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
  36 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
  37                            uuid_t *uuidarray);
  38 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
  39 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  40                                 struct ptlrpc_request *req);
  41 static int mds_cleanup(struct obd_device * obddev);
  42
  43 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  44 {
  45         return &req->rq_export->exp_obd->u.mds;
  46 }
  47
  48 static int mds_bulk_timeout(void *data)
  49 {
  50         struct ptlrpc_bulk_desc *desc = data;
  51
  52         ENTRY;
  53         CERROR("(not yet) starting recovery of client %p\n", desc->b_client);
  54         RETURN(1);
  55 }
  56
  57 /* Assumes caller has already pushed into the kernel filesystem context */
  58 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  59                         __u64 offset)
  60 {
  61         int rc = 0;
  62         struct mds_obd *mds = mds_req2mds(req);
  63         struct ptlrpc_bulk_desc *desc;
  64         struct ptlrpc_bulk_page *bulk;
  65         struct l_wait_info lwi;
  66         char *buf;
  67         ENTRY;
  68
  69         desc = ptlrpc_prep_bulk(req->rq_connection);
  70         if (desc == NULL)
  71                 GOTO(out, rc = -ENOMEM);
  72
  73         bulk = ptlrpc_prep_bulk_page(desc);
  74         if (bulk == NULL)
  75                 GOTO(cleanup_bulk, rc = -ENOMEM);
  76
  77         OBD_ALLOC(buf, PAGE_SIZE);
  78         if (buf == NULL)
  79                 GOTO(cleanup_bulk, rc = -ENOMEM);
  80
  81         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  82
  83         if (rc != PAGE_SIZE)
  84                 GOTO(cleanup_buf, rc = -EIO);
  85
  86         bulk->b_xid = req->rq_xid;
  87         bulk->b_buf = buf;
  88         bulk->b_buflen = PAGE_SIZE;
  89         desc->b_portal = MDS_BULK_PORTAL;
  90
  91         rc = ptlrpc_send_bulk(desc);
  92         if (rc)
  93                 GOTO(cleanup_buf, rc);
  94
  95         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  96                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  97                        OBD_FAIL_MDS_SENDPAGE, rc);
  98                 ptlrpc_abort_bulk(desc);
  99                 GOTO(cleanup_buf, rc);
 100         }
 101
 102         lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
 103         rc = l_wait_event(desc->b_waitq, desc->b_flags & PTL_BULK_FL_SENT, &lwi);
 104         if (rc) {
 105                 if (rc != -ETIMEDOUT)
 106                         LBUG();
 107                 GOTO(cleanup_buf, rc);
 108         }
 109
 110         EXIT;
 111  cleanup_buf:
 112         OBD_FREE(buf, PAGE_SIZE);
 113  cleanup_bulk:
 114         ptlrpc_free_bulk(desc);
 115  out:
 116         return rc;
 117 }
 118
 119 /*
 120  * Look up a named entry in a directory, and get an LDLM lock on it.
 121  * 'dir' is a inode for which an LDLM lock has already been taken.
 122  *
 123  * If we do not need an exclusive or write lock on this entry (e.g.
 124  * a read lock for attribute lookup only) then we do not hold the
 125  * directory on return.  It is up to the caller to know what type
 126  * of lock it is getting, and clean up appropriately.
 127  */
 128 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 129                                       struct dentry *dir, struct vfsmount **mnt,
 130                                       char *name, int namelen, int lock_mode,
 131                                       struct lustre_handle *lockh,
 132                                       int dir_lock_mode)
 133 {
 134         struct dentry *dchild;
 135         int flags, rc;
 136         __u64 res_id[3] = {0};
 137         ENTRY;
 138
 139         down(&dir->d_inode->i_sem);
 140         dchild = lookup_one_len(name, dir, namelen);
 141         if (IS_ERR(dchild)) {
 142                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 143                 up(&dir->d_inode->i_sem);
 144                 LBUG();
 145                 RETURN(dchild);
 146         }
 147         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 148                 up(&dir->d_inode->i_sem);
 149                 ldlm_lock_decref(lockh, dir_lock_mode);
 150         }
 151
 152         if (lock_mode == 0 || !dchild->d_inode)
 153                 RETURN(dchild);
 154
 155         res_id[0] = dchild->d_inode->i_ino;
 156         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 157                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 158                                    &flags, ldlm_completion_ast,
 159                                    mds_blocking_ast, NULL, 0, lockh);
 160         if (rc != ELDLM_OK) {
 161                 l_dput(dchild);
 162                 up(&dir->d_inode->i_sem);
 163                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 164         }
 165
 166         RETURN(dchild);
 167 }
 168
 169 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 170                                      struct vfsmount **mnt, int lock_mode,
 171                                      struct lustre_handle *lockh)
 172 {
 173         struct mds_obd *mds = &obd->u.mds;
 174         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 175         int flags, rc;
 176         __u64 res_id[3] = {0};
 177         ENTRY;
 178
 179         if (IS_ERR(de))
 180                 RETURN(de);
 181
 182         res_id[0] = de->d_inode->i_ino;
 183         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 184                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 185                                    &flags, ldlm_completion_ast,
 186                                    mds_blocking_ast, NULL, 0, lockh);
 187         if (rc != ELDLM_OK) {
 188                 l_dput(de);
 189                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 190         }
 191
 192         RETURN(retval);
 193 }
 194
