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 #include <linux/random.h>
  37 #include <linux/locks.h>
  38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  39 #include <linux/buffer_head.h>
  40 #endif
  41 #include <linux/lprocfs_status.h>
  42
  43 static kmem_cache_t *mds_file_cache;
  44
  45 extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
  46                            obd_uuid_t *uuidarray);
  47 extern int mds_get_lovdesc(struct mds_obd  *obd, struct lov_desc *desc);
  48 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  49                                 struct ptlrpc_request *req);
  50 static int mds_cleanup(struct obd_device * obddev);
  51
  52 extern lprocfs_vars_t status_var_nm_1[];
  53 extern lprocfs_vars_t status_class_var[];
  54
  55 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  56 {
  57         return &req->rq_export->exp_obd->u.mds;
  58 }
  59
  60 static int mds_bulk_timeout(void *data)
  61 {
  62         struct ptlrpc_bulk_desc *desc = data;
  63
  64         ENTRY;
  65         CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
  66         RETURN(1);
  67 }
  68
  69 /* Assumes caller has already pushed into the kernel filesystem context */
  70 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  71                         __u64 offset)
  72 {
  73         int rc = 0;
  74         struct mds_obd *mds = mds_req2mds(req);
  75         struct ptlrpc_bulk_desc *desc;
  76         struct ptlrpc_bulk_page *bulk;
  77         struct l_wait_info lwi;
  78         char *buf;
  79         ENTRY;
  80
  81         desc = ptlrpc_prep_bulk(req->rq_connection);
  82         if (desc == NULL)
  83                 GOTO(out, rc = -ENOMEM);
  84
  85         bulk = ptlrpc_prep_bulk_page(desc);
  86         if (bulk == NULL)
  87                 GOTO(cleanup_bulk, rc = -ENOMEM);
  88
  89         OBD_ALLOC(buf, PAGE_SIZE);
  90         if (buf == NULL)
  91                 GOTO(cleanup_bulk, rc = -ENOMEM);
  92
  93         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  94
  95         if (rc != PAGE_SIZE)
  96                 GOTO(cleanup_buf, rc = -EIO);
  97
  98         bulk->bp_xid = req->rq_xid;
  99         bulk->bp_buf = buf;
 100         bulk->bp_buflen = PAGE_SIZE;
 101         desc->bd_portal = MDS_BULK_PORTAL;
 102
 103         rc = ptlrpc_send_bulk(desc);
 104         if (rc)
 105                 GOTO(cleanup_buf, rc);
 106
 107         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 108                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 109                        OBD_FAIL_MDS_SENDPAGE, rc);
 110                 ptlrpc_abort_bulk(desc);
 111                 GOTO(cleanup_buf, rc);
 112         }
 113
 114         lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
 115         rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
 116         if (rc) {
 117                 if (rc != -ETIMEDOUT)
 118                         LBUG();
 119                 GOTO(cleanup_buf, rc);
 120         }
 121
 122         EXIT;
 123  cleanup_buf:
 124         OBD_FREE(buf, PAGE_SIZE);
 125  cleanup_bulk:
 126         ptlrpc_free_bulk(desc);
 127  out:
 128         return rc;
 129 }
 130
 131 /*
 132  * Look up a named entry in a directory, and get an LDLM lock on it.
 133  * 'dir' is a inode for which an LDLM lock has already been taken.
 134  *
 135  * If we do not need an exclusive or write lock on this entry (e.g.
 136  * a read lock for attribute lookup only) then we do not hold the
 137  * directory on return.  It is up to the caller to know what type
 138  * of lock it is getting, and clean up appropriately.
 139  */
 140 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 141                                       struct dentry *dir, struct vfsmount **mnt,
 142                                       char *name, int namelen, int lock_mode,
 143                                       struct lustre_handle *lockh,
 144                                       int dir_lock_mode)
 145 {
 146         struct dentry *dchild;
 147         int flags = 0, rc;
 148         __u64 res_id[3] = {0};
 149         ENTRY;
 150
 151         down(&dir->d_inode->i_sem);
 152         dchild = lookup_one_len(name, dir, namelen);
 153         if (IS_ERR(dchild)) {
 154                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 155                 up(&dir->d_inode->i_sem);
 156                 LBUG();
 157                 RETURN(dchild);
 158         }
 159         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 160                 up(&dir->d_inode->i_sem);
 161                 ldlm_lock_decref(lockh, dir_lock_mode);
 162         }
 163
 164         if (lock_mode == 0 || !dchild->d_inode)
 165                 RETURN(dchild);
 166
 167         res_id[0] = dchild->d_inode->i_ino;
 168         res_id[1] = dchild->d_inode->i_generation;
 169         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 170                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 171                                    &flags, ldlm_completion_ast,
 172                                    mds_blocking_ast, NULL, 0, lockh);
 173         if (rc != ELDLM_OK) {
 174                 l_dput(dchild);
 175                 up(&dir->d_inode->i_sem);
 176                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 177         }
 178
 179         RETURN(dchild);
 180 }
 181
 182 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 183                                      struct vfsmount **mnt, int lock_mode,
 184                                      struct lustre_handle *lockh)
 185 {
 186         struct mds_obd *mds = &obd->u.mds;
 187         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 188         int flags = 0, rc;
 189         __u64 res_id[3] = {0};
 190         ENTRY;
 191
 192         if (IS_ERR(de))
 193                 RETURN(de);
 194
 195         res_id[0] = de->d_inode->i_ino;
 196         res_id[1] = de->d_inode->i_generation;
 197         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 198                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 199                                    &flags, ldlm_completion_ast,
 200                                    mds_blocking_ast, NULL, 0, lockh);
 201         if (rc != ELDLM_OK) {
 202                 l_dput(de);
 203                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 204         }
 205
 206         RETURN(retval);
 207 }
 208
 209 #ifndef DCACHE_DISCONNECTED
 210 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 211 #endif
 212
 213 /* Look up an entry by inode number. */
 214 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 215                               struct vfsmount **mnt)
 216 {
 217         /* stolen from NFS */
 218         struct super_block *sb = mds->mds_sb;
 219         unsigned long ino = fid->id;
 220         __u32 generation = fid->generation;
 221         struct inode *inode;
 222         struct list_head *lp;
 223         struct dentry *result;
 224
 225         if (ino == 0)
 226                 RETURN(ERR_PTR(-ESTALE));
 227
 228         inode = iget(sb, ino);
 229         if (inode == NULL)
 230                 RETURN(ERR_PTR(-ENOMEM));
 231
 232         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 233
 234         if (is_bad_inode(inode) ||
 235             (generation && inode->i_generation != generation)) {
 236                 /* we didn't find the right inode.. */
 237                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 238                        inode->i_ino, inode->i_nlink,
 239                        atomic_read(&inode->i_count), inode->i_generation,
 240                        generation);
 241                 LBUG();
 242                 iput(inode);
 243                 RETURN(ERR_PTR(-ESTALE));
 244         }
 245
 246         /* now to find a dentry. If possible, get a well-connected one */
 247         if (mnt)
 248                 *mnt = mds->mds_vfsmnt;
 249         spin_lock(&dcache_lock);
 250         list_for_each(lp, &inode->i_dentry) {
 251                 result = list_entry(lp, struct dentry, d_alias);
 252                 if (!(result->d_flags & DCACHE_DISCONNECTED)) {
 253                         dget_locked(result);
 254                         result->d_vfs_flags |= DCACHE_REFERENCED;
 255                         spin_unlock(&dcache_lock);
 256                         iput(inode);
 257                         if (mnt)
 258                                 mntget(*mnt);
 259                         return result;
 260                 }
 261         }
 262         spin_unlock(&dcache_lock);
 263         result = d_alloc_root(inode);
 264         if (result == NULL) {
 265                 iput(inode);
 266                 return ERR_PTR(-ENOMEM);
 267         }
 268         if (mnt)
 269                 mntget(*mnt);
 270         result->d_flags |= DCACHE_DISCONNECTED;
 271         return result;
 272 }
 273
 274 /* Establish a connection to the MDS.
