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) 2001-2003 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  *   Author: Mike Shaver <shaver@clusterfs.com>
  12  *
  13  *   This file is part of Lustre, http://www.lustre.org.
  14  *
  15  *   Lustre is free software; you can redistribute it and/or
  16  *   modify it under the terms of version 2 of the GNU General Public
  17  *   License as published by the Free Software Foundation.
  18  *
  19  *   Lustre is distributed in the hope that it will be useful,
  20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   GNU General Public License for more details.
  23  *
  24  *   You should have received a copy of the GNU General Public License
  25  *   along with Lustre; if not, write to the Free Software
  26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/lustre_mds.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/init.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/random.h>
  40 #include <linux/fs.h>
  41 #include <linux/jbd.h>
  42 #include <linux/ext3_fs.h>
  43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  44 # include <linux/smp_lock.h>
  45 # include <linux/buffer_head.h>
  46 # include <linux/workqueue.h>
  47 # include <linux/mount.h>
  48 #else
  49 # include <linux/locks.h>
  50 #endif
  51 #include <linux/obd_lov.h>
  52 #include <linux/lustre_mds.h>
  53 #include <linux/lustre_fsfilt.h>
  54 #include <linux/lprocfs_status.h>
  55 #include <linux/lustre_commit_confd.h>
  56
  57 #include "mds_internal.h"
  58
  59 static int mds_cleanup(struct obd_device *obd, int flags);
  60
  61 static int mds_bulk_timeout(void *data)
  62 {
  63         struct ptlrpc_bulk_desc *desc = data;
  64         struct obd_export *exp = desc->bd_export;
  65
  66         DEBUG_REQ(D_ERROR, desc->bd_req,"bulk send timed out: evicting %s@%s\n",
  67                   exp->exp_client_uuid.uuid,
  68                   exp->exp_connection->c_remote_uuid.uuid);
  69         ptlrpc_fail_export(exp);
  70         ptlrpc_abort_bulk (desc);
  71         RETURN(1);
  72 }
  73
  74 /* Assumes caller has already pushed into the kernel filesystem context */
  75 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  76                         loff_t offset, int count)
  77 {
  78         struct ptlrpc_bulk_desc *desc;
  79         struct l_wait_info lwi;
  80         struct page **pages;
  81         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  82         ENTRY;
  83
  84         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  85
  86         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  87         OBD_ALLOC(pages, sizeof(*pages) * npages);
  88         if (!pages)
  89                 GOTO(out, rc = -ENOMEM);
  90
  91         desc = ptlrpc_prep_bulk_exp (req, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
  92         if (desc == NULL)
  93                 GOTO(out_free, rc = -ENOMEM);
  94
  95         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  96                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  97
  98                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  99                 if (pages[i] == NULL)
 100                         GOTO(cleanup_buf, rc = -ENOMEM);
 101
 102                 rc = ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 103                 if (rc != 0)
 104                         GOTO(cleanup_buf, rc);
 105         }
 106
 107         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 108                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 109                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 110                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 111                        file->f_dentry->d_inode->i_size);
 112
 113                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 114                                      page_address(pages[i]), tmpsize, &offset);
 115
 116                 if (rc != tmpsize)
 117                         GOTO(cleanup_buf, rc = -EIO);
 118         }
 119
 120         rc = ptlrpc_bulk_put(desc);
 121         if (rc)
 122                 GOTO(cleanup_buf, rc);
 123
 124         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 125                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 126                        OBD_FAIL_MDS_SENDPAGE, rc);
 127                 ptlrpc_abort_bulk(desc);
 128                 GOTO(cleanup_buf, rc);
 129         }
 130
 131         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, mds_bulk_timeout, desc);
 132         rc = l_wait_event(desc->bd_waitq, ptlrpc_bulk_complete (desc), &lwi);
 133         if (rc) {
 134                 LASSERT (rc == -ETIMEDOUT);
 135                 GOTO(cleanup_buf, rc);
 136         }
 137
 138         EXIT;
 139  cleanup_buf:
 140         for (i = 0; i < npages; i++)
 141                 if (pages[i])