 195 /* Look up an entry by inode number. */
 196 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 197                               struct vfsmount **mnt)
 198 {
 199         /* stolen from NFS */
 200         struct super_block *sb = mds->mds_sb;
 201         unsigned long ino = fid->id;
 202         __u32 generation = fid->generation;
 203         struct inode *inode;
 204         struct list_head *lp;
 205         struct dentry *result;
 206
 207         if (ino == 0)
 208                 RETURN(ERR_PTR(-ESTALE));
 209
 210         inode = iget(sb, ino);
 211         if (inode == NULL)
 212                 RETURN(ERR_PTR(-ENOMEM));
 213
 214         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 215
 216         if (is_bad_inode(inode) ||
 217             (generation && inode->i_generation != generation)) {
 218                 /* we didn't find the right inode.. */
 219                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 220                        inode->i_ino, inode->i_nlink,
 221                        atomic_read(&inode->i_count), inode->i_generation,
 222                        generation);
 223                 LBUG();
 224                 iput(inode);
 225                 RETURN(ERR_PTR(-ESTALE));
 226         }
 227
 228         /* now to find a dentry. If possible, get a well-connected one */
 229         if (mnt)
 230                 *mnt = mds->mds_vfsmnt;
 231         spin_lock(&dcache_lock);
 232         list_for_each(lp, &inode->i_dentry) {
 233                 result = list_entry(lp, struct dentry, d_alias);
 234                 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 235                         dget_locked(result);
 236                         result->d_vfs_flags |= DCACHE_REFERENCED;
 237                         spin_unlock(&dcache_lock);
 238                         iput(inode);
 239                         if (mnt)
 240                                 mntget(*mnt);
 241                         return result;
 242                 }
 243         }
 244         spin_unlock(&dcache_lock);
 245         result = d_alloc_root(inode);
 246         if (result == NULL) {
 247                 iput(inode);
 248                 return ERR_PTR(-ENOMEM);
 249         }
 250         if (mnt)
 251                 mntget(*mnt);
 252         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 253         return result;
 254 }
 255
 256 /* Establish a connection to the MDS.
 257  *
 258  * This will set up an export structure for the client to hold state data
 259  * about that client, like open files, the last operation number it did
 260  * on the server, etc.
 261  */
 262 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 263                        char *cluuid)
 264 {
 265         struct obd_export *exp;
 266         struct mds_client_data *mcd;
 267         struct list_head *p;
 268         int rc;
 269         ENTRY;
 270
 271         if (!conn || !obd || !cluuid)
 272                 RETURN(-EINVAL);
 273
 274         MOD_INC_USE_COUNT;
 275
 276         list_for_each(p, &obd->obd_exports) {
 277                 exp = list_entry(p, struct obd_export, exp_chain);
 278                 mcd = exp->exp_mds_data.med_mcd;
 279                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 280                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 281                                cluuid, exp);
 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                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 289                                (long long)conn->addr, (long long)conn->cookie);
 290                         RETURN(0);
 291                 }
 292         }
 293         rc = class_connect(conn, obd, cluuid);
 294         if (rc)
 295                 GOTO(out_dec, rc);
 296         exp = class_conn2export(conn);
 297         LASSERT(exp);
 298
 299         OBD_ALLOC(mcd, sizeof(*mcd));
 300         if (!mcd) {
 301                 CERROR("mds: out of memory for client data\n");
 302                 GOTO(out_export, rc = -ENOMEM);
 303         }
 304         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 305         exp->exp_mds_data.med_mcd = mcd;
 306         rc = mds_client_add(&exp->exp_mds_data, -1);
 307         if (rc)
 308                 GOTO(out_mdc, rc);
 309
 310         RETURN(0);
 311
 312 out_mdc:
 313         OBD_FREE(mcd, sizeof(*mcd));
 314 out_export:
 315         class_disconnect(conn);
 316 out_dec:
 317         MOD_DEC_USE_COUNT;
 318
 319         return rc;
 320 }
 321
 322 static int mds_disconnect(struct lustre_handle *conn)
 323 {
 324         struct obd_export *exp;
 325         int rc;
 326
 327         exp = class_conn2export(conn);
 328         if (!exp)
 329                 RETURN(-EINVAL);
 330
 331         rc = mds_client_free(&exp->exp_mds_data);
 332         if (rc)
 333                 CERROR("error freeing client data: rc = %d\n", rc);
 334
 335         rc = class_disconnect(conn);
 336         if (!rc)
 337                 MOD_DEC_USE_COUNT;
 338
 339         return rc;
 340 }
 341
 342 static int mds_getstatus(struct ptlrpc_request *req)
 343 {
 344         struct mds_obd *mds = mds_req2mds(req);
 345         struct mds_body *body;
 346         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 347         int rc, size = sizeof(*body);
 348         ENTRY;
 349
 350         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 351         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 352                 CERROR("mds: out of memory for message: size=%d\n", size);
 353                 req->rq_status = -ENOMEM;
 354                 RETURN(0);
 355         }
 356
 357         body = lustre_msg_buf(req->rq_reqmsg, 0);
 358         mds_unpack_body(body);
 359
 360         /* Anything we need to do here with the client's trans no or so? */
 361         body = lustre_msg_buf(req->rq_repmsg, 0);
 362         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 363
 364         LASSERT(med->med_mcd);
 365
