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 initial,
 566                             unsigned long flags)
 567 {
 568         struct obd_device *obd = exp->exp_obd;
 569         struct mds_obd *mds = &obd->u.mds;
 570         struct mds_export_data *med;
 571         struct mds_client_data *mcd;
 572         int rc = 0;
 573         ENTRY;
 574
 575         med = &exp->exp_mds_data;
 576         mcd = med->med_mcd;
 577
 578         if (initial) {
 579                 /* some one reconnect initially, we have to reset
 580                  * data existing export can have. bug 6102 */
 581                 if (mcd->mcd_last_xid != 0)
 582                         CDEBUG(D_HA, "initial reconnect to existing export\n");
 583                 mcd->mcd_last_transno = 0;
 584                 mcd->mcd_last_xid = 0;
 585                 mcd->mcd_last_result = 0;
 586                 mcd->mcd_last_data = 0;
 587         }
 588
 589         if (!(flags & OBD_OPT_MDS_CONNECTION)) {
 590                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 591                         atomic_inc(&mds->mds_real_clients);
 592                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 593                                obd->obd_name, exp->exp_client_uuid.uuid,
 594                                atomic_read(&mds->mds_real_clients));
 595                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 596                 }
 597                 if (mds->mds_md_name)
 598                         rc = mds_md_connect(obd, mds->mds_md_name);
 599         }
 600         RETURN(rc);
 601 }
 602
 603 static int mds_init_export(struct obd_export *exp)
 604 {
 605         struct mds_export_data *med = &exp->exp_mds_data;
 606
 607         INIT_LIST_HEAD(&med->med_open_head);
 608         spin_lock_init(&med->med_open_lock);
 609         return 0;
 610 }
 611
 612 static int mds_destroy_export(struct obd_export *export)
 613 {
 614         struct obd_device *obd = export->exp_obd;
 615         struct mds_export_data *med = &export->exp_mds_data;
 616         struct lvfs_run_ctxt saved;
 617         int rc = 0;
 618         ENTRY;
 619
 620         mds_free_export_data(med);
 621         target_destroy_export(export);
 622
 623         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 624                 GOTO(out, 0);
 625
 626         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 627
 628         /* Close any open files (which may also cause orphan unlinking). */
 629         spin_lock(&med->med_open_lock);
 630         while (!list_empty(&med->med_open_head)) {
 631                 struct list_head *tmp = med->med_open_head.next;
 632                 struct mds_file_data *mfd =
 633                         list_entry(tmp, struct mds_file_data, mfd_list);
 634                 struct lustre_id sid;
 635
 636                 BDEVNAME_DECLARE_STORAGE(btmp);
 637
 638                 /* bug 1579: fix force-closing for 2.5 */
 639                 struct dentry *dentry = mfd->mfd_dentry;
 640
 641                 list_del(&mfd->mfd_list);
 642                 spin_unlock(&med->med_open_lock);
 643
 644                 down(&dentry->d_inode->i_sem);
 645                 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
 646                 up(&dentry->d_inode->i_sem);
 647                 if (rc) {
 648                         CERROR("Can't read inode self id, inode %lu, "
 649                                "rc %d\n", dentry->d_inode->i_ino, rc);
 650                         memset(&sid, 0, sizeof(sid));
 651                 }
 652
 653                 /* If you change this message, be sure to update
 654                  * replay_single:test_46 */
 655                 CERROR("force closing client file handle for %.*s (%s:"
 656                        DLID4")\n", dentry->d_name.len, dentry->d_name.name,
 657                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 658                        OLID4(&sid));
 659
 660                 /* child inode->i_alloc_sem protects orphan_dec_test and
 661                  * is_orphan race, mds_mfd_close drops it */
 662                 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 663                 rc = mds_mfd_close(NULL, 0, obd, mfd,
 664                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 665                 if (rc)
 666                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 667                 spin_lock(&med->med_open_lock);
 668         }
 669         spin_unlock(&med->med_open_lock);
 670         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 671
 672         EXIT;
 673 out:
 674         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 675         return rc;
 676 }
 677
 678 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
 679 {
 680         unsigned long irqflags;
 681         struct obd_device *obd;
 682         struct mds_obd *mds;
 683         int rc;
 684         ENTRY;
 685
 686         LASSERT(exp != NULL);
 687         obd = class_exp2obd(exp);
 688         if (obd == NULL) {
 689                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 690                        exp->exp_handle.h_cookie);
 691                 RETURN(-EINVAL);
 692         }
 693         mds = &obd->u.mds;
 694
 695         /*
 696          * suppress any inter-mds requests durring disconnecting lmv if this is
 697          * detected --force mode. This is needed to avoid endless recovery.
 698          */
 699         if (atomic_read(&mds->mds_real_clients) > 0 &&
 700             !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
 701                 flags |= OBD_OPT_FORCE;
 702
 703         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 704             && !atomic_read(&mds->mds_real_clients)) {
 705                 /* there was no client at all */
 706                 mds_md_disconnect(obd, flags);
 707         }
 708
 709         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 710             && atomic_dec_and_test(&mds->mds_real_clients)) {
 711                 /* time to drop LMV connections */
 712                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 713                        "Disconnnect from LMV now\n",
 714                        obd->obd_name, exp->exp_client_uuid.uuid);
 715                 mds_md_disconnect(obd, flags);
 716         }
 717
 718         spin_lock_irqsave(&exp->exp_lock, irqflags);
 719         exp->exp_flags = flags;
 720         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 721
 722         /* disconnect early so that clients can't keep using export */
 723         rc = class_disconnect(exp, flags);
 724         ldlm_cancel_locks_for_export(exp);
 725
 726         /* complete all outstanding replies */
 727         spin_lock_irqsave(&exp->exp_lock, irqflags);
 728         while (!list_empty(&exp->exp_outstanding_replies)) {
 729                 struct ptlrpc_reply_state *rs =
 730                         list_entry(exp->exp_outstanding_replies.next,
 731                                    struct ptlrpc_reply_state, rs_exp_list);
 732                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 733
 734                 spin_lock(&svc->srv_lock);
 735                 list_del_init(&rs->rs_exp_list);
 736                 ptlrpc_schedule_difficult_reply(rs);
 737                 spin_unlock(&svc->srv_lock);
 738         }
 739         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 740         RETURN(rc);
 741 }
 742
 743 static int mds_getstatus(struct ptlrpc_request *req)
 744 {
 745         struct mds_obd *mds = mds_req2mds(req);
 746         struct mds_body *body;
 747         int rc, size;
 748         ENTRY;
 749
 750         size = sizeof(*body);
 751
 752         rc = lustre_pack_reply(req, 1, &size, NULL);
 753         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 754                 CERROR("mds: out of memory for message: size=%d\n", size);
 755                 req->rq_status = -ENOMEM;       /* superfluous? */
 756                 RETURN(-ENOMEM);
 757         }
 758
 759         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 760         body->valid |= OBD_MD_FID;
 761
 762         memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
 763
 764         /*
 765          * the last_committed and last_xid fields are filled in for all replies
 766          * already - no need to do so here also.
 767          */
 768         RETURN(0);
 769 }
 770
 771 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 772                      void *data, int flag)
 773 {
 774         int do_ast;
 775         ENTRY;
 776
 777         if (flag == LDLM_CB_CANCELING) {
 778                 /* Don't need to do anything here. */
 779                 RETURN(0);
 780         }
 781
 782         /* XXX layering violation!  -phil */
 783         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 784
 785         /*
 786          * get this: if mds_blocking_ast is racing with mds_intent_policy, such
 787          * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
 788          * then by the time we get the lock, we might not be the correct
 789          * blocking function anymore.  So check, and return early, if so.
 790          */
 791         if (lock->l_blocking_ast != mds_blocking_ast) {
 792                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 793                 RETURN(0);
 794         }
 795
 796         lock->l_flags |= LDLM_FL_CBPENDING;
 797         do_ast = (!lock->l_readers && !lock->l_writers);
 798         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 799
 800         if (do_ast) {
 801                 struct lustre_handle lockh;
 802                 int rc;
 803
 804                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 805                 ldlm_lock2handle(lock, &lockh);
 806                 rc = ldlm_cli_cancel(&lockh);
 807                 if (rc < 0)
 808                         CERROR("ldlm_cli_cancel: %d\n", rc);
 809         } else {
 810                 LDLM_DEBUG(lock, "Lock still has references, will be "
 811                            "cancelled later");
 812         }
 813         RETURN(0);
 814 }
 815
 816 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
 817                           void *md, int size, int mea)
 818 {
 819         int rc = size;
 820
 821         if (S_ISREG(inode->i_mode)) {
 822                 rc = mds_convert_lov_ea(obd, inode, md, size);
 823         } else if (S_ISDIR(inode->i_mode)) {
 824                 if (mea) {
 825                         rc = mds_convert_mea_ea(obd, inode, md, size);
 826                 } else {
 827                         rc = mds_convert_lov_ea(obd, inode, md, size);
 828                 }
 829                 if (rc == -EINVAL) {
 830                         CERROR("Invalid EA format (nor LOV or MEA) "
 831                                "is detected. Inode %lu/%u\n",
 832                                inode->i_ino, inode->i_generation);
 833                 }
 834         }
 835         return rc;
 836 }
 837
 838 int mds_get_md(struct obd_device *obd, struct inode *inode,
 839                void *md, int *size, int lock, int mea)
 840 {
 841         int lmm_size;
 842         int rc = 0;
 843         ENTRY;
 844
 845         if (lock)
 846                 down(&inode->i_sem);
 847
 848         rc = fsfilt_get_md(obd, inode, md, *size,
 849                            (mea ? EA_MEA : EA_LOV));
 850         if (rc < 0) {
 851                 CERROR("Error %d reading eadata for ino %lu\n",
 852                        rc, inode->i_ino);
 853         } else if (rc > 0) {
 854                 lmm_size = rc;
 855                 rc = mds_convert_md(obd, inode, md,
 856                                     lmm_size, mea);
 857                 if (rc == 0) {
 858                         *size = lmm_size;
 859                         rc = lmm_size;
 860                 } else if (rc > 0) {
 861                         *size = rc;
 862                 }
 863         }
 864         if (lock)
 865                 up(&inode->i_sem);
 866
 867         RETURN(rc);
 868 }
 869
 870
 871 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 872  * Call with lock=0 if the caller has already taken the i_sem. */
 873 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 874                 struct mds_body *body, struct inode *inode, int lock, int mea)
 875 {
 876         struct mds_obd *mds = &obd->u.mds;
 877         void *lmm;
 878         int lmm_size;
 879         int rc;
 880         ENTRY;
 881
 882         lmm = lustre_msg_buf(msg, offset, 0);
 883         if (lmm == NULL) {
 884                 /* Some problem with getting eadata when I sized the reply
 885                  * buffer... */
 886                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 887                        inode->i_ino);
 888                 RETURN(0);
 889         }
 890         lmm_size = msg->buflens[offset];
 891
 892         /* I don't really like this, but it is a sanity check on the client
 893          * MD request.  However, if the client doesn't know how much space
 894          * to reserve for the MD, it shouldn't be bad to have too much space.
 895          */
 896         if (lmm_size > mds->mds_max_mdsize) {
 897                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 898                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 899                 // RETURN(-EINVAL);
 900         }
 901
 902         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
 903         if (rc > 0) {
 904                 if (S_ISDIR(inode->i_mode))
 905                         body->valid |= OBD_MD_FLDIREA;
 906                 else
 907                         body->valid |= OBD_MD_FLEASIZE;
 908
 909                 if (mea)
 910                         body->valid |= OBD_MD_MEA;
 911
 912                 body->eadatasize = lmm_size;
 913                 rc = 0;
 914         }
 915
 916         RETURN(rc);
 917 }
 918 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
 919                   struct mds_body *repbody, int reply_off)
 920 {
 921         struct inode *inode = dentry->d_inode;
 922         char *symname;
 923         int len, rc;
 924         ENTRY;
 925
 926         symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 927         LASSERT(symname != NULL);
 928         len = req->rq_repmsg->buflens[reply_off + 1];
 929
 930         rc = inode->i_op->readlink(dentry, symname, len);
 931         if (rc < 0) {
 932                 CERROR("readlink failed: %d\n", rc);
 933         } else if (rc != len - 1) {
 934                 CERROR ("Unexpected readlink rc %d: expecting %d\n",
 935                         rc, len - 1);
 936                 rc = -EINVAL;
 937         } else {
 938                 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 939                 repbody->valid |= OBD_MD_LINKNAME;
 940                 repbody->eadatasize = rc + 1;
 941                 symname[rc] = 0;        /* NULL terminate */
 942                 rc = 0;
 943         }
 944
 945         RETURN(rc);
 946 }
 947
 948 int mds_pack_ea(struct dentry *dentry, struct ptlrpc_request *req,
 949                 struct mds_body *repbody, int req_off, int reply_off)
 950 {
 951         struct inode *inode = dentry->d_inode;
 952         char *ea_name;
 953         void *value = NULL;
 954         int len, rc;
 955         ENTRY;
 956
 957         ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
 958         len = req->rq_repmsg->buflens[reply_off + 1];
 959         if (len != 0)
 960                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 961
 962         rc = -EOPNOTSUPP;
 963         if (inode->i_op && inode->i_op->getxattr)
 964                 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
 965
 966         if (rc < 0) {
 967                 if (rc != -ENODATA && rc != -EOPNOTSUPP)
 968                         CERROR("getxattr failed: %d", rc);
 969         } else {
 970                 repbody->valid |= OBD_MD_FLEA;
 971                 repbody->eadatasize = rc;
 972                 rc = 0;
 973         }
 974
 975         RETURN(rc);
 976 }
 977
 978 int mds_pack_ealist(struct dentry *dentry, struct ptlrpc_request *req,
 979                     struct mds_body *repbody, int reply_off)
 980 {
 981         struct inode *inode = dentry->d_inode;
 982         void *value = NULL;
 983         int len, rc;
 984         ENTRY;
 985
 986         len = req->rq_repmsg->buflens[reply_off + 1];
 987         if (len != 0)
 988                 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
 989
 990         rc = -EOPNOTSUPP;
 991         if (inode->i_op && inode->i_op->getxattr)
 992                 rc = inode->i_op->listxattr(dentry, value, len);
 993
 994         if (rc < 0) {
 995                 CERROR("listxattr failed: %d", rc);
 996         } else {
 997                 repbody->valid |= OBD_MD_FLEALIST;
 998                 repbody->eadatasize = rc;
 999                 rc = 0;
1000         }
1001         RETURN(rc);
1002 }
1003
1004 int mds_pack_acl(struct obd_device *obd, struct lustre_msg *repmsg, int offset,
1005                  struct mds_body *body, struct inode *inode)
1006 {
1007         struct dentry de = { .d_inode = inode };
1008         __u32 buflen, *sizep;
1009         void *buf;
1010         int size;
1011         ENTRY;
1012
1013         if (!inode->i_op->getxattr)
1014                 RETURN(0);
1015
1016         buflen = repmsg->buflens[offset + 1];
1017         buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1018
1019         size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1020         if (size == -ENODATA || size == -EOPNOTSUPP)
1021                 RETURN(0);
1022         if (size < 0)
1023                 RETURN(size);
1024         LASSERT(size);
1025
1026         sizep = lustre_msg_buf(repmsg, offset, 4);
1027         if (!sizep) {
1028                 CERROR("can't locate returned acl size buf\n");
1029                 RETURN(-EPROTO);
1030         }
1031
1032         *sizep = cpu_to_le32(size);
1033         body->valid |= OBD_MD_FLACL_ACCESS;
1034
1035         RETURN(0);
1036 }
1037
1038 /*
1039  * here we take simple rule: once uid/fsuid is root, we also squash
1040  * the gid/fsgid, don't care setuid/setgid attributes.
