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->bd_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->bp_xid = req->rq_xid;
  88         bulk->bp_buf = buf;
  89         bulk->bp_buflen = PAGE_SIZE;
  90         desc->bd_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->bd_waitq, desc->bd_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_obd_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                         LASSERT(list_empty(&exp->exp_conn_chain));
 285                         conn->addr = (__u64) (unsigned long)exp;
 286                         conn->cookie = exp->exp_cookie;
 287                         spin_unlock(&obd->obd_dev_lock);
 288                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 289                                cluuid, exp);
 290                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 291                                (long long)conn->addr, (long long)conn->cookie);
 292                         RETURN(0);
 293                 }
 294         }
 295         spin_unlock(&obd->obd_dev_lock);
 296         /* XXX There is a small race between checking the list and adding a
 297          * new connection for the same UUID, but the real threat (list
 298          * corruption when multiple different clients connect) is solved.
 299          */
 300         rc = class_connect(conn, obd, cluuid);
 301         if (rc)
 302                 GOTO(out_dec, rc);
 303         exp = class_conn2export(conn);
 304         LASSERT(exp);
 305
 306         OBD_ALLOC(mcd, sizeof(*mcd));
 307         if (!mcd) {
 308                 CERROR("mds: out of memory for client data\n");
 309                 GOTO(out_export, rc = -ENOMEM);
 310         }
 311         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 312         exp->exp_mds_data.med_mcd = mcd;
 313         rc = mds_client_add(&exp->exp_mds_data, -1);
 314         if (rc)
 315                 GOTO(out_mdc, rc);
 316
 317         RETURN(0);
 318
 319 out_mdc:
 320         OBD_FREE(mcd, sizeof(*mcd));
 321 out_export:
 322         class_disconnect(conn);
 323 out_dec:
 324         MOD_DEC_USE_COUNT;
 325
 326         return rc;
 327 }
 328
 329 static int mds_disconnect(struct lustre_handle *conn)
 330 {
 331         int rc;
 332
 333         rc = class_disconnect(conn);
 334         if (!rc)
 335                 MOD_DEC_USE_COUNT;
 336
 337         return rc;
 338 }
 339
 340 static int mds_getstatus(struct ptlrpc_request *req)
 341 {
 342         struct mds_obd *mds = mds_req2mds(req);
 343         struct mds_body *body;
 344         int rc, size = sizeof(*body);
 345         ENTRY;
 346
 347         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 348         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 349                 CERROR("mds: out of memory for message: size=%d\n", size);
 350                 req->rq_status = -ENOMEM;
 351                 RETURN(0);
 352         }
 353
 354         /* Anything we need to do here with the client's trans no or so? */
 355         body = lustre_msg_buf(req->rq_repmsg, 0);
 356         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 357
 358         /* the last_committed and last_xid fields are filled in for all
 359          * replies already - no need to do so here also.