 142                         __free_pages(pages[i], 0);
 143
 144         ptlrpc_free_bulk(desc);
 145  out_free:
 146         OBD_FREE(pages, sizeof(*pages) * npages);
 147  out:
 148         return rc;
 149 }
 150
 151 /* only valid locked dentries or errors should be returned */
 152 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 153                                      struct vfsmount **mnt, int lock_mode,
 154                                      struct lustre_handle *lockh,
 155                                      char *name, int namelen)
 156 {
 157         struct mds_obd *mds = &obd->u.mds;
 158         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 159         struct ldlm_res_id res_id = { .name = {0} };
 160         int flags = 0, rc;
 161         ENTRY;
 162
 163         if (IS_ERR(de))
 164                 RETURN(de);
 165
 166         res_id.name[0] = de->d_inode->i_ino;
 167         res_id.name[1] = de->d_inode->i_generation;
 168         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 169                               res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 170                               &flags, ldlm_completion_ast,
 171                               mds_blocking_ast, NULL, lockh);
 172         if (rc != ELDLM_OK) {
 173                 l_dput(de);
 174                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 175         }
 176
 177         RETURN(retval);
 178 }
 179
 180 #ifndef DCACHE_DISCONNECTED
 181 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 182 #endif
 183
 184
 185 /* Look up an entry by inode number. */
 186 /* this function ONLY returns valid dget'd dentries with an initialized inode
 187    or errors */
 188 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 189                               struct vfsmount **mnt)
 190 {
 191         char fid_name[32];
 192         unsigned long ino = fid->id;
 193         __u32 generation = fid->generation;
 194         struct inode *inode;
 195         struct dentry *result;
 196
 197         if (ino == 0)
 198                 RETURN(ERR_PTR(-ESTALE));
 199
 200         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 201
 202         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 203                ino, generation, mds->mds_sb);
 204
 205         /* under ext3 this is neither supposed to return bad inodes
 206            nor NULL inodes. */
 207         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 208         if (IS_ERR(result))
 209                 RETURN(result);
 210
 211         inode = result->d_inode;
 212         if (!inode)
 213                 RETURN(ERR_PTR(-ENOENT));
 214
 215         if (generation && inode->i_generation != generation) {
 216                 /* we didn't find the right inode.. */
 217                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 218                        inode->i_ino, (unsigned long)inode->i_nlink,
 219                        atomic_read(&inode->i_count), inode->i_generation,
 220                        generation);
 221                 dput(result);
 222                 RETURN(ERR_PTR(-ENOENT));
 223         }
 224
 225         if (mnt) {
 226                 *mnt = mds->mds_vfsmnt;
 227                 mntget(*mnt);
 228         }
 229
 230         RETURN(result);
 231 }
 232
 233
 234 /* Establish a connection to the MDS.
 235  *
 236  * This will set up an export structure for the client to hold state data
 237  * about that client, like open files, the last operation number it did
 238  * on the server, etc.
 239  */
 240 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 241                        struct obd_uuid *cluuid)
 242 {
 243         struct obd_export *exp;
 244         struct mds_export_data *med; /*  */
 245         struct mds_client_data *mcd;
 246         int rc, abort_recovery;
 247         ENTRY;
 248
 249         if (!conn || !obd || !cluuid)
 250                 RETURN(-EINVAL);
 251
 252         /* Check for aborted recovery. */
 253         spin_lock_bh(&obd->obd_processing_task_lock);
 254         abort_recovery = obd->obd_abort_recovery;
 255         spin_unlock_bh(&obd->obd_processing_task_lock);
 256         if (abort_recovery)
 257                 target_abort_recovery(obd);
 258
 259         /* XXX There is a small race between checking the list and adding a
 260          * new connection for the same UUID, but the real threat (list
 261          * corruption when multiple different clients connect) is solved.
 262          *
 263          * There is a second race between adding the export to the list,
 264          * and filling in the client data below.  Hence skipping the case
 265          * of NULL mcd above.  We should already be controlling multiple
 266          * connects at the client, and we can't hold the spinlock over
 267          * memory allocations without risk of deadlocking.
