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, 2002 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  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #define EXPORT_SYMTAB
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_dlm.h>
  34 #include <linux/init.h>
  35 #include <linux/obd_class.h>
  36
  37 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
  38                            uuid_t *uuidarray);
  39 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
  40 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  41                                 struct ptlrpc_request *req);
  42 static int mds_cleanup(struct obd_device * obddev);
  43
  44 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  45 {
  46         return &req->rq_export->exp_obd->u.mds;
  47 }
  48
  49 static int mds_bulk_timeout(void *data)
  50 {
  51         struct ptlrpc_bulk_desc *desc = data;
  52
  53         ENTRY;
  54         CERROR("(not yet) starting recovery of client %p\n", desc->b_client);
  55         RETURN(1);
  56 }
  57
  58 /* Assumes caller has already pushed into the kernel filesystem context */
  59 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  60                         __u64 offset)
  61 {
  62         int rc = 0;
  63         struct mds_obd *mds = mds_req2mds(req);
  64         struct ptlrpc_bulk_desc *desc;
  65         struct ptlrpc_bulk_page *bulk;
  66         struct l_wait_info lwi;
  67         char *buf;
  68         ENTRY;
  69
  70         desc = ptlrpc_prep_bulk(req->rq_connection);
  71         if (desc == NULL)
  72                 GOTO(out, rc = -ENOMEM);
  73
  74         bulk = ptlrpc_prep_bulk_page(desc);
  75         if (bulk == NULL)
  76                 GOTO(cleanup_bulk, rc = -ENOMEM);
  77
  78         OBD_ALLOC(buf, PAGE_SIZE);
  79         if (buf == NULL)
  80                 GOTO(cleanup_bulk, rc = -ENOMEM);
  81
  82         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  83
  84         if (rc != PAGE_SIZE)
  85                 GOTO(cleanup_buf, rc = -EIO);
  86
  87         bulk->b_xid = req->rq_xid;
  88         bulk->b_buf = buf;
  89         bulk->b_buflen = PAGE_SIZE;
  90         desc->b_portal = MDS_BULK_PORTAL;
  91
  92         rc = ptlrpc_send_bulk(desc);
  93         if (rc)
  94                 GOTO(cleanup_buf, rc);
  95
  96         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  97                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  98                        OBD_FAIL_MDS_SENDPAGE, rc);
  99                 ptlrpc_abort_bulk(desc);
 100                 GOTO(cleanup_buf, rc);
 101         }
 102
 103         lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
 104         rc = l_wait_event(desc->b_waitq, desc->b_flags & PTL_BULK_FL_SENT, &lwi);
 105         if (rc) {
 106                 if (rc != -ETIMEDOUT)
 107                         LBUG();
 108                 GOTO(cleanup_buf, rc);
 109         }
 110
 111         EXIT;
 112  cleanup_buf:
 113         OBD_FREE(buf, PAGE_SIZE);
 114  cleanup_bulk:
 115         ptlrpc_free_bulk(desc);
 116  out:
 117         return rc;
 118 }
 119
 120 /*
 121  * Look up a named entry in a directory, and get an LDLM lock on it.
 122  * 'dir' is a inode for which an LDLM lock has already been taken.
 123  *
 124  * If we do not need an exclusive or write lock on this entry (e.g.
 125  * a read lock for attribute lookup only) then we do not hold the
 126  * directory on return.  It is up to the caller to know what type
 127  * of lock it is getting, and clean up appropriately.
 128  */
 129 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 130                                       struct dentry *dir, struct vfsmount **mnt,
 131                                       char *name, int namelen, int lock_mode,
 132                                       struct lustre_handle *lockh,
 133                                       int dir_lock_mode)
 134 {
 135         struct dentry *dchild;
 136         int flags, rc;
 137         __u64 res_id[3] = {0};
 138         ENTRY;
 139
 140         down(&dir->d_inode->i_sem);
 141         dchild = lookup_one_len(name, dir, namelen);
 142         if (IS_ERR(dchild)) {
 143                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 144                 up(&dir->d_inode->i_sem);
 145                 LBUG();
 146                 RETURN(dchild);
 147         }
 148         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 149                 up(&dir->d_inode->i_sem);
 150                 ldlm_lock_decref(lockh, dir_lock_mode);
 151         }
 152
 153         if (lock_mode == 0 || !dchild->d_inode)
 154                 RETURN(dchild);
 155
 156         res_id[0] = dchild->d_inode->i_ino;
 157         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 158                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 159                                    &flags, ldlm_completion_ast,
 160                                    mds_blocking_ast, NULL, 0, lockh);
 161         if (rc != ELDLM_OK) {
 162                 l_dput(dchild);
 163                 up(&dir->d_inode->i_sem);
 164                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 165         }
 166
 167         RETURN(dchild);
 168 }
 169
 170 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 171                                      struct vfsmount **mnt, int lock_mode,
 172                                      struct lustre_handle *lockh)
 173 {
 174         struct mds_obd *mds = &obd->u.mds;
 175         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 176         int flags, rc;
 177         __u64 res_id[3] = {0};
 178         ENTRY;
 179
 180         if (IS_ERR(de))
 181                 RETURN(de);
 182
 183         res_id[0] = de->d_inode->i_ino;
 184         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 185                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 186                                    &flags, ldlm_completion_ast,
 187                                    mds_blocking_ast, NULL, 0, lockh);
 188         if (rc != ELDLM_OK) {
 189                 l_dput(de);
 190                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 191         }
 192
 193         RETURN(retval);
 194 }
 195
 196 /* Look up an entry by inode number. */