 360          */
 361         RETURN(0);
 362 }
 363
 364 static int mds_getlovinfo(struct ptlrpc_request *req)
 365 {
 366         struct mds_obd *mds = mds_req2mds(req);
 367         struct mds_status_req *streq;
 368         struct lov_desc *desc;
 369         int tgt_count;
 370         int rc, size[2] = {sizeof(*desc)};
 371         ENTRY;
 372
 373         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 374         streq->flags = NTOH__u32(streq->flags);
 375         streq->repbuf = NTOH__u32(streq->repbuf);
 376         size[1] = streq->repbuf;
 377
 378         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 379         if (rc) {
 380                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 381                 req->rq_status = -ENOMEM;
 382                 RETURN(0);
 383         }
 384
 385         desc = lustre_msg_buf(req->rq_repmsg, 0);
 386         rc = mds_get_lovdesc(req->rq_obd, desc);
 387         if (rc != 0 ) {
 388                 CERROR("mds_get_lovdesc error %d", rc);
 389                 req->rq_status = rc;
 390                 RETURN(0);
 391         }
 392
 393         tgt_count = NTOH__u32(desc->ld_tgt_count);
 394         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 395                 CERROR("too many targets, enlarge client buffers\n");
 396                 req->rq_status = -ENOSPC;
 397                 RETURN(0);
 398         }
 399
 400         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
 401                 tgt_count * sizeof(struct lov_object_id);
 402         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 403                              lustre_msg_buf(req->rq_repmsg, 1));
 404         if (rc) {
 405                 CERROR("get_lovtgts error %d\n", rc);
 406                 req->rq_status = rc;
 407                 RETURN(0);
 408         }
 409         RETURN(0);
 410 }
 411
 412 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 413                      void *data, __u32 data_len)
 414 {
 415         int do_ast;
 416         ENTRY;
 417
 418         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 419         lock->l_flags |= LDLM_FL_CBPENDING;
 420         do_ast = (!lock->l_readers && !lock->l_writers);
 421         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 422
 423         if (do_ast) {
 424                 struct lustre_handle lockh;
 425                 int rc;
 426
 427                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 428                 ldlm_lock2handle(lock, &lockh);
 429                 rc = ldlm_cli_cancel(&lockh);
 430                 if (rc < 0)
 431                         CERROR("ldlm_cli_cancel: %d\n", rc);
 432         } else
 433                 LDLM_DEBUG(lock, "Lock still has references, will be"
 434                            "cancelled later");
 435         RETURN(0);
 436 }
 437
 438 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 439                                 struct ptlrpc_request *req,
 440                                 struct mds_body *reqbody, int reply_off)
 441 {
 442         struct mds_body *body;
 443         struct inode *inode = dentry->d_inode;
 444         int rc;
 445         ENTRY;
 446
 447         if (inode == NULL)
 448                 RETURN(-ENOENT);
 449
 450         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 451
 452         mds_pack_inode2fid(&body->fid1, inode);
 453         mds_pack_inode2body(body, inode);
 454
 455         if (S_ISREG(inode->i_mode)) {
 456                 struct lov_mds_md *md;
 457
 458                 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 459                 md->lmd_easize = mds->mds_max_mdsize;
 460                 rc = mds_fs_get_md(mds, inode, md);
 461
 462                 if (rc < 0) {
 463                         if (rc == -ENODATA)
 464                                 RETURN(0);
 465                         CERROR("mds_fs_get_md failed: %d\n", rc);
 466                         RETURN(rc);
 467                 }
 468                 body->valid |= OBD_MD_FLEASIZE;
 469         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 470                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 471                 int len = req->rq_repmsg->buflens[reply_off + 1];
 472
 473                 rc = inode->i_op->readlink(dentry, symname, len);
 474                 if (rc < 0) {
 475                         CERROR("readlink failed: %d\n", rc);
 476                         RETURN(rc);
 477                 } else
 478                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 479
 480                 body->valid |= OBD_MD_LINKNAME;
 481         }
 482         RETURN(0);
 483 }
 484
 485 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 486 {
 487         struct mds_obd *mds = mds_req2mds(req);
 488         struct obd_device *obd = req->rq_export->exp_obd;
 489         struct obd_run_ctxt saved;
 490         struct mds_body *body;
 491         struct dentry *de = NULL, *dchild = NULL;
 492         struct inode *dir;
 493         struct lustre_handle lockh;
 494         char *name;
 495         int namelen, flags, lock_mode, rc = 0;
 496         __u64 res_id[3] = {0, 0, 0};
 497         ENTRY;
 498
 499         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 500
 501         if (req->rq_reqmsg->bufcount <= offset + 1) {
