
2182         parent_inode = mds->mds_unnamed_dir->d_inode;
2183
2184         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2185                                   lustre_swab_ost_body);
2186         if (body == NULL)
2187                 RETURN(-EFAULT);
2188
2189         /* acl data is packed transparently, no swab here */
2190         LASSERT(req->rq_reqmsg->bufcount >= 2);
2191         acl_size = req->rq_reqmsg->buflens[1];
2192         if (acl_size) {
2193                 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2194                 if (!acl) {
2195                         CERROR("No default acl buf?\n");
2196                         RETURN(-EFAULT);
2197                 }
2198         }
2199
2200         rc = lustre_pack_reply(req, 1, &size, NULL);
2201         if (rc)
2202                 RETURN(rc);
2203
2204         MDS_CHECK_RESENT(req, reconstruct_create(req));
2205
2206         uc.luc_lsd = NULL;
2207         uc.luc_ginfo = NULL;
2208         uc.luc_uid = body->oa.o_uid;
2209         uc.luc_gid = body->oa.o_gid;
2210         uc.luc_fsuid = body->oa.o_uid;
2211         uc.luc_fsgid = body->oa.o_gid;
2212
2213         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2214         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2215
2216         /* in REPLAY case inum should be given (client or other MDS fills it) */
2217         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2218             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2219                 /*
2220                  * this is re-create request from MDS holding directory name.
2221                  * we have to lookup given ino/gen first. if it exists (good
2222                  * case) then there is nothing to do. if it does not then we
2223                  * have to recreate it.
2224                  */
2225                 id_ino(&id) = body->oa.o_id;
2226                 id_gen(&id) = body->oa.o_generation;
2227
2228                 new = mds_id2dentry(obd, &id, NULL);
2229                 if (!IS_ERR(new) && new->d_inode) {
2230                         struct lustre_id sid;
2231
2232                         CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2233                                (unsigned long)id_ino(&id),
2234                                (unsigned long)id_gen(&id));
2235
2236                         obdo_from_inode(&repbody->oa, new->d_inode,
2237                                         FILTER_VALID_FLAGS);
2238
2239                         repbody->oa.o_id = new->d_inode->i_ino;
2240                         repbody->oa.o_generation = new->d_inode->i_generation;
2241                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2242                         cleanup_phase = 1;
2243
2244                         down(&new->d_inode->i_sem);
2245                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2246                         up(&new->d_inode->i_sem);
2247                         if (rc) {
2248                                 CERROR("Can't read inode self id "
2249                                        "inode %lu, rc %d.\n",
2250                                        new->d_inode->i_ino, rc);
2251                                 GOTO(cleanup, rc);
2252                         }
2253
2254                         repbody->oa.o_fid = id_fid(&sid);
2255                         repbody->oa.o_mds = id_group(&sid);
2256                         LASSERT(id_fid(&sid) != 0);
2257
2258                         /*
2259                          * here we could use fid passed in body->oa.o_fid and
2260                          * thus avoid mds_read_inode_sid().
2261                          */
2262                         cr_inum = new->d_inode->i_ino;
2263                         GOTO(cleanup, rc = 0);
2264                 }
2265         }
2266
2267         down(&parent_inode->i_sem);
2268         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2269         if (IS_ERR(handle)) {
2270                 up(&parent_inode->i_sem);
2271                 CERROR("fsfilt_start() failed, rc = %d\n",
2272                        (int)PTR_ERR(handle));
2273                 GOTO(cleanup, rc = PTR_ERR(handle));
2274         }
2275         cleanup_phase = 1; /* transaction */
2276
2277 repeat:
2278         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2279         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2280         if (IS_ERR(new)) {
2281                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2282                        obd->obd_name, idname, (int) PTR_ERR(new));
2283                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2284                 up(&parent_inode->i_sem);
2285                 RETURN(PTR_ERR(new));
2286         } else if (new->d_inode) {
2287                 CERROR("%s: name exists. repeat\n", obd->obd_name);
2288                 goto repeat;
2289         }
2290
2291         new->d_fsdata = (void *)&dp;
2292         dp.p_inum = 0;
2293         dp.p_ptr = req;
2294
2295         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2296             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2297                 LASSERT(body->oa.o_id != 0);
2298                 dp.p_inum = body->oa.o_id;
2299                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2300                           (unsigned long)body->oa.o_id,
2301                           (unsigned long)body->oa.o_generation);
2302         }
2303
2304         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2305         if (rc == 0) {
2306                 if (acl) {
2307                         rc = mds_inode_init_acl(obd, handle, new,
2308                                                 acl, acl_size);
2309                         if (rc) {
2310                                 up(&parent_inode->i_sem);
2311                                 GOTO(cleanup, rc);
2312                         }
2313                 }
2314                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2315                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2316                         new->d_inode->i_generation = body->oa.o_generation;
2317                         mark_inode_dirty(new->d_inode);
2318
2319                         /*
2320                          * avoiding asserts in cache flush case, as
2321                          * @body->oa.o_id should be zero.
2322                          */
2323                         if (body->oa.o_id) {
2324                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2325                                          "BUG 3550: failed to recreate obj "
2326                                          LPU64" -> %lu\n", body->oa.o_id,
2327                                          new->d_inode->i_ino);
2328
2329                                 LASSERTF(body->oa.o_generation ==
2330                                          new->d_inode->i_generation,
2331                                          "BUG 3550: failed to recreate obj/gen "
2332                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2333                                          body->oa.o_generation,
2334                                          new->d_inode->i_ino,
2335                                          new->d_inode->i_generation);
2336                         }
2337                 }
2338
2339                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2340                 repbody->oa.o_id = new->d_inode->i_ino;
2341                 repbody->oa.o_generation = new->d_inode->i_generation;
2342                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2343
2344                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2345                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2346                         id_group(&id) = mds->mds_num;
2347
2348                         LASSERT(body->oa.o_fid != 0);
2349                         id_fid(&id) = body->oa.o_fid;
2350
2351                         LASSERT(body->oa.o_id != 0);
2352                         id_ino(&id) = repbody->oa.o_id;
2353                         id_gen(&id) = repbody->oa.o_generation;
2354
2355                         down(&new->d_inode->i_sem);
2356                         rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2357                         up(&new->d_inode->i_sem);
2358
2359                         /*
2360                          * make sure, that fid is up-to-date.
2361                          */
2362                         mds_set_last_fid(obd, id_fid(&id));
2363                 } else {
2364                         /*
2365                          * allocate new sid, as object is created from scratch
2366                          * and this is not replay.