1041  */
1042 int mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
1043                     ptl_nid_t *peernid)
1044 {
1045         if (!mds->mds_squash_uid || *peernid == mds->mds_nosquash_nid)
1046                 return 0;
1047
1048         if (rsd->rsd_uid && rsd->rsd_fsuid)
1049                 return 0;
1050
1051         CDEBUG(D_SEC, "squash req from "LPX64":"
1052                "(%u:%u-%u:%u/%x)=>(%u:%u-%u:%u/%x)\n", *peernid,
1053                 rsd->rsd_uid, rsd->rsd_gid,
1054                 rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
1055                 rsd->rsd_uid ? rsd->rsd_uid : mds->mds_squash_uid,
1056                 rsd->rsd_uid ? rsd->rsd_gid : mds->mds_squash_gid,
1057                 rsd->rsd_fsuid ? rsd->rsd_fsuid : mds->mds_squash_uid,
1058                 rsd->rsd_fsuid ? rsd->rsd_fsgid : mds->mds_squash_gid,
1059                 rsd->rsd_cap & ~CAP_FS_MASK);
1060
1061         if (rsd->rsd_uid == 0) {
1062                 rsd->rsd_uid = mds->mds_squash_uid;
1063                 rsd->rsd_gid = mds->mds_squash_gid;
1064         }
1065         if (rsd->rsd_fsuid == 0) {
1066                 rsd->rsd_fsuid = mds->mds_squash_uid;
1067                 rsd->rsd_fsgid = mds->mds_squash_gid;
1068         }
1069         rsd->rsd_cap &= ~CAP_FS_MASK;
1070
1071         return 1;
1072 }
1073
1074 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1075                                 struct ptlrpc_request *req, int req_off,
1076                                 struct mds_body *reqbody, int reply_off)
1077 {
1078         struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1079         struct inode *inode = dentry->d_inode;
1080         struct mds_body *body;
1081         int rc = 0;
1082         ENTRY;
1083
1084         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1085                 RETURN(-ENOENT);
1086
1087         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1088         LASSERT(body != NULL);                 /* caller prepped reply */
1089
1090         if (dentry->d_flags & DCACHE_CROSS_REF) {
1091                 mds_pack_dentry2body(obd, body, dentry,
1092                                      (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1093                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1094                        OLID4(&body->id1));
1095                 RETURN(0);
1096         }
1097
1098         mds_pack_inode2body(obd, body, inode,
1099                             (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1100
1101         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1102             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1103
1104                 /* guessing what kind og attribute do we need. */
1105                 int is_mea = (S_ISDIR(inode->i_mode) &&
1106                     (reqbody->valid & OBD_MD_MEA) != 0);
1107
1108                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1109                                  body, inode, 1, is_mea);
1110
1111                 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1112                 if (!(body->valid & OBD_MD_FLEASIZE) &&
1113                     !(body->valid & OBD_MD_FLDIREA))
1114                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1115                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
1116         } else if (S_ISLNK(inode->i_mode) &&
1117                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1118                 rc = mds_pack_link(dentry, req, body, reply_off);
1119         } else if (reqbody->valid & OBD_MD_FLEA) {
1120                 rc = mds_pack_ea(dentry, req, body, req_off, reply_off);
1121         } else if (reqbody->valid & OBD_MD_FLEALIST) {
1122                 rc = mds_pack_ealist(dentry, req, body, reply_off);
1123         }
1124
1125         if (reqbody->valid & OBD_MD_FLACL_ACCESS) {
1126                 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1127                 rc = mds_pack_acl(obd, req->rq_repmsg, reply_off + inc,
1128                                   body, inode);
1129         }
1130
1131         if (rc == 0)
1132                 mds_body_do_reverse_map(med, body);
1133
1134         RETURN(rc);
1135 }
1136
1137 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1138                                    struct dentry *dentry,
1139                                    int offset)
1140 {
1141         int rc = 0, size[1] = {sizeof(struct mds_body)};
1142         ENTRY;
1143
1144         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1145                 CERROR("failed MDS_GETATTR_PACK test\n");
1146                 req->rq_status = -ENOMEM;
1147                 RETURN(-ENOMEM);
1148         }
1149
1150         rc = lustre_pack_reply(req, 1, size, NULL);
1151         if (rc) {
1152                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1153                 GOTO(out, req->rq_status = rc);
1154         }
1155
1156         EXIT;
1157 out:
1158         return rc;
1159 }
1160
1161 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1162                                 int offset)
1163 {
1164         struct inode *inode = de->d_inode;
1165         struct mds_obd *mds = mds_req2mds(req);
1166         struct mds_body *body;
1167         int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1168         ENTRY;
1169
1170         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1171         LASSERT(body != NULL);                 /* checked by caller */
1172         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
1173
1174         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1175             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1176                 int rc;
1177
1178                 down(&inode->i_sem);
1179                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1180                                    ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1181                 up(&inode->i_sem);
1182                 if (rc < 0) {
1183                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1184                                 CERROR("error getting inode %lu MD: rc = %d\n",
1185                                        inode->i_ino, rc);
1186                         size[bufcount] = 0;
1187                 } else if (rc > mds->mds_max_mdsize) {
1188                         size[bufcount] = 0;
1189                         CERROR("MD size %d larger than maximum possible %u\n",
1190                                rc, mds->mds_max_mdsize);
1191                 } else {
1192                         size[bufcount] = rc;
1193                 }
1194                 bufcount++;
1195         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1196                 if (inode->i_size + 1 != body->eadatasize)
1197                         CERROR("symlink size: %Lu, reply space: %d\n",
1198                                inode->i_size + 1, body->eadatasize);
1199                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1200                 bufcount++;
1201                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1202                        inode->i_size + 1, body->eadatasize);
1203         } else if ((body->valid & OBD_MD_FLEA)) {
1204                 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1205                                                   offset + 1, 0);
1206                 rc = -EOPNOTSUPP;
1207                 if (inode->i_op && inode->i_op->getxattr)
1208                         rc = inode->i_op->getxattr(de, ea_name, NULL, 0);
1209
1210                 if (rc < 0) {
1211                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1212                                 CERROR("error getting inode %lu EA: rc = %d\n",
1213                                        inode->i_ino, rc);
1214                         size[bufcount] = 0;
1215                 } else {
1216                         size[bufcount] = min_t(int, body->eadatasize, rc);
1217                 }
1218                 bufcount++;
1219         } else if (body->valid & OBD_MD_FLEALIST) {
1220                 rc = -EOPNOTSUPP;
1221                 if (inode->i_op && inode->i_op->getxattr)
1222                         rc = inode->i_op->listxattr(de, NULL, 0);
1223
1224                 if (rc < 0) {
1225                         if (rc != -ENODATA && rc != -EOPNOTSUPP)
1226                                 CERROR("error getting inode %lu EA: rc = %d\n",
1227                                        inode->i_ino, rc);
1228                         size[bufcount] = 0;
1229                 } else {
1230                         size[bufcount] = min_t(int, body->eadatasize, rc);
1231                 }
1232                 bufcount++;
1233         }
1234
1235         /* may co-exist with OBD_MD_FLEASIZE */
1236         if (body->valid & OBD_MD_FLACL_ACCESS) {
1237                 size[bufcount++] = 4;
1238                 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1239         }
1240
1241         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1242                 CERROR("failed MDS_GETATTR_PACK test\n");
1243                 req->rq_status = -ENOMEM;
1244                 GOTO(out, rc = -ENOMEM);
1245         }
1246
1247         rc = lustre_pack_reply(req, bufcount, size, NULL);
1248         if (rc) {
1249                 CERROR("out of memory\n");
1250                 GOTO(out, req->rq_status = rc);
1251         }
1252
1253         EXIT;
1254  out:
1255         return rc;
1256 }
1257
1258 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1259                       char *name, int namelen)
1260 {
1261         struct mea *mea = NULL;
1262         int mea_size, rc = 0;
1263         ENTRY;
1264
1265         rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1266         if (rc)
1267                 RETURN(rc);
1268         if (mea != NULL) {
1269                 /*
1270                  * dir is already splitted, check if requested filename should
1271                  * live at this MDS or at another one.
1272                  */
1273                 int i = mea_name2idx(mea, name, namelen - 1);
1274                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1275                         CDEBUG(D_OTHER,
1276                                "inapropriate MDS(%d) for %s. should be "
1277                                "%lu(%d)\n", mea->mea_master, name,
1278                                (unsigned long)id_group(&mea->mea_ids[i]), i);
1279                         rc = -ERESTART;
1280                 }
1281         }
1282
1283         if (mea)
1284                 OBD_FREE(mea, mea_size);
1285         RETURN(rc);
1286 }
1287
1288 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1289                             struct lustre_handle *child_lockh, int child_part)
1290 {
1291         struct obd_device *obd = req->rq_export->exp_obd;
1292         struct mds_obd *mds = &obd->u.mds;
1293         struct ldlm_reply *rep = NULL;
1294         struct lvfs_run_ctxt saved;
1295         struct mds_req_sec_desc *rsd;
1296         struct mds_body *body;
1297         struct dentry *dparent = NULL, *dchild = NULL;
1298         struct lvfs_ucred uc = {NULL, NULL,};
1299         struct lustre_handle parent_lockh[2] = {{0}, {0}};
1300         unsigned int namesize;
1301         int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1302         char *name = NULL;
1303         ENTRY;
1304
1305         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
1306         MD_COUNTER_INCREMENT(obd, getattr_lock);
1307
1308         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1309         if (!rsd) {
1310                 CERROR("Can't unpack security desc\n");
1311                 RETURN(-EFAULT);
1312         }
1313
1314         /* swab now, before anyone looks inside the request. */
1315         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1316                                   lustre_swab_mds_body);
1317         if (body == NULL) {
1318                 CERROR("Can't swab mds_body\n");
1319                 GOTO(cleanup, rc = -EFAULT);
1320         }
1321
1322         LASSERT_REQSWAB(req, offset + 1);
1323         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1324         if (name == NULL) {
1325                 CERROR("Can't unpack name\n");
1326                 GOTO(cleanup, rc = -EFAULT);
1327         }
1328         namesize = req->rq_reqmsg->buflens[offset + 1];
1329
1330         /* namesize less than 2 means we have empty name, probably came from
1331            revalidate by cfid, so no point in having name to be set */
1332         if (namesize <= 1)
1333                 name = NULL;
1334
1335         LASSERT (offset == 1 || offset == 3);
1336         if (offset == 3) {
1337                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1338                 reply_offset = 1;
1339         } else {
1340                 reply_offset = 0;
1341         }
1342
1343         rc = mds_init_ucred(&uc, req, rsd);
1344         if (rc) {
1345                 GOTO(cleanup, rc);
1346         }
1347
1348         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1349         cleanup_phase = 1; /* kernel context */
1350         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1351
1352         LASSERT(namesize > 0);
1353         if (child_lockh->cookie != 0) {
1354                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1355                 resent_req = 1;
1356         }
1357 #if 0
1358 #if HAVE_LOOKUP_RAW
1359         if (body->valid == OBD_MD_FLID) {
1360                 struct mds_body *mds_reply;
1361                 int size = sizeof(*mds_reply);
1362                 struct inode *dir;
1363                 ino_t inum;
1364
1365                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1366                 if (IS_ERR(dparent)) {
1367                         rc = PTR_ERR(dparent);
1368                         GOTO(cleanup, rc);
1369                 }
1370
1371                 /*
1372                  * the user requested ONLY the inode number, so do a raw lookup.
1373                  */
1374                 rc = lustre_pack_reply(req, 1, &size, NULL);
1375                 if (rc) {
1376                         CERROR("out of memory\n");
1377                         l_dput(dparent);
1378                         GOTO(cleanup, rc);
1379                 }
1380                 dir  = dparent->d_inode;
1381                 LASSERT(dir->i_op->lookup_raw != NULL);
1382                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1383                 l_dput(dparent);
1384                 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1385                                            sizeof(*mds_reply));
1386
1387                 id_ino(&mds_reply->id1) = inum;
1388                 mds_reply->valid = OBD_MD_FLID;
1389                 GOTO(cleanup, rc);
1390         }
1391 #endif
1392 #endif
1393         if (resent_req == 0) {
1394                 LASSERT(id_fid(&body->id1) != 0);
1395                 if (name) {
1396                         rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1397                                                          parent_lockh, &dparent,
1398                                                          LCK_PR,
1399                                                          MDS_INODELOCK_UPDATE,
1400                                                          &update_mode,
1401                                                          name, namesize,
1402                                                          child_lockh, &dchild,
1403                                                          LCK_PR, child_part);
1404                         if (rc)
1405                                 GOTO(cleanup, rc);
1406
1407                         cleanup_phase = 2; /* dchild, dparent, locks */
1408
1409                         /*
1410                          * let's make sure this name should leave on this mds
1411                          * node.
