lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001, 2002 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #define EXPORT_SYMTAB
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_dlm.h>
  34 #include <linux/init.h>
  35 #include <linux/obd_class.h>
  36 #include <linux/random.h>
  37 #include <linux/locks.h>
  38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  39 #include <linux/buffer_head.h>
  40 #endif
  41 #include <linux/obd_lov.h>
  42 #include <linux/lprocfs_status.h>
  43
  44 static kmem_cache_t *mds_file_cache;
  45
  46 extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
  47                            obd_uuid_t *uuidarray);
  48 extern int mds_get_lovdesc(struct mds_obd  *obd, struct lov_desc *desc);
  49 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  50                                 struct ptlrpc_request *req);
  51 static int mds_cleanup(struct obd_device * obddev);
  52
  53 extern struct lprocfs_vars status_var_nm_1[];
  54 extern struct lprocfs_vars status_class_var[];
  55
  56 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  57 {
  58         return &req->rq_export->exp_obd->u.mds;
  59 }
  60
  61 static int mds_bulk_timeout(void *data)
  62 {
  63         struct ptlrpc_bulk_desc *desc = data;
  64
  65         ENTRY;
  66         CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
  67         RETURN(1);
  68 }
  69
  70 /* Assumes caller has already pushed into the kernel filesystem context */
  71 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  72                         __u64 offset)
  73 {
  74         int rc = 0;
  75         struct mds_obd *mds = mds_req2mds(req);
  76         struct ptlrpc_bulk_desc *desc;
  77         struct ptlrpc_bulk_page *bulk;
  78         struct l_wait_info lwi;
  79         char *buf;
  80         ENTRY;
  81
  82         desc = ptlrpc_prep_bulk(req->rq_connection);
  83         if (desc == NULL)
  84                 GOTO(out, rc = -ENOMEM);
  85
  86         bulk = ptlrpc_prep_bulk_page(desc);
  87         if (bulk == NULL)
  88                 GOTO(cleanup_bulk, rc = -ENOMEM);
  89
  90         OBD_ALLOC(buf, PAGE_SIZE);
  91         if (buf == NULL)
  92                 GOTO(cleanup_bulk, rc = -ENOMEM);
  93
  94         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  95
  96         if (rc != PAGE_SIZE)
  97                 GOTO(cleanup_buf, rc = -EIO);
  98
  99         bulk->bp_xid = req->rq_xid;
 100         bulk->bp_buf = buf;
 101         bulk->bp_buflen = PAGE_SIZE;
 102         desc->bd_portal = MDS_BULK_PORTAL;
 103
 104         rc = ptlrpc_send_bulk(desc);
 105         if (rc)
 106                 GOTO(cleanup_buf, rc);
 107
 108         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 109                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 110                        OBD_FAIL_MDS_SENDPAGE, rc);
 111                 ptlrpc_abort_bulk(desc);
 112                 GOTO(cleanup_buf, rc);
 113         }
 114
 115         lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
 116         rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
 117         if (rc) {
 118                 if (rc != -ETIMEDOUT)
 119                         LBUG();
 120                 GOTO(cleanup_buf, rc);
 121         }
 122
 123         EXIT;
 124  cleanup_buf:
 125         OBD_FREE(buf, PAGE_SIZE);
 126  cleanup_bulk:
 127         ptlrpc_free_bulk(desc);
 128  out:
 129         return rc;
 130 }
 131
 132 /*
 133  * Look up a named entry in a directory, and get an LDLM lock on it.
 134  * 'dir' is a inode for which an LDLM lock has already been taken.
 135  *
 136  * If we do not need an exclusive or write lock on this entry (e.g.
 137  * a read lock for attribute lookup only) then we do not hold the
 138  * directory on return.  It is up to the caller to know what type
 139  * of lock it is getting, and clean up appropriately.
 140  */
 141 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 142                                       struct dentry *dir, struct vfsmount **mnt,
 143                                       char *name, int namelen, int lock_mode,
 144                                       struct lustre_handle *lockh,
 145                                       int dir_lock_mode)
 146 {
 147         struct dentry *dchild;
 148         int flags = 0, rc;
 149         __u64 res_id[3] = {0};
 150         ENTRY;
 151
 152         down(&dir->d_inode->i_sem);
 153         dchild = lookup_one_len(name, dir, namelen);
 154         if (IS_ERR(dchild)) {
 155                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 156                 up(&dir->d_inode->i_sem);
 157                 LBUG();
 158                 RETURN(dchild);
 159         }
 160         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 161                 up(&dir->d_inode->i_sem);
 162                 ldlm_lock_decref(lockh, dir_lock_mode);
 163         }
 164
 165         if (lock_mode == 0 || !dchild->d_inode)
 166                 RETURN(dchild);
 167
 168         res_id[0] = dchild->d_inode->i_ino;
 169         res_id[1] = dchild->d_inode->i_generation;
 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(dchild);
 176                 up(&dir->d_inode->i_sem);
 177                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 178         }
 179
 180         RETURN(dchild);
 181 }
 182
 183 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 184                                      struct vfsmount **mnt, int lock_mode,
 185                                      struct lustre_handle *lockh)
 186 {
 187         struct mds_obd *mds = &obd->u.mds;
 188         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 189         int flags = 0, rc;
 190         __u64 res_id[3] = {0};
 191         ENTRY;
 192
 193         if (IS_ERR(de))
 194                 RETURN(de);
 195
 196         res_id[0] = de->d_inode->i_ino;
 197         res_id[1] = de->d_inode->i_generation;
 198         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 199                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 200                                    &flags, ldlm_completion_ast,
 201                                    mds_blocking_ast, NULL, 0, lockh);
 202         if (rc != ELDLM_OK) {
 203                 l_dput(de);
 204                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 205         }
 206
 207         RETURN(retval);
 208 }
 209
 210 #ifndef DCACHE_DISCONNECTED
 211 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 212 #endif
 213
 214 /* Look up an entry by inode number. */
 215 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 216                               struct vfsmount **mnt)
 217 {
 218         /* stolen from NFS */
 219         struct super_block *sb = mds->mds_sb;
 220         unsigned long ino = fid->id;
 221         __u32 generation = fid->generation;
 222         struct inode *inode;
 223         struct list_head *lp;
 224         struct dentry *result;
 225
 226         if (ino == 0)
 227                 RETURN(ERR_PTR(-ESTALE));
 228
 229         inode = iget(sb, ino);
 230         if (inode == NULL)
 231                 RETURN(ERR_PTR(-ENOMEM));
 232
 233         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 234
 235         if (is_bad_inode(inode) ||
 236             (generation && inode->i_generation != generation)) {
 237                 /* we didn't find the right inode.. */
 238                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 239                        inode->i_ino, inode->i_nlink,
 240                        atomic_read(&inode->i_count), inode->i_generation,
 241                        generation);
 242                 LBUG();
 243                 iput(inode);
 244                 RETURN(ERR_PTR(-ESTALE));
 245         }
 246
 247         /* now to find a dentry. If possible, get a well-connected one */
 248         if (mnt)
 249                 *mnt = mds->mds_vfsmnt;
 250         spin_lock(&dcache_lock);
 251         list_for_each(lp, &inode->i_dentry) {
 252                 result = list_entry(lp, struct dentry, d_alias);
 253                 if (!(result->d_flags & DCACHE_DISCONNECTED)) {
 254                         dget_locked(result);
 255                         result->d_vfs_flags |= DCACHE_REFERENCED;
 256                         spin_unlock(&dcache_lock);
 257                         iput(inode);
 258                         if (mnt)
 259                                 mntget(*mnt);
 260                         return result;
 261                 }
 262         }
 263         spin_unlock(&dcache_lock);
 264         result = d_alloc_root(inode);
 265         if (result == NULL) {
 266                 iput(inode);
 267                 return ERR_PTR(-ENOMEM);
 268         }
 269         if (mnt)
 270                 mntget(*mnt);
 271         result->d_flags |= DCACHE_DISCONNECTED;
 272         return result;
 273 }
 274
 275 /* Establish a connection to the MDS.
