lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *   Author: Mike Shaver <shaver@clusterfs.com>
  12  *
  13  *   This file is part of Lustre, http://www.lustre.org.
  14  *
  15  *   Lustre is free software; you can redistribute it and/or
  16  *   modify it under the terms of version 2 of the GNU General Public
  17  *   License as published by the Free Software Foundation.
  18  *
  19  *   Lustre is distributed in the hope that it will be useful,
  20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   GNU General Public License for more details.
  23  *
  24  *   You should have received a copy of the GNU General Public License
  25  *   along with Lustre; if not, write to the Free Software
  26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/lustre_mds.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/init.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/random.h>
  40 #include <linux/fs.h>
  41 #include <linux/jbd.h>
  42 #include <linux/namei.h>
  43 #include <linux/ext3_fs.h>
  44 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  45 # include <linux/smp_lock.h>
  46 # include <linux/buffer_head.h>
  47 # include <linux/workqueue.h>
  48 # include <linux/mount.h>
  49 #else
  50 # include <linux/locks.h>
  51 #endif
  52 #include <linux/obd_lov.h>
  53 #include <linux/obd_ost.h>
  54 #include <linux/lustre_mds.h>
  55 #include <linux/lustre_fsfilt.h>
  56 #include <linux/lprocfs_status.h>
  57 #include <linux/lustre_commit_confd.h>
  58 #include <linux/lustre_acl.h>
  59 #include "mds_internal.h"
  60
  61 static int mds_intent_policy(struct ldlm_namespace *ns,
  62                              struct ldlm_lock **lockp, void *req_cookie,
  63                              ldlm_mode_t mode, int flags, void *data);
  64 static int mds_postsetup(struct obd_device *obd);
  65 static int mds_cleanup(struct obd_device *obd, int flags);
  66
  67
  68 /* Assumes caller has already pushed into the kernel filesystem context */
  69 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  70                         loff_t offset, int count)
  71 {
  72         struct ptlrpc_bulk_desc *desc;
  73         struct l_wait_info lwi;
  74         struct page **pages;
  75         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  76         ENTRY;
  77
  78         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  79
  80         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  81         OBD_ALLOC(pages, sizeof(*pages) * npages);
  82         if (!pages)
  83                 GOTO(out, rc = -ENOMEM);
  84
  85         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  86                                     MDS_BULK_PORTAL);
  87         if (desc == NULL)
  88                 GOTO(out_free, rc = -ENOMEM);
  89
  90         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  91                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  92
  93                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  94                 if (pages[i] == NULL)
  95                         GOTO(cleanup_buf, rc = -ENOMEM);
  96
  97                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  98         }
  99
 100         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 101                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 102                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 103                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 104                        file->f_dentry->d_inode->i_size);
 105
 106                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 107                                      kmap(pages[i]), tmpsize, &offset);
 108                 kunmap(pages[i]);
 109
 110                 if (rc != tmpsize)
 111                         GOTO(cleanup_buf, rc = -EIO);
 112         }
 113
 114         LASSERT(desc->bd_nob == count);
 115
 116         rc = ptlrpc_start_bulk_transfer(desc);
 117         if (rc)
 118                 GOTO(cleanup_buf, rc);
 119
 120         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 121                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 122                        OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
 123                 GOTO(abort_bulk, rc);
 124         }
 125
 126         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 127         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 128         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 129
 130         if (rc == 0) {
 131                 if (desc->bd_success &&
 132                     desc->bd_nob_transferred == count)
 133                         GOTO(cleanup_buf, rc);
 134
 135                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 136         }
 137
 138         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 139                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 140                   desc->bd_nob_transferred, count,
 141                   req->rq_export->exp_client_uuid.uuid,
 142                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 143
 144         ptlrpc_fail_export(req->rq_export);
 145
 146         EXIT;
 147  abort_bulk:
 148         ptlrpc_abort_bulk (desc);
 149  cleanup_buf:
 150         for (i = 0; i < npages; i++)
 151                 if (pages[i])
 152                         __free_pages(pages[i], 0);
 153
 154         ptlrpc_free_bulk(desc);
 155  out_free:
 156         OBD_FREE(pages, sizeof(*pages) * npages);
 157  out:
 158         return rc;
 159 }
 160
 161 extern char *ldlm_lockname[];
 162
 163 int mds_lock_mode_for_dir(struct obd_device *obd,
 164                           struct dentry *dentry, int mode)
 165 {
 166         int ret_mode = 0, split;
 167
 168         /* any dir access needs couple locks:
 169          * 1) on part of dir we gonna lookup/modify in
 170          * 2) on a whole dir to protect it from concurrent splitting
 171          *    and to flush client's cache for readdir()
 172          * so, for a given mode and dentry this routine decides what
 173          * lock mode to use for lock #2:
 174          * 1) if caller's gonna lookup in dir then we need to protect
 175          *    dir from being splitted only - LCK_CR
 176          * 2) if caller's gonna modify dir then we need to protect
 177          *    dir from being splitted and to flush cache - LCK_CW
 178          * 3) if caller's gonna modify dir and that dir seems ready
 179          *    for splitting then we need to protect it from any
 180          *    type of access (lookup/modify/split) - LCK_EX -bzzz */
 181
 182         split = mds_splitting_expected(obd, dentry);
 183
 184         /*
 185          * it is important to check here only for MDS_NO_SPLITTABLE. The reason
 186          * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
 187          * and another thread will not split it on the quiet. But if we have
 188          * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
 189          * but not now (for current thread) and we should consider that it can
 190          * happen soon and go that branch which can yield LCK_EX to protect from
 191          * possible splitting.
 192          */
 193         if (split == MDS_NO_SPLITTABLE) {
 194                 /*
 195                  * this inode won't be splitted. so we need not to protect from
 196                  * just flush client's cache on modification.
 197                  */
 198                 if (mode == LCK_PW)
 199                         ret_mode = LCK_CW;
 200                 else
 201                         ret_mode = 0;
 202         } else {
 203                 if (mode == LCK_EX) {
 204                         ret_mode = LCK_EX;
 205                 } else if (mode == LCK_PR) {
 206                         ret_mode = LCK_CR;
 207                 } else if (mode == LCK_PW) {
 208                         /*
 209                          * caller gonna modify directory. We use concurrent
 210                          * write lock here to retract client's cache for
 211                          * readdir.
 212                          */
 213                         if (split == MDS_EXPECT_SPLIT) {
 214                                 /*
 215                                  * splitting possible. serialize any access the
 216                                  * idea is that first one seen dir is splittable
 217                                  * is given exclusive lock and split
 218                                  * directory. caller passes lock mode to
 219                                  * mds_try_to_split_dir() and splitting would be
 220                                  * done with exclusive lock only -bzzz.
 221                                  */
 222                                 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
 223                                        obd->obd_name,
 224                                        (unsigned long)dentry->d_inode->i_ino,
 225                                        (unsigned long)dentry->d_inode->i_generation);
 226                                 ret_mode = LCK_EX;
 227                         } else {
 228                                 ret_mode = LCK_CW;
 229                         }
 230                 }
 231         }
 232
 233         return ret_mode;
 234 }
 235
 236 /* only valid locked dentries or errors should be returned */
 237 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
 238                                     struct vfsmount **mnt, int lock_mode,
 239                                     struct lustre_handle *lockh, int *mode,
 240                                     char *name, int namelen, __u64 lockpart)
 241 {
 242         struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
 243         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 244         struct ldlm_res_id res_id = { .name = {0} };
 245         int flags = 0, rc;
 246         ENTRY;
 247
 248         if (IS_ERR(de))
 249                 RETURN(de);
 250
 251         lockh[1].cookie = 0;
 252         res_id.name[0] = id_fid(id);
 253         res_id.name[1] = id_group(id);
 254
 255 #ifdef S_PDIROPS
 256         if (name && IS_PDIROPS(de->d_inode)) {
 257                 ldlm_policy_data_t cpolicy =
 258                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 259                 LASSERT(mode != NULL);
 260                 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
 261                 if (*mode) {
 262                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 263                                               res_id, LDLM_IBITS,
 264                                               &cpolicy, *mode, &flags,
 265                                               mds_blocking_ast,
 266                                               ldlm_completion_ast, NULL, NULL,
 267                                               NULL, 0, NULL, lockh + 1);
 268                         if (rc != ELDLM_OK) {
 269                                 l_dput(de);
 270                                 RETURN(ERR_PTR(-ENOLCK));
 271                         }
 272                 }
 273                 flags = 0;
 274
 275                 res_id.name[2] = full_name_hash(name, namelen);
 276
 277                 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
 278                        OLID4(id), res_id.name[2]);
 279         }
 280 #else
 281 #warning "No PDIROPS support in the kernel"
 282 #endif
 283         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 284                               LDLM_IBITS, &policy, lock_mode, &flags,
 285                               mds_blocking_ast, ldlm_completion_ast,
 286                               NULL, NULL, NULL, 0, NULL, lockh);
 287         if (rc != ELDLM_OK) {
 288                 l_dput(de);
 289                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 290 #ifdef S_PDIROPS
 291                 if (lockh[1].cookie)
 292                         ldlm_lock_decref(lockh + 1, *mode);
 293 #endif
 294         }
 295
 296         RETURN(retval);
 297 }
 298
 299 #ifndef DCACHE_DISCONNECTED
 300 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 301 #endif
 302
 303 /* Look up an entry by inode number. This function ONLY returns valid dget'd
 304  * dentries with an initialized inode or errors */
 305 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
 306                              struct vfsmount **mnt)
 307 {
 308         unsigned long ino = (unsigned long)id_ino(id);
 309         __u32 generation = (__u32)id_gen(id);
 310         struct mds_obd *mds = &obd->u.mds;
 311         struct dentry *result;
 312         struct inode *inode;
 313         char idname[32];
 314
 315         if (ino == 0)
 316                 RETURN(ERR_PTR(-ESTALE));
 317
 318         snprintf(idname, sizeof(idname), "0x%lx", ino);
 319
 320         CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
 321                ino, generation, mds->mds_sb);
 322
 323         /* under ext3 this is neither supposed to return bad inodes nor NULL
 324            inodes. */
 325         result = ll_lookup_one_len(idname, mds->mds_id_de,
 326                                    strlen(idname));
 327         if (IS_ERR(result))
 328                 RETURN(result);
 329
 330         inode = result->d_inode;
 331         if (!inode)
 332                 RETURN(ERR_PTR(-ENOENT));
 333
 334         if (is_bad_inode(inode)) {
 335                 CERROR("bad inode returned %lu/%u\n",
 336                        inode->i_ino, inode->i_generation);
 337                 dput(result);
 338                 RETURN(ERR_PTR(-ENOENT));
 339         }
 340
 341         /* here we disabled generation check, as root inode i_generation
 342          * of cache mds and real mds are different. */
 343         if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
 344             inode->i_generation != generation) {
 345                 /* we didn't find the right inode.. */
 346                 CERROR("bad inode %lu, link: %lu, ct: %d, generation %u/%u\n",
 347                        inode->i_ino, (unsigned long)inode->i_nlink,
 348                        atomic_read(&inode->i_count), inode->i_generation,
 349                        generation);
 350                 dput(result);
 351                 RETURN(ERR_PTR(-ENOENT));
 352         }
 353
 354         if (mnt) {
 355                 *mnt = mds->mds_vfsmnt;
 356                 mntget(*mnt);
 357         }
 358
 359         RETURN(result);
 360 }
 361
 362 static
 363 int mds_req_add_idmapping(struct ptlrpc_request *req,
 364                           struct mds_export_data *med)
 365 {
 366         struct mds_req_sec_desc *rsd;
 367         struct lustre_sec_desc  *lsd;
 368         int rc;
 369
 370         if (!med->med_remote)
 371                 return 0;
 372
 373         /* maybe we should do it more completely: invalidate the gss ctxt? */
 374         if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
 375                 CWARN("didn't find mapped uid\n");
 376                 return -EPERM;
 377         }
 378
 379         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 380         if (!rsd) {
 381                 CERROR("Can't unpack security desc\n");
 382                 return -EPROTO;
 383         }
 384
 385         lsd = mds_get_lsd(req->rq_mapped_uid);
 386         if (!lsd) {
 387                 CERROR("can't get LSD(%u), no mapping added\n",
 388                        req->rq_mapped_uid);
 389                 return -EPERM;
 390         }
 391
 392         rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 393                            rsd->rsd_gid, lsd->lsd_gid);
 394         mds_put_lsd(lsd);
 395         return rc;
 396 }
 397
 398 static
 399 int mds_req_del_idmapping(struct ptlrpc_request *req,
 400                           struct mds_export_data *med)
 401 {
 402         struct mds_req_sec_desc *rsd;
 403         struct lustre_sec_desc  *lsd;
 404         int rc;
 405
 406         if (!med->med_remote)
 407                 return 0;
 408
 409         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 410         if (!rsd) {
 411                 CERROR("Can't unpack security desc\n");
 412                 return -EPROTO;
 413         }
 414
 415         LASSERT(req->rq_mapped_uid != -1);
 416         lsd = mds_get_lsd(req->rq_mapped_uid);
 417         if (!lsd) {
 418                 CERROR("can't get LSD(%u), no idmapping deleted\n",
 419                        req->rq_mapped_uid);
 420                 return -EPERM;
 421         }
 422
 423         rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
 424                            rsd->rsd_gid, lsd->lsd_gid);
 425         mds_put_lsd(lsd);
 426         return rc;
 427 }
 428
 429 static int mds_init_export_data(struct ptlrpc_request *req,
 430                                 struct mds_export_data *med)
 431 {
 432         struct obd_connect_data *data, *reply;
 433         int ask_remote, ask_local;
 434         ENTRY;
 435
 436         data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
 437         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
 438         LASSERT(data && reply);
 439
 440         if (med->med_initialized) {
 441                 CWARN("med already initialized, reconnect?\n");
 442                 goto reply;
 443         }
 444
 445         ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
 446         ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
 447
 448         /* currently the policy is simple: satisfy client as possible
 449          * as we can.
 450          */
 451         if (req->rq_auth_uid == -1) {
 452                 if (ask_remote)
 453                         CWARN("null sec is used, force to be local\n");
 454                 med->med_remote = 0;
 455         } else {
 456                 if (ask_remote) {
 457                         if (!req->rq_remote_realm)
 458                                 CWARN("local realm asked to be remote\n");
 459                         med->med_remote = 1;
 460                 } else if (ask_local) {
 461                         if (req->rq_remote_realm)
 462                                 CWARN("remote realm asked to be local\n");
 463                         med->med_remote = 0;
 464                 } else
 465                         med->med_remote = (req->rq_remote_realm != 0);
 466         }
 467
 468         med->med_nllu = data->ocd_nllu[0];
 469         med->med_nllg = data->ocd_nllu[1];
 470
 471         med->med_initialized = 1;
 472 reply:
 473         reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
 474         if (med->med_remote) {
 475                 if (!med->med_idmap)
 476                         med->med_idmap = mds_idmap_alloc();
 477
 478                 if (!med->med_idmap)
 479                         CERROR("Failed to alloc idmap, following request from "
 480                                "this client will be refused\n");
 481
 482                 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
 483                 CDEBUG(D_SEC, "set client as remote\n");
 484         } else {
 485                 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
 486                 CDEBUG(D_SEC, "set client as local\n");
 487         }
 488
 489         RETURN(0);
 490 }
 491
 492 static void mds_free_export_data(struct mds_export_data *med)
 493 {
 494         if (!med->med_idmap)
 495                 return;
 496
 497         LASSERT(med->med_remote);
 498         mds_idmap_free(med->med_idmap);
 499         med->med_idmap = NULL;
 500 }
 501
 502 /* Establish a connection to the MDS.
 503  *
 504  * This will set up an export structure for the client to hold state data about
 505  * that client, like open files, the last operation number it did on the server,
 506  * etc.
 507  */
 508 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 509                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 510                        unsigned long flags)
 511 {
 512         struct mds_export_data *med;
 513         struct mds_client_data *mcd;
 514         struct obd_export *exp;
 515         int rc;
 516         ENTRY;
 517
 518         if (!conn || !obd || !cluuid)
 519                 RETURN(-EINVAL);
 520
 521         /* XXX There is a small race between checking the list and adding a new
 522          * connection for the same UUID, but the real threat (list corruption
 523          * when multiple different clients connect) is solved.