1412                          */
1413                         rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1414                         if (rc)
1415                                 GOTO(cleanup, rc);
1416                 } else {
1417                         /* we have no dentry here, drop LOOKUP bit */
1418                         /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1419                         child_part &= ~MDS_INODELOCK_LOOKUP;
1420                         CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1421                                obd->obd_name, OLID4(&body->id1));
1422
1423                         dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1424                                                       LCK_PR, parent_lockh,
1425                                                       &update_mode,
1426                                                       NULL, 0,
1427                                                       MDS_INODELOCK_UPDATE);
1428                         if (IS_ERR(dchild)) {
1429                                 CERROR("can't find inode with id "DLID4", err = %d\n",
1430                                        OLID4(&body->id1), (int)PTR_ERR(dchild));
1431                                 GOTO(cleanup, rc = PTR_ERR(dchild));
1432                         }
1433                         memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1434                 }
1435         } else {
1436                 struct ldlm_lock *granted_lock;
1437
1438                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1439                 granted_lock = ldlm_handle2lock(child_lockh);
1440
1441                 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1442                          id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1443                          child_lockh->cookie);
1444
1445                 if (name) {
1446                         /* usual named request */
1447                         dparent = mds_id2dentry(obd, &body->id1, NULL);
1448                         LASSERT(!IS_ERR(dparent));
1449                         dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1450                         if (IS_ERR(dchild)) {
1451                                 DEBUG_REQ(D_ERROR, req, "resent, not enqueuing new locks");
1452                                 CDEBUG(D_ERROR, "lock against [%lu:%lu]/%*s\n",
1453                                                 (unsigned long) id_ino(&body->id1),
1454                                                 (unsigned long) id_gen(&body->id1),
1455                                                 namesize - 1, name);
1456                         }
1457                         LASSERT(!IS_ERR(dchild));
1458                 } else {
1459                         /* client wants to get attr. by id */
1460                         dchild = mds_id2dentry(obd, &body->id1, NULL);
1461                         if (IS_ERR(dchild)) {
1462                                 DEBUG_REQ(D_ERROR, req, "resent, not enqueuing new locks");
1463                                 CDEBUG(D_ERROR, "lock against [%lu:%lu]\n",
1464                                                 (unsigned long) id_ino(&body->id1),
1465                                                 (unsigned long) id_gen(&body->id1));
1466                         }
1467                         LASSERT(!IS_ERR(dchild));
1468                 }
1469                 LDLM_LOCK_PUT(granted_lock);
1470         }
1471
1472         cleanup_phase = 2; /* dchild, dparent, locks */
1473
1474         if (!DENTRY_VALID(dchild)) {
1475                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1476                 /*
1477                  * in the intent case, the policy clears this error: the
1478                  * disposition is enough.
1479                  */
1480                 rc = -ENOENT;
1481                 GOTO(cleanup, rc);
1482         } else {
1483                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1484         }
1485
1486         if (req->rq_repmsg == NULL) {
1487                 if (dchild->d_flags & DCACHE_CROSS_REF)
1488                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1489                 else
1490                         rc = mds_getattr_pack_msg(req, dchild, offset);
1491                 if (rc != 0) {
1492                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1493                         GOTO (cleanup, rc);
1494                 }
1495         }
1496
1497         rc = mds_getattr_internal(obd, dchild, req, offset, body, reply_offset);
1498         GOTO(cleanup, rc); /* returns the lock to the client */
1499
1500  cleanup:
1501         switch (cleanup_phase) {
1502         case 2:
1503                 if (resent_req == 0) {
1504                         if (rc && DENTRY_VALID(dchild))
1505                                 ldlm_lock_decref(child_lockh, LCK_PR);
1506                         if (name)
1507                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1508 #ifdef S_PDIROPS
1509                         if (parent_lockh[1].cookie != 0)
1510                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1511 #endif
1512                         if (dparent)
1513                                 l_dput(dparent);
1514                 }
1515                 l_dput(dchild);
1516         case 1:
1517                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1518         default:
1519                 mds_exit_ucred(&uc);
1520         }
1521         return rc;
1522 }
1523
1524 static int mds_getattr(struct ptlrpc_request *req, int offset)
1525 {
1526         struct obd_device *obd = req->rq_export->exp_obd;
1527         struct lvfs_run_ctxt saved;
1528         struct dentry *de;
1529         struct mds_req_sec_desc *rsd;
1530         struct mds_body *body;
1531         struct lvfs_ucred uc = {NULL, NULL,};
1532         int rc = 0;
1533         ENTRY;
1534
1535         MD_COUNTER_INCREMENT(obd, getattr);
1536
1537         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1538         if (!rsd) {
1539                 CERROR("Can't unpack security desc\n");
1540                 RETURN(-EFAULT);
1541         }
1542
1543         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1544                                   lustre_swab_mds_body);
1545         if (body == NULL) {
1546                 CERROR ("Can't unpack body\n");
1547                 RETURN (-EFAULT);
1548         }
1549
1550         rc = mds_init_ucred(&uc, req, rsd);
1551         if (rc) {
1552                 mds_exit_ucred(&uc);
1553                 RETURN(rc);
1554         }
1555
1556         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1557         de = mds_id2dentry(obd, &body->id1, NULL);
1558         if (IS_ERR(de)) {
1559                 rc = req->rq_status = PTR_ERR(de);
1560                 GOTO(out_pop, rc);
1561         }
1562
1563         rc = mds_getattr_pack_msg(req, de, offset);
1564         if (rc != 0) {
1565                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1566                 GOTO(out_pop, rc);
1567         }
1568
1569         req->rq_status = mds_getattr_internal(obd, de, req, offset, body, 0);
1570         l_dput(de);
1571
1572         EXIT;
1573 out_pop:
1574         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1575         mds_exit_ucred(&uc);
1576         return rc;
1577 }
1578
1579 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1580                           unsigned long max_age)
1581 {
1582         int rc;
1583         ENTRY;
1584
1585         spin_lock(&obd->obd_osfs_lock);
1586         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1587         if (rc == 0)
1588                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1589         spin_unlock(&obd->obd_osfs_lock);
1590
1591         RETURN(rc);
1592 }
1593
1594 static int mds_statfs(struct ptlrpc_request *req)
1595 {
1596         struct obd_device *obd = req->rq_export->exp_obd;
1597         int rc, size = sizeof(struct obd_statfs);
1598         ENTRY;
1599
1600         /* This will trigger a watchdog timeout */
1601         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1602                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1603
1604         rc = lustre_pack_reply(req, 1, &size, NULL);
1605         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1606                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1607                 GOTO(out, rc);
1608         }
1609
1610         OBD_COUNTER_INCREMENT(obd, statfs);
1611
1612         /* We call this so that we can cache a bit - 1 jiffie worth */
1613         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1614                             jiffies - HZ);
1615         if (rc) {
1616                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1617                 GOTO(out, rc);
1618         }
1619
1620         EXIT;
1621 out:
1622         req->rq_status = rc;
1623         return rc;
1624 }
1625
1626 static int mds_sync(struct ptlrpc_request *req, int offset)
1627 {
1628         struct obd_device *obd = req->rq_export->exp_obd;
1629         struct mds_obd *mds = &obd->u.mds;
1630         struct mds_body *body;
1631         int rc, size = sizeof(*body);
1632         ENTRY;
1633
1634         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1635                                   lustre_swab_mds_body);
1636         if (body == NULL)
1637                 GOTO(out, rc = -EPROTO);
1638
1639         rc = lustre_pack_reply(req, 1, &size, NULL);
1640         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1641                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1642                 GOTO(out, rc);
1643         }
1644
1645         if (id_ino(&body->id1) == 0) {
1646                 /* an id of zero is taken to mean "sync whole filesystem" */
1647                 rc = fsfilt_sync(obd, mds->mds_sb);
1648                 if (rc)
1649                         GOTO(out, rc);
1650         } else {
1651                 /* just any file to grab fsync method - "file" arg unused */
1652                 struct file *file = mds->mds_rcvd_filp;
1653                 struct mds_body *rep_body;
1654                 struct dentry *de;
1655
1656                 de = mds_id2dentry(obd, &body->id1, NULL);
1657                 if (IS_ERR(de))
1658                         GOTO(out, rc = PTR_ERR(de));
1659
1660                 rc = file->f_op->fsync(NULL, de, 1);
1661                 if (rc)
1662                         GOTO(out, rc);
1663
1664                 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1665                 mds_pack_inode2body(obd, rep_body, de->d_inode,
1666                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1667                 l_dput(de);
1668         }
1669
1670         EXIT;
1671 out:
1672         req->rq_status = rc;
1673         return rc;
1674 }
1675
1676 /* mds_readpage does not take a DLM lock on the inode, because the client must
1677  * already have a PR lock.
1678  *
1679  * If we were to take another one here, a deadlock will result, if another
1680  * thread is already waiting for a PW lock. */
1681 static int mds_readpage(struct ptlrpc_request *req, int offset)
1682 {
1683         struct obd_device *obd = req->rq_export->exp_obd;
1684         struct vfsmount *mnt;
1685         struct dentry *de;
1686         struct file *file;
1687         struct mds_req_sec_desc *rsd;
1688         struct mds_body *body, *repbody;
1689         struct lvfs_run_ctxt saved;
1690         int rc, size = sizeof(*repbody);
1691         struct lvfs_ucred uc = {NULL, NULL,};
1692         ENTRY;
1693
1694         rc = lustre_pack_reply(req, 1, &size, NULL);
1695         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1696                 CERROR("mds: out of memory\n");
1697                 GOTO(out, rc = -ENOMEM);
1698         }
1699
1700         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1701         if (!rsd) {
1702                 CERROR("Can't unpack security desc\n");
1703                 GOTO (out, rc = -EFAULT);
1704         }
1705
1706         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1707                                   lustre_swab_mds_body);
1708         if (body == NULL) {
1709                 CERROR("Can't unpack body\n");
1710                 GOTO (out, rc = -EFAULT);
1711         }
1712
1713         rc = mds_init_ucred(&uc, req, rsd);
1714         if (rc) {
1715                 GOTO(out, rc);
1716         }
1717
1718         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1719         de = mds_id2dentry(obd, &body->id1, &mnt);
1720         if (IS_ERR(de))
1721                 GOTO(out_pop, rc = PTR_ERR(de));
1722
1723         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1724
1725         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1726         /* note: in case of an error, dentry_open puts dentry */
1727         if (IS_ERR(file))
1728                 GOTO(out_pop, rc = PTR_ERR(file));
1729
1730         /* body->size is actually the offset -eeb */
1731         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1732                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1733                        body->size, de->d_inode->i_blksize);
1734                 GOTO(out_file, rc = -EFAULT);
1735         }
1736
1737         /* body->nlink is actually the #bytes to read -eeb */
1738         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1739                 CERROR("size %u is not multiple of blocksize %lu\n",
1740                        body->nlink, de->d_inode->i_blksize);
1741                 GOTO(out_file, rc = -EFAULT);
1742         }
1743
1744         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1745         repbody->size = file->f_dentry->d_inode->i_size;
1746         repbody->valid = OBD_MD_FLSIZE;
1747
1748         /* to make this asynchronous make sure that the handling function
1749            doesn't send a reply when this function completes. Instead a
1750            callback function would send the reply */
1751         /* body->size is actually the offset -eeb */
1752         rc = mds_sendpage(req, file, body->size, body->nlink);
1753
1754         EXIT;
1755 out_file:
1756         filp_close(file, 0);
1757 out_pop:
1758         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1759 out:
1760         mds_exit_ucred(&uc);
1761         req->rq_status = rc;
1762         return 0;
1763 }
1764
1765 /* update master MDS ID, which is stored in local inode EA. */
1766 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1767                    void *data, int data_len)
1768 {
1769         struct mds_obd *mds = &obd->u.mds;
1770         struct dentry *dentry;
1771         void *handle;
1772         int rc = 0;
1773         ENTRY;
1774
1775         LASSERT(id);
1776         LASSERT(obd);
1777
1778         dentry = mds_id2dentry(obd, id, NULL);
1779         if (IS_ERR(dentry))
1780                 GOTO(out, rc = PTR_ERR(dentry));
1781
1782         if (!dentry->d_inode) {
1783                 CERROR("Can't find object "DLID4".\n",
1784                        OLID4(id));
1785                 GOTO(out_dentry, rc = -EINVAL);
1786         }
1787
1788         handle = fsfilt_start(obd, dentry->d_inode,
1789                               FSFILT_OP_SETATTR, NULL);
1790         if (IS_ERR(handle))
1791                 GOTO(out_dentry, rc = PTR_ERR(handle));