 276  *
 277  * This will set up an export structure for the client to hold state data
 278  * about that client, like open files, the last operation number it did
 279  * on the server, etc.
 280  */
 281 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 282                        obd_uuid_t cluuid, struct recovd_obd *recovd,
 283                        ptlrpc_recovery_cb_t recover)
 284 {
 285         struct obd_export *exp;
 286         struct mds_export_data *med;
 287         struct mds_client_data *mcd;
 288         struct list_head *p;
 289         int rc;
 290         ENTRY;
 291
 292         if (!conn || !obd || !cluuid)
 293                 RETURN(-EINVAL);
 294
 295         MOD_INC_USE_COUNT;
 296
 297         spin_lock(&obd->obd_dev_lock);
 298         list_for_each(p, &obd->obd_exports) {
 299                 exp = list_entry(p, struct obd_export, exp_obd_chain);
 300                 mcd = exp->exp_mds_data.med_mcd;
 301                 if (!mcd) {
 302                         CERROR("FYI: NULL mcd - simultaneous connects\n");
 303                         continue;
 304                 }
 305                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 306                         LASSERT(exp->exp_obd == obd);
 307
 308                         if (!list_empty(&exp->exp_conn_chain)) {
 309                                 CERROR("existing uuid/export, list not empty!\n");
 310                                 spin_unlock(&obd->obd_dev_lock);
 311                                 MOD_DEC_USE_COUNT;
 312                                 RETURN(-EALREADY);
 313                         }
 314                         conn->addr = (__u64) (unsigned long)exp;
 315                         conn->cookie = exp->exp_cookie;
 316                         spin_unlock(&obd->obd_dev_lock);
 317                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 318                                cluuid, exp);
 319                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 320                                (long long)conn->addr, (long long)conn->cookie);
 321                         MOD_DEC_USE_COUNT;
 322                         RETURN(0);
 323                 }
 324         }
 325         spin_unlock(&obd->obd_dev_lock);
 326         /* XXX There is a small race between checking the list and adding a
 327          * new connection for the same UUID, but the real threat (list
 328          * corruption when multiple different clients connect) is solved.
 329          *
 330          * There is a second race between adding the export to the list,
 331          * and filling in the client data below.  Hence skipping the case
 332          * of NULL mcd above.  We should already be controlling multiple
 333          * connects at the client, and we can't hold the spinlock over
 334          * memory allocations without risk of deadlocking.
 335          */
 336         rc = class_connect(conn, obd, cluuid);
 337         if (rc)
 338                 GOTO(out_dec, rc);
 339         exp = class_conn2export(conn);
 340         LASSERT(exp);
 341         med = &exp->exp_mds_data;
 342
 343         OBD_ALLOC(mcd, sizeof(*mcd));
 344         if (!mcd) {
 345                 CERROR("mds: out of memory for client data\n");
 346                 GOTO(out_export, rc = -ENOMEM);
 347         }
 348
 349         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 350         med->med_mcd = mcd;
 351
 352         INIT_LIST_HEAD(&med->med_open_head);
 353         spin_lock_init(&med->med_open_lock);
 354
 355         rc = mds_client_add(med, -1);
 356         if (rc)
 357                 GOTO(out_mcd, rc);
 358
 359         RETURN(0);
 360
 361 out_mcd:
 362         OBD_FREE(mcd, sizeof(*mcd));
 363 out_export:
 364         class_disconnect(conn);
 365 out_dec:
 366         MOD_DEC_USE_COUNT;
 367
 368         return rc;
 369 }
 370
 371 /* Call with med->med_open_lock held, please. */
 372 inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
 373 {
 374         struct file *file = mfd->mfd_file;
 375         LASSERT(file->private_data == mfd);
 376
 377         list_del(&mfd->mfd_list);
 378         mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
 379         kmem_cache_free(mds_file_cache, mfd);
 380
 381         return filp_close(file, 0);
 382 }
 383
 384 static int mds_disconnect(struct lustre_handle *conn)
 385 {
 386         struct obd_export *export = class_conn2export(conn);
 387         struct list_head *tmp, *n;
 388         struct mds_export_data *med = &export->exp_mds_data;
 389         int rc;
 390         ENTRY;
 391
 392         /*
 393          * Close any open files.
 394          */
 395         spin_lock(&med->med_open_lock);
 396         list_for_each_safe(tmp, n, &med->med_open_head) {
 397                 struct mds_file_data *mfd =
 398                         list_entry(tmp, struct mds_file_data, mfd_list);
 399                 rc = mds_close_mfd(mfd, med);
 400                 if (rc) {
 401                         /* XXX better diagnostics, with file path and stuff */
 402                         CDEBUG(D_INODE, "Error %d closing mfd %p\n", rc, mfd);
 403                 }
 404         }
 405         spin_unlock(&med->med_open_lock);
 406
 407         ldlm_cancel_locks_for_export(export);
 408         mds_client_free(export);
 409
 410         rc = class_disconnect(conn);
 411         if (!rc)
 412                 MOD_DEC_USE_COUNT;
 413
 414         RETURN(rc);
 415 }
 416
 417 /*
 418  * XXX This is NOT guaranteed to flush all transactions to disk (even though
 419  *     it is equivalent to calling sync()) because it only _starts_ the flush
 420  *     and does not wait for completion.  It's better than nothing though.
