lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001, 2002 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #define EXPORT_SYMTAB
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_dlm.h>
  34 #include <linux/init.h>
  35
  36 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
  37                            uuid_t *uuidarray);
  38 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
  39 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  40                                 struct ptlrpc_request *req);
  41 static int mds_cleanup(struct obd_device * obddev);
  42
  43 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  44 {
  45         return &req->rq_export->exp_obd->u.mds;
  46 }
  47
  48
  49 /* Assumes caller has already pushed into the kernel filesystem context */
  50 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  51                         __u64 offset)
  52 {
  53         int rc = 0;
  54         struct mds_obd *mds = mds_req2mds(req);
  55         struct ptlrpc_bulk_desc *desc;
  56         struct ptlrpc_bulk_page *bulk;
  57         char *buf;
  58         ENTRY;
  59
  60         desc = ptlrpc_prep_bulk(req->rq_connection);
  61         if (desc == NULL)
  62                 GOTO(out, rc = -ENOMEM);
  63
  64         bulk = ptlrpc_prep_bulk_page(desc);
  65         if (bulk == NULL)
  66                 GOTO(cleanup_bulk, rc = -ENOMEM);
  67
  68         OBD_ALLOC(buf, PAGE_SIZE);
  69         if (buf == NULL)
  70                 GOTO(cleanup_bulk, rc = -ENOMEM);
  71
  72         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  73
  74         if (rc != PAGE_SIZE)
  75                 GOTO(cleanup_buf, rc = -EIO);
  76
  77         bulk->b_xid = req->rq_xid;
  78         bulk->b_buf = buf;
  79         bulk->b_buflen = PAGE_SIZE;
  80         desc->b_portal = MDS_BULK_PORTAL;
  81
  82         rc = ptlrpc_send_bulk(desc);
  83         if (rc)
  84                 GOTO(cleanup_buf, rc);
  85
  86         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  87                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  88                        OBD_FAIL_MDS_SENDPAGE, rc);
  89                 ptlrpc_abort_bulk(desc);
  90                 GOTO(cleanup_buf, rc);
  91         }
  92
  93         wait_event(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
  94         if (desc->b_flags & PTL_RPC_FL_INTR)
  95                 GOTO(cleanup_buf, rc = -EINTR);
  96
  97         EXIT;
  98  cleanup_buf:
  99         OBD_FREE(buf, PAGE_SIZE);
 100  cleanup_bulk:
 101         ptlrpc_free_bulk(desc);
 102  out:
 103         return rc;
 104 }
 105
 106 /*
 107  * Look up a named entry in a directory, and get an LDLM lock on it.
 108  * 'dir' is a inode for which an LDLM lock has already been taken.
 109  *
 110  * If we do not need an exclusive or write lock on this entry (e.g.
 111  * a read lock for attribute lookup only) then we do not hold the
 112  * directory on return.  It is up to the caller to know what type
 113  * of lock it is getting, and clean up appropriately.
 114  */
 115 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 116                                       struct dentry *dir, struct vfsmount **mnt,
 117                                       char *name, int namelen, int lock_mode,
 118                                       struct lustre_handle *lockh,
 119                                       int dir_lock_mode)
 120 {
 121         struct dentry *dchild;
 122         int flags, rc;
 123         __u64 res_id[3] = {0};
 124         ENTRY;
 125
 126         down(&dir->d_inode->i_sem);
 127         dchild = lookup_one_len(name, dir, namelen);
 128         if (IS_ERR(dchild)) {
 129                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 130                 up(&dir->d_inode->i_sem);
 131                 LBUG();
 132                 RETURN(dchild);
 133         }
 134         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 135                 up(&dir->d_inode->i_sem);
 136                 ldlm_lock_decref(lockh, dir_lock_mode);
 137         }
 138
 139         if (lock_mode == 0 || !dchild->d_inode)
 140                 RETURN(dchild);
 141
 142         res_id[0] = dchild->d_inode->i_ino;
 143         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 144                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 145                                    &flags, ldlm_completion_ast,
 146                                    mds_blocking_ast, NULL, 0, lockh);
 147         if (rc != ELDLM_OK) {
 148                 l_dput(dchild);
 149                 up(&dir->d_inode->i_sem);
 150                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 151         }
 152
 153         RETURN(dchild);
 154 }
 155
 156 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 157                                      struct vfsmount **mnt, int lock_mode,
 158                                      struct lustre_handle *lockh)
 159 {
 160         struct mds_obd *mds = &obd->u.mds;
 161         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 162         int flags, rc;
 163         __u64 res_id[3] = {0};
 164         ENTRY;
 165
 166         if (IS_ERR(de))
 167                 RETURN(de);
 168
 169         res_id[0] = de->d_inode->i_ino;
 170         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 171                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 172                                    &flags, ldlm_completion_ast,
 173                                    mds_blocking_ast, NULL, 0, lockh);
 174         if (rc != ELDLM_OK) {
 175                 l_dput(de);
 176                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 177         }
 178
 179         RETURN(retval);
 180 }
 181
 182 /* Look up an entry by inode number. */
 183 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 184                               struct vfsmount **mnt)
 185 {
 186         /* stolen from NFS */
 187         struct super_block *sb = mds->mds_sb;
 188         unsigned long ino = fid->id;
 189         __u32 generation = fid->generation;
 190         struct inode *inode;
 191         struct list_head *lp;
 192         struct dentry *result;
 193
 194         if (ino == 0)
 195                 RETURN(ERR_PTR(-ESTALE));
 196
 197         inode = iget(sb, ino);
 198         if (inode == NULL)
 199                 RETURN(ERR_PTR(-ENOMEM));
 200
 201         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 202
 203         if (is_bad_inode(inode) ||
 204             (generation && inode->i_generation != generation)) {
 205                 /* we didn't find the right inode.. */
 206                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 207                        inode->i_ino, inode->i_nlink,
 208                        atomic_read(&inode->i_count), inode->i_generation,
 209                        generation);
 210                 LBUG();
 211                 iput(inode);
 212                 RETURN(ERR_PTR(-ESTALE));
 213         }
 214
 215         /* now to find a dentry. If possible, get a well-connected one */
 216         if (mnt)
 217                 *mnt = mds->mds_vfsmnt;
 218         spin_lock(&dcache_lock);
 219         list_for_each(lp, &inode->i_dentry) {
 220                 result = list_entry(lp, struct dentry, d_alias);
 221                 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 222                         dget_locked(result);
 223                         result->d_vfs_flags |= DCACHE_REFERENCED;
 224                         spin_unlock(&dcache_lock);
 225                         iput(inode);
 226                         if (mnt)
 227                                 mntget(*mnt);
 228                         return result;
 229                 }
 230         }
 231         spin_unlock(&dcache_lock);
 232         result = d_alloc_root(inode);
 233         if (result == NULL) {
 234                 iput(inode);
 235                 return ERR_PTR(-ENOMEM);
 236         }
 237         if (mnt)
 238                 mntget(*mnt);
 239         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 240         return result;
 241 }
 242
 243 /* Establish a connection to the MDS.