 275  *
 276  * This will set up an export structure for the client to hold state data
 277  * about that client, like open files, the last operation number it did
 278  * on the server, etc.
 279  */
 280 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 281                        obd_uuid_t cluuid, struct recovd_obd *recovd,
 282                        ptlrpc_recovery_cb_t recover)
 283 {
 284         struct obd_export *exp;
 285         struct mds_export_data *med;
 286         struct mds_client_data *mcd;
 287         struct list_head *p;
 288         int rc;
 289         ENTRY;
 290
 291         if (!conn || !obd || !cluuid)
 292                 RETURN(-EINVAL);
 293
 294         MOD_INC_USE_COUNT;
 295
 296         spin_lock(&obd->obd_dev_lock);
 297         list_for_each(p, &obd->obd_exports) {
 298                 exp = list_entry(p, struct obd_export, exp_obd_chain);
 299                 mcd = exp->exp_mds_data.med_mcd;
 300                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 301                         LASSERT(exp->exp_obd == obd);
 302
 303                         if (!list_empty(&exp->exp_conn_chain)) {
 304                                 CERROR("existing uuid/export, list not empty!\n");
 305                                 spin_unlock(&obd->obd_dev_lock);
 306                                 /* XXX should we MOD_DEC_USE_COUNT; here? */
 307                                 RETURN(-EALREADY);
 308                         }
 309                         conn->addr = (__u64) (unsigned long)exp;
 310                         conn->cookie = exp->exp_cookie;
 311                         spin_unlock(&obd->obd_dev_lock);
 312                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 313                                cluuid, exp);
 314                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 315                                (long long)conn->addr, (long long)conn->cookie);
 316                         MOD_DEC_USE_COUNT;
 317                         RETURN(0);
 318                 }
 319         }
 320         spin_unlock(&obd->obd_dev_lock);
 321         /* XXX There is a small race between checking the list and adding a
 322          * new connection for the same UUID, but the real threat (list
 323          * corruption when multiple different clients connect) is solved.
 324          */
 325         rc = class_connect(conn, obd, cluuid);
 326         if (rc)
 327                 GOTO(out_dec, rc);
 328         exp = class_conn2export(conn);
 329         LASSERT(exp);
 330         med = &exp->exp_mds_data;
 331
 332         OBD_ALLOC(mcd, sizeof(*mcd));
 333         if (!mcd) {
 334                 CERROR("mds: out of memory for client data\n");
 335                 GOTO(out_export, rc = -ENOMEM);
 336         }
 337
 338         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 339         med->med_mcd = mcd;
 340
 341         INIT_LIST_HEAD(&med->med_open_head);
 342         spin_lock_init(&med->med_open_lock);
 343
 344         rc = mds_client_add(med, -1);
 345         if (rc)
 346                 GOTO(out_mdc, rc);
 347
 348         RETURN(0);
 349
 350 out_mdc:
 351         OBD_FREE(mcd, sizeof(*mcd));
 352 out_export:
 353         class_disconnect(conn);
 354 out_dec:
 355         MOD_DEC_USE_COUNT;
 356
 357         return rc;
 358 }
 359
 360 /* Call with med->med_open_lock held, please. */
 361 inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
 362 {
 363         struct file *file = mfd->mfd_file;
 364         LASSERT(file->private_data == mfd);
 365
 366         list_del(&mfd->mfd_list);
 367         mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
 368         kmem_cache_free(mds_file_cache, mfd);
 369
 370         return filp_close(file, 0);
 371 }
 372
 373 static int mds_disconnect(struct lustre_handle *conn)
 374 {
 375         struct obd_export *export = class_conn2export(conn);
 376         struct list_head *tmp, *n;
 377         struct mds_export_data *med = &export->exp_mds_data;
 378         int rc;
 379         ENTRY;
 380
 381         /*
 382          * Close any open files.
 383          */
 384         spin_lock(&med->med_open_lock);
 385         list_for_each_safe(tmp, n, &med->med_open_head) {
 386                 struct mds_file_data *mfd =
 387                         list_entry(tmp, struct mds_file_data, mfd_list);
 388                 rc = mds_close_mfd(mfd, med);
 389                 if (rc) {
 390                         /* XXX better diagnostics, with file path and stuff */
 391                         CDEBUG(D_INODE, "Error %d closing mfd %p\n", rc, mfd);
 392                 }
 393         }
 394         spin_unlock(&med->med_open_lock);
 395
 396         ldlm_cancel_locks_for_export(export);
 397         mds_client_free(export);
 398
 399         rc = class_disconnect(conn);
 400         if (!rc)
 401                 MOD_DEC_USE_COUNT;
 402
 403         RETURN(rc);
 404 }
 405
 406 /*
 407  * XXX This is NOT guaranteed to flush all transactions to disk (even though
 408  *     it is equivalent to calling sync()) because it only _starts_ the flush
 409  *     and does not wait for completion.  It's better than nothing though.