 421  *     What we really want is a mild form of fsync_dev_lockfs(), but it is
 422  *     non-standard, or enabling do_sync_supers in ext3, just for this call.
 423  */
 424 static void mds_fsync_super(struct super_block *sb)
 425 {
 426         lock_kernel();
 427         lock_super(sb);
 428         if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
 429                 sb->s_op->write_super(sb);
 430         unlock_super(sb);
 431         unlock_kernel();
 432 }
 433
 434 static int mds_getstatus(struct ptlrpc_request *req)
 435 {
 436         struct mds_obd *mds = mds_req2mds(req);
 437         struct mds_body *body;
 438         int rc, size = sizeof(*body);
 439         ENTRY;
 440
 441         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 442         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 443                 CERROR("mds: out of memory for message: size=%d\n", size);
 444                 req->rq_status = -ENOMEM;
 445                 RETURN(0);
 446         }
 447
 448         /* Flush any outstanding transactions to disk so the client will
 449          * get the latest last_committed value and can drop their local
 450          * requests if they have any.  This would be fsync_super() if it
 451          * was exported.
 452          */
 453         mds_fsync_super(mds->mds_sb);
 454
 455         body = lustre_msg_buf(req->rq_repmsg, 0);
 456         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 457
 458         /* the last_committed and last_xid fields are filled in for all
 459          * replies already - no need to do so here also.
 460          */
 461         RETURN(0);
 462 }
 463
 464 static int mds_getlovinfo(struct ptlrpc_request *req)
 465 {
 466         struct mds_obd *mds = mds_req2mds(req);
 467         struct mds_status_req *streq;
 468         struct lov_desc *desc;
 469         int tgt_count;
 470         int rc, size[2] = {sizeof(*desc)};
 471         ENTRY;
 472
 473         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 474         streq->flags = NTOH__u32(streq->flags);
 475         streq->repbuf = NTOH__u32(streq->repbuf);
 476         size[1] = streq->repbuf;
 477
 478         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 479         if (rc) {
 480                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 481                 req->rq_status = -ENOMEM;
 482                 RETURN(0);
 483         }
 484
 485         desc = lustre_msg_buf(req->rq_repmsg, 0);
 486         rc = mds_get_lovdesc(mds, desc);
 487         if (rc) {
 488                 CERROR("mds_get_lovdesc error %d", rc);
 489                 req->rq_status = rc;
 490                 RETURN(0);
 491         }
 492
 493         tgt_count = le32_to_cpu(desc->ld_tgt_count);
 494         if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
 495                 CERROR("too many targets, enlarge client buffers\n");
 496                 req->rq_status = -ENOSPC;
 497                 RETURN(0);
 498         }
 499
 500         /* XXX the MDS should not really know about this */
 501         mds->mds_max_mdsize = lov_mds_md_size(tgt_count);
 502         rc = mds_get_lovtgts(mds, tgt_count,
 503                              lustre_msg_buf(req->rq_repmsg, 1));
 504         if (rc) {
 505                 CERROR("get_lovtgts error %d\n", rc);
 506                 req->rq_status = rc;
 507                 RETURN(0);
 508         }
 509         RETURN(0);
 510 }
 511
 512 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 513                      void *data, __u32 data_len, int flag)
 514 {
 515         int do_ast;
 516         ENTRY;
 517
 518         if (flag == LDLM_CB_CANCELING) {
 519                 /* Don't need to do anything here. */
 520                 RETURN(0);
 521         }
 522
 523         /* XXX layering violation!  -phil */
 524         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 525         lock->l_flags |= LDLM_FL_CBPENDING;
 526         do_ast = (!lock->l_readers && !lock->l_writers);
 527         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 528
 529         if (do_ast) {
 530                 struct lustre_handle lockh;
 531                 int rc;
 532
 533                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 534                 ldlm_lock2handle(lock, &lockh);
 535                 rc = ldlm_cli_cancel(&lockh);
 536                 if (rc < 0)
 537                         CERROR("ldlm_cli_cancel: %d\n", rc);
 538         } else
 539                 LDLM_DEBUG(lock, "Lock still has references, will be"
 540                            "cancelled later");
 541         RETURN(0);
 542 }
 543
 544 int mds_pack_md(struct mds_obd *mds, struct ptlrpc_request *req,
 545                 int offset, struct mds_body *body, struct inode *inode)
 546 {
 547         struct lov_mds_md *lmm;
 548         int lmm_size = req->rq_repmsg->buflens[offset];
 549         int rc;
 550
 551         if (lmm_size == 0) {
 552                 CERROR("no space reserved for inode %u MD\n", inode->i_ino);
 553                 RETURN(0);
 554         }
 555
 556         lmm = lustre_msg_buf(req->rq_repmsg, offset);
 557
 558         /* I don't really like this, but it is a sanity check on the client
 559          * MD request.  However, if the client doesn't know how much space
 560          * to reserve for the MD, this shouldn't be fatal either...
 561          */
 562         if (lmm_size > mds->mds_max_mdsize) {
 563                 CERROR("Reading MD for inode %u of %d bytes > max %d\n",
 564                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 565                 // RETURN(-EINVAL);
 566         }
 567
 568         /* We don't need to store the reply size, because this buffer is
 569          * discarded right after unpacking, and the LOV can figure out the
 570          * size itself from the ost count.