 244  *
 245  * This will set up an export structure for the client to hold state data
 246  * about that client, like open files, the last operation number it did
 247  * on the server, etc.
 248  */
 249 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 250                        char *cluuid)
 251 {
 252         struct obd_export *exp;
 253         struct mds_client_data *mcd;
 254         struct list_head *p;
 255         int rc;
 256         ENTRY;
 257
 258         if (!conn || !obd || !cluuid)
 259                 RETURN(-EINVAL);
 260
 261         MOD_INC_USE_COUNT;
 262
 263         list_for_each(p, &obd->obd_exports) {
 264                 exp = list_entry(p, struct obd_export, exp_chain);
 265                 mcd = exp->exp_mds_data.med_mcd;
 266                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 267                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 268                                cluuid, exp);
 269                         LASSERT(exp->exp_obd == obd);
 270
 271                         exp->exp_rconnh.addr = conn->addr;
 272                         exp->exp_rconnh.cookie = conn->cookie;
 273                         conn->addr = (__u64) (unsigned long)exp;
 274                         conn->cookie = exp->exp_cookie;
 275                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 276                                (long long)conn->addr, (long long)conn->cookie);
 277                         RETURN(0);
 278                 }
 279         }
 280
 281 #warning shaver: we might need a real cluuid here
 282         rc = class_connect(conn, obd, NULL);
 283         if (rc)
 284                 GOTO(out_dec, rc);
 285         exp = class_conn2export(conn);
 286         LASSERT(exp);
 287
 288         OBD_ALLOC(mcd, sizeof(*mcd));
 289         if (!mcd) {
 290                 CERROR("mds: out of memory for client data\n");
 291                 GOTO(out_export, rc = -ENOMEM);
 292         }
 293         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 294         exp->exp_mds_data.med_mcd = mcd;
 295         rc = mds_client_add(&exp->exp_mds_data, -1);
 296         if (rc)
 297                 GOTO(out_mdc, rc);
 298
 299         RETURN(0);
 300
 301 out_mdc:
 302         OBD_FREE(mcd, sizeof(*mcd));
 303 out_export:
 304         class_disconnect(conn);
 305 out_dec:
 306         MOD_DEC_USE_COUNT;
 307
 308         return rc;
 309 }
 310
 311 static int mds_disconnect(struct lustre_handle *conn)
 312 {
 313         struct obd_export *exp;
 314         int rc;
 315
 316         exp = class_conn2export(conn);
 317         if (!exp)
 318                 RETURN(-EINVAL);
 319
 320         rc = mds_client_free(&exp->exp_mds_data);
 321         if (rc)
 322                 CERROR("error freeing client data: rc = %d\n", rc);
 323
 324         rc = class_disconnect(conn);
 325         if (!rc)
 326                 MOD_DEC_USE_COUNT;
 327
 328         return rc;
 329 }
 330
 331 static int mds_getstatus(struct ptlrpc_request *req)
 332 {
 333         struct mds_obd *mds = mds_req2mds(req);
 334         struct mds_body *body;
 335         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 336         int rc, size = sizeof(*body);
 337         ENTRY;
 338
 339         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 340         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 341                 CERROR("mds: out of memory for message: size=%d\n", size);
 342                 req->rq_status = -ENOMEM;
 343                 RETURN(0);
 344         }
 345
 346         body = lustre_msg_buf(req->rq_reqmsg, 0);
 347         mds_unpack_body(body);
 348
 349         /* Anything we need to do here with the client's trans no or so? */
 350         body = lustre_msg_buf(req->rq_repmsg, 0);
 351         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 352
 353         LASSERT(med->med_mcd);
 354
 355         /* mcd_last_xid is is stored in little endian on the disk and
 356            mds_pack_rep_body converts it to network order */
 357         req->rq_repmsg->last_xid = le32_to_cpu(med->med_mcd->mcd_last_xid);
 358         mds_pack_rep_body(req);
 359         RETURN(0);
 360 }
 361
 362 static int mds_getlovinfo(struct ptlrpc_request *req)
 363 {
 364         struct mds_obd *mds = mds_req2mds(req);
 365         struct mds_status_req *streq;
 366         struct lov_desc *desc;
 367         int tgt_count;
 368         int rc, size[2] = {sizeof(*desc)};
 369         ENTRY;
 370
 371         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 372         streq->flags = NTOH__u32(streq->flags);
 373         streq->repbuf = NTOH__u32(streq->repbuf);
 374         size[1] = streq->repbuf;
 375
 376         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 377         if (rc) {
 378                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 379                 req->rq_status = -ENOMEM;
 380                 RETURN(0);
 381         }
 382
 383         desc = lustre_msg_buf(req->rq_repmsg, 0);
 384         rc = mds_get_lovdesc(req->rq_obd, desc);
 385         if (rc != 0 ) {
 386                 CERROR("mds_get_lovdesc error %d", rc);
 387                 req->rq_status = rc;
 388                 RETURN(0);
 389         }
 390
 391         tgt_count = NTOH__u32(desc->ld_tgt_count);
 392         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 393                 CERROR("too many targets, enlarge client buffers\n");
 394                 req->rq_status = -ENOSPC;
 395                 RETURN(0);
 396         }
 397
 398         mds->mds_max_mdsize = sizeof(struct lov_stripe_md) +
 399                 tgt_count * sizeof(struct lov_object_id);
