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                         LASSERT(!IS_ERR(dchild));
1451                 } else {
1452                         /* client wants to get attr. by id */
1453                         dchild = mds_id2dentry(obd, &body->id1, NULL);
1454                         LASSERT(!IS_ERR(dchild));
1455                 }
1456                 LDLM_LOCK_PUT(granted_lock);
1457         }
1458
1459         cleanup_phase = 2; /* dchild, dparent, locks */
1460
1461         if (!DENTRY_VALID(dchild)) {
1462                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1463                 /*
1464                  * in the intent case, the policy clears this error: the
1465                  * disposition is enough.
1466                  */
1467                 rc = -ENOENT;
1468                 GOTO(cleanup, rc);
1469         } else {
1470                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1471         }
1472
1473         if (req->rq_repmsg == NULL) {
1474                 if (dchild->d_flags & DCACHE_CROSS_REF)
1475                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1476                 else
1477                         rc = mds_getattr_pack_msg(req, dchild, offset);
1478                 if (rc != 0) {
1479                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1480                         GOTO (cleanup, rc);
1481                 }
1482         }
1483
1484         rc = mds_getattr_internal(obd, dchild, req, offset, body, reply_offset);
1485         GOTO(cleanup, rc); /* returns the lock to the client */
1486
1487  cleanup:
1488         switch (cleanup_phase) {
1489         case 2:
1490                 if (resent_req == 0) {
1491                         if (rc && DENTRY_VALID(dchild))
1492                                 ldlm_lock_decref(child_lockh, LCK_PR);
1493                         if (name)
1494                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1495 #ifdef S_PDIROPS
1496                         if (parent_lockh[1].cookie != 0)
1497                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1498 #endif
1499                         if (dparent)
1500                                 l_dput(dparent);
1501                 }
1502                 l_dput(dchild);
1503         case 1:
1504                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1505         default:
1506                 mds_exit_ucred(&uc);
1507         }
1508         return rc;
1509 }
1510
1511 static int mds_getattr(struct ptlrpc_request *req, int offset)
1512 {
1513         struct obd_device *obd = req->rq_export->exp_obd;
1514         struct lvfs_run_ctxt saved;
1515         struct dentry *de;
1516         struct mds_req_sec_desc *rsd;
1517         struct mds_body *body;
1518         struct lvfs_ucred uc = {NULL, NULL,};
1519         int rc = 0;
1520         ENTRY;
1521
1522         MD_COUNTER_INCREMENT(obd, getattr);
1523
1524         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1525         if (!rsd) {
1526                 CERROR("Can't unpack security desc\n");
1527                 RETURN(-EFAULT);
1528         }
1529
1530         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1531                                   lustre_swab_mds_body);
1532         if (body == NULL) {
1533                 CERROR ("Can't unpack body\n");
1534                 RETURN (-EFAULT);
1535         }
1536
1537         rc = mds_init_ucred(&uc, req, rsd);
1538         if (rc) {
1539                 mds_exit_ucred(&uc);
1540                 RETURN(rc);
1541         }
1542
1543         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1544         de = mds_id2dentry(obd, &body->id1, NULL);
1545         if (IS_ERR(de)) {
1546                 rc = req->rq_status = PTR_ERR(de);
1547                 GOTO(out_pop, rc);
1548         }
1549
1550         rc = mds_getattr_pack_msg(req, de, offset);
1551         if (rc != 0) {
1552                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1553                 GOTO(out_pop, rc);
1554         }
1555
1556         req->rq_status = mds_getattr_internal(obd, de, req, offset, body, 0);
1557         l_dput(de);
1558
1559         EXIT;
1560 out_pop:
1561         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1562         mds_exit_ucred(&uc);
1563         return rc;
1564 }
1565
1566 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1567                           unsigned long max_age)
1568 {
1569         int rc;
1570         ENTRY;
1571
1572         spin_lock(&obd->obd_osfs_lock);
1573         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1574         if (rc == 0)
1575                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1576         spin_unlock(&obd->obd_osfs_lock);
1577
1578         RETURN(rc);
1579 }
1580
1581 static int mds_statfs(struct ptlrpc_request *req)
1582 {
1583         struct obd_device *obd = req->rq_export->exp_obd;
1584         int rc, size = sizeof(struct obd_statfs);
1585         ENTRY;
1586
1587         /* This will trigger a watchdog timeout */
1588         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1589                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1590
1591         rc = lustre_pack_reply(req, 1, &size, NULL);
1592         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1593                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1594                 GOTO(out, rc);
1595         }
1596
1597         OBD_COUNTER_INCREMENT(obd, statfs);
1598
1599         /* We call this so that we can cache a bit - 1 jiffie worth */
1600         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1601                             jiffies - HZ);
1602         if (rc) {
1603                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1604                 GOTO(out, rc);
1605         }
1606
1607         EXIT;
1608 out:
1609         req->rq_status = rc;
1610         return rc;
1611 }
1612
1613 static int mds_sync(struct ptlrpc_request *req, int offset)
1614 {
1615         struct obd_device *obd = req->rq_export->exp_obd;
1616         struct mds_obd *mds = &obd->u.mds;
1617         struct mds_body *body;
1618         int rc, size = sizeof(*body);
1619         ENTRY;
1620
1621         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1622                                   lustre_swab_mds_body);
1623         if (body == NULL)
1624                 GOTO(out, rc = -EPROTO);
1625
1626         rc = lustre_pack_reply(req, 1, &size, NULL);
1627         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1628                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1629                 GOTO(out, rc);
1630         }
1631
1632         if (id_ino(&body->id1) == 0) {
1633                 /* an id of zero is taken to mean "sync whole filesystem" */
1634                 rc = fsfilt_sync(obd, mds->mds_sb);
1635                 if (rc)
1636                         GOTO(out, rc);
1637         } else {
1638                 /* just any file to grab fsync method - "file" arg unused */
1639                 struct file *file = mds->mds_rcvd_filp;
1640                 struct mds_body *rep_body;
1641                 struct dentry *de;
1642
1643                 de = mds_id2dentry(obd, &body->id1, NULL);
1644                 if (IS_ERR(de))
1645                         GOTO(out, rc = PTR_ERR(de));
1646
1647                 rc = file->f_op->fsync(NULL, de, 1);
1648                 if (rc)
1649                         GOTO(out, rc);
1650
1651                 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1652                 mds_pack_inode2body(obd, rep_body, de->d_inode,
1653                                     (body->valid & OBD_MD_FID) ? 1 : 0);
1654                 l_dput(de);
1655         }
1656
1657         EXIT;
1658 out:
1659         req->rq_status = rc;
1660         return rc;
1661 }
1662
1663 /* mds_readpage does not take a DLM lock on the inode, because the client must
1664  * already have a PR lock.
1665  *
1666  * If we were to take another one here, a deadlock will result, if another
1667  * thread is already waiting for a PW lock. */
1668 static int mds_readpage(struct ptlrpc_request *req, int offset)
1669 {
1670         struct obd_device *obd = req->rq_export->exp_obd;
1671         struct vfsmount *mnt;
1672         struct dentry *de;
1673         struct file *file;
1674         struct mds_req_sec_desc *rsd;
1675         struct mds_body *body, *repbody;
1676         struct lvfs_run_ctxt saved;
1677         int rc, size = sizeof(*repbody);
1678         struct lvfs_ucred uc = {NULL, NULL,};
1679         ENTRY;
1680
1681         rc = lustre_pack_reply(req, 1, &size, NULL);
1682         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1683                 CERROR("mds: out of memory\n");
1684                 GOTO(out, rc = -ENOMEM);
1685         }
1686
1687         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1688         if (!rsd) {
1689                 CERROR("Can't unpack security desc\n");
1690                 GOTO (out, rc = -EFAULT);
1691         }
1692
1693         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1694                                   lustre_swab_mds_body);
1695         if (body == NULL) {
1696                 CERROR("Can't unpack body\n");
1697                 GOTO (out, rc = -EFAULT);
1698         }
1699
1700         rc = mds_init_ucred(&uc, req, rsd);
1701         if (rc) {
1702                 GOTO(out, rc);
1703         }
1704
1705         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1706         de = mds_id2dentry(obd, &body->id1, &mnt);
1707         if (IS_ERR(de))
1708                 GOTO(out_pop, rc = PTR_ERR(de));
1709
1710         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1711
1712         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1713         /* note: in case of an error, dentry_open puts dentry */
1714         if (IS_ERR(file))
1715                 GOTO(out_pop, rc = PTR_ERR(file));
1716
1717         /* body->size is actually the offset -eeb */
1718         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1719                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1720                        body->size, de->d_inode->i_blksize);
1721                 GOTO(out_file, rc = -EFAULT);
1722         }
1723
1724         /* body->nlink is actually the #bytes to read -eeb */
1725         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1726                 CERROR("size %u is not multiple of blocksize %lu\n",
1727                        body->nlink, de->d_inode->i_blksize);
1728                 GOTO(out_file, rc = -EFAULT);
1729         }
1730
1731         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1732         repbody->size = file->f_dentry->d_inode->i_size;
1733         repbody->valid = OBD_MD_FLSIZE;
1734
1735         /* to make this asynchronous make sure that the handling function
1736            doesn't send a reply when this function completes. Instead a
1737            callback function would send the reply */
1738         /* body->size is actually the offset -eeb */
1739         rc = mds_sendpage(req, file, body->size, body->nlink);
1740
1741         EXIT;
1742 out_file:
1743         filp_close(file, 0);
1744 out_pop:
1745         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1746 out:
1747         mds_exit_ucred(&uc);
1748         req->rq_status = rc;
1749         return 0;
1750 }
1751
1752 /* update master MDS ID, which is stored in local inode EA. */
1753 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1754                    void *data, int data_len)
1755 {
1756         struct mds_obd *mds = &obd->u.mds;
1757         struct dentry *dentry;
1758         void *handle;
1759         int rc = 0;
1760         ENTRY;
1761
1762         LASSERT(id);
1763         LASSERT(obd);
1764
1765         dentry = mds_id2dentry(obd, id, NULL);
1766         if (IS_ERR(dentry))
1767                 GOTO(out, rc = PTR_ERR(dentry));
1768
1769         if (!dentry->d_inode) {
1770                 CERROR("Can't find object "DLID4".\n",
1771                        OLID4(id));
1772                 GOTO(out_dentry, rc = -EINVAL);
1773         }
1774
1775         handle = fsfilt_start(obd, dentry->d_inode,
1776                               FSFILT_OP_SETATTR, NULL);
1777         if (IS_ERR(handle))
1778                 GOTO(out_dentry, rc = PTR_ERR(handle));
1779
1780         rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1781                                   (struct lustre_id *)data);
1782         if (rc) {
1783                 CERROR("Can't update inode "DLID4" master id, "
1784                        "error = %d.\n", OLID4(id), rc);
1785                 GOTO(out_commit, rc);
1786         }
1787