 571          */
 572         if ((rc = mds_fs_get_md(mds, inode, lmm, lmm_size)) < 0) {
 573                 CDEBUG(D_INFO, "No md for ino %u: rc = %d\n", inode->i_ino, rc);
 574         } else {
 575                 body->valid |= OBD_MD_FLEASIZE;
 576                 rc = 0;
 577         }
 578
 579         return rc;
 580 }
 581
 582 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 583                                 struct ptlrpc_request *req,
 584                                 struct mds_body *reqbody, int reply_off)
 585 {
 586         struct mds_body *body;
 587         struct inode *inode = dentry->d_inode;
 588         int rc = 0;
 589         ENTRY;
 590
 591         if (inode == NULL)
 592                 RETURN(-ENOENT);
 593
 594         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 595
 596         mds_pack_inode2fid(&body->fid1, inode);
 597         mds_pack_inode2body(body, inode);
 598
 599         if (S_ISREG(inode->i_mode)) {
 600                 rc = mds_pack_md(mds, req, reply_off + 1, body, inode);
 601         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 602                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 603                 int len = req->rq_repmsg->buflens[reply_off + 1];
 604
 605                 rc = inode->i_op->readlink(dentry, symname, len);
 606                 if (rc < 0) {
 607                         CERROR("readlink failed: %d\n", rc);
 608                 } else {
 609                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 610                         body->valid |= OBD_MD_LINKNAME;
 611                 }
 612         }
 613         RETURN(rc);
 614 }
 615
 616 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 617 {
 618         struct mds_obd *mds = mds_req2mds(req);
 619         struct obd_device *obd = req->rq_export->exp_obd;
 620         struct obd_run_ctxt saved;
 621         struct mds_body *body;
 622         struct dentry *de = NULL, *dchild = NULL;
 623         struct inode *dir;
 624         struct lustre_handle lockh;
 625         char *name;
 626         int namelen, flags = 0, lock_mode, rc = 0;
 627         struct obd_ucred uc;
 628         __u64 res_id[3] = {0, 0, 0};
 629         ENTRY;
 630
 631         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 632
 633         if (req->rq_reqmsg->bufcount <= offset + 1) {
 634                 LBUG();
 635                 GOTO(out_pre_de, rc = -EINVAL);
 636         }
 637
 638         body = lustre_msg_buf(req->rq_reqmsg, offset);
 639         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 640         namelen = req->rq_reqmsg->buflens[offset + 1];
 641         /* requests were at offset 2, replies go back at 1 */
 642         if (offset)
 643                 offset = 1;
 644
 645         uc.ouc_fsuid = body->fsuid;
 646         uc.ouc_fsgid = body->fsgid;
 647         uc.ouc_cap = body->capability;
 648         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 649         de = mds_fid2dentry(mds, &body->fid1, NULL);
 650         if (IS_ERR(de)) {
 651                 LBUG();
 652                 GOTO(out_pre_de, rc = -ESTALE);
 653         }
 654
 655         dir = de->d_inode;
 656         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 657
 658         lock_mode = LCK_PR;
 659         res_id[0] = dir->i_ino;
 660         res_id[1] = dir->i_generation;
 661
 662         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 663                              NULL, 0, lock_mode, &lockh);
 664         if (rc == 0) {
 665                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 666                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 667                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 668                                       &flags, ldlm_completion_ast,
 669                                       mds_blocking_ast, NULL, 0, &lockh);
 670                 if (rc != ELDLM_OK) {
 671                         CERROR("lock enqueue: err: %d\n", rc);
 672                         GOTO(out_create_de, rc = -EIO);
 673                 }
 674         }
 675         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 676
 677         down(&dir->i_sem);
 678         dchild = lookup_one_len(name, de, namelen - 1);
 679         if (IS_ERR(dchild)) {
 680                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 681                 up(&dir->i_sem);
 682                 GOTO(out_create_dchild, rc = PTR_ERR(dchild));
 683         }
 684
 685         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 686
 687         EXIT;
 688 out_create_dchild:
 689         l_dput(dchild);
 690         up(&dir->i_sem);
 691         ldlm_lock_decref(&lockh, lock_mode);
 692 out_create_de:
 693         l_dput(de);
 694 out_pre_de:
 695         req->rq_status = rc;
 696         pop_ctxt(&saved);
 697         return 0;
 698 }
 699
 700 static int mds_getattr(int offset, struct ptlrpc_request *req)
 701 {
 702         struct mds_obd *mds = mds_req2mds(req);
 703         struct obd_run_ctxt saved;
 704         struct dentry *de;
 705         struct inode *inode;
 706         struct mds_body *body;
 707         struct obd_ucred uc;
 708         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 709         ENTRY;
 710
 711         body = lustre_msg_buf(req->rq_reqmsg, offset);
 712         uc.ouc_fsuid = body->fsuid;
 713         uc.ouc_fsgid = body->fsgid;
 714         uc.ouc_cap = body->capability;
 715         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 716         de = mds_fid2dentry(mds, &body->fid1, NULL);
 717         if (IS_ERR(de)) {
 718                 rc = req->rq_status = -ENOENT;
 719                 GOTO(out_pop, PTR_ERR(de));
 720         }
 721
 722         inode = de->d_inode;
 723         if (S_ISREG(body->fid1.f_type)) {
 724                 int rc = mds_fs_get_md(mds, inode, NULL, 0);
 725                 CDEBUG(D_INODE, "got %d bytes MD data for inode %u\n",
 726                        rc, inode->i_ino);
 727                 if (rc < 0) {
 728                         if (rc != -ENODATA)
 729                                 CERROR("error getting inode %u MD: rc = %d\n",
 730                                        inode->i_ino, rc);
 731                         size[bufcount] = 0;
 732                 } else if (rc < mds->mds_max_mdsize) {
 733                         size[bufcount] = 0;
 734                         CERROR("MD size %d larger than maximum possible %u\n",
 735                                rc, mds->mds_max_mdsize);
 736                 } else
 737                         size[bufcount] = rc;
 738                 bufcount++;
 739         } else if (body->valid & OBD_MD_LINKNAME) {
 740                 size[bufcount] = MIN(inode->i_size + 1, body->size);
 741                 bufcount++;
 742                 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
 743                        inode->i_size + 1, body->size);
 744         }
 745
 746         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 747                 CERROR("failed MDS_GETATTR_PACK test\n");
 748                 req->rq_status = -ENOMEM;
 749                 GOTO(out, rc = -ENOMEM);
 750         }
 751
 752         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 753                              &req->rq_repmsg);
 754         if (rc) {
 755                 CERROR("out of memoryK\n");
 756                 req->rq_status = rc;
 757                 GOTO(out, rc);
 758         }
 759
 760         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 761
 762 out:
 763         l_dput(de);
 764 out_pop:
 765         pop_ctxt(&saved);
 766         RETURN(rc);
 767 }
 768
 769 static int mds_statfs(struct ptlrpc_request *req)
 770 {
 771         struct mds_obd *mds = mds_req2mds(req);
 772         struct obd_statfs *osfs;
 773         struct statfs sfs;
 774         int rc, size = sizeof(*osfs);
 775         ENTRY;
 776
 777         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 778         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 779                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 780                 GOTO(out, rc);
 781         }
 782
 783         rc = mds_fs_statfs(mds, &sfs);
 784         if (rc) {
 785                 CERROR("mds: statfs failed: rc %d\n", rc);
 786                 GOTO(out, rc);
 787         }
 788         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 789         memset(osfs, 0, size);
 790         statfs_pack(osfs, &sfs);
 791         obd_statfs_pack(osfs, osfs);
 792
 793 out:
 794         req->rq_status = rc;
 795         RETURN(0);
 796 }
 797
 798 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
 799 {
 800         struct mds_file_data *mfd = NULL;
 801
 802         if (!handle || !handle->addr)
 803                 RETURN(NULL);
 804
 805         mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
 806         if (!kmem_cache_validate(mds_file_cache, mfd))
 807                 RETURN(NULL);
 808
 809         if (mfd->mfd_servercookie != handle->cookie)
 810                 RETURN(NULL);
 811
 812         return mfd;
 813 }
 814
 815 static int mds_store_md(struct mds_obd *mds, struct ptlrpc_request *req,
 816                         int offset, struct mds_body *body, struct inode *inode)
 817 {
 818         struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, offset);
 819         int lmm_size = req->rq_reqmsg->buflens[offset];
 820         struct obd_run_ctxt saved;
 821         struct obd_ucred uc;
 822         void *handle;
 823         int rc, rc2;
 824         ENTRY;
 825
 826         /* I don't really like this, but it is a sanity check on the client
 827          * MD request.