 400         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 401                              lustre_msg_buf(req->rq_repmsg, 1));
 402         if (rc) {
 403                 CERROR("get_lovtgts error %d\n", rc);
 404                 req->rq_status = rc;
 405                 RETURN(0);
 406         }
 407         RETURN(0);
 408 }
 409
 410 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 411                      void *data, __u32 data_len)
 412 {
 413         struct lustre_handle lockh;
 414         ENTRY;
 415
 416         ldlm_lock2handle(lock, &lockh);
 417         if (ldlm_cli_cancel(&lockh) < 0)
 418                 LBUG();
 419         RETURN(0);
 420 }
 421
 422 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 423 {
 424         struct mds_obd *mds = mds_req2mds(req);
 425         struct obd_device *obd = req->rq_export->exp_obd;
 426         struct obd_run_ctxt saved;
 427         struct mds_body *body;
 428         struct dentry *de = NULL, *dchild = NULL;
 429         struct inode *dir;
 430         struct lustre_handle lockh;
 431         char *name;
 432         int namelen, flags, lock_mode, rc = 0;
 433         __u64 res_id[3] = {0, 0, 0};
 434         ENTRY;
 435
 436         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 437
 438         if (req->rq_reqmsg->bufcount <= offset + 1) {
 439                 LBUG();
 440                 GOTO(out_pre_de, rc = -EINVAL);
 441         }
 442
 443         body = lustre_msg_buf(req->rq_reqmsg, offset);
 444         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 445         namelen = req->rq_reqmsg->buflens[offset + 1];
 446         /* requests were at offset 2, replies go back at 1 */
 447         if (offset)
 448                 offset = 1;
 449
 450         push_ctxt(&saved, &mds->mds_ctxt);
 451         de = mds_fid2dentry(mds, &body->fid1, NULL);
 452         if (IS_ERR(de)) {
 453                 LBUG();
 454                 GOTO(out_pre_de, rc = -ESTALE);
 455         }
 456
 457         dir = de->d_inode;
 458         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 459
 460         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 461         res_id[0] = dir->i_ino;
 462
 463         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 464                              NULL, 0, lock_mode, &lockh);
 465         if (rc == 0) {
 466                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 467                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 468                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 469                                       &flags, ldlm_completion_ast,
 470                                       mds_blocking_ast, NULL, 0, &lockh);
 471                 if (rc != ELDLM_OK) {
 472                         CERROR("lock enqueue: err: %d\n", rc);
 473                         GOTO(out_create_de, rc = -EIO);
 474                 }
 475         }
 476         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 477
 478         down(&dir->i_sem);
 479         dchild = lookup_one_len(name, de, namelen - 1);
 480         if (IS_ERR(dchild)) {
 481                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 482                 up(&dir->i_sem);
 483                 LBUG();
 484                 GOTO(out_create_dchild, rc = -ESTALE);
 485         }
 486
 487         if (dchild->d_inode) {
 488                 struct mds_body *body;
 489                 struct inode *inode = dchild->d_inode;
 490                 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
 491                        dir->i_ino, name, dchild->d_inode->i_ino);
 492
 493                 body = lustre_msg_buf(req->rq_repmsg, offset);
 494                 mds_pack_inode2fid(&body->fid1, inode);
 495                 mds_pack_inode2body(body, inode);
 496                 if (S_ISREG(inode->i_mode)) {
 497                         struct lov_stripe_md *md;
 498                         md = lustre_msg_buf(req->rq_repmsg, offset + 1);
 499                         md->lmd_easize = mds->mds_max_mdsize;
 500                         /* FIXME: why don't we check (or use) rc of get_md? */
 501                         mds_fs_get_md(mds, inode, md);
 502                 }
 503                 /* now a normal case for intent locking */
 504                 rc = 0;
 505         } else
 506                 rc = -ENOENT;
 507
 508         EXIT;
 509 out_create_dchild:
 510         l_dput(dchild);
 511         up(&dir->i_sem);
 512         ldlm_lock_decref(&lockh, lock_mode);
 513 out_create_de:
 514         l_dput(de);
 515 out_pre_de:
 516         req->rq_status = rc;
 517         pop_ctxt(&saved);
 518         return 0;
 519 }
 520
 521
 522 static int mds_getattr(int offset, struct ptlrpc_request *req)
 523 {
 524         struct mds_obd *mds = mds_req2mds(req);
 525         struct obd_run_ctxt saved;
 526         struct dentry *de;
 527         struct inode *inode;
 528         struct mds_body *body;
 529         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 530         ENTRY;
 531
 532         body = lustre_msg_buf(req->rq_reqmsg, offset);
 533         push_ctxt(&saved, &mds->mds_ctxt);
 534         de = mds_fid2dentry(mds, &body->fid1, NULL);
 535         if (IS_ERR(de)) {
 536                 GOTO(out_pop, rc = -ENOENT);
 537         }
 538
 539         inode = de->d_inode;
 540         if (S_ISREG(body->fid1.f_type)) {
 541                 bufcount = 2;
 542                 size[1] = mds->mds_max_mdsize;
 543         } else if (body->valid & OBD_MD_LINKNAME) {
 544                 bufcount = 2;
 545                 size[1] = inode->i_size;
 546         }
 547
 548         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 549                              &req->rq_repmsg);