 268          */
 269         rc = class_connect(conn, obd, cluuid);
 270         if (rc)
 271                 RETURN(rc);
 272         exp = class_conn2export(conn);
 273         LASSERT(exp);
 274         med = &exp->exp_mds_data;
 275
 276         OBD_ALLOC(mcd, sizeof(*mcd));
 277         if (!mcd) {
 278                 CERROR("mds: out of memory for client data\n");
 279                 GOTO(out, rc = -ENOMEM);
 280         }
 281
 282         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 283         med->med_mcd = mcd;
 284         mcd->mcd_mount_count = cpu_to_le64(obd->u.mds.mds_mount_count);
 285
 286         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 287         if (rc == 0)
 288                 EXIT;
 289 out:
 290         if (rc) {
 291                 OBD_FREE(mcd, sizeof(*mcd));
 292                 class_disconnect(exp, 0);
 293         }
 294         class_export_put(exp);
 295
 296         return rc;
 297 }
 298
 299 static int mds_init_export(struct obd_export *exp)
 300 {
 301         struct mds_export_data *med = &exp->exp_mds_data;
 302
 303         INIT_LIST_HEAD(&med->med_open_head);
 304         spin_lock_init(&med->med_open_lock);
 305         RETURN(0);
 306 }
 307
 308 static int mds_destroy_export(struct obd_export *export)
 309 {
 310         struct mds_export_data *med;
 311         struct obd_device *obd = export->exp_obd;
 312         struct obd_run_ctxt saved;
 313         int rc = 0;
 314         ENTRY;
 315
 316         med = &export->exp_mds_data;
 317         target_destroy_export(export);
 318
 319         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 320         /* Close any open files (which may also cause orphan unlinking). */
 321         spin_lock(&med->med_open_lock);
 322         while (!list_empty(&med->med_open_head)) {
 323                 struct list_head *tmp = med->med_open_head.next;
 324                 struct mds_file_data *mfd =
 325                         list_entry(tmp, struct mds_file_data, mfd_list);
 326                 BDEVNAME_DECLARE_STORAGE(btmp);
 327
 328                 /* bug 1579: fix force-closing for 2.5 */
 329                 struct dentry *dentry = mfd->mfd_dentry;
 330
 331                 list_del(&mfd->mfd_list);
 332                 spin_unlock(&med->med_open_lock);
 333
 334                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 335                        dentry->d_name.len, dentry->d_name.name,
 336                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 337                        dentry->d_inode->i_ino);
 338                 rc = mds_mfd_close(NULL, obd, mfd,
 339                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 340
 341                 if (rc)
 342                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 343                 spin_lock(&med->med_open_lock);
 344         }
 345         spin_unlock(&med->med_open_lock);
 346         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 347
 348         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 349
 350         RETURN(rc);
 351 }
 352
 353 static int mds_disconnect(struct obd_export *export, int flags)
 354 {
 355         unsigned long irqflags;
 356         int rc;
 357         ENTRY;
 358
 359         ldlm_cancel_locks_for_export(export);
 360
 361         spin_lock_irqsave(&export->exp_lock, irqflags);
 362         export->exp_flags = flags;
 363         spin_unlock_irqrestore(&export->exp_lock, irqflags);
 364
 365         rc = class_disconnect(export, flags);
 366         RETURN(rc);
 367 }
 368
 369 static int mds_getstatus(struct ptlrpc_request *req)
 370 {
 371         struct mds_obd *mds = mds_req2mds(req);
 372         struct mds_body *body;
 373         int rc, size = sizeof(*body);
 374         ENTRY;
 375
 376         rc = lustre_pack_reply(req, 1, &size, NULL);
 377         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 378                 CERROR("mds: out of memory for message: size=%d\n", size);
 379                 req->rq_status = -ENOMEM;       /* superfluous? */
 380                 RETURN(-ENOMEM);
 381         }
 382
 383         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 384         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 385
 386         /* the last_committed and last_xid fields are filled in for all
 387          * replies already - no need to do so here also.