1792
1793         rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1794                                   (struct lustre_id *)data);
1795         if (rc) {
1796                 CERROR("Can't update inode "DLID4" master id, "
1797                        "error = %d.\n", OLID4(id), rc);
1798                 GOTO(out_commit, rc);
1799         }
1800
1801         EXIT;
1802 out_commit:
1803         fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1804                       handle, 0);
1805 out_dentry:
1806         l_dput(dentry);
1807 out:
1808         return rc;
1809 }
1810 EXPORT_SYMBOL(mds_update_mid);
1811
1812 /* read master MDS ID, which is stored in local inode EA. */
1813 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1814                  void *data, int data_len)
1815 {
1816         struct dentry *dentry;
1817         int rc = 0;
1818         ENTRY;
1819
1820         LASSERT(id);
1821         LASSERT(obd);
1822
1823         dentry = mds_id2dentry(obd, id, NULL);
1824         if (IS_ERR(dentry))
1825                 GOTO(out, rc = PTR_ERR(dentry));
1826
1827         if (!dentry->d_inode) {
1828                 CERROR("Can't find object "DLID4".\n",
1829                        OLID4(id));
1830                 GOTO(out_dentry, rc = -EINVAL);
1831         }
1832
1833         down(&dentry->d_inode->i_sem);
1834         rc = mds_read_inode_mid(obd, dentry->d_inode,
1835                                 (struct lustre_id *)data);
1836         up(&dentry->d_inode->i_sem);
1837         if (rc) {
1838                 CERROR("Can't read inode "DLID4" master id, "
1839                        "error = %d.\n", OLID4(id), rc);
1840                 GOTO(out_dentry, rc);
1841         }
1842
1843         EXIT;
1844 out_dentry:
1845         l_dput(dentry);
1846 out:
1847         return rc;
1848 }
1849 EXPORT_SYMBOL(mds_read_mid);
1850
1851 int mds_reint(struct ptlrpc_request *req, int offset,
1852               struct lustre_handle *lockh)
1853 {
1854         struct mds_update_record *rec;
1855         struct mds_req_sec_desc *rsd;
1856         int rc;
1857         ENTRY;
1858
1859         OBD_ALLOC(rec, sizeof(*rec));
1860         if (rec == NULL)
1861                 RETURN(-ENOMEM);
1862
1863         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1864         if (!rsd) {
1865                 CERROR("Can't unpack security desc\n");
1866                 GOTO(out, rc = -EFAULT);
1867         }
1868
1869         rc = mds_update_unpack(req, offset, rec);
1870         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1871                 CERROR("invalid record\n");
1872                 GOTO(out, req->rq_status = -EINVAL);
1873         }
1874
1875         rc = mds_init_ucred(&rec->ur_uc, req, rsd);
1876         if (rc) {
1877                 GOTO(out, rc);
1878         }
1879
1880         /* rc will be used to interrupt a for loop over multiple records */
1881         rc = mds_reint_rec(rec, offset, req, lockh);
1882
1883  out:
1884         mds_exit_ucred(&rec->ur_uc);
1885         OBD_FREE(rec, sizeof(*rec));
1886         RETURN(rc);
1887 }
1888
1889 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1890                                        struct obd_device *obd, int *process)
1891 {
1892         switch (req->rq_reqmsg->opc) {
1893         case MDS_CONNECT: /* This will never get here, but for completeness. */
1894         case OST_CONNECT: /* This will never get here, but for completeness. */
1895         case MDS_DISCONNECT:
1896         case OST_DISCONNECT:
1897                *process = 1;
1898                RETURN(0);
1899
1900         case MDS_CLOSE:
1901         case MDS_SYNC: /* used in unmounting */
1902         case OBD_PING:
1903         case MDS_REINT:
1904         case LDLM_ENQUEUE:
1905         case OST_CREATE:
1906                 *process = target_queue_recovery_request(req, obd);
1907                 RETURN(0);
1908
1909         default:
1910                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1911                 *process = 0;
1912                 /* XXX what should we set rq_status to here? */
1913                 req->rq_status = -EAGAIN;
1914                 RETURN(ptlrpc_error(req));
1915         }
1916 }
1917
1918 static char *reint_names[] = {
1919         [REINT_SETATTR] "setattr",
1920         [REINT_CREATE]  "create",
1921         [REINT_LINK]    "link",
1922         [REINT_UNLINK]  "unlink",
1923         [REINT_RENAME]  "rename",
1924         [REINT_OPEN]    "open",
1925 };
1926
1927 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
1928                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1929                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1930                             OBD_MD_FLID)
1931
1932 static void reconstruct_create(struct ptlrpc_request *req)
1933 {
1934         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1935         struct mds_client_data *mcd = med->med_mcd;
1936         struct dentry *dentry;
1937         struct ost_body *body;
1938         struct lustre_id id;
1939         int rc;
1940         ENTRY;
1941
1942         /* copy rc, transno and disp; steal locks */
1943         mds_req_from_mcd(req, mcd);
1944         if (req->rq_status) {
1945                 EXIT;
1946                 return;
1947         }
1948
1949         id_gen(&id) = 0;
1950         id_group(&id) = 0;
1951
1952         id_ino(&id) = mcd->mcd_last_data;
1953         LASSERT(id_ino(&id) != 0);
1954
1955         dentry = mds_id2dentry(req2obd(req), &id, NULL);
1956         if (IS_ERR(dentry)) {
1957                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1958                 req->rq_status = PTR_ERR(dentry);
1959                 EXIT;
1960                 return;
1961         }
1962
1963         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1964               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1965               dentry->d_inode->i_generation);
1966
1967         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1968         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1969         body->oa.o_id = dentry->d_inode->i_ino;
1970         body->oa.o_generation = dentry->d_inode->i_generation;
1971         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1972
1973         down(&dentry->d_inode->i_sem);
1974         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1975         up(&dentry->d_inode->i_sem);
1976         if (rc) {
1977                 CERROR("Can't read inode self id, inode %lu, "
1978                        "rc %d\n", dentry->d_inode->i_ino, rc);
1979                 id_fid(&id) = 0;
1980         }
1981
1982         body->oa.o_fid = id_fid(&id);
1983         body->oa.o_mds = id_group(&id);
1984         l_dput(dentry);
1985
1986         EXIT;
1987 }
1988
1989 static int mdt_obj_create(struct ptlrpc_request *req)
1990 {
1991         struct obd_device *obd = req->rq_export->exp_obd;
1992         struct mds_obd *mds = &obd->u.mds;
1993         struct ost_body *body, *repbody;
1994         char idname[LL_ID_NAMELEN];
1995         int size = sizeof(*repbody);
1996         struct inode *parent_inode;
1997         struct lvfs_run_ctxt saved;
1998         int rc, cleanup_phase = 0;
1999         struct dentry *new = NULL;
2000         struct dentry_params dp;
2001         int mealen, flags = 0;
2002         struct lvfs_ucred uc;
2003         struct lustre_id id;
2004         struct mea *mea;
2005         void *handle = NULL;
2006         unsigned long cr_inum = 0;
2007         ENTRY;
2008
2009         DEBUG_REQ(D_HA, req, "create remote object");
2010         parent_inode = mds->mds_unnamed_dir->d_inode;
2011
2012         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2013                                   lustre_swab_ost_body);
2014         if (body == NULL)
2015                 RETURN(-EFAULT);
2016
2017         rc = lustre_pack_reply(req, 1, &size, NULL);
2018         if (rc)
2019                 RETURN(rc);
2020
2021         MDS_CHECK_RESENT(req, reconstruct_create(req));
2022
2023         uc.luc_lsd = NULL;
2024         uc.luc_ginfo = NULL;
2025         uc.luc_uid = body->oa.o_uid;
2026         uc.luc_gid = body->oa.o_gid;
2027         uc.luc_fsuid = body->oa.o_uid;
2028         uc.luc_fsgid = body->oa.o_gid;
2029
2030         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2031         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2032
2033         /* in REPLAY case inum should be given (client or other MDS fills it) */
2034         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2035             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2036                 /*
2037                  * this is re-create request from MDS holding directory name.
2038                  * we have to lookup given ino/gen first. if it exists (good
2039                  * case) then there is nothing to do. if it does not then we
2040                  * have to recreate it.
2041                  */
2042                 id_ino(&id) = body->oa.o_id;
2043                 id_gen(&id) = body->oa.o_generation;
2044
2045                 new = mds_id2dentry(obd, &id, NULL);
2046                 if (!IS_ERR(new) && new->d_inode) {
2047                         struct lustre_id sid;
2048
2049                         CWARN("mkdir() repairing is on its way: %lu/%lu\n",
2050                               (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
2051
2052                         obdo_from_inode(&repbody->oa, new->d_inode,
2053                                         FILTER_VALID_FLAGS);
2054
2055                         repbody->oa.o_id = new->d_inode->i_ino;
2056                         repbody->oa.o_generation = new->d_inode->i_generation;
2057                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2058                         cleanup_phase = 1;
2059
2060                         down(&new->d_inode->i_sem);
2061                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2062                         up(&new->d_inode->i_sem);
2063                         if (rc) {
2064                                 CERROR("Can't read inode self id "
2065                                        "inode %lu, rc %d.\n",
2066                                        new->d_inode->i_ino, rc);
2067                                 GOTO(cleanup, rc);
2068                         }
2069
2070                         repbody->oa.o_fid = id_fid(&sid);
2071                         repbody->oa.o_mds = id_group(&sid);
2072                         LASSERT(id_fid(&sid) != 0);
2073
2074                         /*
2075                          * here we could use fid passed in body->oa.o_fid and
2076                          * thus avoid mds_read_inode_sid().
2077                          */
2078                         cr_inum = new->d_inode->i_ino;
2079                         GOTO(cleanup, rc = 0);
2080                 }
2081         }
2082
2083         down(&parent_inode->i_sem);
2084         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2085         if (IS_ERR(handle)) {
2086                 up(&parent_inode->i_sem);
2087                 CERROR("fsfilt_start() failed, rc = %d\n",
2088                        (int)PTR_ERR(handle));
2089                 GOTO(cleanup, rc = PTR_ERR(handle));
2090         }
2091         cleanup_phase = 1; /* transaction */
2092
2093 repeat:
2094         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2095         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2096         if (IS_ERR(new)) {
2097                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2098                        obd->obd_name, idname, (int) PTR_ERR(new));
2099                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2100                 up(&parent_inode->i_sem);
2101                 RETURN(PTR_ERR(new));
2102         } else if (new->d_inode) {
2103                 CERROR("%s: name exists. repeat\n", obd->obd_name);
2104                 goto repeat;
2105         }
2106
2107         new->d_fsdata = (void *)&dp;
2108         dp.p_inum = 0;
2109         dp.p_ptr = req;
2110
2111         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2112             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2113                 LASSERT(body->oa.o_id != 0);
2114                 dp.p_inum = body->oa.o_id;
2115                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2116                           (unsigned long)body->oa.o_id,
2117                           (unsigned long)body->oa.o_generation);
2118         }
2119
2120         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2121         if (rc == 0) {
2122                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2123                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2124                         new->d_inode->i_generation = body->oa.o_generation;
2125                         mark_inode_dirty(new->d_inode);
2126
2127                         /*
2128                          * avoiding asserts in cache flush case, as
2129                          * @body->oa.o_id should be zero.
2130                          */
2131                         if (body->oa.o_id) {
2132                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2133                                          "BUG 3550: failed to recreate obj "
2134                                          LPU64" -> %lu\n", body->oa.o_id,
2135                                          new->d_inode->i_ino);
2136
2137                                 LASSERTF(body->oa.o_generation ==
2138                                          new->d_inode->i_generation,
2139                                          "BUG 3550: failed to recreate obj/gen "
2140                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2141                                          body->oa.o_generation,
2142                                          new->d_inode->i_ino,
2143                                          new->d_inode->i_generation);
2144                         }
2145                 }
2146
2147                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2148                 repbody->oa.o_id = new->d_inode->i_ino;
2149                 repbody->oa.o_generation = new->d_inode->i_generation;
2150                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2151
2152                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2153                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2154                         id_group(&id) = mds->mds_num;
2155
2156                         LASSERT(body->oa.o_fid != 0);
2157                         id_fid(&id) = body->oa.o_fid;
2158
2159                         LASSERT(body->oa.o_id != 0);
2160                         id_ino(&id) = repbody->oa.o_id;
2161                         id_gen(&id) = repbody->oa.o_generation;
2162
2163                         down(&new->d_inode->i_sem);
2164                         rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2165                         up(&new->d_inode->i_sem);
2166
2167                         /*
2168                          * make sure, that fid is up-to-date.
2169                          */
2170                         mds_set_last_fid(obd, id_fid(&id));
2171                 } else {
2172                         /*
2173                          * allocate new sid, as object is created from scratch
2174                          * and this is not replay.