 550         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 551                 CERROR("mds: out of memory\n");
 552                 GOTO(out, rc);
 553         }
 554
 555         if (body->valid & OBD_MD_LINKNAME) {
 556                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 557
 558                 rc = inode->i_op->readlink(de, tmp, size[1]);
 559
 560                 if (rc < 0) {
 561                         CERROR("readlink failed: %d\n", rc);
 562                         GOTO(out, rc);
 563                 }
 564         }
 565
 566         body = lustre_msg_buf(req->rq_repmsg, 0);
 567         body->ino = inode->i_ino;
 568         body->generation = inode->i_generation;
 569         body->atime = inode->i_atime;
 570         body->ctime = inode->i_ctime;
 571         body->mtime = inode->i_mtime;
 572         body->uid = inode->i_uid;
 573         body->gid = inode->i_gid;
 574         body->size = inode->i_size;
 575         body->mode = inode->i_mode;
 576         body->nlink = inode->i_nlink;
 577         body->valid = (OBD_MD_FLID | OBD_MD_FLATIME | OBD_MD_FLCTIME |
 578                        OBD_MD_FLMTIME | OBD_MD_FLSIZE | OBD_MD_FLUID |
 579                        OBD_MD_FLGID | OBD_MD_FLNLINK | OBD_MD_FLGENER |
 580                        OBD_MD_FLMODE);
 581
 582         if (S_ISREG(inode->i_mode)) {
 583                 rc = mds_fs_get_md(mds, inode,
 584                                    lustre_msg_buf(req->rq_repmsg, 1));
 585                 if (rc < 0) {
 586                         CERROR("mds_fs_get_md failed: %d\n", rc);
 587                         GOTO(out, rc);
 588                 }
 589                 body->valid |= OBD_MD_FLEASIZE;
 590         } else if (S_ISLNK(inode->i_mode))
 591                 body->valid |= OBD_MD_LINKNAME;
 592
 593 out:
 594         l_dput(de);
 595 out_pop:
 596         pop_ctxt(&saved);
 597         req->rq_status = rc;
 598         RETURN(0);
 599 }
 600
 601 static int mds_statfs(struct ptlrpc_request *req)
 602 {
 603         struct mds_obd *mds = mds_req2mds(req);
 604         struct obd_statfs *osfs;
 605         struct statfs sfs;
 606         int rc, size = sizeof(*osfs);
 607         ENTRY;
 608
 609         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 610                              &req->rq_repmsg);
 611         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 612                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 613                 GOTO(out, rc);
 614         }
 615
 616         rc = vfs_statfs(mds->mds_sb, &sfs);
 617         if (rc) {
 618                 CERROR("mds: statfs failed: rc %d\n", rc);
 619                 GOTO(out, rc);
 620         }
 621         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 622         memset(osfs, 0, size);
 623         obd_statfs_pack(osfs, &sfs);
 624
 625 out:
 626         req->rq_status = rc;
 627         RETURN(0);
 628 }
 629
 630 static int mds_open(struct ptlrpc_request *req)
 631 {
 632         struct dentry *de;
 633         struct mds_body *body;
 634         struct file *file;
 635         struct vfsmount *mnt;
 636         struct mds_obd *mds = mds_req2mds(req);
 637         struct mds_export_data *med;
 638         __u32 flags;
 639         struct list_head *tmp;
 640         struct mds_file_data *mfd;
 641         int rc, size = sizeof(*body);
 642         ENTRY;
 643
 644         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 645         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 646                 CERROR("mds: out of memory\n");
 647                 req->rq_status = -ENOMEM;
 648                 RETURN(0);
 649         }
 650
 651         body = lustre_msg_buf(req->rq_reqmsg, 0);
 652
 653         /* was this animal open already? */
 654         /* XXX we should only check on re-open, or do a refcount... */
 655         med = &req->rq_export->exp_mds_data;
 656         list_for_each(tmp, &med->med_open_head) {
 657                 struct mds_file_data *fd;
 658                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 659                 if (body->extra == fd->mfd_clientfd &&
 660                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 661                         CERROR("Re opening %Ld\n", body->fid1.id);
 662                         RETURN(0);
 663                 }
 664         }
 665
 666         OBD_ALLOC(mfd, sizeof(*mfd));
 667         if (!mfd) {
 668                 CERROR("mds: out of memory\n");
 669                 req->rq_status = -ENOMEM;
 670                 RETURN(0);
 671         }
 672
 673         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 674         if (IS_ERR(de)) {
 675                 req->rq_status = -ENOENT;
 676                 RETURN(0);
 677         }
 678
 679         /* check if this inode has seen a delayed object creation */
 680         if (req->rq_reqmsg->bufcount > 1) {
 681                 void *handle;
 682                 struct lov_stripe_md *md;
 683                 struct inode *inode = de->d_inode;
 684                 int rc;
 685
 686                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 687
 688                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 689                 if (!handle) {
 690                         req->rq_status = -ENOMEM;
 691                         RETURN(0);
 692                 }
 693
 694                 /* XXX error handling */
 695                 rc = mds_fs_set_md(mds, inode, handle, md);
 696                 if (!rc) {
 697                         struct obd_run_ctxt saved;
 698                         push_ctxt(&saved, &mds->mds_ctxt);
 699                         rc = mds_update_last_rcvd(mds, handle, req);