 366         /* mcd_last_xid is is stored in little endian on the disk and
 367            mds_pack_rep_body converts it to network order */
 368         req->rq_repmsg->last_xid = le32_to_cpu(med->med_mcd->mcd_last_xid);
 369         mds_pack_rep_body(req);
 370         RETURN(0);
 371 }
 372
 373 static int mds_getlovinfo(struct ptlrpc_request *req)
 374 {
 375         struct mds_obd *mds = mds_req2mds(req);
 376         struct mds_status_req *streq;
 377         struct lov_desc *desc;
 378         int tgt_count;
 379         int rc, size[2] = {sizeof(*desc)};
 380         ENTRY;
 381
 382         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 383         streq->flags = NTOH__u32(streq->flags);
 384         streq->repbuf = NTOH__u32(streq->repbuf);
 385         size[1] = streq->repbuf;
 386
 387         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 388         if (rc) {
 389                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 390                 req->rq_status = -ENOMEM;
 391                 RETURN(0);
 392         }
 393
 394         desc = lustre_msg_buf(req->rq_repmsg, 0);
 395         rc = mds_get_lovdesc(req->rq_obd, desc);
 396         if (rc != 0 ) {
 397                 CERROR("mds_get_lovdesc error %d", rc);
 398                 req->rq_status = rc;
 399                 RETURN(0);
 400         }
 401
 402         tgt_count = NTOH__u32(desc->ld_tgt_count);
 403         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 404                 CERROR("too many targets, enlarge client buffers\n");
 405                 req->rq_status = -ENOSPC;
 406                 RETURN(0);
 407         }
 408
 409         mds->mds_max_mdsize = sizeof(struct lov_stripe_md) +
 410                 tgt_count * sizeof(struct lov_object_id);
 411         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 412                              lustre_msg_buf(req->rq_repmsg, 1));
 413         if (rc) {
 414                 CERROR("get_lovtgts error %d\n", rc);
 415                 req->rq_status = rc;
 416                 RETURN(0);
 417         }
 418         RETURN(0);
 419 }
 420
 421 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 422                      void *data, __u32 data_len)
 423 {
 424         struct lustre_handle lockh;
 425         ENTRY;
 426
 427         ldlm_lock2handle(lock, &lockh);
 428         if (ldlm_cli_cancel(&lockh) < 0)
 429                 LBUG();
 430         RETURN(0);
 431 }
 432
 433 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 434                                 struct ptlrpc_request *req,
 435                                 int request_off, int reply_off)
 436 {
 437         struct mds_body *request_body, *body;
 438         struct inode *inode = dentry->d_inode;
 439         int rc;
 440         ENTRY;
 441
 442         if (inode == NULL)
 443                 RETURN(-ENOENT);
 444
 445         /* Did the client request the link name? */
 446         request_body = lustre_msg_buf(req->rq_reqmsg, request_off);
 447         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 448         if ((body->valid & OBD_MD_LINKNAME) && S_ISLNK(inode->i_mode)) {
 449                 char *tmp = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 450
 451                 rc = inode->i_op->readlink(dentry, tmp, req->rq_repmsg->
 452                                            buflens[reply_off + 1]);
 453                 if (rc < 0) {
 454                         CERROR("readlink failed: %d\n", rc);
 455                         RETURN(rc);
 456                 }
 457
 458                 body->valid |= OBD_MD_LINKNAME;
 459         }
 460
 461         mds_pack_inode2fid(&body->fid1, inode);
 462         mds_pack_inode2body(body, inode);
 463         if (S_ISREG(inode->i_mode)) {
 464                 struct lov_stripe_md *md;
 465
 466                 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 467                 md->lmd_easize = mds->mds_max_mdsize;
 468                 rc = mds_fs_get_md(mds, inode, md);
 469
 470                 if (rc < 0) {
 471                         if (rc == -ENODATA)
 472                                 RETURN(0);
 473                         CERROR("mds_fs_get_md failed: %d\n", rc);
 474                         RETURN(rc);
 475                 }
 476                 body->valid |= OBD_MD_FLEASIZE;
 477         }
 478         RETURN(0);
 479 }
 480
 481 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 482 {
 483         struct mds_obd *mds = mds_req2mds(req);
 484         struct obd_device *obd = req->rq_export->exp_obd;
 485         struct obd_run_ctxt saved;
 486         struct mds_body *body;
 487         struct dentry *de = NULL, *dchild = NULL;
 488         struct inode *dir;
 489         struct lustre_handle lockh;
 490         char *name;
 491         int namelen, flags, lock_mode, rc = 0, old_offset = offset;
 492         __u64 res_id[3] = {0, 0, 0};
 493         ENTRY;
 494
 495         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 496
 497         if (req->rq_reqmsg->bufcount <= offset + 1) {
 498                 LBUG();
 499                 GOTO(out_pre_de, rc = -EINVAL);
 500         }
 501
 502         body = lustre_msg_buf(req->rq_reqmsg, offset);
 503         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 504         namelen = req->rq_reqmsg->buflens[offset + 1];
 505         /* requests were at offset 2, replies go back at 1 */
 506         if (offset)
 507                 offset = 1;
 508
 509         push_ctxt(&saved, &mds->mds_ctxt);
 510         de = mds_fid2dentry(mds, &body->fid1, NULL);
 511         if (IS_ERR(de)) {
 512                 LBUG();
 513                 GOTO(out_pre_de, rc = -ESTALE);
 514         }
 515
 516         dir = de->d_inode;
 517         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 518
 519         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 520         res_id[0] = dir->i_ino;
 521
 522         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 523                              NULL, 0, lock_mode, &lockh);