2367                          */
2368                         down(&new->d_inode->i_sem);
2369                         rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2370                         up(&new->d_inode->i_sem);
2371                 }
2372                 if (rc) {
2373                         CERROR("Can't update lustre ID for inode %lu, "
2374                                "error = %d\n", new->d_inode->i_ino, rc);
2375                         GOTO(cleanup, rc);
2376                 }
2377
2378                 /* initializing o_fid after it is allocated. */
2379                 repbody->oa.o_fid = id_fid(&id);
2380                 repbody->oa.o_mds = id_group(&id);
2381
2382                 rc = fsfilt_del_dir_entry(obd, new);
2383                 up(&parent_inode->i_sem);
2384                 if (rc) {
2385                         CERROR("can't remove name for object: %d\n", rc);
2386                         GOTO(cleanup, rc);
2387                 }
2388
2389                 cleanup_phase = 2; /* created directory object */
2390
2391                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2392                        (unsigned long)new->d_inode->i_ino,
2393                        (unsigned long)new->d_inode->i_generation,
2394                        (unsigned)new->d_inode->i_mode);
2395                 cr_inum = new->d_inode->i_ino;
2396         } else {
2397                 up(&parent_inode->i_sem);
2398                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2399                 GOTO(cleanup, rc);
2400         }
2401
2402         if (body->oa.o_valid & OBD_MD_FLID) {
2403                 /* this is new object for splitted dir. We have to prevent
2404                  * recursive splitting on it -bzzz */
2405                 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2406
2407                 OBD_ALLOC(mea, mealen);
2408                 if (mea == NULL)
2409                         GOTO(cleanup, rc = -ENOMEM);
2410
2411                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2412                 mea->mea_master = 0;
2413                 mea->mea_count = 0;
2414
2415                 down(&new->d_inode->i_sem);
2416                 rc = fsfilt_set_md(obd, new->d_inode, handle,
2417                                    mea, mealen, EA_MEA);
2418                 up(&new->d_inode->i_sem);
2419                 if (rc)
2420                         CERROR("fsfilt_set_md() failed, "
2421                                "rc = %d\n", rc);
2422
2423                 OBD_FREE(mea, mealen);
2424
2425                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2426                        obd->obd_name, new->d_inode->i_ino,
2427                        new->d_inode->i_generation, flags);
2428         } else if (body->oa.o_easize) {
2429                 /* we pass LCK_EX to split routine to signal that we have
2430                  * exclusive access to the directory. simple because nobody
2431                  * knows it already exists -bzzz */
2432                 rc = mds_try_to_split_dir(obd, new, NULL,
2433                                           body->oa.o_easize, LCK_EX);
2434                 if (rc < 0) {
2435                         CERROR("Can't split directory %lu, error = %d.\n",
2436                                new->d_inode->i_ino, rc);
2437                 } else {
2438                         rc = 0;
2439                 }
2440         }
2441
2442         EXIT;
2443 cleanup:
2444         switch (cleanup_phase) {
2445         case 2: /* object has been created, but we'll may want to replay it later */
2446                 if (rc == 0)
2447                         ptlrpc_require_repack(req);
2448         case 1: /* transaction */
2449                 rc = mds_finish_transno(mds, parent_inode, handle,
2450                                         req, rc, cr_inum);
2451         }
2452
2453         l_dput(new);
2454         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2455         return rc;
2456 }
2457
2458 static int mdt_get_info(struct ptlrpc_request *req)
2459 {
2460         struct obd_export *exp = req->rq_export;
2461         int keylen, rc = 0;
2462         char *key;
2463         ENTRY;
2464
2465         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2466         if (key == NULL) {
2467                 DEBUG_REQ(D_HA, req, "no get_info key");
2468                 RETURN(-EFAULT);
2469         }
2470         keylen = req->rq_reqmsg->buflens[0];
2471
2472         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2473             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2474             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2475                 RETURN(-EPROTO);
2476
2477         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2478                 struct lustre_id *reply;
2479                 int size = sizeof(*reply);
2480
2481                 rc = lustre_pack_reply(req, 1, &size, NULL);
2482                 if (rc)
2483                         RETURN(rc);
2484
2485                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2486                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2487         } else {
2488                 obd_id *reply;
2489                 int size = sizeof(*reply);
2490
2491                 rc = lustre_pack_reply(req, 1, &size, NULL);
2492                 if (rc)
2493                         RETURN(rc);
2494
2495                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2496                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2497         }
2498
2499         req->rq_repmsg->status = 0;
2500         RETURN(rc);
2501 }
2502
2503 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2504                         void *key, __u32 vallen, void *val)
2505 {
2506         struct obd_device *obd;
2507         struct mds_obd *mds;
2508         int rc = 0;
2509         ENTRY;
2510
2511         obd = class_exp2obd(exp);
2512         if (obd == NULL) {
2513                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2514                        exp->exp_handle.h_cookie);
2515                 RETURN(-EINVAL);
2516         }
2517
2518         mds = &obd->u.mds;
2519         if (keylen >= strlen("mds_type") &&
2520              memcmp(key, "mds_type", keylen) == 0) {
2521                 int valsize;
2522                 __u32 group;
2523
2524                 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2525
2526                 mds->mds_obd_type = *(int*)val;
2527                 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2528                 valsize = sizeof(group);
2529
2530                 /* mds number has been changed, so the corresponding obdfilter
2531                  * exp need to be changed too. */
2532                 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2533                                   "mds_conn", valsize, &group);
2534                 RETURN(rc);
2535         }
2536         CDEBUG(D_IOCTL, "invalid key\n");
2537         RETURN(-EINVAL);
2538 }
2539
2540 static int mdt_set_info(struct ptlrpc_request *req)
2541 {
2542         char *key, *val;
2543         struct obd_export *exp = req->rq_export;
2544         int keylen, rc = 0, vallen;
2545         ENTRY;
2546
2547         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2548         if (key == NULL) {
2549                 DEBUG_REQ(D_HA, req, "no set_info key");
2550                 RETURN(-EFAULT);
2551         }
2552         keylen = req->rq_reqmsg->buflens[0];
2553
2554         if (keylen == strlen("mds_type") &&
2555             memcmp(key, "mds_type", keylen) == 0) {
2556                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2557                 if (rc)
2558                         RETURN(rc);
2559
2560                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2561                 vallen = req->rq_reqmsg->buflens[1];
2562
2563                 rc = obd_set_info(exp, keylen, key, vallen, val);
2564                 req->rq_repmsg->status = 0;
2565                 RETURN(rc);
2566         }
2567         CDEBUG(D_IOCTL, "invalid key\n");
2568         RETURN(-EINVAL);
2569 }
2570
2571 static void mds_revoke_export_locks(struct obd_export *exp)
2572 {
2573         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
2574         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2575         struct list_head work;
2576         struct ldlm_lock *lock, *next;
2577         struct ldlm_lock_desc desc;
2578
2579         if (!exp->u.eu_mds_data.med_remote)
2580                 return;
2581
2582         ENTRY;
2583         CERROR("implement right locking here! -bzzz\n");
2584         INIT_LIST_HEAD(&work);
2585         spin_lock(&exp->exp_ldlm_data.led_lock);
2586         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2587
2588                 lock_res(lock->l_resource);
2589                 if (lock->l_req_mode != lock->l_granted_mode) {
2590                         unlock_res(lock->l_resource);
2591                         continue;
2592                 }
2593
2594                 LASSERT(lock->l_resource);
2595                 if (lock->l_resource->lr_type != LDLM_IBITS &&
2596                     lock->l_resource->lr_type != LDLM_PLAIN) {
2597                         unlock_res(lock->l_resource);
2598                         continue;
2599                 }
2600
2601                 if (lock->l_flags & LDLM_FL_AST_SENT) {
2602                         unlock_res(lock->l_resource);
2603                         continue;
2604                 }
2605
2606                 lock->l_flags |= LDLM_FL_AST_SENT;
2607                 unlock_res(lock->l_resource);
2608
2609                 /* the desc just pretend to exclusive */
2610                 ldlm_lock2desc(lock, &desc);
2611                 desc.l_req_mode = LCK_EX;
2612                 desc.l_granted_mode = 0;
2613
2614                 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2615         }
2616         spin_unlock(&exp->exp_ldlm_data.led_lock);
2617
2618         EXIT;
2619 }
2620
2621 static int mds_msg_check_version(struct lustre_msg *msg)
2622 {
2623         int rc;
2624
2625         switch (msg->opc) {
2626         case MDS_CONNECT:
2627         case MDS_DISCONNECT:
2628         case OBD_PING:
2629                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2630                 if (rc)
2631                         CERROR("bad opc %u version %08x, expecting %08x\n",
2632                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2633                 break;
2634         case MDS_STATFS:
2635         case MDS_GETSTATUS:
2636         case MDS_GETATTR:
2637         case MDS_GETATTR_LOCK:
2638         case MDS_READPAGE:
2639         case MDS_REINT:
2640         case MDS_CLOSE:
2641         case MDS_DONE_WRITING:
2642         case MDS_PIN:
2643         case MDS_SYNC:
2644                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2645                 if (rc)
2646                         CERROR("bad opc %u version %08x, expecting %08x\n",
2647                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2648                 break;
2649         case LDLM_ENQUEUE:
2650         case LDLM_CONVERT:
2651         case LDLM_BL_CALLBACK:
2652         case LDLM_CP_CALLBACK:
2653                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2654                 if (rc)