 410  *     What we really want is a mild form of fsync_dev_lockfs(), but it is
 411  *     non-standard, or enabling do_sync_supers in ext3, just for this call.
 412  */
 413 static void mds_fsync_super(struct super_block *sb)
 414 {
 415         lock_kernel();
 416         lock_super(sb);
 417         if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
 418                 sb->s_op->write_super(sb);
 419         unlock_super(sb);
 420         unlock_kernel();
 421 }
 422
 423 static int mds_getstatus(struct ptlrpc_request *req)
 424 {
 425         struct mds_obd *mds = mds_req2mds(req);
 426         struct mds_body *body;
 427         int rc, size = sizeof(*body);
 428         ENTRY;
 429
 430         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 431         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 432                 CERROR("mds: out of memory for message: size=%d\n", size);
 433                 req->rq_status = -ENOMEM;
 434                 RETURN(0);
 435         }
 436
 437         /* Flush any outstanding transactions to disk so the client will
 438          * get the latest last_committed value and can drop their local
 439          * requests if they have any.  This would be fsync_super() if it
 440          * was exported.
 441          */
 442         mds_fsync_super(mds->mds_sb);
 443
 444         body = lustre_msg_buf(req->rq_repmsg, 0);
 445         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 446
 447         /* the last_committed and last_xid fields are filled in for all
 448          * replies already - no need to do so here also.
 449          */
 450         RETURN(0);
 451 }
 452
 453 static int mds_getlovinfo(struct ptlrpc_request *req)
 454 {
 455         struct mds_obd *mds = mds_req2mds(req);
 456         struct mds_status_req *streq;
 457         struct lov_desc *desc;
 458         int tgt_count;
 459         int rc, size[2] = {sizeof(*desc)};
 460         ENTRY;
 461
 462         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 463         streq->flags = NTOH__u32(streq->flags);
 464         streq->repbuf = NTOH__u32(streq->repbuf);
 465         size[1] = streq->repbuf;
 466
 467         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 468         if (rc) {
 469                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 470                 req->rq_status = -ENOMEM;
 471                 RETURN(0);
 472         }
 473
 474         desc = lustre_msg_buf(req->rq_repmsg, 0);
 475         rc = mds_get_lovdesc(mds, desc);
 476         if (rc) {
 477                 CERROR("mds_get_lovdesc error %d", rc);
 478                 req->rq_status = rc;
 479                 RETURN(0);
 480         }
 481
 482         tgt_count = le32_to_cpu(desc->ld_tgt_count);
 483         if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
 484                 CERROR("too many targets, enlarge client buffers\n");
 485                 req->rq_status = -ENOSPC;
 486                 RETURN(0);
 487         }
 488
 489         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
 490                 tgt_count * sizeof(struct lov_object_id);
 491         rc = mds_get_lovtgts(mds, tgt_count,
 492                              lustre_msg_buf(req->rq_repmsg, 1));
 493         if (rc) {
 494                 CERROR("get_lovtgts error %d\n", rc);
 495                 req->rq_status = rc;
 496                 RETURN(0);
 497         }
 498         RETURN(0);
 499 }
 500
 501 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 502                      void *data, __u32 data_len, int flag)
 503 {
 504         int do_ast;
 505         ENTRY;
 506
 507         if (flag == LDLM_CB_CANCELING) {
 508                 /* Don't need to do anything here. */
 509                 RETURN(0);
 510         }
 511
 512         /* XXX layering violation!  -phil */
 513         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 514         lock->l_flags |= LDLM_FL_CBPENDING;
 515         do_ast = (!lock->l_readers && !lock->l_writers);
 516         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 517
 518         if (do_ast) {
 519                 struct lustre_handle lockh;
 520                 int rc;
 521
 522                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 523                 ldlm_lock2handle(lock, &lockh);
 524                 rc = ldlm_cli_cancel(&lockh);
 525                 if (rc < 0)
 526                         CERROR("ldlm_cli_cancel: %d\n", rc);
 527         } else
 528                 LDLM_DEBUG(lock, "Lock still has references, will be"
 529                            "cancelled later");
 530         RETURN(0);
 531 }
 532
 533 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 534                                 struct ptlrpc_request *req,
 535                                 struct mds_body *reqbody, int reply_off)
 536 {
 537         struct mds_body *body;
 538         struct inode *inode = dentry->d_inode;
 539         int rc;
 540         ENTRY;
 541
 542         if (inode == NULL)
 543                 RETURN(-ENOENT);
 544
 545         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 546
 547         mds_pack_inode2fid(&body->fid1, inode);
 548         mds_pack_inode2body(body, inode);
 549
 550         if (S_ISREG(inode->i_mode)) {
 551                 struct lov_mds_md *lmm;
 552
 553                 lmm = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 554                 lmm->lmm_easize = mds->mds_max_mdsize;
 555                 rc = mds_fs_get_md(mds, inode, lmm);
 556
 557                 if (rc < 0) {
 558                         if (rc == -ENODATA)
 559                                 RETURN(0);
 560                         CERROR("mds_fs_get_md failed: %d\n", rc);
 561                         RETURN(rc);
 562                 }
 563                 body->valid |= OBD_MD_FLEASIZE;
 564         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 565                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 566                 int len = req->rq_repmsg->buflens[reply_off + 1];
 567
 568                 rc = inode->i_op->readlink(dentry, symname, len);
 569                 if (rc < 0) {
 570                         CERROR("readlink failed: %d\n", rc);
 571                         RETURN(rc);
 572                 } else
 573                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 574
 575                 body->valid |= OBD_MD_LINKNAME;
 576         }
 577         RETURN(0);
 578 }
 579
 580 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 581 {
 582         struct mds_obd *mds = mds_req2mds(req);
 583         struct obd_device *obd = req->rq_export->exp_obd;
 584         struct obd_run_ctxt saved;
 585         struct mds_body *body;
 586         struct dentry *de = NULL, *dchild = NULL;
 587         struct inode *dir;
 588         struct lustre_handle lockh;
 589         char *name;
 590         int namelen, flags = 0, lock_mode, rc = 0;
 591         struct obd_ucred uc;
 592         __u64 res_id[3] = {0, 0, 0};
 593         ENTRY;
 594