 502                 LBUG();
 503                 GOTO(out_pre_de, rc = -EINVAL);
 504         }
 505
 506         body = lustre_msg_buf(req->rq_reqmsg, offset);
 507         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 508         namelen = req->rq_reqmsg->buflens[offset + 1];
 509         /* requests were at offset 2, replies go back at 1 */
 510         if (offset)
 511                 offset = 1;
 512
 513         push_ctxt(&saved, &mds->mds_ctxt);
 514         de = mds_fid2dentry(mds, &body->fid1, NULL);
 515         if (IS_ERR(de)) {
 516                 LBUG();
 517                 GOTO(out_pre_de, rc = -ESTALE);
 518         }
 519
 520         dir = de->d_inode;
 521         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 522
 523         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 524         res_id[0] = dir->i_ino;
 525
 526         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 527                              NULL, 0, lock_mode, &lockh);
 528         if (rc == 0) {
 529                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 530                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 531                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 532                                       &flags, ldlm_completion_ast,
 533                                       mds_blocking_ast, NULL, 0, &lockh);
 534                 if (rc != ELDLM_OK) {
 535                         CERROR("lock enqueue: err: %d\n", rc);
 536                         GOTO(out_create_de, rc = -EIO);
 537                 }
 538         }
 539         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 540
 541         down(&dir->i_sem);
 542         dchild = lookup_one_len(name, de, namelen - 1);
 543         if (IS_ERR(dchild)) {
 544                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 545                 up(&dir->i_sem);
 546                 LBUG();
 547                 GOTO(out_create_dchild, rc = -ESTALE);
 548         }
 549
 550         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 551
 552         EXIT;
 553 out_create_dchild:
 554         l_dput(dchild);
 555         up(&dir->i_sem);
 556         ldlm_lock_decref(&lockh, lock_mode);
 557 out_create_de:
 558         l_dput(de);
 559 out_pre_de:
 560         req->rq_status = rc;
 561         pop_ctxt(&saved);
 562         return 0;
 563 }
 564
 565 static int mds_getattr(int offset, struct ptlrpc_request *req)
 566 {
 567         struct mds_obd *mds = mds_req2mds(req);
 568         struct obd_run_ctxt saved;
 569         struct dentry *de;
 570         struct inode *inode;
 571         struct mds_body *body;
 572         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 573         ENTRY;
 574
 575         body = lustre_msg_buf(req->rq_reqmsg, offset);
 576         push_ctxt(&saved, &mds->mds_ctxt);
 577         de = mds_fid2dentry(mds, &body->fid1, NULL);
 578         if (IS_ERR(de)) {
 579                 req->rq_status = -ENOENT;
 580                 GOTO(out_pop, rc = -ENOENT);
 581         }
 582
 583         inode = de->d_inode;
 584         if (S_ISREG(body->fid1.f_type)) {
 585                 bufcount = 2;
 586                 size[1] = mds->mds_max_mdsize;
 587         } else if (body->valid & OBD_MD_LINKNAME) {
 588                 bufcount = 2;
 589                 size[1] = inode->i_size + 1;
 590         }
 591
 592         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 593                              &req->rq_repmsg);
 594         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 595                 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
 596                 req->rq_status = rc;
 597                 GOTO(out, rc = 0);
 598         }
 599
 600         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 601
 602 out:
 603         l_dput(de);
 604 out_pop:
 605         pop_ctxt(&saved);
 606         RETURN(rc);
 607 }
 608
 609 static int mds_statfs(struct ptlrpc_request *req)
 610 {
 611         struct mds_obd *mds = mds_req2mds(req);
 612         struct obd_statfs *osfs;
 613         struct statfs sfs;
 614         int rc, size = sizeof(*osfs);
 615         ENTRY;
 616
 617         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 618                              &req->rq_repmsg);
 619         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 620                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 621                 GOTO(out, rc);
 622         }
 623
 624         rc = vfs_statfs(mds->mds_sb, &sfs);
 625         if (rc) {
 626                 CERROR("mds: statfs failed: rc %d\n", rc);
 627                 GOTO(out, rc);
 628         }
 629         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 630         memset(osfs, 0, size);
 631         statfs_pack(osfs, &sfs);
 632         obd_statfs_pack(osfs, osfs);
 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->opc != MDS_CONNECT && req->rq_export == NULL)
 854                 GOTO(out, rc = -ENOTCONN);
 855
 856         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 857                 GOTO(out, rc = -EINVAL);
 858
 859         switch (req->rq_reqmsg->opc) {
 860         case MDS_CONNECT:
 861                 CDEBUG(D_INODE, "connect\n");