 197 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 198                               struct vfsmount **mnt)
 199 {
 200         /* stolen from NFS */
 201         struct super_block *sb = mds->mds_sb;
 202         unsigned long ino = fid->id;
 203         __u32 generation = fid->generation;
 204         struct inode *inode;
 205         struct list_head *lp;
 206         struct dentry *result;
 207
 208         if (ino == 0)
 209                 RETURN(ERR_PTR(-ESTALE));
 210
 211         inode = iget(sb, ino);
 212         if (inode == NULL)
 213                 RETURN(ERR_PTR(-ENOMEM));
 214
 215         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 216
 217         if (is_bad_inode(inode) ||
 218             (generation && inode->i_generation != generation)) {
 219                 /* we didn't find the right inode.. */
 220                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 221                        inode->i_ino, inode->i_nlink,
 222                        atomic_read(&inode->i_count), inode->i_generation,
 223                        generation);
 224                 LBUG();
 225                 iput(inode);
 226                 RETURN(ERR_PTR(-ESTALE));
 227         }
 228
 229         /* now to find a dentry. If possible, get a well-connected one */
 230         if (mnt)
 231                 *mnt = mds->mds_vfsmnt;
 232         spin_lock(&dcache_lock);
 233         list_for_each(lp, &inode->i_dentry) {
 234                 result = list_entry(lp, struct dentry, d_alias);
 235                 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 236                         dget_locked(result);
 237                         result->d_vfs_flags |= DCACHE_REFERENCED;
 238                         spin_unlock(&dcache_lock);
 239                         iput(inode);
 240                         if (mnt)
 241                                 mntget(*mnt);
 242                         return result;
 243                 }
 244         }
 245         spin_unlock(&dcache_lock);
 246         result = d_alloc_root(inode);
 247         if (result == NULL) {
 248                 iput(inode);
 249                 return ERR_PTR(-ENOMEM);
 250         }
 251         if (mnt)
 252                 mntget(*mnt);
 253         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 254         return result;
 255 }
 256
 257 /* Establish a connection to the MDS.
 258  *
 259  * This will set up an export structure for the client to hold state data
 260  * about that client, like open files, the last operation number it did
 261  * on the server, etc.
 262  */
 263 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 264                        char *cluuid)
 265 {
 266         struct obd_export *exp;
 267         struct mds_client_data *mcd;
 268         struct list_head *p;
 269         int rc;
 270         ENTRY;
 271
 272         if (!conn || !obd || !cluuid)
 273                 RETURN(-EINVAL);
 274
 275         MOD_INC_USE_COUNT;
 276
 277         spin_lock(&obd->obd_dev_lock);
 278         list_for_each(p, &obd->obd_exports) {
 279                 exp = list_entry(p, struct obd_export, exp_chain);
 280                 mcd = exp->exp_mds_data.med_mcd;
 281                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 282                         LASSERT(exp->exp_obd == obd);
 283
 284                         exp->exp_rconnh.addr = conn->addr;
 285                         exp->exp_rconnh.cookie = conn->cookie;
 286                         conn->addr = (__u64) (unsigned long)exp;
 287                         conn->cookie = exp->exp_cookie;
 288                         spin_unlock(&obd->obd_dev_lock);
 289                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 290                                cluuid, exp);
 291                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 292                                (long long)conn->addr, (long long)conn->cookie);
 293                         RETURN(0);
 294                 }
 295         }
 296         spin_unlock(&obd->obd_dev_lock);
 297         /* XXX There is a small race between checking the list and adding a
 298          * new connection for the same UUID, but the real threat (list
 299          * corruption when multiple different clients connect) is solved.
 300          */
 301         rc = class_connect(conn, obd, cluuid);
 302         if (rc)
 303                 GOTO(out_dec, rc);
 304         exp = class_conn2export(conn);
 305         LASSERT(exp);
 306
 307         OBD_ALLOC(mcd, sizeof(*mcd));
 308         if (!mcd) {
 309                 CERROR("mds: out of memory for client data\n");
 310                 GOTO(out_export, rc = -ENOMEM);
 311         }
 312         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 313         exp->exp_mds_data.med_mcd = mcd;
 314         rc = mds_client_add(&exp->exp_mds_data, -1);
 315         if (rc)
 316                 GOTO(out_mdc, rc);
 317
 318         RETURN(0);
 319
 320 out_mdc:
 321         OBD_FREE(mcd, sizeof(*mcd));
 322 out_export:
 323         class_disconnect(conn);
 324 out_dec:
 325         MOD_DEC_USE_COUNT;
 326
 327         return rc;
 328 }
 329
 330 static int mds_disconnect(struct lustre_handle *conn)
 331 {
 332         int rc;
 333
 334         rc = class_disconnect(conn);
 335         if (!rc)
 336                 MOD_DEC_USE_COUNT;
 337
 338         return rc;
 339 }
 340
 341 static int mds_getstatus(struct ptlrpc_request *req)
 342 {
 343         struct mds_obd *mds = mds_req2mds(req);
 344         struct mds_body *body;
 345         int rc, size = sizeof(*body);
 346         ENTRY;
 347
 348         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 349         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 350                 CERROR("mds: out of memory for message: size=%d\n", size);
 351                 req->rq_status = -ENOMEM;
 352                 RETURN(0);
 353         }
 354
 355         /* Anything we need to do here with the client's trans no or so? */
 356         body = lustre_msg_buf(req->rq_repmsg, 0);
 357         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 358
 359         /* the last_committed and last_xid fields are filled in for all
 360          * replies already - no need to do so here also.