2655                         CERROR("bad opc %u version %08x, expecting %08x\n",
2656                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2657                 break;
2658         case OBD_LOG_CANCEL:
2659         case LLOG_ORIGIN_HANDLE_OPEN:
2660         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2661         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2662         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2663         case LLOG_ORIGIN_HANDLE_CLOSE:
2664         case LLOG_CATINFO:
2665                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2666                 if (rc)
2667                         CERROR("bad opc %u version %08x, expecting %08x\n",
2668                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2669                 break;
2670         case OST_CREATE:
2671         case OST_WRITE:
2672         case OST_GET_INFO:
2673         case OST_SET_INFO:
2674                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2675                 if (rc)
2676                         CERROR("bad opc %u version %08x, expecting %08x\n",
2677                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2678                 break;
2679         case SEC_INIT:
2680         case SEC_INIT_CONTINUE:
2681         case SEC_FINI:
2682                 rc = 0;
2683                 break;
2684         default:
2685                 CERROR("MDS unknown opcode %d\n", msg->opc);
2686                 rc = -ENOTSUPP;
2687                 break;
2688         }
2689
2690         return rc;
2691 }
2692
2693 int mds_handle(struct ptlrpc_request *req)
2694 {
2695         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2696         struct obd_device *obd = NULL;
2697         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2698         int rc = 0;
2699         ENTRY;
2700
2701         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2702
2703         rc = mds_msg_check_version(req->rq_reqmsg);
2704         if (rc) {
2705                 CERROR("MDS drop mal-formed request\n");
2706                 RETURN(rc);
2707         }
2708
2709         /* Security opc should NOT trigger any recovery events */
2710         if (req->rq_reqmsg->opc == SEC_INIT ||
2711             req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2712                 if (req->rq_export) {
2713                         mds_req_add_idmapping(req,
2714                                               &req->rq_export->exp_mds_data);
2715                         mds_revoke_export_locks(req->rq_export);
2716                 }
2717                 GOTO(out, rc = 0);
2718         } else if (req->rq_reqmsg->opc == SEC_FINI) {
2719                 if (req->rq_export) {
2720                         mds_req_del_idmapping(req,
2721                                               &req->rq_export->exp_mds_data);
2722                         mds_revoke_export_locks(req->rq_export);
2723                 }
2724                 GOTO(out, rc = 0);
2725         }
2726
2727         LASSERT(current->journal_info == NULL);
2728         /* XXX identical to OST */
2729         if (req->rq_reqmsg->opc != MDS_CONNECT) {
2730                 struct mds_export_data *med;
2731                 int recovering;
2732
2733                 if (req->rq_export == NULL) {
2734                         CERROR("operation %d on unconnected MDS from %s\n",
2735                                req->rq_reqmsg->opc,
2736                                req->rq_peerstr);
2737                         req->rq_status = -ENOTCONN;
2738                         GOTO(out, rc = -ENOTCONN);
2739                 }
2740
2741                 med = &req->rq_export->exp_mds_data;
2742                 obd = req->rq_export->exp_obd;
2743                 mds = &obd->u.mds;
2744
2745                 /* sanity check: if the xid matches, the request must
2746                  * be marked as a resent or replayed */
2747                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2748                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2749                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2750                                  (MSG_RESENT | MSG_REPLAY),
2751                                  "rq_xid "LPU64" matches last_xid, "
2752                                  "expected RESENT flag\n",
2753                                  req->rq_xid);
2754                 }
2755                 /* else: note the opposite is not always true; a
2756                  * RESENT req after a failover will usually not match
2757                  * the last_xid, since it was likely never
2758                  * committed. A REPLAYed request will almost never
2759                  * match the last xid, however it could for a
2760                  * committed, but still retained, open. */
2761
2762                 spin_lock_bh(&obd->obd_processing_task_lock);
2763                 recovering = obd->obd_recovering;
2764                 spin_unlock_bh(&obd->obd_processing_task_lock);
2765                 if (recovering) {
2766                         rc = mds_filter_recovery_request(req, obd,
2767                                                          &should_process);
2768                         if (rc || should_process == 0) {
2769                                 RETURN(rc);
2770                         } else if (should_process < 0) {
2771                                 req->rq_status = should_process;
2772                                 rc = ptlrpc_error(req);
2773                                 RETURN(rc);
2774                         }
2775                 }
2776         }
2777
2778         switch (req->rq_reqmsg->opc) {
2779         case MDS_CONNECT:
2780                 DEBUG_REQ(D_INODE, req, "connect");
2781                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2782                 rc = target_handle_connect(req);
2783                 if (!rc) {
2784                         struct mds_export_data *med;
2785
2786                         LASSERT(req->rq_export);
2787                         med = &req->rq_export->u.eu_mds_data;
2788                         mds_init_export_data(req, med);
2789                         mds_req_add_idmapping(req, med);
2790
2791                         /* Now that we have an export, set mds. */
2792                         obd = req->rq_export->exp_obd;
2793                         mds = mds_req2mds(req);
2794                 }
2795                 break;
2796
2797         case MDS_DISCONNECT:
2798                 DEBUG_REQ(D_INODE, req, "disconnect");
2799                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2800                 rc = target_handle_disconnect(req);
2801                 req->rq_status = rc;            /* superfluous? */
2802                 break;
2803
2804         case MDS_GETSTATUS:
2805                 DEBUG_REQ(D_INODE, req, "getstatus");
2806                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2807                 rc = mds_getstatus(req);
2808                 break;
2809
2810         case MDS_GETATTR:
2811                 DEBUG_REQ(D_INODE, req, "getattr");
2812                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2813                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2814                 break;
2815
2816         case MDS_GETATTR_LOCK: {
2817                 struct lustre_handle lockh;
2818                 DEBUG_REQ(D_INODE, req, "getattr_lock");
2819                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2820
2821                 /* If this request gets a reconstructed reply, we won't be
2822                  * acquiring any new locks in mds_getattr_lock, so we don't
2823                  * want to cancel.
2824                  */
2825                 lockh.cookie = 0;
2826                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2827                                       MDS_INODELOCK_UPDATE);
2828                 /* this non-intent call (from an ioctl) is special */
2829                 req->rq_status = rc;
2830                 if (rc == 0 && lockh.cookie)
2831                         ldlm_lock_decref(&lockh, LCK_PR);
2832                 break;
2833         }
2834         case MDS_STATFS:
2835                 DEBUG_REQ(D_INODE, req, "statfs");
2836                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2837                 rc = mds_statfs(req);
2838                 break;
2839
2840         case MDS_READPAGE:
2841                 DEBUG_REQ(D_INODE, req, "readpage");
2842                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2843                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2844
2845                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2846                         if (req->rq_reply_state) {
2847                                 lustre_free_reply_state (req->rq_reply_state);
2848                                 req->rq_reply_state = NULL;
2849                         }
2850                         RETURN(0);
2851                 }
2852
2853                 break;
2854         case MDS_REINT: {
2855                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2856                                              sizeof (*opcp));
2857                 __u32  opc;
2858                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2859                                mds->mds_max_cookiesize};
2860                 int bufcount;
2861
2862                 /* NB only peek inside req now; mds_reint() will swab it */
2863                 if (opcp == NULL) {
2864                         CERROR ("Can't inspect opcode\n");
2865                         rc = -EINVAL;
2866                         break;
2867                 }
2868                 opc = *opcp;
2869                 if (lustre_msg_swabbed (req->rq_reqmsg))
2870                         __swab32s(&opc);
2871
2872                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2873                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2874                            reint_names[opc] == NULL) ? reint_names[opc] :
2875                                                        "unknown opcode");
2876
2877                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2878
2879                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2880                         bufcount = 3;
2881                 else if (opc == REINT_OPEN)
2882                         bufcount = 2;
2883                 else
2884                         bufcount = 1;
2885
2886                 rc = lustre_pack_reply(req, bufcount, size, NULL);
2887                 if (rc)
2888                         break;
2889
2890                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2891                 fail = OBD_FAIL_MDS_REINT_NET_REP;
2892                 break;
2893         }
2894
2895         case MDS_CLOSE:
2896                 DEBUG_REQ(D_INODE, req, "close");
2897                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2898                 rc = mds_close(req, MDS_REQ_REC_OFF);
2899                 break;
2900
2901         case MDS_DONE_WRITING:
2902                 DEBUG_REQ(D_INODE, req, "done_writing");
2903                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2904                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2905                 break;