 524         if (rc == 0) {
 525                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 526                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 527                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 528                                       &flags, ldlm_completion_ast,
 529                                       mds_blocking_ast, NULL, 0, &lockh);
 530                 if (rc != ELDLM_OK) {
 531                         CERROR("lock enqueue: err: %d\n", rc);
 532                         GOTO(out_create_de, rc = -EIO);
 533                 }
 534         }
 535         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 536
 537         down(&dir->i_sem);
 538         dchild = lookup_one_len(name, de, namelen - 1);
 539         if (IS_ERR(dchild)) {
 540                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 541                 up(&dir->i_sem);
 542                 LBUG();
 543                 GOTO(out_create_dchild, rc = -ESTALE);
 544         }
 545
 546         rc = mds_getattr_internal(mds, dchild, req, old_offset, offset);
 547
 548         EXIT;
 549 out_create_dchild:
 550         l_dput(dchild);
 551         up(&dir->i_sem);
 552         ldlm_lock_decref(&lockh, lock_mode);
 553 out_create_de:
 554         l_dput(de);
 555 out_pre_de:
 556         req->rq_status = rc;
 557         pop_ctxt(&saved);
 558         return 0;
 559 }
 560
 561 static int mds_getattr(int offset, struct ptlrpc_request *req)
 562 {
 563         struct mds_obd *mds = mds_req2mds(req);
 564         struct obd_run_ctxt saved;
 565         struct dentry *de;
 566         struct inode *inode;
 567         struct mds_body *body;
 568         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 569         ENTRY;
 570
 571         body = lustre_msg_buf(req->rq_reqmsg, offset);
 572         push_ctxt(&saved, &mds->mds_ctxt);
 573         de = mds_fid2dentry(mds, &body->fid1, NULL);
 574         if (IS_ERR(de))
 575                 GOTO(out_pop, rc = -ENOENT);
 576
 577         inode = de->d_inode;
 578         if (S_ISREG(body->fid1.f_type)) {
 579                 bufcount = 2;
 580                 size[1] = mds->mds_max_mdsize;
 581         } else if (body->valid & OBD_MD_LINKNAME) {
 582                 bufcount = 2;
 583                 size[1] = inode->i_size;
 584         }
 585
 586         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 587                              &req->rq_repmsg);
 588         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 589                 CERROR("mds: out of memory\n");
 590                 GOTO(out, rc);
 591         }
 592
 593         rc = mds_getattr_internal(mds, de, req, offset, 0);
 594
 595 out:
 596         l_dput(de);
 597 out_pop:
 598         pop_ctxt(&saved);
 599         req->rq_status = rc;
 600         RETURN(0);
 601 }
 602
 603 static int mds_statfs(struct ptlrpc_request *req)
 604 {
 605         struct mds_obd *mds = mds_req2mds(req);
 606         struct obd_statfs *osfs;
 607         struct statfs sfs;
 608         int rc, size = sizeof(*osfs);
 609         ENTRY;
 610
 611         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 612                              &req->rq_repmsg);
 613         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 614                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 615                 GOTO(out, rc);
 616         }
 617
 618         rc = vfs_statfs(mds->mds_sb, &sfs);
 619         if (rc) {
 620                 CERROR("mds: statfs failed: rc %d\n", rc);
 621                 GOTO(out, rc);
 622         }
 623         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 624         memset(osfs, 0, size);
 625         obd_statfs_pack(osfs, &sfs);
 626
 627 out:
 628         req->rq_status = rc;
 629         RETURN(0);
 630 }
 631
 632 static int mds_open(struct ptlrpc_request *req)
 633 {
 634         struct dentry *de;
 635         struct mds_body *body;
 636         struct file *file;
 637         struct vfsmount *mnt;
 638         struct mds_obd *mds = mds_req2mds(req);
 639         struct mds_export_data *med;
 640         __u32 flags;
 641         struct list_head *tmp;
 642         struct mds_file_data *mfd;
 643         int rc, size = sizeof(*body);
 644         ENTRY;
 645
 646         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 647         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 648                 CERROR("mds: out of memory\n");
 649                 req->rq_status = -ENOMEM;
 650                 RETURN(0);
 651         }
 652
 653         body = lustre_msg_buf(req->rq_reqmsg, 0);
 654
 655         /* was this animal open already? */
 656         /* XXX we should only check on re-open, or do a refcount... */
 657         med = &req->rq_export->exp_mds_data;
 658         list_for_each(tmp, &med->med_open_head) {
 659                 struct mds_file_data *fd;
 660                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 661                 if (body->extra == fd->mfd_clientfd &&
 662                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 663                         CERROR("Re opening %Ld\n", body->fid1.id);
 664                         RETURN(0);
 665                 }
 666         }
 667
 668         OBD_ALLOC(mfd, sizeof(*mfd));
 669         if (!mfd) {
 670                 CERROR("mds: out of memory\n");
 671                 req->rq_status = -ENOMEM;
 672                 RETURN(0);
 673         }
 674
 675         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 676         if (IS_ERR(de)) {
 677                 req->rq_status = -ENOENT;
 678                 RETURN(0);
 679         }
 680
 681         /* check if this inode has seen a delayed object creation */
 682         if (req->rq_reqmsg->bufcount > 1) {
 683                 void *handle;
 684                 struct lov_stripe_md *md;
 685                 struct inode *inode = de->d_inode;
 686                 int rc;
 687