 361          */
 362         RETURN(0);
 363 }
 364
 365 static int mds_getlovinfo(struct ptlrpc_request *req)
 366 {
 367         struct mds_obd *mds = mds_req2mds(req);
 368         struct mds_status_req *streq;
 369         struct lov_desc *desc;
 370         int tgt_count;
 371         int rc, size[2] = {sizeof(*desc)};
 372         ENTRY;
 373
 374         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 375         streq->flags = NTOH__u32(streq->flags);
 376         streq->repbuf = NTOH__u32(streq->repbuf);
 377         size[1] = streq->repbuf;
 378
 379         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 380         if (rc) {
 381                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 382                 req->rq_status = -ENOMEM;
 383                 RETURN(0);
 384         }
 385
 386         desc = lustre_msg_buf(req->rq_repmsg, 0);
 387         rc = mds_get_lovdesc(req->rq_obd, desc);
 388         if (rc != 0 ) {
 389                 CERROR("mds_get_lovdesc error %d", rc);
 390                 req->rq_status = rc;
 391                 RETURN(0);
 392         }
 393
 394         tgt_count = NTOH__u32(desc->ld_tgt_count);
 395         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 396                 CERROR("too many targets, enlarge client buffers\n");
 397                 req->rq_status = -ENOSPC;
 398                 RETURN(0);
 399         }
 400
 401         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
 402                 tgt_count * sizeof(struct lov_object_id);
 403         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 404                              lustre_msg_buf(req->rq_repmsg, 1));
 405         if (rc) {
 406                 CERROR("get_lovtgts error %d\n", rc);
 407                 req->rq_status = rc;
 408                 RETURN(0);
 409         }
 410         RETURN(0);
 411 }
 412
 413 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 414                      void *data, __u32 data_len)
 415 {
 416         int do_ast;
 417         ENTRY;
 418
 419         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 420         lock->l_flags |= LDLM_FL_CBPENDING;
 421         do_ast = (!lock->l_readers && !lock->l_writers);
 422         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 423
 424         if (do_ast) {
 425                 struct lustre_handle lockh;
 426                 int rc;
 427
 428                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 429                 ldlm_lock2handle(lock, &lockh);
 430                 rc = ldlm_cli_cancel(&lockh);
 431                 if (rc < 0)
 432                         CERROR("ldlm_cli_cancel: %d\n", rc);
 433         } else
 434                 LDLM_DEBUG(lock, "Lock still has references, will be"
 435                            "cancelled later");
 436         RETURN(0);
 437 }
 438
 439 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 440                                 struct ptlrpc_request *req,
 441                                 struct mds_body *reqbody, int reply_off)
 442 {
 443         struct mds_body *body;
 444         struct inode *inode = dentry->d_inode;
 445         int rc;
 446         ENTRY;
 447
 448         if (inode == NULL)
 449                 RETURN(-ENOENT);
 450
 451         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 452
 453         mds_pack_inode2fid(&body->fid1, inode);
 454         mds_pack_inode2body(body, inode);
 455
 456         if (S_ISREG(inode->i_mode)) {
 457                 struct lov_mds_md *md;
 458
 459                 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 460                 md->lmd_easize = mds->mds_max_mdsize;
 461                 rc = mds_fs_get_md(mds, inode, md);
 462
 463                 if (rc < 0) {
 464                         if (rc == -ENODATA)
 465                                 RETURN(0);
 466                         CERROR("mds_fs_get_md failed: %d\n", rc);
 467                         RETURN(rc);
 468                 }
 469                 body->valid |= OBD_MD_FLEASIZE;
 470         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 471                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 472                 int len = req->rq_repmsg->buflens[reply_off + 1];
 473
 474                 rc = inode->i_op->readlink(dentry, symname, len);
 475                 if (rc < 0) {
 476                         CERROR("readlink failed: %d\n", rc);
 477                         RETURN(rc);
 478                 } else
 479                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 480
 481                 body->valid |= OBD_MD_LINKNAME;
 482         }
 483         RETURN(0);
 484 }
 485
 486 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 487 {
 488         struct mds_obd *mds = mds_req2mds(req);
 489         struct obd_device *obd = req->rq_export->exp_obd;
 490         struct obd_run_ctxt saved;
 491         struct mds_body *body;
 492         struct dentry *de = NULL, *dchild = NULL;
 493         struct inode *dir;
 494         struct lustre_handle lockh;
 495         char *name;
 496         int namelen, flags, lock_mode, rc = 0;
 497         __u64 res_id[3] = {0, 0, 0};
 498         ENTRY;
 499
 500         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 501
 502         if (req->rq_reqmsg->bufcount <= offset + 1) {
 503                 LBUG();
 504                 GOTO(out_pre_de, rc = -EINVAL);
 505         }
 506
 507         body = lustre_msg_buf(req->rq_reqmsg, offset);
 508         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 509         namelen = req->rq_reqmsg->buflens[offset + 1];
 510         /* requests were at offset 2, replies go back at 1 */
 511         if (offset)
 512                 offset = 1;
 513
 514         push_ctxt(&saved, &mds->mds_ctxt);
 515         de = mds_fid2dentry(mds, &body->fid1, NULL);
 516         if (IS_ERR(de)) {
 517                 LBUG();
 518                 GOTO(out_pre_de, rc = -ESTALE);
 519         }
 520
 521         dir = de->d_inode;