 595         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 596
 597         if (req->rq_reqmsg->bufcount <= offset + 1) {
 598                 LBUG();
 599                 GOTO(out_pre_de, rc = -EINVAL);
 600         }
 601
 602         body = lustre_msg_buf(req->rq_reqmsg, offset);
 603         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 604         namelen = req->rq_reqmsg->buflens[offset + 1];
 605         /* requests were at offset 2, replies go back at 1 */
 606         if (offset)
 607                 offset = 1;
 608
 609         uc.ouc_fsuid = body->fsuid;
 610         uc.ouc_fsgid = body->fsgid;
 611         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 612         de = mds_fid2dentry(mds, &body->fid1, NULL);
 613         if (IS_ERR(de)) {
 614                 LBUG();
 615                 GOTO(out_pre_de, rc = -ESTALE);
 616         }
 617
 618         dir = de->d_inode;
 619         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 620
 621         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 622         res_id[0] = dir->i_ino;
 623         res_id[1] = dir->i_generation;
 624
 625         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 626                              NULL, 0, lock_mode, &lockh);
 627         if (rc == 0) {
 628                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 629                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 630                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 631                                       &flags, ldlm_completion_ast,
 632                                       mds_blocking_ast, NULL, 0, &lockh);
 633                 if (rc != ELDLM_OK) {
 634                         CERROR("lock enqueue: err: %d\n", rc);
 635                         GOTO(out_create_de, rc = -EIO);
 636                 }
 637         }
 638         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 639
 640         down(&dir->i_sem);
 641         dchild = lookup_one_len(name, de, namelen - 1);
 642         if (IS_ERR(dchild)) {
 643                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 644                 up(&dir->i_sem);
 645                 LBUG();
 646                 GOTO(out_create_dchild, rc = -ESTALE);
 647         }
 648
 649         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 650
 651         EXIT;
 652 out_create_dchild:
 653         l_dput(dchild);
 654         up(&dir->i_sem);
 655         ldlm_lock_decref(&lockh, lock_mode);
 656 out_create_de:
 657         l_dput(de);
 658 out_pre_de:
 659         req->rq_status = rc;
 660         pop_ctxt(&saved);
 661         return 0;
 662 }
 663
 664 static int mds_getattr(int offset, struct ptlrpc_request *req)
 665 {
 666         struct mds_obd *mds = mds_req2mds(req);
 667         struct obd_run_ctxt saved;
 668         struct dentry *de;
 669         struct inode *inode;
 670         struct mds_body *body;
 671         struct obd_ucred uc;
 672         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 673         ENTRY;
 674
 675         body = lustre_msg_buf(req->rq_reqmsg, offset);
 676         uc.ouc_fsuid = body->fsuid;
 677         uc.ouc_fsgid = body->fsgid;
 678         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 679         de = mds_fid2dentry(mds, &body->fid1, NULL);
 680         if (IS_ERR(de)) {
 681                 rc = req->rq_status = -ENOENT;
 682                 GOTO(out_pop, PTR_ERR(de));
 683         }
 684
 685         inode = de->d_inode;
 686         if (S_ISREG(body->fid1.f_type)) {
 687                 bufcount = 2;
 688                 size[1] = mds->mds_max_mdsize;
 689         } else if (body->valid & OBD_MD_LINKNAME) {
 690                 bufcount = 2;
 691                 size[1] = MIN(inode->i_size + 1, body->size);
 692                 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
 693                        inode->i_size + 1, body->size);
 694         }
 695
 696         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 697                 CERROR("failed GETATTR_PACK test\n");
 698                 req->rq_status = -ENOMEM;
 699                 GOTO(out, rc = -ENOMEM);
 700         }
 701
 702         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 703                              &req->rq_repmsg);
 704         if (rc) {
 705                 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
 706                 req->rq_status = rc;
 707                 GOTO(out, rc);
 708         }
 709
 710         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 711
 712 out:
 713         l_dput(de);
 714 out_pop:
 715         pop_ctxt(&saved);
 716         RETURN(rc);
 717 }
 718
 719 static int mds_statfs(struct ptlrpc_request *req)
 720 {
 721         struct mds_obd *mds = mds_req2mds(req);
 722         struct obd_statfs *osfs;
 723         struct statfs sfs;
 724         int rc, size = sizeof(*osfs);
 725         ENTRY;
 726
 727         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 728                              &req->rq_repmsg);
 729         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 730                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 731                 GOTO(out, rc);
 732         }
 733
 734         rc = vfs_statfs(mds->mds_sb, &sfs);
 735         if (rc) {
 736                 CERROR("mds: statfs failed: rc %d\n", rc);
 737                 GOTO(out, rc);
 738         }
 739         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 740         memset(osfs, 0, size);
 741         statfs_pack(osfs, &sfs);
 742         obd_statfs_pack(osfs, osfs);
 743
 744 out:
 745         req->rq_status = rc;
 746         RETURN(0);
 747 }
 748
 749 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
 750 {
 751         struct mds_file_data *mfd = NULL;
 752
 753         if (!handle || !handle->addr)
 754                 RETURN(NULL);
 755
 756         mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
 757         if (!kmem_cache_validate(mds_file_cache, mfd))
 758                 RETURN(NULL);
 759
 760         if (mfd->mfd_servercookie != handle->cookie)
 761                 RETURN(NULL);
 762
 763         return mfd;
 764 }
 765
 766 static int mds_store_ea(struct mds_obd *mds, struct ptlrpc_request *req,
 767                         struct mds_body *body, struct dentry *de,
 768                         struct lov_mds_md *lmm)
 769 {
 770         struct obd_run_ctxt saved;
 771         struct obd_ucred uc;
 772         void *handle;
 773         int rc, rc2;
 774
 775         uc.ouc_fsuid = body->fsuid;
 776         uc.ouc_fsgid = body->fsgid;
 777         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 778         handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 779         if (!handle)
 780                 GOTO(out_ea, rc = -ENOMEM);
 781
 782         rc = mds_fs_set_md(mds, de->d_inode, handle, lmm);
 783         if (!rc)
 784                 rc = mds_update_last_rcvd(mds, handle, req);
 785
 786         rc2 = mds_fs_commit(mds, de->d_inode, handle);