 828          */
 829         if (lmm_size > mds->mds_max_mdsize) {
 830                 CERROR("Saving MD for inode %u of %d bytes > max %d\n",
 831                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 832                 //RETURN(-EINVAL);
 833         }
 834
 835         CDEBUG(D_INODE, "storing %d bytes MD for inode %u\n",
 836                lmm_size, inode->i_ino);
 837         uc.ouc_fsuid = body->fsuid;
 838         uc.ouc_fsgid = body->fsgid;
 839         uc.ouc_cap = body->capability;
 840         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 841         handle = mds_fs_start(mds, inode, MDS_FSOP_SETATTR);
 842         if (!handle)
 843                 GOTO(out_ea, rc = -ENOMEM);
 844
 845         rc = mds_fs_set_md(mds, inode, handle, lmm, lmm_size);
 846         if (!rc)
 847                 rc = mds_update_last_rcvd(mds, handle, req);
 848
 849         rc2 = mds_fs_commit(mds, inode, handle);
 850         if (rc2 && !rc)
 851                 rc = rc2;
 852 out_ea:
 853         pop_ctxt(&saved);
 854
 855         RETURN(rc);
 856 }
 857
 858 static int mds_open(struct ptlrpc_request *req)
 859 {
 860         struct mds_obd *mds = mds_req2mds(req);
 861         struct mds_body *body;
 862         struct mds_export_data *med;
 863         struct mds_file_data *mfd;
 864         struct dentry *de;
 865         struct file *file;
 866         struct vfsmount *mnt;
 867         __u32 flags;
 868         struct list_head *tmp;
 869         int rc, size = sizeof(*body);
 870         ENTRY;
 871
 872         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 873                 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
 874                 req->rq_status = -ENOMEM;
 875                 RETURN(-ENOMEM);
 876         }
 877
 878         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 879         if (rc) {
 880                 CERROR("mds: pack error: rc = %d\n", rc);
 881                 req->rq_status = rc;
 882                 RETURN(rc);
 883         }
 884
 885         body = lustre_msg_buf(req->rq_reqmsg, 0);
 886
 887         /* was this animal open already and the client lost the reply? */
 888         /* XXX need some way to detect a reopen, to avoid locked list walks */
 889         med = &req->rq_export->exp_mds_data;
 890         spin_lock(&med->med_open_lock);
 891         list_for_each(tmp, &med->med_open_head) {
 892                 mfd = list_entry(tmp, typeof(*mfd), mfd_list);
 893                 if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
 894                             sizeof(mfd->mfd_clienthandle)) &&
 895                     body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
 896                         de = mfd->mfd_file->f_dentry;
 897                         spin_unlock(&med->med_open_lock);
 898                         CERROR("Re opening "LPD64"\n", body->fid1.id);
 899                         GOTO(out_pack, rc = 0);
 900                 }
 901         }
 902         spin_unlock(&med->med_open_lock);
 903
 904         mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
 905         if (!mfd) {
 906                 CERROR("mds: out of memory\n");
 907                 req->rq_status = -ENOMEM;
 908                 RETURN(0);
 909         }
 910
 911         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 912         if (IS_ERR(de))
 913                 GOTO(out_free, rc = PTR_ERR(de));
 914
 915         /* check if this inode has seen a delayed object creation */
 916         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MDS_OPEN_HAS_EA) {
 917                 rc = mds_store_md(mds, req, 1, body, de->d_inode);
 918                 if (rc) {
 919                         l_dput(de);
 920                         mntput(mnt);
 921                         GOTO(out_free, rc);
 922                 }
 923         }
 924
 925         flags = body->flags;
 926         /* dentry_open does a dput(de) and mntput(mnt) on error */
 927         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 928         if (IS_ERR(file)) {
 929                 rc = PTR_ERR(file);
 930                 GOTO(out_free, 0);
 931         }
 932
 933         file->private_data = mfd;
 934         mfd->mfd_file = file;
 935         memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
 936         get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
 937         spin_lock(&med->med_open_lock);
 938         list_add(&mfd->mfd_list, &med->med_open_head);
 939         spin_unlock(&med->med_open_lock);
 940
 941 out_pack:
 942         body = lustre_msg_buf(req->rq_repmsg, 0);
 943         mds_pack_inode2fid(&body->fid1, de->d_inode);
 944         mds_pack_inode2body(body, de->d_inode);
 945         body->handle.addr = (__u64)(unsigned long)mfd;
 946         body->handle.cookie = mfd->mfd_servercookie;
 947         CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
 948                mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
 949         RETURN(0);
 950
 951 out_free:
 952         mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
 953         kmem_cache_free(mds_file_cache, mfd);
 954         req->rq_status = rc;
 955         RETURN(0);
 956 }
 957
 958 static int mds_close(struct ptlrpc_request *req)
 959 {
 960         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 961         struct mds_body *body;
 962         struct mds_file_data *mfd;
 963         int rc;
 964         ENTRY;
 965
 966         body = lustre_msg_buf(req->rq_reqmsg, 0);
 967
 968         mfd = mds_handle2mfd(&body->handle);
 969         if (!mfd) {
 970                 CERROR("no handle for file close "LPD64
 971                        ": addr "LPX64", cookie "LPX64"\n",
 972                        body->fid1.id, body->handle.addr, body->handle.cookie);
 973                 RETURN(-ESTALE);
 974         }
 975
 976         spin_lock(&med->med_open_lock);
 977         req->rq_status = mds_close_mfd(mfd, med);
 978         spin_unlock(&med->med_open_lock);
 979
 980         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 981                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
 982                 req->rq_status = -ENOMEM;
 983                 RETURN(-ENOMEM);
 984         }
 985
 986         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 987         if (rc) {
 988                 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
 989                 req->rq_status = rc;
 990         }
 991
 992         RETURN(0);
 993 }
 994
 995 static int mds_readpage(struct ptlrpc_request *req)
 996 {
 997         struct mds_obd *mds = mds_req2mds(req);
 998         struct vfsmount *mnt;
 999         struct dentry *de;
1000         struct file *file;
1001         struct mds_body *body, *repbody;
1002         struct obd_run_ctxt saved;
1003         int rc, size = sizeof(*body);
1004         struct obd_ucred uc;
1005         ENTRY;
1006
1007         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
1008         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1009                 CERROR("mds: out of memory\n");
1010                 GOTO(out, rc = -ENOMEM);
1011         }
1012
1013         body = lustre_msg_buf(req->rq_reqmsg, 0);
1014         uc.ouc_fsuid = body->fsuid;
1015         uc.ouc_fsgid = body->fsgid;
1016         uc.ouc_cap = body->capability;
1017         push_ctxt(&saved, &mds->mds_ctxt, &uc);