 524          *
 525          * There is a second race between adding the export to the list, and
 526          * filling in the client data below.  Hence skipping the case of NULL
 527          * mcd above.  We should already be controlling multiple connects at the
 528          * client, and we can't hold the spinlock over memory allocations
 529          * without risk of deadlocking.
 530          */
 531         rc = class_connect(conn, obd, cluuid);
 532         if (rc)
 533                 RETURN(rc);
 534         exp = class_conn2export(conn);
 535
 536         LASSERT(exp != NULL);
 537         med = &exp->exp_mds_data;
 538
 539         OBD_ALLOC(mcd, sizeof(*mcd));
 540         if (!mcd) {
 541                 CERROR("%s: out of memory for client data.\n",
 542                         obd->obd_name);
 543                 GOTO(out, rc = -ENOMEM);
 544         }
 545
 546         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 547         med->med_mcd = mcd;
 548
 549         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 550         if (rc)
 551                 GOTO(out, rc);
 552
 553         EXIT;
 554 out:
 555         if (rc) {
 556                 if (mcd)
 557                         OBD_FREE(mcd, sizeof(*mcd));
 558                 class_disconnect(exp, 0);
 559         } else {
 560                 class_export_put(exp);
 561         }
 562         return rc;
 563 }
 564
 565 static int mds_connect_post(struct obd_export *exp, unsigned long flags)
 566 {
 567         struct obd_device *obd = exp->exp_obd;
 568         struct mds_obd *mds = &obd->u.mds;
 569         int rc = 0;
 570         ENTRY;
 571
 572         if (!(flags & OBD_OPT_MDS_CONNECTION)) {
 573                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 574                         atomic_inc(&mds->mds_real_clients);
 575                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 576                                obd->obd_name, exp->exp_client_uuid.uuid,
 577                                atomic_read(&mds->mds_real_clients));
 578                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 579                 }
 580                 if (mds->mds_md_name)
 581                         rc = mds_md_connect(obd, mds->mds_md_name);
 582         }
 583         RETURN(rc);
 584 }
 585
 586 static int mds_init_export(struct obd_export *exp)
 587 {
 588         struct mds_export_data *med = &exp->exp_mds_data;
 589
 590         INIT_LIST_HEAD(&med->med_open_head);
 591         spin_lock_init(&med->med_open_lock);
 592         return 0;
 593 }
 594
 595 static int mds_destroy_export(struct obd_export *export)
 596 {
 597         struct obd_device *obd = export->exp_obd;
 598         struct mds_export_data *med = &export->exp_mds_data;
 599         struct lvfs_run_ctxt saved;
 600         int rc = 0;
 601         ENTRY;
 602
 603         mds_free_export_data(med);
 604         target_destroy_export(export);
 605
 606         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 607                 GOTO(out, 0);
 608
 609         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 610
 611         /* Close any open files (which may also cause orphan unlinking). */
 612         spin_lock(&med->med_open_lock);
 613         while (!list_empty(&med->med_open_head)) {
 614                 struct list_head *tmp = med->med_open_head.next;
 615                 struct mds_file_data *mfd =
 616                         list_entry(tmp, struct mds_file_data, mfd_list);
 617                 struct lustre_id sid;
 618
 619                 BDEVNAME_DECLARE_STORAGE(btmp);
 620
 621                 /* bug 1579: fix force-closing for 2.5 */
 622                 struct dentry *dentry = mfd->mfd_dentry;
 623
 624                 list_del(&mfd->mfd_list);
 625                 spin_unlock(&med->med_open_lock);
 626
 627                 down(&dentry->d_inode->i_sem);
 628                 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
 629                 up(&dentry->d_inode->i_sem);
 630                 if (rc) {
 631                         CERROR("Can't read inode self id, inode %lu, "
 632                                "rc %d\n", dentry->d_inode->i_ino, rc);
 633                         memset(&sid, 0, sizeof(sid));
 634                 }
 635
 636                 /* If you change this message, be sure to update
 637                  * replay_single:test_46 */
 638                 CERROR("force closing client file handle for %.*s (%s:"
 639                        DLID4")\n", dentry->d_name.len, dentry->d_name.name,
 640                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 641                        OLID4(&sid));
 642
 643                 /* child inode->i_alloc_sem protects orphan_dec_test and
 644                  * is_orphan race, mds_mfd_close drops it */
 645                 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 646                 rc = mds_mfd_close(NULL, 0, obd, mfd,
 647                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 648                 if (rc)
 649                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 650                 spin_lock(&med->med_open_lock);
 651         }
 652         spin_unlock(&med->med_open_lock);
 653         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 654
 655         EXIT;
 656 out:
 657         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 658         return rc;
 659 }
 660
 661 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
 662 {
 663         unsigned long irqflags;
 664         struct obd_device *obd;
 665         struct mds_obd *mds;
 666         int rc;
 667         ENTRY;
 668
 669         LASSERT(exp != NULL);
 670         obd = class_exp2obd(exp);
 671         if (obd == NULL) {
 672                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 673                        exp->exp_handle.h_cookie);
 674                 RETURN(-EINVAL);
 675         }
 676         mds = &obd->u.mds;
 677
 678         /*
 679          * suppress any inter-mds requests durring disconnecting lmv if this is
 680          * detected --force mode. This is needed to avoid endless recovery.
 681          */
 682         if (atomic_read(&mds->mds_real_clients) > 0 &&
 683             !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
 684                 flags |= OBD_OPT_FORCE;
 685
 686         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 687             && !atomic_read(&mds->mds_real_clients)) {
 688                 /* there was no client at all */
 689                 mds_md_disconnect(obd, flags);
 690         }
 691
 692         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 693             && atomic_dec_and_test(&mds->mds_real_clients)) {
 694                 /* time to drop LMV connections */
 695                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 696                        "Disconnnect from LMV now\n",
 697                        obd->obd_name, exp->exp_client_uuid.uuid);
 698                 mds_md_disconnect(obd, flags);
 699         }
 700
 701         spin_lock_irqsave(&exp->exp_lock, irqflags);
 702         exp->exp_flags = flags;
 703         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 704
 705         /* disconnect early so that clients can't keep using export */
 706         rc = class_disconnect(exp, flags);
 707         ldlm_cancel_locks_for_export(exp);
 708
 709         /* complete all outstanding replies */
 710         spin_lock_irqsave(&exp->exp_lock, irqflags);
 711         while (!list_empty(&exp->exp_outstanding_replies)) {
 712                 struct ptlrpc_reply_state *rs =
 713                         list_entry(exp->exp_outstanding_replies.next,
 714                                    struct ptlrpc_reply_state, rs_exp_list);
 715                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 716
 717                 spin_lock(&svc->srv_lock);
 718                 list_del_init(&rs->rs_exp_list);
 719                 ptlrpc_schedule_difficult_reply(rs);
 720                 spin_unlock(&svc->srv_lock);
 721         }
 722         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 723         RETURN(rc);
 724 }
 725
 726 static int mds_getstatus(struct ptlrpc_request *req)
 727 {
 728         struct mds_obd *mds = mds_req2mds(req);
 729         struct mds_body *body;
 730         int rc, size;
 731         ENTRY;
 732
 733         size = sizeof(*body);
 734
 735         rc = lustre_pack_reply(req, 1, &size, NULL);
 736         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 737                 CERROR("mds: out of memory for message: size=%d\n", size);
 738                 req->rq_status = -ENOMEM;       /* superfluous? */
 739                 RETURN(-ENOMEM);
 740         }
 741
 742         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 743         body->valid |= OBD_MD_FID;
 744
 745         memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
 746
 747         /*
 748          * the last_committed and last_xid fields are filled in for all replies
 749          * already - no need to do so here also.
 750          */
 751         RETURN(0);
 752 }
 753
 754 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 755                      void *data, int flag)
 756 {
 757         int do_ast;
 758         ENTRY;
 759
 760         if (flag == LDLM_CB_CANCELING) {
 761                 /* Don't need to do anything here. */
 762                 RETURN(0);
 763         }
 764
 765         /* XXX layering violation!  -phil */
 766         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 767
 768         /*
 769          * get this: if mds_blocking_ast is racing with mds_intent_policy, such
 770          * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
 771          * then by the time we get the lock, we might not be the correct
 772          * blocking function anymore.  So check, and return early, if so.
 773          */
 774         if (lock->l_blocking_ast != mds_blocking_ast) {
 775                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 776                 RETURN(0);
 777         }
 778
 779         lock->l_flags |= LDLM_FL_CBPENDING;
 780         do_ast = (!lock->l_readers && !lock->l_writers);
 781         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 782
 783         if (do_ast) {
 784                 struct lustre_handle lockh;
 785                 int rc;
 786
 787                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 788                 ldlm_lock2handle(lock, &lockh);
 789                 rc = ldlm_cli_cancel(&lockh);
 790                 if (rc < 0)
 791                         CERROR("ldlm_cli_cancel: %d\n", rc);
 792         } else {
 793                 LDLM_DEBUG(lock, "Lock still has references, will be "
 794                            "cancelled later");
 795         }
 796         RETURN(0);
 797 }
 798
 799 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
 800                           void *md, int size, int mea)
 801 {
 802         int rc = size;
 803
 804         if (S_ISREG(inode->i_mode)) {
 805                 rc = mds_convert_lov_ea(obd, inode, md, size);
 806         } else if (S_ISDIR(inode->i_mode)) {
 807                 if (mea) {
 808                         rc = mds_convert_mea_ea(obd, inode, md, size);
 809                 } else {
 810                         rc = mds_convert_lov_ea(obd, inode, md, size);
 811                 }
 812                 if (rc == -EINVAL) {
 813                         CERROR("Invalid EA format (nor LOV or MEA) "
 814                                "is detected. Inode %lu/%u\n",
 815                                inode->i_ino, inode->i_generation);
 816                 }
 817         }
 818         return rc;
 819 }
 820
 821 int mds_get_md(struct obd_device *obd, struct inode *inode,
 822                void *md, int *size, int lock, int mea)
 823 {
 824         int lmm_size;
 825         int rc = 0;
 826         ENTRY;
 827
 828         if (lock)
 829                 down(&inode->i_sem);
 830
 831         rc = fsfilt_get_md(obd, inode, md, *size,
 832                            (mea ? EA_MEA : EA_LOV));
 833         if (rc < 0) {
 834                 CERROR("Error %d reading eadata for ino %lu\n",
 835                        rc, inode->i_ino);
 836         } else if (rc > 0) {
 837                 lmm_size = rc;
 838                 rc = mds_convert_md(obd, inode, md,
 839                                     lmm_size, mea);
 840                 if (rc == 0) {
 841                         *size = lmm_size;
 842                         rc = lmm_size;
 843                 } else if (rc > 0) {
 844                         *size = rc;
 845                 }
 846         }
 847         if (lock)
 848                 up(&inode->i_sem);
 849
 850         RETURN(rc);
 851 }
 852
 853
 854 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 855  * Call with lock=0 if the caller has already taken the i_sem. */
 856 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 857                 struct mds_body *body, struct inode *inode, int lock, int mea)
 858 {
 859         struct mds_obd *mds = &obd->u.mds;
 860         void *lmm;
 861         int lmm_size;
 862         int rc;
 863         ENTRY;
 864
 865         lmm = lustre_msg_buf(msg, offset, 0);
 866         if (lmm == NULL) {
 867                 /* Some problem with getting eadata when I sized the reply
 868                  * buffer... */
 869                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 870                        inode->i_ino);
 871                 RETURN(0);
 872         }
 873         lmm_size = msg->buflens[offset];
 874
 875         /* I don't really like this, but it is a sanity check on the client
 876          * MD request.  However, if the client doesn't know how much space
 877          * to reserve for the MD, it shouldn't be bad to have too much space.
 878          */
 879         if (lmm_size > mds->mds_max_mdsize) {
 880                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 881                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 882                 // RETURN(-EINVAL);
 883         }
 884
 885         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
 886         if (rc > 0) {
 887                 if (S_ISDIR(inode->i_mode))
 888                         body->valid |= OBD_MD_FLDIREA;
 889                 else
 890                         body->valid |= OBD_MD_FLEASIZE;
 891
 892                 if (mea)
 893                         body->valid |= OBD_MD_MEA;
 894
 895                 body->eadatasize = lmm_size;
 896                 rc = 0;
 897         }
 898
 899         RETURN(rc);
 900 }
 901 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
 902                   struct mds_body *repbody, int reply_off)
 903 {
 904         struct inode *inode = dentry->d_inode;
 905         char *symname;
 906         int len, rc;
 907         ENTRY;
 908
 909         symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 910         LASSERT(symname != NULL);
 911         len = req->rq_repmsg->buflens[reply_off + 1];
 912
 913         rc = inode->i_op->readlink(dentry, symname, len);
 914         if (rc < 0) {
 915                 CERROR("readlink failed: %d\n", rc);
 916         } else if (rc != len - 1) {
 917                 CERROR ("Unexpected readlink rc %d: expecting %d\n",
 918                         rc, len - 1);
 919                 rc = -EINVAL;
 920         } else {
 921                 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 922                 repbody->valid |= OBD_MD_LINKNAME;
 923                 repbody->eadatasize = rc + 1;
 924                 symname[rc] = 0;        /* NULL terminate */
 925                 rc = 0;
 926         }
 927
 928         RETURN(rc);
 929 }
 930
 931 int mds_pack_ea(struct dentry *dentry, struct ptlrpc_request *req,
 932                 struct mds_body *repbody, int req_off, int reply_off)
 933 {
 934         struct inode *inode = dentry->d_inode;
 935         char *ea_name;
 936         void *value = NULL;
 937         int len, rc;
 938         ENTRY;
 939
 940         ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
 941         len = req->rq_repmsg->buflens[reply_off + 1];
 942         if (len != 0)
 943                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 944
 945         rc = -EOPNOTSUPP;
 946         if (inode->i_op && inode->i_op->getxattr)
 947                 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
 948
 949         if (rc < 0) {
 950                 if (rc != -ENODATA && rc != -EOPNOTSUPP)
 951                         CERROR("getxattr failed: %d", rc);
 952         } else {
 953                 repbody->valid |= OBD_MD_FLEA;
 954                 repbody->eadatasize = rc;
 955                 rc = 0;
 956         }
 957
 958         RETURN(rc);
 959 }
 960
 961 int mds_pack_ealist(struct dentry *dentry, struct ptlrpc_request *req,
 962                     struct mds_body *repbody, int reply_off)
 963 {
 964         struct inode *inode = dentry->d_inode;
 965         void *value = NULL;
 966         int len, rc;
 967         ENTRY;
 968
 969         len = req->rq_repmsg->buflens[reply_off + 1];
 970         if (len != 0)
 971                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 972
 973         rc = -EOPNOTSUPP;
 974         if (inode->i_op && inode->i_op->getxattr)
 975                 rc = inode->i_op->listxattr(dentry, value, len);
 976
 977         if (rc < 0) {
 978                 CERROR("listxattr failed: %d", rc);
 979         } else {
 980                 repbody->valid |= OBD_MD_FLEALIST;
 981                 repbody->eadatasize = rc;
 982                 rc = 0;
 983         }
 984         RETURN(rc);
 985 }
 986
 987 int mds_pack_acl(struct obd_device *obd, struct lustre_msg *repmsg, int offset,
 988                  struct mds_body *body, struct inode *inode)
 989 {
 990         struct dentry de = { .d_inode = inode };
 991         __u32 buflen, *sizep;
 992         void *buf;
 993         int size;
 994         ENTRY;
 995
 996         if (!inode->i_op->getxattr)
 997                 RETURN(0);
 998
 999         buflen = repmsg->buflens[offset + 1];
1000         buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1001
1002         size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1003         if (size == -ENODATA || size == -EOPNOTSUPP)
1004                 RETURN(0);
1005         if (size < 0)
1006                 RETURN(size);
1007         LASSERT(size);
1008
1009         sizep = lustre_msg_buf(repmsg, offset, 4);
1010         if (!sizep) {
1011                 CERROR("can't locate returned acl size buf\n");
1012                 RETURN(-EPROTO);
1013         }
1014
1015         *sizep = cpu_to_le32(size);
1016         body->valid |= OBD_MD_FLACL_ACCESS;
1017
1018         RETURN(0);
1019 }
1020
1021 /*
1022  * here we take simple rule: once uid/fsuid is root, we also squash
1023  * the gid/fsgid, don't care setuid/setgid attributes.
