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 <linux/lustre_sec.h>
  60 #include <linux/lustre_gs.h>
  61 #include "mds_internal.h"
  62
  63 static int mds_intent_policy(struct ldlm_namespace *ns,
  64                              struct ldlm_lock **lockp, void *req_cookie,
  65                              ldlm_mode_t mode, int flags, void *data);
  66 static int mds_postsetup(struct obd_device *obd);
  67 static int mds_cleanup(struct obd_device *obd, int flags);
  68
  69
  70 /* Assumes caller has already pushed into the kernel filesystem context */
  71 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  72                         loff_t offset, int count)
  73 {
  74         struct ptlrpc_bulk_desc *desc;
  75         struct l_wait_info lwi;
  76         struct page **pages;
  77         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  78         ENTRY;
  79
  80         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  81
  82         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  83         OBD_ALLOC(pages, sizeof(*pages) * npages);
  84         if (!pages)
  85                 GOTO(out, rc = -ENOMEM);
  86
  87         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  88                                     MDS_BULK_PORTAL);
  89         if (desc == NULL)
  90                 GOTO(out_free, rc = -ENOMEM);
  91
  92         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  93                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  94
  95                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  96                 if (pages[i] == NULL)
  97                         GOTO(cleanup_buf, rc = -ENOMEM);
  98
  99                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 100         }
 101
 102         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 103                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 104                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 105                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 106                        file->f_dentry->d_inode->i_size);
 107
 108                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 109                                      kmap(pages[i]), tmpsize, &offset);
 110                 kunmap(pages[i]);
 111
 112                 if (rc != tmpsize)
 113                         GOTO(cleanup_buf, rc = -EIO);
 114         }
 115
 116         LASSERT(desc->bd_nob == count);
 117
 118         rc = ptlrpc_start_bulk_transfer(desc);
 119         if (rc)
 120                 GOTO(cleanup_buf, rc);
 121
 122         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 123                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 124                        OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
 125                 GOTO(abort_bulk, rc);
 126         }
 127
 128         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 129         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 130         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 131
 132         if (rc == 0) {
 133                 if (desc->bd_success &&
 134                     desc->bd_nob_transferred == count)
 135                         GOTO(cleanup_buf, rc);
 136
 137                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 138         }
 139
 140         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 141                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 142                   desc->bd_nob_transferred, count,
 143                   req->rq_export->exp_client_uuid.uuid,
 144                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 145
 146         ptlrpc_fail_export(req->rq_export);
 147
 148         EXIT;
 149  abort_bulk:
 150         ptlrpc_abort_bulk (desc);
 151  cleanup_buf:
 152         for (i = 0; i < npages; i++)
 153                 if (pages[i])
 154                         __free_pages(pages[i], 0);
 155
 156         ptlrpc_free_bulk(desc);
 157  out_free:
 158         OBD_FREE(pages, sizeof(*pages) * npages);
 159  out:
 160         return rc;
 161 }
 162
 163 extern char *ldlm_lockname[];
 164
 165 int mds_lock_mode_for_dir(struct obd_device *obd,
 166                           struct dentry *dentry, int mode)
 167 {
 168         int ret_mode = 0, split;
 169
 170         /* any dir access needs couple locks:
 171          * 1) on part of dir we gonna lookup/modify in
 172          * 2) on a whole dir to protect it from concurrent splitting
 173          *    and to flush client's cache for readdir()
 174          * so, for a given mode and dentry this routine decides what
 175          * lock mode to use for lock #2:
 176          * 1) if caller's gonna lookup in dir then we need to protect
 177          *    dir from being splitted only - LCK_CR
 178          * 2) if caller's gonna modify dir then we need to protect
 179          *    dir from being splitted and to flush cache - LCK_CW
 180          * 3) if caller's gonna modify dir and that dir seems ready
 181          *    for splitting then we need to protect it from any
 182          *    type of access (lookup/modify/split) - LCK_EX -bzzz */
 183
 184         split = mds_splitting_expected(obd, dentry);
 185
 186         /*
 187          * it is important to check here only for MDS_NO_SPLITTABLE. The reason
 188          * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
 189          * and another thread will not split it on the quiet. But if we have
 190          * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
 191          * but not now (for current thread) and we should consider that it can
 192          * happen soon and go that branch which can yield LCK_EX to protect from
 193          * possible splitting.
 194          */
 195         if (split == MDS_NO_SPLITTABLE) {
 196                 /*
 197                  * this inode won't be splitted. so we need not to protect from
 198                  * just flush client's cache on modification.
 199                  */
 200                 if (mode == LCK_PW)
 201                         ret_mode = LCK_CW;
 202                 else
 203                         ret_mode = 0;
 204         } else {
 205                 if (mode == LCK_EX) {
 206                         ret_mode = LCK_EX;
 207                 } else if (mode == LCK_PR) {
 208                         ret_mode = LCK_CR;
 209                 } else if (mode == LCK_PW) {
 210                         /*
 211                          * caller gonna modify directory. We use concurrent
 212                          * write lock here to retract client's cache for
 213                          * readdir.
 214                          */
 215                         if (split == MDS_EXPECT_SPLIT) {
 216                                 /*
 217                                  * splitting possible. serialize any access the
 218                                  * idea is that first one seen dir is splittable
 219                                  * is given exclusive lock and split
 220                                  * directory. caller passes lock mode to
 221                                  * mds_try_to_split_dir() and splitting would be
 222                                  * done with exclusive lock only -bzzz.
 223                                  */
 224                                 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
 225                                        obd->obd_name,
 226                                        (unsigned long)dentry->d_inode->i_ino,
 227                                        (unsigned long)dentry->d_inode->i_generation);
 228                                 ret_mode = LCK_EX;
 229                         } else {
 230                                 ret_mode = LCK_CW;
 231                         }
 232                 }
 233         }
 234
 235         return ret_mode;
 236 }
 237
 238 /* only valid locked dentries or errors should be returned */
 239 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
 240                                     struct vfsmount **mnt, int lock_mode,
 241                                     struct lustre_handle *lockh, int *mode,
 242                                     char *name, int namelen, __u64 lockpart)
 243 {
 244         struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
 245         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 246         struct ldlm_res_id res_id = { .name = {0} };
 247         int flags = LDLM_FL_ATOMIC_CB, rc;
 248         ENTRY;
 249
 250         if (IS_ERR(de))
 251                 RETURN(de);
 252
 253         lockh[1].cookie = 0;
 254         res_id.name[0] = id_fid(id);
 255         res_id.name[1] = id_group(id);
 256
 257 #ifdef S_PDIROPS
 258         if (name && IS_PDIROPS(de->d_inode)) {
 259                 ldlm_policy_data_t cpolicy =
 260                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 261                 LASSERT(mode != NULL);
 262                 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
 263                 if (*mode) {
 264                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 265                                               res_id, LDLM_IBITS,
 266                                               &cpolicy, *mode, &flags,
 267                                               mds_blocking_ast,
 268                                               ldlm_completion_ast, NULL, NULL,
 269                                               NULL, 0, NULL, lockh + 1);
 270                         if (rc != ELDLM_OK) {
 271                                 l_dput(de);
 272                                 RETURN(ERR_PTR(-ENOLCK));
 273                         }
 274                 }
 275                 flags = LDLM_FL_ATOMIC_CB;
 276
 277                 res_id.name[2] = full_name_hash((unsigned char *)name, namelen);
 278
 279                 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
 280                        OLID4(id), res_id.name[2]);
 281         }
 282 #else
 283 #warning "No PDIROPS support in the kernel"
 284 #endif
 285         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 286                               LDLM_IBITS, &policy, lock_mode, &flags,
 287                               mds_blocking_ast, ldlm_completion_ast,
 288                               NULL, NULL, NULL, 0, NULL, lockh);
 289         if (rc != ELDLM_OK) {
 290                 l_dput(de);
 291                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 292 #ifdef S_PDIROPS
 293                 if (lockh[1].cookie)
 294                         ldlm_lock_decref(lockh + 1, *mode);
 295 #endif
 296         } else if (de->d_inode && de->d_inode->i_nlink == 0) {
 297                 /* as sometimes we lookup inode by ino/generation through
 298                    iopen mechanism, it's possible to find already unlinked
 299                    inode with nlink == 0. let's interpretate the case as
 300                    ENOENT -bzzz */
 301                 CWARN("found already unlinked inode %lu/%u\n",
 302                       de->d_inode->i_ino, de->d_inode->i_generation);
 303                 l_dput(de);
 304                 retval = ERR_PTR(-ENOENT);
 305                 ldlm_lock_decref(lockh, lock_mode);
 306 #ifdef S_PDIROPS
 307                 if (lockh[1].cookie)
 308                         ldlm_lock_decref(lockh + 1, *mode);
 309 #endif
 310         }
 311
 312         RETURN(retval);
 313 }
 314
 315 #ifndef DCACHE_DISCONNECTED
 316 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 317 #endif
 318
 319 /* Look up an entry by inode number. This function ONLY returns valid dget'd
 320  * dentries with an initialized inode or errors */
 321 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
 322                              struct vfsmount **mnt)
 323 {
 324         struct mds_obd *mds = &obd->u.mds;
 325         struct dentry *result;
 326         struct inode *inode;
 327         unsigned long ino = 0;
 328         __u32 generation = 0;
 329         char idname[32];
 330
 331         if (!id_ino(id) && id_fid(id)) {
 332                 struct lustre_id *lid;
 333
 334                 /* if this is reint case we should use fidmap for resolving
 335                  * correct local store cookie. */
 336                 lid = mds_fidmap_lookup(obd, id);
 337                 if (!lid)
 338                         RETURN(ERR_PTR(-ENOENT));
 339
 340                 ino = (unsigned long)id_ino(lid);
 341                 generation = (__u32)id_gen(lid);
 342
 343                 CDEBUG(D_DENTRY, "fidmap resolved "DLID4"->"DLID4"\n",
 344                        OLID4(id), OLID4(lid));
 345
 346         } else if (id_ino(id)) {
 347                 ino = (unsigned long)id_ino(id);
 348                 generation = (__u32)id_gen(id);
 349         } else {
 350                 CERROR("invalid id for lookup "
 351                        DLID4"\n", OLID4(id));
 352                 LBUG();
 353         }
 354
 355         if (ino == 0)
 356                 RETURN(ERR_PTR(-ESTALE));
 357
 358         snprintf(idname, sizeof(idname), "0x%lx", ino);
 359
 360         CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
 361                ino, generation, mds->mds_sb);
 362
 363         /* under ext3 this is neither supposed to return bad inodes nor NULL
 364          * inodes. */
 365         result = ll_lookup_one_len(idname, mds->mds_id_de,
 366                                    strlen(idname));
 367         if (IS_ERR(result))
 368                 RETURN(result);
 369
 370         inode = result->d_inode;
 371         if (!inode)
 372                 RETURN(ERR_PTR(-ENOENT));
 373
 374         if (is_bad_inode(inode)) {
 375                 CERROR("bad inode returned %lu/%u\n",
 376                        inode->i_ino, inode->i_generation);
 377                 dput(result);
 378                 RETURN(ERR_PTR(-ENOENT));
 379         }
 380
 381         /* here we disabled generation check, as root inode i_generation
 382          * of cache mds and real mds are different. */
 383         if (id_fid(id) != id_fid(&mds->mds_rootid) && generation != 0 &&
 384             inode->i_generation != generation) {
 385                 /* we didn't find the right inode.. */
 386                 if (id_group(id) != mds->mds_num) {
 387                         CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
 388                                "%u != %u, mds %u != %u, request to wrong MDS?\n",
 389                                inode->i_ino, (unsigned long)inode->i_nlink,
 390                                atomic_read(&inode->i_count), inode->i_generation,
 391                                generation, mds->mds_num, (unsigned)id_group(id));
 392                 } else {
 393                         CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
 394                                "%u != %u, inode is recreated while request handled?\n",
 395                                inode->i_ino, (unsigned long)inode->i_nlink,
 396                                atomic_read(&inode->i_count), inode->i_generation,
 397                                generation);
 398                 }
 399                 dput(result);
 400                 RETURN(ERR_PTR(-ENOENT));
 401         }
 402
 403         if (mnt) {
 404                 *mnt = mds->mds_vfsmnt;
 405                 mntget(*mnt);
 406         }
 407
 408         RETURN(result);
 409 }
 410
 411 static
 412 int mds_req_add_idmapping(struct ptlrpc_request *req,
 413                           struct mds_export_data *med)
 414 {
 415         struct mds_req_sec_desc *rsd;
 416         struct lustre_sec_desc  *lsd;
 417         int rc;
 418
 419         if (!med->med_remote)
 420                 return 0;
 421
 422         /* maybe we should do it more completely: invalidate the gss ctxt? */
 423         if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
 424                 CWARN("didn't find mapped uid\n");
 425                 return -EPERM;
 426         }
 427
 428         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 429         if (!rsd) {
 430                 CERROR("Can't unpack security desc\n");
 431                 return -EPROTO;
 432         }
 433
 434         lsd = mds_get_lsd(req->rq_mapped_uid);
 435         if (!lsd) {
 436                 CERROR("can't get LSD(%u), no mapping added\n",
 437                        req->rq_mapped_uid);
 438                 return -EPERM;
 439         }
 440
 441         rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 442                            rsd->rsd_gid, lsd->lsd_gid);
 443         mds_put_lsd(lsd);
 444         return rc;
 445 }
 446
 447 static
 448 int mds_req_del_idmapping(struct ptlrpc_request *req,
 449                           struct mds_export_data *med)
 450 {
 451         struct mds_req_sec_desc *rsd;
 452         struct lustre_sec_desc  *lsd;
 453         int rc;
 454
 455         if (!med->med_remote)
 456                 return 0;
 457
 458         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 459         if (!rsd) {
 460                 CERROR("Can't unpack security desc\n");
 461                 return -EPROTO;
 462         }
 463
 464         LASSERT(req->rq_mapped_uid != -1);
 465         lsd = mds_get_lsd(req->rq_mapped_uid);
 466         if (!lsd) {
 467                 CERROR("can't get LSD(%u), no idmapping deleted\n",
 468                        req->rq_mapped_uid);
 469                 return -EPERM;
 470         }
 471
 472         rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 473                            rsd->rsd_gid, lsd->lsd_gid);
 474         mds_put_lsd(lsd);
 475         return rc;
 476 }
 477
 478 static int mds_init_export_data(struct ptlrpc_request *req,
 479                                 struct mds_export_data *med)
 480 {
 481         struct obd_connect_data *data, *reply;
 482         int ask_remote, ask_local;
 483         ENTRY;
 484
 485         data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
 486         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
 487         LASSERT(data && reply);
 488
 489         if (med->med_initialized) {
 490                 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
 491                 goto reply;
 492         }
 493
 494         ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
 495         ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
 496
 497         /* currently the policy is simple: satisfy client as possible
 498          * as we can.
 499          */
 500         if (req->rq_auth_uid == -1) {
 501                 if (ask_remote)
 502                         CWARN("null sec is used, force to be local\n");
 503                 med->med_remote = 0;
 504         } else {
 505                 if (ask_remote) {
 506                         if (!req->rq_remote_realm)
 507                                 CWARN("local realm asked to be remote\n");
 508                         med->med_remote = 1;
 509                 } else if (ask_local) {
 510                         if (req->rq_remote_realm)
 511                                 CWARN("remote realm asked to be local\n");
 512                         med->med_remote = 0;
 513                 } else
 514                         med->med_remote = (req->rq_remote_realm != 0);
 515         }
 516
 517         med->med_nllu = data->ocd_nllu[0];
 518         med->med_nllg = data->ocd_nllu[1];
 519
 520         med->med_initialized = 1;
 521 reply:
 522         reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
 523         if (med->med_remote) {
 524                 if (!med->med_idmap)
 525                         med->med_idmap = mds_idmap_alloc();
 526
 527                 if (!med->med_idmap)
 528                         CERROR("Failed to alloc idmap, following request from "
 529                                "this client will be refused\n");
 530
 531                 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
 532                 CDEBUG(D_SEC, "set client as remote\n");
 533         } else {
 534                 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
 535                 CDEBUG(D_SEC, "set client as local\n");
 536         }
 537
 538         RETURN(0);
 539 }
 540
 541 static void mds_free_export_data(struct mds_export_data *med)
 542 {
 543         if (!med->med_idmap)
 544                 return;
 545
 546         LASSERT(med->med_remote);
 547         mds_idmap_free(med->med_idmap);
 548         med->med_idmap = NULL;
 549 }
 550
 551 /* Establish a connection to the MDS.
 552  *
 553  * This will set up an export structure for the client to hold state data about
 554  * that client, like open files, the last operation number it did on the server,
 555  * etc.
 556  */
 557 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 558                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 559                        unsigned long flags)
 560 {
 561         struct mds_export_data *med;
 562         struct mds_client_data *mcd;
 563         struct obd_export *exp;
 564         int rc;
 565         ENTRY;
 566
 567         if (!conn || !obd || !cluuid)
 568                 RETURN(-EINVAL);
 569
 570         /* XXX There is a small race between checking the list and adding a new
 571          * connection for the same UUID, but the real threat (list corruption
 572          * when multiple different clients connect) is solved.
 573          *
 574          * There is a second race between adding the export to the list, and
 575          * filling in the client data below.  Hence skipping the case of NULL
 576          * mcd above.  We should already be controlling multiple connects at the
 577          * client, and we can't hold the spinlock over memory allocations
 578          * without risk of deadlocking.