1018         de = mds_fid2dentry(mds, &body->fid1, &mnt);
1019         if (IS_ERR(de))
1020                 GOTO(out_pop, rc = PTR_ERR(de));
1021
1022         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
1023
1024         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1025         /* note: in case of an error, dentry_open puts dentry */
1026         if (IS_ERR(file))
1027                 GOTO(out_pop, rc = PTR_ERR(file));
1028
1029         repbody = lustre_msg_buf(req->rq_repmsg, 0);
1030         repbody->size = file->f_dentry->d_inode->i_size;
1031         repbody->valid = OBD_MD_FLSIZE;
1032
1033         /* to make this asynchronous make sure that the handling function
1034            doesn't send a reply when this function completes. Instead a
1035            callback function would send the reply */
1036         /* note: in case of an error, dentry_open puts dentry */
1037         rc = mds_sendpage(req, file, body->size);
1038
1039         filp_close(file, 0);
1040 out_pop:
1041         pop_ctxt(&saved);
1042 out:
1043         req->rq_status = rc;
1044         RETURN(0);
1045 }
1046
1047 int mds_reint(struct ptlrpc_request *req, int offset)
1048 {
1049         int rc;
1050         struct mds_update_record rec;
1051
1052         rc = mds_update_unpack(req, offset, &rec);
1053         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1054                 CERROR("invalid record\n");
1055                 req->rq_status = -EINVAL;
1056                 RETURN(0);
1057         }
1058         /* rc will be used to interrupt a for loop over multiple records */
1059         rc = mds_reint_rec(&rec, offset, req);
1060         return rc;
1061 }
1062
1063 int mds_handle(struct ptlrpc_request *req)
1064 {
1065         int rc;
1066         ENTRY;
1067
1068         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1069         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
1070                 CERROR("lustre_mds: Invalid request\n");
1071                 GOTO(out, rc);
1072         }
1073
1074         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
1075                 GOTO(out, rc = -ENOTCONN);
1076
1077         LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
1078
1079         switch (req->rq_reqmsg->opc) {
1080         case MDS_CONNECT:
1081                 CDEBUG(D_INODE, "connect\n");
1082                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1083                 rc = target_handle_connect(req);
1084                 break;
1085
1086         case MDS_DISCONNECT:
1087                 CDEBUG(D_INODE, "disconnect\n");
1088                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1089                 rc = target_handle_disconnect(req);
1090                 goto out;
1091
1092         case MDS_GETSTATUS:
1093                 CDEBUG(D_INODE, "getstatus\n");
1094                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1095                 rc = mds_getstatus(req);
1096                 break;
1097
1098         case MDS_GETLOVINFO:
1099                 CDEBUG(D_INODE, "getlovinfo\n");
1100                 rc = mds_getlovinfo(req);
1101                 break;
1102
1103         case MDS_GETATTR:
1104                 CDEBUG(D_INODE, "getattr\n");
1105                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1106                 rc = mds_getattr(0, req);
1107                 break;
1108
1109         case MDS_STATFS:
1110                 CDEBUG(D_INODE, "statfs\n");
1111                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1112                 rc = mds_statfs(req);
1113                 break;
1114
1115         case MDS_READPAGE:
1116                 CDEBUG(D_INODE, "readpage\n");
1117                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1118                 rc = mds_readpage(req);
1119
1120                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1121                         return 0;
1122                 break;
1123
1124         case MDS_REINT: {
1125                 int size = sizeof(struct mds_body);
1126                 CDEBUG(D_INODE, "reint\n");
1127                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1128
1129                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1130                                      &req->rq_repmsg);
1131                 if (rc) {
1132                         req->rq_status = rc;
1133                         break;
1134                 }
1135                 rc = mds_reint(req, 0);
1136                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1137                 break;
1138                 }
1139
1140         case MDS_OPEN:
1141                 CDEBUG(D_INODE, "open\n");
1142                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
1143                 rc = mds_open(req);
1144                 break;
1145
1146         case MDS_CLOSE:
1147                 CDEBUG(D_INODE, "close\n");
1148                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1149                 rc = mds_close(req);
1150                 break;
1151
1152         case LDLM_ENQUEUE:
1153                 CDEBUG(D_INODE, "enqueue\n");
1154                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1155                 rc = ldlm_handle_enqueue(req);
1156                 break;
1157         case LDLM_CONVERT:
1158                 CDEBUG(D_INODE, "convert\n");
1159                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1160                 rc = ldlm_handle_convert(req);
1161                 break;
1162         case LDLM_BL_CALLBACK:
1163         case LDLM_CP_CALLBACK:
1164                 CDEBUG(D_INODE, "callback\n");
1165                 CERROR("callbacks should not happen on MDS\n");
1166                 LBUG();
1167                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1168                 break;
1169         default:
1170                 rc = ptlrpc_error(req->rq_svc, req);
1171                 RETURN(rc);
1172         }
1173
1174         EXIT;
1175
1176         if (!rc) {
1177                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1178                 struct mds_obd *mds = mds_req2mds(req);
1179
1180                 req->rq_repmsg->last_xid =
1181                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1182                 req->rq_repmsg->last_committed =
1183                         HTON__u64(mds->mds_last_committed);
1184                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1185                        (unsigned long long)mds->mds_last_rcvd,
1186                        (unsigned long long)mds->mds_last_committed,
1187                        cpu_to_le32(req->rq_xid));
1188         }
1189  out:
1190         if (rc) {
1191                 CERROR("mds: processing error (opcode %d): %d\n",
1192                        req->rq_reqmsg->opc, rc);
1193                 ptlrpc_error(req->rq_svc, req);
1194         } else {
1195                 CDEBUG(D_NET, "sending reply\n");
1196                 ptlrpc_reply(req->rq_svc, req);
1197         }
1198         return 0;
1199 }
1200
1201 /* Update the server data on disk.  This stores the new mount_count and
1202  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1203  * then the server last_rcvd value may be less than that of the clients.
1204  * This will alert us that we may need to do client recovery.
1205  *
1206  * Assumes we are already in the server filesystem context.