1024  */
1025 int mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
1026                     ptl_nid_t *peernid)
1027 {
1028         if (!mds->mds_squash_uid || *peernid == mds->mds_nosquash_nid)
1029                 return 0;
1030
1031         if (rsd->rsd_uid && rsd->rsd_fsuid)
1032                 return 0;
1033
1034         CDEBUG(D_SEC, "squash req from "LPX64":"
1035                "(%u:%u-%u:%u/%x)=>(%u:%u-%u:%u/%x)\n", *peernid,
1036                 rsd->rsd_uid, rsd->rsd_gid,
1037                 rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
1038                 rsd->rsd_uid ? rsd->rsd_uid : mds->mds_squash_uid,
1039                 rsd->rsd_uid ? rsd->rsd_gid : mds->mds_squash_gid,
1040                 rsd->rsd_fsuid ? rsd->rsd_fsuid : mds->mds_squash_uid,
1041                 rsd->rsd_fsuid ? rsd->rsd_fsgid : mds->mds_squash_gid,
1042                 rsd->rsd_cap & ~CAP_FS_MASK);
1043
1044         if (rsd->rsd_uid == 0) {
1045                 rsd->rsd_uid = mds->mds_squash_uid;
1046                 rsd->rsd_gid = mds->mds_squash_gid;
1047         }
1048         if (rsd->rsd_fsuid == 0) {
1049                 rsd->rsd_fsuid = mds->mds_squash_uid;
1050                 rsd->rsd_fsgid = mds->mds_squash_gid;
1051         }
1052         rsd->rsd_cap &= ~CAP_FS_MASK;
1053
1054         return 1;
1055 }
1056
1057 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1058                                 struct ptlrpc_request *req, int req_off,
1059                                 struct mds_body *reqbody, int reply_off)
1060 {
1061         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1062         struct inode *inode = dentry->d_inode;
1063         struct mds_body *body;
1064         int rc = 0;
1065         ENTRY;
1066
1067         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1068                 RETURN(-ENOENT);
1069
1070         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1071         LASSERT(body != NULL);                 /* caller prepped reply */
1072
1073         if (dentry->d_flags & DCACHE_CROSS_REF) {
1074                 mds_pack_dentry2body(obd, body, dentry,
1075                                      (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1076                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1077                        OLID4(&body->id1));
1078                 RETURN(0);
1079         }
1080
1081         mds_pack_inode2body(obd, body, inode,
1082                             (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1083
1084         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1085             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1086                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
1087                                  inode, 1, (reqbody->valid & OBD_MD_MEA) ? 1 : 0);
1088
1089                 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1090                 if (!(body->valid & OBD_MD_FLEASIZE) &&
1091                     !(body->valid & OBD_MD_FLDIREA))
1092                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1093                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1094         } else if (S_ISLNK(inode->i_mode) &&
1095                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1096                 rc = mds_pack_link(dentry, req, body, reply_off);
1097         } else if (reqbody->valid & OBD_MD_FLEA) {
1098                 rc = mds_pack_ea(dentry, req, body, req_off, reply_off);
1099         } else if (reqbody->valid & OBD_MD_FLEALIST) {
1100                 rc = mds_pack_ealist(dentry, req, body, reply_off);
1101         }
1102
1103         if (reqbody->valid & OBD_MD_FLACL_ACCESS) {
1104                 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1105                 rc = mds_pack_acl(obd, req->rq_repmsg, reply_off + inc,
1106                                   body, inode);
1107         }
1108
1109         if (rc == 0)
1110                 mds_body_do_reverse_map(med, body);
1111
1112         RETURN(rc);
1113 }
1114
1115 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1116                                    struct dentry *dentry,
1117                                    int offset)
1118 {
1119         int rc = 0, size[1] = {sizeof(struct mds_body)};
1120         ENTRY;
1121
1122         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1123                 CERROR("failed MDS_GETATTR_PACK test\n");
1124                 req->rq_status = -ENOMEM;
1125                 RETURN(-ENOMEM);
1126         }
1127
1128         rc = lustre_pack_reply(req, 1, size, NULL);
1129         if (rc) {
1130                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1131                 GOTO(out, req->rq_status = rc);
1132         }
1133
1134         EXIT;
1135 out:
1136         return rc;
1137 }
1138
1139 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1140                                 int offset)
1141 {
1142         struct inode *inode = de->d_inode;
1143         struct mds_obd *mds = mds_req2mds(req);
1144         struct mds_body *body;
1145         int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1146         ENTRY;
1147
1148         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1149         LASSERT(body != NULL);                 /* checked by caller */
1150         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
1151
1152         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1153             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1154                 int rc;
1155
1156                 down(&inode->i_sem);
1157                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1158                                    ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1159                 up(&inode->i_sem);
1160                 if (rc < 0) {
1161                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1162                                 CERROR("error getting inode %lu MD: rc = %d\n",
1163                                        inode->i_ino, rc);
1164                         size[bufcount] = 0;
1165                 } else if (rc > mds->mds_max_mdsize) {
1166                         size[bufcount] = 0;
1167                         CERROR("MD size %d larger than maximum possible %u\n",
1168                                rc, mds->mds_max_mdsize);
1169                 } else {
1170                         size[bufcount] = rc;
1171                 }
1172                 bufcount++;
1173         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1174                 if (inode->i_size + 1 != body->eadatasize)
1175                         CERROR("symlink size: %Lu, reply space: %d\n",
1176                                inode->i_size + 1, body->eadatasize);
1177                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1178                 bufcount++;
1179                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1180                        inode->i_size + 1, body->eadatasize);
1181         } else if ((body->valid & OBD_MD_FLEA)) {
1182                 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1183                                                   offset + 1, 0);
1184                 rc = -EOPNOTSUPP;
1185                 if (inode->i_op && inode->i_op->getxattr)
1186                         rc = inode->i_op->getxattr(de, ea_name, NULL, 0);
1187
1188                 if (rc < 0) {
1189                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1190                                 CERROR("error getting inode %lu EA: rc = %d\n",
1191                                        inode->i_ino, rc);
1192                         size[bufcount] = 0;
1193                 } else {
1194                         size[bufcount] = min_t(int, body->eadatasize, rc);
1195                 }
1196                 bufcount++;
1197         } else if (body->valid & OBD_MD_FLEALIST) {
1198                 rc = -EOPNOTSUPP;
1199                 if (inode->i_op && inode->i_op->getxattr)
1200                         rc = inode->i_op->listxattr(de, NULL, 0);
1201
1202                 if (rc < 0) {
1203                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1204                                 CERROR("error getting inode %lu EA: rc = %d\n",
1205                                        inode->i_ino, rc);
1206                         size[bufcount] = 0;
1207                 } else {
1208                         size[bufcount] = min_t(int, body->eadatasize, rc);
1209                 }
1210                 bufcount++;
1211         }
1212
1213         /* may co-exist with OBD_MD_FLEASIZE */
1214         if (body->valid & OBD_MD_FLACL_ACCESS) {
1215                 size[bufcount++] = 4;
1216                 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1217         }
1218
1219         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1220                 CERROR("failed MDS_GETATTR_PACK test\n");
1221                 req->rq_status = -ENOMEM;
1222                 GOTO(out, rc = -ENOMEM);
1223         }
1224
1225         rc = lustre_pack_reply(req, bufcount, size, NULL);
1226         if (rc) {
1227                 CERROR("out of memory\n");
1228                 GOTO(out, req->rq_status = rc);
1229         }
1230
1231         EXIT;
1232  out:
1233         return rc;
1234 }
1235
1236 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1237                       char *name, int namelen)
1238 {
1239         struct mea *mea = NULL;
1240         int mea_size, rc = 0;
1241         ENTRY;
1242
1243         rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1244         if (rc)
1245                 RETURN(rc);
1246         if (mea != NULL) {
1247                 /*
1248                  * dir is already splitted, check if requested filename should
1249                  * live at this MDS or at another one.
1250                  */
1251                 int i = mea_name2idx(mea, name, namelen - 1);
1252                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1253                         CDEBUG(D_OTHER,
1254                                "inapropriate MDS(%d) for %s. should be "
1255                                "%lu(%d)\n", mea->mea_master, name,
1256                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1257                         rc = -ERESTART;
1258                 }
1259         }
1260
1261         if (mea)
1262                 OBD_FREE(mea, mea_size);
1263         RETURN(rc);
1264 }
1265
1266 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1267                             struct lustre_handle *child_lockh, int child_part)
1268 {
1269         struct obd_device *obd = req->rq_export->exp_obd;
1270         struct mds_obd *mds = &obd->u.mds;
1271         struct ldlm_reply *rep = NULL;
1272         struct lvfs_run_ctxt saved;
1273         struct mds_req_sec_desc *rsd;
1274         struct mds_body *body;
1275         struct dentry *dparent = NULL, *dchild = NULL;
1276         struct lvfs_ucred uc = {NULL, NULL,};
1277         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1278         unsigned int namesize;
1279         int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1280         char *name = NULL;
1281         ENTRY;
1282
1283         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
1284         MD_COUNTER_INCREMENT(obd, getattr_lock);
1285
1286         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1287         if (!rsd) {
1288                 CERROR("Can't unpack security desc\n");
1289                 RETURN(-EFAULT);
1290         }
1291
1292         /* swab now, before anyone looks inside the request. */
1293         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1294                                   lustre_swab_mds_body);
1295         if (body == NULL) {
1296                 CERROR("Can't swab mds_body\n");
1297                 GOTO(cleanup, rc = -EFAULT);
1298         }
1299
1300         LASSERT_REQSWAB(req, offset + 1);
1301         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1302         if (name == NULL) {
1303                 CERROR("Can't unpack name\n");
1304                 GOTO(cleanup, rc = -EFAULT);
1305         }
1306         namesize = req->rq_reqmsg->buflens[offset + 1];
1307
1308         /* namesize less than 2 means we have empty name, probably came from
1309            revalidate by cfid, so no point in having name to be set */
1310         if (namesize <= 1)
1311                 name = NULL;
1312
1313         LASSERT (offset == 1 || offset == 3);
1314         if (offset == 3) {
1315                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1316                 reply_offset = 1;
1317         } else {
1318                 reply_offset = 0;
1319         }
1320
1321         rc = mds_init_ucred(&uc, req, rsd);
1322         if (rc) {
1323                 GOTO(cleanup, rc);
1324         }
1325
1326         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1327         cleanup_phase = 1; /* kernel context */
1328         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1329
1330         LASSERT(namesize > 0);
1331         if (child_lockh->cookie != 0) {
1332                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1333                 resent_req = 1;
1334         }
1335 #if 0
1336 #if HAVE_LOOKUP_RAW
1337         if (body->valid == OBD_MD_FLID) {
1338                 struct mds_body *mds_reply;
1339                 int size = sizeof(*mds_reply);
1340                 struct inode *dir;
1341                 ino_t inum;
1342
1343                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1344                 if (IS_ERR(dparent)) {
1345                         rc = PTR_ERR(dparent);
1346                         GOTO(cleanup, rc);
1347                 }
1348
1349                 /*
1350                  * the user requested ONLY the inode number, so do a raw lookup.
1351                  */
1352                 rc = lustre_pack_reply(req, 1, &size, NULL);
1353                 if (rc) {
1354                         CERROR("out of memory\n");
1355                         l_dput(dparent);
1356                         GOTO(cleanup, rc);
1357                 }
1358                 dir  = dparent->d_inode;
1359                 LASSERT(dir->i_op->lookup_raw != NULL);
1360                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1361                 l_dput(dparent);
1362                 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1363                                            sizeof(*mds_reply));
1364
1365                 id_ino(&mds_reply->id1) = inum;
1366                 mds_reply->valid = OBD_MD_FLID;
1367                 GOTO(cleanup, rc);
1368         }
1369 #endif
1370 #endif
1371         if (resent_req == 0) {
1372                 LASSERT(id_fid(&body->id1) != 0);
1373                 if (name) {
1374                         rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1375                                                          parent_lockh, &dparent,
1376                                                          LCK_PR,
1377                                                          MDS_INODELOCK_UPDATE,
1378                                                          &update_mode,
1379                                                          name, namesize,
1380                                                          child_lockh, &dchild,
1381                                                          LCK_PR, child_part);
1382                         if (rc)
1383                                 GOTO(cleanup, rc);
1384
1385                         cleanup_phase = 2; /* dchild, dparent, locks */
1386
1387                         /*
1388                          * let's make sure this name should leave on this mds
1389                          * node.
1390                          */
1391                         rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1392                         if (rc)
1393                                 GOTO(cleanup, rc);
1394                 } else {
1395                         /* we have no dentry here, drop LOOKUP bit */
1396                         /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1397                         child_part &= ~MDS_INODELOCK_LOOKUP;
1398                         CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1399                                obd->obd_name, OLID4(&body->id1));
1400
1401                         dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1402                                                       LCK_PR, parent_lockh,
1403                                                       &update_mode,
1404                                                       NULL, 0,
1405                                                       MDS_INODELOCK_UPDATE);
1406                         if (IS_ERR(dchild)) {
1407                                 CERROR("can't find inode with id "DLID4", err = %d\n",
1408                                        OLID4(&body->id1), (int)PTR_ERR(dchild));
1409                                 GOTO(cleanup, rc = PTR_ERR(dchild));
1410                         }
1411                         memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1412                 }
1413         } else {
1414                 struct ldlm_lock *granted_lock;
1415
1416                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1417                 granted_lock = ldlm_handle2lock(child_lockh);
1418
1419                 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1420                          id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1421                          child_lockh->cookie);
1422
1423                 if (name) {
1424                         /* usual named request */
1425                         dparent = mds_id2dentry(obd, &body->id1, NULL);
1426                         LASSERT(!IS_ERR(dparent));
1427                         dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1428                         if (IS_ERR(dchild)) {
1429                                 DEBUG_REQ(D_ERROR, req, "resent, not enqueuing new locks");
1430                                 CDEBUG(D_ERROR, "lock against [%lu:%lu]/%*s\n",
1431                                                 (unsigned long) id_ino(&body->id1),
1432                                                 (unsigned long) id_gen(&body->id1),
1433                                                 namesize - 1, name);
1434                         }
1435                         LASSERT(!IS_ERR(dchild));
1436                 } else {
1437                         /* client wants to get attr. by id */
1438                         dchild = mds_id2dentry(obd, &body->id1, NULL);
1439                         if (IS_ERR(dchild)) {
1440                                 DEBUG_REQ(D_ERROR, req, "resent, not enqueuing new locks");
1441                                 CDEBUG(D_ERROR, "lock against [%lu:%lu]\n",
1442                                                 (unsigned long) id_ino(&body->id1),
1443                                                 (unsigned long) id_gen(&body->id1));
1444                         }
1445                         LASSERT(!IS_ERR(dchild));
1446                 }
1447                 LDLM_LOCK_PUT(granted_lock);
1448         }
1449
1450         cleanup_phase = 2; /* dchild, dparent, locks */
1451
1452         if (!DENTRY_VALID(dchild)) {
1453                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1454                 /*
1455                  * in the intent case, the policy clears this error: the
1456                  * disposition is enough.