 700                         pop_ctxt(&saved);
 701                 } else {
 702                         req->rq_status = rc;
 703                         RETURN(0);
 704                 }
 705                 /* FIXME: need to return last_rcvd, last_committed */
 706
 707                 /* FIXME: keep rc intact */
 708                 rc = mds_fs_commit(mds, de->d_inode, handle);
 709                 if (rc) {
 710                         req->rq_status = rc;
 711                         RETURN(0);
 712                 }
 713         }
 714
 715         flags = body->flags;
 716         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 717         if (!file || IS_ERR(file)) {
 718                 req->rq_status = -EINVAL;
 719                 OBD_FREE(mfd, sizeof(*mfd));
 720                 RETURN(0);
 721         }
 722
 723         file->private_data = mfd;
 724         mfd->mfd_file = file;
 725         mfd->mfd_clientfd = body->extra;
 726         list_add(&mfd->mfd_list, &med->med_open_head);
 727
 728         body = lustre_msg_buf(req->rq_repmsg, 0);
 729         /* FIXME: need to have cookies involved here */
 730         body->extra = (__u64) (unsigned long)file;
 731         RETURN(0);
 732 }
 733
 734 static int mds_close(struct ptlrpc_request *req)
 735 {
 736         struct dentry *de;
 737         struct mds_body *body;
 738         struct file *file;
 739         struct mds_obd *mds = mds_req2mds(req);
 740         struct vfsmount *mnt;
 741         struct mds_file_data *mfd;
 742         int rc;
 743         ENTRY;
 744
 745         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 746         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 747                 CERROR("mds: out of memory\n");
 748                 req->rq_status = -ENOMEM;
 749                 RETURN(0);
 750         }
 751
 752         body = lustre_msg_buf(req->rq_reqmsg, 0);
 753         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 754         if (IS_ERR(de)) {
 755                 req->rq_status = -ENOENT;
 756                 RETURN(0);
 757         }
 758
 759         /* FIXME: need to have cookies involved here */
 760         file = (struct file *)(unsigned long)body->extra;
 761         if (!file->f_dentry)
 762                 LBUG();
 763         mfd = (struct mds_file_data *)file->private_data;
 764         list_del(&mfd->mfd_list);
 765         OBD_FREE(mfd, sizeof(*mfd));
 766
 767         req->rq_status = filp_close(file, 0);
 768         l_dput(de);
 769         mntput(mnt);
 770
 771         RETURN(0);
 772 }
 773
 774 static int mds_readpage(struct ptlrpc_request *req)
 775 {
 776         struct mds_obd *mds = mds_req2mds(req);
 777         struct vfsmount *mnt;
 778         struct dentry *de;
 779         struct file *file;
 780         struct mds_body *body;
 781         struct obd_run_ctxt saved;
 782         int rc, size = sizeof(*body);
 783         ENTRY;
 784
 785         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 786         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 787                 CERROR("mds: out of memory\n");
 788                 GOTO(out, rc = -ENOMEM);
 789         }
 790
 791         body = lustre_msg_buf(req->rq_reqmsg, 0);
 792         push_ctxt(&saved, &mds->mds_ctxt);
 793         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 794         if (IS_ERR(de))
 795                 GOTO(out_pop, rc = PTR_ERR(de));
 796
 797         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 798
 799         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 800         /* note: in case of an error, dentry_open puts dentry */
 801         if (IS_ERR(file))
 802                 GOTO(out_pop, rc = PTR_ERR(file));
 803
 804         /* to make this asynchronous make sure that the handling function
 805            doesn't send a reply when this function completes. Instead a
 806            callback function would send the reply */
 807         rc = mds_sendpage(req, file, body->size);
 808
 809         filp_close(file, 0);
 810 out_pop:
 811         pop_ctxt(&saved);
 812 out:
 813         req->rq_status = rc;
 814         RETURN(0);
 815 }
 816
 817 int mds_reint(int offset, struct ptlrpc_request *req)
 818 {
 819         int rc;
 820         struct mds_update_record rec;
 821
 822         rc = mds_update_unpack(req, offset, &rec);
 823         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 824                 CERROR("invalid record\n");
 825                 req->rq_status = -EINVAL;
 826                 RETURN(0);
 827         }
 828         /* rc will be used to interrupt a for loop over multiple records */
 829         rc = mds_reint_rec(&rec, offset, req);
 830         return rc;
 831 }
 832
 833 int mds_handle(struct ptlrpc_request *req)
 834 {
 835         int rc;
 836         ENTRY;
 837
 838         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 839         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 840                 CERROR("lustre_mds: Invalid request\n");
 841                 GOTO(out, rc);
 842         }
 843
 844         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 845                 CERROR("lustre_mds: wrong packet type sent %d\n",
 846                        req->rq_reqmsg->type);
 847                 GOTO(out, rc = -EINVAL);
 848         }
 849
 850         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
 851                 GOTO(out, rc = -ENOTCONN);
 852
 853         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 854                 GOTO(out, rc = -EINVAL);
 855
 856         switch (req->rq_reqmsg->opc) {
 857         case MDS_CONNECT:
 858                 CDEBUG(D_INODE, "connect\n");