2175                          */
2176                         down(&new->d_inode->i_sem);
2177                         rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2178                         up(&new->d_inode->i_sem);
2179                 }
2180                 if (rc) {
2181                         CERROR("Can't update lustre ID for inode %lu, "
2182                                "error = %d\n", new->d_inode->i_ino, rc);
2183                         GOTO(cleanup, rc);
2184                 }
2185
2186                 /* initializing o_fid after it is allocated. */
2187                 repbody->oa.o_fid = id_fid(&id);
2188                 repbody->oa.o_mds = id_group(&id);
2189
2190                 rc = fsfilt_del_dir_entry(obd, new);
2191                 up(&parent_inode->i_sem);
2192                 if (rc) {
2193                         CERROR("can't remove name for object: %d\n", rc);
2194                         GOTO(cleanup, rc);
2195                 }
2196
2197                 cleanup_phase = 2; /* created directory object */
2198
2199                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2200                        (unsigned long)new->d_inode->i_ino,
2201                        (unsigned long)new->d_inode->i_generation,
2202                        (unsigned)new->d_inode->i_mode);
2203                 cr_inum = new->d_inode->i_ino;
2204         } else {
2205                 up(&parent_inode->i_sem);
2206                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2207                 GOTO(cleanup, rc);
2208         }
2209
2210         if (body->oa.o_valid & OBD_MD_FLID) {
2211                 /* this is new object for splitted dir. We have to prevent
2212                  * recursive splitting on it -bzzz */
2213                 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2214
2215                 OBD_ALLOC(mea, mealen);
2216                 if (mea == NULL)
2217                         GOTO(cleanup, rc = -ENOMEM);
2218
2219                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2220                 mea->mea_master = 0;
2221                 mea->mea_count = 0;
2222
2223                 down(&new->d_inode->i_sem);
2224                 rc = fsfilt_set_md(obd, new->d_inode, handle,
2225                                    mea, mealen, EA_MEA);
2226                 up(&new->d_inode->i_sem);
2227                 if (rc)
2228                         CERROR("fsfilt_set_md() failed, "
2229                                "rc = %d\n", rc);
2230
2231                 OBD_FREE(mea, mealen);
2232
2233                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2234                        obd->obd_name, new->d_inode->i_ino,
2235                        new->d_inode->i_generation, flags);
2236         } else if (body->oa.o_easize) {
2237                 /* we pass LCK_EX to split routine to signal that we have
2238                  * exclusive access to the directory. simple because nobody
2239                  * knows it already exists -bzzz */
2240                 rc = mds_try_to_split_dir(obd, new, NULL,
2241                                           body->oa.o_easize, LCK_EX);
2242                 if (rc < 0) {
2243                         CERROR("Can't split directory %lu, error = %d.\n",
2244                                new->d_inode->i_ino, rc);
2245                 } else {
2246                         rc = 0;
2247                 }
2248         }
2249
2250         EXIT;
2251 cleanup:
2252         switch (cleanup_phase) {
2253         case 2: /* object has been created, but we'll may want to replay it later */
2254                 if (rc == 0)
2255                         ptlrpc_require_repack(req);
2256         case 1: /* transaction */
2257                 rc = mds_finish_transno(mds, parent_inode, handle,
2258                                         req, rc, cr_inum);
2259         }
2260
2261         l_dput(new);
2262         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2263         return rc;
2264 }
2265
2266 static int mdt_get_info(struct ptlrpc_request *req)
2267 {
2268         struct obd_export *exp = req->rq_export;
2269         int keylen, rc = 0;
2270         char *key;
2271         ENTRY;
2272
2273         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2274         if (key == NULL) {
2275                 DEBUG_REQ(D_HA, req, "no get_info key");
2276                 RETURN(-EFAULT);
2277         }
2278         keylen = req->rq_reqmsg->buflens[0];
2279
2280         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2281             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2282             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2283                 RETURN(-EPROTO);
2284
2285         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2286                 struct lustre_id *reply;
2287                 int size = sizeof(*reply);
2288
2289                 rc = lustre_pack_reply(req, 1, &size, NULL);
2290                 if (rc)
2291                         RETURN(rc);
2292
2293                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2294                 rc = obd_get_info(exp, keylen, key, &size, reply);
2295         } else {
2296                 obd_id *reply;
2297                 int size = sizeof(*reply);
2298
2299                 rc = lustre_pack_reply(req, 1, &size, NULL);
2300                 if (rc)
2301                         RETURN(rc);
2302
2303                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2304                 rc = obd_get_info(exp, keylen, key, &size, reply);
2305         }
2306
2307         req->rq_repmsg->status = 0;
2308         RETURN(rc);
2309 }
2310
2311 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2312                         void *key, __u32 vallen, void *val)
2313 {
2314         struct obd_device *obd;
2315         struct mds_obd *mds;
2316         int rc = 0;
2317         ENTRY;
2318
2319         obd = class_exp2obd(exp);
2320         if (obd == NULL) {
2321                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2322                        exp->exp_handle.h_cookie);
2323                 RETURN(-EINVAL);
2324         }
2325
2326         mds = &obd->u.mds;
2327         if (keylen >= strlen("mds_type") &&
2328              memcmp(key, "mds_type", keylen) == 0) {
2329                 int valsize;
2330                 __u32 group;
2331
2332                 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2333
2334                 mds->mds_obd_type = *(int*)val;
2335                 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2336                 valsize = sizeof(group);
2337
2338                 /* mds number has been changed, so the corresponding obdfilter
2339                  * exp need to be changed too. */
2340                 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2341                                   "mds_conn", valsize, &group);
2342                 RETURN(rc);
2343         }
2344         CDEBUG(D_IOCTL, "invalid key\n");
2345         RETURN(-EINVAL);
2346 }
2347
2348 static int mdt_set_info(struct ptlrpc_request *req)
2349 {
2350         char *key, *val;
2351         struct obd_export *exp = req->rq_export;
2352         int keylen, rc = 0, vallen;
2353         ENTRY;
2354
2355         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2356         if (key == NULL) {
2357                 DEBUG_REQ(D_HA, req, "no set_info key");
2358                 RETURN(-EFAULT);
2359         }
2360         keylen = req->rq_reqmsg->buflens[0];
2361
2362         if (keylen == strlen("mds_type") &&
2363             memcmp(key, "mds_type", keylen) == 0) {
2364                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2365                 if (rc)
2366                         RETURN(rc);
2367
2368                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2369                 vallen = req->rq_reqmsg->buflens[1];
2370
2371                 rc = obd_set_info(exp, keylen, key, vallen, val);
2372                 req->rq_repmsg->status = 0;
2373                 RETURN(rc);
2374         }
2375         CDEBUG(D_IOCTL, "invalid key\n");
2376         RETURN(-EINVAL);
2377 }
2378
2379 static void mds_revoke_export_locks(struct obd_export *exp)
2380 {
2381         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
2382         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2383         struct ldlm_lock *lock, *next;
2384         struct ldlm_lock_desc desc;
2385
2386         if (!exp->u.eu_mds_data.med_remote)
2387                 return;
2388
2389         ENTRY;
2390         l_lock(&ns->ns_lock);
2391         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2392                 if (lock->l_req_mode != lock->l_granted_mode)
2393                         continue;
2394
2395                 LASSERT(lock->l_resource);
2396                 if (lock->l_resource->lr_type != LDLM_IBITS &&
2397                     lock->l_resource->lr_type != LDLM_PLAIN)
2398                         continue;
2399
2400                 if (lock->l_flags & LDLM_FL_AST_SENT)
2401                         continue;
2402
2403                 lock->l_flags |= LDLM_FL_AST_SENT;
2404
2405                 /* the desc just pretend to exclusive */
2406                 ldlm_lock2desc(lock, &desc);
2407                 desc.l_req_mode = LCK_EX;
2408                 desc.l_granted_mode = 0;
2409
2410                 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2411         }
2412         l_unlock(&ns->ns_lock);
2413         EXIT;
2414 }
2415
2416 static int mds_msg_check_version(struct lustre_msg *msg)
2417 {
2418         int rc;
2419
2420         switch (msg->opc) {
2421         case MDS_CONNECT:
2422         case MDS_DISCONNECT:
2423         case OBD_PING:
2424                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2425                 if (rc)
2426                         CERROR("bad opc %u version %08x, expecting %08x\n",
2427                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2428                 break;
2429         case MDS_STATFS:
2430         case MDS_GETSTATUS:
2431         case MDS_GETATTR:
2432         case MDS_GETATTR_LOCK:
2433         case MDS_READPAGE:
2434         case MDS_REINT:
2435         case MDS_CLOSE:
2436         case MDS_DONE_WRITING:
2437         case MDS_PIN:
2438         case MDS_SYNC:
2439                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2440                 if (rc)
2441                         CERROR("bad opc %u version %08x, expecting %08x\n",
2442                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2443                 break;
2444         case LDLM_ENQUEUE:
2445         case LDLM_CONVERT:
2446         case LDLM_BL_CALLBACK:
2447         case LDLM_CP_CALLBACK:
2448                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2449                 if (rc)
2450                         CERROR("bad opc %u version %08x, expecting %08x\n",
2451                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2452                 break;
2453         case OBD_LOG_CANCEL:
2454         case LLOG_ORIGIN_HANDLE_OPEN:
2455         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2456         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2457         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2458         case LLOG_ORIGIN_HANDLE_CLOSE:
2459         case LLOG_CATINFO:
2460                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2461                 if (rc)
2462                         CERROR("bad opc %u version %08x, expecting %08x\n",
2463                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2464                 break;
2465         case OST_CREATE:
2466         case OST_WRITE:
2467         case OST_GET_INFO:
2468         case OST_SET_INFO:
2469                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2470                 if (rc)
2471                         CERROR("bad opc %u version %08x, expecting %08x\n",
2472                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2473                 break;
2474         case SEC_INIT:
2475         case SEC_INIT_CONTINUE:
2476         case SEC_FINI:
2477                 rc = 0;
2478                 break;
2479         default:
2480                 CERROR("MDS unknown opcode %d\n", msg->opc);
2481                 rc = -ENOTSUPP;
2482                 break;
2483         }
2484
2485         return rc;
2486 }
2487
2488 int mds_handle(struct ptlrpc_request *req)
2489 {
2490         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2491         struct obd_device *obd = NULL;
2492         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2493         int rc = 0;
2494         ENTRY;
2495
2496         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2497
2498         rc = mds_msg_check_version(req->rq_reqmsg);
2499         if (rc) {
2500                 CERROR("MDS drop mal-formed request\n");
2501                 RETURN(rc);
2502         }
2503
2504         /* Security opc should NOT trigger any recovery events */
2505         if (req->rq_reqmsg->opc == SEC_INIT ||
2506             req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2507                 if (!req->rq_export)
2508                         GOTO(out, rc = 0);
2509
2510                 mds_req_add_idmapping(req, &req->rq_export->exp_mds_data);
2511                 mds_revoke_export_locks(req->rq_export);
2512                 GOTO(out, rc = 0);
2513         } else if (req->rq_reqmsg->opc == SEC_FINI) {
2514                 if (!req->rq_export)
2515                         GOTO(out, rc = 0);
2516
2517                 mds_req_del_idmapping(req, &req->rq_export->exp_mds_data);
2518                 mds_revoke_export_locks(req->rq_export);
2519                 GOTO(out, rc = 0);
2520         }
2521
2522         LASSERT(current->journal_info == NULL);
2523         /* XXX identical to OST */
2524         if (req->rq_reqmsg->opc != MDS_CONNECT) {
2525                 struct mds_export_data *med;
2526                 int recovering;
2527
2528                 if (req->rq_export == NULL) {
2529                         CERROR("operation %d on unconnected MDS from %s\n",
2530                                req->rq_reqmsg->opc,
2531                                req->rq_peerstr);
2532                         req->rq_status = -ENOTCONN;
2533                         GOTO(out, rc = -ENOTCONN);
2534                 }
2535
2536                 med = &req->rq_export->exp_mds_data;
2537                 obd = req->rq_export->exp_obd;
2538                 mds = &obd->u.mds;
2539
2540                 /* sanity check: if the xid matches, the request must
2541                  * be marked as a resent or replayed */
2542                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2543                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2544                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2545                                  (MSG_RESENT | MSG_REPLAY),
2546                                  "rq_xid "LPU64" matches last_xid, "
2547                                  "expected RESENT flag\n",
2548                                  req->rq_xid);
2549                 }
2550                 /* else: note the opposite is not always true; a
2551                  * RESENT req after a failover will usually not match
2552                  * the last_xid, since it was likely never
2553                  * committed. A REPLAYed request will almost never
2554                  * match the last xid, however it could for a
2555                  * committed, but still retained, open. */
2556
2557                 spin_lock_bh(&obd->obd_processing_task_lock);
2558                 recovering = obd->obd_recovering;
2559                 spin_unlock_bh(&obd->obd_processing_task_lock);
2560                 if (recovering) {
2561                         rc = mds_filter_recovery_request(req, obd,
2562                                                          &should_process);
2563                         if (rc || should_process == 0) {
2564                                 RETURN(rc);
2565                         } else if (should_process < 0) {
2566                                 req->rq_status = should_process;
2567                                 rc = ptlrpc_error(req);
2568                                 RETURN(rc);
2569                         }
2570                 }
2571         }
2572
2573         switch (req->rq_reqmsg->opc) {
2574         case MDS_CONNECT:
2575                 DEBUG_REQ(D_INODE, req, "connect");
2576                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2577                 rc = target_handle_connect(req);
2578                 if (!rc) {
2579                         struct mds_export_data *med;
2580
2581                         LASSERT(req->rq_export);
2582                         med = &req->rq_export->u.eu_mds_data;
2583                         mds_init_export_data(req, med);
2584                         mds_req_add_idmapping(req, med);
2585
2586                         /* Now that we have an export, set mds. */
2587                         obd = req->rq_export->exp_obd;
2588                         mds = mds_req2mds(req);
2589                 }
2590                 break;
2591
2592         case MDS_DISCONNECT:
2593                 DEBUG_REQ(D_INODE, req, "disconnect");
2594                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2595                 rc = target_handle_disconnect(req);
2596                 req->rq_status = rc;            /* superfluous? */
2597                 break;
2598
2599         case MDS_GETSTATUS:
2600                 DEBUG_REQ(D_INODE, req, "getstatus");
2601                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2602                 rc = mds_getstatus(req);
2603                 break;
2604
2605         case MDS_GETATTR:
2606                 DEBUG_REQ(D_INODE, req, "getattr");
2607                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2608                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2609                 break;
2610
2611         case MDS_GETATTR_LOCK: {
2612                 struct lustre_handle lockh;
2613                 DEBUG_REQ(D_INODE, req, "getattr_lock");
2614                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2615
2616                 /* If this request gets a reconstructed reply, we won't be
2617                  * acquiring any new locks in mds_getattr_lock, so we don't
2618                  * want to cancel.