1788         EXIT;
1789 out_commit:
1790         fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1791                       handle, 0);
1792 out_dentry:
1793         l_dput(dentry);
1794 out:
1795         return rc;
1796 }
1797 EXPORT_SYMBOL(mds_update_mid);
1798
1799 /* read master MDS ID, which is stored in local inode EA. */
1800 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1801                  void *data, int data_len)
1802 {
1803         struct dentry *dentry;
1804         int rc = 0;
1805         ENTRY;
1806
1807         LASSERT(id);
1808         LASSERT(obd);
1809
1810         dentry = mds_id2dentry(obd, id, NULL);
1811         if (IS_ERR(dentry))
1812                 GOTO(out, rc = PTR_ERR(dentry));
1813
1814         if (!dentry->d_inode) {
1815                 CERROR("Can't find object "DLID4".\n",
1816                        OLID4(id));
1817                 GOTO(out_dentry, rc = -EINVAL);
1818         }
1819
1820         down(&dentry->d_inode->i_sem);
1821         rc = mds_read_inode_mid(obd, dentry->d_inode,
1822                                 (struct lustre_id *)data);
1823         up(&dentry->d_inode->i_sem);
1824         if (rc) {
1825                 CERROR("Can't read inode "DLID4" master id, "
1826                        "error = %d.\n", OLID4(id), rc);
1827                 GOTO(out_dentry, rc);
1828         }
1829
1830         EXIT;
1831 out_dentry:
1832         l_dput(dentry);
1833 out:
1834         return rc;
1835 }
1836 EXPORT_SYMBOL(mds_read_mid);
1837
1838 int mds_reint(struct ptlrpc_request *req, int offset,
1839               struct lustre_handle *lockh)
1840 {
1841         struct mds_update_record *rec;
1842         struct mds_req_sec_desc *rsd;
1843         int rc;
1844         ENTRY;
1845
1846         OBD_ALLOC(rec, sizeof(*rec));
1847         if (rec == NULL)
1848                 RETURN(-ENOMEM);
1849
1850         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1851         if (!rsd) {
1852                 CERROR("Can't unpack security desc\n");
1853                 GOTO(out, rc = -EFAULT);
1854         }
1855
1856         rc = mds_update_unpack(req, offset, rec);
1857         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1858                 CERROR("invalid record\n");
1859                 GOTO(out, req->rq_status = -EINVAL);
1860         }
1861
1862         rc = mds_init_ucred(&rec->ur_uc, req, rsd);
1863         if (rc) {
1864                 GOTO(out, rc);
1865         }
1866
1867         /* rc will be used to interrupt a for loop over multiple records */
1868         rc = mds_reint_rec(rec, offset, req, lockh);
1869
1870  out:
1871         mds_exit_ucred(&rec->ur_uc);
1872         OBD_FREE(rec, sizeof(*rec));
1873         RETURN(rc);
1874 }
1875
1876 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1877                                        struct obd_device *obd, int *process)
1878 {
1879         switch (req->rq_reqmsg->opc) {
1880         case MDS_CONNECT: /* This will never get here, but for completeness. */
1881         case OST_CONNECT: /* This will never get here, but for completeness. */
1882         case MDS_DISCONNECT:
1883         case OST_DISCONNECT:
1884                *process = 1;
1885                RETURN(0);
1886
1887         case MDS_CLOSE:
1888         case MDS_SYNC: /* used in unmounting */
1889         case OBD_PING:
1890         case MDS_REINT:
1891         case LDLM_ENQUEUE:
1892         case OST_CREATE:
1893                 *process = target_queue_recovery_request(req, obd);
1894                 RETURN(0);
1895
1896         default:
1897                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1898                 *process = 0;
1899                 /* XXX what should we set rq_status to here? */
1900                 req->rq_status = -EAGAIN;
1901                 RETURN(ptlrpc_error(req));
1902         }
1903 }
1904
1905 static char *reint_names[] = {
1906         [REINT_SETATTR] "setattr",
1907         [REINT_CREATE]  "create",
1908         [REINT_LINK]    "link",
1909         [REINT_UNLINK]  "unlink",
1910         [REINT_RENAME]  "rename",
1911         [REINT_OPEN]    "open",
1912 };
1913
1914 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
1915                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1916                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1917                             OBD_MD_FLID)
1918
1919 static void reconstruct_create(struct ptlrpc_request *req)
1920 {
1921         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1922         struct mds_client_data *mcd = med->med_mcd;
1923         struct dentry *dentry;
1924         struct ost_body *body;
1925         struct lustre_id id;
1926         int rc;
1927         ENTRY;
1928
1929         /* copy rc, transno and disp; steal locks */
1930         mds_req_from_mcd(req, mcd);
1931         if (req->rq_status) {
1932                 EXIT;
1933                 return;
1934         }
1935
1936         id_gen(&id) = 0;
1937         id_group(&id) = 0;
1938
1939         id_ino(&id) = mcd->mcd_last_data;
1940         LASSERT(id_ino(&id) != 0);
1941
1942         dentry = mds_id2dentry(req2obd(req), &id, NULL);
1943         if (IS_ERR(dentry)) {
1944                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1945                 req->rq_status = PTR_ERR(dentry);
1946                 EXIT;
1947                 return;
1948         }
1949
1950         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1951               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1952               dentry->d_inode->i_generation);
1953
1954         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1955         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1956         body->oa.o_id = dentry->d_inode->i_ino;
1957         body->oa.o_generation = dentry->d_inode->i_generation;
1958         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1959
1960         down(&dentry->d_inode->i_sem);
1961         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1962         up(&dentry->d_inode->i_sem);
1963         if (rc) {
1964                 CERROR("Can't read inode self id, inode %lu, "
1965                        "rc %d\n", dentry->d_inode->i_ino, rc);
1966                 id_fid(&id) = 0;
1967         }
1968
1969         body->oa.o_fid = id_fid(&id);
1970         body->oa.o_mds = id_group(&id);
1971         l_dput(dentry);
1972
1973         EXIT;
1974 }
1975
1976 static int mdt_obj_create(struct ptlrpc_request *req)
1977 {
1978         struct obd_device *obd = req->rq_export->exp_obd;
1979         struct mds_obd *mds = &obd->u.mds;
1980         struct ost_body *body, *repbody;
1981         char idname[LL_ID_NAMELEN];
1982         int size = sizeof(*repbody);
1983         struct inode *parent_inode;
1984         struct lvfs_run_ctxt saved;
1985         int rc, cleanup_phase = 0;
1986         struct dentry *new = NULL;
1987         struct dentry_params dp;
1988         int mealen, flags = 0;
1989         struct lvfs_ucred uc;
1990         struct lustre_id id;
1991         struct mea *mea;
1992         void *handle = NULL;
1993         unsigned long cr_inum = 0;
1994         ENTRY;
1995
1996         DEBUG_REQ(D_HA, req, "create remote object");
1997         parent_inode = mds->mds_unnamed_dir->d_inode;
1998
1999         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2000                                   lustre_swab_ost_body);
2001         if (body == NULL)
2002                 RETURN(-EFAULT);
2003
2004         rc = lustre_pack_reply(req, 1, &size, NULL);
2005         if (rc)
2006                 RETURN(rc);
2007
2008         MDS_CHECK_RESENT(req, reconstruct_create(req));
2009
2010         uc.luc_lsd = NULL;
2011         uc.luc_ginfo = NULL;
2012         uc.luc_uid = body->oa.o_uid;
2013         uc.luc_gid = body->oa.o_gid;
2014         uc.luc_fsuid = body->oa.o_uid;
2015         uc.luc_fsgid = body->oa.o_gid;
2016
2017         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2018         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2019
2020         /* in REPLAY case inum should be given (client or other MDS fills it) */
2021         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2022             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2023                 /*
2024                  * this is re-create request from MDS holding directory name.
2025                  * we have to lookup given ino/gen first. if it exists (good
2026                  * case) then there is nothing to do. if it does not then we
2027                  * have to recreate it.
2028                  */
2029                 id_ino(&id) = body->oa.o_id;
2030                 id_gen(&id) = body->oa.o_generation;
2031
2032                 new = mds_id2dentry(obd, &id, NULL);
2033                 if (!IS_ERR(new) && new->d_inode) {
2034                         struct lustre_id sid;
2035
2036                         CWARN("mkdir() repairing is on its way: %lu/%lu\n",
2037                               (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
2038
2039                         obdo_from_inode(&repbody->oa, new->d_inode,
2040                                         FILTER_VALID_FLAGS);
2041
2042                         repbody->oa.o_id = new->d_inode->i_ino;
2043                         repbody->oa.o_generation = new->d_inode->i_generation;
2044                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2045                         cleanup_phase = 1;
2046
2047                         down(&new->d_inode->i_sem);
2048                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2049                         up(&new->d_inode->i_sem);
2050                         if (rc) {
2051                                 CERROR("Can't read inode self id "
2052                                        "inode %lu, rc %d.\n",
2053                                        new->d_inode->i_ino, rc);
2054                                 GOTO(cleanup, rc);
2055                         }
2056
2057                         repbody->oa.o_fid = id_fid(&sid);
2058                         repbody->oa.o_mds = id_group(&sid);
2059                         LASSERT(id_fid(&sid) != 0);
2060
2061                         /*
2062                          * here we could use fid passed in body->oa.o_fid and
2063                          * thus avoid mds_read_inode_sid().
2064                          */
2065                         cr_inum = new->d_inode->i_ino;
2066                         GOTO(cleanup, rc = 0);
2067                 }
2068         }
2069
2070         down(&parent_inode->i_sem);
2071         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2072         if (IS_ERR(handle)) {
2073                 up(&parent_inode->i_sem);
2074                 CERROR("fsfilt_start() failed, rc = %d\n",
2075                        (int)PTR_ERR(handle));
2076                 GOTO(cleanup, rc = PTR_ERR(handle));
2077         }
2078         cleanup_phase = 1; /* transaction */
2079
2080 repeat:
2081         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2082         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2083         if (IS_ERR(new)) {
2084                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2085                        obd->obd_name, idname, (int) PTR_ERR(new));
2086                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2087                 up(&parent_inode->i_sem);
2088                 RETURN(PTR_ERR(new));
2089         } else if (new->d_inode) {
2090                 CERROR("%s: name exists. repeat\n", obd->obd_name);
2091                 goto repeat;
2092         }
2093
2094         new->d_fsdata = (void *)&dp;
2095         dp.p_inum = 0;
2096         dp.p_ptr = req;
2097
2098         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2099             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2100                 LASSERT(body->oa.o_id != 0);
2101                 dp.p_inum = body->oa.o_id;
2102                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2103                           (unsigned long)body->oa.o_id,
2104                           (unsigned long)body->oa.o_generation);
2105         }
2106
2107         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2108         if (rc == 0) {
2109                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2110                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2111                         new->d_inode->i_generation = body->oa.o_generation;
2112                         mark_inode_dirty(new->d_inode);
2113
2114                         /*
2115                          * avoiding asserts in cache flush case, as
2116                          * @body->oa.o_id should be zero.