 579          */
 580         rc = class_connect(conn, obd, cluuid);
 581         if (rc)
 582                 RETURN(rc);
 583         exp = class_conn2export(conn);
 584
 585         LASSERT(exp != NULL);
 586         med = &exp->exp_mds_data;
 587
 588         OBD_ALLOC(mcd, sizeof(*mcd));
 589         if (!mcd) {
 590                 CERROR("%s: out of memory for client data.\n",
 591                         obd->obd_name);
 592                 GOTO(out, rc = -ENOMEM);
 593         }
 594
 595         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 596         med->med_mcd = mcd;
 597
 598         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 599         if (rc)
 600                 GOTO(out, rc);
 601
 602         EXIT;
 603 out:
 604         if (rc) {
 605                 if (mcd)
 606                         OBD_FREE(mcd, sizeof(*mcd));
 607                 class_disconnect(exp, 0);
 608         } else {
 609                 class_export_put(exp);
 610         }
 611         return rc;
 612 }
 613
 614 static int mds_connect_post(struct obd_export *exp, unsigned initial,
 615                             unsigned long flags)
 616 {
 617         struct obd_device *obd = exp->exp_obd;
 618         struct mds_obd *mds = &obd->u.mds;
 619         struct mds_export_data *med;
 620         struct mds_client_data *mcd;
 621         int rc = 0;
 622         ENTRY;
 623
 624         med = &exp->exp_mds_data;
 625         mcd = med->med_mcd;
 626
 627         if (initial) {
 628                 /* some one reconnect initially, we have to reset
 629                  * data existing export can have. bug 6102 */
 630                 if (mcd->mcd_last_xid != 0)
 631                         CDEBUG(D_HA, "initial reconnect to existing export\n");
 632                 mcd->mcd_last_transno = 0;
 633                 mcd->mcd_last_xid = 0;
 634                 mcd->mcd_last_close_xid = 0;
 635                 mcd->mcd_last_result = 0;
 636                 mcd->mcd_last_data = 0;
 637         }
 638
 639         if (!(flags & OBD_OPT_MDS_CONNECTION)) {
 640                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 641                         atomic_inc(&mds->mds_real_clients);
 642                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 643                                obd->obd_name, exp->exp_client_uuid.uuid,
 644                                atomic_read(&mds->mds_real_clients));
 645                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 646                 }
 647                 if (mds->mds_md_name)
 648                         rc = mds_md_connect(obd, mds->mds_md_name);
 649         }
 650         RETURN(rc);
 651 }
 652
 653 static int mds_init_export(struct obd_export *exp)
 654 {
 655         struct mds_export_data *med = &exp->exp_mds_data;
 656
 657         INIT_LIST_HEAD(&med->med_open_head);
 658         spin_lock_init(&med->med_open_lock);
 659         return 0;
 660 }
 661
 662 static int mds_destroy_export(struct obd_export *export)
 663 {
 664         struct obd_device *obd = export->exp_obd;
 665         struct mds_export_data *med = &export->exp_mds_data;
 666         struct lvfs_run_ctxt saved;
 667         int rc = 0;
 668         ENTRY;
 669
 670         mds_free_export_data(med);
 671         target_destroy_export(export);
 672
 673         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 674                 GOTO(out, 0);
 675
 676         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 677
 678         /* Close any open files (which may also cause orphan unlinking). */
 679         spin_lock(&med->med_open_lock);
 680         while (!list_empty(&med->med_open_head)) {
 681                 struct list_head *tmp = med->med_open_head.next;
 682                 struct mds_file_data *mfd =
 683                         list_entry(tmp, struct mds_file_data, mfd_list);
 684                 struct lustre_id sid;
 685
 686                 BDEVNAME_DECLARE_STORAGE(btmp);
 687
 688                 /* bug 1579: fix force-closing for 2.5 */
 689                 struct dentry *dentry = mfd->mfd_dentry;
 690
 691                 list_del(&mfd->mfd_list);
 692                 spin_unlock(&med->med_open_lock);
 693
 694                 down(&dentry->d_inode->i_sem);
 695                 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
 696                 up(&dentry->d_inode->i_sem);
 697                 if (rc) {
 698                         CERROR("Can't read inode self id, inode %lu, "
 699                                "rc %d\n", dentry->d_inode->i_ino, rc);
 700                         memset(&sid, 0, sizeof(sid));
 701                 }
 702
 703                 /* If you change this message, be sure to update
 704                  * replay_single:test_46 */
 705                 CERROR("force closing client file handle for %.*s (%s:"
 706                        DLID4")\n", dentry->d_name.len, dentry->d_name.name,
 707                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 708                        OLID4(&sid));
 709
 710                 /* child inode->i_alloc_sem protects orphan_dec_test and
 711                  * is_orphan race, mds_mfd_close drops it */
 712                 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 713                 rc = mds_mfd_close(NULL, 0, obd, mfd,
 714                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 715                 if (rc)
 716                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 717                 spin_lock(&med->med_open_lock);
 718         }
 719         spin_unlock(&med->med_open_lock);
 720         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 721
 722         EXIT;
 723 out:
 724         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 725         return rc;
 726 }
 727
 728 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
 729 {
 730         unsigned long irqflags;
 731         struct obd_device *obd;
 732         struct mds_obd *mds;
 733         int rc;
 734         ENTRY;
 735
 736         LASSERT(exp != NULL);
 737         obd = class_exp2obd(exp);
 738         if (obd == NULL) {
 739                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 740                        exp->exp_handle.h_cookie);
 741                 RETURN(-EINVAL);
 742         }
 743         mds = &obd->u.mds;
 744
 745         /*
 746          * suppress any inter-mds requests durring disconnecting lmv if this is
 747          * detected --force mode. This is needed to avoid endless recovery.
 748          */
 749         if (atomic_read(&mds->mds_real_clients) > 0 &&
 750             !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
 751                 flags |= OBD_OPT_FORCE;
 752
 753         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 754             && !atomic_read(&mds->mds_real_clients)) {
 755                 /* there was no client at all */
 756                 mds_md_disconnect(obd, flags);
 757         }
 758
 759         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 760             && atomic_dec_and_test(&mds->mds_real_clients)) {
 761                 /* time to drop LMV connections */
 762                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 763                        "Disconnnect from LMV now\n",
 764                        obd->obd_name, exp->exp_client_uuid.uuid);
 765                 mds_md_disconnect(obd, flags);
 766         }
 767
 768         spin_lock_irqsave(&exp->exp_lock, irqflags);
 769         exp->exp_flags = flags;
 770         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 771
 772         /* disconnect early so that clients can't keep using export */
 773         rc = class_disconnect(exp, flags);
 774         ldlm_cancel_locks_for_export(exp);
 775
 776         /* complete all outstanding replies */
 777         spin_lock_irqsave(&exp->exp_lock, irqflags);
 778         while (!list_empty(&exp->exp_outstanding_replies)) {
 779                 struct ptlrpc_reply_state *rs =
 780                         list_entry(exp->exp_outstanding_replies.next,
 781                                    struct ptlrpc_reply_state, rs_exp_list);
 782                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 783
 784                 spin_lock(&svc->srv_lock);
 785                 list_del_init(&rs->rs_exp_list);
 786                 ptlrpc_schedule_difficult_reply(rs);
 787                 spin_unlock(&svc->srv_lock);
 788         }
 789         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 790         RETURN(rc);
 791 }
 792
 793 static int mds_getstatus(struct ptlrpc_request *req)
 794 {
 795         struct mds_obd *mds = mds_req2mds(req);
 796         struct mds_body *body;
 797         int rc, size;
 798         ENTRY;
 799
 800         size = sizeof(*body);
 801
 802         rc = lustre_pack_reply(req, 1, &size, NULL);
 803         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 804                 CERROR("mds: out of memory for message: size=%d\n", size);
 805                 req->rq_status = -ENOMEM;       /* superfluous? */
 806                 RETURN(-ENOMEM);
 807         }
 808
 809         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 810         body->valid |= OBD_MD_FID;
 811
 812         memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
 813
 814         /*
 815          * the last_committed and last_xid fields are filled in for all replies
 816          * already - no need to do so here also.
 817          */
 818         RETURN(0);
 819 }
 820
 821 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 822                      void *data, int flag)
 823 {
 824         int do_ast;
 825         ENTRY;
 826
 827         if (flag == LDLM_CB_CANCELING) {
 828                 /* Don't need to do anything here. */
 829                 RETURN(0);
 830         }
 831
 832         /* XXX layering violation!  -phil */
 833         lock_res_and_lock(lock);
 834
 835         /*
 836          * get this: if mds_blocking_ast is racing with mds_intent_policy, such
 837          * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
 838          * then by the time we get the lock, we might not be the correct
 839          * blocking function anymore.  So check, and return early, if so.
 840          */
 841         if (lock->l_blocking_ast != mds_blocking_ast) {
 842                 unlock_res_and_lock(lock);
 843                 RETURN(0);
 844         }
 845
 846         lock->l_flags |= LDLM_FL_CBPENDING;
 847         do_ast = (!lock->l_readers && !lock->l_writers);
 848         unlock_res_and_lock(lock);
 849
 850         if (do_ast) {
 851                 struct lustre_handle lockh;
 852                 int rc;
 853
 854                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 855                 ldlm_lock2handle(lock, &lockh);
 856                 rc = ldlm_cli_cancel(&lockh);
 857                 if (rc < 0)
 858                         CERROR("ldlm_cli_cancel: %d\n", rc);
 859         } else {
 860                 LDLM_DEBUG(lock, "Lock still has references, will be "
 861                            "cancelled later");
 862         }
 863         RETURN(0);
 864 }
 865
 866 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
 867                           void *md, int size, int mea)
 868 {
 869         int rc = size;
 870
 871         if (S_ISREG(inode->i_mode)) {
 872                 rc = mds_convert_lov_ea(obd, inode, md, size);
 873         } else if (S_ISDIR(inode->i_mode)) {
 874                 if (mea) {
 875                         rc = mds_convert_mea_ea(obd, inode, md, size);
 876                 } else {
 877                         rc = mds_convert_lov_ea(obd, inode, md, size);
 878                 }
 879                 if (rc == -EINVAL) {
 880                         CERROR("Invalid EA format (nor LOV or MEA) "
 881                                "is detected. Inode %lu/%u\n",
 882                                inode->i_ino, inode->i_generation);
 883                 }
 884         }
 885         return rc;
 886 }
 887
 888 int mds_get_md(struct obd_device *obd, struct inode *inode,
 889                void *md, int *size, int lock, int mea)
 890 {
 891         int lmm_size;
 892         int rc = 0;
 893         ENTRY;
 894
 895         if (lock)
 896                 down(&inode->i_sem);
 897
 898         rc = fsfilt_get_md(obd, inode, md, *size,
 899                            (mea ? EA_MEA : EA_LOV));
 900         if (rc < 0) {
 901                 CERROR("Error %d reading eadata for ino %lu\n",
 902                        rc, inode->i_ino);
 903         } else if (rc > 0) {
 904                 lmm_size = rc;
 905                 rc = mds_convert_md(obd, inode, md,
 906                                     lmm_size, mea);
 907                 if (rc == 0) {
 908                         *size = lmm_size;
 909                         rc = lmm_size;
 910                 } else if (rc > 0) {
 911                         *size = rc;
 912                 }
 913         }
 914         if (lock)
 915                 up(&inode->i_sem);
 916
 917         RETURN(rc);
 918 }
 919
 920 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 921  * Call with lock=0 if the caller has already taken the i_sem. */
 922 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 923                 struct mds_body *body, struct inode *inode, int lock, int mea)
 924 {
 925         struct mds_obd *mds = &obd->u.mds;
 926         int rc, lmm_size;
 927         void *lmm;
 928         ENTRY;
 929
 930         lmm = lustre_msg_buf(msg, offset, 0);
 931         if (lmm == NULL) {
 932                 /* Some problem with getting eadata when I sized the reply
 933                  * buffer... */
 934                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 935                        inode->i_ino);
 936                 RETURN(0);
 937         }
 938         lmm_size = msg->buflens[offset];
 939
 940         /* I don't really like this, but it is a sanity check on the client
 941          * MD request.  However, if the client doesn't know how much space
 942          * to reserve for the MD, it shouldn't be bad to have too much space.
 943          */
 944         if (lmm_size > mds->mds_max_mdsize) {
 945                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 946                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 947                 // RETURN(-EINVAL);
 948         }
 949
 950         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
 951         if (rc > 0) {
 952                 body->valid |= S_ISDIR(inode->i_mode) ?
 953                         OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
 954
 955                 if (mea)
 956                         body->valid |= OBD_MD_MEA;
 957
 958                 body->eadatasize = lmm_size;
 959                 rc = 0;
 960         }
 961
 962         RETURN(rc);
 963 }
 964
 965 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
 966                   struct mds_body *repbody, int reply_off)
 967 {
 968         struct inode *inode = dentry->d_inode;
 969         char *symname;
 970         int len, rc;
 971         ENTRY;
 972
 973         symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 974         LASSERT(symname != NULL);
 975         len = req->rq_repmsg->buflens[reply_off + 1];
 976
 977         rc = inode->i_op->readlink(dentry, symname, len);
 978         if (rc < 0) {
 979                 CERROR("readlink failed: %d\n", rc);
 980         } else if (rc != len - 1) {
 981                 CERROR ("Unexpected readlink rc %d: expecting %d\n",
 982                         rc, len - 1);
 983                 rc = -EINVAL;
 984         } else {
 985                 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 986                 repbody->valid |= OBD_MD_LINKNAME;
 987                 repbody->eadatasize = rc + 1;
 988                 symname[rc] = 0;        /* NULL terminate */
 989                 rc = 0;
 990         }
 991
 992         RETURN(rc);
 993 }
 994
 995 int mds_pack_xattr(struct dentry *dentry, struct ptlrpc_request *req,
 996                    struct mds_body *repbody, int req_off, int reply_off)
 997 {
 998         struct inode *inode = dentry->d_inode;
 999         char *ea_name;
1000         void *value = NULL;
1001         int len, rc;
1002         ENTRY;
1003
1004         ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
1005         len = req->rq_repmsg->buflens[reply_off + 1];
1006         if (len != 0)
1007                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1008
1009         rc = -EOPNOTSUPP;
1010
1011         if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1012                 struct rmtacl_upcall_desc desc;
1013
1014                 if (len != LUSTRE_ACL_SIZE_MAX || !value) {
1015                         CERROR("no reply buffer prepared\n");
1016                         RETURN(-EFAULT);
1017                 }
1018
1019                 memset(&desc, 0, sizeof(desc));
1020                 desc.get = 1;
1021                 desc.cmd = lustre_msg_string(req->rq_reqmsg, req_off + 2, 0);
1022                 desc.cmdlen =  req->rq_reqmsg->buflens[req_off + 2];
1023                 desc.res = (char *) value;
1024                 desc.reslen = LUSTRE_ACL_SIZE_MAX;
1025
1026                 mds_do_remote_acl_upcall(&desc);
1027
1028                 if (desc.upcall_status)
1029                         RETURN(desc.upcall_status);
1030
1031                 if (desc.reslen > LUSTRE_ACL_SIZE_MAX) {
1032                         CERROR("downcall claim reslen %u\n", desc.reslen);
1033                         RETURN(-EINVAL);
1034                 }
1035                 /* like remote setfacl, steal "flags" in mds_body as the
1036                  * exececution status
1037                  */
1038                 repbody->flags = desc.status;
1039                 repbody->valid |= OBD_MD_FLXATTR;
1040                 repbody->eadatasize = desc.reslen;
1041
1042                 RETURN(0);
1043         }
1044
1045         if (inode->i_op && inode->i_op->getxattr)
1046                 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
1047
1048         if (rc < 0) {
1049                 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1050                         CERROR("getxattr failed: %d", rc);
1051         } else {
1052                 repbody->valid |= OBD_MD_FLXATTR;
1053                 repbody->eadatasize = rc;
1054                 rc = 0;
1055         }
1056
1057         RETURN(rc);
1058 }
1059
1060 int mds_pack_xattr_list(struct dentry *dentry, struct ptlrpc_request *req,
1061                         struct mds_body *repbody, int reply_off)
1062 {
1063         struct inode *inode = dentry->d_inode;
1064         void *value = NULL;
1065         int len, rc;
1066         ENTRY;
1067
1068         len = req->rq_repmsg->buflens[reply_off + 1];
1069         if (len != 0)
1070                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1071
1072         rc = -EOPNOTSUPP;
1073         if (inode->i_op && inode->i_op->getxattr)
1074                 rc = inode->i_op->listxattr(dentry, value, len);
1075
1076         if (rc < 0) {
1077                 CERROR("listxattr failed: %d", rc);
1078         } else {
1079                 repbody->valid |= OBD_MD_FLXATTRLIST;
1080                 repbody->eadatasize = rc;
1081                 rc = 0;
1082         }
1083         RETURN(rc);
1084 }
1085
1086 static
1087 int mds_pack_posix_acl(struct lustre_msg *repmsg, int offset,
1088                        struct mds_body *body, struct inode *inode)
1089 {
1090         struct dentry de = { .d_inode = inode };
1091         __u32 buflen, *sizep;
1092         void *buf;
1093         int size;
1094         ENTRY;
1095
1096         if (!inode->i_op->getxattr)
1097                 RETURN(0);
1098
1099         sizep = lustre_msg_buf(repmsg, offset, 4);
1100         if (!sizep) {
1101                 CERROR("can't locate returned acl size buf\n");
1102                 RETURN(-EPROTO);
1103         }
1104
1105         buflen = repmsg->buflens[offset + 1];
1106         buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1107
1108         size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1109         if (size == -ENODATA || size == -EOPNOTSUPP)
1110                 RETURN(0);
1111         if (size < 0)
1112                 RETURN(size);
1113         LASSERT(size);
1114
1115         *sizep = cpu_to_le32(size);
1116         body->valid |= OBD_MD_FLACL;
1117
1118         RETURN(0);
1119 }
1120
1121 int mds_pack_remote_perm(struct ptlrpc_request *req, int reply_off,
1122                          struct mds_body *body, struct inode *inode)
1123 {
1124         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1125         struct lustre_sec_desc *lsd;
1126         struct mds_remote_perm *perm;
1127         __u32 lsd_perms;
1128
1129         LASSERT(inode->i_op);
1130         LASSERT(inode->i_op->permission);
1131         LASSERT(req->rq_export->exp_mds_data.med_remote);
1132
1133         perm = (struct mds_remote_perm *)
1134                        lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(perm));
1135         if (!perm) {
1136                 CERROR("no remote perm buf at offset %d\n", reply_off);
1137                 return -EINVAL;
1138         }
1139
1140         memset(perm, 0, sizeof(*perm));
1141
1142         /* obtain authenticated uid/gid and LSD permissions, which
1143          * might be different from current process context, from LSD
1144          */
1145         lsd = mds_get_lsd(current->uid);
1146         if (!lsd) {
1147                 CWARN("can't LSD of uid %u\n", current->uid);
1148                 RETURN(-EPERM);
1149         }
1150
1151         perm->mrp_auth_uid = lsd->lsd_uid;
1152         perm->mrp_auth_gid = lsd->lsd_gid;
1153
1154         lsd_perms = mds_lsd_get_perms(lsd, 1, 0, req->rq_peer.peer_id.nid);
1155         if (lsd_perms & LSD_PERM_SETUID)
1156                 perm->mrp_allow_setuid = 1;
1157         if (lsd_perms & LSD_PERM_SETGID)
1158                 perm->mrp_allow_setgid = 1;
1159
1160         mds_put_lsd(lsd);
1161
1162         if (mds_remote_perm_do_reverse_map(med, perm))
1163                 RETURN(-EPERM);
1164
1165         /* permission bits of current user
1166          * XXX this is low efficient, could we do it in one blow?