1457                  */
1458                 rc = -ENOENT;
1459                 GOTO(cleanup, rc);
1460         } else {
1461                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1462         }
1463
1464         if (req->rq_repmsg == NULL) {
1465                 if (dchild->d_flags & DCACHE_CROSS_REF)
1466                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1467                 else
1468                         rc = mds_getattr_pack_msg(req, dchild, offset);
1469                 if (rc != 0) {
1470                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1471                         GOTO (cleanup, rc);
1472                 }
1473         }
1474
1475         rc = mds_getattr_internal(obd, dchild, req, offset, body, reply_offset);
1476         GOTO(cleanup, rc); /* returns the lock to the client */
1477
1478  cleanup:
1479         switch (cleanup_phase) {
1480         case 2:
1481                 if (resent_req == 0) {
1482                         if (rc && DENTRY_VALID(dchild))
1483                                 ldlm_lock_decref(child_lockh, LCK_PR);
1484                         if (name)
1485                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1486 #ifdef S_PDIROPS
1487                         if (parent_lockh[1].cookie != 0)
1488                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1489 #endif
1490                         if (dparent)
1491                                 l_dput(dparent);
1492                 }
1493                 l_dput(dchild);
1494         case 1:
1495                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1496         default:
1497                 mds_exit_ucred(&uc);
1498         }
1499         return rc;
1500 }
1501
1502 static int mds_getattr(struct ptlrpc_request *req, int offset)
1503 {
1504         struct obd_device *obd = req->rq_export->exp_obd;
1505         struct lvfs_run_ctxt saved;
1506         struct dentry *de;
1507         struct mds_req_sec_desc *rsd;
1508         struct mds_body *body;
1509         struct lvfs_ucred uc = {NULL, NULL,};
1510         int rc = 0;
1511         ENTRY;
1512
1513         MD_COUNTER_INCREMENT(obd, getattr);
1514
1515         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1516         if (!rsd) {
1517                 CERROR("Can't unpack security desc\n");
1518                 RETURN(-EFAULT);
1519         }
1520
1521         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1522                                   lustre_swab_mds_body);
1523         if (body == NULL) {
1524                 CERROR ("Can't unpack body\n");
1525                 RETURN (-EFAULT);
1526         }
1527
1528         rc = mds_init_ucred(&uc, req, rsd);
1529         if (rc) {
1530                 mds_exit_ucred(&uc);
1531                 RETURN(rc);
1532         }
1533
1534         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1535         de = mds_id2dentry(obd, &body->id1, NULL);
1536         if (IS_ERR(de)) {
1537                 rc = req->rq_status = PTR_ERR(de);
1538                 GOTO(out_pop, rc);
1539         }
1540
1541         rc = mds_getattr_pack_msg(req, de, offset);
1542         if (rc != 0) {
1543                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1544                 GOTO(out_pop, rc);
1545         }
1546
1547         req->rq_status = mds_getattr_internal(obd, de, req, offset, body, 0);
1548         l_dput(de);
1549
1550         EXIT;
1551 out_pop:
1552         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1553         mds_exit_ucred(&uc);
1554         return rc;
1555 }
1556
1557 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1558                           unsigned long max_age)
1559 {
1560         int rc;
1561         ENTRY;
1562
1563         spin_lock(&obd->obd_osfs_lock);
1564         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1565         if (rc == 0)
1566                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1567         spin_unlock(&obd->obd_osfs_lock);
1568
1569         RETURN(rc);
1570 }
1571
1572 static int mds_statfs(struct ptlrpc_request *req)
1573 {
1574         struct obd_device *obd = req->rq_export->exp_obd;
1575         int rc, size = sizeof(struct obd_statfs);
1576         ENTRY;
1577
1578         /* This will trigger a watchdog timeout */
1579         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1580                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1581
1582         rc = lustre_pack_reply(req, 1, &size, NULL);
1583         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1584                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1585                 GOTO(out, rc);
1586         }
1587
1588         OBD_COUNTER_INCREMENT(obd, statfs);
1589
1590         /* We call this so that we can cache a bit - 1 jiffie worth */
1591         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1592                             jiffies - HZ);
1593         if (rc) {
1594                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1595                 GOTO(out, rc);
1596         }
1597
1598         EXIT;
1599 out:
1600         req->rq_status = rc;
1601         return rc;
1602 }
1603
1604 static int mds_sync(struct ptlrpc_request *req, int offset)
1605 {
1606         struct obd_device *obd = req->rq_export->exp_obd;
1607         struct mds_obd *mds = &obd->u.mds;
1608         struct mds_body *body;
1609         int rc, size = sizeof(*body);
1610         ENTRY;
1611
1612         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1613                                   lustre_swab_mds_body);
1614         if (body == NULL)
1615                 GOTO(out, rc = -EPROTO);
1616
1617         rc = lustre_pack_reply(req, 1, &size, NULL);
1618         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1619                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1620                 GOTO(out, rc);
1621         }
1622
1623         if (id_ino(&body->id1) == 0) {
1624                 /* an id of zero is taken to mean "sync whole filesystem" */
1625                 rc = fsfilt_sync(obd, mds->mds_sb);
1626                 if (rc)
1627                         GOTO(out, rc);
1628         } else {
1629                 /* just any file to grab fsync method - "file" arg unused */
1630                 struct file *file = mds->mds_rcvd_filp;
1631                 struct mds_body *rep_body;
1632                 struct dentry *de;
1633
1634                 de = mds_id2dentry(obd, &body->id1, NULL);
1635                 if (IS_ERR(de))
1636                         GOTO(out, rc = PTR_ERR(de));
1637
1638                 rc = file->f_op->fsync(NULL, de, 1);
1639                 if (rc)
1640                         GOTO(out, rc);
1641
1642                 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1643                 mds_pack_inode2body(obd, rep_body, de->d_inode,
1644                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1645                 l_dput(de);
1646         }
1647
1648         EXIT;
1649 out:
1650         req->rq_status = rc;
1651         return rc;
1652 }
1653
1654 /* mds_readpage does not take a DLM lock on the inode, because the client must
1655  * already have a PR lock.
1656  *
1657  * If we were to take another one here, a deadlock will result, if another
1658  * thread is already waiting for a PW lock. */
1659 static int mds_readpage(struct ptlrpc_request *req, int offset)
1660 {
1661         struct obd_device *obd = req->rq_export->exp_obd;
1662         struct vfsmount *mnt;
1663         struct dentry *de;
1664         struct file *file;
1665         struct mds_req_sec_desc *rsd;
1666         struct mds_body *body, *repbody;
1667         struct lvfs_run_ctxt saved;
1668         int rc, size = sizeof(*repbody);
1669         struct lvfs_ucred uc = {NULL, NULL,};
1670         ENTRY;
1671
1672         rc = lustre_pack_reply(req, 1, &size, NULL);
1673         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1674                 CERROR("mds: out of memory\n");
1675                 GOTO(out, rc = -ENOMEM);
1676         }
1677
1678         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1679         if (!rsd) {
1680                 CERROR("Can't unpack security desc\n");
1681                 GOTO (out, rc = -EFAULT);
1682         }
1683
1684         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1685                                   lustre_swab_mds_body);
1686         if (body == NULL) {
1687                 CERROR("Can't unpack body\n");
1688                 GOTO (out, rc = -EFAULT);
1689         }
1690
1691         rc = mds_init_ucred(&uc, req, rsd);
1692         if (rc) {
1693                 GOTO(out, rc);
1694         }
1695
1696         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1697         de = mds_id2dentry(obd, &body->id1, &mnt);
1698         if (IS_ERR(de))
1699                 GOTO(out_pop, rc = PTR_ERR(de));
1700
1701         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1702
1703         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1704         /* note: in case of an error, dentry_open puts dentry */
1705         if (IS_ERR(file))
1706                 GOTO(out_pop, rc = PTR_ERR(file));
1707
1708         /* body->size is actually the offset -eeb */
1709         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1710                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1711                        body->size, de->d_inode->i_blksize);
1712                 GOTO(out_file, rc = -EFAULT);
1713         }
1714
1715         /* body->nlink is actually the #bytes to read -eeb */
1716         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1717                 CERROR("size %u is not multiple of blocksize %lu\n",
1718                        body->nlink, de->d_inode->i_blksize);
1719                 GOTO(out_file, rc = -EFAULT);
1720         }
1721
1722         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1723         repbody->size = file->f_dentry->d_inode->i_size;
1724         repbody->valid = OBD_MD_FLSIZE;
1725
1726         /* to make this asynchronous make sure that the handling function
1727            doesn't send a reply when this function completes. Instead a
1728            callback function would send the reply */
1729         /* body->size is actually the offset -eeb */
1730         rc = mds_sendpage(req, file, body->size, body->nlink);
1731
1732         EXIT;
1733 out_file:
1734         filp_close(file, 0);
1735 out_pop:
1736         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1737 out:
1738         mds_exit_ucred(&uc);
1739         req->rq_status = rc;
1740         return 0;
1741 }
1742
1743 /* update master MDS ID, which is stored in local inode EA. */
1744 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1745                    void *data, int data_len)
1746 {
1747         struct mds_obd *mds = &obd->u.mds;
1748         struct dentry *dentry;
1749         void *handle;
1750         int rc = 0;
1751         ENTRY;
1752
1753         LASSERT(id);
1754         LASSERT(obd);
1755
1756         dentry = mds_id2dentry(obd, id, NULL);
1757         if (IS_ERR(dentry))
1758                 GOTO(out, rc = PTR_ERR(dentry));
1759
1760         if (!dentry->d_inode) {
1761                 CERROR("Can't find object "DLID4".\n",
1762                        OLID4(id));
1763                 GOTO(out_dentry, rc = -EINVAL);
1764         }
1765
1766         handle = fsfilt_start(obd, dentry->d_inode,
1767                               FSFILT_OP_SETATTR, NULL);
1768         if (IS_ERR(handle))
1769                 GOTO(out_dentry, rc = PTR_ERR(handle));
1770
1771         rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1772                                   (struct lustre_id *)data);
1773         if (rc) {
1774                 CERROR("Can't update inode "DLID4" master id, "
1775                        "error = %d.\n", OLID4(id), rc);
1776                 GOTO(out_commit, rc);
1777         }
1778
1779         EXIT;
1780 out_commit:
1781         fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1782                       handle, 0);
1783 out_dentry:
1784         l_dput(dentry);
1785 out:
1786         return rc;
1787 }
1788 EXPORT_SYMBOL(mds_update_mid);
1789
1790 /* read master MDS ID, which is stored in local inode EA. */
1791 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1792                  void *data, int data_len)
1793 {
1794         struct dentry *dentry;
1795         int rc = 0;
1796         ENTRY;
1797
1798         LASSERT(id);
1799         LASSERT(obd);
1800
1801         dentry = mds_id2dentry(obd, id, NULL);
1802         if (IS_ERR(dentry))
1803                 GOTO(out, rc = PTR_ERR(dentry));
1804
1805         if (!dentry->d_inode) {
1806                 CERROR("Can't find object "DLID4".\n",
1807                        OLID4(id));
1808                 GOTO(out_dentry, rc = -EINVAL);
1809         }
1810
1811         down(&dentry->d_inode->i_sem);
1812         rc = mds_read_inode_mid(obd, dentry->d_inode,
1813                                 (struct lustre_id *)data);
1814         up(&dentry->d_inode->i_sem);
1815         if (rc) {
1816                 CERROR("Can't read inode "DLID4" master id, "
1817                        "error = %d.\n", OLID4(id), rc);
1818                 GOTO(out_dentry, rc);
1819         }
1820
1821         EXIT;
1822 out_dentry:
1823         l_dput(dentry);
1824 out:
1825         return rc;
1826 }
1827 EXPORT_SYMBOL(mds_read_mid);
1828
1829 int mds_reint(struct ptlrpc_request *req, int offset,
1830               struct lustre_handle *lockh)
1831 {
1832         struct mds_update_record *rec;
1833         struct mds_req_sec_desc *rsd;
1834         int rc;
1835         ENTRY;
1836
1837         OBD_ALLOC(rec, sizeof(*rec));
1838         if (rec == NULL)
1839                 RETURN(-ENOMEM);
1840
1841         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1842         if (!rsd) {
1843                 CERROR("Can't unpack security desc\n");
1844                 GOTO(out, rc = -EFAULT);
1845         }
1846
1847         rc = mds_update_unpack(req, offset, rec);
1848         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1849                 CERROR("invalid record\n");
1850                 GOTO(out, req->rq_status = -EINVAL);
1851         }
1852
1853         rc = mds_init_ucred(&rec->ur_uc, req, rsd);
1854         if (rc) {
1855                 GOTO(out, rc);
1856         }
1857
1858         /* rc will be used to interrupt a for loop over multiple records */
1859         rc = mds_reint_rec(rec, offset, req, lockh);
1860
1861  out:
1862         mds_exit_ucred(&rec->ur_uc);
1863         OBD_FREE(rec, sizeof(*rec));
1864         RETURN(rc);
1865 }
1866
1867 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1868                                        struct obd_device *obd, int *process)
1869 {
1870         switch (req->rq_reqmsg->opc) {
1871         case MDS_CONNECT: /* This will never get here, but for completeness. */
1872         case OST_CONNECT: /* This will never get here, but for completeness. */
1873         case MDS_DISCONNECT:
1874         case OST_DISCONNECT:
1875                *process = 1;
1876                RETURN(0);
1877
1878         case MDS_CLOSE:
1879         case MDS_SYNC: /* used in unmounting */
1880         case OBD_PING:
1881         case MDS_REINT:
1882         case LDLM_ENQUEUE:
1883         case OST_CREATE:
1884                 *process = target_queue_recovery_request(req, obd);
1885                 RETURN(0);
1886
1887         default:
1888                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1889                 *process = 0;
1890                 /* XXX what should we set rq_status to here? */
1891                 req->rq_status = -EAGAIN;
1892                 RETURN(ptlrpc_error(req));
1893         }
1894 }
1895
1896 static char *reint_names[] = {
1897         [REINT_SETATTR] "setattr",
1898         [REINT_CREATE]  "create",
1899         [REINT_LINK]    "link",
1900         [REINT_UNLINK]  "unlink",
1901         [REINT_RENAME]  "rename",
1902         [REINT_OPEN]    "open",
1903 };
1904
1905 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
1906                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1907                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1908                             OBD_MD_FLID)
1909
1910 static void reconstruct_create(struct ptlrpc_request *req)
1911 {
1912         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1913         struct mds_client_data *mcd = med->med_mcd;
1914         struct dentry *dentry;
1915         struct ost_body *body;
1916         struct lustre_id id;
1917         int rc;
1918         ENTRY;
1919
1920         /* copy rc, transno and disp; steal locks */
1921         mds_req_from_mcd(req, mcd);
1922         if (req->rq_status) {
1923                 EXIT;
1924                 return;
1925         }
1926
1927         id_gen(&id) = 0;
1928         id_group(&id) = 0;
1929
1930         id_ino(&id) = mcd->mcd_last_data;
1931         LASSERT(id_ino(&id) != 0);
1932
1933         dentry = mds_id2dentry(req2obd(req), &id, NULL);
1934         if (IS_ERR(dentry)) {
1935                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1936                 req->rq_status = PTR_ERR(dentry);
1937                 EXIT;
1938                 return;
1939         }
1940
1941         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1942               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1943               dentry->d_inode->i_generation);
1944
1945         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1946         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1947         body->oa.o_id = dentry->d_inode->i_ino;
1948         body->oa.o_generation = dentry->d_inode->i_generation;
1949         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1950
1951         down(&dentry->d_inode->i_sem);
1952         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1953         up(&dentry->d_inode->i_sem);
1954         if (rc) {
1955                 CERROR("Can't read inode self id, inode %lu, "
1956                        "rc %d\n", dentry->d_inode->i_ino, rc);
1957                 id_fid(&id) = 0;
1958         }
1959
1960         body->oa.o_fid = id_fid(&id);
1961         body->oa.o_mds = id_group(&id);
1962         l_dput(dentry);
1963
1964         EXIT;
1965 }
1966
1967 static int mdt_obj_create(struct ptlrpc_request *req)
1968 {
1969         struct obd_device *obd = req->rq_export->exp_obd;
1970         struct mds_obd *mds = &obd->u.mds;
1971         struct ost_body *body, *repbody;
1972         char idname[LL_ID_NAMELEN];
1973         int size = sizeof(*repbody);
1974         struct inode *parent_inode;
1975         struct lvfs_run_ctxt saved;
1976         int rc, cleanup_phase = 0;
1977         struct dentry *new = NULL;
1978         struct dentry_params dp;
1979         int mealen, flags = 0;
1980         struct lvfs_ucred uc;
1981         struct lustre_id id;
1982         struct mea *mea;
1983         void *handle = NULL;
1984         unsigned long cr_inum = 0;
1985         ENTRY;
1986
1987         DEBUG_REQ(D_HA, req, "create remote object");
1988         parent_inode = mds->mds_unnamed_dir->d_inode;
1989
1990         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1991                                   lustre_swab_ost_body);
1992         if (body == NULL)
1993                 RETURN(-EFAULT);
1994
1995         rc = lustre_pack_reply(req, 1, &size, NULL);
1996         if (rc)
1997                 RETURN(rc);
1998
1999         MDS_CHECK_RESENT(req, reconstruct_create(req));
2000
2001         uc.luc_lsd = NULL;
2002         uc.luc_ginfo = NULL;
2003         uc.luc_uid = body->oa.o_uid;
2004         uc.luc_gid = body->oa.o_gid;
2005         uc.luc_fsuid = body->oa.o_uid;
2006         uc.luc_fsgid = body->oa.o_gid;
2007
2008         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2009         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2010
2011         /* in REPLAY case inum should be given (client or other MDS fills it) */
2012         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2013             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2014                 /*
2015                  * this is re-create request from MDS holding directory name.