2117                          */
2118                         if (body->oa.o_id) {
2119                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2120                                          "BUG 3550: failed to recreate obj "
2121                                          LPU64" -> %lu\n", body->oa.o_id,
2122                                          new->d_inode->i_ino);
2123
2124                                 LASSERTF(body->oa.o_generation ==
2125                                          new->d_inode->i_generation,
2126                                          "BUG 3550: failed to recreate obj/gen "
2127                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2128                                          body->oa.o_generation,
2129                                          new->d_inode->i_ino,
2130                                          new->d_inode->i_generation);
2131                         }
2132                 }
2133
2134                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2135                 repbody->oa.o_id = new->d_inode->i_ino;
2136                 repbody->oa.o_generation = new->d_inode->i_generation;
2137                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2138
2139                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2140                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2141                         id_group(&id) = mds->mds_num;
2142
2143                         LASSERT(body->oa.o_fid != 0);
2144                         id_fid(&id) = body->oa.o_fid;
2145
2146                         LASSERT(body->oa.o_id != 0);
2147                         id_ino(&id) = repbody->oa.o_id;
2148                         id_gen(&id) = repbody->oa.o_generation;
2149
2150                         down(&new->d_inode->i_sem);
2151                         rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2152                         up(&new->d_inode->i_sem);
2153
2154                         /*
2155                          * make sure, that fid is up-to-date.
2156                          */
2157                         mds_set_last_fid(obd, id_fid(&id));
2158                 } else {
2159                         /*
2160                          * allocate new sid, as object is created from scratch
2161                          * and this is not replay.
2162                          */
2163                         down(&new->d_inode->i_sem);
2164                         rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2165                         up(&new->d_inode->i_sem);
2166                 }
2167                 if (rc) {
2168                         CERROR("Can't update lustre ID for inode %lu, "
2169                                "error = %d\n", new->d_inode->i_ino, rc);
2170                         GOTO(cleanup, rc);
2171                 }
2172
2173                 /* initializing o_fid after it is allocated. */
2174                 repbody->oa.o_fid = id_fid(&id);
2175                 repbody->oa.o_mds = id_group(&id);
2176
2177                 rc = fsfilt_del_dir_entry(obd, new);
2178                 up(&parent_inode->i_sem);
2179                 if (rc) {
2180                         CERROR("can't remove name for object: %d\n", rc);
2181                         GOTO(cleanup, rc);
2182                 }
2183
2184                 cleanup_phase = 2; /* created directory object */
2185
2186                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2187                        (unsigned long)new->d_inode->i_ino,
2188                        (unsigned long)new->d_inode->i_generation,
2189                        (unsigned)new->d_inode->i_mode);
2190                 cr_inum = new->d_inode->i_ino;
2191         } else {
2192                 up(&parent_inode->i_sem);
2193                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2194                 GOTO(cleanup, rc);
2195         }
2196
2197         if (body->oa.o_valid & OBD_MD_FLID) {
2198                 /* this is new object for splitted dir. We have to prevent
2199                  * recursive splitting on it -bzzz */
2200                 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2201
2202                 OBD_ALLOC(mea, mealen);
2203                 if (mea == NULL)
2204                         GOTO(cleanup, rc = -ENOMEM);
2205
2206                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2207                 mea->mea_master = 0;
2208                 mea->mea_count = 0;
2209
2210                 down(&new->d_inode->i_sem);
2211                 rc = fsfilt_set_md(obd, new->d_inode, handle,
2212                                    mea, mealen, EA_MEA);
2213                 up(&new->d_inode->i_sem);
2214                 if (rc)
2215                         CERROR("fsfilt_set_md() failed, "
2216                                "rc = %d\n", rc);
2217
2218                 OBD_FREE(mea, mealen);
2219
2220                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2221                        obd->obd_name, new->d_inode->i_ino,
2222                        new->d_inode->i_generation, flags);
2223         } else if (body->oa.o_easize) {
2224                 /* we pass LCK_EX to split routine to signal that we have
2225                  * exclusive access to the directory. simple because nobody
2226                  * knows it already exists -bzzz */
2227                 rc = mds_try_to_split_dir(obd, new, NULL,
2228                                           body->oa.o_easize, LCK_EX);
2229                 if (rc < 0) {
2230                         CERROR("Can't split directory %lu, error = %d.\n",
2231                                new->d_inode->i_ino, rc);
2232                 } else {
2233                         rc = 0;
2234                 }
2235         }
2236
2237         EXIT;
2238 cleanup:
2239         switch (cleanup_phase) {
2240         case 2: /* object has been created, but we'll may want to replay it later */
2241                 if (rc == 0)
2242                         ptlrpc_require_repack(req);
2243         case 1: /* transaction */
2244                 rc = mds_finish_transno(mds, parent_inode, handle,
2245                                         req, rc, cr_inum);
2246         }
2247
2248         l_dput(new);
2249         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2250         return rc;
2251 }
2252
2253 static int mdt_get_info(struct ptlrpc_request *req)
2254 {
2255         struct obd_export *exp = req->rq_export;
2256         int keylen, rc = 0;
2257         char *key;
2258         ENTRY;
2259
2260         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2261         if (key == NULL) {
2262                 DEBUG_REQ(D_HA, req, "no get_info key");
2263                 RETURN(-EFAULT);
2264         }
2265         keylen = req->rq_reqmsg->buflens[0];
2266
2267         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2268             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2269             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2270                 RETURN(-EPROTO);
2271
2272         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2273                 struct lustre_id *reply;
2274                 int size = sizeof(*reply);
2275
2276                 rc = lustre_pack_reply(req, 1, &size, NULL);
2277                 if (rc)
2278                         RETURN(rc);
2279
2280                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2281                 rc = obd_get_info(exp, keylen, key, &size, reply);
2282         } else {
2283                 obd_id *reply;
2284                 int size = sizeof(*reply);
2285
2286                 rc = lustre_pack_reply(req, 1, &size, NULL);
2287                 if (rc)
2288                         RETURN(rc);
2289
2290                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2291                 rc = obd_get_info(exp, keylen, key, &size, reply);
2292         }
2293
2294         req->rq_repmsg->status = 0;
2295         RETURN(rc);
2296 }
2297
2298 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2299                         void *key, __u32 vallen, void *val)
2300 {
2301         struct obd_device *obd;
2302         struct mds_obd *mds;
2303         int rc = 0;
2304         ENTRY;
2305
2306         obd = class_exp2obd(exp);
2307         if (obd == NULL) {
2308                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2309                        exp->exp_handle.h_cookie);
2310                 RETURN(-EINVAL);
2311         }
2312
2313         mds = &obd->u.mds;
2314         if (keylen >= strlen("mds_type") &&
2315              memcmp(key, "mds_type", keylen) == 0) {
2316                 int valsize;
2317                 __u32 group;
2318
2319                 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2320
2321                 mds->mds_obd_type = *(int*)val;
2322                 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2323                 valsize = sizeof(group);
2324
2325                 /* mds number has been changed, so the corresponding obdfilter
2326                  * exp need to be changed too. */
2327                 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2328                                   "mds_conn", valsize, &group);
2329                 RETURN(rc);
2330         }
2331         CDEBUG(D_IOCTL, "invalid key\n");
2332         RETURN(-EINVAL);
2333 }
2334
2335 static int mdt_set_info(struct ptlrpc_request *req)
2336 {
2337         char *key, *val;
2338         struct obd_export *exp = req->rq_export;
2339         int keylen, rc = 0, vallen;
2340         ENTRY;
2341
2342         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2343         if (key == NULL) {
2344                 DEBUG_REQ(D_HA, req, "no set_info key");
2345                 RETURN(-EFAULT);
2346         }
2347         keylen = req->rq_reqmsg->buflens[0];
2348
2349         if (keylen == strlen("mds_type") &&
2350             memcmp(key, "mds_type", keylen) == 0) {
2351                 rc = lustre_pack_reply(req, 0, NULL, NULL);
2352                 if (rc)
2353                         RETURN(rc);
2354
2355                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2356                 vallen = req->rq_reqmsg->buflens[1];
2357
2358                 rc = obd_set_info(exp, keylen, key, vallen, val);
2359                 req->rq_repmsg->status = 0;
2360                 RETURN(rc);
2361         }
2362         CDEBUG(D_IOCTL, "invalid key\n");
2363         RETURN(-EINVAL);
2364 }
2365
2366 static void mds_revoke_export_locks(struct obd_export *exp)
2367 {
2368         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
2369         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2370         struct ldlm_lock *lock, *next;
2371         struct ldlm_lock_desc desc;
2372
2373         if (!exp->u.eu_mds_data.med_remote)
2374                 return;
2375
2376         ENTRY;
2377         l_lock(&ns->ns_lock);
2378         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2379                 if (lock->l_req_mode != lock->l_granted_mode)
2380                         continue;
2381
2382                 LASSERT(lock->l_resource);
2383                 if (lock->l_resource->lr_type != LDLM_IBITS &&
2384                     lock->l_resource->lr_type != LDLM_PLAIN)
2385                         continue;
2386
2387                 if (lock->l_flags & LDLM_FL_AST_SENT)
2388                         continue;
2389
2390                 lock->l_flags |= LDLM_FL_AST_SENT;
2391
2392                 /* the desc just pretend to exclusive */
2393                 ldlm_lock2desc(lock, &desc);
2394                 desc.l_req_mode = LCK_EX;
2395                 desc.l_granted_mode = 0;
2396
2397                 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2398         }
2399         l_unlock(&ns->ns_lock);
2400         EXIT;
2401 }
2402
2403 static int mds_msg_check_version(struct lustre_msg *msg)
2404 {
2405         int rc;
2406
2407         switch (msg->opc) {
2408         case MDS_CONNECT:
2409         case MDS_DISCONNECT:
2410         case OBD_PING:
2411                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2412                 if (rc)
2413                         CERROR("bad opc %u version %08x, expecting %08x\n",
2414                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2415                 break;
2416         case MDS_STATFS:
2417         case MDS_GETSTATUS:
2418         case MDS_GETATTR:
2419         case MDS_GETATTR_LOCK:
2420         case MDS_READPAGE:
2421         case MDS_REINT:
2422         case MDS_CLOSE:
2423         case MDS_DONE_WRITING:
2424         case MDS_PIN:
2425         case MDS_SYNC:
2426                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2427                 if (rc)