1207  *
1208  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1209  */
1210 static
1211 int mds_update_server_data(struct mds_obd *mds)
1212 {
1213         struct mds_server_data *msd = mds->mds_server_data;
1214         struct file *filp = mds->mds_rcvd_filp;
1215         loff_t off = 0;
1216         int rc;
1217
1218         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1219         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1220
1221         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1222                (unsigned long long)mds->mds_mount_count,
1223                (unsigned long long)mds->mds_last_rcvd);
1224         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1225         if (rc != sizeof(*msd)) {
1226                 CERROR("error writing MDS server data: rc = %d\n", rc);
1227                 if (rc > 0)
1228                         RETURN(-EIO);
1229                 RETURN(rc);
1230         }
1231 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1232         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1233 #else
1234         rc = file_fsync(filp,  filp->f_dentry, 1);
1235 #endif
1236         if (rc)
1237                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1238
1239         return 0;
1240 }
1241
1242 /* Do recovery actions for the MDS */
1243 static int mds_recover(struct obd_device *obddev)
1244 {
1245         struct mds_obd *mds = &obddev->u.mds;
1246         struct obd_run_ctxt saved;
1247         int rc;
1248
1249         /* This happens at the end when recovery is complete */
1250         ++mds->mds_mount_count;
1251         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1252         rc = mds_update_server_data(mds);
1253         pop_ctxt(&saved);
1254
1255         return rc;
1256 }
1257
1258 /* mount the file system (secretly) */
1259 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1260 {
1261         struct obd_ioctl_data* data = buf;
1262         struct mds_obd *mds = &obddev->u.mds;
1263         struct vfsmount *mnt;
1264         int rc = 0;
1265         ENTRY;
1266
1267         MOD_INC_USE_COUNT;
1268 #ifdef CONFIG_DEV_RDONLY
1269         dev_clear_rdonly(2);
1270 #endif
1271         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1272                 GOTO(err_dec, rc = -EINVAL);
1273
1274         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1275
1276         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1277         if (IS_ERR(mnt)) {
1278                 rc = PTR_ERR(mnt);
1279                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1280                 GOTO(err_kfree, rc);
1281         }
1282
1283         CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, mnt);
1284         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1285         if (!mds->mds_sb)
1286                 GOTO(err_put, rc = -ENODEV);
1287
1288         spin_lock_init(&mds->mds_last_lock);
1289         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1290         rc = mds_fs_setup(obddev, mnt);
1291         if (rc) {
1292                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1293                 GOTO(err_put, rc);
1294         }
1295
1296         obddev->obd_namespace =
1297                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1298         if (obddev->obd_namespace == NULL) {
1299                 mds_cleanup(obddev);
1300                 GOTO(err_fs, rc = -ENOMEM);
1301         }
1302
1303
1304         rc = mds_recover(obddev);
1305         if (rc)
1306                 GOTO(err_fs, rc);
1307
1308         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1309                            "mds_ldlm_client", &obddev->obd_ldlm_client);
1310
1311         RETURN(0);
1312
1313 err_fs:
1314         mds_fs_cleanup(obddev);
1315 err_put:
1316         unlock_kernel();
1317         mntput(mds->mds_vfsmnt);
1318         mds->mds_sb = 0;
1319         lock_kernel();
1320 err_kfree:
1321         kfree(mds->mds_fstype);
1322 err_dec:
1323         MOD_DEC_USE_COUNT;
1324         RETURN(rc);
1325 }
1326
1327 static int mds_cleanup(struct obd_device *obddev)
1328 {
1329         struct super_block *sb;
1330         struct mds_obd *mds = &obddev->u.mds;
1331         struct obd_run_ctxt saved;
1332         ENTRY;
1333
1334         sb = mds->mds_sb;
1335         if (!mds->mds_sb)
1336                 RETURN(0);
1337
1338         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1339         mds_update_server_data(mds);
1340
1341         if (mds->mds_rcvd_filp) {
1342                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1343                 mds->mds_rcvd_filp = NULL;
1344
1345                 if (rc)
1346                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1347         }
1348         pop_ctxt(&saved);
1349
1350         unlock_kernel();
1351         mntput(mds->mds_vfsmnt);
1352         mds->mds_sb = 0;
1353         kfree(mds->mds_fstype);
1354
1355         ldlm_namespace_free(obddev->obd_namespace);
1356
1357         lock_kernel();
1358 #ifdef CONFIG_DEV_RDONLY
1359         dev_clear_rdonly(2);
1360 #endif
1361         mds_fs_cleanup(obddev);
1362
1363         MOD_DEC_USE_COUNT;
1364         RETURN(0);
1365 }
1366
1367 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1368                               ldlm_mode_t mode, int flags, void *data)
1369 {
1370         struct ptlrpc_request *req = req_cookie;
1371         int rc = 0;
1372         ENTRY;
1373
1374         if (!req_cookie)
1375                 RETURN(0);
1376
1377         if (req->rq_reqmsg->bufcount > 1) {
1378                 /* an intent needs to be considered */
1379                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1380                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1381                 struct mds_body *mds_rep;
1382                 struct ldlm_reply *rep;
1383                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1384                 int rc, size[3] = {sizeof(struct ldlm_reply),
1385                                                   sizeof(struct mds_body),
1386                                                   mds->mds_max_mdsize};
1387
1388                 it->opc = NTOH__u64(it->opc);
1389
1390                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1391                            ldlm_it2str(it->opc));
1392
1393                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1394                                      &req->rq_repmsg);
1395                 if (rc) {
1396                         rc = req->rq_status = -ENOMEM;
1397                         RETURN(rc);
1398                 }
1399
1400                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1401                 rep->lock_policy_res1 = 1;
1402
1403                 /* execute policy */
1404                 switch ((long)it->opc) {
1405                 case IT_CREAT|IT_OPEN:
1406                         rc = mds_reint(req, 2);
1407                         if (rc || (req->rq_status != 0 &&
1408                                    req->rq_status != -EEXIST)) {
1409                                 rep->lock_policy_res2 = req->rq_status;
1410                                 RETURN(ELDLM_LOCK_ABORTED);
1411                         }
1412                         break;
1413                 case IT_CREAT:
1414                 case IT_MKDIR:
1415                 case IT_MKNOD:
1416                 case IT_RENAME2:
1417                 case IT_LINK2:
1418                 case IT_RMDIR:
1419                 case IT_SYMLINK:
1420                 case IT_UNLINK:
1421                         rc = mds_reint(req, 2);
1422                         if (rc || (req->rq_status != 0 &&
1423                                    req->rq_status != -EISDIR &&
1424                                    req->rq_status != -ENOTDIR)) {
1425                                 rep->lock_policy_res2 = req->rq_status;
1426                                 RETURN(ELDLM_LOCK_ABORTED);
1427                         }
1428                         break;
1429                 case IT_GETATTR:
1430                 case IT_LOOKUP:
1431                 case IT_OPEN:
1432                 case IT_READDIR:
1433                 case IT_READLINK:
1434                 case IT_RENAME:
1435                 case IT_LINK:
1436                 case IT_SETATTR:
1437                         rc = mds_getattr_name(2, req);
1438                         /* FIXME: we need to sit down and decide on who should
1439                          * set req->rq_status, who should return negative and
1440                          * positive return values, and what they all mean. */
1441                         if (rc || req->rq_status != 0) {
1442                                 rep->lock_policy_res2 = req->rq_status;
1443                                 RETURN(ELDLM_LOCK_ABORTED);
1444                         }
1445                         break;
1446                 case IT_READDIR|IT_OPEN:
1447                         LBUG();
1448                         break;
1449                 default:
1450                         CERROR("Unhandled intent "LPD64"\n", it->opc);
1451                         LBUG();
1452                 }
1453
1454                 /* We don't bother returning a lock to the client for a file
1455                  * or directory we are removing.