1167          */
1168         if (inode->i_op->permission(inode, MAY_EXEC, NULL) == 0)
1169                 perm->mrp_perm |= MAY_EXEC;
1170         if (inode->i_op->permission(inode, MAY_WRITE, NULL) == 0)
1171                 perm->mrp_perm |= MAY_WRITE;
1172         if (inode->i_op->permission(inode, MAY_READ, NULL) == 0)
1173                 perm->mrp_perm |= MAY_READ;
1174
1175         body->valid |= (OBD_MD_FLACL | OBD_MD_FLRMTACL);
1176
1177         RETURN(0);
1178 }
1179
1180 int mds_pack_acl(struct ptlrpc_request *req, int reply_off,
1181                  struct mds_body *body, struct inode *inode)
1182 {
1183         int rc;
1184
1185         if (!req->rq_export->exp_mds_data.med_remote)
1186                 rc = mds_pack_posix_acl(req->rq_repmsg, reply_off, body, inode);
1187         else
1188                 rc = mds_pack_remote_perm(req, reply_off + 1, body, inode);
1189
1190         return rc;
1191 }
1192
1193 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1194                                 struct ptlrpc_request *req, int req_off,
1195                                 struct mds_body *reqbody, int reply_off,
1196                                 struct mds_req_sec_desc *rsd)
1197 {
1198         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1199         struct inode *inode = dentry->d_inode;
1200         struct mds_body *body;
1201         int rc = 0, offset = 0;
1202         ENTRY;
1203
1204         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1205                 RETURN(-ENOENT);
1206
1207         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1208         LASSERT(body != NULL);                 /* caller prepped reply */
1209
1210         if (dentry->d_flags & DCACHE_CROSS_REF) {
1211                 mds_pack_dentry2body(obd, body, dentry,
1212                                      (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1213                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1214                        OLID4(&body->id1));
1215                 RETURN(0);
1216         }
1217
1218         mds_pack_inode2body(obd, body, inode,
1219                             (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1220
1221         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1222             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1223
1224                 /* guessing what kind og attribute do we need. */
1225                 int is_mea = (S_ISDIR(inode->i_mode) &&
1226                     (reqbody->valid & OBD_MD_MEA) != 0);
1227
1228                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1229                                  body, inode, 1, is_mea);
1230
1231                 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1232                 if (!(body->valid & OBD_MD_FLEASIZE) &&
1233                     !(body->valid & OBD_MD_FLDIREA))
1234                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1235                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1236         } else if (S_ISLNK(inode->i_mode) &&
1237                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1238                 rc = mds_pack_link(dentry, req, body, reply_off);
1239         } else if (reqbody->valid & OBD_MD_FLXATTR) {
1240                 rc = mds_pack_xattr(dentry, req, body, req_off, reply_off);
1241         } else if (reqbody->valid & OBD_MD_FLXATTRLIST) {
1242                 rc = mds_pack_xattr_list(dentry, req, body, reply_off);
1243         }
1244
1245         offset = reply_off + ((reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1);
1246         if (reqbody->valid & OBD_MD_FLACL) {
1247                 rc = mds_pack_acl(req, offset, body, inode);
1248                 offset += 2;
1249         }
1250
1251         if (reqbody->valid & OBD_MD_FLKEY) {
1252                 rc = mds_pack_gskey(obd, req->rq_repmsg, &offset,
1253                                     body, inode);
1254         }
1255
1256         mds_pack_audit(obd, inode, body);
1257
1258         if (reqbody->valid & OBD_MD_CAPA) {
1259                 struct lustre_capa *req_capa;
1260
1261                 LASSERT(!(reqbody->valid & ~OBD_MD_CAPA));
1262                 LASSERT(S_ISREG(inode->i_mode));
1263
1264                 req_capa = lustre_swab_reqbuf(req, req_off + 1,
1265                                               sizeof(*req_capa),
1266                                               lustre_swab_lustre_capa);
1267                 if (req_capa == NULL) {
1268                         CERROR("Can't unpack capa\n");
1269                         RETURN(-EFAULT);
1270                 }
1271
1272                 offset = reply_off + 1;
1273                 rc = mds_pack_capa(obd, med, reqbody, req_capa, req,
1274                                    &offset, body);
1275         }
1276
1277         if (rc == 0)
1278                 mds_body_do_reverse_map(med, body);
1279
1280         RETURN(rc);
1281 }
1282
1283 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1284                                    struct dentry *dentry,
1285                                    int offset)
1286 {
1287         int rc = 0, size[1] = {sizeof(struct mds_body)};
1288         ENTRY;
1289
1290         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1291                 CERROR("failed MDS_GETATTR_PACK test\n");
1292                 req->rq_status = -ENOMEM;
1293                 RETURN(-ENOMEM);
1294         }
1295
1296         rc = lustre_pack_reply(req, 1, size, NULL);
1297         if (rc) {
1298                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1299                 GOTO(out, req->rq_status = rc);
1300         }
1301
1302         EXIT;
1303 out:
1304         return rc;
1305 }
1306
1307 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1308                                 int offset)
1309 {
1310         struct inode *inode = de->d_inode;
1311         struct mds_obd *mds = mds_req2mds(req);
1312         struct mds_body *body;
1313         int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1314         ENTRY;
1315
1316         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1317         LASSERT(body != NULL);                 /* checked by caller */
1318         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
1319
1320         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1321             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1322                 int rc;
1323
1324                 down(&inode->i_sem);
1325                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1326                                    ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1327                 up(&inode->i_sem);
1328                 if (rc < 0) {
1329                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1330                                 CERROR("error getting inode %lu MD: rc = %d\n",
1331                                        inode->i_ino, rc);
1332                         size[bufcount] = 0;
1333                 } else if (rc > mds->mds_max_mdsize) {
1334                         size[bufcount] = 0;
1335                         CERROR("MD size %d larger than maximum possible %u\n",
1336                                rc, mds->mds_max_mdsize);
1337                 } else {
1338                         size[bufcount] = rc;
1339                 }
1340                 bufcount++;
1341         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1342                 if (inode->i_size + 1 != body->eadatasize)
1343                         CERROR("symlink size: %Lu, reply space: %d\n",
1344                                inode->i_size + 1, body->eadatasize);
1345                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1346                 bufcount++;
1347                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1348                        inode->i_size + 1, body->eadatasize);
1349         } else if ((body->valid & OBD_MD_FLXATTR)) {
1350                 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1351                                                   offset + 1, 0);
1352                 rc = -EOPNOTSUPP;
1353
1354                 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1355                         size[bufcount] = LUSTRE_ACL_SIZE_MAX;
1356                 } else {
1357                         if (inode->i_op && inode->i_op->getxattr)
1358                                 rc = inode->i_op->getxattr(de, ea_name,
1359                                                            NULL, 0);
1360
1361                         if (rc < 0) {
1362                                 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1363                                         CERROR("error get inode %lu EA: %d\n",
1364                                                inode->i_ino, rc);
1365                                 size[bufcount] = 0;
1366                         } else {
1367                                 size[bufcount] = min_t(int,
1368                                                        body->eadatasize, rc);
1369                         }
1370                 }
1371                 bufcount++;
1372         } else if (body->valid & OBD_MD_FLXATTRLIST) {
1373                 rc = -EOPNOTSUPP;
1374                 if (inode->i_op && inode->i_op->getxattr)
1375                         rc = inode->i_op->listxattr(de, NULL, 0);
1376
1377                 if (rc < 0) {
1378                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1379                                 CERROR("error getting inode %lu EA: rc = %d\n",
1380                                        inode->i_ino, rc);
1381                         size[bufcount] = 0;
1382                 } else {
1383                         size[bufcount] = min_t(int, body->eadatasize, rc);
1384                 }
1385                 bufcount++;
1386         }
1387
1388         /* may co-exist with OBD_MD_FLEASIZE */
1389         if (body->valid & OBD_MD_FLACL) {
1390                 if (req->rq_export->exp_mds_data.med_remote) {
1391                         size[bufcount++] = sizeof(int);
1392                         size[bufcount++] = sizeof(struct mds_remote_perm);
1393                 } else {
1394                         size[bufcount++] = sizeof(int);
1395                         size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1396                 }
1397         }
1398
1399         if (body->valid & OBD_MD_FLKEY) {
1400                 size[bufcount++] = sizeof(int);
1401                 size[bufcount++] = sizeof(struct crypto_key);
1402         }
1403
1404         if (body->valid & OBD_MD_CAPA)
1405                 size[bufcount++] = sizeof(struct lustre_capa);
1406
1407         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1408                 CERROR("failed MDS_GETATTR_PACK test\n");
1409                 req->rq_status = -ENOMEM;
1410                 GOTO(out, rc = -ENOMEM);
1411         }
1412
1413         rc = lustre_pack_reply(req, bufcount, size, NULL);
1414         if (rc) {
1415                 CERROR("out of memory\n");
1416                 GOTO(out, req->rq_status = rc);
1417         }
1418
1419         EXIT;
1420  out:
1421         return rc;
1422 }
1423
1424 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1425                       char *name, int namelen)
1426 {
1427         struct mea *mea = NULL;
1428         int mea_size, rc = 0;
1429         ENTRY;
1430
1431         rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1432         if (rc)
1433                 RETURN(rc);
1434         if (mea != NULL && mea->mea_count) {
1435                 /*
1436                  * dir is already splitted, check if requested filename should
1437                  * live at this MDS or at another one.
1438                  */
1439                 int i = mea_name2idx(mea, name, namelen - 1);
1440                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1441                         CDEBUG(D_OTHER,
1442                                "inapropriate MDS(%d) for %s. should be "
1443                                "%lu(%d)\n", mea->mea_master, name,
1444                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1445                         rc = -ERESTART;
1446                 }
1447         }
1448
1449         if (mea)
1450                 OBD_FREE(mea, mea_size);
1451         RETURN(rc);
1452 }
1453
1454 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1455                      struct ptlrpc_request *req, struct mds_body *body)
1456 {
1457         struct inode *inode = dentry->d_inode;
1458         ENTRY;
1459
1460         LASSERT(body != NULL);
1461
1462         if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1463                 RETURN(0);
1464
1465         /* XXX: quite a ugly hack, need to check old code
1466          * drop FLSIZE/FLBLOCKS prior any checking to */
1467         body->valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
1468
1469         if (obd->obd_recovering) {
1470                 CDEBUG(D_INODE, "size for "DLID4" is unknown yet (recovering)\n",
1471                        OLID4(&body->id1));
1472                 RETURN(0);
1473         }
1474
1475         if (atomic_read(&inode->i_writecount)) {
1476                 /* some one has opened the file for write.
1477                  * mds doesn't know actual size */
1478                 CDEBUG(D_INODE, "MDS doesn't know actual size for "DLID4"\n",
1479                        OLID4(&body->id1));
1480                 RETURN(0);
1481         }
1482         CDEBUG(D_INODE, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1483                body->size, body->blocks, OLID4(&body->id1));
1484         body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1485         RETURN(0);
1486 }
1487
1488 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1489                             struct lustre_handle *child_lockh, int child_part)
1490 {
1491         struct obd_device *obd = req->rq_export->exp_obd;
1492         struct mds_obd *mds = &obd->u.mds;
1493         struct ldlm_reply *rep = NULL;
1494         struct lvfs_run_ctxt saved;
1495         struct mds_req_sec_desc *rsd;
1496         struct mds_body *body;
1497         struct dentry *dparent = NULL, *dchild = NULL;
1498         struct lvfs_ucred uc = {NULL, NULL,};
1499         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1500         unsigned int namesize = 0;
1501         int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1502         char *name = NULL;
1503         ENTRY;
1504
1505         LASSERT(!strcmp(obd->obd_type->typ_name, OBD_MDS_DEVICENAME));
1506         MD_COUNTER_INCREMENT(obd, getattr_lock);
1507
1508         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1509         if (!rsd) {
1510                 CERROR("Can't unpack security desc\n");
1511                 RETURN(-EFAULT);
1512         }
1513
1514         /* swab now, before anyone looks inside the request. */
1515         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1516                                   lustre_swab_mds_body);
1517         if (body == NULL) {
1518                 CERROR("Can't swab mds_body\n");
1519                 GOTO(cleanup, rc = -EFAULT);
1520         }
1521
1522         LASSERT_REQSWAB(req, offset + 1);
1523         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1524         if (name == NULL) {
1525                 CERROR("Can't unpack name\n");
1526                 GOTO(cleanup, rc = -EFAULT);
1527         }
1528         namesize = req->rq_reqmsg->buflens[offset + 1];
1529
1530         /* namesize less than 2 means we have empty name, probably came from
1531            revalidate by cfid, so no point in having name to be set */
1532         if (namesize <= 1)
1533                 name = NULL;
1534
1535         LASSERT (offset == 1 || offset == 3);
1536         if (offset == 3) {
1537                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1538                 reply_offset = 1;
1539         } else {
1540                 reply_offset = 0;
1541         }
1542
1543         rc = mds_init_ucred(&uc, req, rsd);
1544         if (rc) {
1545                 if (child_lockh->cookie == 0)
1546                         mds_audit_auth(req, &uc, AUDIT_STAT, &body->id1,
1547                                        name, namesize - 1);
1548                 GOTO(cleanup, rc);
1549         }
1550
1551         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1552         cleanup_phase = 1; /* kernel context */
1553         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1554
1555         LASSERT(namesize > 0);
1556         if (child_lockh->cookie != 0) {
1557                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1558                 resent_req = 1;
1559         }
1560 #if HAVE_LOOKUP_RAW
1561         if (body->valid == OBD_MD_FLID) {
1562                 struct mds_body *mds_reply;
1563                 int size = sizeof(*mds_reply);
1564                 struct inode *dir;
1565                 ino_t inum;
1566
1567                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1568                 if (IS_ERR(dparent)) {
1569                         rc = PTR_ERR(dparent);
1570                         GOTO(cleanup, rc);
1571                 }
1572                 /*
1573                  * the user requested ONLY the inode number, so do a raw lookup.
1574                  */
1575                 rc = lustre_pack_reply(req, 1, &size, NULL);
1576                 if (rc) {
1577                         CERROR("out of memory\n");
1578                         l_dput(dparent);
1579                         GOTO(cleanup, rc);
1580                 }
1581                 dir  = dparent->d_inode;
1582                 LASSERT(dir->i_op->lookup_raw != NULL);
1583                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1584                 l_dput(dparent);
1585                 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1586                                            sizeof(*mds_reply));
1587
1588                 id_ino(&mds_reply->id1) = inum;
1589                 mds_reply->valid = OBD_MD_FLID;
1590                 GOTO(cleanup, rc);
1591         }
1592 #endif
1593         if (resent_req == 0) {
1594                 LASSERT(id_fid(&body->id1) != 0);
1595                 if (name) {
1596                         rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1597                                                          parent_lockh, &dparent,
1598                                                          LCK_PR,
1599                                                          MDS_INODELOCK_UPDATE,
1600                                                          &update_mode,
1601                                                          name, namesize,
1602                                                          child_lockh, &dchild,
1603                                                          LCK_PR, child_part);
1604                         if (rc)
1605                                 GOTO(cleanup, rc);
1606
1607                         cleanup_phase = 2; /* dchild, dparent, locks */
1608
1609                         /*
1610                          * let's make sure this name should leave on this mds
1611                          * node.
1612                          */
1613                         rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1614                         if (rc)
1615                                 GOTO(cleanup, rc);
1616                 } else {
1617                         /* we have no dentry here, drop LOOKUP bit */
1618                         /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1619                         child_part &= ~MDS_INODELOCK_LOOKUP;
1620                         CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1621                                obd->obd_name, OLID4(&body->id1));
1622
1623                         dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1624                                                       LCK_PR, parent_lockh,
1625                                                       &update_mode, NULL, 0,
1626                                                       MDS_INODELOCK_UPDATE);
1627                         if (IS_ERR(dchild)) {
1628                                 CERROR("can't find inode with id "DLID4", err = %d\n",
1629                                        OLID4(&body->id1), (int)PTR_ERR(dchild));
1630                                 rc = PTR_ERR(dchild);
1631                                 dchild = NULL;
1632                                 GOTO(cleanup, rc);
1633                         }
1634                         memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1635                 }
1636         } else {
1637                 struct ldlm_lock *granted_lock;
1638
1639                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1640                 granted_lock = ldlm_handle2lock(child_lockh);
1641
1642                 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1643                          id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1644                          child_lockh->cookie);
1645
1646                 if (name) {
1647                         /* usual named request */
1648                         dparent = mds_id2dentry(obd, &body->id1, NULL);
1649                         LASSERT(!IS_ERR(dparent));
1650                         dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1651                         LASSERT(!IS_ERR(dchild));
1652                 } else {
1653                         /* client wants to get attr. by id */
1654                         dchild = mds_id2dentry(obd, &body->id1, NULL);
1655                         LASSERT(!IS_ERR(dchild));
1656                 }
1657                 LDLM_LOCK_PUT(granted_lock);
1658         }
1659
1660         cleanup_phase = 2; /* dchild, dparent, locks */
1661
1662         if (!DENTRY_VALID(dchild)) {
1663                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1664                 /*
1665                  * in the intent case, the policy clears this error: the
1666                  * disposition is enough.