2619                  */
2620                 lockh.cookie = 0;
2621                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2622                                       MDS_INODELOCK_UPDATE);
2623                 /* this non-intent call (from an ioctl) is special */
2624                 req->rq_status = rc;
2625                 if (rc == 0 && lockh.cookie)
2626                         ldlm_lock_decref(&lockh, LCK_PR);
2627                 break;
2628         }
2629         case MDS_STATFS:
2630                 DEBUG_REQ(D_INODE, req, "statfs");
2631                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2632                 rc = mds_statfs(req);
2633                 break;
2634
2635         case MDS_READPAGE:
2636                 DEBUG_REQ(D_INODE, req, "readpage");
2637                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2638                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2639
2640                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2641                         if (req->rq_reply_state) {
2642                                 lustre_free_reply_state (req->rq_reply_state);
2643                                 req->rq_reply_state = NULL;
2644                         }
2645                         RETURN(0);
2646                 }
2647
2648                 break;
2649         case MDS_REINT: {
2650                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2651                                              sizeof (*opcp));
2652                 __u32  opc;
2653                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2654                                mds->mds_max_cookiesize};
2655                 int bufcount;
2656
2657                 /* NB only peek inside req now; mds_reint() will swab it */
2658                 if (opcp == NULL) {
2659                         CERROR ("Can't inspect opcode\n");
2660                         rc = -EINVAL;
2661                         break;
2662                 }
2663                 opc = *opcp;
2664                 if (lustre_msg_swabbed (req->rq_reqmsg))
2665                         __swab32s(&opc);
2666
2667                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2668                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2669                            reint_names[opc] == NULL) ? reint_names[opc] :
2670                                                        "unknown opcode");
2671
2672                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2673
2674                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2675                         bufcount = 3;
2676                 else if (opc == REINT_OPEN)
2677                         bufcount = 2;
2678                 else
2679                         bufcount = 1;
2680
2681                 rc = lustre_pack_reply(req, bufcount, size, NULL);
2682                 if (rc)
2683                         break;
2684
2685                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2686                 fail = OBD_FAIL_MDS_REINT_NET_REP;
2687                 break;
2688         }
2689
2690         case MDS_CLOSE:
2691                 DEBUG_REQ(D_INODE, req, "close");
2692                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2693                 rc = mds_close(req, MDS_REQ_REC_OFF);
2694                 break;
2695
2696         case MDS_DONE_WRITING:
2697                 DEBUG_REQ(D_INODE, req, "done_writing");
2698                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2699                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2700                 break;
2701
2702         case MDS_PIN:
2703                 DEBUG_REQ(D_INODE, req, "pin");
2704                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2705                 rc = mds_pin(req, MDS_REQ_REC_OFF);
2706                 break;
2707
2708         case MDS_SYNC:
2709                 DEBUG_REQ(D_INODE, req, "sync");
2710                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2711                 rc = mds_sync(req, MDS_REQ_REC_OFF);
2712                 break;
2713
2714         case OBD_PING:
2715                 DEBUG_REQ(D_INODE, req, "ping");
2716                 rc = target_handle_ping(req);
2717                 break;
2718
2719         case OBD_LOG_CANCEL:
2720                 CDEBUG(D_INODE, "log cancel\n");
2721                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2722                 rc = -ENOTSUPP; /* la la la */
2723                 break;
2724
2725         case LDLM_ENQUEUE:
2726                 DEBUG_REQ(D_INODE, req, "enqueue");
2727                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2728                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2729                                          ldlm_server_blocking_ast, NULL);
2730                 fail = OBD_FAIL_LDLM_REPLY;
2731                 break;
2732         case LDLM_CONVERT:
2733                 DEBUG_REQ(D_INODE, req, "convert");
2734                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2735                 rc = ldlm_handle_convert(req);
2736                 break;
2737         case LDLM_BL_CALLBACK:
2738         case LDLM_CP_CALLBACK:
2739                 DEBUG_REQ(D_INODE, req, "callback");
2740                 CERROR("callbacks should not happen on MDS\n");
2741                 LBUG();
2742                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2743                 break;
2744         case LLOG_ORIGIN_HANDLE_OPEN:
2745                 DEBUG_REQ(D_INODE, req, "llog_init");
2746                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2747                 rc = llog_origin_handle_open(req);
2748                 break;
2749         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2750                 DEBUG_REQ(D_INODE, req, "llog next block");
2751                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2752                 rc = llog_origin_handle_next_block(req);
2753                 break;
2754         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2755                 DEBUG_REQ(D_INODE, req, "llog prev block");
2756                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2757                 rc = llog_origin_handle_prev_block(req);
2758                 break;
2759         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2760                 DEBUG_REQ(D_INODE, req, "llog read header");
2761                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2762                 rc = llog_origin_handle_read_header(req);
2763                 break;
2764         case LLOG_ORIGIN_HANDLE_CLOSE:
2765                 DEBUG_REQ(D_INODE, req, "llog close");
2766                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2767                 rc = llog_origin_handle_close(req);
2768                 break;
2769         case OST_CREATE:
2770                 DEBUG_REQ(D_INODE, req, "ost_create");
2771                 rc = mdt_obj_create(req);
2772                 break;
2773         case OST_GET_INFO:
2774                 DEBUG_REQ(D_INODE, req, "get_info");
2775                 rc = mdt_get_info(req);
2776                 break;
2777         case OST_SET_INFO:
2778                 DEBUG_REQ(D_INODE, req, "set_info");
2779                 rc = mdt_set_info(req);
2780                 break;
2781         case OST_WRITE:
2782                 CDEBUG(D_INODE, "write\n");
2783                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2784                 rc = ost_brw_write(req, NULL);
2785                 LASSERT(current->journal_info == NULL);
2786                 /* mdt_brw sends its own replies */
2787                 RETURN(rc);
2788                 break;
2789         case LLOG_CATINFO:
2790                 DEBUG_REQ(D_INODE, req, "llog catinfo");
2791                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2792                 rc = llog_catinfo(req);
2793                 break;
2794         default:
2795                 req->rq_status = -ENOTSUPP;
2796                 rc = ptlrpc_error(req);
2797                 RETURN(rc);
2798         }
2799
2800         LASSERT(current->journal_info == NULL);
2801
2802         EXIT;
2803
2804         /* If we're DISCONNECTing, the mds_export_data is already freed */
2805         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2806                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2807                 struct obd_device *obd = list_entry(mds, struct obd_device,
2808                                                     u.mds);
2809                 req->rq_repmsg->last_xid =
2810                         le64_to_cpu(med->med_mcd->mcd_last_xid);
2811
2812                 if (!obd->obd_no_transno) {
2813                         req->rq_repmsg->last_committed =
2814                                 obd->obd_last_committed;
2815                 } else {
2816                         DEBUG_REQ(D_IOCTL, req,
2817                                   "not sending last_committed update");
2818                 }
2819                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2820                        ", xid "LPU64"\n",
2821                        mds->mds_last_transno, obd->obd_last_committed,
2822                        req->rq_xid);
2823         }
2824  out:
2825
2826         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2827                 if (obd && obd->obd_recovering) {
2828                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2829                         return target_queue_final_reply(req, rc);
2830                 }
2831                 /* Lost a race with recovery; let the error path DTRT. */
2832                 rc = req->rq_status = -ENOTCONN;
2833         }
2834
2835         target_send_reply(req, rc, fail);
2836         return 0;
2837 }
2838
2839 /* Update the server data on disk.  This stores the new mount_count and also the
2840  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
2841  * last_rcvd value may be less than that of the clients.  This will alert us
2842  * that we may need to do client recovery.
2843  *
2844  * Also assumes for mds_last_transno that we are not modifying it (no locking).
2845  */
2846 int mds_update_server_data(struct obd_device *obd, int force_sync)
2847 {
2848         struct mds_obd *mds = &obd->u.mds;
2849         struct mds_server_data *msd = mds->mds_server_data;
2850         struct file *filp = mds->mds_rcvd_filp;
2851         struct lvfs_run_ctxt saved;
2852         loff_t off = 0;
2853         int rc;
2854         ENTRY;
2855
2856         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2857         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2858
2859         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2860                mds->mds_mount_count, mds->mds_last_transno);
2861         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2862         if (rc)
2863                 CERROR("error writing MDS server data: rc = %d\n", rc);
2864         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2865
2866         RETURN(rc);
2867 }
2868
2869 /* saves last allocated fid counter to file. */
2870 int mds_update_last_fid(struct obd_device *obd, void *handle,
2871                         int force_sync)
2872 {
2873         struct mds_obd *mds = &obd->u.mds;
2874         struct file *filp = mds->mds_fid_filp;
2875         struct lvfs_run_ctxt saved;
2876         loff_t off = 0;
2877         __u64 last_fid;
2878         int rc = 0;
2879         ENTRY;
2880
2881         spin_lock(&mds->mds_last_fid_lock);
2882         last_fid = mds->mds_last_fid;
2883         spin_unlock(&mds->mds_last_fid_lock);
2884
2885         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2886                last_fid);
2887
2888         if (handle) {
2889                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
2890                                       handle, mds_commit_last_fid_cb,
2891                                       NULL);
2892         }
2893
2894         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2895         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
2896                                  &off, force_sync);
2897         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2898
2899         if (rc) {
2900                 CERROR("error writing MDS last_fid #"LPU64
2901                        ", err = %d\n", last_fid, rc);
2902                 RETURN(rc);
2903         }
2904
2905         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2906                "%llu: err = %d\n", last_fid, off, rc);
2907
2908         RETURN(rc);
2909 }
2910
2911 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2912 {
2913         struct mds_obd *mds = &obd->u.mds;
2914
2915         spin_lock(&mds->mds_last_fid_lock);
2916         if (fid > mds->mds_last_fid)
2917                 mds->mds_last_fid = fid;
2918         spin_unlock(&mds->mds_last_fid_lock);
2919 }
2920
2921 void mds_commit_last_transno_cb(struct obd_device *obd,
2922                                 __u64 transno, void *data,
2923                                 int error)
2924 {
2925         obd_transno_commit_cb(obd, transno, error);
2926 }
2927
2928 void mds_commit_last_fid_cb(struct obd_device *obd,
2929                             __u64 fid, void *data,
2930                             int error)
2931 {
2932         if (error) {
2933                 CERROR("%s: fid "LPD64" commit error: %d\n",
2934                        obd->obd_name, fid, error);
2935                 return;
2936         }
2937
2938         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2939                obd->obd_name, fid);
2940 }
2941
2942 __u64 mds_alloc_fid(struct obd_device *obd)
2943 {
2944         struct mds_obd *mds = &obd->u.mds;
2945         __u64 fid;
2946
2947         spin_lock(&mds->mds_last_fid_lock);
2948         fid = ++mds->mds_last_fid;
2949         spin_unlock(&mds->mds_last_fid_lock);
2950
2951         return fid;
2952 }
2953
2954 /*
2955  * allocates new lustre_id on passed @inode and saves it to inode EA.
2956  */
2957 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2958                         void *handle, struct lustre_id *id)
2959 {
2960         struct mds_obd *mds = &obd->u.mds;
2961         int alloc = 0, rc = 0;
2962         ENTRY;
2963
2964         LASSERT(obd != NULL);
2965         LASSERT(inode != NULL);
2966
2967         if (id == NULL) {
2968                 OBD_ALLOC(id, sizeof(*id));
2969                 if (id == NULL)
2970                         RETURN(-ENOMEM);
2971                 alloc = 1;
2972         }
2973
2974         id_group(id) = mds->mds_num;
2975         id_fid(id) = mds_alloc_fid(obd);
2976         id_ino(id) = inode->i_ino;
2977         id_gen(id) = inode->i_generation;
2978         id_type(id) = (S_IFMT & inode->i_mode);
2979
2980         rc = mds_update_inode_sid(obd, inode, handle, id);
2981         if (rc) {
2982                 CERROR("Can't update inode FID EA, "
2983                        "rc = %d\n", rc);
2984         }
2985
2986         if (alloc)
2987                 OBD_FREE(id, sizeof(*id));
2988         RETURN(rc);
2989 }
2990
2991 /*
2992  * reads inode self id from inode EA. Probably later this should be replaced by
2993  * caching inode self id to avoid raeding it every time it is needed.
2994  */
2995 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2996                        struct lustre_id *id)
2997 {
2998         int rc;
2999         ENTRY;
3000
3001         LASSERT(id != NULL);
3002         LASSERT(obd != NULL);
3003         LASSERT(inode != NULL);
3004
3005         rc = fsfilt_get_md(obd, inode, &id->li_fid,
3006                            sizeof(id->li_fid), EA_SID);
3007         if (rc < 0) {
3008                 CERROR("fsfilt_get_md() failed, "
3009                        "rc = %d\n", rc);
3010                 RETURN(rc);
3011         } else if (!rc) {
3012                 rc = -ENODATA;
3013                 RETURN(rc);
3014         } else {
3015                 rc = 0;
3016         }
3017
3018         RETURN(rc);
3019 }
3020
3021 /* updates inode self id in EA. */
3022 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3023                          void *handle, struct lustre_id *id)
3024 {
3025         int rc = 0;
3026         ENTRY;
3027
3028         LASSERT(id != NULL);
3029         LASSERT(obd != NULL);
3030         LASSERT(inode != NULL);
3031
3032         rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3033                            sizeof(id->li_fid), EA_SID);
3034         if (rc) {
3035                 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3036                 RETURN(rc);
3037         }
3038
3039         RETURN(rc);
3040 }
3041
3042 /*
3043  * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3044  * to master MDS by correct store cookie, needed to find inode on master MDS
3045  * quickly.
3046  */
3047 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3048                        struct lustre_id *id)
3049 {
3050         int rc;
3051         ENTRY;
3052
3053         LASSERT(id != NULL);
3054         LASSERT(obd != NULL);
3055         LASSERT(inode != NULL);
3056
3057         rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3058         if (rc < 0) {
3059                 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3060                 RETURN(rc);
3061         } else if (!rc) {
3062                 rc = -ENODATA;
3063                 RETURN(rc);
3064         } else {
3065                 rc = 0;
3066         }
3067
3068         RETURN(rc);
3069 }
3070
3071 /*
3072  * updates master inode id. Usualy this is done by CMOBD after an inode is
3073  * created and relationship between cache MDS and master one should be
3074  * established.
3075  */
3076 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3077                          void *handle, struct lustre_id *id)
3078 {
3079         int rc = 0;
3080         ENTRY;
3081
3082         LASSERT(id != NULL);
3083         LASSERT(obd != NULL);
3084         LASSERT(inode != NULL);
3085
3086         rc = fsfilt_set_md(obd, inode, handle, id,
3087                            sizeof(*id), EA_MID);
3088         if (rc) {
3089                 CERROR("fsfilt_set_md() failed, "
3090                        "rc = %d\n", rc);
3091                 RETURN(rc);
3092         }
3093
3094         RETURN(rc);
3095 }
3096
3097 /* mount the file system (secretly) */
3098 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3099 {
3100         struct lustre_cfg* lcfg = buf;
3101         struct mds_obd *mds = &obd->u.mds;
3102         struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3103         char *options = NULL;
3104         struct vfsmount *mnt;
3105         char ns_name[48];
3106         unsigned long page;
3107         int rc = 0;
3108         ENTRY;
3109
3110
3111         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
3112                 RETURN(rc = -EINVAL);
3113
3114         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
3115         if (IS_ERR(obd->obd_fsops))
3116                 RETURN(rc = PTR_ERR(obd->obd_fsops));
3117
3118         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3119
3120         page = __get_free_page(GFP_KERNEL);
3121         if (!page)
3122                 RETURN(-ENOMEM);
3123
3124         options = (char *)page;
3125         memset(options, 0, PAGE_SIZE);
3126
3127         /*
3128          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3129          * and the rest of options are passed by mount options. Probably this
3130          * should be moved to somewhere else like startup scripts or lconf. */
3131         sprintf(options, "iopen_nopriv");
3132         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
3133                 sprintf(options + strlen(options), ",%s",
3134                         lcfg->lcfg_inlbuf4);
3135
3136         /* we have to know mdsnum before touching underlying fs -bzzz */
3137         atomic_set(&mds->mds_open_count, 0);
3138         sema_init(&mds->mds_md_sem, 1);
3139         mds->mds_md_connected = 0;
3140         mds->mds_md_name = NULL;
3141
3142         if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
3143             strncmp(lcfg->lcfg_inlbuf5, "dumb", lcfg->lcfg_inllen5)) {
3144                 class_uuid_t uuid;
3145
3146                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3147                        obd->obd_name, lcfg->lcfg_inlbuf5);
3148
3149                 generate_random_uuid(uuid);
3150                 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3151
3152                 OBD_ALLOC(mds->mds_md_name, lcfg->lcfg_inllen5);
3153                 if (mds->mds_md_name == NULL)
3154                         RETURN(rc = -ENOMEM);
3155
3156                 memcpy(mds->mds_md_name, lcfg->lcfg_inlbuf5,
3157                        lcfg->lcfg_inllen5);
3158
3159                 rc = mds_md_connect(obd, mds->mds_md_name);
3160                 if (rc) {
3161                         OBD_FREE(mds->mds_md_name, lcfg->lcfg_inllen5);
3162                         GOTO(err_ops, rc);
3163                 }
3164         }
3165
3166         mds->mds_obd_type = MDS_MASTER_OBD;
3167
3168         if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
3169             strncmp(lcfg->lcfg_inlbuf6, "dumb", lcfg->lcfg_inllen6)) {
3170                 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
3171                         mds->mds_obd_type = MDS_MASTER_OBD;
3172                 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
3173                         mds->mds_obd_type = MDS_CACHE_OBD;
3174                 }
3175         }
3176
3177         rc = lvfs_mount_fs(lcfg->lcfg_inlbuf1, lcfg->lcfg_inlbuf2,
3178                            options, 0, &lvfs_ctxt);
3179
3180         free_page(page);
3181
3182         if (rc || !lvfs_ctxt) {
3183                 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3184                 GOTO(err_ops, rc);
3185         }
3186
3187         mnt = lvfs_ctxt->loc_mnt;
3188         mds->mds_lvfs_ctxt = lvfs_ctxt;
3189         ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3190
3191         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
3192
3193         sema_init(&mds->mds_epoch_sem, 1);
3194         atomic_set(&mds->mds_real_clients, 0);
3195         spin_lock_init(&mds->mds_transno_lock);
3196         spin_lock_init(&mds->mds_last_fid_lock);
3197         sema_init(&mds->mds_orphan_recovery_sem, 1);
3198         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3199
3200         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3201         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3202
3203         if (obd->obd_namespace == NULL) {
3204                 mds_cleanup(obd, 0);
3205                 GOTO(err_put, rc = -ENOMEM);
3206         }
3207         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3208
3209         rc = mds_fs_setup(obd, mnt);
3210         if (rc) {
3211                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3212                        obd->obd_name, rc);
3213                 GOTO(err_ns, rc);
3214         }
3215
3216         rc = llog_start_commit_thread();
3217         if (rc < 0)
3218                 GOTO(err_fs, rc);
3219
3220         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
3221             strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)) {
3222                 class_uuid_t uuid;
3223
3224                 generate_random_uuid(uuid);
3225                 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3226
3227                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
3228                 if (mds->mds_profile == NULL)
3229                         GOTO(err_fs, rc = -ENOMEM);
3230
3231                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
3232                        lcfg->lcfg_inllen3);
3233         }
3234
3235         /*
3236          * setup root dir and files ID dir if lmv already connected, or there is
3237          * not lmv at all.