 522         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 523
 524         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 525         res_id[0] = dir->i_ino;
 526
 527         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 528                              NULL, 0, lock_mode, &lockh);
 529         if (rc == 0) {
 530                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 531                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 532                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 533                                       &flags, ldlm_completion_ast,
 534                                       mds_blocking_ast, NULL, 0, &lockh);
 535                 if (rc != ELDLM_OK) {
 536                         CERROR("lock enqueue: err: %d\n", rc);
 537                         GOTO(out_create_de, rc = -EIO);
 538                 }
 539         }
 540         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 541
 542         down(&dir->i_sem);
 543         dchild = lookup_one_len(name, de, namelen - 1);
 544         if (IS_ERR(dchild)) {
 545                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 546                 up(&dir->i_sem);
 547                 LBUG();
 548                 GOTO(out_create_dchild, rc = -ESTALE);
 549         }
 550
 551         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 552
 553         EXIT;
 554 out_create_dchild:
 555         l_dput(dchild);
 556         up(&dir->i_sem);
 557         ldlm_lock_decref(&lockh, lock_mode);
 558 out_create_de:
 559         l_dput(de);
 560 out_pre_de:
 561         req->rq_status = rc;
 562         pop_ctxt(&saved);
 563         return 0;
 564 }
 565
 566 static int mds_getattr(int offset, struct ptlrpc_request *req)
 567 {
 568         struct mds_obd *mds = mds_req2mds(req);
 569         struct obd_run_ctxt saved;
 570         struct dentry *de;
 571         struct inode *inode;
 572         struct mds_body *body;
 573         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 574         ENTRY;
 575
 576         body = lustre_msg_buf(req->rq_reqmsg, offset);
 577         push_ctxt(&saved, &mds->mds_ctxt);
 578         de = mds_fid2dentry(mds, &body->fid1, NULL);
 579         if (IS_ERR(de)) {
 580                 req->rq_status = -ENOENT;
 581                 GOTO(out_pop, rc = -ENOENT);
 582         }
 583
 584         inode = de->d_inode;
 585         if (S_ISREG(body->fid1.f_type)) {
 586                 bufcount = 2;
 587                 size[1] = mds->mds_max_mdsize;
 588         } else if (body->valid & OBD_MD_LINKNAME) {
 589                 bufcount = 2;
 590                 size[1] = inode->i_size + 1;
 591         }
 592
 593         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 594                              &req->rq_repmsg);
 595         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 596                 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
 597                 req->rq_status = rc;
 598                 GOTO(out, rc = 0);
 599         }
 600
 601         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 602
 603 out:
 604         l_dput(de);
 605 out_pop:
 606         pop_ctxt(&saved);
 607         RETURN(rc);
 608 }
 609
 610 static int mds_statfs(struct ptlrpc_request *req)
 611 {
 612         struct mds_obd *mds = mds_req2mds(req);
 613         struct obd_statfs *osfs;
 614         struct statfs sfs;
 615         int rc, size = sizeof(*osfs);
 616         ENTRY;
 617
 618         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 619                              &req->rq_repmsg);
 620         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 621                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 622                 GOTO(out, rc);
 623         }
 624
 625         rc = vfs_statfs(mds->mds_sb, &sfs);
 626         if (rc) {
 627                 CERROR("mds: statfs failed: rc %d\n", rc);
 628                 GOTO(out, rc);
 629         }
 630         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 631         memset(osfs, 0, size);
 632         obd_statfs_pack(osfs, &sfs);
 633
 634 out:
 635         req->rq_status = rc;
 636         RETURN(0);
 637 }
 638
 639 static int mds_open(struct ptlrpc_request *req)
 640 {
 641         struct dentry *de;
 642         struct mds_body *body;
 643         struct file *file;
 644         struct vfsmount *mnt;
 645         struct mds_obd *mds = mds_req2mds(req);
 646         struct mds_export_data *med;
 647         __u32 flags;
 648         struct list_head *tmp;
 649         struct mds_file_data *mfd;
 650         int rc, size = sizeof(*body);
 651         ENTRY;
 652
 653         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 654         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 655                 CERROR("mds: out of memory\n");
 656                 req->rq_status = -ENOMEM;
 657                 RETURN(0);
 658         }
 659
 660         body = lustre_msg_buf(req->rq_reqmsg, 0);
 661
 662         /* was this animal open already? */
 663         /* XXX we should only check on re-open, or do a refcount... */
 664         med = &req->rq_export->exp_mds_data;
 665         list_for_each(tmp, &med->med_open_head) {
 666                 struct mds_file_data *fd;
 667                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 668                 if (body->extra == fd->mfd_clientfd &&
 669                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 670                         CERROR("Re opening %Ld\n", body->fid1.id);
 671                         RETURN(0);
 672                 }
 673         }
 674
 675         OBD_ALLOC(mfd, sizeof(*mfd));
 676         if (!mfd) {
 677                 CERROR("mds: out of memory\n");
 678                 req->rq_status = -ENOMEM;
 679                 RETURN(0);
 680         }
 681
 682         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 683         if (IS_ERR(de)) {
 684                 req->rq_status = -ENOENT;
 685                 RETURN(0);
 686         }
 687