2428                         CERROR("bad opc %u version %08x, expecting %08x\n",
2429                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2430                 break;
2431         case LDLM_ENQUEUE:
2432         case LDLM_CONVERT:
2433         case LDLM_BL_CALLBACK:
2434         case LDLM_CP_CALLBACK:
2435                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2436                 if (rc)
2437                         CERROR("bad opc %u version %08x, expecting %08x\n",
2438                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2439                 break;
2440         case OBD_LOG_CANCEL:
2441         case LLOG_ORIGIN_HANDLE_OPEN:
2442         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2443         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2444         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2445         case LLOG_ORIGIN_HANDLE_CLOSE:
2446         case LLOG_CATINFO:
2447                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2448                 if (rc)
2449                         CERROR("bad opc %u version %08x, expecting %08x\n",
2450                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2451                 break;
2452         case OST_CREATE:
2453         case OST_WRITE:
2454         case OST_GET_INFO:
2455         case OST_SET_INFO:
2456                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2457                 if (rc)
2458                         CERROR("bad opc %u version %08x, expecting %08x\n",
2459                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2460                 break;
2461         case SEC_INIT:
2462         case SEC_INIT_CONTINUE:
2463         case SEC_FINI:
2464                 rc = 0;
2465                 break;
2466         default:
2467                 CERROR("MDS unknown opcode %d\n", msg->opc);
2468                 rc = -ENOTSUPP;
2469                 break;
2470         }
2471
2472         return rc;
2473 }
2474
2475 int mds_handle(struct ptlrpc_request *req)
2476 {
2477         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2478         struct obd_device *obd = NULL;
2479         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2480         int rc = 0;
2481         ENTRY;
2482
2483         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2484
2485         rc = mds_msg_check_version(req->rq_reqmsg);
2486         if (rc) {
2487                 CERROR("MDS drop mal-formed request\n");
2488                 RETURN(rc);
2489         }
2490
2491         /* Security opc should NOT trigger any recovery events */
2492         if (req->rq_reqmsg->opc == SEC_INIT ||
2493             req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2494                 if (!req->rq_export)
2495                         GOTO(out, rc = 0);
2496
2497                 mds_req_add_idmapping(req, &req->rq_export->exp_mds_data);
2498                 mds_revoke_export_locks(req->rq_export);
2499                 GOTO(out, rc = 0);
2500         } else if (req->rq_reqmsg->opc == SEC_FINI) {
2501                 if (!req->rq_export)
2502                         GOTO(out, rc = 0);
2503
2504                 mds_req_del_idmapping(req, &req->rq_export->exp_mds_data);
2505                 mds_revoke_export_locks(req->rq_export);
2506                 GOTO(out, rc = 0);
2507         }
2508
2509         LASSERT(current->journal_info == NULL);
2510         /* XXX identical to OST */
2511         if (req->rq_reqmsg->opc != MDS_CONNECT) {
2512                 struct mds_export_data *med;
2513                 int recovering;
2514
2515                 if (req->rq_export == NULL) {
2516                         CERROR("operation %d on unconnected MDS from %s\n",
2517                                req->rq_reqmsg->opc,
2518                                req->rq_peerstr);
2519                         req->rq_status = -ENOTCONN;
2520                         GOTO(out, rc = -ENOTCONN);
2521                 }
2522
2523                 med = &req->rq_export->exp_mds_data;
2524                 obd = req->rq_export->exp_obd;
2525                 mds = &obd->u.mds;
2526
2527                 /* sanity check: if the xid matches, the request must
2528                  * be marked as a resent or replayed */
2529                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2530                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2531                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2532                                  (MSG_RESENT | MSG_REPLAY),
2533                                  "rq_xid "LPU64" matches last_xid, "
2534                                  "expected RESENT flag\n",
2535                                  req->rq_xid);
2536                 }
2537                 /* else: note the opposite is not always true; a
2538                  * RESENT req after a failover will usually not match
2539                  * the last_xid, since it was likely never
2540                  * committed. A REPLAYed request will almost never
2541                  * match the last xid, however it could for a
2542                  * committed, but still retained, open. */
2543
2544                 spin_lock_bh(&obd->obd_processing_task_lock);
2545                 recovering = obd->obd_recovering;
2546                 spin_unlock_bh(&obd->obd_processing_task_lock);
2547                 if (recovering) {
2548                         rc = mds_filter_recovery_request(req, obd,
2549                                                          &should_process);
2550                         if (rc || should_process == 0) {
2551                                 RETURN(rc);
2552                         } else if (should_process < 0) {
2553                                 req->rq_status = should_process;
2554                                 rc = ptlrpc_error(req);
2555                                 RETURN(rc);
2556                         }
2557                 }
2558         }
2559
2560         switch (req->rq_reqmsg->opc) {
2561         case MDS_CONNECT:
2562                 DEBUG_REQ(D_INODE, req, "connect");
2563                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2564                 rc = target_handle_connect(req);
2565                 if (!rc) {
2566                         struct mds_export_data *med;
2567
2568                         LASSERT(req->rq_export);
2569                         med = &req->rq_export->u.eu_mds_data;
2570                         mds_init_export_data(req, med);
2571                         mds_req_add_idmapping(req, med);
2572
2573                         /* Now that we have an export, set mds. */
2574                         obd = req->rq_export->exp_obd;
2575                         mds = mds_req2mds(req);
2576                 }
2577                 break;
2578
2579         case MDS_DISCONNECT:
2580                 DEBUG_REQ(D_INODE, req, "disconnect");
2581                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2582                 rc = target_handle_disconnect(req);
2583                 req->rq_status = rc;            /* superfluous? */
2584                 break;
2585
2586         case MDS_GETSTATUS:
2587                 DEBUG_REQ(D_INODE, req, "getstatus");
2588                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2589                 rc = mds_getstatus(req);
2590                 break;
2591
2592         case MDS_GETATTR:
2593                 DEBUG_REQ(D_INODE, req, "getattr");
2594                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2595                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2596                 break;
2597
2598         case MDS_GETATTR_LOCK: {
2599                 struct lustre_handle lockh;
2600                 DEBUG_REQ(D_INODE, req, "getattr_lock");
2601                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2602
2603                 /* If this request gets a reconstructed reply, we won't be
2604                  * acquiring any new locks in mds_getattr_lock, so we don't
2605                  * want to cancel.
2606                  */
2607                 lockh.cookie = 0;
2608                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2609                                       MDS_INODELOCK_UPDATE);
2610                 /* this non-intent call (from an ioctl) is special */
2611                 req->rq_status = rc;
2612                 if (rc == 0 && lockh.cookie)
2613                         ldlm_lock_decref(&lockh, LCK_PR);
2614                 break;
2615         }
2616         case MDS_STATFS:
2617                 DEBUG_REQ(D_INODE, req, "statfs");
2618                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2619                 rc = mds_statfs(req);
2620                 break;
2621
2622         case MDS_READPAGE:
2623                 DEBUG_REQ(D_INODE, req, "readpage");
2624                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2625                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2626
2627                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2628                         if (req->rq_reply_state) {
2629                                 lustre_free_reply_state (req->rq_reply_state);
2630                                 req->rq_reply_state = NULL;
2631                         }
2632                         RETURN(0);
2633                 }
2634
2635                 break;
2636         case MDS_REINT: {
2637                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2638                                              sizeof (*opcp));
2639                 __u32  opc;
2640                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2641                                mds->mds_max_cookiesize};
2642                 int bufcount;
2643
2644                 /* NB only peek inside req now; mds_reint() will swab it */
2645                 if (opcp == NULL) {
2646                         CERROR ("Can't inspect opcode\n");
2647                         rc = -EINVAL;
2648                         break;
2649                 }
2650                 opc = *opcp;
2651                 if (lustre_msg_swabbed (req->rq_reqmsg))
2652                         __swab32s(&opc);
2653
2654                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2655                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2656                            reint_names[opc] == NULL) ? reint_names[opc] :
2657                                                        "unknown opcode");
2658
2659                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2660
2661                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2662                         bufcount = 3;
2663                 else if (opc == REINT_OPEN)
2664                         bufcount = 2;
2665                 else
2666                         bufcount = 1;
2667
2668                 rc = lustre_pack_reply(req, bufcount, size, NULL);
2669                 if (rc)
2670                         break;
2671
2672                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2673                 fail = OBD_FAIL_MDS_REINT_NET_REP;
2674                 break;
2675         }
2676
2677         case MDS_CLOSE:
2678                 DEBUG_REQ(D_INODE, req, "close");
2679                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2680                 rc = mds_close(req, MDS_REQ_REC_OFF);
2681                 break;
2682
2683         case MDS_DONE_WRITING:
2684                 DEBUG_REQ(D_INODE, req, "done_writing");
2685                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2686                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2687                 break;
2688
2689         case MDS_PIN:
2690                 DEBUG_REQ(D_INODE, req, "pin");
2691                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2692                 rc = mds_pin(req, MDS_REQ_REC_OFF);
2693                 break;
2694
2695         case MDS_SYNC:
2696                 DEBUG_REQ(D_INODE, req, "sync");
2697                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2698                 rc = mds_sync(req, MDS_REQ_REC_OFF);
2699                 break;
2700
2701         case OBD_PING:
2702                 DEBUG_REQ(D_INODE, req, "ping");
2703                 rc = target_handle_ping(req);
2704                 break;
2705
2706         case OBD_LOG_CANCEL:
2707                 CDEBUG(D_INODE, "log cancel\n");
2708                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2709                 rc = -ENOTSUPP; /* la la la */
2710                 break;
2711
2712         case LDLM_ENQUEUE:
2713                 DEBUG_REQ(D_INODE, req, "enqueue");