1667                  */
1668                 rc = -ENOENT;
1669                 GOTO(cleanup, rc);
1670         } else {
1671                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1672         }
1673
1674         if (req->rq_repmsg == NULL) {
1675                 if (dchild->d_flags & DCACHE_CROSS_REF)
1676                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1677                 else
1678                         rc = mds_getattr_pack_msg(req, dchild, offset);
1679                 if (rc != 0) {
1680                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1681                         GOTO (cleanup, rc);
1682                 }
1683         }
1684
1685         rc = mds_getattr_internal(obd, dchild, req, offset, body,
1686                                   reply_offset, rsd);
1687         if (rc)
1688                 GOTO(cleanup, rc); /* returns the lock to the client */
1689
1690         /* probably MDS knows actual size? */
1691         body = lustre_msg_buf(req->rq_repmsg, reply_offset, sizeof(*body));
1692         LASSERT(body != NULL);
1693         mds_getattr_size(obd, dchild, req, body);
1694
1695         GOTO(cleanup, rc);
1696
1697  cleanup:
1698         /* audit stuff for getattr */
1699         if (resent_req == 0 && (dparent || dchild)) {
1700                 struct inode * au_inode = NULL;
1701
1702                 if (dchild && dchild->d_inode) {
1703                         au_inode = dchild->d_inode;
1704                         mds_audit_stat(req, &body->id1, au_inode,
1705                                        NULL, 0, rc);
1706                 } else {
1707                         au_inode = dparent->d_inode;
1708                         mds_audit_stat(req, &body->id1, au_inode,
1709                                        name, namesize - 1, rc);
1710                 }
1711         }
1712         switch (cleanup_phase) {
1713         case 2:
1714                 if (resent_req == 0) {
1715                         if (rc && DENTRY_VALID(dchild))
1716                                 ldlm_lock_decref(child_lockh, LCK_PR);
1717                         if (name)
1718                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1719 #ifdef S_PDIROPS
1720                         if (parent_lockh[1].cookie != 0)
1721                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1722 #endif
1723                         if (dparent)
1724                                 l_dput(dparent);
1725                 }
1726                 l_dput(dchild);
1727         case 1:
1728                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1729         default:
1730                 mds_exit_ucred(&uc);
1731         }
1732         return rc;
1733 }
1734
1735 static int mds_getattr(struct ptlrpc_request *req, int offset)
1736 {
1737         struct obd_device *obd = req->rq_export->exp_obd;
1738         struct lvfs_run_ctxt saved;
1739         struct dentry *de;
1740         struct mds_req_sec_desc *rsd;
1741         struct mds_body *body;
1742         struct lvfs_ucred uc = {NULL, NULL,};
1743         int rc = 0;
1744         ENTRY;
1745
1746         MD_COUNTER_INCREMENT(obd, getattr);
1747
1748         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1749         if (!rsd) {
1750                 CERROR("Can't unpack security desc\n");
1751                 RETURN(-EFAULT);
1752         }
1753
1754         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1755                                   lustre_swab_mds_body);
1756         if (body == NULL) {
1757                 CERROR ("Can't unpack body\n");
1758                 RETURN (-EFAULT);
1759         }
1760
1761         rc = mds_init_ucred(&uc, req, rsd);
1762         if (rc) {
1763                 mds_exit_ucred(&uc);
1764                 RETURN(rc);
1765         }
1766
1767         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1768         de = mds_id2dentry(obd, &body->id1, NULL);
1769         if (IS_ERR(de)) {
1770                 rc = req->rq_status = PTR_ERR(de);
1771                 GOTO(out_pop, rc);
1772         }
1773
1774         rc = mds_getattr_pack_msg(req, de, offset);
1775         if (rc != 0) {
1776                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1777                 GOTO(out_pop, rc);
1778         }
1779
1780         req->rq_status = mds_getattr_internal(obd, de, req, offset, body,
1781                                               0, rsd);
1782         l_dput(de);
1783
1784         EXIT;
1785 out_pop:
1786         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1787         mds_exit_ucred(&uc);
1788         return rc;
1789 }
1790
1791 static int mds_access_check(struct ptlrpc_request *req, int offset)
1792 {
1793         struct obd_device *obd = req->rq_export->exp_obd;
1794         struct lvfs_run_ctxt saved;
1795         struct dentry *de;
1796         struct mds_req_sec_desc *rsd;
1797         struct mds_body *body;
1798         struct lvfs_ucred uc;
1799         int rep_size[2] = {sizeof(*body),
1800                            sizeof(struct mds_remote_perm)};
1801         int rc = 0;
1802         ENTRY;
1803
1804         if (!req->rq_export->exp_mds_data.med_remote) {
1805                 CERROR("from local client "LPU64"\n", req->rq_peer.peer_id.nid);
1806                 RETURN(-EINVAL);
1807         }
1808
1809         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1810         if (!rsd) {
1811                 CERROR("Can't unpack security desc\n");
1812                 RETURN(-EFAULT);
1813         }
1814
1815         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1816                                   lustre_swab_mds_body);
1817         if (body == NULL) {
1818                 CERROR ("Can't unpack body\n");
1819                 RETURN (-EFAULT);
1820         }
1821
1822         MD_COUNTER_INCREMENT(obd, access_check);
1823
1824         rc = mds_init_ucred(&uc, req, rsd);
1825         if (rc) {
1826                 CERROR("init ucred error: %d\n", rc);
1827                 RETURN(rc);
1828         }
1829         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1830
1831         de = mds_id2dentry(obd, &body->id1, NULL);
1832         if (IS_ERR(de)) {
1833                 CERROR("grab ino "LPU64": err %ld\n",
1834                        body->id1.li_stc.u.e3s.l3s_ino, PTR_ERR(de));
1835                 GOTO(out_pop, rc = PTR_ERR(de));
1836         }
1837
1838         rc = lustre_pack_reply(req, 2, rep_size, NULL);
1839         if (rc) {
1840                 CERROR("pack reply error: %d\n", rc);
1841                 GOTO(out_dput, rc = -EINVAL);
1842         }
1843
1844         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1845         LASSERT(body);
1846
1847         rc = mds_pack_remote_perm(req, 1, body, de->d_inode);
1848
1849         EXIT;
1850
1851 out_dput:
1852         l_dput(de);
1853 out_pop:
1854         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1855         mds_exit_ucred(&uc);
1856         return rc;
1857 }
1858
1859 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1860                           unsigned long max_age)
1861 {
1862         int rc;
1863         ENTRY;
1864
1865         spin_lock(&obd->obd_osfs_lock);
1866         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1867         if (rc == 0)
1868                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1869         spin_unlock(&obd->obd_osfs_lock);
1870
1871         RETURN(rc);
1872 }
1873
1874 static int mds_statfs(struct ptlrpc_request *req)
1875 {
1876         struct obd_device *obd = req->rq_export->exp_obd;
1877         int rc, size = sizeof(struct obd_statfs);
1878         ENTRY;
1879
1880         /* This will trigger a watchdog timeout */
1881         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1882                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1883
1884         rc = lustre_pack_reply(req, 1, &size, NULL);
1885         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1886                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1887                 GOTO(out, rc);
1888         }
1889
1890         OBD_COUNTER_INCREMENT(obd, statfs);
1891
1892         /* We call this so that we can cache a bit - 1 jiffie worth */
1893         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1894                             jiffies - HZ);
1895         if (rc) {
1896                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1897                 GOTO(out, rc);
1898         }
1899
1900         EXIT;
1901 out:
1902         req->rq_status = rc;
1903         return rc;
1904 }
1905
1906 static int mds_sync(struct ptlrpc_request *req, int offset)
1907 {
1908         struct obd_device *obd = req->rq_export->exp_obd;
1909         struct mds_obd *mds = &obd->u.mds;
1910         struct mds_body *body;
1911         int rc, size = sizeof(*body);
1912         ENTRY;
1913
1914         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1915                                   lustre_swab_mds_body);
1916         if (body == NULL)
1917                 GOTO(out, rc = -EPROTO);
1918
1919         rc = lustre_pack_reply(req, 1, &size, NULL);
1920         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1921                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1922                 GOTO(out, rc);
1923         }
1924
1925         if (id_ino(&body->id1) == 0) {
1926                 /* an id of zero is taken to mean "sync whole filesystem" */
1927                 rc = fsfilt_sync(obd, mds->mds_sb);
1928                 if (rc)
1929                         GOTO(out, rc);
1930         } else {
1931                 /* just any file to grab fsync method - "file" arg unused */
1932                 struct file *file = mds->mds_rcvd_filp;
1933                 struct mds_body *rep_body;
1934                 struct dentry *de;
1935
1936                 de = mds_id2dentry(obd, &body->id1, NULL);
1937                 if (IS_ERR(de))
1938                         GOTO(out, rc = PTR_ERR(de));
1939
1940                 rc = file->f_op->fsync(NULL, de, 1);
1941                 if (rc)
1942                         GOTO(out, rc);
1943
1944                 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1945                 mds_pack_inode2body(obd, rep_body, de->d_inode,
1946                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1947                 l_dput(de);
1948         }
1949
1950         EXIT;
1951 out:
1952         req->rq_status = rc;
1953         return rc;
1954 }
1955
1956 /* mds_readpage does not take a DLM lock on the inode, because the client must
1957  * already have a PR lock.
1958  *
1959  * If we were to take another one here, a deadlock will result, if another
1960  * thread is already waiting for a PW lock. */
1961 static int mds_readpage(struct ptlrpc_request *req, int offset)
1962 {
1963         struct obd_device *obd = req->rq_export->exp_obd;
1964         struct vfsmount *mnt;
1965         struct dentry *de;
1966         struct file *file;
1967         struct mds_req_sec_desc *rsd;
1968         struct mds_body *body, *repbody;
1969         struct lvfs_run_ctxt saved;
1970         int rc, size = sizeof(*repbody);
1971         struct lvfs_ucred uc = {NULL, NULL,};
1972         ENTRY;
1973
1974         rc = lustre_pack_reply(req, 1, &size, NULL);
1975         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1976                 CERROR("mds: out of memory\n");
1977                 GOTO(out, rc = -ENOMEM);
1978         }
1979
1980         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1981         if (!rsd) {
1982                 CERROR("Can't unpack security desc\n");
1983                 GOTO (out, rc = -EFAULT);
1984         }
1985
1986         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1987                                   lustre_swab_mds_body);
1988         if (body == NULL) {
1989                 CERROR("Can't unpack body\n");
1990                 GOTO (out, rc = -EFAULT);
1991         }
1992
1993         rc = mds_init_ucred(&uc, req, rsd);
1994         if (rc) {
1995                 mds_audit_auth(req, &uc, AUDIT_READDIR, &body->id1,
1996                                NULL, 0);
1997                 GOTO(out, rc);
1998         }
1999
2000         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2001         de = mds_id2dentry(obd, &body->id1, &mnt);
2002         if (IS_ERR(de))
2003                 GOTO(out_pop, rc = PTR_ERR(de));
2004
2005         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
2006
2007         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
2008         /* note: in case of an error, dentry_open puts dentry */
2009         if (IS_ERR(file))
2010                 GOTO(out_pop, rc = PTR_ERR(file));
2011
2012         /* body->size is actually the offset -eeb */
2013         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
2014                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
2015                        body->size, de->d_inode->i_blksize);
2016                 GOTO(out_file, rc = -EFAULT);
2017         }
2018
2019         /* body->nlink is actually the #bytes to read -eeb */
2020         if (body->nlink & (de->d_inode->i_blksize - 1)) {
2021                 CERROR("size %u is not multiple of blocksize %lu\n",
2022                        body->nlink, de->d_inode->i_blksize);
2023                 GOTO(out_file, rc = -EFAULT);
2024         }
2025
2026         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
2027         repbody->size = file->f_dentry->d_inode->i_size;
2028         repbody->valid = OBD_MD_FLSIZE;
2029
2030         /* to make this asynchronous make sure that the handling function
2031            doesn't send a reply when this function completes. Instead a
2032            callback function would send the reply */
2033         /* body->size is actually the offset -eeb */
2034         rc = mds_sendpage(req, file, body->size, body->nlink);
2035
2036         EXIT;
2037 out_file:
2038         filp_close(file, 0);
2039 out_pop:
2040         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2041 out:
2042         mds_exit_ucred(&uc);
2043         req->rq_status = rc;
2044         return 0;
2045 }
2046
2047 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
2048                 char **data, int *datalen)
2049 {
2050         struct dentry *dentry;
2051         struct mds_obd *mds = &obd->u.mds;
2052         int rc = 0, mea = 0;
2053         char *ea;
2054         ENTRY;
2055
2056         LASSERT(id);
2057         LASSERT(obd);
2058
2059         dentry = mds_id2dentry(obd, id, NULL);
2060         if (IS_ERR(dentry))
2061                 GOTO(out, rc = PTR_ERR(dentry));
2062
2063         if (!dentry->d_inode) {
2064                 CERROR("Can't find object "DLID4".\n",
2065                        OLID4(id));
2066                 GOTO(out_dentry, rc = -EINVAL);
2067         }
2068         if (S_ISDIR(dentry->d_inode->i_mode)) {
2069                 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
2070                 mea = 1;
2071         } else {
2072                 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
2073                 mea = 0;
2074         }
2075         OBD_ALLOC(ea, *datalen);
2076         if (!ea) {
2077                 *datalen = 0;
2078                 GOTO(out_dentry, rc = PTR_ERR(dentry));
2079         }
2080         *data = ea;
2081         down(&dentry->d_inode->i_sem);
2082         rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
2083                            (mea ? EA_MEA : EA_LOV));
2084         up(&dentry->d_inode->i_sem);
2085
2086         if (rc < 0)
2087                 CERROR("Error %d reading eadata for ino %lu\n",
2088                         rc, dentry->d_inode->i_ino);
2089 out_dentry:
2090         l_dput(dentry);
2091 out:
2092         RETURN(rc);
2093 }
2094 EXPORT_SYMBOL(mds_read_md);
2095
2096 int mds_reint(struct ptlrpc_request *req, int offset,
2097               struct lustre_handle *lockh)
2098 {
2099         struct mds_update_record *rec;
2100         struct mds_req_sec_desc *rsd;
2101         int rc;
2102         ENTRY;
2103
2104         OBD_ALLOC(rec, sizeof(*rec));
2105         if (rec == NULL)
2106                 RETURN(-ENOMEM);
2107
2108         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
2109         if (!rsd) {
2110                 CERROR("Can't unpack security desc\n");
2111                 GOTO(out, rc = -EFAULT);
2112         }
2113
2114         rc = mds_update_unpack(req, offset, rec);
2115         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
2116                 CERROR("invalid record\n");
2117                 GOTO(out, req->rq_status = -EINVAL);
2118         }
2119
2120         rc = mds_init_ucred(&rec->ur_uc, req, rsd);
2121         if (rc) {
2122                 mds_audit_reint(req, rec);
2123                 GOTO(out, rc);
2124         }
2125
2126         /* rc will be used to interrupt a for loop over multiple records */
2127         rc = mds_reint_rec(rec, offset, req, lockh);
2128
2129 out:
2130         mds_exit_ucred(&rec->ur_uc);
2131         OBD_FREE(rec, sizeof(*rec));
2132         RETURN(rc);
2133 }
2134
2135 static int mds_filter_recovery_request(struct ptlrpc_request *req,
2136                                        struct obd_device *obd, int *process)
2137 {
2138         switch (req->rq_reqmsg->opc) {
2139         case MDS_CONNECT: /* This will never get here, but for completeness. */
2140         case OST_CONNECT: /* This will never get here, but for completeness. */
2141         case MDS_DISCONNECT:
2142         case OST_DISCONNECT:
2143                *process = 1;
2144                RETURN(0);
2145
2146         case MDS_CLOSE:
2147         case MDS_SYNC: /* used in unmounting */
2148         case OBD_PING:
2149         case MDS_REINT:
2150         case LDLM_ENQUEUE:
2151         case OST_CREATE:
2152                 *process = target_queue_recovery_request(req, obd);
2153                 RETURN(0);
2154
2155         default:
2156                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2157                 *process = 0;
2158                 /* XXX what should we set rq_status to here? */
2159                 req->rq_status = -EAGAIN;
2160                 RETURN(ptlrpc_error(req));
2161         }
2162 }
2163
2164 static char *reint_names[] = {
2165         [REINT_SETATTR] "setattr",
2166         [REINT_CREATE]  "create",
2167         [REINT_LINK]    "link",
2168         [REINT_UNLINK]  "unlink",
2169         [REINT_RENAME]  "rename",
2170         [REINT_OPEN]    "open",
2171 };
2172
2173 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
2174                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2175                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2176                             OBD_MD_FLID)
2177
2178 static void reconstruct_create(struct ptlrpc_request *req)
2179 {
2180         struct mds_export_data *med = &req->rq_export->exp_mds_data;
2181         struct mds_client_data *mcd = med->med_mcd;
2182         struct dentry *dentry;
2183         struct ost_body *body;
2184         struct lustre_id id;
2185         int rc;
2186         ENTRY;
2187
2188         /* copy rc, transno and disp; steal locks */
2189         mds_req_from_mcd(req, mcd);
2190         if (req->rq_status) {
2191                 EXIT;
2192                 return;
2193         }
2194
2195         id_gen(&id) = 0;
2196         id_group(&id) = 0;
2197
2198         id_ino(&id) = mcd->mcd_last_data;
2199         LASSERT(id_ino(&id) != 0);
2200
2201         dentry = mds_id2dentry(req2obd(req), &id, NULL);
2202         if (IS_ERR(dentry)) {
2203                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2204                 req->rq_status = PTR_ERR(dentry);
2205                 EXIT;
2206                 return;
2207         }
2208
2209         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2210               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2211               dentry->d_inode->i_generation);
2212
2213         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2214         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2215         body->oa.o_id = dentry->d_inode->i_ino;
2216         body->oa.o_generation = dentry->d_inode->i_generation;
2217         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2218
2219         down(&dentry->d_inode->i_sem);
2220         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2221         up(&dentry->d_inode->i_sem);
2222         if (rc) {
2223                 CERROR("Can't read inode self id, inode %lu, "
2224                        "rc %d\n", dentry->d_inode->i_ino, rc);
2225                 id_fid(&id) = 0;
2226         }
2227
2228         body->oa.o_fid = id_fid(&id);
2229         body->oa.o_mds = id_group(&id);
2230         l_dput(dentry);
2231
2232         EXIT;
2233 }
2234
2235 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2236                               struct dentry *de, void *xattr, int xattr_size)
2237 {
2238         struct inode *inode = de->d_inode;
2239         struct posix_acl *acl;
2240         mode_t mode;
2241         int rc = 0;
2242
2243         LASSERT(handle);
2244         LASSERT(inode);
2245         LASSERT(xattr);
2246         LASSERT(xattr_size > 0);
2247
2248         if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2249                 CERROR("backend fs dosen't support xattr\n");
2250                 return -EOPNOTSUPP;
2251         }
2252
2253         /* set default acl */
2254         if (S_ISDIR(inode->i_mode)) {
2255                 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2256                                            xattr, xattr_size, 0);
2257                 if (rc) {
2258                         CERROR("set default acl err: %d\n", rc);
2259                         return rc;
2260                 }
2261         }
2262
2263         /* set access acl */
2264         acl = posix_acl_from_xattr(xattr, xattr_size);
2265         if (acl == NULL || IS_ERR(acl)) {
2266                 CERROR("insane attr data\n");
2267                 return PTR_ERR(acl);
2268         }
2269
2270         if (posix_acl_valid(acl)) {
2271                 CERROR("default acl not valid: %d\n", rc);
2272                 rc = -EFAULT;
2273                 goto out;
2274         }
2275
2276         mode = inode->i_mode;
2277         rc = posix_acl_create_masq(acl, &mode);
2278         if (rc < 0) {
2279                 CERROR("create masq err %d\n", rc);
2280                 goto out;
2281         }
2282
2283         if (inode->i_mode != mode) {
2284                 struct iattr iattr = { .ia_valid = ATTR_MODE,
2285                                        .ia_mode = mode };
2286                 int rc2;
2287
2288                 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2289                 if (rc2) {
2290                         CERROR("setattr mode err: %d\n", rc2);
2291                         rc = rc2;
2292                         goto out;
2293                 }
2294         }
2295
2296         if (rc > 0) {
2297                 /* we didn't change acl except mode bits of some
2298                  * entries, so should be fit into original size.
2299                  */
2300                 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2301                 LASSERT(rc > 0);
2302
2303                 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2304                                            xattr, xattr_size, 0);
2305                 if (rc)
2306                         CERROR("set access acl err: %d\n", rc);
2307         }
2308 out:
2309         posix_acl_release(acl);
2310         return rc;
2311 }
2312
2313 static int mdt_obj_create(struct ptlrpc_request *req)
2314 {
2315         struct obd_device *obd = req->rq_export->exp_obd;
2316         struct mds_obd *mds = &obd->u.mds;
2317         struct ost_body *body, *repbody;
2318         void *acl = NULL;
2319         int acl_size;
2320         char idname[LL_ID_NAMELEN];
2321         int size = sizeof(*repbody);
2322         struct inode *parent_inode;
2323         struct lvfs_run_ctxt saved;
2324         int rc, cleanup_phase = 0;
2325         struct dentry *new = NULL;
2326         struct dentry_params dp;
2327         int mealen, flags = 0;
2328         struct lvfs_ucred uc;
2329         struct lustre_id id;
2330         struct mea *mea;
2331         void *handle = NULL;
2332         unsigned long cr_inum = 0;
2333         __u64 fid = 0;
2334         ENTRY;
2335
2336         DEBUG_REQ(D_HA, req, "create remote object");
2337         parent_inode = mds->mds_unnamed_dir->d_inode;
2338
2339         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2340                                   lustre_swab_ost_body);
2341         if (body == NULL)
2342                 RETURN(-EFAULT);
2343
2344         /* acl data is packed transparently, no swab here */
2345         LASSERT(req->rq_reqmsg->bufcount >= 2);
2346         acl_size = req->rq_reqmsg->buflens[1];
2347         if (acl_size) {
2348                 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2349                 if (!acl) {
2350                         CERROR("No default acl buf?\n");
2351                         RETURN(-EFAULT);
2352                 }
2353         }
2354
2355         rc = lustre_pack_reply(req, 1, &size, NULL);
2356         if (rc)
2357                 RETURN(rc);
2358
2359         MDS_CHECK_RESENT(req, reconstruct_create(req));
2360
2361         uc.luc_lsd = NULL;
2362         uc.luc_ginfo = NULL;
2363         uc.luc_uid = body->oa.o_uid;
2364         uc.luc_gid = body->oa.o_gid;
2365         uc.luc_fsuid = body->oa.o_uid;
2366         uc.luc_fsgid = body->oa.o_gid;
2367
2368         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2369         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2370
2371         /* in REPLAY case inum should be given (client or other MDS fills it) */
2372         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2373             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2374                 /*
2375                  * this is re-create request from MDS holding directory name.