 787         if (rc2 && !rc)
 788                 rc = rc2;
 789 out_ea:
 790         pop_ctxt(&saved);
 791
 792         return rc;
 793 }
 794
 795 static int mds_open(struct ptlrpc_request *req)
 796 {
 797         struct mds_obd *mds = mds_req2mds(req);
 798         struct mds_body *body;
 799         struct mds_export_data *med;
 800         struct mds_file_data *mfd;
 801         struct dentry *de;
 802         struct file *file;
 803         struct vfsmount *mnt;
 804         __u32 flags;
 805         struct list_head *tmp;
 806         int rc, size = sizeof(*body);
 807         ENTRY;
 808
 809         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 810                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 811                 req->rq_status = -ENOMEM;
 812                 RETURN(-ENOMEM);
 813         }
 814
 815         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 816         if (rc) {
 817                 CERROR("mds: pack error: rc = %d\n", rc);
 818                 req->rq_status = rc;
 819                 RETURN(rc);
 820         }
 821
 822         body = lustre_msg_buf(req->rq_reqmsg, 0);
 823
 824         /* was this animal open already and the client lost the reply? */
 825         /* XXX need some way to detect a reopen, to avoid locked list walks */
 826         med = &req->rq_export->exp_mds_data;
 827         spin_lock(&med->med_open_lock);
 828         list_for_each(tmp, &med->med_open_head) {
 829                 mfd = list_entry(tmp, typeof(*mfd), mfd_list);
 830                 if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
 831                             sizeof(mfd->mfd_clienthandle)) &&
 832                     body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
 833                         de = mfd->mfd_file->f_dentry;
 834                         spin_unlock(&med->med_open_lock);
 835                         CERROR("Re opening "LPD64"\n", body->fid1.id);
 836                         GOTO(out_pack, rc = 0);
 837                 }
 838         }
 839         spin_unlock(&med->med_open_lock);
 840
 841         mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
 842         if (!mfd) {
 843                 CERROR("mds: out of memory\n");
 844                 req->rq_status = -ENOMEM;
 845                 RETURN(0);
 846         }
 847
 848         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 849         if (IS_ERR(de))
 850                 GOTO(out_free, rc = PTR_ERR(de));
 851
 852         /* check if this inode has seen a delayed object creation */
 853         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MDS_OPEN_HAS_EA) {
 854                 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, 1);
 855
 856                 rc = mds_store_ea(mds, req, body, de, lmm);
 857                 if (rc) {
 858                         l_dput(de);
 859                         mntput(mnt);
 860                         GOTO(out_free, rc);
 861                 }
 862         }
 863
 864         flags = body->flags;
 865         /* dentry_open does a dput(de) and mntput(mnt) on error */
 866         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 867         if (IS_ERR(file)) {
 868                 rc = PTR_ERR(file);
 869                 GOTO(out_free, 0);
 870         }
 871
 872         file->private_data = mfd;
 873         mfd->mfd_file = file;
 874         memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
 875         get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
 876         spin_lock(&med->med_open_lock);
 877         list_add(&mfd->mfd_list, &med->med_open_head);
 878         spin_unlock(&med->med_open_lock);
 879
 880 out_pack:
 881         body = lustre_msg_buf(req->rq_repmsg, 0);
 882         mds_pack_inode2fid(&body->fid1, de->d_inode);
 883         mds_pack_inode2body(body, de->d_inode);
 884         body->handle.addr = (__u64)(unsigned long)mfd;
 885         body->handle.cookie = mfd->mfd_servercookie;
 886         CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
 887                mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
 888         RETURN(0);
 889
 890 out_free:
 891         mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
 892         kmem_cache_free(mds_file_cache, mfd);
 893         req->rq_status = rc;
 894         RETURN(0);
 895 }
 896
 897 static int mds_close(struct ptlrpc_request *req)
 898 {
 899         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 900         struct mds_body *body;
 901         struct mds_file_data *mfd;
 902         int rc;
 903         ENTRY;
 904
 905         body = lustre_msg_buf(req->rq_reqmsg, 0);
 906
 907         mfd = mds_handle2mfd(&body->handle);
 908         if (!mfd) {
 909                 CERROR("no handle for file close "LPD64
 910                        ": addr "LPX64", cookie "LPX64"\n",
 911                        body->fid1.id, body->handle.addr, body->handle.cookie);
 912                 RETURN(-ESTALE);
 913         }
 914
 915         spin_lock(&med->med_open_lock);
 916         req->rq_status = mds_close_mfd(mfd, med);
 917         spin_unlock(&med->med_open_lock);
 918
 919         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 920                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
 921                 req->rq_status = -ENOMEM;
 922                 RETURN(-ENOMEM);
 923         }
 924
 925         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 926         if (rc) {
 927                 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
 928                 req->rq_status = rc;
 929         }
 930
 931         RETURN(0);
 932 }
 933
 934 static int mds_readpage(struct ptlrpc_request *req)
 935 {
 936         struct mds_obd *mds = mds_req2mds(req);
 937         struct vfsmount *mnt;
 938         struct dentry *de;
 939         struct file *file;
 940         struct mds_body *body, *repbody;
 941         struct obd_run_ctxt saved;
 942         int rc, size = sizeof(*body);
 943         struct obd_ucred uc;
 944         ENTRY;
 945
 946         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 947         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 948                 CERROR("mds: out of memory\n");
 949                 GOTO(out, rc = -ENOMEM);
 950         }
 951
 952         body = lustre_msg_buf(req->rq_reqmsg, 0);
 953         uc.ouc_fsuid = body->fsuid;
 954         uc.ouc_fsgid = body->fsgid;
 955         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 956         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 957         if (IS_ERR(de))
 958                 GOTO(out_pop, rc = PTR_ERR(de));
 959
 960         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 961
 962         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 963         /* note: in case of an error, dentry_open puts dentry */
 964         if (IS_ERR(file))
 965                 GOTO(out_pop, rc = PTR_ERR(file));
 966
 967         repbody = lustre_msg_buf(req->rq_repmsg, 0);
 968         repbody->size = file->f_dentry->d_inode->i_size;