 862                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 863                 rc = target_handle_connect(req);
 864                 break;
 865
 866         case MDS_DISCONNECT:
 867                 CDEBUG(D_INODE, "disconnect\n");
 868                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 869                 rc = target_handle_disconnect(req);
 870                 goto out;
 871
 872         case MDS_GETSTATUS:
 873                 CDEBUG(D_INODE, "getstatus\n");
 874                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 875                 rc = mds_getstatus(req);
 876                 break;
 877
 878         case MDS_GETLOVINFO:
 879                 CDEBUG(D_INODE, "getlovinfo\n");
 880                 rc = mds_getlovinfo(req);
 881                 break;
 882
 883         case MDS_GETATTR:
 884                 CDEBUG(D_INODE, "getattr\n");
 885                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 886                 rc = mds_getattr(0, req);
 887                 break;
 888
 889         case MDS_STATFS:
 890                 CDEBUG(D_INODE, "statfs\n");
 891                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 892                 rc = mds_statfs(req);
 893                 break;
 894
 895         case MDS_READPAGE:
 896                 CDEBUG(D_INODE, "readpage\n");
 897                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 898                 rc = mds_readpage(req);
 899
 900                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 901                         return 0;
 902                 break;
 903
 904         case MDS_REINT: {
 905                 int size = sizeof(struct mds_body);
 906                 CDEBUG(D_INODE, "reint\n");
 907                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 908
 909                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 910                                      &req->rq_repmsg);
 911                 if (rc) {
 912                         req->rq_status = rc;
 913                         break;
 914                 }
 915                 rc = mds_reint(0, req);
 916                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 917                 break;
 918                 }
 919
 920         case MDS_OPEN:
 921                 CDEBUG(D_INODE, "open\n");
 922                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 923                 rc = mds_open(req);
 924                 break;
 925
 926         case MDS_CLOSE:
 927                 CDEBUG(D_INODE, "close\n");
 928                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 929                 rc = mds_close(req);
 930                 break;
 931
 932         case LDLM_ENQUEUE:
 933                 CDEBUG(D_INODE, "enqueue\n");
 934                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 935                 rc = ldlm_handle_enqueue(req);
 936                 if (rc)
 937                         break;
 938                 RETURN(0);
 939
 940         case LDLM_CONVERT:
 941                 CDEBUG(D_INODE, "convert\n");
 942                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 943                 rc = ldlm_handle_convert(req);
 944                 if (rc)
 945                         break;
 946                 RETURN(0);
 947
 948         case LDLM_CANCEL:
 949                 CDEBUG(D_INODE, "cancel\n");
 950                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 951                 rc = ldlm_handle_cancel(req);
 952                 if (rc)
 953                         break;
 954                 RETURN(0);
 955         case LDLM_BL_CALLBACK:
 956         case LDLM_CP_CALLBACK:
 957                 CDEBUG(D_INODE, "callback\n");
 958                 CERROR("callbacks should not happen on MDS\n");
 959                 LBUG();
 960                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 961                 break;
 962         default:
 963                 rc = ptlrpc_error(req->rq_svc, req);
 964                 RETURN(rc);
 965         }
 966
 967         EXIT;
 968
 969         if (!rc) {
 970                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
 971                 struct mds_obd *mds = mds_req2mds(req);
 972
 973                 req->rq_repmsg->last_xid =
 974                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
 975                 req->rq_repmsg->last_committed =
 976                         HTON__u64(mds->mds_last_committed);
 977                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
 978                        (unsigned long long)mds->mds_last_rcvd,
 979                        (unsigned long long)mds->mds_last_committed,
 980                        cpu_to_le32(req->rq_xid));
 981         }
 982  out:
 983         if (rc) {
 984                 CERROR("mds: processing error %d\n", rc);
 985                 ptlrpc_error(req->rq_svc, req);
 986         } else {
 987                 CDEBUG(D_NET, "sending reply\n");
 988                 ptlrpc_reply(req->rq_svc, req);
 989         }
 990         return 0;
 991 }
 992
 993 /* Update the server data on disk.  This stores the new mount_count and
 994  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 995  * then the server last_rcvd value may be less than that of the clients.