2376                  * we have to lookup given ino/gen first. if it exists (good
2377                  * case) then there is nothing to do. if it does not then we
2378                  * have to recreate it.
2379                  */
2380                 id_ino(&id) = body->oa.o_id;
2381                 id_gen(&id) = body->oa.o_generation;
2382
2383                 new = mds_id2dentry(obd, &id, NULL);
2384                 if (!IS_ERR(new) && new->d_inode) {
2385                         struct lustre_id sid;
2386
2387                         CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2388                                (unsigned long)id_ino(&id),
2389                                (unsigned long)id_gen(&id));
2390
2391                         obdo_from_inode(&repbody->oa, new->d_inode,
2392                                         FILTER_VALID_FLAGS);
2393
2394                         repbody->oa.o_id = new->d_inode->i_ino;
2395                         repbody->oa.o_generation = new->d_inode->i_generation;
2396                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2397                         cleanup_phase = 1;
2398
2399                         down(&new->d_inode->i_sem);
2400                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2401                         up(&new->d_inode->i_sem);
2402                         if (rc) {
2403                                 CERROR("Can't read inode self id "
2404                                        "inode %lu, rc %d.\n",
2405                                        new->d_inode->i_ino, rc);
2406                                 GOTO(cleanup, rc);
2407                         }
2408
2409                         repbody->oa.o_fid = id_fid(&sid);
2410                         repbody->oa.o_mds = id_group(&sid);
2411                         LASSERT(id_fid(&sid) != 0);
2412
2413                         /*
2414                          * here we could use fid passed in body->oa.o_fid and
2415                          * thus avoid mds_read_inode_sid().
2416                          */
2417                         cr_inum = new->d_inode->i_ino;
2418                         GOTO(cleanup, rc = 0);
2419                 }
2420         }
2421
2422         down(&parent_inode->i_sem);
2423         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2424         if (IS_ERR(handle)) {
2425                 up(&parent_inode->i_sem);
2426                 CERROR("fsfilt_start() failed, rc = %d\n",
2427                        (int)PTR_ERR(handle));
2428                 GOTO(cleanup, rc = PTR_ERR(handle));
2429         }
2430         cleanup_phase = 1; /* transaction */
2431
2432 repeat:
2433         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2434         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2435         if (IS_ERR(new)) {
2436                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2437                        obd->obd_name, idname, (int) PTR_ERR(new));
2438                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2439                 up(&parent_inode->i_sem);
2440                 RETURN(PTR_ERR(new));
2441         } else if (new->d_inode) {
2442                 CERROR("%s: name exists. repeat\n", obd->obd_name);
2443                 goto repeat;
2444         }
2445         if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2446              lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2447                 fid = body->oa.o_fid;
2448         } else {
2449                 fid = mds_alloc_fid(obd);
2450         }
2451         new->d_fsdata = (void *)&dp;
2452         dp.p_inum = 0;
2453         dp.p_ptr = req;
2454
2455         dp.p_fid = fid;
2456         dp.p_group = mds->mds_num;
2457
2458         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2459             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2460                 LASSERT(body->oa.o_id != 0);
2461                 dp.p_inum = body->oa.o_id;
2462                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2463                           (unsigned long)body->oa.o_id,
2464                           (unsigned long)body->oa.o_generation);
2465         }
2466
2467         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2468         if (rc == 0) {
2469                 if (acl) {
2470                         rc = mds_inode_init_acl(obd, handle, new,
2471                                                 acl, acl_size);
2472                         if (rc) {
2473                                 up(&parent_inode->i_sem);
2474                                 GOTO(cleanup, rc);
2475                         }
2476                 }
2477                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2478                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2479                         new->d_inode->i_generation = body->oa.o_generation;
2480                         mark_inode_dirty(new->d_inode);
2481
2482                         /*
2483                          * avoiding asserts in cache flush case, as
2484                          * @body->oa.o_id should be zero.
2485                          */
2486                         if (body->oa.o_id) {
2487                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2488                                          "BUG 3550: failed to recreate obj "
2489                                          LPU64" -> %lu\n", body->oa.o_id,
2490                                          new->d_inode->i_ino);
2491
2492                                 LASSERTF(body->oa.o_generation ==
2493                                          new->d_inode->i_generation,
2494                                          "BUG 3550: failed to recreate obj/gen "
2495                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2496                                          body->oa.o_generation,
2497                                          new->d_inode->i_ino,
2498                                          new->d_inode->i_generation);
2499                         }
2500                 }
2501
2502                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2503                 repbody->oa.o_id = new->d_inode->i_ino;
2504                 repbody->oa.o_generation = new->d_inode->i_generation;
2505                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2506
2507                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2508                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2509                         LASSERT(body->oa.o_id != 0);
2510                         LASSERT(body->oa.o_fid != 0);
2511                 }
2512
2513                 mds_inode2id(obd, &id, new->d_inode, fid);
2514                 mds_update_inode_ids(obd, new->d_inode, handle, &id,
2515                                      body->oa.o_valid & OBD_MD_FLID ?
2516                                      NULL : obdo_id(&body->oa));
2517
2518                 /* initializing o_fid after it is allocated. */
2519                 repbody->oa.o_fid = id_fid(&id);
2520                 repbody->oa.o_mds = id_group(&id);
2521
2522                 rc = fsfilt_del_dir_entry(obd, new);
2523                 up(&parent_inode->i_sem);
2524                 if (rc) {
2525                         CERROR("can't remove name for object: %d\n", rc);
2526                         GOTO(cleanup, rc);
2527                 }
2528
2529                 cleanup_phase = 2; /* created directory object */
2530
2531                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2532                        (unsigned long)new->d_inode->i_ino,
2533                        (unsigned long)new->d_inode->i_generation,
2534                        (unsigned)new->d_inode->i_mode);
2535                 cr_inum = new->d_inode->i_ino;
2536         } else {
2537                 up(&parent_inode->i_sem);
2538                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2539                 GOTO(cleanup, rc);
2540         }
2541
2542         if (body->oa.o_valid & OBD_MD_FLID) {
2543                 /* this is new object for splitted dir. We have to prevent
2544                  * recursive splitting on it -bzzz */
2545                 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2546
2547                 OBD_ALLOC(mea, mealen);
2548                 if (mea == NULL)
2549                         GOTO(cleanup, rc = -ENOMEM);
2550
2551                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2552                 mea->mea_master = body->oa.o_mds;      /* master mds num */
2553                 mea->mea_count = 0;
2554
2555                 obdo2id(&mea->mea_ids[body->oa.o_mds], &body->oa);
2556
2557                 down(&new->d_inode->i_sem);
2558                 rc = fsfilt_set_md(obd, new->d_inode, handle,
2559                                    mea, mealen, EA_MEA);
2560                 up(&new->d_inode->i_sem);
2561                 if (rc)
2562                         CERROR("fsfilt_set_md() failed, "
2563                                "rc = %d\n", rc);
2564
2565                 OBD_FREE(mea, mealen);
2566
2567                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2568                        obd->obd_name, new->d_inode->i_ino,
2569                        new->d_inode->i_generation, flags);
2570         } else if (body->oa.o_easize) {
2571                 /* we pass LCK_EX to split routine to signal that we have
2572                  * exclusive access to the directory. simple because nobody
2573                  * knows it already exists -bzzz */
2574                 rc = mds_try_to_split_dir(obd, new, NULL,
2575                                           body->oa.o_easize, LCK_EX);
2576                 if (rc < 0) {
2577                         CERROR("Can't split directory %lu, error = %d.\n",
2578                                new->d_inode->i_ino, rc);
2579                 } else {
2580                         rc = 0;
2581                 }
2582         }
2583
2584         EXIT;
2585 cleanup:
2586         if (rc == 0 && (body->oa.o_flags & OBD_FL_REINT)) {
2587                 rc = mds_fidmap_add(obd, &id);
2588                 if (rc < 0) {
2589                         CERROR("can't create fid->ino mapping, "
2590                                "err %d\n", rc);
2591                 } else {
2592                         rc = 0;
2593                 }
2594         }
2595
2596         switch (cleanup_phase) {
2597         case 2: /* object has been created, but we'll may want to replay it later */
2598                 if (rc == 0)
2599                         ptlrpc_require_repack(req);
2600         case 1: /* transaction */
2601                 rc = mds_finish_transno(mds, parent_inode, handle,
2602                                         req, rc, cr_inum);
2603         }
2604
2605         l_dput(new);
2606         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2607         return rc;
2608 }
2609
2610 static int mdt_get_info(struct ptlrpc_request *req)
2611 {
2612         struct obd_export *exp = req->rq_export;
2613         int keylen, rc = 0;
2614         char *key;
2615         ENTRY;
2616
2617         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2618         if (key == NULL) {
2619                 DEBUG_REQ(D_HA, req, "no get_info key");
2620                 RETURN(-EFAULT);
2621         }
2622         keylen = req->rq_reqmsg->buflens[0];
2623
2624         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2625             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2626             (keylen < strlen("lovdesc") || strcmp(key, "lovdesc") != 0) &&
2627             (keylen < strlen("getext") || strcmp(key, "getext") != 0) &&
2628             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2629                 RETURN(-EPROTO);
2630
2631         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2632                 struct lustre_id *reply;
2633                 int size = sizeof(*reply);
2634
2635                 rc = lustre_pack_reply(req, 1, &size, NULL);
2636                 if (rc)
2637                         RETURN(rc);
2638
2639                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2640                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2641         } else if (keylen >= strlen("lovdesc") && !strcmp(key, "lovdesc")) {
2642                 struct lov_desc *reply;
2643                 int size = sizeof(*reply);
2644
2645                 rc = lustre_pack_reply(req, 1, &size, NULL);
2646                 if (rc)
2647                         RETURN(rc);
2648
2649                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2650                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2651         } else if (keylen >= strlen("getext") && !strcmp(key, "getext")) {
2652                 struct fid_extent *reply;
2653                 int size = sizeof(*reply);
2654
2655                 rc = lustre_pack_reply(req, 1, &size, NULL);
2656                 if (rc)
2657                         RETURN(rc);
2658
2659                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2660                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2661         } else {
2662                 obd_id *reply;
2663                 int size = sizeof(*reply);
2664
2665                 rc = lustre_pack_reply(req, 1, &size, NULL);
2666                 if (rc)
2667                         RETURN(rc);
2668
2669                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2670                 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2671         }
2672
2673         req->rq_repmsg->status = 0;
2674         RETURN(rc);
2675 }
2676
2677 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2678                         void *key, __u32 vallen, void *val)
2679 {
2680         struct obd_device *obd;
2681         struct mds_obd *mds;
2682         int rc = 0;
2683         ENTRY;
2684
2685         obd = class_exp2obd(exp);
2686         if (obd == NULL) {
2687                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2688                        exp->exp_handle.h_cookie);
2689                 RETURN(-EINVAL);
2690         }
2691
2692         mds = &obd->u.mds;
2693         if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2694                 rc = mds_set_audit(obd, val);
2695                 RETURN(rc);
2696         } else if (keylen >= strlen("ids") && memcmp(key, "ids", keylen) == 0) {
2697                 struct lustre_id *ids = (struct lustre_id *)val;
2698                 struct dentry *de;
2699                 struct inode *inode;
2700                 void *handle;
2701                 int err;
2702
2703                 de = mds_id2dentry(obd, ids, NULL);
2704                 if (IS_ERR(de)) {
2705                         rc = PTR_ERR(de);
2706                         CERROR("lookup by an id error rc=%d\n ", rc);
2707                         RETURN(rc);
2708                 }
2709                 inode = de->d_inode;
2710                 if (inode == NULL)
2711                         GOTO(out_put, rc = -ENOENT);
2712
2713                 down(&inode->i_sem);
2714                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
2715                 if (IS_ERR(handle)) {
2716                         up(&inode->i_sem);
2717                         GOTO(out_put, rc = PTR_ERR(handle));
2718                 }
2719
2720                 rc = mds_update_inode_ids(obd, inode, handle, NULL, ids + 1);
2721
2722                 err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
2723                                     exp->exp_sync);
2724                 if (err) {
2725                         CERROR("error committing transaction: %d\n", err);
2726                         if (!rc) rc = err;
2727                 }
2728                 up(&inode->i_sem);
2729 out_put:
2730                 l_dput(de);
2731                 RETURN(rc);
2732         }
2733
2734         if (keylen >= strlen("crypto_type") &&
2735              memcmp(key, "crypto_type", keylen) == 0) {
2736                 rc = mds_set_crypto_type(obd, val, vallen);
2737                 RETURN(rc);
2738         }
2739
2740         if (keylen >= strlen("setext") && !memcmp(key, "setext", keylen)) {
2741                 struct fid_extent *ext = val;
2742
2743                 CDEBUG(D_IOCTL, "set last fid to extent ["LPD64"-"LPD64"]\n",
2744                        ext->fe_start, ext->fe_width);
2745
2746                 /* set lastfid into fid extent start. All next object creates
2747                  * will use that fid. */
2748                 mds_set_last_fid(obd, ext->fe_start);
2749
2750                 /* setting the same extent to OSC to avoid ids intersecting in
2751                  * object ids, as all cache MDSs have the same group 0. */
2752                 rc = obd_set_info(mds->mds_dt_exp, strlen("setext"),
2753                                   "setext", sizeof(*ext), ext);
2754                 if (rc) {
2755                         CERROR("can't set extent ["LPD64"-"LPD64"] to %s, "
2756                                "err %d\n", ext->fe_start, ext->fe_width,
2757                                mds->mds_dt_exp->exp_obd->obd_name, rc);
2758                 }
2759                 RETURN(rc);
2760         }
2761         CDEBUG(D_IOCTL, "invalid key\n");
2762         RETURN(-EINVAL);
2763 }
2764
2765 static int mdt_set_info(struct ptlrpc_request *req)
2766 {
2767         char *key, *val;
2768         struct obd_export *exp = req->rq_export;
2769         int keylen, rc = 0, vallen = 0;
2770         ENTRY;
2771
2772         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2773         if (key == NULL) {
2774                 DEBUG_REQ(D_HA, req, "no set_info key");
2775                 RETURN(-EFAULT);
2776         }
2777         keylen = req->rq_reqmsg->buflens[0];
2778
2779         if ((keylen == strlen("crypto_type") &&
2780             memcmp(key, "crypto_type", keylen) == 0)) {
2781                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2782                 if (rc)
2783                         RETURN(rc);
2784
2785                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2786                 vallen = req->rq_reqmsg->buflens[1];
2787
2788                 rc = obd_set_info(exp, keylen, key, vallen, val);
2789                 req->rq_repmsg->status = 0;
2790                 RETURN(rc);
2791         } else if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2792                 struct audit_attr_msg msg, *p;
2793                 int rc = 0;
2794
2795                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2796                 if (rc)
2797                         RETURN(rc);
2798
2799                 p = lustre_swab_reqbuf(req, 1, sizeof(msg),
2800                                        lustre_swab_audit_attr);
2801
2802                 msg = *p;
2803                 CDEBUG(D_INFO, "Get new audit setting 0x%x\n", (__u32)msg.attr);
2804                 rc = obd_set_info(exp, keylen, key, sizeof(msg), &msg);
2805
2806                 req->rq_repmsg->status = rc;
2807                 RETURN(rc);
2808         } else if (keylen == strlen("ids") &&
2809                    memcmp(key, "ids", keylen) == 0) {
2810                 struct lustre_id *id, ids[2];
2811
2812                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2813                 if (rc)
2814                         RETURN(rc);
2815                 id = lustre_swab_reqbuf(req, 1, sizeof(struct lustre_id),
2816                                         lustre_swab_lustre_id);
2817                 ids[0] = *id;
2818                 id = lustre_swab_reqbuf(req, 2, sizeof(struct lustre_id),
2819                                         lustre_swab_lustre_id);
2820                 ids[1] = *id;
2821
2822                 rc = obd_set_info(exp, keylen, key, vallen, ids);
2823                 req->rq_repmsg->status = rc;
2824                 RETURN(rc);
2825         } else if (keylen == strlen("setext") &&
2826                    memcmp(key, "setext", keylen) == 0) {
2827                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2828                 if (rc)
2829                         RETURN(rc);
2830
2831                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2832                 vallen = req->rq_reqmsg->buflens[1];
2833
2834                 rc = obd_set_info(exp, keylen, key, vallen, val);
2835                 req->rq_repmsg->status = 0;
2836                 RETURN(rc);
2837         }
2838
2839         CDEBUG(D_IOCTL, "invalid key\n");
2840         RETURN(-EINVAL);
2841 }
2842
2843 static void mds_revoke_export_locks(struct obd_export *exp)
2844 {
2845         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2846         struct list_head  rpc_list;
2847         struct ldlm_lock *lock, *next;
2848         struct ldlm_lock_desc desc;
2849
2850         /* don't do this for local client */
2851         if (!exp->u.eu_mds_data.med_remote)
2852                 return;
2853
2854         /* don't revoke locks during recovery */
2855         if (exp->exp_obd->obd_recovering)
2856                 return;
2857
2858         ENTRY;
2859         INIT_LIST_HEAD(&rpc_list);
2860
2861         spin_lock(&exp->exp_ldlm_data.led_lock);
2862         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2863
2864                 lock_res_and_lock(lock);
2865                 if (lock->l_req_mode != lock->l_granted_mode) {
2866                         unlock_res_and_lock(lock);
2867                         continue;
2868                 }
2869
2870                 LASSERT(lock->l_resource);
2871                 if (lock->l_resource->lr_type != LDLM_IBITS &&
2872                     lock->l_resource->lr_type != LDLM_PLAIN) {
2873                         unlock_res_and_lock(lock);
2874                         continue;
2875                 }
2876
2877                 if (lock->l_flags & LDLM_FL_AST_SENT) {
2878                         unlock_res_and_lock(lock);
2879                         continue;
2880                 }
2881
2882                 LASSERT(lock->l_blocking_ast);
2883                 LASSERT(!lock->l_blocking_lock);
2884
2885                 lock->l_flags |= LDLM_FL_AST_SENT;
2886                 unlock_res_and_lock(lock);
2887
2888                 list_move(&lock->l_export_chain, &rpc_list);
2889         }
2890         spin_unlock(&exp->exp_ldlm_data.led_lock);
2891
2892         while (!list_empty(&rpc_list)) {
2893                 lock = list_entry(rpc_list.next, struct ldlm_lock,
2894                                   l_export_chain);
2895                 list_del_init(&lock->l_export_chain);
2896
2897                 /* the desc just pretend to exclusive */
2898                 ldlm_lock2desc(lock, &desc);
2899                 desc.l_req_mode = LCK_EX;
2900                 desc.l_granted_mode = 0;
2901
2902                 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2903         }
2904         EXIT;
2905 }
2906
2907 static int mds_msg_check_version(struct lustre_msg *msg)
2908 {
2909         int rc;
2910
2911         switch (msg->opc) {
2912         case MDS_CONNECT:
2913         case MDS_DISCONNECT:
2914         case OBD_PING:
2915                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2916                 if (rc)
