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