 688         /* check if this inode has seen a delayed object creation */
 689         if (req->rq_reqmsg->bufcount > 1) {
 690                 void *handle;
 691                 struct lov_mds_md *md;
 692                 struct inode *inode = de->d_inode;
 693                 int rc;
 694
 695                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 696
 697                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 698                 if (!handle) {
 699                         req->rq_status = -ENOMEM;
 700                         RETURN(0);
 701                 }
 702
 703                 /* XXX error handling */
 704                 rc = mds_fs_set_md(mds, inode, handle, md);
 705                 if (!rc) {
 706                         struct obd_run_ctxt saved;
 707                         push_ctxt(&saved, &mds->mds_ctxt);
 708                         rc = mds_update_last_rcvd(mds, handle, req);
 709                         pop_ctxt(&saved);
 710                 } else {
 711                         req->rq_status = rc;
 712                         RETURN(0);
 713                 }
 714                 /* FIXME: need to return last_rcvd, last_committed */
 715
 716                 /* FIXME: keep rc intact */
 717                 rc = mds_fs_commit(mds, de->d_inode, handle);
 718                 if (rc) {
 719                         req->rq_status = rc;
 720                         RETURN(0);
 721                 }
 722         }
 723
 724         flags = body->flags;
 725         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 726         if (!file || IS_ERR(file)) {
 727                 req->rq_status = -EINVAL;
 728                 OBD_FREE(mfd, sizeof(*mfd));
 729                 RETURN(0);
 730         }
 731
 732         file->private_data = mfd;
 733         mfd->mfd_file = file;
 734         mfd->mfd_clientfd = body->extra;
 735         list_add(&mfd->mfd_list, &med->med_open_head);
 736
 737         body = lustre_msg_buf(req->rq_repmsg, 0);
 738         /* FIXME: need to have cookies involved here */
 739         body->extra = (__u64) (unsigned long)file;
 740         RETURN(0);
 741 }
 742
 743 static int mds_close(struct ptlrpc_request *req)
 744 {
 745         struct dentry *de;
 746         struct mds_body *body;
 747         struct file *file;
 748         struct mds_obd *mds = mds_req2mds(req);
 749         struct vfsmount *mnt;
 750         struct mds_file_data *mfd;
 751         int rc;
 752         ENTRY;
 753
 754         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 755         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 756                 CERROR("mds: out of memory\n");
 757                 req->rq_status = -ENOMEM;
 758                 RETURN(0);
 759         }
 760
 761         body = lustre_msg_buf(req->rq_reqmsg, 0);
 762         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 763         if (IS_ERR(de)) {
 764                 req->rq_status = -ENOENT;
 765                 RETURN(0);
 766         }
 767
 768         /* FIXME: need to have cookies involved here */
 769         file = (struct file *)(unsigned long)body->extra;
 770         if (!file->f_dentry)
 771                 LBUG();
 772         mfd = (struct mds_file_data *)file->private_data;
 773         list_del(&mfd->mfd_list);
 774         OBD_FREE(mfd, sizeof(*mfd));
 775
 776         req->rq_status = filp_close(file, 0);
 777         l_dput(de);
 778         mntput(mnt);
 779
 780         RETURN(0);
 781 }
 782
 783 static int mds_readpage(struct ptlrpc_request *req)
 784 {
 785         struct mds_obd *mds = mds_req2mds(req);
 786         struct vfsmount *mnt;
 787         struct dentry *de;
 788         struct file *file;
 789         struct mds_body *body;
 790         struct obd_run_ctxt saved;
 791         int rc, size = sizeof(*body);
 792         ENTRY;
 793
 794         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 795         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 796                 CERROR("mds: out of memory\n");
 797                 GOTO(out, rc = -ENOMEM);
 798         }
 799
 800         body = lustre_msg_buf(req->rq_reqmsg, 0);
 801         push_ctxt(&saved, &mds->mds_ctxt);
 802         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 803         if (IS_ERR(de))
 804                 GOTO(out_pop, rc = PTR_ERR(de));
 805
 806         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 807
 808         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 809         /* note: in case of an error, dentry_open puts dentry */
 810         if (IS_ERR(file))
 811                 GOTO(out_pop, rc = PTR_ERR(file));
 812
 813         /* to make this asynchronous make sure that the handling function
 814            doesn't send a reply when this function completes. Instead a
 815            callback function would send the reply */
 816         rc = mds_sendpage(req, file, body->size);
 817
 818         filp_close(file, 0);
 819 out_pop:
 820         pop_ctxt(&saved);
 821 out:
 822         req->rq_status = rc;
 823         RETURN(0);
 824 }
 825
 826 int mds_reint(int offset, struct ptlrpc_request *req)
 827 {
 828         int rc;
 829         struct mds_update_record rec;
 830
 831         rc = mds_update_unpack(req, offset, &rec);
 832         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 833                 CERROR("invalid record\n");
 834                 req->rq_status = -EINVAL;
 835                 RETURN(0);
 836         }
 837         /* rc will be used to interrupt a for loop over multiple records */
 838         rc = mds_reint_rec(&rec, offset, req);
 839         return rc;
 840 }
 841
 842 int mds_handle(struct ptlrpc_request *req)
 843 {
 844         int rc;
 845         ENTRY;
 846
 847         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 848         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 849                 CERROR("lustre_mds: Invalid request\n");
 850                 GOTO(out, rc);
 851         }
 852