2714                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2715                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2716                                          ldlm_server_blocking_ast, NULL);
2717                 fail = OBD_FAIL_LDLM_REPLY;
2718                 break;
2719         case LDLM_CONVERT:
2720                 DEBUG_REQ(D_INODE, req, "convert");
2721                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2722                 rc = ldlm_handle_convert(req);
2723                 break;
2724         case LDLM_BL_CALLBACK:
2725         case LDLM_CP_CALLBACK:
2726                 DEBUG_REQ(D_INODE, req, "callback");
2727                 CERROR("callbacks should not happen on MDS\n");
2728                 LBUG();
2729                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2730                 break;
2731         case LLOG_ORIGIN_HANDLE_OPEN:
2732                 DEBUG_REQ(D_INODE, req, "llog_init");
2733                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2734                 rc = llog_origin_handle_open(req);
2735                 break;
2736         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2737                 DEBUG_REQ(D_INODE, req, "llog next block");
2738                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2739                 rc = llog_origin_handle_next_block(req);
2740                 break;
2741         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2742                 DEBUG_REQ(D_INODE, req, "llog prev block");
2743                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2744                 rc = llog_origin_handle_prev_block(req);
2745                 break;
2746         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2747                 DEBUG_REQ(D_INODE, req, "llog read header");
2748                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2749                 rc = llog_origin_handle_read_header(req);
2750                 break;
2751         case LLOG_ORIGIN_HANDLE_CLOSE:
2752                 DEBUG_REQ(D_INODE, req, "llog close");
2753                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2754                 rc = llog_origin_handle_close(req);
2755                 break;
2756         case OST_CREATE:
2757                 DEBUG_REQ(D_INODE, req, "ost_create");
2758                 rc = mdt_obj_create(req);
2759                 break;
2760         case OST_GET_INFO:
2761                 DEBUG_REQ(D_INODE, req, "get_info");
2762                 rc = mdt_get_info(req);
2763                 break;
2764         case OST_SET_INFO:
2765                 DEBUG_REQ(D_INODE, req, "set_info");
2766                 rc = mdt_set_info(req);
2767                 break;
2768         case OST_WRITE:
2769                 CDEBUG(D_INODE, "write\n");
2770                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2771                 rc = ost_brw_write(req, NULL);
2772                 LASSERT(current->journal_info == NULL);
2773                 /* mdt_brw sends its own replies */
2774                 RETURN(rc);
2775                 break;
2776         case LLOG_CATINFO:
2777                 DEBUG_REQ(D_INODE, req, "llog catinfo");
2778                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2779                 rc = llog_catinfo(req);
2780                 break;
2781         default:
2782                 req->rq_status = -ENOTSUPP;
2783                 rc = ptlrpc_error(req);
2784                 RETURN(rc);
2785         }
2786
2787         LASSERT(current->journal_info == NULL);
2788
2789         EXIT;
2790
2791         /* If we're DISCONNECTing, the mds_export_data is already freed */
2792         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2793                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2794                 struct obd_device *obd = list_entry(mds, struct obd_device,
2795                                                     u.mds);
2796                 req->rq_repmsg->last_xid =
2797                         le64_to_cpu(med->med_mcd->mcd_last_xid);
2798
2799                 if (!obd->obd_no_transno) {
2800                         req->rq_repmsg->last_committed =
2801                                 obd->obd_last_committed;
2802                 } else {
2803                         DEBUG_REQ(D_IOCTL, req,
2804                                   "not sending last_committed update");
2805                 }
2806                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2807                        ", xid "LPU64"\n",
2808                        mds->mds_last_transno, obd->obd_last_committed,
2809                        req->rq_xid);
2810         }
2811  out:
2812
2813
2814         target_send_reply(req, rc, fail);
2815         return 0;
2816 }
2817
2818 /* Update the server data on disk.  This stores the new mount_count and also the
2819  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
2820  * last_rcvd value may be less than that of the clients.  This will alert us
2821  * that we may need to do client recovery.
2822  *
2823  * Also assumes for mds_last_transno that we are not modifying it (no locking).
2824  */
2825 int mds_update_server_data(struct obd_device *obd, int force_sync)
2826 {
2827         struct mds_obd *mds = &obd->u.mds;
2828         struct mds_server_data *msd = mds->mds_server_data;
2829         struct file *filp = mds->mds_rcvd_filp;
2830         struct lvfs_run_ctxt saved;
2831         loff_t off = 0;
2832         int rc;
2833         ENTRY;
2834
2835         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2836         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2837
2838         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2839                mds->mds_mount_count, mds->mds_last_transno);
2840         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2841         if (rc)
2842                 CERROR("error writing MDS server data: rc = %d\n", rc);
2843         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2844
2845         RETURN(rc);
2846 }
2847
2848 /* saves last allocated fid counter to file. */
2849 int mds_update_last_fid(struct obd_device *obd, void *handle,
2850                         int force_sync)
2851 {
2852         struct mds_obd *mds = &obd->u.mds;
2853         struct file *filp = mds->mds_fid_filp;
2854         struct lvfs_run_ctxt saved;
2855         loff_t off = 0;
2856         __u64 last_fid;
2857         int rc = 0;
2858         ENTRY;
2859
2860         spin_lock(&mds->mds_last_fid_lock);
2861         last_fid = mds->mds_last_fid;
2862         spin_unlock(&mds->mds_last_fid_lock);
2863
2864         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2865                last_fid);
2866
2867         if (handle) {
2868                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
2869                                       handle, mds_commit_last_fid_cb,
2870                                       NULL);
2871         }
2872
2873         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2874         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
2875                                  &off, force_sync);
2876         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2877
2878         if (rc) {
2879                 CERROR("error writing MDS last_fid #"LPU64
2880                        ", err = %d\n", last_fid, rc);
2881                 RETURN(rc);
2882         }
2883
2884         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2885                "%llu: err = %d\n", last_fid, off, rc);
2886
2887         RETURN(rc);
2888 }
2889
2890 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2891 {
2892         struct mds_obd *mds = &obd->u.mds;
2893
2894         spin_lock(&mds->mds_last_fid_lock);
2895         if (fid > mds->mds_last_fid)
2896                 mds->mds_last_fid = fid;
2897         spin_unlock(&mds->mds_last_fid_lock);
2898 }
2899
2900 void mds_commit_last_transno_cb(struct obd_device *obd,
2901                                 __u64 transno, void *data,
2902                                 int error)
2903 {
2904         obd_transno_commit_cb(obd, transno, error);
2905 }
2906
2907 void mds_commit_last_fid_cb(struct obd_device *obd,
2908                             __u64 fid, void *data,
2909                             int error)
2910 {
2911         if (error) {
2912                 CERROR("%s: fid "LPD64" commit error: %d\n",
2913                        obd->obd_name, fid, error);
2914                 return;
2915         }
2916
2917         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2918                obd->obd_name, fid);
2919 }
2920
2921 __u64 mds_alloc_fid(struct obd_device *obd)
2922 {
2923         struct mds_obd *mds = &obd->u.mds;
2924         __u64 fid;
2925
2926         spin_lock(&mds->mds_last_fid_lock);
2927         fid = ++mds->mds_last_fid;
2928         spin_unlock(&mds->mds_last_fid_lock);
2929
2930         return fid;
2931 }
2932
2933 /*
2934  * allocates new lustre_id on passed @inode and saves it to inode EA.
2935  */
2936 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2937                         void *handle, struct lustre_id *id)
2938 {
2939         struct mds_obd *mds = &obd->u.mds;
2940         int alloc = 0, rc = 0;
2941         ENTRY;
2942
2943         LASSERT(obd != NULL);
2944         LASSERT(inode != NULL);
2945
2946         if (id == NULL) {
2947                 OBD_ALLOC(id, sizeof(*id));
2948                 if (id == NULL)
2949                         RETURN(-ENOMEM);
2950                 alloc = 1;
2951         }
2952
2953         id_group(id) = mds->mds_num;
2954         id_fid(id) = mds_alloc_fid(obd);
2955         id_ino(id) = inode->i_ino;
2956         id_gen(id) = inode->i_generation;
2957         id_type(id) = (S_IFMT & inode->i_mode);
2958
2959         rc = mds_update_inode_sid(obd, inode, handle, id);
2960         if (rc) {
2961                 CERROR("Can't update inode FID EA, "
2962                        "rc = %d\n", rc);
2963         }
2964
2965         if (alloc)
2966                 OBD_FREE(id, sizeof(*id));
2967         RETURN(rc);
2968 }
2969
2970 /*
2971  * reads inode self id from inode EA. Probably later this should be replaced by
2972  * caching inode self id to avoid raeding it every time it is needed.
2973  */
2974 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2975                        struct lustre_id *id)
2976 {
2977         int rc;
2978         ENTRY;
2979
2980         LASSERT(id != NULL);
2981         LASSERT(obd != NULL);
2982         LASSERT(inode != NULL);
2983
2984         rc = fsfilt_get_md(obd, inode, &id->li_fid,
2985                            sizeof(id->li_fid), EA_SID);
2986         if (rc < 0) {
2987                 CERROR("fsfilt_get_md() failed, "
2988                        "rc = %d\n", rc);
2989                 RETURN(rc);
2990         } else if (!rc) {
2991                 rc = -ENODATA;
2992                 RETURN(rc);
2993         } else {
2994                 rc = 0;
2995         }
2996
2997         RETURN(rc);
2998 }
2999
3000 /* updates inode self id in EA. */
3001 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3002                          void *handle, struct lustre_id *id)
3003 {
3004         int rc = 0;
3005         ENTRY;
3006
3007         LASSERT(id != NULL);
3008         LASSERT(obd != NULL);
3009         LASSERT(inode != NULL);
3010
3011         rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3012                            sizeof(id->li_fid), EA_SID);
3013         if (rc) {
3014                 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3015                 RETURN(rc);
3016         }
3017
3018         RETURN(rc);
3019 }
3020
3021 /*
3022  * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3023  * to master MDS by correct store cookie, needed to find inode on master MDS
3024  * quickly.
3025  */
3026 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3027                        struct lustre_id *id)
3028 {
3029         int rc;
3030         ENTRY;
3031
3032         LASSERT(id != NULL);
3033         LASSERT(obd != NULL);
3034         LASSERT(inode != NULL);
3035
3036         rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3037         if (rc < 0) {
3038                 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3039                 RETURN(rc);
3040         } else if (!rc) {
3041                 rc = -ENODATA;
3042                 RETURN(rc);
3043         } else {
3044                 rc = 0;
3045         }
3046
3047         RETURN(rc);
3048 }
3049
3050 /*
3051  * updates master inode id. Usualy this is done by CMOBD after an inode is
3052  * created and relationship between cache MDS and master one should be
3053  * established.