2917                         CERROR("bad opc %u version %08x, expecting %08x\n",
2918                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2919                 break;
2920         case MDS_STATFS:
2921         case MDS_GETSTATUS:
2922         case MDS_GETATTR:
2923         case MDS_GETATTR_LOCK:
2924         case MDS_ACCESS_CHECK:
2925         case MDS_READPAGE:
2926         case MDS_REINT:
2927         case MDS_CLOSE:
2928         case MDS_DONE_WRITING:
2929         case MDS_PIN:
2930         case MDS_SYNC:
2931         case MDS_PARSE_ID:
2932                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2933                 if (rc)
2934                         CERROR("bad opc %u version %08x, expecting %08x\n",
2935                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2936                 break;
2937         case LDLM_ENQUEUE:
2938         case LDLM_CONVERT:
2939         case LDLM_BL_CALLBACK:
2940         case LDLM_CP_CALLBACK:
2941                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2942                 if (rc)
2943                         CERROR("bad opc %u version %08x, expecting %08x\n",
2944                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2945                 break;
2946         case OBD_LOG_CANCEL:
2947         case LLOG_ORIGIN_HANDLE_OPEN:
2948         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2949         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2950         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2951         case LLOG_ORIGIN_HANDLE_CLOSE:
2952         case LLOG_CATINFO:
2953                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2954                 if (rc)
2955                         CERROR("bad opc %u version %08x, expecting %08x\n",
2956                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2957                 break;
2958         case OST_CREATE:
2959         case OST_WRITE:
2960         case OST_GET_INFO:
2961         case OST_SET_INFO:
2962                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2963                 if (rc)
2964                         CERROR("bad opc %u version %08x, expecting %08x\n",
2965                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2966                 break;
2967         case SEC_INIT:
2968         case SEC_INIT_CONTINUE:
2969         case SEC_FINI:
2970                 rc = 0;
2971                 break;
2972         default:
2973                 CERROR("MDS unknown opcode %d\n", msg->opc);
2974                 rc = -ENOTSUPP;
2975                 break;
2976         }
2977
2978         return rc;
2979 }
2980
2981 int mds_handle(struct ptlrpc_request *req)
2982 {
2983         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2984         struct obd_device *obd = NULL;
2985         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2986         int rc = 0;
2987         ENTRY;
2988
2989         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2990
2991         rc = mds_msg_check_version(req->rq_reqmsg);
2992         if (rc) {
2993                 CERROR("MDS drop mal-formed request\n");
2994                 RETURN(rc);
2995         }
2996
2997         /* Security opc should NOT trigger any recovery events */
2998         if (req->rq_reqmsg->opc == SEC_INIT ||
2999             req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
3000                 if (req->rq_export) {
3001                         mds_req_add_idmapping(req,
3002                                               &req->rq_export->exp_mds_data);
3003                         mds_revoke_export_locks(req->rq_export);
3004                 }
3005                 GOTO(out, rc = 0);
3006         } else if (req->rq_reqmsg->opc == SEC_FINI) {
3007                 if (req->rq_export) {
3008                         mds_req_del_idmapping(req,
3009                                               &req->rq_export->exp_mds_data);
3010                         mds_revoke_export_locks(req->rq_export);
3011                 }
3012                 GOTO(out, rc = 0);
3013         }
3014
3015         LASSERT(current->journal_info == NULL);
3016         /* XXX identical to OST */
3017         if (req->rq_reqmsg->opc != MDS_CONNECT) {
3018                 struct mds_export_data *med;
3019                 int recovering;
3020
3021                 if (req->rq_export == NULL) {
3022                         CERROR("operation %d on unconnected MDS from %s\n",
3023                                req->rq_reqmsg->opc,
3024                                req->rq_peerstr);
3025                         req->rq_status = -ENOTCONN;
3026                         GOTO(out, rc = -ENOTCONN);
3027                 }
3028
3029                 med = &req->rq_export->exp_mds_data;
3030                 obd = req->rq_export->exp_obd;
3031                 mds = &obd->u.mds;
3032
3033                 /* sanity check: if the xid matches, the request must
3034                  * be marked as a resent or replayed */
3035                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
3036                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
3037                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
3038                                  (MSG_RESENT | MSG_REPLAY),
3039                                  "rq_xid "LPU64" matches last_xid, "
3040                                  "expected RESENT flag\n",
3041                                  req->rq_xid);
3042                 }
3043                 /* else: note the opposite is not always true; a
3044                  * RESENT req after a failover will usually not match
3045                  * the last_xid, since it was likely never
3046                  * committed. A REPLAYed request will almost never
3047                  * match the last xid, however it could for a
3048                  * committed, but still retained, open. */
3049
3050                 spin_lock_bh(&obd->obd_processing_task_lock);
3051                 recovering = obd->obd_recovering;
3052                 spin_unlock_bh(&obd->obd_processing_task_lock);
3053                 if (recovering) {
3054                         rc = mds_filter_recovery_request(req, obd,
3055                                                          &should_process);
3056                         if (rc || should_process == 0) {
3057                                 RETURN(rc);
3058                         } else if (should_process < 0) {
3059                                 req->rq_status = should_process;
3060                                 rc = ptlrpc_error(req);
3061                                 RETURN(rc);
3062                         }
3063                 }
3064         }
3065
3066         switch (req->rq_reqmsg->opc) {
3067         case MDS_CONNECT:
3068                 DEBUG_REQ(D_INODE, req, "connect");
3069                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
3070                 rc = target_handle_connect(req);
3071                 if (!rc) {
3072                         struct mds_export_data *med;
3073
3074                         LASSERT(req->rq_export);
3075                         med = &req->rq_export->u.eu_mds_data;
3076                         mds_init_export_data(req, med);
3077                         mds_req_add_idmapping(req, med);
3078
3079                         /* Now that we have an export, set mds. */
3080                         obd = req->rq_export->exp_obd;
3081                         mds = mds_req2mds(req);
3082                 }
3083                 break;
3084
3085         case MDS_DISCONNECT:
3086                 DEBUG_REQ(D_INODE, req, "disconnect");
3087                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
3088                 rc = target_handle_disconnect(req);
3089                 req->rq_status = rc;            /* superfluous? */
3090                 break;
3091
3092         case MDS_GETSTATUS:
3093                 DEBUG_REQ(D_INODE, req, "getstatus");
3094                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
3095                 rc = mds_getstatus(req);
3096                 break;
3097
3098         case MDS_GETATTR:
3099                 DEBUG_REQ(D_INODE, req, "getattr");
3100                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
3101                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
3102                 break;
3103
3104         case MDS_ACCESS_CHECK:
3105                 DEBUG_REQ(D_INODE, req, "access_check");
3106                 OBD_FAIL_RETURN(OBD_FAIL_MDS_ACCESS_CHECK_NET, 0);
3107                 rc = mds_access_check(req, MDS_REQ_REC_OFF);
3108                 break;
3109
3110         case MDS_GETATTR_LOCK: {
3111                 struct lustre_handle lockh;
3112                 DEBUG_REQ(D_INODE, req, "getattr_lock");
3113                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
3114
3115                 /* If this request gets a reconstructed reply, we won't be
3116                  * acquiring any new locks in mds_getattr_lock, so we don't
3117                  * want to cancel.
3118                  */
3119                 lockh.cookie = 0;
3120                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
3121                                       MDS_INODELOCK_UPDATE);
3122                 /* this non-intent call (from an ioctl) is special */
3123                 req->rq_status = rc;
3124                 if (rc == 0 && lockh.cookie)
3125                         ldlm_lock_decref(&lockh, LCK_PR);
3126                 break;
3127         }
3128         case MDS_STATFS:
3129                 DEBUG_REQ(D_INODE, req, "statfs");
3130                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
3131                 rc = mds_statfs(req);
3132                 break;
3133
3134         case MDS_READPAGE:
3135                 DEBUG_REQ(D_INODE, req, "readpage");
3136                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
3137                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
3138
3139                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
3140                         if (req->rq_reply_state) {
3141                                 lustre_free_reply_state (req->rq_reply_state);
3142                                 req->rq_reply_state = NULL;
3143                         }
3144                         RETURN(0);
3145                 }
3146
3147                 break;
3148         case MDS_REINT: {
3149                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
3150                                              sizeof (*opcp));
3151                 __u32  opc;
3152                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
3153                                mds->mds_max_cookiesize};
3154                 int bufcount;
3155
3156                 /* NB only peek inside req now; mds_reint() will swab it */
3157                 if (opcp == NULL) {
3158                         CERROR ("Can't inspect opcode\n");
3159                         rc = -EINVAL;
3160                         break;
3161                 }
3162                 opc = *opcp;
3163                 if (lustre_msg_swabbed (req->rq_reqmsg))
3164                         __swab32s(&opc);
3165
3166                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
3167                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
3168                            reint_names[opc] == NULL) ? reint_names[opc] :
3169                                                        "unknown opcode");
3170
3171                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
3172
3173                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
3174                         bufcount = 3;
3175                 else if (opc == REINT_OPEN)
3176                         bufcount = 2;
3177                 else
3178                         bufcount = 1;
3179
3180                 /* for SETATTR: I have different reply setting for
3181                  * remote setfacl, so delay the reply buffer allocation.
3182                  */
3183                 if (opc != REINT_SETATTR) {
3184                         rc = lustre_pack_reply(req, bufcount, size, NULL);
3185                         if (rc)
3186                                 break;
3187                 }
3188
3189                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
3190                 fail = OBD_FAIL_MDS_REINT_NET_REP;
3191                 break;
3192         }
3193
3194         case MDS_CLOSE:
3195                 DEBUG_REQ(D_INODE, req, "close");
3196                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
3197                 rc = mds_close(req, MDS_REQ_REC_OFF);
3198                 break;
3199
3200         case MDS_DONE_WRITING:
3201                 DEBUG_REQ(D_INODE, req, "done_writing");
3202                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
3203                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
3204                 break;
3205
3206         case MDS_PIN:
3207                 DEBUG_REQ(D_INODE, req, "pin");
3208                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
3209                 rc = mds_pin(req, MDS_REQ_REC_OFF);
3210                 break;
3211
3212         case MDS_SYNC:
3213                 DEBUG_REQ(D_INODE, req, "sync");
3214                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
3215                 rc = mds_sync(req, MDS_REQ_REC_OFF);
3216                 break;
3217         case MDS_PARSE_ID:
3218                 DEBUG_REQ(D_INODE, req, "parseid");
3219                 rc = mds_parse_id(req);
3220                 break;
3221         case OBD_PING:
3222                 DEBUG_REQ(D_INODE, req, "ping");
3223                 rc = target_handle_ping(req);
3224                 break;
3225
3226         case OBD_LOG_CANCEL:
3227                 CDEBUG(D_INODE, "log cancel\n");
3228                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
3229                 rc = -ENOTSUPP; /* la la la */
3230                 break;
3231
3232         case LDLM_ENQUEUE:
3233                 DEBUG_REQ(D_INODE, req, "enqueue");
3234                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
3235                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
3236                                          ldlm_server_blocking_ast, NULL);
3237                 fail = OBD_FAIL_LDLM_REPLY;
3238                 break;
3239         case LDLM_CONVERT:
3240                 DEBUG_REQ(D_INODE, req, "convert");
3241                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
3242                 rc = ldlm_handle_convert(req);
3243                 break;
3244         case LDLM_BL_CALLBACK:
3245         case LDLM_CP_CALLBACK:
3246                 DEBUG_REQ(D_INODE, req, "callback");
3247                 CERROR("callbacks should not happen on MDS\n");
3248                 LBUG();
3249                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
3250                 break;
3251         case LLOG_ORIGIN_HANDLE_OPEN:
3252                 DEBUG_REQ(D_INODE, req, "llog_init");
3253                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3254                 rc = llog_origin_handle_open(req);
3255                 break;
3256         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
3257                 DEBUG_REQ(D_INODE, req, "llog next block");
3258                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3259                 rc = llog_origin_handle_next_block(req);
3260                 break;
3261         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
3262                 DEBUG_REQ(D_INODE, req, "llog prev block");
3263                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3264                 rc = llog_origin_handle_prev_block(req);
3265                 break;
3266         case LLOG_ORIGIN_HANDLE_READ_HEADER:
3267                 DEBUG_REQ(D_INODE, req, "llog read header");
3268                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3269                 rc = llog_origin_handle_read_header(req);
3270                 break;
3271         case LLOG_ORIGIN_HANDLE_CLOSE:
3272                 DEBUG_REQ(D_INODE, req, "llog close");
3273                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3274                 rc = llog_origin_handle_close(req);
3275                 break;
3276         case OST_CREATE:
3277                 DEBUG_REQ(D_INODE, req, "ost_create");
3278                 rc = mdt_obj_create(req);
3279                 break;
3280         case OST_GET_INFO:
3281                 DEBUG_REQ(D_INODE, req, "get_info");
3282                 rc = mdt_get_info(req);
3283                 break;
3284         case OST_SET_INFO:
3285                 DEBUG_REQ(D_INODE, req, "set_info");
3286                 rc = mdt_set_info(req);
3287                 break;
3288         case OST_WRITE:
3289                 CDEBUG(D_INODE, "write\n");
3290                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
3291                 rc = ost_brw_write(req, NULL);
3292                 LASSERT(current->journal_info == NULL);
3293                 /* mdt_brw sends its own replies */
3294                 RETURN(rc);
3295                 break;
3296         case LLOG_CATINFO:
3297                 DEBUG_REQ(D_INODE, req, "llog catinfo");
3298                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3299                 rc = llog_catinfo(req);
3300                 break;
3301         default:
3302                 req->rq_status = -ENOTSUPP;
3303                 rc = ptlrpc_error(req);
3304                 RETURN(rc);
3305         }
3306
3307         LASSERT(current->journal_info == NULL);
3308
3309         /* If we're DISCONNECTing, the mds_export_data is already freed */
3310         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3311                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3312                 struct obd_device *obd = list_entry(mds, struct obd_device,
3313                                                     u.mds);
3314                 req->rq_repmsg->last_xid =
3315                         le64_to_cpu(med->med_mcd->mcd_last_xid);
3316
3317                 if (!obd->obd_no_transno) {
3318                         req->rq_repmsg->last_committed =
3319                                 obd->obd_last_committed;
3320                 } else {
3321                         DEBUG_REQ(D_IOCTL, req,
3322                                   "not sending last_committed update");
3323                 }
3324                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3325                        ", xid "LPU64"\n",
3326                        mds->mds_last_transno, obd->obd_last_committed,
3327                        req->rq_xid);
3328         }
3329  out:
3330
3331
3332         target_send_reply(req, rc, fail);
3333         RETURN(0);
3334 }
3335
3336 /* Update the server data on disk.  This stores the new mount_count and also the
3337  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
3338  * last_rcvd value may be less than that of the clients.  This will alert us
3339  * that we may need to do client recovery.
3340  *
3341  * Also assumes for mds_last_transno that we are not modifying it (no locking).
3342  */
3343 int mds_update_server_data(struct obd_device *obd, int force_sync)
3344 {
3345         struct mds_obd *mds = &obd->u.mds;
3346         struct mds_server_data *msd = mds->mds_server_data;
3347         struct file *filp = mds->mds_rcvd_filp;
3348         struct lvfs_run_ctxt saved;
3349         loff_t off = 0;
3350         int rc;
3351         ENTRY;
3352
3353         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3354         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3355
3356         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3357                mds->mds_mount_count, mds->mds_last_transno);
3358         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3359         if (rc)
3360                 CERROR("error writing MDS server data: rc = %d\n", rc);
3361         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3362
3363         RETURN(rc);
3364 }
3365
3366 /* saves last allocated fid counter to file. */
3367 int mds_update_last_fid(struct obd_device *obd, void *handle,
3368                         int force_sync)
3369 {
3370         struct mds_obd *mds = &obd->u.mds;
3371         struct file *filp = mds->mds_fid_filp;
3372         struct lvfs_run_ctxt saved;
3373         loff_t off = 0;
3374         __u64 last_fid;
3375         int rc = 0;
3376         ENTRY;
3377
3378         spin_lock(&mds->mds_last_fid_lock);
3379         last_fid = mds->mds_last_fid;
3380         spin_unlock(&mds->mds_last_fid_lock);
3381
3382         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3383                last_fid);
3384
3385         if (handle) {
3386                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3387                                       handle, mds_commit_last_fid_cb,
3388                                       NULL);
3389         }
3390
3391         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3392         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3393                                  &off, force_sync);
3394         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3395
3396         if (rc) {
3397                 CERROR("error writing MDS last_fid #"LPU64
3398                        ", err = %d\n", last_fid, rc);
3399                 RETURN(rc);
3400         }
3401
3402         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3403                "%llu: err = %d\n", last_fid, off, rc);
3404
3405         RETURN(rc);
3406 }
3407
3408 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3409 {
3410         struct mds_obd *mds = &obd->u.mds;
3411
3412         spin_lock(&mds->mds_last_fid_lock);
3413         if (fid > mds->mds_last_fid)
3414                 mds->mds_last_fid = fid;
3415         spin_unlock(&mds->mds_last_fid_lock);
3416 }
3417
3418 void mds_commit_last_transno_cb(struct obd_device *obd,
3419                                 __u64 transno, void *data,
3420                                 int error)
3421 {
3422         obd_transno_commit_cb(obd, transno, error);
3423 }
3424
3425 void mds_commit_last_fid_cb(struct obd_device *obd,
3426                             __u64 fid, void *data,
3427                             int error)
3428 {
3429         if (error) {
3430                 CERROR("%s: fid "LPD64" commit error: %d\n",
3431                        obd->obd_name, fid, error);
3432                 return;
3433         }
3434
3435         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3436                obd->obd_name, fid);
3437 }
3438
3439 __u64 mds_alloc_fid(struct obd_device *obd)
3440 {
3441         struct mds_obd *mds = &obd->u.mds;
3442         __u64 fid;
3443
3444         spin_lock(&mds->mds_last_fid_lock);
3445         fid = ++mds->mds_last_fid;
3446         spin_unlock(&mds->mds_last_fid_lock);
3447
3448         return fid;
3449 }
3450
3451 /*
3452  * reads inode self id from inode EA. Probably later this should be replaced by
3453  * caching inode self id to avoid raeding it every time it is needed.