3238          */
3239         if (mds->mds_md_exp || (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
3240                                 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)))
3241         {
3242                 rc = mds_fs_setup_rootid(obd);
3243                 if (rc)
3244                         GOTO(err_fs, rc);
3245
3246                 rc = mds_fs_setup_virtid(obd);
3247                 if (rc)
3248                         GOTO(err_fs, rc);
3249         }
3250
3251         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3252                            "mds_ldlm_client", &obd->obd_ldlm_client);
3253         obd->obd_replayable = 1;
3254
3255         rc = mds_postsetup(obd);
3256         if (rc)
3257                 GOTO(err_fs, rc);
3258
3259         RETURN(0);
3260
3261 err_fs:
3262         /* No extra cleanup needed for llog_init_commit_thread() */
3263         mds_fs_cleanup(obd, 0);
3264 err_ns:
3265         ldlm_namespace_free(obd->obd_namespace, 0);
3266         obd->obd_namespace = NULL;
3267 err_put:
3268         unlock_kernel();
3269         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3270         mds->mds_sb = 0;
3271         lock_kernel();
3272 err_ops:
3273         fsfilt_put_ops(obd->obd_fsops);
3274         return rc;
3275 }
3276
3277 static int mds_postsetup(struct obd_device *obd)
3278 {
3279         struct mds_obd *mds = &obd->u.mds;
3280         int rc = 0;
3281         ENTRY;
3282
3283         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3284                             obd, 0, NULL, &llog_lvfs_ops);
3285         if (rc)
3286                 RETURN(rc);
3287
3288         if (mds->mds_profile) {
3289                 struct llog_ctxt *lgctxt;
3290                 struct lvfs_run_ctxt saved;
3291                 struct lustre_profile *lprof;
3292                 struct config_llog_instance cfg;
3293
3294                 cfg.cfg_instance = NULL;
3295                 cfg.cfg_uuid = mds->mds_dt_uuid;
3296                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3297
3298                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3299                 if (!lgctxt) {
3300                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3301                         GOTO(err_llog, rc = -EINVAL);
3302                 }
3303
3304                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3305                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3306
3307                 if (rc)
3308                         GOTO(err_llog, rc);
3309
3310                 lprof = class_get_profile(mds->mds_profile);
3311                 if (lprof == NULL) {
3312                         CERROR("No profile found: %s\n", mds->mds_profile);
3313                         GOTO(err_cleanup, rc = -ENOENT);
3314                 }
3315                 rc = mds_dt_connect(obd, lprof->lp_lov);
3316                 if (rc)
3317                         GOTO(err_cleanup, rc);
3318
3319                 rc = mds_md_postsetup(obd);
3320                 if (rc)
3321                         GOTO(err_cleanup, rc);
3322         }
3323
3324         RETURN(rc);
3325 err_cleanup:
3326         mds_dt_clean(obd);
3327 err_llog:
3328         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3329                                           LLOG_CONFIG_ORIG_CTXT));
3330         return rc;
3331 }
3332
3333 int mds_postrecov_common(struct obd_device *obd)
3334 {
3335         struct mds_obd *mds = &obd->u.mds;
3336         struct llog_ctxt *ctxt;
3337         int rc, item = 0, valsize;
3338          __u32 group;
3339         ENTRY;
3340
3341         LASSERT(!obd->obd_recovering);
3342         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3343         LASSERT(ctxt != NULL);
3344
3345         /* set nextid first, so we are sure it happens */
3346         rc = mds_dt_set_nextid(obd);
3347         if (rc) {
3348                 CERROR("%s: mds_dt_set_nextid() failed\n", obd->obd_name);
3349                 GOTO(out, rc);
3350         }
3351
3352         /* clean PENDING dir */
3353         rc = mds_cleanup_orphans(obd);
3354         if (rc < 0)
3355                 GOTO(out, rc);
3356         item = rc;
3357
3358         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3359         valsize = sizeof(group);
3360         rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"), "mds_conn",
3361                           valsize, &group);
3362         if (rc)
3363                 GOTO(out, rc);
3364
3365         rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3366                           NULL, NULL, NULL);
3367         if (rc) {
3368                 CERROR("%s: failed at llog_origin_connect: %d\n",
3369                        obd->obd_name, rc);
3370                 GOTO(out, rc);
3371         }
3372
3373         /* remove the orphaned precreated objects */
3374         rc = mds_dt_clearorphans(mds, NULL /* all OSTs */);
3375         if (rc)
3376                 GOTO(err_llog, rc);
3377
3378 out:
3379         RETURN(rc < 0 ? rc : item);
3380
3381 err_llog:
3382         /* cleanup all llogging subsystems */
3383         rc = obd_llog_finish(obd, &obd->obd_llogs,
3384                              mds->mds_dt_desc.ld_tgt_count);
3385         if (rc)
3386                 CERROR("%s: failed to cleanup llogging subsystems\n",
3387                         obd->obd_name);
3388         goto out;
3389 }
3390
3391 int mds_postrecov(struct obd_device *obd)
3392 {
3393         int rc;
3394         ENTRY;
3395         rc = mds_postrecov_common(obd);
3396         if (rc == 0)
3397                 rc = mds_md_reconnect(obd);
3398         RETURN(rc);
3399 }
3400
3401 int mds_dt_clean(struct obd_device *obd)
3402 {
3403         struct mds_obd *mds = &obd->u.mds;
3404         ENTRY;
3405
3406         if (mds->mds_profile) {
3407                 char * cln_prof;
3408                 struct llog_ctxt *llctx;
3409                 struct lvfs_run_ctxt saved;
3410                 struct config_llog_instance cfg;
3411                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3412
3413                 OBD_ALLOC(cln_prof, len);
3414                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3415
3416                 cfg.cfg_instance = NULL;
3417                 cfg.cfg_uuid = mds->mds_dt_uuid;
3418
3419                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3420                 llctx = llog_get_context(&obd->obd_llogs,
3421                                          LLOG_CONFIG_ORIG_CTXT);
3422                 class_config_process_llog(llctx, cln_prof, &cfg);
3423                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3424
3425                 OBD_FREE(cln_prof, len);
3426                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3427                 mds->mds_profile = NULL;
3428         }
3429         RETURN(0);
3430 }
3431
3432 int mds_md_clean(struct obd_device *obd)
3433 {
3434         struct mds_obd *mds = &obd->u.mds;
3435         ENTRY;
3436
3437         if (mds->mds_md_name) {
3438                 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3439                 mds->mds_md_name = NULL;
3440         }
3441         RETURN(0);
3442 }
3443
3444 static int mds_precleanup(struct obd_device *obd, int flags)
3445 {
3446         int rc = 0;
3447         ENTRY;
3448
3449         mds_md_clean(obd);
3450         mds_dt_disconnect(obd, flags);
3451         mds_dt_clean(obd);
3452         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3453         RETURN(rc);
3454 }
3455
3456 extern void lgss_svc_cache_purge_all(void);
3457 static int mds_cleanup(struct obd_device *obd, int flags)
3458 {
3459         struct mds_obd *mds = &obd->u.mds;
3460         ENTRY;
3461
3462         if (mds->mds_sb == NULL)
3463                 RETURN(0);
3464
3465         mds_update_server_data(obd, 1);
3466         mds_update_last_fid(obd, NULL, 1);
3467
3468         if (mds->mds_dt_objids != NULL) {
3469                 int size = mds->mds_dt_desc.ld_tgt_count *
3470                         sizeof(obd_id);
3471                 OBD_FREE(mds->mds_dt_objids, size);
3472         }
3473         mds_fs_cleanup(obd, flags);
3474
3475         unlock_kernel();
3476
3477         /* 2 seems normal on mds, (may_umount() also expects 2
3478           fwiw), but we only see 1 at this point in obdfilter. */
3479         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3480
3481         mds->mds_sb = 0;
3482
3483         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3484
3485         spin_lock_bh(&obd->obd_processing_task_lock);
3486         if (obd->obd_recovering) {
3487                 target_cancel_recovery_timer(obd);
3488                 obd->obd_recovering = 0;
3489         }
3490         spin_unlock_bh(&obd->obd_processing_task_lock);
3491
3492         lock_kernel();
3493         fsfilt_put_ops(obd->obd_fsops);
3494
3495 #ifdef ENABLE_GSS
3496         /* XXX */
3497         lgss_svc_cache_purge_all();
3498 #endif
3499         RETURN(0);
3500 }
3501
3502 static int set_security(const char *value, char **sec)
3503 {
3504         int rc = 0;
3505
3506         if (!strcmp(value, "null"))
3507                 *sec = "null";
3508         else if (!strcmp(value, "krb5i"))
3509                 *sec = "krb5i";
3510         else if (!strcmp(value, "krb5p"))
3511                 *sec = "krb5p";
3512         else {
3513                 CERROR("Unrecognized value, force use NULL\n");
3514                 rc = -EINVAL;
3515         }
3516
3517         return rc;
3518 }
3519
3520 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3521 {
3522         struct lustre_cfg *lcfg = buf;
3523         struct mds_obd *mds = &obd->u.mds;
3524         int rc = 0;
3525         ENTRY;
3526
3527         switch(lcfg->lcfg_command) {
3528         case LCFG_SET_SECURITY: {
3529                 if (!lcfg->lcfg_inllen1 || !lcfg->lcfg_inllen2)
3530                         GOTO(out, rc = -EINVAL);
3531
3532                 if (!strcmp(lcfg->lcfg_inlbuf1, "mds_mds_sec"))
3533                         rc = set_security(lcfg->lcfg_inlbuf2,
3534                                           &mds->mds_mds_sec);
3535                 else if (!strcmp(lcfg->lcfg_inlbuf1, "mds_ost_sec"))
3536                         rc = set_security(lcfg->lcfg_inlbuf2,
3537                                           &mds->mds_ost_sec);
3538                 else {
3539                         CERROR("Unrecognized key\n");
3540                         rc = -EINVAL;
3541                 }
3542                 break;
3543         }
3544         default: {
3545                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3546                 GOTO(out, rc = -EINVAL);
3547
3548         }
3549         }
3550 out:
3551         RETURN(rc);
3552 }
3553
3554 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3555                                         int offset,
3556                                         struct ldlm_lock *new_lock,
3557                                         struct ldlm_lock **old_lock,
3558                                         struct lustre_handle *lockh)
3559 {
3560         struct obd_export *exp = req->rq_export;
3561         struct obd_device *obd = exp->exp_obd;
3562         struct ldlm_request *dlmreq =
3563                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3564         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3565         struct list_head *iter;
3566
3567         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3568                 return;
3569
3570         l_lock(&obd->obd_namespace->ns_lock);
3571         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3572                 struct ldlm_lock *lock;
3573                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3574                 if (lock == new_lock)
3575                         continue;
3576                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3577                         lockh->cookie = lock->l_handle.h_cookie;
3578                         LDLM_DEBUG(lock, "restoring lock cookie");
3579                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3580                                   lockh->cookie);
3581                         if (old_lock)
3582                                 *old_lock = LDLM_LOCK_GET(lock);
3583                         l_unlock(&obd->obd_namespace->ns_lock);
3584                         return;
3585                 }
3586         }
3587         l_unlock(&obd->obd_namespace->ns_lock);
3588
3589         /* If the xid matches, then we know this is a resent request,
3590          * and allow it. (It's probably an OPEN, for which we don't
3591          * send a lock */
3592         if (req->rq_xid ==
3593             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3594                 return;
3595
3596         if (req->rq_xid ==
3597             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3598                 return;
3599
3600         /* This remote handle isn't enqueued, so we never received or
3601          * processed this request.  Clear MSG_RESENT, because it can
3602          * be handled like any normal request now. */
3603
3604         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3605
3606         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3607                   remote_hdl.cookie);
3608 }
3609
3610 int intent_disposition(struct ldlm_reply *rep, int flag)
3611 {
3612         if (!rep)
3613                 return 0;
3614         return (rep->lock_policy_res1 & flag);
3615 }
3616
3617 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3618 {
3619         if (!rep)
3620                 return;
3621         rep->lock_policy_res1 |= flag;
3622 }
3623
3624 static int mds_intent_policy(struct ldlm_namespace *ns,
3625                              struct ldlm_lock **lockp, void *req_cookie,
3626                              ldlm_mode_t mode, int flags, void *data)
3627 {
3628         struct ptlrpc_request *req = req_cookie;
3629         struct ldlm_lock *lock = *lockp;
3630         struct ldlm_intent *it;
3631         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3632         struct ldlm_reply *rep;
3633         struct lustre_handle lockh[2] = {{0}, {0}};
3634         struct ldlm_lock *new_lock = NULL;
3635         int getattr_part = MDS_INODELOCK_UPDATE;
3636         int rc, reply_buffers;
3637         int repsize[5] = {sizeof(struct ldlm_reply),
3638                           sizeof(struct mds_body),
3639                           mds->mds_max_mdsize};
3640
3641         int offset = MDS_REQ_INTENT_REC_OFF;
3642         ENTRY;
3643
3644         LASSERT(req != NULL);
3645         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3646
3647         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3648                 /* No intent was provided */
3649                 int size = sizeof(struct ldlm_reply);
3650                 rc = lustre_pack_reply(req, 1, &size, NULL);
3651                 LASSERT(rc == 0);
3652                 RETURN(0);
3653         }
3654
3655         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3656                                 lustre_swab_ldlm_intent);
3657         if (it == NULL) {
3658                 CERROR("Intent missing\n");
3659                 RETURN(req->rq_status = -EFAULT);
3660         }
3661
3662         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3663
3664         reply_buffers = 3;
3665         if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3666                 reply_buffers = 5;
3667                 repsize[3] = 4;
3668                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
3669         }
3670
3671         rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
3672         if (rc)
3673                 RETURN(req->rq_status = rc);
3674
3675         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3676         LASSERT(rep != NULL);
3677
3678         intent_set_disposition(rep, DISP_IT_EXECD);
3679
3680         /* execute policy */
3681         switch ((long)it->opc) {
3682         case IT_OPEN:
3683         case IT_CREAT|IT_OPEN:
3684                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3685                                             lock, NULL, lockh);
3686                 /* XXX swab here to assert that an mds_open reint
3687                  * packet is following */
3688                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3689                                             lock, NULL, lockh);
3690                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3691
3692                 if (rep->lock_policy_res2) {
3693                         /*
3694                          * mds_open() returns ENOLCK where it should return
3695                          * zero, but it has no lock to return.