3054  */
3055 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3056                          void *handle, struct lustre_id *id)
3057 {
3058         int rc = 0;
3059         ENTRY;
3060
3061         LASSERT(id != NULL);
3062         LASSERT(obd != NULL);
3063         LASSERT(inode != NULL);
3064
3065         rc = fsfilt_set_md(obd, inode, handle, id,
3066                            sizeof(*id), EA_MID);
3067         if (rc) {
3068                 CERROR("fsfilt_set_md() failed, "
3069                        "rc = %d\n", rc);
3070                 RETURN(rc);
3071         }
3072
3073         RETURN(rc);
3074 }
3075
3076 /* mount the file system (secretly) */
3077 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3078 {
3079         struct lustre_cfg* lcfg = buf;
3080         struct mds_obd *mds = &obd->u.mds;
3081         struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3082         char *options = NULL;
3083         struct vfsmount *mnt;
3084         char ns_name[48];
3085         unsigned long page;
3086         int rc = 0;
3087         ENTRY;
3088
3089
3090         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
3091                 RETURN(rc = -EINVAL);
3092
3093         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
3094         if (IS_ERR(obd->obd_fsops))
3095                 RETURN(rc = PTR_ERR(obd->obd_fsops));
3096
3097         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3098
3099         page = __get_free_page(GFP_KERNEL);
3100         if (!page)
3101                 RETURN(-ENOMEM);
3102
3103         options = (char *)page;
3104         memset(options, 0, PAGE_SIZE);
3105
3106         /*
3107          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3108          * and the rest of options are passed by mount options. Probably this
3109          * should be moved to somewhere else like startup scripts or lconf. */
3110         sprintf(options, "iopen_nopriv");
3111         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
3112                 sprintf(options + strlen(options), ",%s",
3113                         lcfg->lcfg_inlbuf4);
3114
3115         /* we have to know mdsnum before touching underlying fs -bzzz */
3116         atomic_set(&mds->mds_open_count, 0);
3117         sema_init(&mds->mds_md_sem, 1);
3118         mds->mds_md_connected = 0;
3119         mds->mds_md_name = NULL;
3120
3121         if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
3122             strncmp(lcfg->lcfg_inlbuf5, "dumb", lcfg->lcfg_inllen5)) {
3123                 class_uuid_t uuid;
3124
3125                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3126                        obd->obd_name, lcfg->lcfg_inlbuf5);
3127
3128                 generate_random_uuid(uuid);
3129                 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3130
3131                 OBD_ALLOC(mds->mds_md_name, lcfg->lcfg_inllen5);
3132                 if (mds->mds_md_name == NULL)
3133                         RETURN(rc = -ENOMEM);
3134
3135                 memcpy(mds->mds_md_name, lcfg->lcfg_inlbuf5,
3136                        lcfg->lcfg_inllen5);
3137
3138                 rc = mds_md_connect(obd, mds->mds_md_name);
3139                 if (rc) {
3140                         OBD_FREE(mds->mds_md_name, lcfg->lcfg_inllen5);
3141                         GOTO(err_ops, rc);
3142                 }
3143         }
3144
3145         mds->mds_obd_type = MDS_MASTER_OBD;
3146
3147         if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
3148             strncmp(lcfg->lcfg_inlbuf6, "dumb", lcfg->lcfg_inllen6)) {
3149                 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
3150                         mds->mds_obd_type = MDS_MASTER_OBD;
3151                 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
3152                         mds->mds_obd_type = MDS_CACHE_OBD;
3153                 }
3154         }
3155
3156         rc = lvfs_mount_fs(lcfg->lcfg_inlbuf1, lcfg->lcfg_inlbuf2,
3157                            options, 0, &lvfs_ctxt);
3158
3159         free_page(page);
3160
3161         if (rc || !lvfs_ctxt) {
3162                 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3163                 GOTO(err_ops, rc);
3164         }
3165
3166         mnt = lvfs_ctxt->loc_mnt;
3167         mds->mds_lvfs_ctxt = lvfs_ctxt;
3168         ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3169
3170         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
3171
3172         sema_init(&mds->mds_epoch_sem, 1);
3173         atomic_set(&mds->mds_real_clients, 0);
3174         spin_lock_init(&mds->mds_transno_lock);
3175         spin_lock_init(&mds->mds_last_fid_lock);
3176         sema_init(&mds->mds_orphan_recovery_sem, 1);
3177         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3178
3179         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3180         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3181
3182         if (obd->obd_namespace == NULL) {
3183                 mds_cleanup(obd, 0);
3184                 GOTO(err_put, rc = -ENOMEM);
3185         }
3186         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3187
3188         rc = mds_fs_setup(obd, mnt);
3189         if (rc) {
3190                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3191                        obd->obd_name, rc);
3192                 GOTO(err_ns, rc);
3193         }
3194
3195         rc = llog_start_commit_thread();
3196         if (rc < 0)
3197                 GOTO(err_fs, rc);
3198
3199         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
3200             strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)) {
3201                 class_uuid_t uuid;
3202
3203                 generate_random_uuid(uuid);
3204                 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3205
3206                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
3207                 if (mds->mds_profile == NULL)
3208                         GOTO(err_fs, rc = -ENOMEM);
3209
3210                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
3211                        lcfg->lcfg_inllen3);
3212         }
3213
3214         /*
3215          * setup root dir and files ID dir if lmv already connected, or there is
3216          * not lmv at all.
3217          */
3218         if (mds->mds_md_exp || (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
3219                                 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)))
3220         {
3221                 rc = mds_fs_setup_rootid(obd);
3222                 if (rc)
3223                         GOTO(err_fs, rc);
3224
3225                 rc = mds_fs_setup_virtid(obd);
3226                 if (rc)
3227                         GOTO(err_fs, rc);
3228         }
3229
3230         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3231                            "mds_ldlm_client", &obd->obd_ldlm_client);
3232         obd->obd_replayable = 1;
3233
3234         rc = mds_postsetup(obd);
3235         if (rc)
3236                 GOTO(err_fs, rc);
3237
3238         RETURN(0);
3239
3240 err_fs:
3241         /* No extra cleanup needed for llog_init_commit_thread() */
3242         mds_fs_cleanup(obd, 0);
3243 err_ns:
3244         ldlm_namespace_free(obd->obd_namespace, 0);
3245         obd->obd_namespace = NULL;
3246 err_put:
3247         unlock_kernel();
3248         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3249         mds->mds_sb = 0;
3250         lock_kernel();
3251 err_ops:
3252         fsfilt_put_ops(obd->obd_fsops);
3253         return rc;
3254 }
3255
3256 static int mds_postsetup(struct obd_device *obd)
3257 {
3258         struct mds_obd *mds = &obd->u.mds;
3259         int rc = 0;
3260         ENTRY;
3261
3262         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3263                             obd, 0, NULL, &llog_lvfs_ops);
3264         if (rc)
3265                 RETURN(rc);
3266
3267         if (mds->mds_profile) {
3268                 struct llog_ctxt *lgctxt;
3269                 struct lvfs_run_ctxt saved;
3270                 struct lustre_profile *lprof;
3271                 struct config_llog_instance cfg;
3272
3273                 cfg.cfg_instance = NULL;
3274                 cfg.cfg_uuid = mds->mds_dt_uuid;
3275                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3276
3277                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3278                 if (!lgctxt) {
3279                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3280                         GOTO(err_llog, rc = -EINVAL);
3281                 }
3282
3283                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3284                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3285
3286                 if (rc)
3287                         GOTO(err_llog, rc);
3288
3289                 lprof = class_get_profile(mds->mds_profile);
3290                 if (lprof == NULL) {
3291                         CERROR("No profile found: %s\n", mds->mds_profile);
3292                         GOTO(err_cleanup, rc = -ENOENT);
3293                 }
3294                 rc = mds_dt_connect(obd, lprof->lp_lov);
3295                 if (rc)
3296                         GOTO(err_cleanup, rc);
3297
3298                 rc = mds_md_postsetup(obd);
3299                 if (rc)
3300                         GOTO(err_cleanup, rc);
3301         }
3302
3303         RETURN(rc);
3304 err_cleanup:
3305         mds_dt_clean(obd);
3306 err_llog:
3307         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3308                                           LLOG_CONFIG_ORIG_CTXT));
3309         return rc;
3310 }
3311
3312 int mds_postrecov_common(struct obd_device *obd)
3313 {
3314         struct mds_obd *mds = &obd->u.mds;
3315         struct llog_ctxt *ctxt;
3316         int rc, item = 0, valsize;
3317          __u32 group;
3318         ENTRY;
3319
3320         LASSERT(!obd->obd_recovering);
3321         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3322         LASSERT(ctxt != NULL);
3323
3324         /* set nextid first, so we are sure it happens */
3325         rc = mds_dt_set_nextid(obd);
3326         if (rc) {
3327                 CERROR("%s: mds_dt_set_nextid() failed\n", obd->obd_name);
3328                 GOTO(out, rc);
3329         }
3330
3331         /* clean PENDING dir */
3332         rc = mds_cleanup_orphans(obd);
3333         if (rc < 0)
3334                 GOTO(out, rc);
3335         item = rc;
3336
3337         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3338         valsize = sizeof(group);
3339         rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"), "mds_conn",
3340                           valsize, &group);
3341         if (rc)
3342                 GOTO(out, rc);
3343
3344         rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3345                           NULL, NULL, NULL);
3346         if (rc) {
3347                 CERROR("%s: failed at llog_origin_connect: %d\n",
3348                        obd->obd_name, rc);
3349                 GOTO(out, rc);
3350         }
3351
3352         /* remove the orphaned precreated objects */
3353         rc = mds_dt_clearorphans(mds, NULL /* all OSTs */);
3354         if (rc)
3355                 GOTO(err_llog, rc);
3356
3357 out:
3358         RETURN(rc < 0 ? rc : item);
3359
3360 err_llog:
3361         /* cleanup all llogging subsystems */
3362         rc = obd_llog_finish(obd, &obd->obd_llogs,
3363                              mds->mds_dt_desc.ld_tgt_count);
3364         if (rc)
3365                 CERROR("%s: failed to cleanup llogging subsystems\n",
3366                         obd->obd_name);
3367         goto out;
3368 }
3369
3370 int mds_postrecov(struct obd_device *obd)
3371 {
3372         int rc;
3373         ENTRY;
3374         rc = mds_postrecov_common(obd);
3375         if (rc == 0)
3376                 rc = mds_md_reconnect(obd);
3377         RETURN(rc);
3378 }
3379
3380 int mds_dt_clean(struct obd_device *obd)
3381 {
3382         struct mds_obd *mds = &obd->u.mds;
3383         ENTRY;
3384
3385         if (mds->mds_profile) {
3386                 char * cln_prof;