 688                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 689
 690                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 691                 if (!handle) {
 692                         req->rq_status = -ENOMEM;
 693                         RETURN(0);
 694                 }
 695
 696                 /* XXX error handling */
 697                 rc = mds_fs_set_md(mds, inode, handle, md);
 698                 if (!rc) {
 699                         struct obd_run_ctxt saved;
 700                         push_ctxt(&saved, &mds->mds_ctxt);
 701                         rc = mds_update_last_rcvd(mds, handle, req);
 702                         pop_ctxt(&saved);
 703                 } else {
 704                         req->rq_status = rc;
 705                         RETURN(0);
 706                 }
 707                 /* FIXME: need to return last_rcvd, last_committed */
 708
 709                 /* FIXME: keep rc intact */
 710                 rc = mds_fs_commit(mds, de->d_inode, handle);
 711                 if (rc) {
 712                         req->rq_status = rc;
 713                         RETURN(0);
 714                 }
 715         }
 716
 717         flags = body->flags;
 718         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 719         if (!file || IS_ERR(file)) {
 720                 req->rq_status = -EINVAL;
 721                 OBD_FREE(mfd, sizeof(*mfd));
 722                 RETURN(0);
 723         }
 724
 725         file->private_data = mfd;
 726         mfd->mfd_file = file;
 727         mfd->mfd_clientfd = body->extra;
 728         list_add(&mfd->mfd_list, &med->med_open_head);
 729
 730         body = lustre_msg_buf(req->rq_repmsg, 0);
 731         /* FIXME: need to have cookies involved here */
 732         body->extra = (__u64) (unsigned long)file;
 733         RETURN(0);
 734 }
 735
 736 static int mds_close(struct ptlrpc_request *req)
 737 {
 738         struct dentry *de;
 739         struct mds_body *body;
 740         struct file *file;
 741         struct mds_obd *mds = mds_req2mds(req);
 742         struct vfsmount *mnt;
 743         struct mds_file_data *mfd;
 744         int rc;
 745         ENTRY;
 746
 747         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 748         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 749                 CERROR("mds: out of memory\n");
 750                 req->rq_status = -ENOMEM;
 751                 RETURN(0);
 752         }
 753
 754         body = lustre_msg_buf(req->rq_reqmsg, 0);
 755         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 756         if (IS_ERR(de)) {
 757                 req->rq_status = -ENOENT;
 758                 RETURN(0);
 759         }
 760
 761         /* FIXME: need to have cookies involved here */
 762         file = (struct file *)(unsigned long)body->extra;
 763         if (!file->f_dentry)
 764                 LBUG();
 765         mfd = (struct mds_file_data *)file->private_data;
 766         list_del(&mfd->mfd_list);
 767         OBD_FREE(mfd, sizeof(*mfd));
 768
 769         req->rq_status = filp_close(file, 0);
 770         l_dput(de);
 771         mntput(mnt);
 772
 773         RETURN(0);
 774 }
 775
 776 static int mds_readpage(struct ptlrpc_request *req)
 777 {
 778         struct mds_obd *mds = mds_req2mds(req);
 779         struct vfsmount *mnt;
 780         struct dentry *de;
 781         struct file *file;
 782         struct mds_body *body;
 783         struct obd_run_ctxt saved;
 784         int rc, size = sizeof(*body);
 785         ENTRY;
 786
 787         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 788         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 789                 CERROR("mds: out of memory\n");
 790                 GOTO(out, rc = -ENOMEM);
 791         }
 792
 793         body = lustre_msg_buf(req->rq_reqmsg, 0);
 794         push_ctxt(&saved, &mds->mds_ctxt);
 795         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 796         if (IS_ERR(de))
 797                 GOTO(out_pop, rc = PTR_ERR(de));
 798
 799         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 800
 801         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 802         /* note: in case of an error, dentry_open puts dentry */
 803         if (IS_ERR(file))
 804                 GOTO(out_pop, rc = PTR_ERR(file));
 805
 806         /* to make this asynchronous make sure that the handling function
 807            doesn't send a reply when this function completes. Instead a
 808            callback function would send the reply */
 809         rc = mds_sendpage(req, file, body->size);
 810
 811         filp_close(file, 0);
 812 out_pop:
 813         pop_ctxt(&saved);
 814 out:
 815         req->rq_status = rc;
 816         RETURN(0);
 817 }
 818
 819 int mds_reint(int offset, struct ptlrpc_request *req)
 820 {
 821         int rc;
 822         struct mds_update_record rec;
 823
 824         rc = mds_update_unpack(req, offset, &rec);
 825         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 826                 CERROR("invalid record\n");
 827                 req->rq_status = -EINVAL;
 828                 RETURN(0);
 829         }
 830         /* rc will be used to interrupt a for loop over multiple records */
 831         rc = mds_reint_rec(&rec, offset, req);
 832         return rc;
 833 }
 834
 835 int mds_handle(struct ptlrpc_request *req)
 836 {
 837         int rc;
 838         ENTRY;
 839
 840         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 841         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 842                 CERROR("lustre_mds: Invalid request\n");
 843                 GOTO(out, rc);
 844         }
 845
 846         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 847                 CERROR("lustre_mds: wrong packet type sent %d\n",
 848                        req->rq_reqmsg->type);
 849                 GOTO(out, rc = -EINVAL);
 850         }
 851
 852         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