3454  */
3455 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3456                        struct lustre_id *id)
3457 {
3458         int rc;
3459         ENTRY;
3460
3461         LASSERT(id != NULL);
3462         LASSERT(obd != NULL);
3463         LASSERT(inode != NULL);
3464
3465         rc = fsfilt_get_md(obd, inode, &id->li_fid,
3466                            sizeof(id->li_fid), EA_SID);
3467         if (rc < 0) {
3468                 CERROR("fsfilt_get_md() failed, "
3469                        "rc = %d\n", rc);
3470                 RETURN(rc);
3471         } else if (!rc) {
3472                 rc = -ENODATA;
3473                 RETURN(rc);
3474         } else {
3475                 rc = 0;
3476         }
3477         id_ino(id) = inode->i_ino;
3478         id_gen(id) = inode->i_generation;
3479         id_type(id) = S_IFMT & inode->i_mode;
3480
3481         RETURN(rc);
3482 }
3483
3484 int mds_read_inode_pid(struct obd_device *obd, struct inode *inode,
3485                        struct lustre_id *id)
3486 {
3487         int rc;
3488         ENTRY;
3489
3490         LASSERT(inode && id);
3491
3492         rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_PID);
3493         if (rc < 0)
3494                CERROR("get parent id from EA failed, rc=%d\n", rc);
3495         else if (!rc)
3496                 rc = -ENODATA;
3497         else
3498                 rc = 0;
3499
3500         RETURN(rc);
3501 }
3502
3503 /* updates inode self id in EA. */
3504 int mds_update_inode_ids(struct obd_device *obd, struct inode *inode,
3505                          void *handle, struct lustre_id *id,
3506                          struct lustre_id *pid)
3507 {
3508         int rc = 0;
3509         ENTRY;
3510
3511         LASSERT(id || pid);
3512         LASSERT(id == NULL || id_fid(id) != 0);
3513         LASSERT(pid == NULL || id_fid(pid) != 0);
3514         LASSERT(obd != NULL);
3515         LASSERT(inode != NULL);
3516
3517         if (id) {
3518                 mds_set_last_fid(obd, id_fid(id));
3519                 rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3520                                    sizeof(id->li_fid), EA_SID);
3521                 LASSERTF(rc == 0, "failed to update fid:  %d\n", rc);
3522         }
3523         if (pid) {
3524                 rc = fsfilt_set_md(obd, inode, handle, pid,
3525                                    sizeof(*pid), EA_PID);
3526                 LASSERTF(rc == 0, "failed to update parent fid:  %d\n", rc);
3527         }
3528
3529         RETURN(rc);
3530 }
3531
3532 /* mount the file system (secretly) */
3533 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3534 {
3535         struct lustre_cfg* lcfg = buf;
3536         struct mds_obd *mds = &obd->u.mds;
3537         struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3538         char *options = NULL;
3539         struct vfsmount *mnt;
3540         char ns_name[48];
3541         unsigned long page;
3542         struct crypto_tfm *tfm = NULL;
3543         int rc = 0;
3544         ENTRY;
3545
3546         if (lcfg->lcfg_bufcount < 3)
3547                 RETURN(rc = -EINVAL);
3548
3549         if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3550                 RETURN(rc = -EINVAL);
3551
3552         obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3553         if (IS_ERR(obd->obd_fsops))
3554                 RETURN(rc = PTR_ERR(obd->obd_fsops));
3555
3556         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3557
3558         page = get_zeroed_page(GFP_KERNEL);
3559         if (!page)
3560                 RETURN(-ENOMEM);
3561
3562         options = (char *)page;
3563
3564         /*
3565          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3566          * and the rest of options are passed by mount options. Probably this
3567          * should be moved to somewhere else like startup scripts or lconf. */
3568         sprintf(options, "iopen_nopriv");
3569
3570         if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3571                 sprintf(options + strlen(options), ",%s",
3572                         lustre_cfg_string(lcfg, 4));
3573
3574         /* we have to know mdsnum before touching underlying fs -bzzz */
3575         atomic_set(&mds->mds_open_count, 0);
3576         sema_init(&mds->mds_md_sem, 1);
3577         mds->mds_md_connected = 0;
3578         mds->mds_md_name = NULL;
3579
3580         if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3581             strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3582                 class_uuid_t uuid;
3583
3584                 generate_random_uuid(uuid);
3585                 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3586
3587                 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3588                 if (mds->mds_md_name == NULL)
3589                         RETURN(rc = -ENOMEM);
3590
3591                 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3592                        LUSTRE_CFG_BUFLEN(lcfg, 5));
3593
3594                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3595                        obd->obd_name, mds->mds_md_name);
3596
3597                 rc = mds_md_connect(obd, mds->mds_md_name);
3598                 if (rc) {
3599                         OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3600                         GOTO(err_ops, rc);
3601                 }
3602         }
3603
3604         mds->mds_obd_type = MDS_MASTER_OBD;
3605
3606         if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3607             strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3608                     LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3609                 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3610                             strlen("master"))) {
3611                         mds->mds_obd_type = MDS_MASTER_OBD;
3612                 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3613                                    strlen("cache"))) {
3614                         mds->mds_obd_type = MDS_CACHE_OBD;
3615                 }
3616         }
3617
3618         rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3619                            lustre_cfg_string(lcfg, 2),
3620                            options, 0, &lvfs_ctxt);
3621
3622         free_page(page);
3623
3624         if (rc || !lvfs_ctxt) {
3625                 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3626                 GOTO(err_ops, rc);
3627         }
3628
3629         mnt = lvfs_ctxt->loc_mnt;
3630         mds->mds_lvfs_ctxt = lvfs_ctxt;
3631         ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3632
3633         CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3634
3635         mds->mds_fidext_thumb = 0;
3636         sema_init(&mds->mds_epoch_sem, 1);
3637         atomic_set(&mds->mds_real_clients, 0);
3638         spin_lock_init(&mds->mds_fidext_lock);
3639         spin_lock_init(&mds->mds_fidmap_lock);
3640         spin_lock_init(&mds->mds_transno_lock);
3641         spin_lock_init(&mds->mds_last_fid_lock);
3642         sema_init(&mds->mds_orphan_recovery_sem, 1);
3643         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3644
3645         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3646         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3647
3648         if (obd->obd_namespace == NULL) {
3649                 mds_cleanup(obd, 0);
3650                 GOTO(err_put, rc = -ENOMEM);
3651         }
3652         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3653
3654         tfm = crypto_alloc_tfm(CAPA_HMAC_ALG, 0);
3655         if (!tfm)
3656                 GOTO(err_ns, rc = -ENOSYS);
3657
3658         mds->mds_capa_hmac = tfm;
3659         mds->mds_capa_timeout = CAPA_TIMEOUT;
3660         mds->mds_capa_key_timeout = CAPA_KEY_TIMEOUT;
3661
3662         rc = mds_fs_setup(obd, mnt);
3663         if (rc) {
3664                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3665                        obd->obd_name, rc);
3666                 GOTO(err_capa, rc);
3667         }
3668
3669         rc = llog_start_commit_thread();
3670         if (rc < 0)
3671
3672                 GOTO(err_fs, rc);
3673
3674         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3675             strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3676                     LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3677                 class_uuid_t uuid;
3678
3679                 generate_random_uuid(uuid);
3680                 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3681
3682                 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3683                 if (mds->mds_profile == NULL)
3684                         GOTO(err_fs, rc = -ENOMEM);
3685
3686                 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3687                         LUSTRE_CFG_BUFLEN(lcfg, 3));
3688         }
3689
3690         /*
3691          * setup root dir and files ID dir if lmv already connected, or there is
3692          * not lmv at all.
3693          */
3694         if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3695                                 lustre_cfg_buf(lcfg, 3) &&
3696                                 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3697                                         LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3698                 rc = mds_fs_setup_rootid(obd);
3699                 if (rc)
3700                         GOTO(err_fs, rc);
3701
3702                 rc = mds_fs_setup_virtid(obd);
3703                 if (rc)
3704                         GOTO(err_fs, rc);
3705
3706         }
3707
3708         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3709                            "mds_ldlm_client", &obd->obd_ldlm_client);
3710         obd->obd_replayable = 1;
3711
3712         mds->mds_crypto_type = NO_CRYPTO;
3713
3714         rc = mds_postsetup(obd);
3715         if (rc)
3716                 GOTO(err_fs, rc);
3717
3718         RETURN(0);
3719
3720 err_fs:
3721         /* No extra cleanup needed for llog_init_commit_thread() */
3722         mds_fs_cleanup(obd, 0);
3723 err_capa:
3724         crypto_free_tfm(mds->mds_capa_hmac);
3725 err_ns:
3726         ldlm_namespace_free(obd->obd_namespace, 0);
3727         obd->obd_namespace = NULL;
3728 err_put:
3729         unlock_kernel();
3730         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3731         mds->mds_sb = 0;
3732         lock_kernel();
3733 err_ops:
3734         fsfilt_put_ops(obd->obd_fsops);
3735         return rc;
3736 }
3737
3738 static int mds_fs_post_setup(struct obd_device *obd)
3739 {
3740         struct mds_obd *mds = &obd->u.mds;
3741         struct dentry *dentry;
3742         int rc = 0;
3743         ENTRY;
3744
3745         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
3746         if (IS_ERR(dentry)) {
3747                 CERROR("Can't find ROOT, err = %d\n",
3748                        (int)PTR_ERR(dentry));
3749                 RETURN(PTR_ERR(dentry));
3750         }
3751         rc = fsfilt_post_setup(obd, dentry);
3752         //set id2name function handler
3753         fsfilt_set_info(obd, mds->mds_sb, NULL, 7, "id2name",
3754                         sizeof(mds_audit_id2name), mds_audit_id2name);
3755
3756         l_dput(dentry);
3757         RETURN(rc);
3758 }
3759
3760 static int mds_postsetup(struct obd_device *obd)
3761 {
3762         struct mds_obd *mds = &obd->u.mds;
3763         int rc = 0;
3764         ENTRY;
3765
3766         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3767                             obd, 0, NULL, &llog_lvfs_ops);
3768         if (rc)
3769                 RETURN(rc);
3770
3771         if (mds->mds_profile) {
3772                 struct llog_ctxt *lgctxt;
3773                 struct lvfs_run_ctxt saved;
3774                 struct lustre_profile *lprof;
3775                 struct config_llog_instance cfg;
3776
3777                 cfg.cfg_instance = NULL;
3778                 cfg.cfg_uuid = mds->mds_dt_uuid;
3779                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3780
3781                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3782                 if (!lgctxt) {
3783                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3784                         GOTO(err_llog, rc = -EINVAL);
3785                 }
3786
3787                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3788                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3789
3790                 if (rc)
3791                         GOTO(err_llog, rc);
3792
3793                 lprof = class_get_profile(mds->mds_profile);
3794                 if (lprof == NULL) {
3795                         CERROR("No profile found: %s\n", mds->mds_profile);
3796                         GOTO(err_cleanup, rc = -ENOENT);
3797                 }
3798                 rc = mds_dt_connect(obd, lprof->lp_lov);
3799                 if (rc)
3800                         GOTO(err_cleanup, rc);
3801
3802                 rc = mds_md_postsetup(obd);
3803                 if (rc)
3804                         GOTO(err_cleanup, rc);
3805         }
3806         rc = mds_fs_post_setup(obd);
3807         if (rc)
3808                 CERROR("can not post setup fsfilt\n");
3809         RETURN(rc);
3810 err_cleanup:
3811         mds_dt_clean(obd);
3812 err_llog:
3813         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3814                                           LLOG_CONFIG_ORIG_CTXT));
3815         return rc;
3816 }
3817
3818 int mds_postrecov_common(struct obd_device *obd)
3819 {
3820         struct mds_obd *mds = &obd->u.mds;
3821         struct llog_ctxt *ctxt;
3822         int rc, item = 0, valsize;
3823          __u32 group;
3824         ENTRY;
3825
3826         LASSERT(!obd->obd_recovering);
3827         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3828         LASSERT(ctxt != NULL);
3829
3830         /* clean PENDING dir */
3831         rc = mds_cleanup_orphans(obd);
3832         if (rc < 0)
3833                 GOTO(out, rc);
3834         item = rc;
3835
3836         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3837         valsize = sizeof(group);
3838         rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3839                           "mds_conn", valsize, &group);
3840         if (rc)
3841                 GOTO(out, rc);
3842
3843         rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3844                           NULL, NULL, NULL);
3845         if (rc) {
3846                 CERROR("%s: failed at llog_origin_connect: %d\n",
3847                        obd->obd_name, rc);
3848                 GOTO(out, rc);
3849         }
3850
3851         /* remove the orphaned precreated objects */
3852         rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3853         if (rc)
3854                 GOTO(err_llog, rc);
3855
3856 out:
3857         RETURN(rc < 0 ? rc : item);
3858
3859 err_llog:
3860         /* cleanup all llogging subsystems */
3861         rc = obd_llog_finish(obd, &obd->obd_llogs,
3862                              mds->mds_dt_desc.ld_tgt_count);
3863         if (rc)
3864                 CERROR("%s: failed to cleanup llogging subsystems\n",
3865                         obd->obd_name);
3866         goto out;
3867 }
3868
3869 int mds_postrecov(struct obd_device *obd)
3870 {
3871         int rc;
3872         ENTRY;
3873         rc = mds_postrecov_common(obd);
3874         if (rc == 0)
3875                 rc = mds_md_reconnect(obd);
3876         RETURN(rc);
3877 }
3878
3879 int mds_dt_clean(struct obd_device *obd)
3880 {
3881         struct mds_obd *mds = &obd->u.mds;
3882         ENTRY;
3883
3884         if (mds->mds_profile) {
3885                 char * cln_prof;
3886                 struct llog_ctxt *llctx;
3887                 struct lvfs_run_ctxt saved;
3888                 struct config_llog_instance cfg;
3889                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3890
3891                 OBD_ALLOC(cln_prof, len);
3892                 if (!cln_prof) {
3893                         CERROR("can't allocate memory, processing cleanup "
3894                                "profile is skipped\n");
3895                         goto out;
3896                 }
3897
3898                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3899
3900                 cfg.cfg_instance = NULL;
3901                 cfg.cfg_uuid = mds->mds_dt_uuid;
3902
3903                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3904                 llctx = llog_get_context(&obd->obd_llogs,
3905                                          LLOG_CONFIG_ORIG_CTXT);
3906                 class_config_process_llog(llctx, cln_prof, &cfg);
3907                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3908
3909                 OBD_FREE(cln_prof, len);
3910         out:
3911                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3912                 mds->mds_profile = NULL;
3913         }
3914         RETURN(0);
3915 }
3916
3917 int mds_md_clean(struct obd_device *obd)
3918 {
3919         struct mds_obd *mds = &obd->u.mds;
3920         ENTRY;
3921
3922         if (mds->mds_md_name) {
3923                 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3924                 mds->mds_md_name = NULL;
3925         }
3926         RETURN(0);
3927 }
3928
3929 static int mds_precleanup(struct obd_device *obd, int flags)
3930 {
3931         int rc = 0;
3932         ENTRY;
3933
3934         mds_md_clean(obd);
3935         mds_dt_disconnect(obd, flags);
3936         mds_dt_clean(obd);
3937         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3938         RETURN(rc);
3939 }
3940
3941 extern void lgss_svc_cache_purge_all(void);
3942 static int mds_cleanup(struct obd_device *obd, int flags)
3943 {
3944         struct mds_obd *mds = &obd->u.mds;
3945         ENTRY;
3946
3947         if (mds->mds_sb == NULL)
3948                 RETURN(0);
3949
3950         mds_update_server_data(obd, 1);
3951         mds_update_last_fid(obd, NULL, 1);
3952
3953         if (mds->mds_dt_objids != NULL) {
3954                 int size = mds->mds_dt_desc.ld_tgt_count *
3955                         sizeof(obd_id);
3956                 OBD_FREE(mds->mds_dt_objids, size);
3957         }
3958         mds_fs_cleanup(obd, flags);
3959
3960         unlock_kernel();
3961
3962         /* 2 seems normal on mds, (may_umount() also expects 2
3963           fwiw), but we only see 1 at this point in obdfilter. */
3964         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3965
3966         mds->mds_sb = 0;
3967
3968         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3969
3970         spin_lock_bh(&obd->obd_processing_task_lock);
3971         if (obd->obd_recovering) {
3972                 target_cancel_recovery_timer(obd);
3973                 obd->obd_recovering = 0;
3974         }
3975         spin_unlock_bh(&obd->obd_processing_task_lock);
3976
3977         lock_kernel();
3978         fsfilt_put_ops(obd->obd_fsops);
3979
3980 #ifdef ENABLE_GSS
3981         /* XXX */
3982         lgss_svc_cache_purge_all();
3983 #endif
3984
3985         spin_lock(&mds->mds_denylist_lock);
3986         while (!list_empty( &mds->mds_denylist ) ) {
3987                 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3988                                                     deny_sec_t, list);
3989                 list_del(&p_deny_sec->list);
3990                 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3991         }
3992         spin_unlock(&mds->mds_denylist_lock);
3993
3994         mds_capa_keys_cleanup(obd);
3995
3996         if (mds->mds_capa_hmac)
3997                 crypto_free_tfm(mds->mds_capa_hmac);
3998         RETURN(0);
3999 }
4000
4001 static int set_security(const char *value, char **sec)
4002 {
4003         if (!strcmp(value, "null"))
4004                 *sec = "null";
4005         else if (!strcmp(value, "krb5i"))
4006                 *sec = "krb5i";
4007         else if (!strcmp(value, "krb5p"))
4008                 *sec = "krb5p";
4009         else {
4010                 CERROR("Unrecognized security flavor %s\n", value);
4011                 return -EINVAL;
4012         }
4013
4014         return 0;
4015 }
4016
4017 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
4018 {
4019         struct lustre_cfg *lcfg = buf;
4020         struct mds_obd *mds = &obd->u.mds;
4021         int rc = 0;
4022         ENTRY;
4023
4024         switch(lcfg->lcfg_command) {
4025         case LCFG_SET_SECURITY: {
4026                 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
4027                     (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
4028                         GOTO(out, rc = -EINVAL);
4029
4030                 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
4031                         rc = set_security(lustre_cfg_string(lcfg, 2),
4032                                           &mds->mds_mds_sec);
4033                 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
4034                         rc = set_security(lustre_cfg_string(lcfg, 2),
4035                                           &mds->mds_ost_sec);
4036                 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
4037                         spin_lock(&mds->mds_denylist_lock);
4038                         rc = add_deny_security(lustre_cfg_string(lcfg, 2),
4039                                                &mds->mds_denylist);
4040                         spin_unlock(&mds->mds_denylist_lock);
4041                 } else {
4042                         CERROR("Unrecognized key\n");
4043                         rc = -EINVAL;
4044                 }
4045                 break;
4046         }
4047         default:
4048                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
4049                 GOTO(out, rc = -EINVAL);
4050         }
4051 out:
4052         RETURN(rc);
4053 }
4054
4055 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
4056                                         int offset,
4057                                         struct ldlm_lock *new_lock,
4058                                         struct ldlm_lock **old_lock,
4059                                         struct lustre_handle *lockh)
4060 {
4061         struct obd_export *exp = req->rq_export;
4062         struct obd_device *obd = exp->exp_obd;
4063         struct ldlm_request *dlmreq =
4064                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
4065         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
4066         struct list_head *iter;
4067
4068         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4069                 return;
4070
4071         spin_lock(&obd->obd_namespace->ns_hash_lock);
4072         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
4073                 struct ldlm_lock *lock;
4074                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
4075                 if (lock == new_lock)
4076                         continue;
4077                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
4078                         lockh->cookie = lock->l_handle.h_cookie;
4079                         LDLM_DEBUG(lock, "restoring lock cookie");
4080                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
4081                                   lockh->cookie);
4082                         if (old_lock)
4083                                 *old_lock = LDLM_LOCK_GET(lock);
4084                         spin_unlock(&obd->obd_namespace->ns_hash_lock);
4085                         return;
4086                 }
4087         }
4088         spin_unlock(&obd->obd_namespace->ns_hash_lock);
4089
4090         /* If the xid matches, then we know this is a resent request,
4091          * and allow it. (It's probably an OPEN, for which we don't
4092          * send a lock */
4093         if (req->rq_xid ==
4094             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
4095                 return;
4096
4097         if (req->rq_xid ==
4098             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
4099                 return;
4100
4101         /* This remote handle isn't enqueued, so we never received or
4102          * processed this request.  Clear MSG_RESENT, because it can
4103          * be handled like any normal request now. */
4104
4105         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4106
4107         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
4108                   remote_hdl.cookie);
4109 }
4110
4111 int intent_disposition(struct ldlm_reply *rep, int flag)
4112 {
4113         if (!rep)
4114                 return 0;
4115         return (rep->lock_policy_res1 & flag);
4116 }
4117
4118 void intent_set_disposition(struct ldlm_reply *rep, int flag)
4119 {
4120         if (!rep)
4121                 return;
4122         rep->lock_policy_res1 |= flag;
4123 }
4124
4125 static int mds_intent_prepare_reply_buffers(struct ptlrpc_request *req,
4126                                             struct ldlm_intent *it)
4127 {
4128         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
4129         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
4130         int rc, reply_buffers;
4131         int repsize[8] = {sizeof(struct ldlm_reply),
4132                           sizeof(struct mds_body),
4133                           mds->mds_max_mdsize};
4134         ENTRY;
4135
4136         reply_buffers = 3;
4137         if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
4138                 /*Ugly here, Actually, we should prepare the reply buffer
4139                  *after we know whether these stuff exist or not, which should
4140                  * be fixed in future, Now each item is in the fix position,
4141                  * the sequence is lsm, acl, crypto ea, capa.*/
4142                 repsize[reply_buffers++] = sizeof(int);
4143                 if (med->med_remote)
4144                         repsize[reply_buffers++] =
4145                                 sizeof(struct mds_remote_perm);
4146                 else
4147                         repsize[reply_buffers++] =
4148                                 xattr_acl_size(LL_ACL_MAX_ENTRIES);
4149
4150                 repsize[reply_buffers++] = sizeof(int);
4151                 repsize[reply_buffers++] = sizeof(struct crypto_key);
4152
4153                 /* XXX: if new buffer is to be added, capability reply
4154                  *      buffer should always been reserved. */
4155                 if (it->opc & IT_OPEN)
4156                         repsize[reply_buffers++] = sizeof(struct lustre_capa);
4157         }
4158
4159         rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
4160
4161         RETURN(rc);
4162 }
4163
4164 static int mds_intent_policy(struct ldlm_namespace *ns,
4165                              struct ldlm_lock **lockp, void *req_cookie,
4166                              ldlm_mode_t mode, int flags, void *data)
4167 {
4168         struct ptlrpc_request *req = req_cookie;
4169         struct ldlm_lock *lock = *lockp;
4170         struct ldlm_intent *it;
4171         struct ldlm_reply *rep;
4172         struct lustre_handle lockh[2] = {{0}, {0}};
4173         struct ldlm_lock *new_lock = NULL;
4174         int getattr_part = MDS_INODELOCK_UPDATE;
4175         int rc;
4176
4177         int offset = MDS_REQ_INTENT_REC_OFF;
4178         ENTRY;
4179
4180         LASSERT(req != NULL);
4181         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
4182
4183         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
4184                 /* No intent was provided */
4185                 int size = sizeof(struct ldlm_reply);
4186                 rc = lustre_pack_reply(req, 1, &size, NULL);
4187                 LASSERT(rc == 0);
4188                 RETURN(0);
4189         }
4190
4191         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
4192                                 lustre_swab_ldlm_intent);
4193         if (it == NULL) {
4194                 CERROR("Intent missing\n");
4195                 RETURN(req->rq_status = -EFAULT);
4196         }
4197
4198         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
4199
4200         rc = mds_intent_prepare_reply_buffers(req, it);
4201
4202         if (rc)
4203                 RETURN(req->rq_status = rc);
4204
4205         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
4206         LASSERT(rep != NULL);
4207
4208         intent_set_disposition(rep, DISP_IT_EXECD);
4209
4210         /* execute policy */
4211         switch ((long)it->opc) {
4212         case IT_OPEN:
4213         case IT_CREAT|IT_OPEN:
4214                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4215                                             lock, NULL, lockh);
4216                 /* XXX swab here to assert that an mds_open reint
4217                  * packet is following */
4218                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4219                                             lock, NULL, lockh);
4220                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
4221
4222                 if (rep->lock_policy_res2) {
4223                         /*
4224                          * mds_open() returns ENOLCK where it should return
4225                          * zero, but it has no lock to return.