 859                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 860                 rc = target_handle_connect(req);
 861                 break;
 862
 863         case MDS_DISCONNECT:
 864                 CDEBUG(D_INODE, "disconnect\n");
 865                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 866                 rc = target_handle_disconnect(req);
 867                 goto out;
 868
 869         case MDS_GETSTATUS:
 870                 CDEBUG(D_INODE, "getstatus\n");
 871                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 872                 rc = mds_getstatus(req);
 873                 break;
 874
 875         case MDS_GETLOVINFO:
 876                 CDEBUG(D_INODE, "getlovinfo\n");
 877                 rc = mds_getlovinfo(req);
 878                 break;
 879
 880         case MDS_GETATTR:
 881                 CDEBUG(D_INODE, "getattr\n");
 882                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 883                 rc = mds_getattr(0, req);
 884                 break;
 885
 886         case MDS_STATFS:
 887                 CDEBUG(D_INODE, "statfs\n");
 888                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 889                 rc = mds_statfs(req);
 890                 break;
 891
 892         case MDS_READPAGE:
 893                 CDEBUG(D_INODE, "readpage\n");
 894                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 895                 rc = mds_readpage(req);
 896
 897                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 898                         return 0;
 899                 break;
 900
 901         case MDS_REINT: {
 902                 int size = sizeof(struct mds_body);
 903                 CDEBUG(D_INODE, "reint\n");
 904                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 905
 906                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 907                                      &req->rq_repmsg);
 908                 if (rc) {
 909                         req->rq_status = rc;
 910                         break;
 911                 }
 912                 rc = mds_reint(0, req);
 913                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 914                 break;
 915                 }
 916
 917         case MDS_OPEN:
 918                 CDEBUG(D_INODE, "open\n");
 919                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 920                 rc = mds_open(req);
 921                 break;
 922
 923         case MDS_CLOSE:
 924                 CDEBUG(D_INODE, "close\n");
 925                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 926                 rc = mds_close(req);
 927                 break;
 928
 929         case LDLM_ENQUEUE:
 930                 CDEBUG(D_INODE, "enqueue\n");
 931                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 932                 rc = ldlm_handle_enqueue(req);
 933                 if (rc)
 934                         break;
 935                 RETURN(0);
 936
 937         case LDLM_CONVERT:
 938                 CDEBUG(D_INODE, "convert\n");
 939                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 940                 rc = ldlm_handle_convert(req);
 941                 if (rc)
 942                         break;
 943                 RETURN(0);
 944
 945         case LDLM_CANCEL:
 946                 CDEBUG(D_INODE, "cancel\n");
 947                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 948                 rc = ldlm_handle_cancel(req);
 949                 if (rc)
 950                         break;
 951                 RETURN(0);
 952         case LDLM_BL_CALLBACK:
 953         case LDLM_CP_CALLBACK:
 954                 CDEBUG(D_INODE, "callback\n");
 955                 CERROR("callbacks should not happen on MDS\n");
 956                 LBUG();
 957                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 958                 break;
 959         default:
 960                 rc = ptlrpc_error(req->rq_svc, req);
 961                 RETURN(rc);
 962         }
 963
 964         EXIT;
 965
 966         if (!rc) {
 967                 struct mds_obd *mds = mds_req2mds(req);
 968                 req->rq_repmsg->last_xid = HTON__u64(mds->mds_last_rcvd);
 969                 req->rq_repmsg->last_committed =
 970                         HTON__u64(mds->mds_last_committed);
 971                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 972                        (unsigned long long)mds->mds_last_rcvd,
 973                        (unsigned long long)mds->mds_last_committed,
 974                        cpu_to_le32(req->rq_xid));
 975         }
 976  out:
 977         /* Still not 100% sure whether we should reply with the server
 978          * last_rcvd or that of this client.  I'm not sure it even makes
 979          * a difference on a per-client basis, because last_rcvd is global
 980          * and we are not supposed to allow transactions while in recovery.
 981          */
 982         if (rc) {
 983                 CERROR("mds: processing error %d\n", rc);
 984                 ptlrpc_error(req->rq_svc, req);
 985         } else {
 986                 CDEBUG(D_NET, "sending reply\n");
 987                 ptlrpc_reply(req->rq_svc, req);
 988         }
 989         return 0;
 990 }
 991
 992 /* Update the server data on disk.  This stores the new mount_count and
 993  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 994  * then the server last_rcvd value may be less than that of the clients.
 995  * This will alert us that we may need to do client recovery.
 996  *
 997  * Assumes we are already in the server filesystem context.