 969         repbody->valid = OBD_MD_FLSIZE;
 970
 971         /* to make this asynchronous make sure that the handling function
 972            doesn't send a reply when this function completes. Instead a
 973            callback function would send the reply */
 974         /* note: in case of an error, dentry_open puts dentry */
 975         rc = mds_sendpage(req, file, body->size);
 976
 977         filp_close(file, 0);
 978 out_pop:
 979         pop_ctxt(&saved);
 980 out:
 981         req->rq_status = rc;
 982         RETURN(0);
 983 }
 984
 985 int mds_reint(int offset, struct ptlrpc_request *req)
 986 {
 987         int rc;
 988         struct mds_update_record rec;
 989
 990         rc = mds_update_unpack(req, offset, &rec);
 991         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 992                 CERROR("invalid record\n");
 993                 req->rq_status = -EINVAL;
 994                 RETURN(0);
 995         }
 996         /* rc will be used to interrupt a for loop over multiple records */
 997         rc = mds_reint_rec(&rec, offset, req);
 998         return rc;
 999 }
1000
1001 int mds_handle(struct ptlrpc_request *req)
1002 {
1003         int rc;
1004         ENTRY;
1005
1006         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1007         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
1008                 CERROR("lustre_mds: Invalid request\n");
1009                 GOTO(out, rc);
1010         }
1011
1012         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
1013                 GOTO(out, rc = -ENOTCONN);
1014
1015         LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
1016
1017         switch (req->rq_reqmsg->opc) {
1018         case MDS_CONNECT:
1019                 CDEBUG(D_INODE, "connect\n");
1020                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1021                 rc = target_handle_connect(req);
1022                 break;
1023
1024         case MDS_DISCONNECT:
1025                 CDEBUG(D_INODE, "disconnect\n");
1026                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1027                 rc = target_handle_disconnect(req);
1028                 goto out;
1029
1030         case MDS_GETSTATUS:
1031                 CDEBUG(D_INODE, "getstatus\n");
1032                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1033                 rc = mds_getstatus(req);
1034                 break;
1035
1036         case MDS_GETLOVINFO:
1037                 CDEBUG(D_INODE, "getlovinfo\n");
1038                 rc = mds_getlovinfo(req);
1039                 break;
1040
1041         case MDS_GETATTR:
1042                 CDEBUG(D_INODE, "getattr\n");
1043                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1044                 rc = mds_getattr(0, req);
1045                 break;
1046
1047         case MDS_STATFS:
1048                 CDEBUG(D_INODE, "statfs\n");
1049                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1050                 rc = mds_statfs(req);
1051                 break;
1052
1053         case MDS_READPAGE:
1054                 CDEBUG(D_INODE, "readpage\n");
1055                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1056                 rc = mds_readpage(req);
1057
1058                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1059                         return 0;
1060                 break;
1061
1062         case MDS_REINT: {
1063                 int size = sizeof(struct mds_body);
1064                 CDEBUG(D_INODE, "reint\n");
1065                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1066
1067                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1068                                      &req->rq_repmsg);
1069                 if (rc) {
1070                         req->rq_status = rc;
1071                         break;
1072                 }
1073                 rc = mds_reint(0, req);
1074                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1075                 break;
1076                 }
1077
1078         case MDS_OPEN:
1079                 CDEBUG(D_INODE, "open\n");
1080                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
1081                 rc = mds_open(req);
1082                 break;
1083
1084         case MDS_CLOSE:
1085                 CDEBUG(D_INODE, "close\n");
1086                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1087                 rc = mds_close(req);
1088                 break;
1089
1090         case LDLM_ENQUEUE:
1091                 CDEBUG(D_INODE, "enqueue\n");
1092                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1093                 rc = ldlm_handle_enqueue(req);
1094                 if (rc)
1095                         break;
1096                 RETURN(0);
1097         case LDLM_CONVERT:
1098                 CDEBUG(D_INODE, "convert\n");
1099                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1100                 rc = ldlm_handle_convert(req);
1101                 if (rc)
1102                         break;
1103                 RETURN(0);
1104         case LDLM_BL_CALLBACK:
1105         case LDLM_CP_CALLBACK:
1106                 CDEBUG(D_INODE, "callback\n");
1107                 CERROR("callbacks should not happen on MDS\n");
1108                 LBUG();
1109                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1110                 break;
1111         default:
1112                 rc = ptlrpc_error(req->rq_svc, req);
1113                 RETURN(rc);
1114         }
1115
1116         EXIT;
1117
1118         if (!rc) {
1119                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1120                 struct mds_obd *mds = mds_req2mds(req);
1121
1122                 req->rq_repmsg->last_xid =
1123                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1124                 req->rq_repmsg->last_committed =
1125                         HTON__u64(mds->mds_last_committed);
1126                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1127                        (unsigned long long)mds->mds_last_rcvd,
1128                        (unsigned long long)mds->mds_last_committed,
1129                        cpu_to_le32(req->rq_xid));
1130         }
1131  out:
1132         if (rc) {
1133                 CERROR("mds: processing error (opcode %d): %d\n",
1134                        req->rq_reqmsg->opc, rc);
1135                 ptlrpc_error(req->rq_svc, req);
1136         } else {
1137                 CDEBUG(D_NET, "sending reply\n");
1138                 ptlrpc_reply(req->rq_svc, req);
1139         }
1140         return 0;
1141 }
1142
1143 /* Update the server data on disk.  This stores the new mount_count and
1144  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1145  * then the server last_rcvd value may be less than that of the clients.
1146  * This will alert us that we may need to do client recovery.
1147  *
1148  * Assumes we are already in the server filesystem context.