 996  * This will alert us that we may need to do client recovery.
 997  *
 998  * Assumes we are already in the server filesystem context.
 999  *
1000  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1001  */
1002 static
1003 int mds_update_server_data(struct mds_obd *mds)
1004 {
1005         struct mds_server_data *msd = mds->mds_server_data;
1006         struct file *filp = mds->mds_rcvd_filp;
1007         loff_t off = 0;
1008         int rc;
1009
1010         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1011         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1012
1013         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1014                (unsigned long long)mds->mds_mount_count,
1015                (unsigned long long)mds->mds_last_rcvd);
1016         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1017         if (rc != sizeof(*msd)) {
1018                 CERROR("error writing MDS server data: rc = %d\n", rc);
1019                 if (rc > 0)
1020                         RETURN(-EIO);
1021                 RETURN(rc);
1022         }
1023         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1024         if (rc)
1025                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1026
1027         return 0;
1028 }
1029
1030 /* Do recovery actions for the MDS */
1031 static int mds_recover(struct obd_device *obddev)
1032 {
1033         struct mds_obd *mds = &obddev->u.mds;
1034         struct obd_run_ctxt saved;
1035         int rc;
1036
1037         /* This happens at the end when recovery is complete */
1038         ++mds->mds_mount_count;
1039         push_ctxt(&saved, &mds->mds_ctxt);
1040         rc = mds_update_server_data(mds);
1041         pop_ctxt(&saved);
1042
1043         return rc;
1044 }
1045
1046 #define MDS_NUM_THREADS 8
1047 /* mount the file system (secretly) */
1048 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1049 {
1050         int i;
1051         struct obd_ioctl_data* data = buf;
1052         struct mds_obd *mds = &obddev->u.mds;
1053         struct vfsmount *mnt;
1054         int rc = 0;
1055         ENTRY;
1056
1057         MOD_INC_USE_COUNT;
1058 #ifdef CONFIG_DEV_RDONLY
1059         dev_clear_rdonly(2);
1060 #endif
1061         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1062                 GOTO(err_dec, rc = -EINVAL);
1063
1064         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1065
1066         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1067         if (IS_ERR(mnt)) {
1068                 rc = PTR_ERR(mnt);
1069                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1070                 GOTO(err_kfree, rc);
1071         }
1072
1073         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1074         if (!mds->mds_sb)
1075                 GOTO(err_put, rc = -ENODEV);
1076
1077         spin_lock_init(&mds->mds_last_lock);
1078         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1079         rc = mds_fs_setup(obddev, mnt);
1080         if (rc) {
1081                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1082                 GOTO(err_put, rc);
1083         }
1084
1085         mds->mds_service = ptlrpc_init_svc(1024, 640, MDS_REQUEST_PORTAL,
1086                                            MDC_REPLY_PORTAL, "self",mds_handle,
1087                                            "mds");
1088         if (!mds->mds_service) {
1089                 CERROR("failed to start service\n");
1090                 GOTO(err_fs, rc = -EINVAL);
1091         }
1092
1093         obddev->obd_namespace =
1094                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1095         if (obddev->obd_namespace == NULL) {
1096                 mds_cleanup(obddev);
1097                 GOTO(err_svc, rc = -ENOMEM);
1098         }
1099
1100         for (i = 0; i < MDS_NUM_THREADS; i++) {
1101                 char name[32];
1102                 sprintf(name, "lustre_MDS_%02d", i);
1103                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1104                 if (rc) {
1105                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1106                         GOTO(err_thread, rc);
1107                 }
1108         }
1109
1110         rc = mds_recover(obddev);
1111         if (rc)
1112                 GOTO(err_thread, rc);
1113
1114         mds_destroy_export = mds_client_free;
1115