4226                          */
4227                         if (rep->lock_policy_res2 == ENOLCK)
4228                                 rep->lock_policy_res2 = 0;
4229
4230                         RETURN(ELDLM_LOCK_ABORTED);
4231                 }
4232
4233                 /*
4234                  * IT_OPEN may return lock on cross-node dentry that we want to
4235                  * hold during attr retrival -bzzz
4236                  */
4237                 if (lockh[0].cookie == 0)
4238                         RETURN(ELDLM_LOCK_ABORTED);
4239
4240                 break;
4241         case IT_LOOKUP:
4242                 getattr_part = MDS_INODELOCK_LOOKUP;
4243         case IT_CHDIR:
4244         case IT_GETATTR:
4245                 getattr_part |= MDS_INODELOCK_LOOKUP;
4246         case IT_READDIR:
4247                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4248                                             lock, &new_lock, lockh);
4249                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
4250                                                          getattr_part);
4251                 /* FIXME: LDLM can set req->rq_status. MDS sets
4252                    policy_res{1,2} with disposition and status.
4253                    - replay: returns 0 & req->status is old status
4254                    - otherwise: returns req->status */
4255                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
4256                         rep->lock_policy_res2 = 0;
4257                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
4258                     rep->lock_policy_res2)
4259                         RETURN(ELDLM_LOCK_ABORTED);
4260                 if (req->rq_status != 0) {
4261                         LBUG();
4262                         rep->lock_policy_res2 = req->rq_status;
4263                         RETURN(ELDLM_LOCK_ABORTED);
4264                 }
4265                 break;
4266         case IT_UNLINK:
4267                 rc = mds_lock_and_check_slave(offset, req, lockh);
4268                 if ((rep->lock_policy_res2 = rc)) {
4269                         if (rc == ENOLCK)
4270                                 rep->lock_policy_res2 = 0;
4271                         RETURN(ELDLM_LOCK_ABORTED);
4272                 }
4273                 break;
4274         default:
4275                 CERROR("Unhandled intent "LPD64"\n", it->opc);
4276                 LBUG();
4277         }
4278
4279         /* By this point, whatever function we called above must have either
4280          * filled in 'lockh', been an intent replay, or returned an error.  We
4281          * want to allow replayed RPCs to not get a lock, since we would just
4282          * drop it below anyways because lock replay is done separately by the
4283          * client afterwards.  For regular RPCs we want to give the new lock to
4284          * the client instead of whatever lock it was about to get. */
4285         if (new_lock == NULL)
4286                 new_lock = ldlm_handle2lock(&lockh[0]);
4287         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
4288                 RETURN(0);
4289
4290         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
4291                  it->opc, lockh[0].cookie);
4292
4293         /* If we've already given this lock to a client once, then we should
4294          * have no readers or writers.  Otherwise, we should have one reader
4295          * _or_ writer ref (which will be zeroed below) before returning the
4296          * lock to a client. */
4297         if (new_lock->l_export == req->rq_export) {
4298                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4299         } else {
4300                 LASSERT(new_lock->l_export == NULL);
4301                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4302         }
4303
4304         *lockp = new_lock;
4305
4306         if (new_lock->l_export == req->rq_export) {
4307                 /* Already gave this to the client, which means that we
4308                  * reconstructed a reply. */
4309                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4310                         MSG_RESENT);
4311                 RETURN(ELDLM_LOCK_REPLACED);
4312         }
4313
4314         /* Fixup the lock to be given to the client */
4315         lock_res_and_lock(new_lock);
4316         new_lock->l_readers = 0;
4317         new_lock->l_writers = 0;
4318
4319         new_lock->l_export = class_export_get(req->rq_export);
4320
4321         spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4322         list_add(&new_lock->l_export_chain,
4323                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
4324         spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4325
4326         new_lock->l_blocking_ast = lock->l_blocking_ast;
4327         new_lock->l_completion_ast = lock->l_completion_ast;
4328
4329         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4330                sizeof(lock->l_remote_handle));
4331
4332         new_lock->l_flags &= ~LDLM_FL_LOCAL;
4333
4334         unlock_res_and_lock(new_lock);
4335         LDLM_LOCK_PUT(new_lock);
4336
4337         RETURN(ELDLM_LOCK_REPLACED);
4338 }
4339
4340 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4341 {
4342         struct lprocfs_static_vars lvars;
4343         int rc = 0;
4344         struct mds_obd *mds = &dev->u.mds;
4345
4346         spin_lock_init(&mds->mds_denylist_lock);
4347         INIT_LIST_HEAD(&mds->mds_denylist);
4348
4349         lprocfs_init_multi_vars(0, &lvars);
4350
4351         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4352         if (rc)
4353                 return rc;
4354
4355         return lprocfs_alloc_md_stats(dev, 0);
4356 }
4357
4358 int mds_detach(struct obd_device *dev)
4359 {
4360         lprocfs_free_md_stats(dev);
4361         return lprocfs_obd_detach(dev);
4362 }
4363
4364 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4365 {
4366         struct lprocfs_static_vars lvars;
4367
4368         lprocfs_init_multi_vars(1, &lvars);
4369         return lprocfs_obd_attach(dev, lvars.obd_vars);
4370 }
4371
4372 int mdt_detach(struct obd_device *dev)
4373 {
4374         return lprocfs_obd_detach(dev);
4375 }
4376
4377 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4378 {
4379         struct mds_obd *mds = &obd->u.mds;
4380         int rc = 0;
4381         ENTRY;
4382
4383         mds->mds_service =
4384                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4385                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4386                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4387                                 mds_handle, "mds", obd->obd_proc_entry);
4388
4389         if (!mds->mds_service) {
4390                 CERROR("failed to start service\n");
4391                 RETURN(-ENOMEM);
4392         }
4393
4394         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4395                                     "ll_mdt");
4396         if (rc)
4397                 GOTO(err_thread, rc);
4398
4399         mds->mds_setattr_service =
4400                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4401                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4402                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4403                                 mds_handle, "mds_setattr",
4404                                 obd->obd_proc_entry);
4405         if (!mds->mds_setattr_service) {
4406                 CERROR("failed to start getattr service\n");
4407                 GOTO(err_thread, rc = -ENOMEM);
4408         }
4409
4410         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4411                                     MDT_NUM_THREADS, "ll_mdt_attr");
4412         if (rc)
4413                 GOTO(err_thread2, rc);
4414
4415         mds->mds_readpage_service =
4416                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4417                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4418                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4419                                 mds_handle, "mds_readpage",
4420                                 obd->obd_proc_entry);
4421         if (!mds->mds_readpage_service) {
4422                 CERROR("failed to start readpage service\n");
4423                 GOTO(err_thread2, rc = -ENOMEM);
4424         }
4425
4426         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4427                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
4428
4429         if (rc)
4430                 GOTO(err_thread3, rc);
4431
4432         mds->mds_close_service =
4433                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4434                                 MDS_CLOSE_PORTAL, MDC_REPLY_PORTAL,
4435                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
4436                                 mds_handle, "mds_close",
4437                                 obd->obd_proc_entry);
4438         if (!mds->mds_close_service) {
4439                 CERROR("failed to start close service\n");
4440                 GOTO(err_thread3, rc = -ENOMEM);
4441         }
4442
4443         rc = ptlrpc_start_n_threads(obd, mds->mds_close_service,
4444                                     MDT_NUM_THREADS, "ll_mdt_clos");
4445
4446         if (rc)
4447                 GOTO(err_thread4, rc);
4448         RETURN(0);
4449
4450 err_thread4:
4451         ptlrpc_unregister_service(mds->mds_close_service);
4452 err_thread3:
4453         ptlrpc_unregister_service(mds->mds_readpage_service);
4454 err_thread2:
4455         ptlrpc_unregister_service(mds->mds_setattr_service);
4456 err_thread:
4457         ptlrpc_unregister_service(mds->mds_service);
4458         return rc;
4459 }
4460
4461 static int mdt_cleanup(struct obd_device *obd, int flags)
4462 {
4463         struct mds_obd *mds = &obd->u.mds;
4464         ENTRY;
4465
4466         ptlrpc_stop_all_threads(mds->mds_close_service);
4467         ptlrpc_unregister_service(mds->mds_close_service);
4468
4469         ptlrpc_stop_all_threads(mds->mds_readpage_service);
4470         ptlrpc_unregister_service(mds->mds_readpage_service);
4471
4472         ptlrpc_stop_all_threads(mds->mds_setattr_service);
4473         ptlrpc_unregister_service(mds->mds_setattr_service);
4474
4475         ptlrpc_stop_all_threads(mds->mds_service);
4476         ptlrpc_unregister_service(mds->mds_service);
4477
4478         RETURN(0);
4479 }
4480
4481 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4482                                          __u64 gr, void *data)
4483 {
4484         struct lustre_id id;
4485         struct obd_device *obd = data;
4486
4487         id_ino(&id) = ino;
4488         id_gen(&id) = gen;
4489         return mds_id2dentry(obd, &id, NULL);
4490 }
4491
4492 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4493                         void *key, __u32 *valsize, void *val)
4494 {
4495         struct obd_device *obd;
4496         struct mds_obd *mds;
4497         int rc = 0;
4498         ENTRY;
4499
4500         obd = class_exp2obd(exp);
4501         mds = &obd->u.mds;
4502
4503         if (obd == NULL) {
4504                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4505                        exp->exp_handle.h_cookie);
4506                 RETURN(-EINVAL);
4507         }
4508
4509         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4510                 /* get log_context handle. */
4511                 struct llog_ctxt *ctxt;
4512                 unsigned long *llh_handle = val;
4513                 *valsize = sizeof(unsigned long);
4514                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_REINT_ORIG_CTXT);
4515                 if (!ctxt) {
4516                         CERROR("Cannot get REINT llog context\n");
4517                         RETURN(-ENOENT);
4518                 }
4519                 *llh_handle = (unsigned long)ctxt;
4520                 RETURN(0);
4521         }
4522         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4523                 /* get log_context handle. */
4524                 unsigned long *sb = val;
4525                 *valsize = sizeof(unsigned long);
4526                 *sb = (unsigned long)obd->u.mds.mds_sb;
4527                 RETURN(0);
4528         }
4529
4530         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4531                 __u32 *mdsize = val;
4532                 *valsize = sizeof(*mdsize);
4533                 *mdsize = mds->mds_max_mdsize;
4534                 RETURN(0);
4535         }
4536
4537         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4538                 __u32 *mdsnum = val;
4539                 *valsize = sizeof(*mdsnum);
4540                 *mdsnum = mds->mds_num;
4541                 RETURN(0);
4542         }
4543
4544         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4545                 struct lustre_id *rootid = val;
4546                 *valsize = sizeof(*rootid);
4547                 *rootid = mds->mds_rootid;
4548                 RETURN(0);
4549         }
4550
4551         if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
4552                 struct lov_desc *desc = val;
4553                 *valsize = sizeof(*desc);
4554                 *desc = mds->mds_dt_desc;
4555                 RETURN(0);
4556         }
4557
4558         if (keylen >= strlen("getext") && strcmp(key, "getext") == 0) {
4559                 struct fid_extent *ext = val;
4560                 *valsize = sizeof(*ext);
4561
4562                 spin_lock(&mds->mds_fidext_lock);
4563                 ext->fe_width = MDS_FIDEXT_SIZE;
4564                 ext->fe_start = mds->mds_fidext_thumb + 1;
4565                 mds->mds_fidext_thumb += MDS_FIDEXT_SIZE;
4566                 spin_unlock(&mds->mds_fidext_lock);
4567
4568                 RETURN(0);
4569         }
4570
4571         rc = fsfilt_get_info(obd, mds->mds_sb, NULL, keylen, key, valsize, val);
4572         if (rc)
4573                 CDEBUG(D_IOCTL, "invalid key\n");
4574
4575         RETURN(rc);
4576 }
4577
4578 struct lvfs_callback_ops mds_lvfs_ops = {
4579         l_id2dentry:     mds_lvfs_id2dentry,
4580 };
4581
4582 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4583                 int objcount, struct obd_ioobj *obj,
4584                 int niocount, struct niobuf_remote *nb,
4585                 struct niobuf_local *res,
4586                 struct obd_trans_info *oti, struct lustre_capa *capa);
4587
4588 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4589                  int objcount, struct obd_ioobj *obj, int niocount,
4590                  struct niobuf_local *res, struct obd_trans_info *oti,
4591                  int rc);
4592
4593 /* use obd ops to offer management infrastructure */
4594 static struct obd_ops mds_obd_ops = {
4595         .o_owner           = THIS_MODULE,
4596         .o_attach          = mds_attach,
4597         .o_detach          = mds_detach,
4598         .o_connect         = mds_connect,
4599         .o_connect_post    = mds_connect_post,
4600         .o_init_export     = mds_init_export,
4601         .o_destroy_export  = mds_destroy_export,
4602         .o_disconnect      = mds_disconnect,
4603         .o_setup           = mds_setup,
4604         .o_precleanup      = mds_precleanup,
4605         .o_cleanup         = mds_cleanup,
4606         .o_process_config  = mds_process_config,
4607         .o_postrecov       = mds_postrecov,
4608         .o_statfs          = mds_obd_statfs,
4609         .o_iocontrol       = mds_iocontrol,
4610         .o_create          = mds_obd_create,
4611         .o_destroy         = mds_obd_destroy,
4612         .o_llog_init       = mds_llog_init,
4613         .o_llog_finish     = mds_llog_finish,
4614         .o_notify          = mds_notify,
4615         .o_get_info        = mds_get_info,
4616         .o_set_info        = mds_set_info,
4617         .o_preprw          = mds_preprw,
4618         .o_commitrw        = mds_commitrw,
4619 };
4620
4621 static struct obd_ops mdt_obd_ops = {
4622         .o_owner           = THIS_MODULE,
4623         .o_attach          = mdt_attach,
4624         .o_detach          = mdt_detach,
4625         .o_setup           = mdt_setup,
4626         .o_cleanup         = mdt_cleanup,
4627 };
4628
4629 static int __init mds_init(void)
4630 {
4631         struct lprocfs_static_vars lvars;
4632         int rc = 0;
4633         ENTRY;
4634
4635         mds_init_lsd_cache();
4636         mds_init_rmtacl_upcall_cache();
4637
4638         lprocfs_init_multi_vars(0, &lvars);
4639         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4640                             OBD_MDS_DEVICENAME);
4641         lprocfs_init_multi_vars(1, &lvars);
4642         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4643                             OBD_MDT_DEVICENAME);
4644
4645         rc = mds_capa_key_start_thread();
4646         if (rc) {
4647                 class_unregister_type(OBD_MDT_DEVICENAME);
4648                 class_unregister_type(OBD_MDS_DEVICENAME);
4649                 mds_cleanup_lsd_cache();
4650         }
4651
4652         mds_eck_timer.function = mds_capa_key_timer_callback;
4653         mds_eck_timer.data = 0;
4654         init_timer(&mds_eck_timer);
4655
4656         RETURN(rc);
4657 }
4658
4659 static void /*__exit*/ mds_exit(void)
4660 {
4661         mds_capa_key_stop_thread();
4662         mds_cleanup_rmtacl_upcall_cache();
4663         mds_cleanup_lsd_cache();
4664
4665         class_unregister_type(OBD_MDS_DEVICENAME);
4666         class_unregister_type(OBD_MDT_DEVICENAME);
4667 }
4668
4669 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4670 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4671 MODULE_LICENSE("GPL");
4672
4673 module_init(mds_init);
4674 module_exit(mds_exit);