 388          */
 389         RETURN(0);
 390 }
 391
 392 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 393                      void *data, int flag)
 394 {
 395         int do_ast;
 396         ENTRY;
 397
 398         if (flag == LDLM_CB_CANCELING) {
 399                 /* Don't need to do anything here. */
 400                 RETURN(0);
 401         }
 402
 403         /* XXX layering violation!  -phil */
 404         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 405         /* Get this: if mds_blocking_ast is racing with ldlm_intent_policy,
 406          * such that mds_blocking_ast is called just before l_i_p takes the
 407          * ns_lock, then by the time we get the lock, we might not be the
 408          * correct blocking function anymore.  So check, and return early, if
 409          * so. */
 410         if (lock->l_blocking_ast != mds_blocking_ast) {
 411                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 412                 RETURN(0);
 413         }
 414
 415         lock->l_flags |= LDLM_FL_CBPENDING;
 416         do_ast = (!lock->l_readers && !lock->l_writers);
 417         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 418
 419         if (do_ast) {
 420                 struct lustre_handle lockh;
 421                 int rc;
 422
 423                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 424                 ldlm_lock2handle(lock, &lockh);
 425                 rc = ldlm_cli_cancel(&lockh);
 426                 if (rc < 0)
 427                         CERROR("ldlm_cli_cancel: %d\n", rc);
 428         } else {
 429                 LDLM_DEBUG(lock, "Lock still has references, will be "
 430                            "cancelled later");
 431         }
 432         RETURN(0);
 433 }
 434
 435 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 436  * Call with lock=0 if the caller has already taken the i_sem. */
 437 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 438                 struct mds_body *body, struct inode *inode, int lock)
 439 {
 440         struct mds_obd *mds = &obd->u.mds;
 441         void *lmm;
 442         int lmm_size;
 443         int rc;
 444         ENTRY;
 445
 446         lmm = lustre_msg_buf(msg, offset, 0);
 447         if (lmm == NULL) {
 448                 /* Some problem with getting eadata when I sized the reply
 449                  * buffer... */
 450                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 451                        inode->i_ino);
 452                 RETURN(0);
 453         }
 454         lmm_size = msg->buflens[offset];
 455
 456         /* I don't really like this, but it is a sanity check on the client
 457          * MD request.  However, if the client doesn't know how much space
 458          * to reserve for the MD, it shouldn't be bad to have too much space.
 459          */
 460         if (lmm_size > mds->mds_max_mdsize) {
 461                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 462                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 463                 // RETURN(-EINVAL);
 464         }
 465
 466         if (lock)
 467                 down(&inode->i_sem);
 468         rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
 469         if (lock)
 470                 up(&inode->i_sem);
 471         if (rc < 0) {
 472                 CERROR("Error %d reading eadata for ino %lu\n",
 473                        rc, inode->i_ino);
 474         } else if (rc > 0) {
 475                 lmm_size = rc;
 476                 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
 477
 478                 if (rc > 0)
 479                         lmm_size = rc;
 480                 body->valid |= OBD_MD_FLEASIZE;
 481                 body->eadatasize = lmm_size;
 482                 rc = 0;
 483         }
 484
 485         RETURN(rc);
 486 }
 487
 488 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 489                                 struct ptlrpc_request *req,
 490                                 struct mds_body *reqbody, int reply_off)
 491 {
 492         struct mds_body *body;
 493         struct inode *inode = dentry->d_inode;
 494         int rc = 0;
 495         ENTRY;
 496
 497         if (inode == NULL)
 498                 RETURN(-ENOENT);
 499
 500         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 501         LASSERT(body != NULL);                 /* caller prepped reply */
 502
 503         mds_pack_inode2fid(&body->fid1, inode);
 504         mds_pack_inode2body(body, inode);
 505
 506         if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
 507                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 508                                  inode, 1);
 509
 510                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 511                 if (!(body->valid & OBD_MD_FLEASIZE))