3696                          */
3697                         if (rep->lock_policy_res2 == ENOLCK)
3698                                 rep->lock_policy_res2 = 0;
3699
3700                         RETURN(ELDLM_LOCK_ABORTED);
3701                 }
3702
3703                 /*
3704                  * IT_OPEN may return lock on cross-node dentry that we want to
3705                  * hold during attr retrival -bzzz
3706                  */
3707                 if (lockh[0].cookie == 0)
3708                         RETURN(ELDLM_LOCK_ABORTED);
3709
3710                 break;
3711         case IT_LOOKUP:
3712                 getattr_part = MDS_INODELOCK_LOOKUP;
3713         case IT_CHDIR:
3714         case IT_GETATTR:
3715                 getattr_part |= MDS_INODELOCK_LOOKUP;
3716         case IT_READDIR:
3717                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3718                                             lock, &new_lock, lockh);
3719                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3720                                                          getattr_part);
3721                 /* FIXME: LDLM can set req->rq_status. MDS sets
3722                    policy_res{1,2} with disposition and status.
3723                    - replay: returns 0 & req->status is old status
3724                    - otherwise: returns req->status */
3725                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3726                         rep->lock_policy_res2 = 0;
3727                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3728                     rep->lock_policy_res2)
3729                         RETURN(ELDLM_LOCK_ABORTED);
3730                 if (req->rq_status != 0) {
3731                         LBUG();
3732                         rep->lock_policy_res2 = req->rq_status;
3733                         RETURN(ELDLM_LOCK_ABORTED);
3734                 }
3735                 break;
3736         case IT_UNLINK:
3737                 rc = mds_lock_and_check_slave(offset, req, lockh);
3738                 if ((rep->lock_policy_res2 = rc)) {
3739                         if (rc == ENOLCK)
3740                                 rep->lock_policy_res2 = 0;
3741                         RETURN(ELDLM_LOCK_ABORTED);
3742                 }
3743                 break;
3744         default:
3745                 CERROR("Unhandled intent "LPD64"\n", it->opc);
3746                 LBUG();
3747         }
3748
3749         /* By this point, whatever function we called above must have either
3750          * filled in 'lockh', been an intent replay, or returned an error.  We
3751          * want to allow replayed RPCs to not get a lock, since we would just
3752          * drop it below anyways because lock replay is done separately by the
3753          * client afterwards.  For regular RPCs we want to give the new lock to
3754          * the client instead of whatever lock it was about to get. */
3755         if (new_lock == NULL)
3756                 new_lock = ldlm_handle2lock(&lockh[0]);
3757         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3758                 RETURN(0);
3759
3760         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3761                  it->opc, lockh[0].cookie);
3762
3763         /* If we've already given this lock to a client once, then we should
3764          * have no readers or writers.  Otherwise, we should have one reader
3765          * _or_ writer ref (which will be zeroed below) before returning the
3766          * lock to a client. */
3767         if (new_lock->l_export == req->rq_export) {
3768                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3769         } else {
3770                 LASSERT(new_lock->l_export == NULL);
3771                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3772         }
3773
3774         *lockp = new_lock;
3775
3776         if (new_lock->l_export == req->rq_export) {
3777                 /* Already gave this to the client, which means that we
3778                  * reconstructed a reply. */
3779                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3780                         MSG_RESENT);
3781                 RETURN(ELDLM_LOCK_REPLACED);
3782         }
3783
3784         /* Fixup the lock to be given to the client */
3785         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3786         new_lock->l_readers = 0;
3787         new_lock->l_writers = 0;
3788
3789         new_lock->l_export = class_export_get(req->rq_export);
3790         list_add(&new_lock->l_export_chain,
3791                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
3792
3793         new_lock->l_blocking_ast = lock->l_blocking_ast;
3794         new_lock->l_completion_ast = lock->l_completion_ast;
3795
3796         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3797                sizeof(lock->l_remote_handle));
3798
3799         new_lock->l_flags &= ~LDLM_FL_LOCAL;
3800
3801         LDLM_LOCK_PUT(new_lock);
3802         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3803
3804         RETURN(ELDLM_LOCK_REPLACED);
3805 }
3806
3807 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3808 {
3809         struct lprocfs_static_vars lvars;
3810         int rc = 0;
3811
3812         lprocfs_init_multi_vars(0, &lvars);
3813
3814         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3815         if (rc)
3816                 return rc;
3817
3818         return lprocfs_alloc_md_stats(dev, 0);
3819 }
3820
3821 int mds_detach(struct obd_device *dev)
3822 {
3823         lprocfs_free_md_stats(dev);
3824         return lprocfs_obd_detach(dev);
3825 }
3826
3827 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3828 {
3829         struct lprocfs_static_vars lvars;
3830
3831         lprocfs_init_multi_vars(1, &lvars);
3832         return lprocfs_obd_attach(dev, lvars.obd_vars);
3833 }
3834
3835 int mdt_detach(struct obd_device *dev)
3836 {
3837         return lprocfs_obd_detach(dev);
3838 }
3839
3840 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3841 {
3842         struct mds_obd *mds = &obd->u.mds;
3843         int rc = 0;
3844         ENTRY;
3845
3846         mds->mds_service =
3847                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3848                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3849                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3850                                 mds_handle, "mds", obd->obd_proc_entry);
3851
3852         if (!mds->mds_service) {
3853                 CERROR("failed to start service\n");
3854                 RETURN(-ENOMEM);
3855         }
3856
3857         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3858                                     "ll_mdt");
3859         if (rc)
3860                 GOTO(err_thread, rc);
3861
3862         mds->mds_setattr_service =
3863                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3864                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3865                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3866                                 mds_handle, "mds_setattr",
3867                                 obd->obd_proc_entry);
3868         if (!mds->mds_setattr_service) {
3869                 CERROR("failed to start getattr service\n");
3870                 GOTO(err_thread, rc = -ENOMEM);
3871         }
3872
3873         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3874                                     MDT_NUM_THREADS, "ll_mdt_attr");
3875         if (rc)
3876                 GOTO(err_thread2, rc);
3877
3878         mds->mds_readpage_service =
3879                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3880                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3881                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3882                                 mds_handle, "mds_readpage",
3883                                 obd->obd_proc_entry);
3884         if (!mds->mds_readpage_service) {
3885                 CERROR("failed to start readpage service\n");
3886                 GOTO(err_thread2, rc = -ENOMEM);
3887         }
3888
3889         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3890                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
3891
3892         if (rc)
3893                 GOTO(err_thread3, rc);
3894
3895         RETURN(0);
3896
3897 err_thread3:
3898         ptlrpc_unregister_service(mds->mds_readpage_service);
3899 err_thread2:
3900         ptlrpc_unregister_service(mds->mds_setattr_service);
3901 err_thread:
3902         ptlrpc_unregister_service(mds->mds_service);
3903         return rc;
3904 }
3905
3906 static int mdt_cleanup(struct obd_device *obd, int flags)
3907 {
3908         struct mds_obd *mds = &obd->u.mds;
3909         ENTRY;
3910
3911         ptlrpc_stop_all_threads(mds->mds_readpage_service);
3912         ptlrpc_unregister_service(mds->mds_readpage_service);
3913
3914         ptlrpc_stop_all_threads(mds->mds_setattr_service);
3915         ptlrpc_unregister_service(mds->mds_setattr_service);
3916
3917         ptlrpc_stop_all_threads(mds->mds_service);
3918         ptlrpc_unregister_service(mds->mds_service);
3919
3920         RETURN(0);
3921 }
3922
3923 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3924                                          __u64 gr, void *data)
3925 {
3926         struct lustre_id id;
3927         struct obd_device *obd = data;
3928
3929         id_ino(&id) = ino;
3930         id_gen(&id) = gen;
3931         return mds_id2dentry(obd, &id, NULL);
3932 }
3933
3934 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3935                         void *key, __u32 *valsize, void *val)
3936 {
3937         struct obd_device *obd;
3938         struct mds_obd *mds;
3939         ENTRY;
3940
3941         obd = class_exp2obd(exp);
3942         mds = &obd->u.mds;
3943
3944         if (obd == NULL) {
3945                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3946                        exp->exp_handle.h_cookie);
3947                 RETURN(-EINVAL);
3948         }
3949
3950         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3951                 /* get log_context handle. */
3952                 unsigned long *llh_handle = val;
3953                 *valsize = sizeof(unsigned long);
3954                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3955                 RETURN(0);
3956         }
3957         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3958                 /* get log_context handle. */
3959                 unsigned long *sb = val;
3960                 *valsize = sizeof(unsigned long);
3961                 *sb = (unsigned long)obd->u.mds.mds_sb;
3962                 RETURN(0);
3963         }
3964
3965         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3966                 __u32 *mdsize = val;
3967                 *valsize = sizeof(*mdsize);
3968                 *mdsize = mds->mds_max_mdsize;
3969                 RETURN(0);
3970         }
3971
3972         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3973                 __u32 *mdsnum = val;
3974                 *valsize = sizeof(*mdsnum);
3975                 *mdsnum = mds->mds_num;
3976                 RETURN(0);
3977         }
3978
3979         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3980                 struct lustre_id *rootid = val;
3981                 *valsize = sizeof(struct lustre_id);
3982                 *rootid = mds->mds_rootid;
3983                 RETURN(0);
3984         }
3985
3986         CDEBUG(D_IOCTL, "invalid key\n");
3987         RETURN(-EINVAL);
3988
3989 }
3990 struct lvfs_callback_ops mds_lvfs_ops = {
3991         l_id2dentry:     mds_lvfs_id2dentry,
3992 };
3993
3994 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3995                 int objcount, struct obd_ioobj *obj,
3996                 int niocount, struct niobuf_remote *nb,
3997                 struct niobuf_local *res,
3998                 struct obd_trans_info *oti);
3999
4000 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4001                  int objcount, struct obd_ioobj *obj, int niocount,
4002                  struct niobuf_local *res, struct obd_trans_info *oti,
4003                  int rc);
4004
4005 /* use obd ops to offer management infrastructure */
4006 static struct obd_ops mds_obd_ops = {
4007         .o_owner           = THIS_MODULE,
4008         .o_attach          = mds_attach,
4009         .o_detach          = mds_detach,
4010         .o_connect         = mds_connect,
4011         .o_connect_post    = mds_connect_post,
4012         .o_init_export     = mds_init_export,
4013         .o_destroy_export  = mds_destroy_export,
4014         .o_disconnect      = mds_disconnect,
4015         .o_setup           = mds_setup,
4016         .o_precleanup      = mds_precleanup,
4017         .o_cleanup         = mds_cleanup,
4018         .o_process_config  = mds_process_config,
4019         .o_postrecov       = mds_postrecov,
4020         .o_statfs          = mds_obd_statfs,
4021         .o_iocontrol       = mds_iocontrol,
4022         .o_create          = mds_obd_create,
4023         .o_destroy         = mds_obd_destroy,
4024         .o_llog_init       = mds_llog_init,
4025         .o_llog_finish     = mds_llog_finish,
4026         .o_notify          = mds_notify,
4027         .o_get_info        = mds_get_info,
4028         .o_set_info        = mds_set_info,
4029         .o_preprw          = mds_preprw,
4030         .o_commitrw        = mds_commitrw,
4031 };
4032
4033 static struct obd_ops mdt_obd_ops = {
4034         .o_owner           = THIS_MODULE,
4035         .o_attach          = mdt_attach,
4036         .o_detach          = mdt_detach,
4037         .o_setup           = mdt_setup,
4038         .o_cleanup         = mdt_cleanup,
4039 };
4040
4041 static int __init mds_init(void)
4042 {
4043         struct lprocfs_static_vars lvars;
4044
4045         mds_init_lsd_cache();
4046
4047         lprocfs_init_multi_vars(0, &lvars);
4048         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4049                             LUSTRE_MDS_NAME);
4050         lprocfs_init_multi_vars(1, &lvars);
4051         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4052                             LUSTRE_MDT_NAME);
4053
4054         return 0;
4055 }
4056
4057 static void /*__exit*/ mds_exit(void)
4058 {
4059         mds_cleanup_lsd_cache();
4060
4061         class_unregister_type(LUSTRE_MDS_NAME);
4062         class_unregister_type(LUSTRE_MDT_NAME);
4063 }
4064
4065 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4066 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4067 MODULE_LICENSE("GPL");
4068
4069 module_init(mds_init);
4070 module_exit(mds_exit);