 853         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 854                 CERROR("lustre_mds: wrong packet type sent %d\n",
 855                        req->rq_reqmsg->type);
 856                 GOTO(out, rc = -EINVAL);
 857         }
 858
 859         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
 860                 GOTO(out, rc = -ENOTCONN);
 861
 862         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 863                 GOTO(out, rc = -EINVAL);
 864
 865         switch (req->rq_reqmsg->opc) {
 866         case MDS_CONNECT:
 867                 CDEBUG(D_INODE, "connect\n");
 868                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 869                 rc = target_handle_connect(req);
 870                 break;
 871
 872         case MDS_DISCONNECT:
 873                 CDEBUG(D_INODE, "disconnect\n");
 874                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 875                 rc = target_handle_disconnect(req);
 876                 goto out;
 877
 878         case MDS_GETSTATUS:
 879                 CDEBUG(D_INODE, "getstatus\n");
 880                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 881                 rc = mds_getstatus(req);
 882                 break;
 883
 884         case MDS_GETLOVINFO:
 885                 CDEBUG(D_INODE, "getlovinfo\n");
 886                 rc = mds_getlovinfo(req);
 887                 break;
 888
 889         case MDS_GETATTR:
 890                 CDEBUG(D_INODE, "getattr\n");
 891                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 892                 rc = mds_getattr(0, req);
 893                 break;
 894
 895         case MDS_STATFS:
 896                 CDEBUG(D_INODE, "statfs\n");
 897                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 898                 rc = mds_statfs(req);
 899                 break;
 900
 901         case MDS_READPAGE:
 902                 CDEBUG(D_INODE, "readpage\n");
 903                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 904                 rc = mds_readpage(req);
 905
 906                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 907                         return 0;
 908                 break;
 909
 910         case MDS_REINT: {
 911                 int size = sizeof(struct mds_body);
 912                 CDEBUG(D_INODE, "reint\n");
 913                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 914
 915                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 916                                      &req->rq_repmsg);
 917                 if (rc) {
 918                         req->rq_status = rc;
 919                         break;
 920                 }
 921                 rc = mds_reint(0, req);
 922                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 923                 break;
 924                 }
 925
 926         case MDS_OPEN:
 927                 CDEBUG(D_INODE, "open\n");
 928                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 929                 rc = mds_open(req);
 930                 break;
 931
 932         case MDS_CLOSE:
 933                 CDEBUG(D_INODE, "close\n");
 934                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 935                 rc = mds_close(req);
 936                 break;
 937
 938         case LDLM_ENQUEUE:
 939                 CDEBUG(D_INODE, "enqueue\n");
 940                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 941                 rc = ldlm_handle_enqueue(req);
 942                 if (rc)
 943                         break;
 944                 RETURN(0);
 945
 946         case LDLM_CONVERT:
 947                 CDEBUG(D_INODE, "convert\n");
 948                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 949                 rc = ldlm_handle_convert(req);
 950                 if (rc)
 951                         break;
 952                 RETURN(0);
 953
 954         case LDLM_CANCEL:
 955                 CDEBUG(D_INODE, "cancel\n");
 956                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 957                 rc = ldlm_handle_cancel(req);
 958                 if (rc)
 959                         break;
 960                 RETURN(0);
 961         case LDLM_BL_CALLBACK:
 962         case LDLM_CP_CALLBACK:
 963                 CDEBUG(D_INODE, "callback\n");
 964                 CERROR("callbacks should not happen on MDS\n");
 965                 LBUG();
 966                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 967                 break;
 968         default:
 969                 rc = ptlrpc_error(req->rq_svc, req);
 970                 RETURN(rc);
 971         }
 972
 973         EXIT;
 974
 975         if (!rc) {
 976                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
 977                 struct mds_obd *mds = mds_req2mds(req);
 978
 979                 req->rq_repmsg->last_xid =
 980                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
 981                 req->rq_repmsg->last_committed =
 982                         HTON__u64(mds->mds_last_committed);
 983                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
 984                        (unsigned long long)mds->mds_last_rcvd,
 985                        (unsigned long long)mds->mds_last_committed,
 986                        cpu_to_le32(req->rq_xid));
 987         }
 988  out:
 989         if (rc) {
 990                 CERROR("mds: processing error %d\n", rc);
 991                 ptlrpc_error(req->rq_svc, req);
 992         } else {
 993                 CDEBUG(D_NET, "sending reply\n");
 994                 ptlrpc_reply(req->rq_svc, req);
 995         }
 996         return 0;
 997 }
 998
 999 /* Update the server data on disk.  This stores the new mount_count and
1000  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1001  * then the server last_rcvd value may be less than that of the clients.
1002  * This will alert us that we may need to do client recovery.
1003  *
1004  * Assumes we are already in the server filesystem context.