2906
2907         case MDS_PIN:
2908                 DEBUG_REQ(D_INODE, req, "pin");
2909                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2910                 rc = mds_pin(req, MDS_REQ_REC_OFF);
2911                 break;
2912
2913         case MDS_SYNC:
2914                 DEBUG_REQ(D_INODE, req, "sync");
2915                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2916                 rc = mds_sync(req, MDS_REQ_REC_OFF);
2917                 break;
2918
2919         case OBD_PING:
2920                 DEBUG_REQ(D_INODE, req, "ping");
2921                 rc = target_handle_ping(req);
2922                 break;
2923
2924         case OBD_LOG_CANCEL:
2925                 CDEBUG(D_INODE, "log cancel\n");
2926                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2927                 rc = -ENOTSUPP; /* la la la */
2928                 break;
2929
2930         case LDLM_ENQUEUE:
2931                 DEBUG_REQ(D_INODE, req, "enqueue");
2932                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2933                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2934                                          ldlm_server_blocking_ast, NULL);
2935                 fail = OBD_FAIL_LDLM_REPLY;
2936                 break;
2937         case LDLM_CONVERT:
2938                 DEBUG_REQ(D_INODE, req, "convert");
2939                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2940                 rc = ldlm_handle_convert(req);
2941                 break;
2942         case LDLM_BL_CALLBACK:
2943         case LDLM_CP_CALLBACK:
2944                 DEBUG_REQ(D_INODE, req, "callback");
2945                 CERROR("callbacks should not happen on MDS\n");
2946                 LBUG();
2947                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2948                 break;
2949         case LLOG_ORIGIN_HANDLE_OPEN:
2950                 DEBUG_REQ(D_INODE, req, "llog_init");
2951                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2952                 rc = llog_origin_handle_open(req);
2953                 break;
2954         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2955                 DEBUG_REQ(D_INODE, req, "llog next block");
2956                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2957                 rc = llog_origin_handle_next_block(req);
2958                 break;
2959         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2960                 DEBUG_REQ(D_INODE, req, "llog prev block");
2961                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2962                 rc = llog_origin_handle_prev_block(req);
2963                 break;
2964         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2965                 DEBUG_REQ(D_INODE, req, "llog read header");
2966                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2967                 rc = llog_origin_handle_read_header(req);
2968                 break;
2969         case LLOG_ORIGIN_HANDLE_CLOSE:
2970                 DEBUG_REQ(D_INODE, req, "llog close");
2971                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2972                 rc = llog_origin_handle_close(req);
2973                 break;
2974         case OST_CREATE:
2975                 DEBUG_REQ(D_INODE, req, "ost_create");
2976                 rc = mdt_obj_create(req);
2977                 break;
2978         case OST_GET_INFO:
2979                 DEBUG_REQ(D_INODE, req, "get_info");
2980                 rc = mdt_get_info(req);
2981                 break;
2982         case OST_SET_INFO:
2983                 DEBUG_REQ(D_INODE, req, "set_info");
2984                 rc = mdt_set_info(req);
2985                 break;
2986         case OST_WRITE:
2987                 CDEBUG(D_INODE, "write\n");
2988                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2989                 rc = ost_brw_write(req, NULL);
2990                 LASSERT(current->journal_info == NULL);
2991                 /* mdt_brw sends its own replies */
2992                 RETURN(rc);
2993                 break;
2994         case LLOG_CATINFO:
2995                 DEBUG_REQ(D_INODE, req, "llog catinfo");
2996                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2997                 rc = llog_catinfo(req);
2998                 break;
2999         default:
3000                 req->rq_status = -ENOTSUPP;
3001                 rc = ptlrpc_error(req);
3002                 RETURN(rc);
3003         }
3004
3005         LASSERT(current->journal_info == NULL);
3006
3007         EXIT;
3008
3009         /* If we're DISCONNECTing, the mds_export_data is already freed */
3010         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3011                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3012                 struct obd_device *obd = list_entry(mds, struct obd_device,
3013                                                     u.mds);
3014                 req->rq_repmsg->last_xid =
3015                         le64_to_cpu(med->med_mcd->mcd_last_xid);
3016
3017                 if (!obd->obd_no_transno) {
3018                         req->rq_repmsg->last_committed =
3019                                 obd->obd_last_committed;
3020                 } else {
3021                         DEBUG_REQ(D_IOCTL, req,
3022                                   "not sending last_committed update");
3023                 }
3024                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3025                        ", xid "LPU64"\n",
3026                        mds->mds_last_transno, obd->obd_last_committed,
3027                        req->rq_xid);
3028         }
3029  out:
3030
3031
3032         target_send_reply(req, rc, fail);
3033         return 0;
3034 }
3035
3036 /* Update the server data on disk.  This stores the new mount_count and also the
3037  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
3038  * last_rcvd value may be less than that of the clients.  This will alert us
3039  * that we may need to do client recovery.
3040  *
3041  * Also assumes for mds_last_transno that we are not modifying it (no locking).
3042  */
3043 int mds_update_server_data(struct obd_device *obd, int force_sync)
3044 {
3045         struct mds_obd *mds = &obd->u.mds;
3046         struct mds_server_data *msd = mds->mds_server_data;
3047         struct file *filp = mds->mds_rcvd_filp;
3048         struct lvfs_run_ctxt saved;
3049         loff_t off = 0;
3050         int rc;
3051         ENTRY;
3052
3053         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3054         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3055
3056         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3057                mds->mds_mount_count, mds->mds_last_transno);
3058         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3059         if (rc)
3060                 CERROR("error writing MDS server data: rc = %d\n", rc);
3061         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3062
3063         RETURN(rc);
3064 }
3065
3066 /* saves last allocated fid counter to file. */
3067 int mds_update_last_fid(struct obd_device *obd, void *handle,
3068                         int force_sync)
3069 {
3070         struct mds_obd *mds = &obd->u.mds;
3071         struct file *filp = mds->mds_fid_filp;
3072         struct lvfs_run_ctxt saved;
3073         loff_t off = 0;
3074         __u64 last_fid;
3075         int rc = 0;
3076         ENTRY;
3077
3078         spin_lock(&mds->mds_last_fid_lock);
3079         last_fid = mds->mds_last_fid;
3080         spin_unlock(&mds->mds_last_fid_lock);
3081
3082         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3083                last_fid);
3084
3085         if (handle) {
3086                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3087                                       handle, mds_commit_last_fid_cb,
3088                                       NULL);
3089         }
3090
3091         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3092         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3093                                  &off, force_sync);
3094         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3095
3096         if (rc) {
3097                 CERROR("error writing MDS last_fid #"LPU64
3098                        ", err = %d\n", last_fid, rc);
3099                 RETURN(rc);
3100         }
3101
3102         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3103                "%llu: err = %d\n", last_fid, off, rc);
3104
3105         RETURN(rc);
3106 }
3107
3108 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3109 {
3110         struct mds_obd *mds = &obd->u.mds;
3111
3112         spin_lock(&mds->mds_last_fid_lock);
3113         if (fid > mds->mds_last_fid)
3114                 mds->mds_last_fid = fid;
3115         spin_unlock(&mds->mds_last_fid_lock);
3116 }
3117
3118 void mds_commit_last_transno_cb(struct obd_device *obd,
3119                                 __u64 transno, void *data,
3120                                 int error)
3121 {
3122         obd_transno_commit_cb(obd, transno, error);
3123 }
3124
3125 void mds_commit_last_fid_cb(struct obd_device *obd,
3126                             __u64 fid, void *data,
3127                             int error)
3128 {
3129         if (error) {
3130                 CERROR("%s: fid "LPD64" commit error: %d\n",
3131                        obd->obd_name, fid, error);
3132                 return;
3133         }
3134
3135         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3136                obd->obd_name, fid);
3137 }
3138
3139 __u64 mds_alloc_fid(struct obd_device *obd)
3140 {
3141         struct mds_obd *mds = &obd->u.mds;
3142         __u64 fid;
3143
3144         spin_lock(&mds->mds_last_fid_lock);
3145         fid = ++mds->mds_last_fid;
3146         spin_unlock(&mds->mds_last_fid_lock);
3147
3148         return fid;
3149 }
3150
3151 /*
3152  * allocates new lustre_id on passed @inode and saves it to inode EA.
3153  */
3154 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
3155                         void *handle, struct lustre_id *id)
3156 {
3157         struct mds_obd *mds = &obd->u.mds;
3158         int alloc = 0, rc = 0;
3159         ENTRY;
3160
3161         LASSERT(obd != NULL);
3162         LASSERT(inode != NULL);
3163
3164         if (id == NULL) {
3165                 OBD_ALLOC(id, sizeof(*id));
3166                 if (id == NULL)
3167                         RETURN(-ENOMEM);
3168                 alloc = 1;
3169         }
3170
3171         id_group(id) = mds->mds_num;
3172         id_fid(id) = mds_alloc_fid(obd);
3173         id_ino(id) = inode->i_ino;
3174         id_gen(id) = inode->i_generation;
3175         id_type(id) = (S_IFMT & inode->i_mode);
3176
3177         rc = mds_update_inode_sid(obd, inode, handle, id);
3178         if (rc) {
3179                 CERROR("Can't update inode FID EA, "
3180                        "rc = %d\n", rc);
3181         }
3182
3183         if (alloc)
3184                 OBD_FREE(id, sizeof(*id));
3185         RETURN(rc);
3186 }
3187
3188 /*
3189  * reads inode self id from inode EA. Probably later this should be replaced by
3190  * caching inode self id to avoid raeding it every time it is needed.