1149  *
1150  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1151  */
1152 static
1153 int mds_update_server_data(struct mds_obd *mds)
1154 {
1155         struct mds_server_data *msd = mds->mds_server_data;
1156         struct file *filp = mds->mds_rcvd_filp;
1157         loff_t off = 0;
1158         int rc;
1159
1160         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1161         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1162
1163         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1164                (unsigned long long)mds->mds_mount_count,
1165                (unsigned long long)mds->mds_last_rcvd);
1166         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1167         if (rc != sizeof(*msd)) {
1168                 CERROR("error writing MDS server data: rc = %d\n", rc);
1169                 if (rc > 0)
1170                         RETURN(-EIO);
1171                 RETURN(rc);
1172         }
1173 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1174         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1175 #else
1176         rc = file_fsync(filp,  filp->f_dentry, 1);
1177 #endif
1178         if (rc)
1179                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1180
1181         return 0;
1182 }
1183
1184 /* Do recovery actions for the MDS */
1185 static int mds_recover(struct obd_device *obddev)
1186 {
1187         struct mds_obd *mds = &obddev->u.mds;
1188         struct obd_run_ctxt saved;
1189         int rc;
1190
1191         /* This happens at the end when recovery is complete */
1192         ++mds->mds_mount_count;
1193         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1194         rc = mds_update_server_data(mds);
1195         pop_ctxt(&saved);
1196
1197         return rc;
1198 }
1199
1200 /* mount the file system (secretly) */
1201 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1202 {
1203         struct obd_ioctl_data* data = buf;
1204         struct mds_obd *mds = &obddev->u.mds;
1205         struct vfsmount *mnt;
1206         int rc = 0;
1207         ENTRY;
1208
1209         MOD_INC_USE_COUNT;
1210 #ifdef CONFIG_DEV_RDONLY
1211         dev_clear_rdonly(2);
1212 #endif
1213         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1214                 GOTO(err_dec, rc = -EINVAL);
1215
1216         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1217
1218         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1219         if (IS_ERR(mnt)) {
1220                 rc = PTR_ERR(mnt);
1221                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1222                 GOTO(err_kfree, rc);
1223         }
1224
1225         CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, mnt);
1226         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1227         if (!mds->mds_sb)
1228                 GOTO(err_put, rc = -ENODEV);
1229
1230         spin_lock_init(&mds->mds_last_lock);
1231         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1232         rc = mds_fs_setup(obddev, mnt);
1233         if (rc) {
1234                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1235                 GOTO(err_put, rc);
1236         }
1237
1238         obddev->obd_namespace =
1239                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1240         if (obddev->obd_namespace == NULL) {
1241                 mds_cleanup(obddev);
1242                 GOTO(err_fs, rc = -ENOMEM);
1243         }
1244
1245
1246         rc = mds_recover(obddev);
1247         if (rc)
1248                 GOTO(err_fs, rc);
1249
1250         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1251                            "mds_ldlm_client", &obddev->obd_ldlm_client);
1252
1253         RETURN(0);
1254
1255 err_fs:
1256         mds_fs_cleanup(obddev);
1257 err_put:
1258         unlock_kernel();
1259         mntput(mds->mds_vfsmnt);
1260         mds->mds_sb = 0;
1261         lock_kernel();
1262 err_kfree:
1263         kfree(mds->mds_fstype);
1264 err_dec:
1265         MOD_DEC_USE_COUNT;
1266         RETURN(rc);
1267 }
1268
1269 static int mds_cleanup(struct obd_device *obddev)
1270 {
1271         struct super_block *sb;
1272         struct mds_obd *mds = &obddev->u.mds;
1273         struct obd_run_ctxt saved;
1274         ENTRY;
1275
1276         sb = mds->mds_sb;
1277         if (!mds->mds_sb)
1278                 RETURN(0);
1279
1280         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1281         mds_update_server_data(mds);
1282
1283         if (mds->mds_rcvd_filp) {
1284                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1285                 mds->mds_rcvd_filp = NULL;
1286
1287                 if (rc)
1288                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1289         }
1290         pop_ctxt(&saved);
1291
1292         unlock_kernel();
1293         mntput(mds->mds_vfsmnt);
1294         mds->mds_sb = 0;
1295         kfree(mds->mds_fstype);
1296
1297         ldlm_namespace_free(obddev->obd_namespace);
1298
1299         lock_kernel();
1300 #ifdef CONFIG_DEV_RDONLY
1301         dev_clear_rdonly(2);
1302 #endif
1303         mds_fs_cleanup(obddev);
1304
1305         MOD_DEC_USE_COUNT;
1306         RETURN(0);
1307 }
1308
1309 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1310                               ldlm_mode_t mode, int flags, void *data)
1311 {
1312         struct ptlrpc_request *req = req_cookie;
1313         int rc = 0;
1314         ENTRY;
1315
1316         if (!req_cookie)
1317                 RETURN(0);
1318
1319         if (req->rq_reqmsg->bufcount > 1) {
1320                 /* an intent needs to be considered */
1321                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1322                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1323                 struct mds_body *mds_rep;
1324                 struct ldlm_reply *rep;
1325                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1326                 int rc, size[3] = {sizeof(struct ldlm_reply),
1327                                                   sizeof(struct mds_body),
1328                                                   mds->mds_max_mdsize};
1329
1330                 it->opc = NTOH__u64(it->opc);
1331
1332                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1333                            ldlm_it2str(it->opc));
1334
1335                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1336                                      &req->rq_repmsg);
1337                 if (rc) {
1338                         rc = req->rq_status = -ENOMEM;
1339                         RETURN(rc);
1340                 }
1341
1342                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1343                 rep->lock_policy_res1 = 1;
1344
1345                 /* execute policy */
1346                 switch ((long)it->opc) {
1347                 case IT_CREAT|IT_OPEN:
1348                         rc = mds_reint(2, req);
1349                         if (rc || (req->rq_status != 0 &&
1350                                    req->rq_status != -EEXIST)) {
1351                                 rep->lock_policy_res2 = req->rq_status;
1352                                 RETURN(ELDLM_LOCK_ABORTED);
1353                         }
1354                         break;
1355                 case IT_CREAT:
1356                 case IT_MKDIR:
1357                 case IT_MKNOD:
1358                 case IT_RENAME2:
1359                 case IT_LINK2:
1360                 case IT_RMDIR:
1361                 case IT_SYMLINK:
1362                 case IT_UNLINK:
1363                         rc = mds_reint(2, req);
1364                         if (rc || (req->rq_status != 0 &&
1365                                    req->rq_status != -EISDIR &&
1366                                    req->rq_status != -ENOTDIR)) {
1367                                 rep->lock_policy_res2 = req->rq_status;
1368                                 RETURN(ELDLM_LOCK_ABORTED);
1369                         }
1370                         break;
1371                 case IT_GETATTR:
1372                 case IT_LOOKUP:
1373                 case IT_OPEN:
1374                 case IT_READDIR:
1375                 case IT_READLINK:
1376                 case IT_RENAME:
1377                 case IT_LINK:
1378                 case IT_SETATTR:
1379                         rc = mds_getattr_name(2, req);
1380                         /* FIXME: we need to sit down and decide on who should
1381                          * set req->rq_status, who should return negative and
1382                          * positive return values, and what they all mean. */
1383                         if (rc || req->rq_status != 0) {
1384                                 rep->lock_policy_res2 = req->rq_status;
1385                                 RETURN(ELDLM_LOCK_ABORTED);
1386                         }
1387                         break;
1388                 case IT_READDIR|IT_OPEN:
1389                         LBUG();
1390                         break;
1391                 default:
1392                         CERROR("Unhandled intent\n");
1393                         LBUG();
1394                 }
1395
1396                 /* We don't bother returning a lock to the client for a file
1397                  * or directory we are removing.