1005  *
1006  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1007  */
1008 static
1009 int mds_update_server_data(struct mds_obd *mds)
1010 {
1011         struct mds_server_data *msd = mds->mds_server_data;
1012         struct file *filp = mds->mds_rcvd_filp;
1013         loff_t off = 0;
1014         int rc;
1015
1016         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1017         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1018
1019         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1020                (unsigned long long)mds->mds_mount_count,
1021                (unsigned long long)mds->mds_last_rcvd);
1022         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1023         if (rc != sizeof(*msd)) {
1024                 CERROR("error writing MDS server data: rc = %d\n", rc);
1025                 if (rc > 0)
1026                         RETURN(-EIO);
1027                 RETURN(rc);
1028         }
1029         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1030         if (rc)
1031                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1032
1033         return 0;
1034 }
1035
1036 /* Do recovery actions for the MDS */
1037 static int mds_recover(struct obd_device *obddev)
1038 {
1039         struct mds_obd *mds = &obddev->u.mds;
1040         struct obd_run_ctxt saved;
1041         int rc;
1042
1043         /* This happens at the end when recovery is complete */
1044         ++mds->mds_mount_count;
1045         push_ctxt(&saved, &mds->mds_ctxt);
1046         rc = mds_update_server_data(mds);
1047         pop_ctxt(&saved);
1048
1049         return rc;
1050 }
1051
1052 #define MDS_NUM_THREADS 8
1053 /* mount the file system (secretly) */
1054 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1055 {
1056         int i;
1057         struct obd_ioctl_data* data = buf;
1058         struct mds_obd *mds = &obddev->u.mds;
1059         struct vfsmount *mnt;
1060         int rc = 0;
1061         ENTRY;
1062
1063         MOD_INC_USE_COUNT;
1064 #ifdef CONFIG_DEV_RDONLY
1065         dev_clear_rdonly(2);
1066 #endif
1067         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1068                 GOTO(err_dec, rc = -EINVAL);
1069
1070         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1071
1072         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1073         if (IS_ERR(mnt)) {
1074                 rc = PTR_ERR(mnt);
1075                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1076                 GOTO(err_kfree, rc);
1077         }
1078
1079         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1080         if (!mds->mds_sb)
1081                 GOTO(err_put, rc = -ENODEV);
1082
1083         spin_lock_init(&mds->mds_last_lock);
1084         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1085         rc = mds_fs_setup(obddev, mnt);
1086         if (rc) {
1087                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1088                 GOTO(err_put, rc);
1089         }
1090
1091         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1092                                            MDC_REPLY_PORTAL, "self",mds_handle,
1093                                            "mds");
1094         if (!mds->mds_service) {
1095                 CERROR("failed to start service\n");
1096                 GOTO(err_fs, rc = -EINVAL);
1097         }
1098
1099         obddev->obd_namespace =
1100                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1101         if (obddev->obd_namespace == NULL) {
1102                 mds_cleanup(obddev);
1103                 GOTO(err_svc, rc = -ENOMEM);
1104         }
1105
1106         for (i = 0; i < MDS_NUM_THREADS; i++) {
1107                 char name[32];
1108                 sprintf(name, "lustre_MDS_%02d", i);
1109                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1110                 if (rc) {
1111                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1112                         GOTO(err_thread, rc);
1113                 }
1114         }
1115
1116         rc = mds_recover(obddev);
1117         if (rc)
1118                 GOTO(err_thread, rc);
1119
1120         mds_destroy_export = mds_client_free;
1121
1122         RETURN(0);
1123
1124 err_thread:
1125         ptlrpc_stop_all_threads(mds->mds_service);
1126 err_svc:
1127         ptlrpc_unregister_service(mds->mds_service);
1128 err_fs:
1129         mds_fs_cleanup(obddev);
1130 err_put:
1131         unlock_kernel();
1132         mntput(mds->mds_vfsmnt);
1133         mds->mds_sb = 0;
1134         lock_kernel();
1135 err_kfree:
1136         kfree(mds->mds_fstype);
1137 err_dec:
1138         MOD_DEC_USE_COUNT;
1139         return rc;
1140 }
1141
1142 static int mds_cleanup(struct obd_device *obddev)
1143 {
1144         struct super_block *sb;
1145         struct mds_obd *mds = &obddev->u.mds;
1146         struct obd_run_ctxt saved;
1147         ENTRY;
1148
1149         ptlrpc_stop_all_threads(mds->mds_service);
1150         ptlrpc_unregister_service(mds->mds_service);
1151
1152         sb = mds->mds_sb;
1153         if (!mds->mds_sb)
1154                 RETURN(0);
1155
1156         push_ctxt(&saved, &mds->mds_ctxt);
1157         mds_update_server_data(mds);
1158
1159         if (mds->mds_rcvd_filp) {
1160                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1161                 mds->mds_rcvd_filp = NULL;
1162
1163                 if (rc)
1164                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1165         }
1166         pop_ctxt(&saved);
1167
1168         unlock_kernel();
1169         mntput(mds->mds_vfsmnt);
1170         mds->mds_sb = 0;
1171         kfree(mds->mds_fstype);
1172
1173         ldlm_namespace_free(obddev->obd_namespace);
1174
1175         lock_kernel();
1176 #ifdef CONFIG_DEV_RDONLY
1177         dev_clear_rdonly(2);
1178 #endif
1179         mds_fs_cleanup(obddev);
1180
1181         MOD_DEC_USE_COUNT;
1182         RETURN(0);
1183 }
1184
1185 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1186                               ldlm_mode_t mode, void *data)
1187 {
1188         struct ptlrpc_request *req = req_cookie;
1189         int rc = 0;
1190         ENTRY;
1191
1192         if (!req_cookie)
1193                 RETURN(0);
1194
1195         if (req->rq_reqmsg->bufcount > 1) {
1196                 /* an intent needs to be considered */
1197                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1198                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1199                 struct mds_body *mds_rep;