3387                 struct llog_ctxt *llctx;
3388                 struct lvfs_run_ctxt saved;
3389                 struct config_llog_instance cfg;
3390                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3391
3392                 OBD_ALLOC(cln_prof, len);
3393                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3394
3395                 cfg.cfg_instance = NULL;
3396                 cfg.cfg_uuid = mds->mds_dt_uuid;
3397
3398                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3399                 llctx = llog_get_context(&obd->obd_llogs,
3400                                          LLOG_CONFIG_ORIG_CTXT);
3401                 class_config_process_llog(llctx, cln_prof, &cfg);
3402                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3403
3404                 OBD_FREE(cln_prof, len);
3405                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3406                 mds->mds_profile = NULL;
3407         }
3408         RETURN(0);
3409 }
3410
3411 int mds_md_clean(struct obd_device *obd)
3412 {
3413         struct mds_obd *mds = &obd->u.mds;
3414         ENTRY;
3415
3416         if (mds->mds_md_name) {
3417                 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3418                 mds->mds_md_name = NULL;
3419         }
3420         RETURN(0);
3421 }
3422
3423 static int mds_precleanup(struct obd_device *obd, int flags)
3424 {
3425         int rc = 0;
3426         ENTRY;
3427
3428         mds_md_clean(obd);
3429         mds_dt_disconnect(obd, flags);
3430         mds_dt_clean(obd);
3431         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3432         RETURN(rc);
3433 }
3434
3435 extern void lgss_svc_cache_purge_all(void);
3436 static int mds_cleanup(struct obd_device *obd, int flags)
3437 {
3438         struct mds_obd *mds = &obd->u.mds;
3439         ENTRY;
3440
3441         if (mds->mds_sb == NULL)
3442                 RETURN(0);
3443
3444         mds_update_server_data(obd, 1);
3445         mds_update_last_fid(obd, NULL, 1);
3446
3447         if (mds->mds_dt_objids != NULL) {
3448                 int size = mds->mds_dt_desc.ld_tgt_count *
3449                         sizeof(obd_id);
3450                 OBD_FREE(mds->mds_dt_objids, size);
3451         }
3452         mds_fs_cleanup(obd, flags);
3453
3454         unlock_kernel();
3455
3456         /* 2 seems normal on mds, (may_umount() also expects 2
3457           fwiw), but we only see 1 at this point in obdfilter. */
3458         lvfs_umount_fs(mds->mds_lvfs_ctxt);
3459
3460         mds->mds_sb = 0;
3461
3462         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3463
3464         spin_lock_bh(&obd->obd_processing_task_lock);
3465         if (obd->obd_recovering) {
3466                 target_cancel_recovery_timer(obd);
3467                 obd->obd_recovering = 0;
3468         }
3469         spin_unlock_bh(&obd->obd_processing_task_lock);
3470
3471         lock_kernel();
3472         fsfilt_put_ops(obd->obd_fsops);
3473
3474 #ifdef ENABLE_GSS
3475         /* XXX */
3476         lgss_svc_cache_purge_all();
3477 #endif
3478         RETURN(0);
3479 }
3480
3481 static int set_security(const char *value, char **sec)
3482 {
3483         int rc = 0;
3484
3485         if (!strcmp(value, "null"))
3486                 *sec = "null";
3487         else if (!strcmp(value, "krb5i"))
3488                 *sec = "krb5i";
3489         else if (!strcmp(value, "krb5p"))
3490                 *sec = "krb5p";
3491         else {
3492                 CERROR("Unrecognized value, force use NULL\n");
3493                 rc = -EINVAL;
3494         }
3495
3496         return rc;
3497 }
3498
3499 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3500 {
3501         struct lustre_cfg *lcfg = buf;
3502         struct mds_obd *mds = &obd->u.mds;
3503         int rc = 0;
3504         ENTRY;
3505
3506         switch(lcfg->lcfg_command) {
3507         case LCFG_SET_SECURITY: {
3508                 if (!lcfg->lcfg_inllen1 || !lcfg->lcfg_inllen2)
3509                         GOTO(out, rc = -EINVAL);
3510
3511                 if (!strcmp(lcfg->lcfg_inlbuf1, "mds_mds_sec"))
3512                         rc = set_security(lcfg->lcfg_inlbuf2,
3513                                           &mds->mds_mds_sec);
3514                 else if (!strcmp(lcfg->lcfg_inlbuf1, "mds_ost_sec"))
3515                         rc = set_security(lcfg->lcfg_inlbuf2,
3516                                           &mds->mds_ost_sec);
3517                 else {
3518                         CERROR("Unrecognized key\n");
3519                         rc = -EINVAL;
3520                 }
3521                 break;
3522         }
3523         default: {
3524                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3525                 GOTO(out, rc = -EINVAL);
3526
3527         }
3528         }
3529 out:
3530         RETURN(rc);
3531 }
3532
3533 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3534                                         int offset,
3535                                         struct ldlm_lock *new_lock,
3536                                         struct ldlm_lock **old_lock,
3537                                         struct lustre_handle *lockh)
3538 {
3539         struct obd_export *exp = req->rq_export;
3540         struct obd_device *obd = exp->exp_obd;
3541         struct ldlm_request *dlmreq =
3542                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3543         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3544         struct list_head *iter;
3545
3546         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3547                 return;
3548
3549         l_lock(&obd->obd_namespace->ns_lock);
3550         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3551                 struct ldlm_lock *lock;
3552                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3553                 if (lock == new_lock)
3554                         continue;
3555                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3556                         lockh->cookie = lock->l_handle.h_cookie;
3557                         LDLM_DEBUG(lock, "restoring lock cookie");
3558                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3559                                   lockh->cookie);
3560                         if (old_lock)
3561                                 *old_lock = LDLM_LOCK_GET(lock);
3562                         l_unlock(&obd->obd_namespace->ns_lock);
3563                         return;
3564                 }
3565         }
3566         l_unlock(&obd->obd_namespace->ns_lock);
3567
3568         /* If the xid matches, then we know this is a resent request,
3569          * and allow it. (It's probably an OPEN, for which we don't
3570          * send a lock */
3571         if (req->rq_xid ==
3572             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3573                 return;
3574
3575         if (req->rq_xid ==
3576             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3577                 return;
3578
3579         /* This remote handle isn't enqueued, so we never received or
3580          * processed this request.  Clear MSG_RESENT, because it can
3581          * be handled like any normal request now. */
3582
3583         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3584
3585         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3586                   remote_hdl.cookie);
3587 }
3588
3589 int intent_disposition(struct ldlm_reply *rep, int flag)
3590 {
3591         if (!rep)
3592                 return 0;
3593         return (rep->lock_policy_res1 & flag);
3594 }
3595
3596 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3597 {
3598         if (!rep)
3599                 return;
3600         rep->lock_policy_res1 |= flag;
3601 }
3602
3603 static int mds_intent_policy(struct ldlm_namespace *ns,
3604                              struct ldlm_lock **lockp, void *req_cookie,
3605                              ldlm_mode_t mode, int flags, void *data)
3606 {
3607         struct ptlrpc_request *req = req_cookie;
3608         struct ldlm_lock *lock = *lockp;
3609         struct ldlm_intent *it;
3610         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3611         struct ldlm_reply *rep;
3612         struct lustre_handle lockh[2] = {{0}, {0}};
3613         struct ldlm_lock *new_lock = NULL;
3614         int getattr_part = MDS_INODELOCK_UPDATE;
3615         int rc, reply_buffers;
3616         int repsize[5] = {sizeof(struct ldlm_reply),
3617                           sizeof(struct mds_body),
3618                           mds->mds_max_mdsize};
3619
3620         int offset = MDS_REQ_INTENT_REC_OFF;
3621         ENTRY;
3622
3623         LASSERT(req != NULL);
3624         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3625
3626         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3627                 /* No intent was provided */
3628                 int size = sizeof(struct ldlm_reply);
3629                 rc = lustre_pack_reply(req, 1, &size, NULL);
3630                 LASSERT(rc == 0);
3631                 RETURN(0);
3632         }
3633
3634         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3635                                 lustre_swab_ldlm_intent);
3636         if (it == NULL) {
3637                 CERROR("Intent missing\n");
3638                 RETURN(req->rq_status = -EFAULT);
3639         }
3640
3641         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3642
3643         reply_buffers = 3;
3644         if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3645                 reply_buffers = 5;
3646                 repsize[3] = 4;
3647                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
3648         }
3649
3650         rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
3651         if (rc)
3652                 RETURN(req->rq_status = rc);
3653
3654         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3655         LASSERT(rep != NULL);
3656
3657         intent_set_disposition(rep, DISP_IT_EXECD);
3658
3659         /* execute policy */
3660         switch ((long)it->opc) {
3661         case IT_OPEN:
3662         case IT_CREAT|IT_OPEN:
3663                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3664                                             lock, NULL, lockh);
3665                 /* XXX swab here to assert that an mds_open reint
3666                  * packet is following */
3667                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3668                                             lock, NULL, lockh);
3669                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3670
3671                 if (rep->lock_policy_res2) {
3672                         /*
3673                          * mds_open() returns ENOLCK where it should return
3674                          * zero, but it has no lock to return.
3675                          */
3676                         if (rep->lock_policy_res2 == ENOLCK)
3677                                 rep->lock_policy_res2 = 0;
3678
3679                         RETURN(ELDLM_LOCK_ABORTED);
3680                 }
3681
3682                 /*
3683                  * IT_OPEN may return lock on cross-node dentry that we want to
3684                  * hold during attr retrival -bzzz
3685                  */
3686                 if (lockh[0].cookie == 0)
3687                         RETURN(ELDLM_LOCK_ABORTED);
3688
3689                 break;
3690         case IT_LOOKUP:
3691                 getattr_part = MDS_INODELOCK_LOOKUP;
3692         case IT_CHDIR:
3693         case IT_GETATTR:
3694                 getattr_part |= MDS_INODELOCK_LOOKUP;
3695         case IT_READDIR:
3696                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3697                                             lock, &new_lock, lockh);
3698                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3699                                                          getattr_part);
3700                 /* FIXME: LDLM can set req->rq_status. MDS sets
3701                    policy_res{1,2} with disposition and status.