 998  */
 999 static
1000 int mds_update_server_data(struct mds_obd *mds)
1001 {
1002         struct mds_server_data *msd = mds->mds_server_data;
1003         struct file *filp = mds->mds_rcvd_filp;
1004         loff_t off = 0;
1005         int rc;
1006
1007         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1008         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1009
1010         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1011                (unsigned long long)mds->mds_mount_count,
1012                (unsigned long long)mds->mds_last_rcvd);
1013         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1014         if (rc != sizeof(*msd)) {
1015                 CERROR("error writing MDS server data: rc = %d\n", rc);
1016                 if (rc > 0)
1017                         RETURN(-EIO);
1018                 RETURN(rc);
1019         }
1020         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1021         if (rc)
1022                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1023
1024         return 0;
1025 }
1026
1027 /* Do recovery actions for the MDS */
1028 static int mds_recover(struct obd_device *obddev)
1029 {
1030         struct mds_obd *mds = &obddev->u.mds;
1031         struct obd_run_ctxt saved;
1032         int rc;
1033
1034         /* This happens at the end when recovery is complete */
1035         ++mds->mds_mount_count;
1036         push_ctxt(&saved, &mds->mds_ctxt);
1037         rc = mds_update_server_data(mds);
1038         pop_ctxt(&saved);
1039
1040         return rc;
1041 }
1042
1043 #define MDS_NUM_THREADS 8
1044 /* mount the file system (secretly) */
1045 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1046 {
1047         int i;
1048         struct obd_ioctl_data* data = buf;
1049         struct mds_obd *mds = &obddev->u.mds;
1050         struct vfsmount *mnt;
1051         int rc = 0;
1052         ENTRY;
1053
1054         MOD_INC_USE_COUNT;
1055 #ifdef CONFIG_DEV_RDONLY
1056         dev_clear_rdonly(2);
1057 #endif
1058         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1059                 GOTO(err_dec, rc = -EINVAL);
1060
1061         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1062
1063         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1064         if (IS_ERR(mnt)) {
1065                 rc = PTR_ERR(mnt);
1066                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1067                 GOTO(err_kfree, rc);
1068         }
1069
1070         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1071         if (!mds->mds_sb)
1072                 GOTO(err_put, rc = -ENODEV);
1073
1074         mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
1075         rc = mds_fs_setup(obddev, mnt);
1076         if (rc) {
1077                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1078                 GOTO(err_put, rc);
1079         }
1080
1081         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1082                                            MDC_REPLY_PORTAL, "self",mds_handle);
1083         if (!mds->mds_service) {
1084                 CERROR("failed to start service\n");
1085                 GOTO(err_fs, rc = -EINVAL);
1086         }
1087
1088         obddev->obd_namespace =
1089                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1090         if (obddev->obd_namespace == NULL) {
1091                 mds_cleanup(obddev);
1092                 GOTO(err_svc, rc = -ENOMEM);
1093         }
1094
1095         for (i = 0; i < MDS_NUM_THREADS; i++) {
1096                 char name[32];
1097                 sprintf(name, "lustre_MDS_%02d", i);
1098                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1099                 if (rc) {
1100                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1101                         GOTO(err_thread, rc);
1102                 }
1103         }
1104
1105         rc = mds_recover(obddev);
1106         if (rc)
1107                 GOTO(err_thread, rc);
1108
1109         RETURN(0);
1110
1111 err_thread:
1112         ptlrpc_stop_all_threads(mds->mds_service);
1113 err_svc:
1114         ptlrpc_unregister_service(mds->mds_service);
1115 err_fs:
1116         mds_fs_cleanup(obddev);
1117 err_put:
1118         unlock_kernel();
1119         mntput(mds->mds_vfsmnt);
1120         mds->mds_sb = 0;
1121         lock_kernel();
1122 err_kfree:
1123         kfree(mds->mds_fstype);
1124 err_dec:
1125         MOD_DEC_USE_COUNT;
1126         return rc;
1127 }
1128
1129 static int mds_cleanup(struct obd_device *obddev)
1130 {
1131         struct super_block *sb;
1132         struct mds_obd *mds = &obddev->u.mds;
1133         struct obd_run_ctxt saved;
1134         ENTRY;
1135
1136         ptlrpc_stop_all_threads(mds->mds_service);
1137         ptlrpc_unregister_service(mds->mds_service);
1138
1139         sb = mds->mds_sb;
1140         if (!mds->mds_sb)
1141                 RETURN(0);
1142
1143         push_ctxt(&saved, &mds->mds_ctxt);
1144         mds_update_server_data(mds);
1145
1146         if (mds->mds_rcvd_filp) {
1147                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1148                 mds->mds_rcvd_filp = NULL;
1149
1150                 if (rc)
1151                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1152         }
1153         pop_ctxt(&saved);
1154
1155         unlock_kernel();
1156         mntput(mds->mds_vfsmnt);
1157         mds->mds_sb = 0;
1158         kfree(mds->mds_fstype);
1159
1160         ldlm_namespace_free(obddev->obd_namespace);
1161
1162         lock_kernel();
1163 #ifdef CONFIG_DEV_RDONLY
1164         dev_clear_rdonly(2);
1165 #endif
1166         mds_fs_cleanup(obddev);
1167
1168         MOD_DEC_USE_COUNT;
1169         RETURN(0);
1170 }
1171
1172 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1173                               ldlm_mode_t mode, void *data)
1174 {
1175         struct ptlrpc_request *req = req_cookie;
1176         int rc = 0;
1177         ENTRY;
1178
1179         if (!req_cookie)
1180                 RETURN(0);
1181
1182         if (req->rq_reqmsg->bufcount > 1) {
1183                 /* an intent needs to be considered */
1184                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1185                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1186                 struct mds_body *mds_rep;
1187                 struct ldlm_reply *rep;
1188                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1189                 int bufcount = -1, rc, size[3] = {sizeof(struct ldlm_reply),
1190                                                   sizeof(struct mds_body),
1191                                                   mds->mds_max_mdsize};
1192
1193                 it->opc = NTOH__u64(it->opc);
1194
1195                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1196                            ldlm_it2str(it->opc));
1197
1198                 /* prepare reply */