3191  */
3192 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3193                        struct lustre_id *id)
3194 {
3195         int rc;
3196         ENTRY;
3197
3198         LASSERT(id != NULL);
3199         LASSERT(obd != NULL);
3200         LASSERT(inode != NULL);
3201
3202         rc = fsfilt_get_md(obd, inode, &id->li_fid,
3203                            sizeof(id->li_fid), EA_SID);
3204         if (rc < 0) {
3205                 CERROR("fsfilt_get_md() failed, "
3206                        "rc = %d\n", rc);
3207                 RETURN(rc);
3208         } else if (!rc) {
3209                 rc = -ENODATA;
3210                 RETURN(rc);
3211         } else {
3212                 rc = 0;
3213         }
3214
3215         RETURN(rc);
3216 }
3217
3218 /* updates inode self id in EA. */
3219 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3220                          void *handle, struct lustre_id *id)
3221 {
3222         int rc = 0;
3223         ENTRY;
3224
3225         LASSERT(id != NULL);
3226         LASSERT(obd != NULL);
3227         LASSERT(inode != NULL);
3228
3229         rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3230                            sizeof(id->li_fid), EA_SID);
3231         if (rc) {
3232                 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3233                 RETURN(rc);
3234         }
3235
3236         RETURN(rc);
3237 }
3238
3239 /*
3240  * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3241  * to master MDS by correct store cookie, needed to find inode on master MDS
3242  * quickly.
3243  */
3244 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3245                        struct lustre_id *id)
3246 {
3247         int rc;
3248         ENTRY;
3249
3250         LASSERT(id != NULL);
3251         LASSERT(obd != NULL);
3252         LASSERT(inode != NULL);
3253
3254         rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3255         if (rc < 0) {
3256                 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3257                 RETURN(rc);
3258         } else if (!rc) {
3259                 rc = -ENODATA;
3260                 RETURN(rc);
3261         } else {
3262                 rc = 0;
3263         }
3264
3265         RETURN(rc);
3266 }
3267
3268 /*
3269  * updates master inode id. Usualy this is done by CMOBD after an inode is
3270  * created and relationship between cache MDS and master one should be
3271  * established.
3272  */
3273 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3274                          void *handle, struct lustre_id *id)
3275 {
3276         int rc = 0;
3277         ENTRY;
3278
3279         LASSERT(id != NULL);
3280         LASSERT(obd != NULL);
3281         LASSERT(inode != NULL);
3282
3283         rc = fsfilt_set_md(obd, inode, handle, id,
3284                            sizeof(*id), EA_MID);
3285         if (rc) {
3286                 CERROR("fsfilt_set_md() failed, "
3287                        "rc = %d\n", rc);
3288                 RETURN(rc);
3289         }
3290
3291         RETURN(rc);
3292 }
3293
3294 /* mount the file system (secretly) */
3295 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3296 {
3297         struct lustre_cfg* lcfg = buf;
3298         struct mds_obd *mds = &obd->u.mds;
3299         struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3300         char *options = NULL;
3301         struct vfsmount *mnt;
3302         char ns_name[48];
3303         unsigned long page;
3304         int rc = 0;
3305         ENTRY;
3306
3307         if (lcfg->lcfg_bufcount < 3)
3308                 RETURN(rc = -EINVAL);
3309
3310         if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3311                 RETURN(rc = -EINVAL);
3312
3313         obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3314         if (IS_ERR(obd->obd_fsops))
3315                 RETURN(rc = PTR_ERR(obd->obd_fsops));
3316
3317         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3318
3319         page = get_zeroed_page(GFP_KERNEL);
3320         if (!page)
3321                 RETURN(-ENOMEM);
3322
3323         options = (char *)page;
3324
3325         /*
3326          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3327          * and the rest of options are passed by mount options. Probably this
3328          * should be moved to somewhere else like startup scripts or lconf. */
3329         sprintf(options, "iopen_nopriv");
3330
3331         if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3332                 sprintf(options + strlen(options), ",%s",
3333                         lustre_cfg_string(lcfg, 4));
3334
3335         /* we have to know mdsnum before touching underlying fs -bzzz */
3336         atomic_set(&mds->mds_open_count, 0);
3337         sema_init(&mds->mds_md_sem, 1);
3338         sema_init(&mds->mds_create_sem, 1);
3339         mds->mds_md_connected = 0;
3340         mds->mds_md_name = NULL;
3341
3342         if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3343             strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3344                 class_uuid_t uuid;
3345
3346                 generate_random_uuid(uuid);
3347                 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3348
3349                 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3350                 if (mds->mds_md_name == NULL)
3351                         RETURN(rc = -ENOMEM);
3352
3353                 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3354                        LUSTRE_CFG_BUFLEN(lcfg, 5));
3355
3356                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3357                        obd->obd_name, mds->mds_md_name);
3358
3359                 rc = mds_md_connect(obd, mds->mds_md_name);
3360                 if (rc) {
3361                         OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3362                         GOTO(err_ops, rc);
3363                 }
3364         }
3365
3366         mds->mds_obd_type = MDS_MASTER_OBD;
3367
3368         if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3369             strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3370                     LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3371                 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3372                             strlen("master"))) {
3373                         mds->mds_obd_type = MDS_MASTER_OBD;
3374                 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3375                                    strlen("cache"))) {
3376                         mds->mds_obd_type = MDS_CACHE_OBD;
3377                 }
3378         }
3379
3380         rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3381                            lustre_cfg_string(lcfg, 2),
3382                            options, 0, &lvfs_ctxt);
3383
3384         free_page(page);
3385
3386         if (rc || !lvfs_ctxt) {
3387                 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3388                 GOTO(err_ops, rc);
3389         }
3390
3391         mnt = lvfs_ctxt->loc_mnt;
3392         mds->mds_lvfs_ctxt = lvfs_ctxt;
3393         ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3394
3395         CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3396
3397         sema_init(&mds->mds_epoch_sem, 1);
3398         atomic_set(&mds->mds_real_clients, 0);
3399         spin_lock_init(&mds->mds_transno_lock);
3400         spin_lock_init(&mds->mds_last_fid_lock);
3401         sema_init(&mds->mds_orphan_recovery_sem, 1);
3402         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3403
3404         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3405         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3406
3407         if (obd->obd_namespace == NULL) {
3408                 mds_cleanup(obd, 0);
3409                 GOTO(err_put, rc = -ENOMEM);
3410         }
3411         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3412
3413         rc = mds_fs_setup(obd, mnt);
3414         if (rc) {
3415                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3416                        obd->obd_name, rc);
3417                 GOTO(err_ns, rc);
3418         }
3419
3420         rc = llog_start_commit_thread();
3421         if (rc < 0)
3422
3423                 GOTO(err_fs, rc);
3424
3425
3426         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3427             strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3428                     LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3429                 class_uuid_t uuid;
3430
3431                 generate_random_uuid(uuid);
3432                 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3433
3434                 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3435                 if (mds->mds_profile == NULL)
3436                         GOTO(err_fs, rc = -ENOMEM);
3437
3438                 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3439                         LUSTRE_CFG_BUFLEN(lcfg, 3));
3440         }
3441
3442         /*
3443          * setup root dir and files ID dir if lmv already connected, or there is
3444          * not lmv at all.