2016                  * we have to lookup given ino/gen first. if it exists (good
2017                  * case) then there is nothing to do. if it does not then we
2018                  * have to recreate it.
2019                  */
2020                 id_ino(&id) = body->oa.o_id;
2021                 id_gen(&id) = body->oa.o_generation;
2022
2023                 new = mds_id2dentry(obd, &id, NULL);
2024                 if (!IS_ERR(new) && new->d_inode) {
2025                         struct lustre_id sid;
2026
2027                         CWARN("mkdir() repairing is on its way: %lu/%lu\n",
2028                               (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
2029
2030                         obdo_from_inode(&repbody->oa, new->d_inode,
2031                                         FILTER_VALID_FLAGS);
2032
2033                         repbody->oa.o_id = new->d_inode->i_ino;
2034                         repbody->oa.o_generation = new->d_inode->i_generation;
2035                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2036                         cleanup_phase = 1;
2037
2038                         down(&new->d_inode->i_sem);
2039                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2040                         up(&new->d_inode->i_sem);
2041                         if (rc) {
2042                                 CERROR("Can't read inode self id "
2043                                        "inode %lu, rc %d.\n",
2044                                        new->d_inode->i_ino, rc);
2045                                 GOTO(cleanup, rc);
2046                         }
2047
2048                         repbody->oa.o_fid = id_fid(&sid);
2049                         repbody->oa.o_mds = id_group(&sid);
2050                         LASSERT(id_fid(&sid) != 0);
2051
2052                         /*
2053                          * here we could use fid passed in body->oa.o_fid and
2054                          * thus avoid mds_read_inode_sid().
2055                          */
2056                         cr_inum = new->d_inode->i_ino;
2057                         GOTO(cleanup, rc = 0);
2058                 }
2059         }
2060
2061         down(&parent_inode->i_sem);
2062         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2063         if (IS_ERR(handle)) {
2064                 up(&parent_inode->i_sem);
2065                 CERROR("fsfilt_start() failed, rc = %d\n",
2066                        (int)PTR_ERR(handle));
2067                 GOTO(cleanup, rc = PTR_ERR(handle));
2068         }
2069         cleanup_phase = 1; /* transaction */
2070
2071 repeat:
2072         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2073         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2074         if (IS_ERR(new)) {
2075                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2076                        obd->obd_name, idname, (int) PTR_ERR(new));
2077                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2078                 up(&parent_inode->i_sem);
2079                 RETURN(PTR_ERR(new));
2080         } else if (new->d_inode) {
2081                 CERROR("%s: name exists. repeat\n", obd->obd_name);
2082                 goto repeat;
2083         }
2084
2085         new->d_fsdata = (void *)&dp;
2086         dp.p_inum = 0;
2087         dp.p_ptr = req;
2088
2089         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2090             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2091                 LASSERT(body->oa.o_id != 0);
2092                 dp.p_inum = body->oa.o_id;
2093                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2094                           (unsigned long)body->oa.o_id,
2095                           (unsigned long)body->oa.o_generation);
2096         }
2097
2098         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2099         if (rc == 0) {
2100                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2101                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2102                         new->d_inode->i_generation = body->oa.o_generation;
2103                         mark_inode_dirty(new->d_inode);
2104
2105                         /*
2106                          * avoiding asserts in cache flush case, as
2107                          * @body->oa.o_id should be zero.
2108                          */
2109                         if (body->oa.o_id) {
2110                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2111                                          "BUG 3550: failed to recreate obj "
2112                                          LPU64" -> %lu\n", body->oa.o_id,
2113                                          new->d_inode->i_ino);
2114
2115                                 LASSERTF(body->oa.o_generation ==
2116                                          new->d_inode->i_generation,
2117                                          "BUG 3550: failed to recreate obj/gen "
2118                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2119                                          body->oa.o_generation,
2120                                          new->d_inode->i_ino,
2121                                          new->d_inode->i_generation);
2122                         }
2123                 }
2124
2125                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2126                 repbody->oa.o_id = new->d_inode->i_ino;
2127                 repbody->oa.o_generation = new->d_inode->i_generation;
2128                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2129
2130                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2131                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2132                         id_group(&id) = mds->mds_num;
2133
2134                         LASSERT(body->oa.o_fid != 0);
2135                         id_fid(&id) = body->oa.o_fid;
2136
2137                         LASSERT(body->oa.o_id != 0);
2138                         id_ino(&id) = repbody->oa.o_id;
2139                         id_gen(&id) = repbody->oa.o_generation;
2140
2141                         down(&new->d_inode->i_sem);
2142                         rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2143                         up(&new->d_inode->i_sem);
2144
2145                         /*
2146                          * make sure, that fid is up-to-date.
2147                          */
2148                         mds_set_last_fid(obd, id_fid(&id));
2149                 } else {
2150                         /*
2151                          * allocate new sid, as object is created from scratch
2152                          * and this is not replay.
2153                          */
2154                         down(&new->d_inode->i_sem);
2155                         rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2156                         up(&new->d_inode->i_sem);
2157                 }
2158                 if (rc) {
2159                         CERROR("Can't update lustre ID for inode %lu, "
2160                                "error = %d\n", new->d_inode->i_ino, rc);
2161                         GOTO(cleanup, rc);
2162                 }
2163
2164                 /* initializing o_fid after it is allocated. */
2165                 repbody->oa.o_fid = id_fid(&id);
2166                 repbody->oa.o_mds = id_group(&id);
2167
2168                 rc = fsfilt_del_dir_entry(obd, new);
2169                 up(&parent_inode->i_sem);
2170                 if (rc) {
2171                         CERROR("can't remove name for object: %d\n", rc);
2172                         GOTO(cleanup, rc);
2173                 }
2174
2175                 cleanup_phase = 2; /* created directory object */
2176
2177                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2178                        (unsigned long)new->d_inode->i_ino,
2179                        (unsigned long)new->d_inode->i_generation,
2180                        (unsigned)new->d_inode->i_mode);
2181                 cr_inum = new->d_inode->i_ino;
2182         } else {
2183                 up(&parent_inode->i_sem);
2184                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2185                 GOTO(cleanup, rc);
2186         }
2187
2188         if (body->oa.o_valid & OBD_MD_FLID) {
2189                 /* this is new object for splitted dir. We have to prevent
2190                  * recursive splitting on it -bzzz */
2191                 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2192
2193                 OBD_ALLOC(mea, mealen);
2194                 if (mea == NULL)
2195                         GOTO(cleanup, rc = -ENOMEM);
2196
2197                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2198                 mea->mea_master = 0;
2199                 mea->mea_count = 0;
2200
2201                 down(&new->d_inode->i_sem);
2202                 rc = fsfilt_set_md(obd, new->d_inode, handle,
2203                                    mea, mealen, EA_MEA);
2204                 up(&new->d_inode->i_sem);
2205                 if (rc)
2206                         CERROR("fsfilt_set_md() failed, "
2207                                "rc = %d\n", rc);
2208
2209                 OBD_FREE(mea, mealen);
2210
2211                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2212                        obd->obd_name, new->d_inode->i_ino,
2213                        new->d_inode->i_generation, flags);
2214         } else if (body->oa.o_easize) {
2215                 /* we pass LCK_EX to split routine to signal that we have
2216                  * exclusive access to the directory. simple because nobody
2217                  * knows it already exists -bzzz */
2218                 rc = mds_try_to_split_dir(obd, new, NULL,
2219                                           body->oa.o_easize, LCK_EX);
2220                 if (rc < 0) {
2221                         CERROR("Can't split directory %lu, error = %d.\n",
2222                                new->d_inode->i_ino, rc);
2223                 } else {
2224                         rc = 0;
2225                 }
2226         }
2227
2228         EXIT;
2229 cleanup:
2230         switch (cleanup_phase) {
2231         case 2: /* object has been created, but we'll may want to replay it later */
2232                 if (rc == 0)
2233                         ptlrpc_require_repack(req);
2234         case 1: /* transaction */
2235                 rc = mds_finish_transno(mds, parent_inode, handle,
2236                                         req, rc, cr_inum);
2237         }
2238
2239         l_dput(new);
2240         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2241         return rc;
2242 }
2243
2244 static int mdt_get_info(struct ptlrpc_request *req)
2245 {
2246         struct obd_export *exp = req->rq_export;
2247         int keylen, rc = 0;
2248         char *key;
2249         ENTRY;
2250
2251         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2252         if (key == NULL) {
2253                 DEBUG_REQ(D_HA, req, "no get_info key");
2254                 RETURN(-EFAULT);
2255         }
2256         keylen = req->rq_reqmsg->buflens[0];
2257
2258         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2259             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2260             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2261                 RETURN(-EPROTO);
2262
2263         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2264                 struct lustre_id *reply;
2265                 int size = sizeof(*reply);
2266
2267                 rc = lustre_pack_reply(req, 1, &size, NULL);
2268                 if (rc)
2269                         RETURN(rc);
2270
2271                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2272                 rc = obd_get_info(exp, keylen, key, &size, reply);
2273         } else {
2274                 obd_id *reply;
2275                 int size = sizeof(*reply);
2276
2277                 rc = lustre_pack_reply(req, 1, &size, NULL);
2278                 if (rc)
2279                         RETURN(rc);
2280
2281                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2282                 rc = obd_get_info(exp, keylen, key, &size, reply);
2283         }
2284
2285         req->rq_repmsg->status = 0;
2286         RETURN(rc);
2287 }
2288
2289 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2290                         void *key, __u32 vallen, void *val)
2291 {
2292         struct obd_device *obd;
2293         struct mds_obd *mds;
2294         int rc = 0;
2295         ENTRY;
2296
2297         obd = class_exp2obd(exp);
2298         if (obd == NULL) {
2299                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2300                        exp->exp_handle.h_cookie);
2301                 RETURN(-EINVAL);
2302         }
2303
2304         mds = &obd->u.mds;
2305         if (keylen >= strlen("mds_type") &&
2306              memcmp(key, "mds_type", keylen) == 0) {
2307                 int valsize;
2308                 __u32 group;
2309
2310                 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2311
2312                 mds->mds_obd_type = *(int*)val;
2313                 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2314                 valsize = sizeof(group);
2315
2316                 /* mds number has been changed, so the corresponding obdfilter
2317                  * exp need to be changed too. */
2318                 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2319                                   "mds_conn", valsize, &group);
2320                 RETURN(rc);
2321         }
2322         CDEBUG(D_IOCTL, "invalid key\n");
2323         RETURN(-EINVAL);
2324 }
2325
2326 static int mdt_set_info(struct ptlrpc_request *req)
2327 {
2328         char *key, *val;
2329         struct obd_export *exp = req->rq_export;
2330         int keylen, rc = 0, vallen;
2331         ENTRY;
2332
2333         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2334         if (key == NULL) {
2335                 DEBUG_REQ(D_HA, req, "no set_info key");
2336                 RETURN(-EFAULT);
2337         }
2338         keylen = req->rq_reqmsg->buflens[0];
2339
2340         if (keylen == strlen("mds_type") &&
2341             memcmp(key, "mds_type", keylen) == 0) {
2342                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2343                 if (rc)
2344                         RETURN(rc);
2345
2346                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2347                 vallen = req->rq_reqmsg->buflens[1];
2348
2349                 rc = obd_set_info(exp, keylen, key, vallen, val);
2350                 req->rq_repmsg->status = 0;
2351                 RETURN(rc);
2352         }
2353         CDEBUG(D_IOCTL, "invalid key\n");
2354         RETURN(-EINVAL);
2355 }
2356
2357 static void mds_revoke_export_locks(struct obd_export *exp)
2358 {
2359         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
2360         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2361         struct ldlm_lock *lock, *next;
2362         struct ldlm_lock_desc desc;
2363
2364         if (!exp->u.eu_mds_data.med_remote)
2365                 return;
2366
2367         ENTRY;
2368         l_lock(&ns->ns_lock);
2369         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2370                 if (lock->l_req_mode != lock->l_granted_mode)
2371                         continue;
2372
2373                 LASSERT(lock->l_resource);
2374                 if (lock->l_resource->lr_type != LDLM_IBITS &&
2375                     lock->l_resource->lr_type != LDLM_PLAIN)
2376                         continue;
2377
2378                 if (lock->l_flags & LDLM_FL_AST_SENT)
2379                         continue;
2380
2381                 lock->l_flags |= LDLM_FL_AST_SENT;
2382
2383                 /* the desc just pretend to exclusive */
2384                 ldlm_lock2desc(lock, &desc);
2385                 desc.l_req_mode = LCK_EX;
2386                 desc.l_granted_mode = 0;
2387
2388                 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2389         }
2390         l_unlock(&ns->ns_lock);
2391         EXIT;
2392 }
2393
2394 static int mds_msg_check_version(struct lustre_msg *msg)
2395 {
2396         int rc;
2397
2398         switch (msg->opc) {
2399         case MDS_CONNECT:
2400         case MDS_DISCONNECT:
2401         case OBD_PING:
2402                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2403                 if (rc)
2404                         CERROR("bad opc %u version %08x, expecting %08x\n",
2405                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2406                 break;
2407         case MDS_STATFS:
2408         case MDS_GETSTATUS:
2409         case MDS_GETATTR:
2410         case MDS_GETATTR_LOCK:
2411         case MDS_READPAGE:
2412         case MDS_REINT:
2413         case MDS_CLOSE:
2414         case MDS_DONE_WRITING:
2415         case MDS_PIN:
2416         case MDS_SYNC:
2417                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2418                 if (rc)
2419                         CERROR("bad opc %u version %08x, expecting %08x\n",
2420                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2421                 break;
2422         case LDLM_ENQUEUE:
2423         case LDLM_CONVERT:
2424         case LDLM_BL_CALLBACK:
2425         case LDLM_CP_CALLBACK:
2426                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2427                 if (rc)
2428                         CERROR("bad opc %u version %08x, expecting %08x\n",
2429                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2430                 break;
2431         case OBD_LOG_CANCEL:
2432         case LLOG_ORIGIN_HANDLE_OPEN:
2433         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2434         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2435         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2436         case LLOG_ORIGIN_HANDLE_CLOSE:
2437         case LLOG_CATINFO:
2438                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2439                 if (rc)
2440                         CERROR("bad opc %u version %08x, expecting %08x\n",
2441                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2442                 break;
2443         case OST_CREATE:
2444         case OST_WRITE:
2445         case OST_GET_INFO:
2446         case OST_SET_INFO:
2447                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2448                 if (rc)
2449                         CERROR("bad opc %u version %08x, expecting %08x\n",
2450                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2451                 break;
2452         case SEC_INIT:
2453         case SEC_INIT_CONTINUE:
2454         case SEC_FINI:
2455                 rc = 0;
2456                 break;
2457         default:
2458                 CERROR("MDS unknown opcode %d\n", msg->opc);
2459                 rc = -ENOTSUPP;
2460                 break;
2461         }
2462
2463         return rc;
2464 }
2465
2466 int mds_handle(struct ptlrpc_request *req)
2467 {
2468         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2469         struct obd_device *obd = NULL;
2470         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2471         int rc = 0;
2472         ENTRY;
2473
2474         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2475
2476         rc = mds_msg_check_version(req->rq_reqmsg);
2477         if (rc) {
2478                 CERROR("MDS drop mal-formed request\n");
2479                 RETURN(rc);
2480         }
2481
2482         /* Security opc should NOT trigger any recovery events */
2483         if (req->rq_reqmsg->opc == SEC_INIT ||
2484             req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2485                 if (!req->rq_export)
2486                         GOTO(out, rc = 0);
2487
2488                 mds_req_add_idmapping(req, &req->rq_export->exp_mds_data);
2489                 mds_revoke_export_locks(req->rq_export);
2490                 GOTO(out, rc = 0);
2491         } else if (req->rq_reqmsg->opc == SEC_FINI) {
2492                 if (!req->rq_export)
2493                         GOTO(out, rc = 0);
2494
2495                 mds_req_del_idmapping(req, &req->rq_export->exp_mds_data);
2496                 mds_revoke_export_locks(req->rq_export);
2497                 GOTO(out, rc = 0);
2498         }
2499
2500         LASSERT(current->journal_info == NULL);
2501         /* XXX identical to OST */
2502         if (req->rq_reqmsg->opc != MDS_CONNECT) {
2503                 struct mds_export_data *med;
2504                 int recovering;
2505
2506                 if (req->rq_export == NULL) {
2507                         CERROR("operation %d on unconnected MDS from %s\n",
2508                                req->rq_reqmsg->opc,
2509                                req->rq_peerstr);
2510                         req->rq_status = -ENOTCONN;
2511                         GOTO(out, rc = -ENOTCONN);
2512                 }
2513
2514                 med = &req->rq_export->exp_mds_data;
2515                 obd = req->rq_export->exp_obd;
2516                 mds = &obd->u.mds;
2517
2518                 /* sanity check: if the xid matches, the request must
2519                  * be marked as a resent or replayed */
2520                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2521                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2522                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2523                                  (MSG_RESENT | MSG_REPLAY),
2524                                  "rq_xid "LPU64" matches last_xid, "
2525                                  "expected RESENT flag\n",
2526                                  req->rq_xid);
2527                 }
2528                 /* else: note the opposite is not always true; a
2529                  * RESENT req after a failover will usually not match
2530                  * the last_xid, since it was likely never
2531                  * committed. A REPLAYed request will almost never
2532                  * match the last xid, however it could for a
2533                  * committed, but still retained, open. */
2534
2535                 spin_lock_bh(&obd->obd_processing_task_lock);
2536                 recovering = obd->obd_recovering;
2537                 spin_unlock_bh(&obd->obd_processing_task_lock);
2538                 if (recovering) {
2539                         rc = mds_filter_recovery_request(req, obd,
2540                                                          &should_process);
2541                         if (rc || should_process == 0) {
2542                                 RETURN(rc);
2543                         } else if (should_process < 0) {
2544                                 req->rq_status = should_process;
2545                                 rc = ptlrpc_error(req);
2546                                 RETURN(rc);
2547                         }
2548                 }
2549         }
2550
2551         switch (req->rq_reqmsg->opc) {
2552         case MDS_CONNECT:
2553                 DEBUG_REQ(D_INODE, req, "connect");
2554                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2555                 rc = target_handle_connect(req);
2556                 if (!rc) {
2557                         struct mds_export_data *med;
2558
2559                         LASSERT(req->rq_export);
2560                         med = &req->rq_export->u.eu_mds_data;
2561                         mds_init_export_data(req, med);
2562                         mds_req_add_idmapping(req, med);
2563
2564                         /* Now that we have an export, set mds. */
2565                         obd = req->rq_export->exp_obd;
2566                         mds = mds_req2mds(req);
2567                 }
2568                 break;
2569
2570         case MDS_DISCONNECT:
2571                 DEBUG_REQ(D_INODE, req, "disconnect");
2572                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2573                 rc = target_handle_disconnect(req);
2574                 req->rq_status = rc;            /* superfluous? */
2575                 break;
2576
2577         case MDS_GETSTATUS:
2578                 DEBUG_REQ(D_INODE, req, "getstatus");
2579                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2580                 rc = mds_getstatus(req);
2581                 break;
2582
2583         case MDS_GETATTR:
2584                 DEBUG_REQ(D_INODE, req, "getattr");
2585                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2586                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2587                 break;
2588
2589         case MDS_GETATTR_LOCK: {
2590                 struct lustre_handle lockh;
2591                 DEBUG_REQ(D_INODE, req, "getattr_lock");
2592                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2593
2594                 /* If this request gets a reconstructed reply, we won't be
2595                  * acquiring any new locks in mds_getattr_lock, so we don't
2596                  * want to cancel.