 512                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 513                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 514         } else if (S_ISLNK(inode->i_mode) &&
 515                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 516                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 517                 int len;
 518
 519                 LASSERT (symname != NULL);       /* caller prepped reply */
 520                 len = req->rq_repmsg->buflens[reply_off + 1];
 521
 522                 rc = inode->i_op->readlink(dentry, symname, len);
 523                 if (rc < 0) {
 524                         CERROR("readlink failed: %d\n", rc);
 525                 } else if (rc != len - 1) {
 526                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 527                                 rc, len - 1);
 528                         rc = -EINVAL;
 529                 } else {
 530                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 531                         body->valid |= OBD_MD_LINKNAME;
 532                         body->eadatasize = rc + 1;
 533                         symname[rc] = 0;        /* NULL terminate */
 534                         rc = 0;
 535                 }
 536         }
 537
 538         RETURN(rc);
 539 }
 540
 541 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 542                                 int offset)
 543 {
 544         struct mds_obd *mds = mds_req2mds(req);
 545         struct mds_body *body;
 546         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 547         ENTRY;
 548
 549         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 550         LASSERT(body != NULL);                 /* checked by caller */
 551         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 552
 553         if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
 554                 int rc;
 555                 down(&inode->i_sem);
 556                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 557                 up(&inode->i_sem);
 558                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 559                        rc, inode->i_ino);
 560                 if (rc < 0) {
 561                         if (rc != -ENODATA)
 562                                 CERROR("error getting inode %lu MD: rc = %d\n",
 563                                        inode->i_ino, rc);
 564                         size[bufcount] = 0;
 565                 } else if (rc > mds->mds_max_mdsize) {
 566                         size[bufcount] = 0;
 567                         CERROR("MD size %d larger than maximum possible %u\n",
 568                                rc, mds->mds_max_mdsize);
 569                 } else {
 570                         size[bufcount] = rc;
 571                 }
 572                 bufcount++;
 573         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 574                 if (inode->i_size + 1 != body->eadatasize)
 575                         CERROR("symlink size: %Lu, reply space: %d\n",
 576                                inode->i_size + 1, body->eadatasize);
 577                 size[bufcount] = MIN(inode->i_size + 1, body->eadatasize);
 578                 bufcount++;
 579                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 580                        inode->i_size + 1, body->eadatasize);
 581         }
 582
 583         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 584                 CERROR("failed MDS_GETATTR_PACK test\n");
 585                 req->rq_status = -ENOMEM;
 586                 GOTO(out, rc = -ENOMEM);
 587         }
 588
 589         rc = lustre_pack_reply(req, bufcount, size, NULL);
 590         if (rc) {
 591                 CERROR("out of memory\n");
 592                 GOTO(out, req->rq_status = rc);
 593         }
 594
 595         EXIT;
 596  out:
 597         return(rc);
 598 }
 599
 600 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
 601                             struct lustre_handle *child_lockh)
 602 {
 603         struct obd_device *obd = req->rq_export->exp_obd;
 604         struct ldlm_reply *rep = NULL;
 605         struct obd_run_ctxt saved;
 606         struct mds_body *body;
 607         struct dentry *dparent = NULL, *dchild = NULL;
 608         struct obd_ucred uc;
 609         struct lustre_handle parent_lockh;
 610         int namesize;
 611         int rc = 0, cleanup_phase = 0, resent_req = 0;
 612         char *name;
 613         ENTRY;
 614
 615         LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
 616
 617         /* Swab now, before anyone looks inside the request */
 618
 619         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 620                                   lustre_swab_mds_body);
 621         if (body == NULL) {
 622                 CERROR("Can't swab mds_body\n");
 623                 GOTO(cleanup, rc = -EFAULT);
 624         }
 625
 626         LASSERT_REQSWAB(req, offset + 1);