3445          */
3446         if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3447                                 lustre_cfg_buf(lcfg, 3) &&
3448                                 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3449                                         LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3450                 rc = mds_fs_setup_rootid(obd);
3451                 if (rc)
3452                         GOTO(err_fs, rc);
3453
3454                 rc = mds_fs_setup_virtid(obd);
3455                 if (rc)
3456                         GOTO(err_fs, rc);
3457         }
3458
3459         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3460                            "mds_ldlm_client", &obd->obd_ldlm_client);
3461         obd->obd_replayable = 1;
3462
3463         rc = mds_postsetup(obd);
3464         if (rc)
3465                 GOTO(err_fs, rc);
3466
3467         RETURN(0);
3468
3469 err_fs:
3470         /* No extra cleanup needed for llog_init_commit_thread() */
3471         mds_fs_cleanup(obd, 0);
3472 err_ns:
3473         ldlm_namespace_free(obd->obd_namespace, 0);
3474         obd->obd_namespace = NULL;
3475 err_put:
3476         unlock_kernel();
3477         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3478         mds->mds_sb = 0;
3479         lock_kernel();
3480 err_ops:
3481         fsfilt_put_ops(obd->obd_fsops);
3482         return rc;
3483 }
3484
3485 static int mds_postsetup(struct obd_device *obd)
3486 {
3487         struct mds_obd *mds = &obd->u.mds;
3488         int rc = 0;
3489         ENTRY;
3490
3491         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3492                             obd, 0, NULL, &llog_lvfs_ops);
3493         if (rc)
3494                 RETURN(rc);
3495
3496         if (mds->mds_profile) {
3497                 struct llog_ctxt *lgctxt;
3498                 struct lvfs_run_ctxt saved;
3499                 struct lustre_profile *lprof;
3500                 struct config_llog_instance cfg;
3501
3502                 cfg.cfg_instance = NULL;
3503                 cfg.cfg_uuid = mds->mds_dt_uuid;
3504                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3505
3506                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3507                 if (!lgctxt) {
3508                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3509                         GOTO(err_llog, rc = -EINVAL);
3510                 }
3511
3512                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3513                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3514
3515                 if (rc)
3516                         GOTO(err_llog, rc);
3517
3518                 lprof = class_get_profile(mds->mds_profile);
3519                 if (lprof == NULL) {
3520                         CERROR("No profile found: %s\n", mds->mds_profile);
3521                         GOTO(err_cleanup, rc = -ENOENT);
3522                 }
3523                 rc = mds_dt_connect(obd, lprof->lp_lov);
3524                 if (rc)
3525                         GOTO(err_cleanup, rc);
3526
3527                 rc = mds_md_postsetup(obd);
3528                 if (rc)
3529                         GOTO(err_cleanup, rc);
3530         }
3531
3532         RETURN(rc);
3533 err_cleanup:
3534         mds_dt_clean(obd);
3535 err_llog:
3536         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3537                                           LLOG_CONFIG_ORIG_CTXT));
3538         return rc;
3539 }
3540
3541 int mds_postrecov_common(struct obd_device *obd)
3542 {
3543         struct mds_obd *mds = &obd->u.mds;
3544         struct llog_ctxt *ctxt;
3545         int rc, item = 0, valsize;
3546          __u32 group;
3547         ENTRY;
3548
3549         LASSERT(!obd->obd_recovering);
3550         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3551         LASSERT(ctxt != NULL);
3552
3553         /* clean PENDING dir */
3554         rc = mds_cleanup_orphans(obd);
3555         if (rc < 0)
3556                 GOTO(out, rc);
3557         item = rc;
3558
3559         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3560         valsize = sizeof(group);
3561         rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3562                           "mds_conn", valsize, &group);
3563         if (rc)
3564                 GOTO(out, rc);
3565
3566         rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3567                           NULL, NULL, NULL);
3568         if (rc) {
3569                 CERROR("%s: failed at llog_origin_connect: %d\n",
3570                        obd->obd_name, rc);
3571                 GOTO(out, rc);
3572         }
3573
3574         /* remove the orphaned precreated objects */
3575         rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3576         if (rc)
3577                 GOTO(err_llog, rc);
3578
3579 out:
3580         RETURN(rc < 0 ? rc : item);
3581
3582 err_llog:
3583         /* cleanup all llogging subsystems */
3584         rc = obd_llog_finish(obd, &obd->obd_llogs,
3585                              mds->mds_dt_desc.ld_tgt_count);
3586         if (rc)
3587                 CERROR("%s: failed to cleanup llogging subsystems\n",
3588                         obd->obd_name);
3589         goto out;
3590 }
3591
3592 int mds_postrecov(struct obd_device *obd)
3593 {
3594         int rc;
3595         ENTRY;
3596         rc = mds_postrecov_common(obd);
3597         if (rc == 0)
3598                 rc = mds_md_reconnect(obd);
3599         RETURN(rc);
3600 }
3601
3602 int mds_dt_clean(struct obd_device *obd)
3603 {
3604         struct mds_obd *mds = &obd->u.mds;
3605         ENTRY;
3606
3607         if (mds->mds_profile) {
3608                 char * cln_prof;
3609                 struct llog_ctxt *llctx;
3610                 struct lvfs_run_ctxt saved;
3611                 struct config_llog_instance cfg;
3612                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3613
3614                 OBD_ALLOC(cln_prof, len);
3615                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3616
3617                 cfg.cfg_instance = NULL;
3618                 cfg.cfg_uuid = mds->mds_dt_uuid;
3619
3620                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3621                 llctx = llog_get_context(&obd->obd_llogs,
3622                                          LLOG_CONFIG_ORIG_CTXT);
3623                 class_config_process_llog(llctx, cln_prof, &cfg);
3624                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3625
3626                 OBD_FREE(cln_prof, len);
3627                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3628                 mds->mds_profile = NULL;
3629         }
3630         RETURN(0);
3631 }
3632
3633 int mds_md_clean(struct obd_device *obd)
3634 {
3635         struct mds_obd *mds = &obd->u.mds;
3636         ENTRY;
3637
3638         if (mds->mds_md_name) {
3639                 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3640                 mds->mds_md_name = NULL;
3641         }
3642         RETURN(0);
3643 }
3644
3645 static int mds_precleanup(struct obd_device *obd, int flags)
3646 {
3647         int rc = 0;
3648         ENTRY;
3649
3650         mds_md_clean(obd);
3651         mds_dt_disconnect(obd, flags);
3652         mds_dt_clean(obd);
3653         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3654         RETURN(rc);
3655 }
3656
3657 extern void lgss_svc_cache_purge_all(void);
3658 static int mds_cleanup(struct obd_device *obd, int flags)
3659 {
3660         struct mds_obd *mds = &obd->u.mds;
3661         ENTRY;
3662
3663         if (mds->mds_sb == NULL)
3664                 RETURN(0);
3665
3666         mds_update_server_data(obd, 1);
3667         mds_update_last_fid(obd, NULL, 1);
3668
3669         if (mds->mds_dt_objids != NULL) {
3670                 int size = mds->mds_dt_desc.ld_tgt_count *
3671                         sizeof(obd_id);
3672                 OBD_FREE(mds->mds_dt_objids, size);
3673         }
3674         mds_fs_cleanup(obd, flags);
3675
3676         unlock_kernel();
3677
3678         /* 2 seems normal on mds, (may_umount() also expects 2
3679           fwiw), but we only see 1 at this point in obdfilter. */
3680         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3681
3682         mds->mds_sb = 0;
3683
3684         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3685
3686         spin_lock_bh(&obd->obd_processing_task_lock);
3687         if (obd->obd_recovering) {
3688                 target_cancel_recovery_timer(obd);
3689                 obd->obd_recovering = 0;
3690         }
3691         spin_unlock_bh(&obd->obd_processing_task_lock);
3692
3693         lock_kernel();
3694         fsfilt_put_ops(obd->obd_fsops);
3695
3696 #ifdef ENABLE_GSS
3697         /* XXX */
3698         lgss_svc_cache_purge_all();
3699 #endif
3700
3701         spin_lock(&mds->mds_denylist_lock);
3702         while (!list_empty( &mds->mds_denylist ) ) {
3703                 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3704                                                     deny_sec_t, list);
3705                 list_del(&p_deny_sec->list);
3706                 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3707         }
3708         spin_unlock(&mds->mds_denylist_lock);
3709
3710         RETURN(0);
3711 }
3712
3713 static int set_security(const char *value, char **sec)
3714 {
3715         if (!strcmp(value, "null"))
3716                 *sec = "null";
3717         else if (!strcmp(value, "krb5i"))
3718                 *sec = "krb5i";
3719         else if (!strcmp(value, "krb5p"))
3720                 *sec = "krb5p";
3721         else {
3722                 CERROR("Unrecognized security flavor %s\n", value);
3723                 return -EINVAL;
3724         }
3725
3726         return 0;
3727 }
3728
3729 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3730 {
3731         struct lustre_cfg *lcfg = buf;
3732         struct mds_obd *mds = &obd->u.mds;
3733         int rc = 0;
3734         ENTRY;
3735
3736         switch(lcfg->lcfg_command) {
3737         case LCFG_SET_SECURITY: {
3738                 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
3739                     (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
3740                         GOTO(out, rc = -EINVAL);
3741
3742                 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
3743                         rc = set_security(lustre_cfg_string(lcfg, 2),
3744                                           &mds->mds_mds_sec);
3745                 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
3746                         rc = set_security(lustre_cfg_string(lcfg, 2),
3747                                           &mds->mds_ost_sec);
3748                 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