2597                  */
2598                 lockh.cookie = 0;
2599                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2600                                       MDS_INODELOCK_UPDATE);
2601                 /* this non-intent call (from an ioctl) is special */
2602                 req->rq_status = rc;
2603                 if (rc == 0 && lockh.cookie)
2604                         ldlm_lock_decref(&lockh, LCK_PR);
2605                 break;
2606         }
2607         case MDS_STATFS:
2608                 DEBUG_REQ(D_INODE, req, "statfs");
2609                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2610                 rc = mds_statfs(req);
2611                 break;
2612
2613         case MDS_READPAGE:
2614                 DEBUG_REQ(D_INODE, req, "readpage");
2615                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2616                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2617
2618                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2619                         if (req->rq_reply_state) {
2620                                 lustre_free_reply_state (req->rq_reply_state);
2621                                 req->rq_reply_state = NULL;
2622                         }
2623                         RETURN(0);
2624                 }
2625
2626                 break;
2627         case MDS_REINT: {
2628                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2629                                              sizeof (*opcp));
2630                 __u32  opc;
2631                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2632                                mds->mds_max_cookiesize};
2633                 int bufcount;
2634
2635                 /* NB only peek inside req now; mds_reint() will swab it */
2636                 if (opcp == NULL) {
2637                         CERROR ("Can't inspect opcode\n");
2638                         rc = -EINVAL;
2639                         break;
2640                 }
2641                 opc = *opcp;
2642                 if (lustre_msg_swabbed (req->rq_reqmsg))
2643                         __swab32s(&opc);
2644
2645                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2646                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2647                            reint_names[opc] == NULL) ? reint_names[opc] :
2648                                                        "unknown opcode");
2649
2650                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2651
2652                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2653                         bufcount = 3;
2654                 else if (opc == REINT_OPEN)
2655                         bufcount = 2;
2656                 else
2657                         bufcount = 1;
2658
2659                 rc = lustre_pack_reply(req, bufcount, size, NULL);
2660                 if (rc)
2661                         break;
2662
2663                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2664                 fail = OBD_FAIL_MDS_REINT_NET_REP;
2665                 break;
2666         }
2667
2668         case MDS_CLOSE:
2669                 DEBUG_REQ(D_INODE, req, "close");
2670                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2671                 rc = mds_close(req, MDS_REQ_REC_OFF);
2672                 break;
2673
2674         case MDS_DONE_WRITING:
2675                 DEBUG_REQ(D_INODE, req, "done_writing");
2676                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2677                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2678                 break;
2679
2680         case MDS_PIN:
2681                 DEBUG_REQ(D_INODE, req, "pin");
2682                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2683                 rc = mds_pin(req, MDS_REQ_REC_OFF);
2684                 break;
2685
2686         case MDS_SYNC:
2687                 DEBUG_REQ(D_INODE, req, "sync");
2688                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2689                 rc = mds_sync(req, MDS_REQ_REC_OFF);
2690                 break;
2691
2692         case OBD_PING:
2693                 DEBUG_REQ(D_INODE, req, "ping");
2694                 rc = target_handle_ping(req);
2695                 break;
2696
2697         case OBD_LOG_CANCEL:
2698                 CDEBUG(D_INODE, "log cancel\n");
2699                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2700                 rc = -ENOTSUPP; /* la la la */
2701                 break;
2702
2703         case LDLM_ENQUEUE:
2704                 DEBUG_REQ(D_INODE, req, "enqueue");
2705                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2706                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2707                                          ldlm_server_blocking_ast, NULL);
2708                 fail = OBD_FAIL_LDLM_REPLY;
2709                 break;
2710         case LDLM_CONVERT:
2711                 DEBUG_REQ(D_INODE, req, "convert");
2712                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2713                 rc = ldlm_handle_convert(req);
2714                 break;
2715         case LDLM_BL_CALLBACK:
2716         case LDLM_CP_CALLBACK:
2717                 DEBUG_REQ(D_INODE, req, "callback");
2718                 CERROR("callbacks should not happen on MDS\n");
2719                 LBUG();
2720                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2721                 break;
2722         case LLOG_ORIGIN_HANDLE_OPEN:
2723                 DEBUG_REQ(D_INODE, req, "llog_init");
2724                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2725                 rc = llog_origin_handle_open(req);
2726                 break;
2727         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2728                 DEBUG_REQ(D_INODE, req, "llog next block");
2729                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2730                 rc = llog_origin_handle_next_block(req);
2731                 break;
2732         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2733                 DEBUG_REQ(D_INODE, req, "llog prev block");
2734                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2735                 rc = llog_origin_handle_prev_block(req);
2736                 break;
2737         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2738                 DEBUG_REQ(D_INODE, req, "llog read header");
2739                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2740                 rc = llog_origin_handle_read_header(req);
2741                 break;
2742         case LLOG_ORIGIN_HANDLE_CLOSE:
2743                 DEBUG_REQ(D_INODE, req, "llog close");
2744                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2745                 rc = llog_origin_handle_close(req);
2746                 break;
2747         case OST_CREATE:
2748                 DEBUG_REQ(D_INODE, req, "ost_create");
2749                 rc = mdt_obj_create(req);
2750                 break;
2751         case OST_GET_INFO:
2752                 DEBUG_REQ(D_INODE, req, "get_info");
2753                 rc = mdt_get_info(req);
2754                 break;
2755         case OST_SET_INFO:
2756                 DEBUG_REQ(D_INODE, req, "set_info");
2757                 rc = mdt_set_info(req);
2758                 break;
2759         case OST_WRITE:
2760                 CDEBUG(D_INODE, "write\n");
2761                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2762                 rc = ost_brw_write(req, NULL);
2763                 LASSERT(current->journal_info == NULL);
2764                 /* mdt_brw sends its own replies */
2765                 RETURN(rc);
2766                 break;
2767         case LLOG_CATINFO:
2768                 DEBUG_REQ(D_INODE, req, "llog catinfo");
2769                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2770                 rc = llog_catinfo(req);
2771                 break;
2772         default:
2773                 req->rq_status = -ENOTSUPP;
2774                 rc = ptlrpc_error(req);
2775                 RETURN(rc);
2776         }
2777
2778         LASSERT(current->journal_info == NULL);
2779
2780         EXIT;
2781
2782         /* If we're DISCONNECTing, the mds_export_data is already freed */
2783         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2784                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2785                 struct obd_device *obd = list_entry(mds, struct obd_device,
2786                                                     u.mds);
2787                 req->rq_repmsg->last_xid =
2788                         le64_to_cpu(med->med_mcd->mcd_last_xid);
2789
2790                 if (!obd->obd_no_transno) {
2791                         req->rq_repmsg->last_committed =
2792                                 obd->obd_last_committed;
2793                 } else {
2794                         DEBUG_REQ(D_IOCTL, req,
2795                                   "not sending last_committed update");
2796                 }
2797                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2798                        ", xid "LPU64"\n",
2799                        mds->mds_last_transno, obd->obd_last_committed,
2800                        req->rq_xid);
2801         }
2802  out:
2803
2804         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2805                 if (obd && obd->obd_recovering) {
2806                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2807                         return target_queue_final_reply(req, rc);
2808                 }
2809                 /* Lost a race with recovery; let the error path DTRT. */
2810                 rc = req->rq_status = -ENOTCONN;
2811         }
2812
2813         target_send_reply(req, rc, fail);
2814         return 0;
2815 }
2816
2817 /* Update the server data on disk.  This stores the new mount_count and also the
2818  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
2819  * last_rcvd value may be less than that of the clients.  This will alert us
2820  * that we may need to do client recovery.
2821  *
2822  * Also assumes for mds_last_transno that we are not modifying it (no locking).
2823  */
2824 int mds_update_server_data(struct obd_device *obd, int force_sync)
2825 {
2826         struct mds_obd *mds = &obd->u.mds;
2827         struct mds_server_data *msd = mds->mds_server_data;
2828         struct file *filp = mds->mds_rcvd_filp;
2829         struct lvfs_run_ctxt saved;
2830         loff_t off = 0;
2831         int rc;
2832         ENTRY;
2833
2834         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2835         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2836
2837         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2838                mds->mds_mount_count, mds->mds_last_transno);
2839         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2840         if (rc)
2841                 CERROR("error writing MDS server data: rc = %d\n", rc);
2842         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2843
2844         RETURN(rc);
2845 }
2846
2847 /* saves last allocated fid counter to file. */
2848 int mds_update_last_fid(struct obd_device *obd, void *handle,
2849                         int force_sync)
2850 {
2851         struct mds_obd *mds = &obd->u.mds;
2852         struct file *filp = mds->mds_fid_filp;
2853         struct lvfs_run_ctxt saved;
2854         loff_t off = 0;
2855         __u64 last_fid;
2856         int rc = 0;
2857         ENTRY;
2858
2859         spin_lock(&mds->mds_last_fid_lock);
2860         last_fid = mds->mds_last_fid;
2861         spin_unlock(&mds->mds_last_fid_lock);
2862
2863         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2864                last_fid);
2865
2866         if (handle) {
2867                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
2868                                       handle, mds_commit_last_fid_cb,
2869                                       NULL);
2870         }
2871
2872         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2873         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
2874                                  &off, force_sync);
2875         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2876
2877         if (rc) {
2878                 CERROR("error writing MDS last_fid #"LPU64
2879                        ", err = %d\n", last_fid, rc);
2880                 RETURN(rc);
2881         }
2882
2883         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2884                "%llu: err = %d\n", last_fid, off, rc);
2885
2886         RETURN(rc);
2887 }
2888
2889 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2890 {
2891         struct mds_obd *mds = &obd->u.mds;
2892
2893         spin_lock(&mds->mds_last_fid_lock);
2894         if (fid > mds->mds_last_fid)
2895                 mds->mds_last_fid = fid;
2896         spin_unlock(&mds->mds_last_fid_lock);
2897 }
2898
2899 void mds_commit_last_transno_cb(struct obd_device *obd,
2900                                 __u64 transno, void *data,
2901                                 int error)
2902 {
2903         obd_transno_commit_cb(obd, transno, error);
2904 }
2905
2906 void mds_commit_last_fid_cb(struct obd_device *obd,
2907                             __u64 fid, void *data,
2908                             int error)
2909 {
2910         if (error) {
2911                 CERROR("%s: fid "LPD64" commit error: %d\n",
2912                        obd->obd_name, fid, error);
2913                 return;
2914         }
2915
2916         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2917                obd->obd_name, fid);
2918 }
2919
2920 __u64 mds_alloc_fid(struct obd_device *obd)
2921 {
2922         struct mds_obd *mds = &obd->u.mds;
2923         __u64 fid;
2924
2925         spin_lock(&mds->mds_last_fid_lock);
2926         fid = ++mds->mds_last_fid;
2927         spin_unlock(&mds->mds_last_fid_lock);
2928
2929         return fid;
2930 }
2931
2932 /*
2933  * allocates new lustre_id on passed @inode and saves it to inode EA.
2934  */
2935 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2936                         void *handle, struct lustre_id *id)
2937 {
2938         struct mds_obd *mds = &obd->u.mds;
2939         int alloc = 0, rc = 0;
2940         ENTRY;
2941
2942         LASSERT(obd != NULL);
2943         LASSERT(inode != NULL);
2944
2945         if (id == NULL) {
2946                 OBD_ALLOC(id, sizeof(*id));
2947                 if (id == NULL)
2948                         RETURN(-ENOMEM);
2949                 alloc = 1;
2950         }
2951
2952         id_group(id) = mds->mds_num;
2953         id_fid(id) = mds_alloc_fid(obd);
2954         id_ino(id) = inode->i_ino;
2955         id_gen(id) = inode->i_generation;
2956         id_type(id) = (S_IFMT & inode->i_mode);
2957
2958         rc = mds_update_inode_sid(obd, inode, handle, id);
2959         if (rc) {
2960                 CERROR("Can't update inode FID EA, "
2961                        "rc = %d\n", rc);
2962         }
2963
2964         if (alloc)
2965                 OBD_FREE(id, sizeof(*id));
2966         RETURN(rc);
2967 }
2968
2969 /*
2970  * reads inode self id from inode EA. Probably later this should be replaced by
2971  * caching inode self id to avoid raeding it every time it is needed.
2972  */
2973 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2974                        struct lustre_id *id)
2975 {
2976         int rc;
2977         ENTRY;
2978
2979         LASSERT(id != NULL);
2980         LASSERT(obd != NULL);
2981         LASSERT(inode != NULL);
2982
2983         rc = fsfilt_get_md(obd, inode, &id->li_fid,
2984                            sizeof(id->li_fid), EA_SID);
2985         if (rc < 0) {
2986                 CERROR("fsfilt_get_md() failed, "
2987                        "rc = %d\n", rc);
2988                 RETURN(rc);
2989         } else if (!rc) {
2990                 rc = -ENODATA;
2991                 RETURN(rc);
2992         } else {
2993                 rc = 0;
2994         }
2995
2996         RETURN(rc);
2997 }
2998
2999 /* updates inode self id in EA. */
3000 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3001                          void *handle, struct lustre_id *id)
3002 {
3003         int rc = 0;
3004         ENTRY;
3005
3006         LASSERT(id != NULL);
3007         LASSERT(obd != NULL);
3008         LASSERT(inode != NULL);
3009
3010         rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3011                            sizeof(id->li_fid), EA_SID);
3012         if (rc) {
3013                 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3014                 RETURN(rc);
3015         }
3016
3017         RETURN(rc);
3018 }
3019
3020 /*
3021  * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3022  * to master MDS by correct store cookie, needed to find inode on master MDS
3023  * quickly.