 853                 GOTO(out, rc = -ENOTCONN);
 854
 855         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 856                 GOTO(out, rc = -EINVAL);
 857
 858         switch (req->rq_reqmsg->opc) {
 859         case MDS_CONNECT:
 860                 CDEBUG(D_INODE, "connect\n");
 861                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 862                 rc = target_handle_connect(req);
 863                 break;
 864
 865         case MDS_DISCONNECT:
 866                 CDEBUG(D_INODE, "disconnect\n");
 867                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 868                 rc = target_handle_disconnect(req);
 869                 goto out;
 870
 871         case MDS_GETSTATUS:
 872                 CDEBUG(D_INODE, "getstatus\n");
 873                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 874                 rc = mds_getstatus(req);
 875                 break;
 876
 877         case MDS_GETLOVINFO:
 878                 CDEBUG(D_INODE, "getlovinfo\n");
 879                 rc = mds_getlovinfo(req);
 880                 break;
 881
 882         case MDS_GETATTR:
 883                 CDEBUG(D_INODE, "getattr\n");
 884                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 885                 rc = mds_getattr(0, req);
 886                 break;
 887
 888         case MDS_STATFS:
 889                 CDEBUG(D_INODE, "statfs\n");
 890                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 891                 rc = mds_statfs(req);
 892                 break;
 893
 894         case MDS_READPAGE:
 895                 CDEBUG(D_INODE, "readpage\n");
 896                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 897                 rc = mds_readpage(req);
 898
 899                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 900                         return 0;
 901                 break;
 902
 903         case MDS_REINT: {
 904                 int size = sizeof(struct mds_body);
 905                 CDEBUG(D_INODE, "reint\n");
 906                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 907
 908                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 909                                      &req->rq_repmsg);
 910                 if (rc) {
 911                         req->rq_status = rc;
 912                         break;
 913                 }
 914                 rc = mds_reint(0, req);
 915                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 916                 break;
 917                 }
 918
 919         case MDS_OPEN:
 920                 CDEBUG(D_INODE, "open\n");
 921                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 922                 rc = mds_open(req);
 923                 break;
 924
 925         case MDS_CLOSE:
 926                 CDEBUG(D_INODE, "close\n");
 927                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 928                 rc = mds_close(req);
 929                 break;
 930
 931         case LDLM_ENQUEUE:
 932                 CDEBUG(D_INODE, "enqueue\n");
 933                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 934                 rc = ldlm_handle_enqueue(req);
 935                 if (rc)
 936                         break;
 937                 RETURN(0);
 938
 939         case LDLM_CONVERT:
 940                 CDEBUG(D_INODE, "convert\n");
 941                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 942                 rc = ldlm_handle_convert(req);
 943                 if (rc)
 944                         break;
 945                 RETURN(0);
 946
 947         case LDLM_CANCEL:
 948                 CDEBUG(D_INODE, "cancel\n");
 949                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 950                 rc = ldlm_handle_cancel(req);
 951                 if (rc)
 952                         break;
 953                 RETURN(0);
 954         case LDLM_BL_CALLBACK:
 955         case LDLM_CP_CALLBACK:
 956                 CDEBUG(D_INODE, "callback\n");
 957                 CERROR("callbacks should not happen on MDS\n");
 958                 LBUG();
 959                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 960                 break;
 961         default:
 962                 rc = ptlrpc_error(req->rq_svc, req);
 963                 RETURN(rc);
 964         }
 965
 966         EXIT;
 967
 968         if (!rc) {
 969                 struct mds_obd *mds = mds_req2mds(req);
 970                 req->rq_repmsg->last_xid = HTON__u64(mds->mds_last_rcvd);
 971                 req->rq_repmsg->last_committed =
 972                         HTON__u64(mds->mds_last_committed);
 973                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 974                        (unsigned long long)mds->mds_last_rcvd,
 975                        (unsigned long long)mds->mds_last_committed,
 976                        cpu_to_le32(req->rq_xid));
 977         }
 978  out:
 979         /* Still not 100% sure whether we should reply with the server
 980          * last_rcvd or that of this client.  I'm not sure it even makes
 981          * a difference on a per-client basis, because last_rcvd is global
 982          * and we are not supposed to allow transactions while in recovery.
 983          */
 984         if (rc) {
 985                 CERROR("mds: processing error %d\n", rc);
 986                 ptlrpc_error(req->rq_svc, req);
 987         } else {
 988                 CDEBUG(D_NET, "sending reply\n");
 989                 ptlrpc_reply(req->rq_svc, req);
 990         }
 991         return 0;
 992 }
 993
 994 /* Update the server data on disk.  This stores the new mount_count and
 995  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 996  * then the server last_rcvd value may be less than that of the clients.
 997  * This will alert us that we may need to do client recovery.
 998  *
 999  * Assumes we are already in the server filesystem context.