1116         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1117                            "mds_ldlm_client", &obddev->obd_ldlm_client);
1118
1119         RETURN(0);
1120
1121 err_thread:
1122         ptlrpc_stop_all_threads(mds->mds_service);
1123 err_svc:
1124         ptlrpc_unregister_service(mds->mds_service);
1125 err_fs:
1126         mds_fs_cleanup(obddev);
1127 err_put:
1128         unlock_kernel();
1129         mntput(mds->mds_vfsmnt);
1130         mds->mds_sb = 0;
1131         lock_kernel();
1132 err_kfree:
1133         kfree(mds->mds_fstype);
1134 err_dec:
1135         MOD_DEC_USE_COUNT;
1136         return rc;
1137 }
1138
1139 static int mds_cleanup(struct obd_device *obddev)
1140 {
1141         struct super_block *sb;
1142         struct mds_obd *mds = &obddev->u.mds;
1143         struct obd_run_ctxt saved;
1144         ENTRY;
1145
1146         ptlrpc_stop_all_threads(mds->mds_service);
1147         ptlrpc_unregister_service(mds->mds_service);
1148
1149         sb = mds->mds_sb;
1150         if (!mds->mds_sb)
1151                 RETURN(0);
1152
1153         push_ctxt(&saved, &mds->mds_ctxt);
1154         mds_update_server_data(mds);
1155
1156         if (mds->mds_rcvd_filp) {
1157                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1158                 mds->mds_rcvd_filp = NULL;
1159
1160                 if (rc)
1161                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1162         }
1163         pop_ctxt(&saved);
1164
1165         unlock_kernel();
1166         mntput(mds->mds_vfsmnt);
1167         mds->mds_sb = 0;
1168         kfree(mds->mds_fstype);
1169
1170         ldlm_namespace_free(obddev->obd_namespace);
1171
1172         lock_kernel();
1173 #ifdef CONFIG_DEV_RDONLY
1174         dev_clear_rdonly(2);
1175 #endif
1176         mds_fs_cleanup(obddev);
1177
1178         MOD_DEC_USE_COUNT;
1179         RETURN(0);
1180 }
1181
1182 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1183                               ldlm_mode_t mode, void *data)
1184 {
1185         struct ptlrpc_request *req = req_cookie;
1186         int rc = 0;
1187         ENTRY;
1188
1189         if (!req_cookie)
1190                 RETURN(0);
1191
1192         if (req->rq_reqmsg->bufcount > 1) {
1193                 /* an intent needs to be considered */
1194                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1195                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1196                 struct mds_body *mds_rep;
1197                 struct ldlm_reply *rep;
1198                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1199                 int rc, size[3] = {sizeof(struct ldlm_reply),
1200                                                   sizeof(struct mds_body),
1201                                                   mds->mds_max_mdsize};
1202
1203                 it->opc = NTOH__u64(it->opc);
1204
1205                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1206                            ldlm_it2str(it->opc));
1207
1208                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1209                                      &req->rq_repmsg);
1210                 if (rc) {
1211                         rc = req->rq_status = -ENOMEM;
1212                         RETURN(rc);
1213                 }
1214
1215                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1216                 rep->lock_policy_res1 = 1;
1217
1218                 /* execute policy */
1219                 switch ((long)it->opc) {
1220                 case IT_CREAT|IT_OPEN:
1221                         rc = mds_reint(2, req);
1222                         if (rc || (req->rq_status != 0 &&
1223                                    req->rq_status != -EEXIST)) {
1224                                 rep->lock_policy_res2 = req->rq_status;
1225                                 RETURN(ELDLM_LOCK_ABORTED);
1226                         }
1227                         break;
1228                 case IT_CREAT:
1229                 case IT_LINK:
1230                 case IT_MKDIR:
1231                 case IT_MKNOD:
1232                 case IT_RENAME2:
1233                 case IT_RMDIR:
1234                 case IT_SYMLINK:
1235                 case IT_UNLINK:
1236                         rc = mds_reint(2, req);
1237                         if (rc || (req->rq_status != 0 &&
1238                                    req->rq_status != -EISDIR &&
1239                                    req->rq_status != -ENOTDIR)) {