 627         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 628         if (name == NULL) {
 629                 CERROR("Can't unpack name\n");
 630                 GOTO(cleanup, rc = -EFAULT);
 631         }
 632         namesize = req->rq_reqmsg->buflens[offset + 1];
 633
 634         LASSERT (offset == 0 || offset == 2);
 635         /* if requests were at offset 2, the getattr reply goes back at 1 */
 636         if (offset) {
 637                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 638                 offset = 1;
 639         }
 640
 641         uc.ouc_fsuid = body->fsuid;
 642         uc.ouc_fsgid = body->fsgid;
 643         uc.ouc_cap = body->capability;
 644         uc.ouc_suppgid1 = body->suppgid;
 645         uc.ouc_suppgid2 = -1;
 646         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 647         cleanup_phase = 1; /* kernel context */
 648         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 649
 650         /* FIXME: handle raw lookup */
 651 #if 0
 652         if (body->valid == OBD_MD_FLID) {
 653                 struct mds_body *mds_reply;
 654                 int size = sizeof(*mds_reply);
 655                 ino_t inum;
 656                 // The user requested ONLY the inode number, so do a raw lookup
 657                 rc = lustre_pack_reply(req, 1, &size, NULL);
 658                 if (rc) {
 659                         CERROR("out of memory\n");
 660                         GOTO(cleanup, rc);
 661                 }
 662
 663                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 664
 665                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 666                                            sizeof(*mds_reply));
 667                 mds_reply->fid1.id = inum;
 668                 mds_reply->valid = OBD_MD_FLID;
 669                 GOTO(cleanup, rc);
 670         }
 671 #endif
 672
 673         if (child_lockh->cookie != 0) {
 674                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 675                 resent_req = 1;
 676         }
 677
 678         if (resent_req == 0) {
 679                 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
 680                                                  &parent_lockh, &dparent,
 681                                                  LCK_PR, name, namesize,
 682                                                  child_lockh, &dchild, LCK_PR);
 683                 if (rc)
 684                         GOTO(cleanup, rc);
 685         } else {
 686                 struct ldlm_lock *granted_lock;
 687                 struct ll_fid child_fid;
 688                 struct ldlm_resource *res;
 689                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 690                 granted_lock = ldlm_handle2lock(child_lockh);
 691                 LASSERT(granted_lock);
 692
 693                 res = granted_lock->l_resource;
 694                 child_fid.id = res->lr_name.name[0];
 695                 child_fid.generation = res->lr_name.name[1];
 696                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 697                 LASSERT(dchild);
 698                 LDLM_LOCK_PUT(granted_lock);
 699         }
 700
 701         cleanup_phase = 2; /* dchild, dparent, locks */
 702
 703         if (dchild->d_inode == NULL) {
 704                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 705                 /* in the intent case, the policy clears this error:
 706                    the disposition is enough */
 707                 GOTO(cleanup, rc = -ENOENT);
 708         } else {
 709                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 710         }
 711
 712         if (req->rq_repmsg == NULL) {
 713                 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 714                 if (rc != 0) {
 715                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 716                         GOTO (cleanup, rc);
 717                 }
 718         }
 719
 720         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 721         GOTO(cleanup, rc); /* returns the lock to the client */
 722
 723  cleanup:
 724         switch (cleanup_phase) {
 725         case 2:
 726                 if (resent_req == 0) {
 727                         if (rc && dchild->d_inode)
 728                                 ldlm_lock_decref(child_lockh, LCK_PR);
 729                         ldlm_lock_decref(&parent_lockh, LCK_PR);
 730                         l_dput(dparent);
 731                 }
 732                 l_dput(dchild);
 733         case 1:
 734                 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 735         default: ;
 736         }
 737         return rc;
 738 }
 739
 740 static int mds_getattr(int offset, struct ptlrpc_request *req)
 741 {
 742         struct mds_obd *mds = mds_req2mds(req);
 743         struct obd_device *obd = req->rq_export->exp_obd;
 744         struct obd_run_ctxt saved;