1398                  *
1399                  * As for link and rename, there is no reason for the client
1400                  * to get a lock on the target at this point.  If they are
1401                  * going to modify the file/directory later they will get a
1402                  * lock at that time.
1403                  */
1404                 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1405                                IT_RENAME | IT_RENAME2))
1406                         RETURN(ELDLM_LOCK_ABORTED);
1407
1408                 rep->lock_policy_res2 = req->rq_status;
1409                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1410
1411                 /* If the client is about to open a file that doesn't have an MD
1412                  * stripe record, it's going to need a write lock. */
1413                 if (it->opc & IT_OPEN) {
1414                         struct lov_mds_md *lmm =
1415                                 lustre_msg_buf(req->rq_repmsg, 2);
1416                         if (lmm->lmm_easize == 0) {
1417                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1418                                            " lock");
1419                                 lock->l_req_mode = LCK_PW;
1420                         }
1421                 }
1422
1423                 if (flags & LDLM_FL_INTENT_ONLY) {
1424                         LDLM_DEBUG(lock, "INTENT_ONLY, aborting lock");
1425                         RETURN(ELDLM_LOCK_ABORTED);
1426                 }
1427                 /* Give the client a lock on the child object, instead of the
1428                  * parent that it requested. */
1429                 new_resid[0] = NTOH__u32(mds_rep->ino);
1430                 new_resid[1] = NTOH__u32(mds_rep->generation);
1431                 if (new_resid[0] == 0)
1432                         LBUG();
1433                 old_res = lock->l_resource->lr_name[0];
1434
1435                 ldlm_lock_change_resource(lock, new_resid);
1436                 if (lock->l_resource == NULL) {
1437                         LBUG();
1438                         RETURN(-ENOMEM);
1439                 }
1440                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1441                            (long)old_res);
1442                 RETURN(ELDLM_LOCK_CHANGED);
1443         } else {
1444                 int size = sizeof(struct ldlm_reply);
1445                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1446                                      &req->rq_repmsg);
1447                 if (rc) {
1448                         LBUG();
1449                         RETURN(-ENOMEM);
1450                 }
1451         }
1452         RETURN(rc);
1453 }
1454
1455 int mds_attach(struct obd_device *dev,
1456                obd_count len, void *data)
1457 {
1458         int rc;
1459         rc = lprocfs_reg_obd(dev, (lprocfs_vars_t*)status_var_nm_1, (void*)dev);
1460         return rc;
1461 }
1462
1463 int mds_detach(struct obd_device *dev)
1464 {
1465         int rc;
1466         rc = lprocfs_dereg_obd(dev);
1467         return rc;
1468
1469 }
1470
1471 #define MDT_NUM_THREADS 8
1472 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1473 {
1474         int i;
1475         //        struct obd_ioctl_data* data = buf;
1476         struct mds_obd *mds = &obddev->u.mds;
1477         int rc = 0;
1478         ENTRY;
1479
1480         MOD_INC_USE_COUNT;
1481
1482         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1483                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1484                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1485                                            "self", mds_handle, "mds");
1486         if (!mds->mds_service) {
1487                 CERROR("failed to start service\n");
1488                 GOTO(err_dec, rc = -EINVAL);
1489         }
1490
1491         for (i = 0; i < MDT_NUM_THREADS; i++) {
1492                 char name[32];
1493                 sprintf(name, "lustre_MDT_%02d", i);
1494                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1495                 if (rc) {
1496                         CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
1497                         GOTO(err_thread, rc);
1498                 }
1499         }
1500
1501         RETURN(0);
1502
1503 err_thread:
1504         ptlrpc_stop_all_threads(mds->mds_service);
1505         ptlrpc_unregister_service(mds->mds_service);
1506 err_dec:
1507         MOD_DEC_USE_COUNT;
1508         RETURN(rc);
1509 }
1510
1511
1512 static int mdt_cleanup(struct obd_device *obddev)
1513 {
1514         struct mds_obd *mds = &obddev->u.mds;
1515         ENTRY;
1516
1517         ptlrpc_stop_all_threads(mds->mds_service);
1518         ptlrpc_unregister_service(mds->mds_service);
1519
1520         MOD_DEC_USE_COUNT;
1521         RETURN(0);
1522 }
1523
1524 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1525                          int len, void *karg, void *uarg);
1526
1527 /* use obd ops to offer management infrastructure */
1528 static struct obd_ops mds_obd_ops = {
1529         o_attach:      mds_attach,
1530         o_detach:      mds_detach,
1531         o_connect:     mds_connect,
1532         o_disconnect:  mds_disconnect,
1533         o_setup:       mds_setup,
1534         o_cleanup:     mds_cleanup,
1535         o_iocontrol:   mds_iocontrol
1536 };
1537
1538 static struct obd_ops mdt_obd_ops = {
1539         o_setup:       mdt_setup,
1540         o_cleanup:     mdt_cleanup,
1541 };
1542
1543
1544 static int __init mds_init(void)
1545 {
1546
1547         mds_file_cache = kmem_cache_create("ll_mds_file_data",
1548                                            sizeof(struct mds_file_data),
1549                                            0, 0, NULL, NULL);
1550         if (mds_file_cache == NULL)
1551                 return -ENOMEM;
1552
1553         class_register_type(&mds_obd_ops, (lprocfs_vars_t*)status_class_var,
1554                             LUSTRE_MDS_NAME);
1555         class_register_type(&mdt_obd_ops, 0, LUSTRE_MDT_NAME);
1556
1557         ldlm_register_intent(ldlm_intent_policy);
1558
1559         return 0;
1560
1561 }
1562
1563 static void __exit mds_exit(void)
1564 {
1565
1566
1567         ldlm_unregister_intent();
1568         class_unregister_type(LUSTRE_MDS_NAME);
1569         class_unregister_type(LUSTRE_MDT_NAME);
1570         if (kmem_cache_destroy(mds_file_cache))
1571                 CERROR("couldn't free MDS file cache\n");
1572
1573 }
1574
1575 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1576 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1577 MODULE_LICENSE("GPL");
1578
1579 module_init(mds_init);
1580 module_exit(mds_exit);