3702                    - replay: returns 0 & req->status is old status
3703                    - otherwise: returns req->status */
3704                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3705                         rep->lock_policy_res2 = 0;
3706                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3707                     rep->lock_policy_res2)
3708                         RETURN(ELDLM_LOCK_ABORTED);
3709                 if (req->rq_status != 0) {
3710                         LBUG();
3711                         rep->lock_policy_res2 = req->rq_status;
3712                         RETURN(ELDLM_LOCK_ABORTED);
3713                 }
3714                 break;
3715         case IT_UNLINK:
3716                 rc = mds_lock_and_check_slave(offset, req, lockh);
3717                 if ((rep->lock_policy_res2 = rc)) {
3718                         if (rc == ENOLCK)
3719                                 rep->lock_policy_res2 = 0;
3720                         RETURN(ELDLM_LOCK_ABORTED);
3721                 }
3722                 break;
3723         default:
3724                 CERROR("Unhandled intent "LPD64"\n", it->opc);
3725                 LBUG();
3726         }
3727
3728         /* By this point, whatever function we called above must have either
3729          * filled in 'lockh', been an intent replay, or returned an error.  We
3730          * want to allow replayed RPCs to not get a lock, since we would just
3731          * drop it below anyways because lock replay is done separately by the
3732          * client afterwards.  For regular RPCs we want to give the new lock to
3733          * the client instead of whatever lock it was about to get. */
3734         if (new_lock == NULL)
3735                 new_lock = ldlm_handle2lock(&lockh[0]);
3736         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3737                 RETURN(0);
3738
3739         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3740                  it->opc, lockh[0].cookie);
3741
3742         /* If we've already given this lock to a client once, then we should
3743          * have no readers or writers.  Otherwise, we should have one reader
3744          * _or_ writer ref (which will be zeroed below) before returning the
3745          * lock to a client. */
3746         if (new_lock->l_export == req->rq_export) {
3747                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3748         } else {
3749                 LASSERT(new_lock->l_export == NULL);
3750                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3751         }
3752
3753         *lockp = new_lock;
3754
3755         if (new_lock->l_export == req->rq_export) {
3756                 /* Already gave this to the client, which means that we
3757                  * reconstructed a reply. */
3758                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3759                         MSG_RESENT);
3760                 RETURN(ELDLM_LOCK_REPLACED);
3761         }
3762
3763         /* Fixup the lock to be given to the client */
3764         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3765         new_lock->l_readers = 0;
3766         new_lock->l_writers = 0;
3767
3768         new_lock->l_export = class_export_get(req->rq_export);
3769         list_add(&new_lock->l_export_chain,
3770                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
3771
3772         new_lock->l_blocking_ast = lock->l_blocking_ast;
3773         new_lock->l_completion_ast = lock->l_completion_ast;
3774
3775         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3776                sizeof(lock->l_remote_handle));
3777
3778         new_lock->l_flags &= ~LDLM_FL_LOCAL;
3779
3780         LDLM_LOCK_PUT(new_lock);
3781         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3782
3783         RETURN(ELDLM_LOCK_REPLACED);
3784 }
3785
3786 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3787 {
3788         struct lprocfs_static_vars lvars;
3789         int rc = 0;
3790
3791         lprocfs_init_multi_vars(0, &lvars);
3792
3793         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3794         if (rc)
3795                 return rc;
3796
3797         return lprocfs_alloc_md_stats(dev, 0);
3798 }
3799
3800 int mds_detach(struct obd_device *dev)
3801 {
3802         lprocfs_free_md_stats(dev);
3803         return lprocfs_obd_detach(dev);
3804 }
3805
3806 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3807 {
3808         struct lprocfs_static_vars lvars;
3809
3810         lprocfs_init_multi_vars(1, &lvars);
3811         return lprocfs_obd_attach(dev, lvars.obd_vars);
3812 }
3813
3814 int mdt_detach(struct obd_device *dev)
3815 {
3816         return lprocfs_obd_detach(dev);
3817 }
3818
3819 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3820 {
3821         struct mds_obd *mds = &obd->u.mds;
3822         int rc = 0;
3823         ENTRY;
3824
3825         mds->mds_service =
3826                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3827                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3828                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3829                                 mds_handle, "mds", obd->obd_proc_entry);
3830
3831         if (!mds->mds_service) {
3832                 CERROR("failed to start service\n");
3833                 RETURN(-ENOMEM);
3834         }
3835
3836         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3837                                     "ll_mdt");
3838         if (rc)
3839                 GOTO(err_thread, rc);
3840
3841         mds->mds_setattr_service =
3842                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3843                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3844                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3845                                 mds_handle, "mds_setattr",
3846                                 obd->obd_proc_entry);
3847         if (!mds->mds_setattr_service) {
3848                 CERROR("failed to start getattr service\n");
3849                 GOTO(err_thread, rc = -ENOMEM);
3850         }
3851
3852         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3853                                     MDT_NUM_THREADS, "ll_mdt_attr");
3854         if (rc)
3855                 GOTO(err_thread2, rc);
3856
3857         mds->mds_readpage_service =
3858                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3859                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3860                                 MDS_SERVICE_WATCHDOG_TIMEOUT,
3861                                 mds_handle, "mds_readpage",
3862                                 obd->obd_proc_entry);
3863         if (!mds->mds_readpage_service) {
3864                 CERROR("failed to start readpage service\n");
3865                 GOTO(err_thread2, rc = -ENOMEM);
3866         }
3867
3868         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3869                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
3870
3871         if (rc)
3872                 GOTO(err_thread3, rc);
3873
3874         RETURN(0);
3875
3876 err_thread3:
3877         ptlrpc_unregister_service(mds->mds_readpage_service);
3878 err_thread2:
3879         ptlrpc_unregister_service(mds->mds_setattr_service);
3880 err_thread:
3881         ptlrpc_unregister_service(mds->mds_service);
3882         return rc;
3883 }
3884
3885 static int mdt_cleanup(struct obd_device *obd, int flags)
3886 {
3887         struct mds_obd *mds = &obd->u.mds;
3888         ENTRY;
3889
3890         ptlrpc_stop_all_threads(mds->mds_readpage_service);
3891         ptlrpc_unregister_service(mds->mds_readpage_service);
3892
3893         ptlrpc_stop_all_threads(mds->mds_setattr_service);
3894         ptlrpc_unregister_service(mds->mds_setattr_service);
3895
3896         ptlrpc_stop_all_threads(mds->mds_service);
3897         ptlrpc_unregister_service(mds->mds_service);
3898
3899         RETURN(0);
3900 }
3901
3902 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3903                                          __u64 gr, void *data)
3904 {
3905         struct lustre_id id;
3906         struct obd_device *obd = data;
3907
3908         id_ino(&id) = ino;
3909         id_gen(&id) = gen;
3910         return mds_id2dentry(obd, &id, NULL);
3911 }
3912
3913 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3914                         void *key, __u32 *valsize, void *val)
3915 {
3916         struct obd_device *obd;
3917         struct mds_obd *mds;
3918         ENTRY;
3919
3920         obd = class_exp2obd(exp);
3921         mds = &obd->u.mds;
3922
3923         if (obd == NULL) {
3924                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3925                        exp->exp_handle.h_cookie);
3926                 RETURN(-EINVAL);
3927         }
3928
3929         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3930                 /* get log_context handle. */
3931                 unsigned long *llh_handle = val;
3932                 *valsize = sizeof(unsigned long);
3933                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3934                 RETURN(0);
3935         }
3936         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3937                 /* get log_context handle. */
3938                 unsigned long *sb = val;
3939                 *valsize = sizeof(unsigned long);
3940                 *sb = (unsigned long)obd->u.mds.mds_sb;
3941                 RETURN(0);
3942         }
3943
3944         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3945                 __u32 *mdsize = val;
3946                 *valsize = sizeof(*mdsize);
3947                 *mdsize = mds->mds_max_mdsize;
3948                 RETURN(0);
3949         }
3950
3951         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3952                 __u32 *mdsnum = val;
3953                 *valsize = sizeof(*mdsnum);
3954                 *mdsnum = mds->mds_num;
3955                 RETURN(0);
3956         }
3957
3958         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3959                 struct lustre_id *rootid = val;
3960                 *valsize = sizeof(struct lustre_id);
3961                 *rootid = mds->mds_rootid;
3962                 RETURN(0);
3963         }
3964
3965         CDEBUG(D_IOCTL, "invalid key\n");
3966         RETURN(-EINVAL);
3967
3968 }
3969 struct lvfs_callback_ops mds_lvfs_ops = {
3970         l_id2dentry:     mds_lvfs_id2dentry,
3971 };
3972
3973 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3974                 int objcount, struct obd_ioobj *obj,
3975                 int niocount, struct niobuf_remote *nb,
3976                 struct niobuf_local *res,
3977                 struct obd_trans_info *oti);
3978
3979 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
3980                  int objcount, struct obd_ioobj *obj, int niocount,
3981                  struct niobuf_local *res, struct obd_trans_info *oti,
3982                  int rc);
3983
3984 /* use obd ops to offer management infrastructure */
3985 static struct obd_ops mds_obd_ops = {
3986         .o_owner           = THIS_MODULE,
3987         .o_attach          = mds_attach,
3988         .o_detach          = mds_detach,
3989         .o_connect         = mds_connect,
3990         .o_connect_post    = mds_connect_post,
3991         .o_init_export     = mds_init_export,
3992         .o_destroy_export  = mds_destroy_export,
3993         .o_disconnect      = mds_disconnect,
3994         .o_setup           = mds_setup,
3995         .o_precleanup      = mds_precleanup,
3996         .o_cleanup         = mds_cleanup,
3997         .o_process_config  = mds_process_config,
3998         .o_postrecov       = mds_postrecov,
3999         .o_statfs          = mds_obd_statfs,
4000         .o_iocontrol       = mds_iocontrol,
4001         .o_create          = mds_obd_create,
4002         .o_destroy         = mds_obd_destroy,
4003         .o_llog_init       = mds_llog_init,
4004         .o_llog_finish     = mds_llog_finish,
4005         .o_notify          = mds_notify,
4006         .o_get_info        = mds_get_info,
4007         .o_set_info        = mds_set_info,
4008         .o_preprw          = mds_preprw,
4009         .o_commitrw        = mds_commitrw,
4010 };
4011
4012 static struct obd_ops mdt_obd_ops = {
4013         .o_owner           = THIS_MODULE,
4014         .o_attach          = mdt_attach,
4015         .o_detach          = mdt_detach,
4016         .o_setup           = mdt_setup,
4017         .o_cleanup         = mdt_cleanup,
4018 };
4019
4020 static int __init mds_init(void)
4021 {
4022         struct lprocfs_static_vars lvars;
4023
4024         mds_init_lsd_cache();
4025
4026         lprocfs_init_multi_vars(0, &lvars);
4027         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4028                             LUSTRE_MDS_NAME);
4029         lprocfs_init_multi_vars(1, &lvars);
4030         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4031                             LUSTRE_MDT_NAME);
4032
4033         return 0;
4034 }
4035
4036 static void /*__exit*/ mds_exit(void)
4037 {
4038         mds_cleanup_lsd_cache();
4039
4040         class_unregister_type(LUSTRE_MDS_NAME);
4041         class_unregister_type(LUSTRE_MDT_NAME);
4042 }
4043
4044 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4045 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4046 MODULE_LICENSE("GPL");
4047
4048 module_init(mds_init);
4049 module_exit(mds_exit);