3024  */
3025 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3026                        struct lustre_id *id)
3027 {
3028         int rc;
3029         ENTRY;
3030
3031         LASSERT(id != NULL);
3032         LASSERT(obd != NULL);
3033         LASSERT(inode != NULL);
3034
3035         rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3036         if (rc < 0) {
3037                 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3038                 RETURN(rc);
3039         } else if (!rc) {
3040                 rc = -ENODATA;
3041                 RETURN(rc);
3042         } else {
3043                 rc = 0;
3044         }
3045
3046         RETURN(rc);
3047 }
3048
3049 /*
3050  * updates master inode id. Usualy this is done by CMOBD after an inode is
3051  * created and relationship between cache MDS and master one should be
3052  * established.
3053  */
3054 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3055                          void *handle, struct lustre_id *id)
3056 {
3057         int rc = 0;
3058         ENTRY;
3059
3060         LASSERT(id != NULL);
3061         LASSERT(obd != NULL);
3062         LASSERT(inode != NULL);
3063
3064         rc = fsfilt_set_md(obd, inode, handle, id,
3065                            sizeof(*id), EA_MID);
3066         if (rc) {
3067                 CERROR("fsfilt_set_md() failed, "
3068                        "rc = %d\n", rc);
3069                 RETURN(rc);
3070         }
3071
3072         RETURN(rc);
3073 }
3074
3075 /* mount the file system (secretly) */
3076 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3077 {
3078         struct lustre_cfg* lcfg = buf;
3079         struct mds_obd *mds = &obd->u.mds;
3080         struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3081         char *options = NULL;
3082         struct vfsmount *mnt;
3083         char ns_name[48];
3084         unsigned long page;
3085         int rc = 0;
3086         ENTRY;
3087
3088
3089         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
3090                 RETURN(rc = -EINVAL);
3091
3092         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
3093         if (IS_ERR(obd->obd_fsops))
3094                 RETURN(rc = PTR_ERR(obd->obd_fsops));
3095
3096         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3097
3098         page = __get_free_page(GFP_KERNEL);
3099         if (!page)
3100                 RETURN(-ENOMEM);
3101
3102         options = (char *)page;
3103         memset(options, 0, PAGE_SIZE);
3104
3105         /*
3106          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3107          * and the rest of options are passed by mount options. Probably this
3108          * should be moved to somewhere else like startup scripts or lconf. */
3109         sprintf(options, "iopen_nopriv");
3110         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
3111                 sprintf(options + strlen(options), ",%s",
3112                         lcfg->lcfg_inlbuf4);
3113
3114         /* we have to know mdsnum before touching underlying fs -bzzz */
3115         atomic_set(&mds->mds_open_count, 0);
3116         sema_init(&mds->mds_md_sem, 1);
3117         mds->mds_md_connected = 0;
3118         mds->mds_md_name = NULL;
3119
3120         if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
3121             strncmp(lcfg->lcfg_inlbuf5, "dumb", lcfg->lcfg_inllen5)) {
3122                 class_uuid_t uuid;
3123
3124                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3125                        obd->obd_name, lcfg->lcfg_inlbuf5);
3126
3127                 generate_random_uuid(uuid);
3128                 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3129
3130                 OBD_ALLOC(mds->mds_md_name, lcfg->lcfg_inllen5);
3131                 if (mds->mds_md_name == NULL)
3132                         RETURN(rc = -ENOMEM);
3133
3134                 memcpy(mds->mds_md_name, lcfg->lcfg_inlbuf5,
3135                        lcfg->lcfg_inllen5);
3136
3137                 rc = mds_md_connect(obd, mds->mds_md_name);
3138                 if (rc) {
3139                         OBD_FREE(mds->mds_md_name, lcfg->lcfg_inllen5);
3140                         GOTO(err_ops, rc);
3141                 }
3142         }
3143
3144         mds->mds_obd_type = MDS_MASTER_OBD;
3145
3146         if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
3147             strncmp(lcfg->lcfg_inlbuf6, "dumb", lcfg->lcfg_inllen6)) {
3148                 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
3149                         mds->mds_obd_type = MDS_MASTER_OBD;
3150                 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
3151                         mds->mds_obd_type = MDS_CACHE_OBD;
3152                 }
3153         }
3154
3155         rc = lvfs_mount_fs(lcfg->lcfg_inlbuf1, lcfg->lcfg_inlbuf2,
3156                            options, 0, &lvfs_ctxt);
3157
3158         free_page(page);
3159
3160         if (rc || !lvfs_ctxt) {
3161                 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3162                 GOTO(err_ops, rc);
3163         }
3164
3165         mnt = lvfs_ctxt->loc_mnt;
3166         mds->mds_lvfs_ctxt = lvfs_ctxt;
3167         ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3168
3169         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
3170
3171         sema_init(&mds->mds_epoch_sem, 1);
3172         atomic_set(&mds->mds_real_clients, 0);
3173         spin_lock_init(&mds->mds_transno_lock);
3174         spin_lock_init(&mds->mds_last_fid_lock);
3175         sema_init(&mds->mds_orphan_recovery_sem, 1);
3176         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3177
3178         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3179         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3180
3181         if (obd->obd_namespace == NULL) {
3182                 mds_cleanup(obd, 0);
3183                 GOTO(err_put, rc = -ENOMEM);
3184         }
3185         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3186
3187         rc = mds_fs_setup(obd, mnt);
3188         if (rc) {
3189                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3190                        obd->obd_name, rc);
3191                 GOTO(err_ns, rc);
3192         }
3193
3194         rc = llog_start_commit_thread();
3195         if (rc < 0)
3196                 GOTO(err_fs, rc);
3197
3198         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
3199             strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)) {
3200                 class_uuid_t uuid;
3201
3202                 generate_random_uuid(uuid);
3203                 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3204
3205                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
3206                 if (mds->mds_profile == NULL)
3207                         GOTO(err_fs, rc = -ENOMEM);
3208
3209                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
3210                        lcfg->lcfg_inllen3);
3211         }
3212
3213         /*
3214          * setup root dir and files ID dir if lmv already connected, or there is
3215          * not lmv at all.
3216          */
3217         if (mds->mds_md_exp || (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
3218                                 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)))
3219         {
3220                 rc = mds_fs_setup_rootid(obd);
3221                 if (rc)
3222                         GOTO(err_fs, rc);
3223
3224                 rc = mds_fs_setup_virtid(obd);
3225                 if (rc)
3226                         GOTO(err_fs, rc);
3227         }
3228
3229         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3230                            "mds_ldlm_client", &obd->obd_ldlm_client);
3231         obd->obd_replayable = 1;
3232
3233         rc = mds_postsetup(obd);
3234         if (rc)
3235                 GOTO(err_fs, rc);
3236
3237         RETURN(0);
3238
3239 err_fs:
3240         /* No extra cleanup needed for llog_init_commit_thread() */
3241         mds_fs_cleanup(obd, 0);
3242 err_ns:
3243         ldlm_namespace_free(obd->obd_namespace, 0);
3244         obd->obd_namespace = NULL;
3245 err_put:
3246         unlock_kernel();
3247         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3248         mds->mds_sb = 0;
3249         lock_kernel();
3250 err_ops:
3251         fsfilt_put_ops(obd->obd_fsops);
3252         return rc;
3253 }
3254
3255 static int mds_postsetup(struct obd_device *obd)
3256 {
3257         struct mds_obd *mds = &obd->u.mds;
3258         int rc = 0;
3259         ENTRY;
3260
3261         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3262                             obd, 0, NULL, &llog_lvfs_ops);
3263         if (rc)
3264                 RETURN(rc);
3265
3266         if (mds->mds_profile) {
3267                 struct llog_ctxt *lgctxt;
3268                 struct lvfs_run_ctxt saved;
3269                 struct lustre_profile *lprof;
3270                 struct config_llog_instance cfg;
3271
3272                 cfg.cfg_instance = NULL;
3273                 cfg.cfg_uuid = mds->mds_dt_uuid;
3274                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3275
3276                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3277                 if (!lgctxt) {
3278                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3279                         GOTO(err_llog, rc = -EINVAL);
3280                 }
3281
3282                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3283                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3284
3285                 if (rc)
3286                         GOTO(err_llog, rc);
3287
3288                 lprof = class_get_profile(mds->mds_profile);
3289                 if (lprof == NULL) {
3290                         CERROR("No profile found: %s\n", mds->mds_profile);
3291                         GOTO(err_cleanup, rc = -ENOENT);
3292                 }
3293                 rc = mds_dt_connect(obd, lprof->lp_lov);
3294                 if (rc)
3295                         GOTO(err_cleanup, rc);
3296
3297                 rc = mds_md_postsetup(obd);
3298                 if (rc)
3299                         GOTO(err_cleanup, rc);
3300         }
3301
3302         RETURN(rc);
3303 err_cleanup:
3304         mds_dt_clean(obd);
3305 err_llog:
3306         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3307                                           LLOG_CONFIG_ORIG_CTXT));
3308         return rc;
3309 }
3310
3311 int mds_postrecov(struct obd_device *obd)
3312 {
3313         struct mds_obd *mds = &obd->u.mds;
3314         struct llog_ctxt *ctxt;
3315         int rc, item = 0, valsize;
3316          __u32 group;
3317         ENTRY;
3318
3319         LASSERT(!obd->obd_recovering);
3320         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3321         LASSERT(ctxt != NULL);
3322
3323         /* set nextid first, so we are sure it happens */
3324         rc = mds_dt_set_nextid(obd);
3325         if (rc) {
3326                 CERROR("%s: mds_dt_set_nextid() failed\n", obd->obd_name);
3327                 GOTO(out, rc);
3328         }
3329
3330         /* clean PENDING dir */
3331         rc = mds_cleanup_orphans(obd);
3332         if (rc < 0)
3333                 GOTO(out, rc);
3334         item = rc;
3335
3336         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3337         valsize = sizeof(group);
3338         rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"), "mds_conn",
3339                           valsize, &group);
3340         if (rc)
3341                 GOTO(out, rc);
3342
3343         rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3344                           NULL, NULL, NULL);
3345         if (rc) {
3346                 CERROR("%s: failed at llog_origin_connect: %d\n",
3347                        obd->obd_name, rc);
3348                 GOTO(out, rc);
3349         }
3350
3351         /* remove the orphaned precreated objects */
3352         rc = mds_dt_clearorphans(mds, NULL /* all OSTs */);
3353         if (rc)
3354                 GOTO(err_llog, rc);
3355
3356 out:
3357         RETURN(rc < 0 ? rc : item);
3358
3359 err_llog:
3360         /* cleanup all llogging subsystems */
3361         rc = obd_llog_finish(obd, &obd->obd_llogs,
3362                              mds->mds_dt_desc.ld_tgt_count);
3363         if (rc)
3364                 CERROR("%s: failed to cleanup llogging subsystems\n",
3365                         obd->obd_name);
3366         goto out;
3367 }
3368
3369 int mds_dt_clean(struct obd_device *obd)
3370 {
3371         struct mds_obd *mds = &obd->u.mds;
3372         ENTRY;
3373
3374         if (mds->mds_profile) {
3375                 char * cln_prof;
3376                 struct llog_ctxt *llctx;
3377                 struct lvfs_run_ctxt saved;
3378                 struct config_llog_instance cfg;
3379                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3380
3381                 OBD_ALLOC(cln_prof, len);
3382                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3383
3384                 cfg.cfg_instance = NULL;
3385                 cfg.cfg_uuid = mds->mds_dt_uuid;
3386
3387                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3388                 llctx = llog_get_context(&obd->obd_llogs,
3389                                          LLOG_CONFIG_ORIG_CTXT);
3390                 class_config_process_llog(llctx, cln_prof, &cfg);
3391                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3392
3393                 OBD_FREE(cln_prof, len);
3394                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3395                 mds->mds_profile = NULL;
3396         }
3397         RETURN(0);
3398 }
3399
3400 int mds_md_clean(struct obd_device *obd)
3401 {
3402         struct mds_obd *mds = &obd->u.mds;
3403         ENTRY;
3404
3405         if (mds->mds_md_name) {
3406                 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3407                 mds->mds_md_name = NULL;
3408         }
3409         RETURN(0);
3410 }
3411
3412 static int mds_precleanup(struct obd_device *obd, int flags)
3413 {
3414         int rc = 0;
3415         ENTRY;
3416
3417         mds_md_clean(obd);
3418         mds_dt_disconnect(obd, flags);
3419         mds_dt_clean(obd);
3420         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3421         RETURN(rc);
3422 }
3423
3424 extern void lgss_svc_cache_purge_all(void);
3425 static int mds_cleanup(struct obd_device *obd, int flags)
3426 {
3427         struct mds_obd *mds = &obd->u.mds;
3428         ENTRY;
3429
3430         if (mds->mds_sb == NULL)
3431                 RETURN(0);
3432
3433         mds_update_server_data(obd, 1);
3434         mds_update_last_fid(obd, NULL, 1);
3435
3436         if (mds->mds_dt_objids != NULL) {
3437                 int size = mds->mds_dt_desc.ld_tgt_count *
3438                         sizeof(obd_id);
3439                 OBD_FREE(mds->mds_dt_objids, size);
3440         }
3441         mds_fs_cleanup(obd, flags);
3442
3443         unlock_kernel();
3444
3445         /* 2 seems normal on mds, (may_umount() also expects 2
3446           fwiw), but we only see 1 at this point in obdfilter. */
3447         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3448
3449         mds->mds_sb = 0;
3450
3451         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3452
3453         spin_lock_bh(&obd->obd_processing_task_lock);
3454         if (obd->obd_recovering) {
3455                 target_cancel_recovery_timer(obd);
3456                 obd->obd_recovering = 0;
3457         }
3458         spin_unlock_bh(&obd->obd_processing_task_lock);
3459
3460         lock_kernel();
3461         fsfilt_put_ops(obd->obd_fsops);
3462
3463 #ifdef ENABLE_GSS
3464         /* XXX */
3465         lgss_svc_cache_purge_all();
3466 #endif
3467         RETURN(0);
3468 }
3469
3470 static int set_security(const char *value, char **sec)
3471 {
3472         int rc = 0;
3473
3474         if (!strcmp(value, "null"))
3475                 *sec = "null";
3476         else if (!strcmp(value, "krb5i"))
3477                 *sec = "krb5i";
3478         else if (!strcmp(value, "krb5p"))
3479                 *sec = "krb5p";
3480         else {
3481                 CERROR("Unrecognized value, force use NULL\n");
3482                 rc = -EINVAL;
3483         }
3484
3485         return rc;
3486 }
3487
3488 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3489 {
3490         struct lustre_cfg *lcfg = buf;
3491         struct mds_obd *mds = &obd->u.mds;
3492         int rc = 0;
3493         ENTRY;
3494
3495         switch(lcfg->lcfg_command) {
3496         case LCFG_SET_SECURITY: {
3497                 if (!lcfg->lcfg_inllen1 || !lcfg->lcfg_inllen2)
3498                         GOTO(out, rc = -EINVAL);
3499
3500                 if (!strcmp(lcfg->lcfg_inlbuf1, "mds_mds_sec"))
3501                         rc = set_security(lcfg->lcfg_inlbuf2,
3502                                           &mds->mds_mds_sec);
3503                 else if (!strcmp(lcfg->lcfg_inlbuf1, "mds_ost_sec"))
3504                         rc = set_security(lcfg->lcfg_inlbuf2,
3505                                           &mds->mds_ost_sec);
3506                 else {
3507                         CERROR("Unrecognized key\n");
3508                         rc = -EINVAL;
3509                 }
3510                 break;
3511         }
3512         default: {
3513                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3514                 GOTO(out, rc = -EINVAL);
3515
3516         }
3517         }
3518 out:
3519         RETURN(rc);
3520 }
3521
3522 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3523                                         int offset,
3524                                         struct ldlm_lock *new_lock,
3525                                         struct ldlm_lock **old_lock,
3526                                         struct lustre_handle *lockh)
3527 {
3528         struct obd_export *exp = req->rq_export;
3529         struct obd_device *obd = exp->exp_obd;
3530         struct ldlm_request *dlmreq =
3531                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3532         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3533         struct list_head *iter;
3534
3535         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3536                 return;
3537
3538         l_lock(&obd->obd_namespace->ns_lock);
3539         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3540                 struct ldlm_lock *lock;
3541                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3542                 if (lock == new_lock)
3543                         continue;
3544                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3545                         lockh->cookie = lock->l_handle.h_cookie;
3546                         LDLM_DEBUG(lock, "restoring lock cookie");
3547                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3548                                   lockh->cookie);
3549                         if (old_lock)
3550                                 *old_lock = LDLM_LOCK_GET(lock);
3551                         l_unlock(&obd->obd_namespace->ns_lock);
3552                         return;
3553                 }
3554         }
3555         l_unlock(&obd->obd_namespace->ns_lock);
3556
3557         /* If the xid matches, then we know this is a resent request,
3558          * and allow it. (It's probably an OPEN, for which we don't
3559          * send a lock */
3560         if (req->rq_xid ==
3561             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3562                 return;
3563
3564         if (req->rq_xid ==
3565             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3566                 return;
3567
3568         /* This remote handle isn't enqueued, so we never received or
3569          * processed this request.  Clear MSG_RESENT, because it can
3570          * be handled like any normal request now. */
3571
3572         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3573
3574         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3575                   remote_hdl.cookie);
3576 }
3577
3578 int intent_disposition(struct ldlm_reply *rep, int flag)
3579 {
3580         if (!rep)
3581                 return 0;
3582         return (rep->lock_policy_res1 & flag);
3583 }
3584
3585 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3586 {
3587         if (!rep)
3588                 return;
3589         rep->lock_policy_res1 |= flag;
3590 }
3591
3592 static int mds_intent_policy(struct ldlm_namespace *ns,
3593                              struct ldlm_lock **lockp, void *req_cookie,
3594                              ldlm_mode_t mode, int flags, void *data)
3595 {
3596         struct ptlrpc_request *req = req_cookie;
3597         struct ldlm_lock *lock = *lockp;
3598         struct ldlm_intent *it;
3599         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3600         struct ldlm_reply *rep;
3601         struct lustre_handle lockh[2] = {{0}, {0}};
3602         struct ldlm_lock *new_lock = NULL;
3603         int getattr_part = MDS_INODELOCK_UPDATE;
3604         int rc, reply_buffers;
3605         int repsize[5] = {sizeof(struct ldlm_reply),
3606                           sizeof(struct mds_body),
3607                           mds->mds_max_mdsize};
3608
3609         int offset = MDS_REQ_INTENT_REC_OFF;
3610         ENTRY;
3611
3612         LASSERT(req != NULL);
3613         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3614
3615         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3616                 /* No intent was provided */
3617                 int size = sizeof(struct ldlm_reply);
3618                 rc = lustre_pack_reply(req, 1, &size, NULL);
3619                 LASSERT(rc == 0);
3620                 RETURN(0);
3621         }
3622
3623         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3624                                 lustre_swab_ldlm_intent);
3625         if (it == NULL) {
3626                 CERROR("Intent missing\n");
3627                 RETURN(req->rq_status = -EFAULT);
3628         }
3629
3630         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3631
3632         reply_buffers = 3;
3633         if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3634                 reply_buffers = 5;
3635                 repsize[3] = 4;
3636                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
3637         }
3638
3639         rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
3640         if (rc)
3641                 RETURN(req->rq_status = rc);
3642
3643         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3644         LASSERT(rep != NULL);
3645
3646         intent_set_disposition(rep, DISP_IT_EXECD);
3647
3648         /* execute policy */
3649         switch ((long)it->opc) {
3650         case IT_OPEN:
3651         case IT_CREAT|IT_OPEN:
3652                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3653                                             lock, NULL, lockh);
3654                 /* XXX swab here to assert that an mds_open reint
3655                  * packet is following */
3656                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3657                                             lock, NULL, lockh);
3658                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3659
3660                 if (rep->lock_policy_res2) {
3661                         /*
3662                          * mds_open() returns ENOLCK where it should return
3663                          * zero, but it has no lock to return.