1000  */
1001 static
1002 int mds_update_server_data(struct mds_obd *mds)
1003 {
1004         struct mds_server_data *msd = mds->mds_server_data;
1005         struct file *filp = mds->mds_rcvd_filp;
1006         loff_t off = 0;
1007         int rc;
1008
1009         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1010         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1011
1012         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1013                (unsigned long long)mds->mds_mount_count,
1014                (unsigned long long)mds->mds_last_rcvd);
1015         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1016         if (rc != sizeof(*msd)) {
1017                 CERROR("error writing MDS server data: rc = %d\n", rc);
1018                 if (rc > 0)
1019                         RETURN(-EIO);
1020                 RETURN(rc);
1021         }
1022         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1023         if (rc)
1024                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1025
1026         return 0;
1027 }
1028
1029 /* Do recovery actions for the MDS */
1030 static int mds_recover(struct obd_device *obddev)
1031 {
1032         struct mds_obd *mds = &obddev->u.mds;
1033         struct obd_run_ctxt saved;
1034         int rc;
1035
1036         /* This happens at the end when recovery is complete */
1037         ++mds->mds_mount_count;
1038         push_ctxt(&saved, &mds->mds_ctxt);
1039         rc = mds_update_server_data(mds);
1040         pop_ctxt(&saved);
1041
1042         return rc;
1043 }
1044
1045 #define MDS_NUM_THREADS 8
1046 /* mount the file system (secretly) */
1047 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1048 {
1049         int i;
1050         struct obd_ioctl_data* data = buf;
1051         struct mds_obd *mds = &obddev->u.mds;
1052         struct vfsmount *mnt;
1053         int rc = 0;
1054         ENTRY;
1055
1056         MOD_INC_USE_COUNT;
1057 #ifdef CONFIG_DEV_RDONLY
1058         dev_clear_rdonly(2);
1059 #endif
1060         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1061                 GOTO(err_dec, rc = -EINVAL);
1062
1063         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1064
1065         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1066         if (IS_ERR(mnt)) {
1067                 rc = PTR_ERR(mnt);
1068                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1069                 GOTO(err_kfree, rc);
1070         }
1071
1072         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1073         if (!mds->mds_sb)
1074                 GOTO(err_put, rc = -ENODEV);
1075
1076         mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
1077         rc = mds_fs_setup(obddev, mnt);
1078         if (rc) {
1079                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1080                 GOTO(err_put, rc);
1081         }
1082
1083         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1084                                            MDC_REPLY_PORTAL, "self",mds_handle);
1085         if (!mds->mds_service) {
1086                 CERROR("failed to start service\n");
1087                 GOTO(err_fs, rc = -EINVAL);
1088         }
1089
1090         obddev->obd_namespace =
1091                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1092         if (obddev->obd_namespace == NULL) {
1093                 mds_cleanup(obddev);
1094                 GOTO(err_svc, rc = -ENOMEM);
1095         }
1096
1097         for (i = 0; i < MDS_NUM_THREADS; i++) {
1098                 char name[32];
1099                 sprintf(name, "lustre_MDS_%02d", i);
1100                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1101                 if (rc) {
1102                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1103                         GOTO(err_thread, rc);
1104                 }
1105         }
1106
1107         rc = mds_recover(obddev);
1108         if (rc)
1109                 GOTO(err_thread, rc);
1110
1111         RETURN(0);
1112
1113 err_thread:
1114         ptlrpc_stop_all_threads(mds->mds_service);
1115 err_svc:
1116         ptlrpc_unregister_service(mds->mds_service);
1117 err_fs:
1118         mds_fs_cleanup(obddev);
1119 err_put:
1120         unlock_kernel();
1121         mntput(mds->mds_vfsmnt);
1122         mds->mds_sb = 0;
1123         lock_kernel();
1124 err_kfree:
1125         kfree(mds->mds_fstype);
1126 err_dec:
1127         MOD_DEC_USE_COUNT;
1128         return rc;
1129 }
1130
1131 static int mds_cleanup(struct obd_device *obddev)
1132 {
1133         struct super_block *sb;
1134         struct mds_obd *mds = &obddev->u.mds;
1135         struct obd_run_ctxt saved;
1136         ENTRY;
1137
1138         ptlrpc_stop_all_threads(mds->mds_service);
1139         ptlrpc_unregister_service(mds->mds_service);
1140
1141         sb = mds->mds_sb;
1142         if (!mds->mds_sb)
1143                 RETURN(0);
1144
1145         push_ctxt(&saved, &mds->mds_ctxt);
1146         mds_update_server_data(mds);
1147
1148         if (mds->mds_rcvd_filp) {
1149                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1150                 mds->mds_rcvd_filp = NULL;
1151
1152                 if (rc)
1153                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1154         }
1155         pop_ctxt(&saved);
1156
1157         unlock_kernel();
1158         mntput(mds->mds_vfsmnt);
1159         mds->mds_sb = 0;
1160         kfree(mds->mds_fstype);
1161
1162         ldlm_namespace_free(obddev->obd_namespace);
1163
1164         lock_kernel();
1165 #ifdef CONFIG_DEV_RDONLY
1166         dev_clear_rdonly(2);
1167 #endif
1168         mds_fs_cleanup(obddev);
1169
1170         MOD_DEC_USE_COUNT;
1171         RETURN(0);
1172 }
1173
1174 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1175                               ldlm_mode_t mode, void *data)
1176 {
1177         struct ptlrpc_request *req = req_cookie;
1178         int rc = 0;
1179         ENTRY;
1180
1181         if (!req_cookie)
1182                 RETURN(0);
1183
1184         if (req->rq_reqmsg->bufcount > 1) {
1185                 /* an intent needs to be considered */
1186                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1187                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1188                 struct mds_body *mds_rep;
1189                 struct ldlm_reply *rep;
1190                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1191                 int bufcount = -1, rc, size[3] = {sizeof(struct ldlm_reply),
1192                                                   sizeof(struct mds_body),
1193                                                   mds->mds_max_mdsize};
1194
1195                 it->opc = NTOH__u64(it->opc);
1196
1197                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1198                            ldlm_it2str(it->opc));
1199
1200                 /* prepare reply */