3749                         spin_lock(&mds->mds_denylist_lock);
3750                         rc = add_deny_security(lustre_cfg_string(lcfg, 2),
3751                                                &mds->mds_denylist);
3752                         spin_unlock(&mds->mds_denylist_lock);
3753                 } else {
3754                         CERROR("Unrecognized key\n");
3755                         rc = -EINVAL;
3756                 }
3757                 break;
3758         }
3759         default:
3760                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3761                 GOTO(out, rc = -EINVAL);
3762         }
3763 out:
3764         RETURN(rc);
3765 }
3766
3767 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3768                                         int offset,
3769                                         struct ldlm_lock *new_lock,
3770                                         struct ldlm_lock **old_lock,
3771                                         struct lustre_handle *lockh)
3772 {
3773         struct obd_export *exp = req->rq_export;
3774         struct obd_device *obd = exp->exp_obd;
3775         struct ldlm_request *dlmreq =
3776                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3777         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3778         struct list_head *iter;
3779
3780         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3781                 return;
3782
3783         spin_lock(&obd->obd_namespace->ns_hash_lock);
3784         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3785                 struct ldlm_lock *lock;
3786                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3787                 if (lock == new_lock)
3788                         continue;
3789                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3790                         lockh->cookie = lock->l_handle.h_cookie;
3791                         LDLM_DEBUG(lock, "restoring lock cookie");
3792                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3793                                   lockh->cookie);
3794                         if (old_lock)
3795                                 *old_lock = LDLM_LOCK_GET(lock);
3796                         spin_unlock(&obd->obd_namespace->ns_hash_lock);
3797                         return;
3798                 }
3799         }
3800         spin_unlock(&obd->obd_namespace->ns_hash_lock);
3801
3802         /* If the xid matches, then we know this is a resent request,
3803          * and allow it. (It's probably an OPEN, for which we don't
3804          * send a lock */
3805         if (req->rq_xid ==
3806             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3807                 return;
3808
3809         if (req->rq_xid ==
3810             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3811                 return;
3812
3813         /* This remote handle isn't enqueued, so we never received or
3814          * processed this request.  Clear MSG_RESENT, because it can
3815          * be handled like any normal request now. */
3816
3817         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3818
3819         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3820                   remote_hdl.cookie);
3821 }
3822
3823 int intent_disposition(struct ldlm_reply *rep, int flag)
3824 {
3825         if (!rep)
3826                 return 0;
3827         return (rep->lock_policy_res1 & flag);
3828 }
3829
3830 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3831 {
3832         if (!rep)
3833                 return;
3834         rep->lock_policy_res1 |= flag;
3835 }
3836
3837 static int mds_intent_policy(struct ldlm_namespace *ns,
3838                              struct ldlm_lock **lockp, void *req_cookie,
3839                              ldlm_mode_t mode, int flags, void *data)
3840 {
3841         struct ptlrpc_request *req = req_cookie;
3842         struct ldlm_lock *lock = *lockp;
3843         struct ldlm_intent *it;
3844         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3845         struct ldlm_reply *rep;
3846         struct lustre_handle lockh[2] = {{0}, {0}};
3847         struct ldlm_lock *new_lock = NULL;
3848         int getattr_part = MDS_INODELOCK_UPDATE;
3849         int rc, reply_buffers;
3850         int repsize[5] = {sizeof(struct ldlm_reply),
3851                           sizeof(struct mds_body),
3852                           mds->mds_max_mdsize};
3853
3854         int offset = MDS_REQ_INTENT_REC_OFF;
3855         ENTRY;
3856
3857         LASSERT(req != NULL);
3858         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3859
3860         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3861                 /* No intent was provided */
3862                 int size = sizeof(struct ldlm_reply);
3863                 rc = lustre_pack_reply(req, 1, &size, NULL);
3864                 LASSERT(rc == 0);
3865                 RETURN(0);
3866         }
3867
3868         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3869                                 lustre_swab_ldlm_intent);
3870         if (it == NULL) {
3871                 CERROR("Intent missing\n");
3872                 RETURN(req->rq_status = -EFAULT);
3873         }
3874
3875         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3876
3877         reply_buffers = 3;
3878         if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3879                 reply_buffers = 5;
3880                 repsize[3] = 4;
3881                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
3882         }
3883
3884         rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
3885         if (rc)
3886                 RETURN(req->rq_status = rc);
3887
3888         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3889         LASSERT(rep != NULL);
3890
3891         intent_set_disposition(rep, DISP_IT_EXECD);
3892
3893         /* execute policy */
3894         switch ((long)it->opc) {
3895         case IT_OPEN:
3896         case IT_CREAT|IT_OPEN:
3897                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3898                                             lock, NULL, lockh);
3899                 /* XXX swab here to assert that an mds_open reint
3900                  * packet is following */
3901                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3902                                             lock, NULL, lockh);
3903                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3904
3905                 if (rep->lock_policy_res2) {
3906                         /*
3907                          * mds_open() returns ENOLCK where it should return
3908                          * zero, but it has no lock to return.
3909                          */
3910                         if (rep->lock_policy_res2 == ENOLCK)
3911                                 rep->lock_policy_res2 = 0;
3912
3913                         RETURN(ELDLM_LOCK_ABORTED);
3914                 }
3915
3916                 /*
3917                  * IT_OPEN may return lock on cross-node dentry that we want to
3918                  * hold during attr retrival -bzzz
3919                  */
3920                 if (lockh[0].cookie == 0)
3921                         RETURN(ELDLM_LOCK_ABORTED);
3922
3923                 break;
3924         case IT_LOOKUP:
3925                 getattr_part = MDS_INODELOCK_LOOKUP;
3926         case IT_CHDIR:
3927         case IT_GETATTR:
3928                 getattr_part |= MDS_INODELOCK_LOOKUP;
3929         case IT_READDIR:
3930                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3931                                             lock, &new_lock, lockh);
3932                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3933                                                          getattr_part);
3934                 /* FIXME: LDLM can set req->rq_status. MDS sets
3935                    policy_res{1,2} with disposition and status.
3936                    - replay: returns 0 & req->status is old status
3937                    - otherwise: returns req->status */
3938                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3939                         rep->lock_policy_res2 = 0;
3940                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3941                     rep->lock_policy_res2)
3942                         RETURN(ELDLM_LOCK_ABORTED);
3943                 if (req->rq_status != 0) {
3944                         LBUG();
3945                         rep->lock_policy_res2 = req->rq_status;
3946                         RETURN(ELDLM_LOCK_ABORTED);
3947                 }
3948                 break;
3949         case IT_UNLINK:
3950                 rc = mds_lock_and_check_slave(offset, req, lockh);
3951                 if ((rep->lock_policy_res2 = rc)) {
3952                         if (rc == ENOLCK)
3953                                 rep->lock_policy_res2 = 0;
3954                         RETURN(ELDLM_LOCK_ABORTED);
3955                 }
3956                 break;
3957         default:
3958                 CERROR("Unhandled intent "LPD64"\n", it->opc);
3959                 LBUG();
3960         }
3961
3962         /* By this point, whatever function we called above must have either
3963          * filled in 'lockh', been an intent replay, or returned an error.  We
3964          * want to allow replayed RPCs to not get a lock, since we would just
3965          * drop it below anyways because lock replay is done separately by the
3966          * client afterwards.  For regular RPCs we want to give the new lock to
3967          * the client instead of whatever lock it was about to get. */
3968         if (new_lock == NULL)
3969                 new_lock = ldlm_handle2lock(&lockh[0]);
3970         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3971                 RETURN(0);
3972
3973         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3974                  it->opc, lockh[0].cookie);
3975
3976         /* If we've already given this lock to a client once, then we should
3977          * have no readers or writers.  Otherwise, we should have one reader
3978          * _or_ writer ref (which will be zeroed below) before returning the
3979          * lock to a client. */
3980         if (new_lock->l_export == req->rq_export) {
3981                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3982         } else {
3983                 LASSERT(new_lock->l_export == NULL);
3984                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3985         }
3986
3987         *lockp = new_lock;
3988
3989         if (new_lock->l_export == req->rq_export) {
3990                 /* Already gave this to the client, which means that we
3991                  * reconstructed a reply. */
3992                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3993                         MSG_RESENT);