3664                          */
3665                         if (rep->lock_policy_res2 == ENOLCK)
3666                                 rep->lock_policy_res2 = 0;
3667
3668                         RETURN(ELDLM_LOCK_ABORTED);
3669                 }
3670
3671                 /*
3672                  * IT_OPEN may return lock on cross-node dentry that we want to
3673                  * hold during attr retrival -bzzz
3674                  */
3675                 if (lockh[0].cookie == 0)
3676                         RETURN(ELDLM_LOCK_ABORTED);
3677
3678                 break;
3679         case IT_LOOKUP:
3680                 getattr_part = MDS_INODELOCK_LOOKUP;
3681         case IT_CHDIR:
3682         case IT_GETATTR:
3683                 getattr_part |= MDS_INODELOCK_LOOKUP;
3684         case IT_READDIR:
3685                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3686                                             lock, &new_lock, lockh);
3687                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3688                                                          getattr_part);
3689                 /* FIXME: LDLM can set req->rq_status. MDS sets
3690                    policy_res{1,2} with disposition and status.
3691                    - replay: returns 0 & req->status is old status
3692                    - otherwise: returns req->status */
3693                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3694                         rep->lock_policy_res2 = 0;
3695                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3696                     rep->lock_policy_res2)
3697                         RETURN(ELDLM_LOCK_ABORTED);
3698                 if (req->rq_status != 0) {
3699                         LBUG();
3700                         rep->lock_policy_res2 = req->rq_status;
3701                         RETURN(ELDLM_LOCK_ABORTED);
3702                 }
3703                 break;
3704         case IT_UNLINK:
3705                 rc = mds_lock_and_check_slave(offset, req, lockh);
3706                 if ((rep->lock_policy_res2 = rc)) {
3707                         if (rc == ENOLCK)
3708                                 rep->lock_policy_res2 = 0;
3709                         RETURN(ELDLM_LOCK_ABORTED);
3710                 }
3711                 break;
3712         default:
3713                 CERROR("Unhandled intent "LPD64"\n", it->opc);
3714                 LBUG();
3715         }
3716
3717         /* By this point, whatever function we called above must have either
3718          * filled in 'lockh', been an intent replay, or returned an error.  We
3719          * want to allow replayed RPCs to not get a lock, since we would just
3720          * drop it below anyways because lock replay is done separately by the
3721          * client afterwards.  For regular RPCs we want to give the new lock to
3722          * the client instead of whatever lock it was about to get. */
3723         if (new_lock == NULL)
3724                 new_lock = ldlm_handle2lock(&lockh[0]);
3725         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3726                 RETURN(0);
3727
3728         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3729                  it->opc, lockh[0].cookie);
3730
3731         /* If we've already given this lock to a client once, then we should
3732          * have no readers or writers.  Otherwise, we should have one reader
3733          * _or_ writer ref (which will be zeroed below) before returning the
3734          * lock to a client. */
3735         if (new_lock->l_export == req->rq_export) {
3736                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3737         } else {
3738                 LASSERT(new_lock->l_export == NULL);
3739                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3740         }
3741
3742         *lockp = new_lock;
3743
3744         if (new_lock->l_export == req->rq_export) {
3745                 /* Already gave this to the client, which means that we
3746                  * reconstructed a reply. */
3747                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3748                         MSG_RESENT);
3749                 RETURN(ELDLM_LOCK_REPLACED);
3750         }
3751
3752         /* Fixup the lock to be given to the client */
3753         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3754         new_lock->l_readers = 0;
3755         new_lock->l_writers = 0;
3756
3757         new_lock->l_export = class_export_get(req->rq_export);
3758         list_add(&new_lock->l_export_chain,
3759                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
3760
3761         new_lock->l_blocking_ast = lock->l_blocking_ast;
3762         new_lock->l_completion_ast = lock->l_completion_ast;
3763
3764         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3765                sizeof(lock->l_remote_handle));
3766
3767         new_lock->l_flags &= ~LDLM_FL_LOCAL;
3768
3769         LDLM_LOCK_PUT(new_lock);
3770         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3771
3772         RETURN(ELDLM_LOCK_REPLACED);
3773 }
3774
3775 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3776 {
3777         struct lprocfs_static_vars lvars;
3778         int rc = 0;
3779
3780         lprocfs_init_multi_vars(0, &lvars);
3781
3782         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3783         if (rc)
3784                 return rc;
3785
3786         return lprocfs_alloc_md_stats(dev, 0);
3787 }
3788
3789 int mds_detach(struct obd_device *dev)
3790 {
3791         lprocfs_free_md_stats(dev);
3792         return lprocfs_obd_detach(dev);
3793 }
3794
3795 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3796 {
3797         struct lprocfs_static_vars lvars;
3798
3799         lprocfs_init_multi_vars(1, &lvars);
3800         return lprocfs_obd_attach(dev, lvars.obd_vars);
3801 }
3802
3803 int mdt_detach(struct obd_device *dev)
3804 {
3805         return lprocfs_obd_detach(dev);
3806 }
3807
3808 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3809 {
3810         struct mds_obd *mds = &obd->u.mds;
3811         int rc = 0;
3812         ENTRY;
3813
3814         mds->mds_service =
3815                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3816                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3817                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3818                                 mds_handle, "mds", obd->obd_proc_entry);
3819
3820         if (!mds->mds_service) {
3821                 CERROR("failed to start service\n");
3822                 RETURN(-ENOMEM);
3823         }
3824
3825         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3826                                     "ll_mdt");
3827         if (rc)
3828                 GOTO(err_thread, rc);
3829
3830         mds->mds_setattr_service =
3831                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3832                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3833                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3834                                 mds_handle, "mds_setattr",
3835                                 obd->obd_proc_entry);
3836         if (!mds->mds_setattr_service) {
3837                 CERROR("failed to start getattr service\n");
3838                 GOTO(err_thread, rc = -ENOMEM);
3839         }
3840
3841         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3842                                     MDT_NUM_THREADS, "ll_mdt_attr");
3843         if (rc)
3844                 GOTO(err_thread2, rc);
3845
3846         mds->mds_readpage_service =
3847                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3848                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3849                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3850                                 mds_handle, "mds_readpage",
3851                                 obd->obd_proc_entry);
3852         if (!mds->mds_readpage_service) {
3853                 CERROR("failed to start readpage service\n");
3854                 GOTO(err_thread2, rc = -ENOMEM);
3855         }
3856
3857         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3858                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
3859
3860         if (rc)
3861                 GOTO(err_thread3, rc);
3862
3863         RETURN(0);
3864
3865 err_thread3:
3866         ptlrpc_unregister_service(mds->mds_readpage_service);
3867 err_thread2:
3868         ptlrpc_unregister_service(mds->mds_setattr_service);
3869 err_thread:
3870         ptlrpc_unregister_service(mds->mds_service);
3871         return rc;
3872 }
3873
3874 static int mdt_cleanup(struct obd_device *obd, int flags)
3875 {
3876         struct mds_obd *mds = &obd->u.mds;
3877         ENTRY;
3878
3879         ptlrpc_stop_all_threads(mds->mds_readpage_service);
3880         ptlrpc_unregister_service(mds->mds_readpage_service);
3881
3882         ptlrpc_stop_all_threads(mds->mds_setattr_service);
3883         ptlrpc_unregister_service(mds->mds_setattr_service);
3884
3885         ptlrpc_stop_all_threads(mds->mds_service);
3886         ptlrpc_unregister_service(mds->mds_service);
3887
3888         RETURN(0);
3889 }
3890
3891 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3892                                          __u64 gr, void *data)
3893 {
3894         struct lustre_id id;
3895         struct obd_device *obd = data;
3896
3897         id_ino(&id) = ino;
3898         id_gen(&id) = gen;
3899         return mds_id2dentry(obd, &id, NULL);
3900 }
3901
3902 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3903                         void *key, __u32 *valsize, void *val)
3904 {
3905         struct obd_device *obd;
3906         struct mds_obd *mds;
3907         ENTRY;
3908
3909         obd = class_exp2obd(exp);
3910         mds = &obd->u.mds;
3911
3912         if (obd == NULL) {
3913                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3914                        exp->exp_handle.h_cookie);
3915                 RETURN(-EINVAL);
3916         }
3917
3918         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3919                 /* get log_context handle. */
3920                 unsigned long *llh_handle = val;
3921                 *valsize = sizeof(unsigned long);
3922                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3923                 RETURN(0);
3924         }
3925         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3926                 /* get log_context handle. */
3927                 unsigned long *sb = val;
3928                 *valsize = sizeof(unsigned long);
3929                 *sb = (unsigned long)obd->u.mds.mds_sb;
3930                 RETURN(0);
3931         }
3932
3933         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3934                 __u32 *mdsize = val;
3935                 *valsize = sizeof(*mdsize);
3936                 *mdsize = mds->mds_max_mdsize;
3937                 RETURN(0);
3938         }
3939
3940         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3941                 __u32 *mdsnum = val;
3942                 *valsize = sizeof(*mdsnum);
3943                 *mdsnum = mds->mds_num;
3944                 RETURN(0);
3945         }
3946
3947         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3948                 struct lustre_id *rootid = val;
3949                 *valsize = sizeof(struct lustre_id);
3950                 *rootid = mds->mds_rootid;
3951                 RETURN(0);
3952         }
3953
3954         CDEBUG(D_IOCTL, "invalid key\n");
3955         RETURN(-EINVAL);
3956
3957 }
3958 struct lvfs_callback_ops mds_lvfs_ops = {
3959         l_id2dentry:     mds_lvfs_id2dentry,
3960 };
3961
3962 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3963                 int objcount, struct obd_ioobj *obj,
3964                 int niocount, struct niobuf_remote *nb,
3965                 struct niobuf_local *res,
3966                 struct obd_trans_info *oti);
3967
3968 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
3969                  int objcount, struct obd_ioobj *obj, int niocount,
3970                  struct niobuf_local *res, struct obd_trans_info *oti,
3971                  int rc);
3972
3973 /* use obd ops to offer management infrastructure */
3974 static struct obd_ops mds_obd_ops = {
3975         .o_owner           = THIS_MODULE,
3976         .o_attach          = mds_attach,
3977         .o_detach          = mds_detach,
3978         .o_connect         = mds_connect,
3979         .o_connect_post    = mds_connect_post,
3980         .o_init_export     = mds_init_export,
3981         .o_destroy_export  = mds_destroy_export,
3982         .o_disconnect      = mds_disconnect,
3983         .o_setup           = mds_setup,
3984         .o_precleanup      = mds_precleanup,
3985         .o_cleanup         = mds_cleanup,
3986         .o_process_config  = mds_process_config,
3987         .o_postrecov       = mds_postrecov,
3988         .o_statfs          = mds_obd_statfs,
3989         .o_iocontrol       = mds_iocontrol,
3990         .o_create          = mds_obd_create,
3991         .o_destroy         = mds_obd_destroy,
3992         .o_llog_init       = mds_llog_init,
3993         .o_llog_finish     = mds_llog_finish,
3994         .o_notify          = mds_notify,
3995         .o_get_info        = mds_get_info,
3996         .o_set_info        = mds_set_info,
3997         .o_preprw          = mds_preprw,
3998         .o_commitrw        = mds_commitrw,
3999 };
4000
4001 static struct obd_ops mdt_obd_ops = {
4002         .o_owner           = THIS_MODULE,
4003         .o_attach          = mdt_attach,
4004         .o_detach          = mdt_detach,
4005         .o_setup           = mdt_setup,
4006         .o_cleanup         = mdt_cleanup,
4007 };
4008
4009 static int __init mds_init(void)
4010 {
4011         struct lprocfs_static_vars lvars;
4012
4013         mds_init_lsd_cache();
4014
4015         lprocfs_init_multi_vars(0, &lvars);
4016         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4017                             LUSTRE_MDS_NAME);
4018         lprocfs_init_multi_vars(1, &lvars);
4019         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4020                             LUSTRE_MDT_NAME);
4021
4022         return 0;
4023 }
4024
4025 static void /*__exit*/ mds_exit(void)
4026 {
4027         mds_cleanup_lsd_cache();
4028
4029         class_unregister_type(LUSTRE_MDS_NAME);
4030         class_unregister_type(LUSTRE_MDT_NAME);
4031 }
4032
4033 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4034 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4035 MODULE_LICENSE("GPL");
4036
4037 module_init(mds_init);
4038 module_exit(mds_exit);