1199                 switch((long)it->opc) {
1200                 case IT_GETATTR:
1201                         /* Note that in the negative case you may be returning
1202                          * a file and its obdo */
1203                 case IT_CREAT:
1204                 case IT_CREAT|IT_OPEN:
1205                 case IT_LINK:
1206                 case IT_LOOKUP:
1207                 case IT_MKDIR:
1208                 case IT_MKNOD:
1209                 case IT_OPEN:
1210                 case IT_READLINK:
1211                 case IT_RENAME:
1212                 case IT_RMDIR:
1213                 case IT_SETATTR:
1214                 case IT_SYMLINK:
1215                 case IT_UNLINK:
1216                         bufcount = 3;
1217                         break;
1218                 case IT_RENAME2:
1219                         bufcount = 1;
1220                         break;
1221                 default:
1222                         LBUG();
1223                 }
1224
1225                 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
1226                                      &req->rq_repmsg);
1227                 if (rc) {
1228                         rc = req->rq_status = -ENOMEM;
1229                         RETURN(rc);
1230                 }
1231
1232                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1233                 rep->lock_policy_res1 = 1;
1234
1235                 /* execute policy */
1236                 switch ((long)it->opc) {
1237                 case IT_CREAT|IT_OPEN:
1238                         rc = mds_reint(2, req);
1239                         if (rc || (req->rq_status != 0 &&
1240                                    req->rq_status != -EEXIST)) {
1241                                 rep->lock_policy_res2 = req->rq_status;
1242                                 RETURN(ELDLM_LOCK_ABORTED);
1243                         }
1244                         break;
1245                 case IT_CREAT:
1246                 case IT_LINK:
1247                 case IT_MKDIR:
1248                 case IT_MKNOD:
1249                 case IT_RENAME2:
1250                 case IT_RMDIR:
1251                 case IT_SYMLINK:
1252                 case IT_UNLINK:
1253                         rc = mds_reint(2, req);
1254                         if (rc || (req->rq_status != 0 &&
1255                                    req->rq_status != -EISDIR &&
1256                                    req->rq_status != -ENOTDIR)) {
1257                                 rep->lock_policy_res2 = req->rq_status;
1258                                 RETURN(ELDLM_LOCK_ABORTED);
1259                         }
1260                         break;
1261                 case IT_GETATTR:
1262                 case IT_LOOKUP:
1263                 case IT_OPEN:
1264                 case IT_READDIR:
1265                 case IT_READLINK:
1266                 case IT_RENAME:
1267                 case IT_SETATTR:
1268                         rc = mds_getattr_name(2, req);
1269                         /* FIXME: we need to sit down and decide on who should
1270                          * set req->rq_status, who should return negative and
1271                          * positive return values, and what they all mean. */
1272                         if (rc || req->rq_status != 0) {
1273                                 rep->lock_policy_res2 = req->rq_status;
1274                                 RETURN(ELDLM_LOCK_ABORTED);
1275                         }
1276                         break;
1277                 case IT_READDIR|IT_OPEN:
1278                         LBUG();
1279                         break;
1280                 default:
1281                         CERROR("Unhandled intent\n");
1282                         LBUG();
1283                 }
1284
1285                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1286                     it->opc == IT_RENAME || it->opc == IT_RENAME2)
1287                         RETURN(ELDLM_LOCK_ABORTED);
1288
1289                 rep->lock_policy_res2 = req->rq_status;
1290                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1291
1292                 /* If the client is about to open a file that doesn't have an MD
1293                  * stripe record, it's going to need a write lock. */
1294                 if (it->opc & IT_OPEN) {
1295                         struct lov_stripe_md *md =
1296                                 lustre_msg_buf(req->rq_repmsg, 2);
1297                         if (md->lmd_easize == 0) {
1298                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1299                                            " lock");
1300                                 lock->l_req_mode = LCK_PW;
1301                         }
1302                 }
1303
1304                 /* Give the client a lock on the child object, instead of the
1305                  * parent that it requested. */
1306                 new_resid[0] = NTOH__u32(mds_rep->ino);
1307                 if (new_resid[0] == 0)
1308                         LBUG();
1309                 old_res = lock->l_resource->lr_name[0];
1310
1311                 ldlm_lock_change_resource(lock, new_resid);
1312                 if (lock->l_resource == NULL) {
1313                         LBUG();
1314                         RETURN(-ENOMEM);
1315                 }
1316                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1317                            (long)old_res);
1318                 RETURN(ELDLM_LOCK_CHANGED);
1319         } else {
1320                 int size = sizeof(struct ldlm_reply);
1321                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1322                                      &req->rq_repmsg);
1323                 if (rc) {
1324                         LBUG();
1325                         RETURN(-ENOMEM);
1326                 }
1327         }
1328         RETURN(rc);
1329 }
1330
1331
1332 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1333                          int len, void *karg, void *uarg);
1334
1335 /* use obd ops to offer management infrastructure */
1336 static struct obd_ops mds_obd_ops = {
1337         o_connect:     mds_connect,
1338         o_disconnect:  mds_disconnect,
1339         o_setup:       mds_setup,
1340         o_cleanup:     mds_cleanup,
1341         o_iocontrol:   mds_iocontrol
1342 };
1343
1344 static int __init mds_init(void)
1345 {
1346         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1347         ldlm_register_intent(ldlm_intent_policy);
1348         return 0;
1349 }
1350
1351 static void __exit mds_exit(void)
1352 {
1353         ldlm_unregister_intent();
1354         class_unregister_type(LUSTRE_MDS_NAME);
1355 }
1356
1357 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1358 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1359 MODULE_LICENSE("GPL");
1360
1361 module_init(mds_init);
1362 module_exit(mds_exit);