3994                 RETURN(ELDLM_LOCK_REPLACED);
3995         }
3996
3997         /* Fixup the lock to be given to the client */
3998         lock_res(new_lock->l_resource);
3999         new_lock->l_readers = 0;
4000         new_lock->l_writers = 0;
4001
4002         new_lock->l_export = class_export_get(req->rq_export);
4003
4004         spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4005         list_add(&new_lock->l_export_chain,
4006                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
4007         spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4008
4009         new_lock->l_blocking_ast = lock->l_blocking_ast;
4010         new_lock->l_completion_ast = lock->l_completion_ast;
4011
4012         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4013                sizeof(lock->l_remote_handle));
4014
4015         new_lock->l_flags &= ~LDLM_FL_LOCAL;
4016
4017         unlock_res(new_lock->l_resource);
4018         LDLM_LOCK_PUT(new_lock);
4019
4020         RETURN(ELDLM_LOCK_REPLACED);
4021 }
4022
4023 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4024 {
4025         struct lprocfs_static_vars lvars;
4026         int rc = 0;
4027         struct mds_obd *mds = &dev->u.mds;
4028
4029         spin_lock_init(&mds->mds_denylist_lock);
4030         INIT_LIST_HEAD(&mds->mds_denylist);
4031
4032         lprocfs_init_multi_vars(0, &lvars);
4033
4034         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4035         if (rc)
4036                 return rc;
4037
4038         return lprocfs_alloc_md_stats(dev, 0);
4039 }
4040
4041 int mds_detach(struct obd_device *dev)
4042 {
4043         lprocfs_free_md_stats(dev);
4044         return lprocfs_obd_detach(dev);
4045 }
4046
4047 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4048 {
4049         struct lprocfs_static_vars lvars;
4050
4051         lprocfs_init_multi_vars(1, &lvars);
4052         return lprocfs_obd_attach(dev, lvars.obd_vars);
4053 }
4054
4055 int mdt_detach(struct obd_device *dev)
4056 {
4057         return lprocfs_obd_detach(dev);
4058 }
4059
4060 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4061 {
4062         struct mds_obd *mds = &obd->u.mds;
4063         int rc = 0;
4064         ENTRY;
4065
4066         mds->mds_service =
4067                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4068                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4069                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4070                                 mds_handle, "mds", obd->obd_proc_entry);
4071
4072         if (!mds->mds_service) {
4073                 CERROR("failed to start service\n");
4074                 RETURN(-ENOMEM);
4075         }
4076
4077         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4078                                     "ll_mdt");
4079         if (rc)
4080                 GOTO(err_thread, rc);
4081
4082         mds->mds_setattr_service =
4083                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4084                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4085                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4086                                 mds_handle, "mds_setattr",
4087                                 obd->obd_proc_entry);
4088         if (!mds->mds_setattr_service) {
4089                 CERROR("failed to start getattr service\n");
4090                 GOTO(err_thread, rc = -ENOMEM);
4091         }
4092
4093         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4094                                     MDT_NUM_THREADS, "ll_mdt_attr");
4095         if (rc)
4096                 GOTO(err_thread2, rc);
4097
4098         mds->mds_readpage_service =
4099                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4100                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4101                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4102                                 mds_handle, "mds_readpage",
4103                                 obd->obd_proc_entry);
4104         if (!mds->mds_readpage_service) {
4105                 CERROR("failed to start readpage service\n");
4106                 GOTO(err_thread2, rc = -ENOMEM);
4107         }
4108
4109         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4110                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
4111
4112         if (rc)
4113                 GOTO(err_thread3, rc);
4114
4115         RETURN(0);
4116
4117 err_thread3:
4118         ptlrpc_unregister_service(mds->mds_readpage_service);
4119 err_thread2:
4120         ptlrpc_unregister_service(mds->mds_setattr_service);
4121 err_thread:
4122         ptlrpc_unregister_service(mds->mds_service);
4123         return rc;
4124 }
4125
4126 static int mdt_cleanup(struct obd_device *obd, int flags)
4127 {
4128         struct mds_obd *mds = &obd->u.mds;
4129         ENTRY;
4130
4131         ptlrpc_stop_all_threads(mds->mds_readpage_service);
4132         ptlrpc_unregister_service(mds->mds_readpage_service);
4133
4134         ptlrpc_stop_all_threads(mds->mds_setattr_service);
4135         ptlrpc_unregister_service(mds->mds_setattr_service);
4136
4137         ptlrpc_stop_all_threads(mds->mds_service);
4138         ptlrpc_unregister_service(mds->mds_service);
4139
4140         RETURN(0);
4141 }
4142
4143 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4144                                          __u64 gr, void *data)
4145 {
4146         struct lustre_id id;
4147         struct obd_device *obd = data;
4148
4149         id_ino(&id) = ino;
4150         id_gen(&id) = gen;
4151         return mds_id2dentry(obd, &id, NULL);
4152 }
4153
4154 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4155                         void *key, __u32 *valsize, void *val)
4156 {
4157         struct obd_device *obd;
4158         struct mds_obd *mds;
4159         ENTRY;
4160
4161         obd = class_exp2obd(exp);
4162         mds = &obd->u.mds;
4163
4164         if (obd == NULL) {
4165                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4166                        exp->exp_handle.h_cookie);
4167                 RETURN(-EINVAL);
4168         }
4169
4170         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4171                 /* get log_context handle. */
4172                 unsigned long *llh_handle = val;
4173                 *valsize = sizeof(unsigned long);
4174                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
4175                 RETURN(0);
4176         }
4177         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4178                 /* get log_context handle. */
4179                 unsigned long *sb = val;
4180                 *valsize = sizeof(unsigned long);
4181                 *sb = (unsigned long)obd->u.mds.mds_sb;
4182                 RETURN(0);
4183         }
4184
4185         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4186                 __u32 *mdsize = val;
4187                 *valsize = sizeof(*mdsize);
4188                 *mdsize = mds->mds_max_mdsize;
4189                 RETURN(0);
4190         }
4191
4192         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4193                 __u32 *mdsnum = val;
4194                 *valsize = sizeof(*mdsnum);
4195                 *mdsnum = mds->mds_num;
4196                 RETURN(0);
4197         }
4198
4199         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4200                 struct lustre_id *rootid = val;
4201                 *valsize = sizeof(struct lustre_id);
4202                 *rootid = mds->mds_rootid;
4203                 RETURN(0);
4204         }
4205
4206         CDEBUG(D_IOCTL, "invalid key\n");
4207         RETURN(-EINVAL);
4208
4209 }
4210 struct lvfs_callback_ops mds_lvfs_ops = {
4211         l_id2dentry:     mds_lvfs_id2dentry,
4212 };
4213
4214 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4215                 int objcount, struct obd_ioobj *obj,
4216                 int niocount, struct niobuf_remote *nb,
4217                 struct niobuf_local *res,
4218                 struct obd_trans_info *oti);
4219
4220 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4221                  int objcount, struct obd_ioobj *obj, int niocount,
4222                  struct niobuf_local *res, struct obd_trans_info *oti,
4223                  int rc);
4224
4225 /* use obd ops to offer management infrastructure */
4226 static struct obd_ops mds_obd_ops = {
4227         .o_owner           = THIS_MODULE,
4228         .o_attach          = mds_attach,
4229         .o_detach          = mds_detach,
4230         .o_connect         = mds_connect,
4231         .o_connect_post    = mds_connect_post,
4232         .o_init_export     = mds_init_export,
4233         .o_destroy_export  = mds_destroy_export,
4234         .o_disconnect      = mds_disconnect,
4235         .o_setup           = mds_setup,
4236         .o_precleanup      = mds_precleanup,
4237         .o_cleanup         = mds_cleanup,
4238         .o_process_config  = mds_process_config,
4239         .o_postrecov       = mds_postrecov,
4240         .o_statfs          = mds_obd_statfs,
4241         .o_iocontrol       = mds_iocontrol,
4242         .o_create          = mds_obd_create,
4243         .o_destroy         = mds_obd_destroy,
4244         .o_llog_init       = mds_llog_init,
4245         .o_llog_finish     = mds_llog_finish,
4246         .o_notify          = mds_notify,
4247         .o_get_info        = mds_get_info,
4248         .o_set_info        = mds_set_info,
4249         .o_preprw          = mds_preprw,
4250         .o_commitrw        = mds_commitrw,
4251 };
4252
4253 static struct obd_ops mdt_obd_ops = {
4254         .o_owner           = THIS_MODULE,
4255         .o_attach          = mdt_attach,
4256         .o_detach          = mdt_detach,
4257         .o_setup           = mdt_setup,
4258         .o_cleanup         = mdt_cleanup,
4259 };
4260
4261 static int __init mds_init(void)
4262 {
4263         struct lprocfs_static_vars lvars;
4264
4265         mds_init_lsd_cache();
4266
4267         lprocfs_init_multi_vars(0, &lvars);
4268         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4269                             LUSTRE_MDS_NAME);
4270         lprocfs_init_multi_vars(1, &lvars);
4271         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4272                             LUSTRE_MDT_NAME);
4273
4274         return 0;
4275 }
4276
4277 static void /*__exit*/ mds_exit(void)
4278 {
4279         mds_cleanup_lsd_cache();
4280
4281         class_unregister_type(LUSTRE_MDS_NAME);
4282         class_unregister_type(LUSTRE_MDT_NAME);
4283 }
4284
4285 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4286 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4287 MODULE_LICENSE("GPL");
4288
4289 module_init(mds_init);
4290 module_exit(mds_exit);