1456                  *
1457                  * As for link and rename, there is no reason for the client
1458                  * to get a lock on the target at this point.  If they are
1459                  * going to modify the file/directory later they will get a
1460                  * lock at that time.
1461                  */
1462                 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1463                                IT_RENAME | IT_RENAME2))
1464                         RETURN(ELDLM_LOCK_ABORTED);
1465
1466                 rep->lock_policy_res2 = req->rq_status;
1467                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1468
1469                 /* If the client is about to open a file that doesn't have an MD
1470                  * stripe record, it's going to need a write lock. */
1471                 if (it->opc & IT_OPEN &&
1472                     !(lustre_msg_get_op_flags(req->rq_reqmsg)&MDS_OPEN_HAS_EA)){
1473                         LDLM_DEBUG(lock, "open with no EA; returning PW lock");
1474                         lock->l_req_mode = LCK_PW;
1475                 }
1476
1477                 if (flags & LDLM_FL_INTENT_ONLY) {
1478                         LDLM_DEBUG(lock, "INTENT_ONLY, aborting lock");
1479                         RETURN(ELDLM_LOCK_ABORTED);
1480                 }
1481                 /* Give the client a lock on the child object, instead of the
1482                  * parent that it requested. */
1483                 new_resid[0] = NTOH__u32(mds_rep->ino);
1484                 new_resid[1] = NTOH__u32(mds_rep->generation);
1485                 if (new_resid[0] == 0)
1486                         LBUG();
1487                 old_res = lock->l_resource->lr_name[0];
1488
1489                 ldlm_lock_change_resource(lock, new_resid);
1490                 if (lock->l_resource == NULL) {
1491                         LBUG();
1492                         RETURN(-ENOMEM);
1493                 }
1494                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1495                            (long)old_res);
1496                 RETURN(ELDLM_LOCK_CHANGED);
1497         } else {
1498                 int size = sizeof(struct ldlm_reply);
1499                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1500                                      &req->rq_repmsg);
1501                 if (rc) {
1502                         LBUG();
1503                         RETURN(-ENOMEM);
1504                 }
1505         }
1506         RETURN(rc);
1507 }
1508
1509 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1510 {
1511         return lprocfs_reg_obd(dev, status_var_nm_1, dev);
1512 }
1513
1514 int mds_detach(struct obd_device *dev)
1515 {
1516         return lprocfs_dereg_obd(dev);
1517 }
1518
1519 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1520 {
1521         int i;
1522         //        struct obd_ioctl_data* data = buf;
1523         struct mds_obd *mds = &obddev->u.mds;
1524         int rc = 0;
1525         ENTRY;
1526
1527         MOD_INC_USE_COUNT;
1528
1529         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1530                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1531                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1532                                            "self", mds_handle, "mds");
1533         if (!mds->mds_service) {
1534                 CERROR("failed to start service\n");
1535                 GOTO(err_dec, rc = -EINVAL);
1536         }
1537
1538         for (i = 0; i < MDT_NUM_THREADS; i++) {
1539                 char name[32];
1540                 sprintf(name, "ll_mdt_%02d", i);
1541                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1542                 if (rc) {
1543                         CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
1544                         GOTO(err_thread, rc);
1545                 }
1546         }
1547
1548         RETURN(0);
1549
1550 err_thread:
1551         ptlrpc_stop_all_threads(mds->mds_service);
1552         ptlrpc_unregister_service(mds->mds_service);
1553 err_dec:
1554         MOD_DEC_USE_COUNT;
1555         RETURN(rc);
1556 }
1557
1558
1559 static int mdt_cleanup(struct obd_device *obddev)
1560 {
1561         struct mds_obd *mds = &obddev->u.mds;
1562         ENTRY;
1563
1564         ptlrpc_stop_all_threads(mds->mds_service);
1565         ptlrpc_unregister_service(mds->mds_service);
1566
1567         MOD_DEC_USE_COUNT;
1568         RETURN(0);
1569 }
1570
1571 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1572                          int len, void *karg, void *uarg);
1573
1574 /* use obd ops to offer management infrastructure */
1575 static struct obd_ops mds_obd_ops = {
1576         o_attach:      mds_attach,
1577         o_detach:      mds_detach,
1578         o_connect:     mds_connect,
1579         o_disconnect:  mds_disconnect,
1580         o_setup:       mds_setup,
1581         o_cleanup:     mds_cleanup,
1582         o_iocontrol:   mds_iocontrol
1583 };
1584
1585 static struct obd_ops mdt_obd_ops = {
1586         o_setup:       mdt_setup,
1587         o_cleanup:     mdt_cleanup,
1588 };
1589
1590
1591 static int __init mds_init(void)
1592 {
1593
1594         mds_file_cache = kmem_cache_create("ll_mds_file_data",
1595                                            sizeof(struct mds_file_data),
1596                                            0, 0, NULL, NULL);
1597         if (mds_file_cache == NULL)
1598                 return -ENOMEM;
1599
1600         class_register_type(&mds_obd_ops, status_class_var, LUSTRE_MDS_NAME);
1601         class_register_type(&mdt_obd_ops, 0, LUSTRE_MDT_NAME);
1602         ldlm_register_intent(ldlm_intent_policy);
1603         return 0;
1604
1605 }
1606
1607 static void __exit mds_exit(void)
1608 {
1609
1610
1611         ldlm_unregister_intent();
1612         class_unregister_type(LUSTRE_MDS_NAME);
1613         class_unregister_type(LUSTRE_MDT_NAME);
1614         if (kmem_cache_destroy(mds_file_cache))
1615                 CERROR("couldn't free MDS file cache\n");
1616
1617 }
1618
1619 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1620 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1621 MODULE_LICENSE("GPL");
1622
1623 module_init(mds_init);
1624 module_exit(mds_exit);