1240                                 rep->lock_policy_res2 = req->rq_status;
1241                                 RETURN(ELDLM_LOCK_ABORTED);
1242                         }
1243                         break;
1244                 case IT_GETATTR:
1245                 case IT_LOOKUP:
1246                 case IT_OPEN:
1247                 case IT_READDIR:
1248                 case IT_READLINK:
1249                 case IT_RENAME:
1250                 case IT_SETATTR:
1251                         rc = mds_getattr_name(2, req);
1252                         /* FIXME: we need to sit down and decide on who should
1253                          * set req->rq_status, who should return negative and
1254                          * positive return values, and what they all mean. */
1255                         if (rc || req->rq_status != 0) {
1256                                 rep->lock_policy_res2 = req->rq_status;
1257                                 RETURN(ELDLM_LOCK_ABORTED);
1258                         }
1259                         break;
1260                 case IT_READDIR|IT_OPEN:
1261                         LBUG();
1262                         break;
1263                 default:
1264                         CERROR("Unhandled intent\n");
1265                         LBUG();
1266                 }
1267
1268                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1269                     it->opc == IT_RENAME || it->opc == IT_RENAME2)
1270                         RETURN(ELDLM_LOCK_ABORTED);
1271
1272                 rep->lock_policy_res2 = req->rq_status;
1273                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1274
1275                 /* If the client is about to open a file that doesn't have an MD
1276                  * stripe record, it's going to need a write lock. */
1277                 if (it->opc & IT_OPEN) {
1278                         struct lov_mds_md *md =
1279                                 lustre_msg_buf(req->rq_repmsg, 2);
1280                         if (md->lmd_easize == 0) {
1281                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1282                                            " lock");
1283                                 lock->l_req_mode = LCK_PW;
1284                         }
1285                 }
1286
1287                 /* Give the client a lock on the child object, instead of the
1288                  * parent that it requested. */
1289                 new_resid[0] = NTOH__u32(mds_rep->ino);
1290                 if (new_resid[0] == 0)
1291                         LBUG();
1292                 old_res = lock->l_resource->lr_name[0];
1293
1294                 ldlm_lock_change_resource(lock, new_resid);
1295                 if (lock->l_resource == NULL) {
1296                         LBUG();
1297                         RETURN(-ENOMEM);
1298                 }
1299                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1300                            (long)old_res);
1301                 RETURN(ELDLM_LOCK_CHANGED);
1302         } else {
1303                 int size = sizeof(struct ldlm_reply);
1304                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1305                                      &req->rq_repmsg);
1306                 if (rc) {
1307                         LBUG();
1308                         RETURN(-ENOMEM);
1309                 }
1310         }
1311         RETURN(rc);
1312 }
1313
1314
1315 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1316                          int len, void *karg, void *uarg);
1317
1318 /* use obd ops to offer management infrastructure */
1319 static struct obd_ops mds_obd_ops = {
1320         o_connect:     mds_connect,
1321         o_disconnect:  mds_disconnect,
1322         o_setup:       mds_setup,
1323         o_cleanup:     mds_cleanup,
1324         o_iocontrol:   mds_iocontrol
1325 };
1326
1327 static int __init mds_init(void)
1328 {
1329         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1330         ldlm_register_intent(ldlm_intent_policy);
1331         return 0;
1332 }
1333
1334 static void __exit mds_exit(void)
1335 {
1336         ldlm_unregister_intent();
1337         class_unregister_type(LUSTRE_MDS_NAME);
1338 }
1339
1340 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1341 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1342 MODULE_LICENSE("GPL");
1343
1344 module_init(mds_init);
1345 module_exit(mds_exit);