1201                 switch((long)it->opc) {
1202                 case IT_GETATTR:
1203                         /* Note that in the negative case you may be returning
1204                          * a file and its obdo */
1205                 case IT_CREAT:
1206                 case IT_CREAT|IT_OPEN:
1207                 case IT_LINK:
1208                 case IT_LOOKUP:
1209                 case IT_MKDIR:
1210                 case IT_MKNOD:
1211                 case IT_OPEN:
1212                 case IT_READLINK:
1213                 case IT_RENAME:
1214                 case IT_RMDIR:
1215                 case IT_SETATTR:
1216                 case IT_SYMLINK:
1217                 case IT_UNLINK:
1218                         bufcount = 3;
1219                         break;
1220                 case IT_RENAME2:
1221                         bufcount = 1;
1222                         break;
1223                 default:
1224                         LBUG();
1225                 }
1226
1227                 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
1228                                      &req->rq_repmsg);
1229                 if (rc) {
1230                         rc = req->rq_status = -ENOMEM;
1231                         RETURN(rc);
1232                 }
1233
1234                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1235                 rep->lock_policy_res1 = 1;
1236
1237                 /* execute policy */
1238                 switch ((long)it->opc) {
1239                 case IT_CREAT|IT_OPEN:
1240                         rc = mds_reint(2, req);
1241                         if (rc || (req->rq_status != 0 &&
1242                                    req->rq_status != -EEXIST)) {
1243                                 rep->lock_policy_res2 = req->rq_status;
1244                                 RETURN(ELDLM_LOCK_ABORTED);
1245                         }
1246                         break;
1247                 case IT_CREAT:
1248                 case IT_LINK:
1249                 case IT_MKDIR:
1250                 case IT_MKNOD:
1251                 case IT_RENAME2:
1252                 case IT_RMDIR:
1253                 case IT_SYMLINK:
1254                 case IT_UNLINK:
1255                         rc = mds_reint(2, req);
1256                         if (rc || (req->rq_status != 0 &&
1257                                    req->rq_status != -EISDIR &&
1258                                    req->rq_status != -ENOTDIR)) {
1259                                 rep->lock_policy_res2 = req->rq_status;
1260                                 RETURN(ELDLM_LOCK_ABORTED);
1261                         }
1262                         break;
1263                 case IT_GETATTR:
1264                 case IT_LOOKUP:
1265                 case IT_OPEN:
1266                 case IT_READDIR:
1267                 case IT_READLINK:
1268                 case IT_RENAME:
1269                 case IT_SETATTR:
1270                         rc = mds_getattr_name(2, req);
1271                         /* FIXME: we need to sit down and decide on who should
1272                          * set req->rq_status, who should return negative and
1273                          * positive return values, and what they all mean. */
1274                         if (rc || req->rq_status != 0) {
1275                                 rep->lock_policy_res2 = req->rq_status;
1276                                 RETURN(ELDLM_LOCK_ABORTED);
1277                         }
1278                         break;
1279                 case IT_READDIR|IT_OPEN:
1280                         LBUG();
1281                         break;
1282                 default:
1283                         CERROR("Unhandled intent\n");
1284                         LBUG();
1285                 }
1286
1287                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1288                     it->opc == IT_RENAME || it->opc == IT_RENAME2)
1289                         RETURN(ELDLM_LOCK_ABORTED);
1290
1291                 rep->lock_policy_res2 = req->rq_status;
1292                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1293
1294                 /* If the client is about to open a file that doesn't have an MD
1295                  * stripe record, it's going to need a write lock. */
1296                 if (it->opc & IT_OPEN) {
1297                         struct lov_stripe_md *md =
1298                                 lustre_msg_buf(req->rq_repmsg, 2);
1299                         if (md->lmd_easize == 0) {
1300                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1301                                            " lock");
1302                                 lock->l_req_mode = LCK_PW;
1303                         }
1304                 }
1305
1306                 /* Give the client a lock on the child object, instead of the
1307                  * parent that it requested. */
1308                 new_resid[0] = NTOH__u32(mds_rep->ino);
1309                 if (new_resid[0] == 0)
1310                         LBUG();
1311                 old_res = lock->l_resource->lr_name[0];
1312
1313                 ldlm_lock_change_resource(lock, new_resid);
1314                 if (lock->l_resource == NULL) {
1315                         LBUG();
1316                         RETURN(-ENOMEM);
1317                 }
1318                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1319                            (long)old_res);
1320                 RETURN(ELDLM_LOCK_CHANGED);
1321         } else {
1322                 int size = sizeof(struct ldlm_reply);
1323                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1324                                      &req->rq_repmsg);
1325                 if (rc) {
1326                         LBUG();
1327                         RETURN(-ENOMEM);
1328                 }
1329         }
1330         RETURN(rc);
1331 }
1332
1333
1334 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1335                          int len, void *karg, void *uarg);
1336
1337 /* use obd ops to offer management infrastructure */
1338 static struct obd_ops mds_obd_ops = {
1339         o_connect:     mds_connect,
1340         o_disconnect:  mds_disconnect,
1341         o_setup:       mds_setup,
1342         o_cleanup:     mds_cleanup,
1343         o_iocontrol:   mds_iocontrol
1344 };
1345
1346 static int __init mds_init(void)
1347 {
1348         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1349         ldlm_register_intent(ldlm_intent_policy);
1350         return 0;
1351 }
1352
1353 static void __exit mds_exit(void)
1354 {
1355         ldlm_unregister_intent();
1356         class_unregister_type(LUSTRE_MDS_NAME);
1357 }
1358
1359 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1360 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1361 MODULE_LICENSE("GPL");
1362
1363 module_init(mds_init);
1364 module_exit(mds_exit);