1200                 struct ldlm_reply *rep;
1201                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1202                 int rc, size[3] = {sizeof(struct ldlm_reply),
1203                                                   sizeof(struct mds_body),
1204                                                   mds->mds_max_mdsize};
1205
1206                 it->opc = NTOH__u64(it->opc);
1207
1208                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1209                            ldlm_it2str(it->opc));
1210
1211                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1212                                      &req->rq_repmsg);
1213                 if (rc) {
1214                         rc = req->rq_status = -ENOMEM;
1215                         RETURN(rc);
1216                 }
1217
1218                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1219                 rep->lock_policy_res1 = 1;
1220
1221                 /* execute policy */
1222                 switch ((long)it->opc) {
1223                 case IT_CREAT|IT_OPEN:
1224                         rc = mds_reint(2, req);
1225                         if (rc || (req->rq_status != 0 &&
1226                                    req->rq_status != -EEXIST)) {
1227                                 rep->lock_policy_res2 = req->rq_status;
1228                                 RETURN(ELDLM_LOCK_ABORTED);
1229                         }
1230                         break;
1231                 case IT_CREAT:
1232                 case IT_LINK:
1233                 case IT_MKDIR:
1234                 case IT_MKNOD:
1235                 case IT_RENAME2:
1236                 case IT_RMDIR:
1237                 case IT_SYMLINK:
1238                 case IT_UNLINK:
1239                         rc = mds_reint(2, req);
1240                         if (rc || (req->rq_status != 0 &&
1241                                    req->rq_status != -EISDIR &&
1242                                    req->rq_status != -ENOTDIR)) {
1243                                 rep->lock_policy_res2 = req->rq_status;
1244                                 RETURN(ELDLM_LOCK_ABORTED);
1245                         }
1246                         break;
1247                 case IT_GETATTR:
1248                 case IT_LOOKUP:
1249                 case IT_OPEN:
1250                 case IT_READDIR:
1251                 case IT_READLINK:
1252                 case IT_RENAME:
1253                 case IT_SETATTR:
1254                         rc = mds_getattr_name(2, req);
1255                         /* FIXME: we need to sit down and decide on who should
1256                          * set req->rq_status, who should return negative and
1257                          * positive return values, and what they all mean. */
1258                         if (rc || req->rq_status != 0) {
1259                                 rep->lock_policy_res2 = req->rq_status;
1260                                 RETURN(ELDLM_LOCK_ABORTED);
1261                         }
1262                         break;
1263                 case IT_READDIR|IT_OPEN:
1264                         LBUG();
1265                         break;
1266                 default:
1267                         CERROR("Unhandled intent\n");
1268                         LBUG();
1269                 }
1270
1271                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1272                     it->opc == IT_RENAME || it->opc == IT_RENAME2)
1273                         RETURN(ELDLM_LOCK_ABORTED);
1274
1275                 rep->lock_policy_res2 = req->rq_status;
1276                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1277
1278                 /* If the client is about to open a file that doesn't have an MD
1279                  * stripe record, it's going to need a write lock. */
1280                 if (it->opc & IT_OPEN) {
1281                         struct lov_mds_md *md =
1282                                 lustre_msg_buf(req->rq_repmsg, 2);
1283                         if (md->lmd_easize == 0) {
1284                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1285                                            " lock");
1286                                 lock->l_req_mode = LCK_PW;
1287                         }
1288                 }
1289
1290                 /* Give the client a lock on the child object, instead of the
1291                  * parent that it requested. */
1292                 new_resid[0] = NTOH__u32(mds_rep->ino);
1293                 if (new_resid[0] == 0)
1294                         LBUG();
1295                 old_res = lock->l_resource->lr_name[0];
1296
1297                 ldlm_lock_change_resource(lock, new_resid);
1298                 if (lock->l_resource == NULL) {
1299                         LBUG();
1300                         RETURN(-ENOMEM);
1301                 }
1302                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1303                            (long)old_res);
1304                 RETURN(ELDLM_LOCK_CHANGED);
1305         } else {
1306                 int size = sizeof(struct ldlm_reply);
1307                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1308                                      &req->rq_repmsg);
1309                 if (rc) {
1310                         LBUG();
1311                         RETURN(-ENOMEM);
1312                 }
1313         }
1314         RETURN(rc);
1315 }
1316
1317
1318 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1319                          int len, void *karg, void *uarg);
1320
1321 /* use obd ops to offer management infrastructure */
1322 static struct obd_ops mds_obd_ops = {
1323         o_connect:     mds_connect,
1324         o_disconnect:  mds_disconnect,
1325         o_setup:       mds_setup,
1326         o_cleanup:     mds_cleanup,
1327         o_iocontrol:   mds_iocontrol
1328 };
1329
1330 static int __init mds_init(void)
1331 {
1332         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1333         ldlm_register_intent(ldlm_intent_policy);
1334         return 0;
1335 }
1336
1337 static void __exit mds_exit(void)
1338 {
1339         ldlm_unregister_intent();
1340         class_unregister_type(LUSTRE_MDS_NAME);
1341 }
1342
1343 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1344 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1345 MODULE_LICENSE("GPL");
1346
1347 module_init(mds_init);
1348 module_exit(mds_exit);