 745         struct dentry *de;
 746         struct mds_body *body;
 747         struct obd_ucred uc;
 748         int rc = 0;
 749         ENTRY;
 750
 751         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
 752                                    lustre_swab_mds_body);
 753         if (body == NULL) {
 754                 CERROR ("Can't unpack body\n");
 755                 RETURN (-EFAULT);
 756         }
 757
 758         uc.ouc_fsuid = body->fsuid;
 759         uc.ouc_fsgid = body->fsgid;
 760         uc.ouc_cap = body->capability;
 761         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 762         de = mds_fid2dentry(mds, &body->fid1, NULL);
 763         if (IS_ERR(de)) {
 764                 rc = req->rq_status = -ENOENT;
 765                 GOTO(out_pop, PTR_ERR(de));
 766         }
 767
 768         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
 769         if (rc != 0) {
 770                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
 771                 GOTO (out_pop, rc);
 772         }
 773
 774         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
 775
 776         l_dput(de);
 777         GOTO(out_pop, rc);
 778 out_pop:
 779         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 780         return rc;
 781 }
 782
 783
 784 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 785                           unsigned long max_age)
 786 {
 787         int rc;
 788
 789         spin_lock(&obd->obd_osfs_lock);
 790         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
 791         if (rc == 0)
 792                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
 793         spin_unlock(&obd->obd_osfs_lock);
 794
 795         return rc;
 796 }
 797
 798 static int mds_statfs(struct ptlrpc_request *req)
 799 {
 800         struct obd_device *obd = req->rq_export->exp_obd;
 801         int rc, size = sizeof(struct obd_statfs);
 802         ENTRY;
 803
 804         rc = lustre_pack_reply(req, 1, &size, NULL);
 805         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 806                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
 807                 GOTO(out, rc);
 808         }
 809
 810         /* We call this so that we can cache a bit - 1 jiffie worth */
 811         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
 812                             jiffies - HZ);
 813         if (rc) {
 814                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
 815                 GOTO(out, rc);
 816         }
 817
 818         EXIT;
 819 out:
 820         req->rq_status = rc;
 821         return 0;
 822 }
 823
 824 static int mds_sync(struct ptlrpc_request *req)
 825 {
 826         struct obd_device *obd = req->rq_export->exp_obd;
 827         struct mds_obd *mds = &obd->u.mds;
 828         struct mds_body *body;
 829         int rc, size = sizeof(*body);
 830         ENTRY;
 831
 832         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 833         if (body == NULL)
 834                 GOTO(out, rc = -EPROTO);
 835
 836         rc = lustre_pack_reply(req, 1, &size, NULL);
 837         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
 838                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
 839                 GOTO(out, rc);
 840         }
 841
 842         if (body->fid1.id == 0) {
 843                 /* a fid of zero is taken to mean "sync whole filesystem" */
 844                 rc = fsfilt_sync(obd, mds->mds_sb);
 845                 if (rc)
 846                         GOTO(out, rc);
 847         } else {
 848                 /* just any file to grab fsync method - "file" arg unused */
 849                 struct file *file = mds->mds_rcvd_filp;
 850                 struct dentry *de;
 851
 852                 de = mds_fid2dentry(mds, &body->fid1, NULL);
 853                 if (IS_ERR(de))
 854                         GOTO(out, rc = PTR_ERR(de));
 855
 856                 rc = file->f_op->fsync(NULL, de, 1);
 857                 l_dput(de);
 858                 if (rc)
 859                         GOTO(out, rc);
 860
 861                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 862                 mds_pack_inode2fid(&body->fid1, de->d_inode);
 863                 mds_pack_inode2body(body, de->d_inode);
 864         }
 865 out:
 866         req->rq_status = rc;
 867         return 0;
 868 }
 869
 870 /* mds_readpage does not take a DLM lock on the inode, because the client must
 871  * already have a PR lock.
 872  *
 873  * If we were to take another one here, a deadlock will result, if another
 874  * thread is already waiting for a PW lock. */
 875 static int mds_readpage(struct ptlrpc_request *req)
 876 {
 877         struct obd_device *obd = req->rq_export->exp_obd;
 878         struct vfsmount *mnt;
 879         struct dentry *de;
 880         struct file *file;
 881         struct mds_body *body, *repbody;
 882         struct obd_run_ctxt saved;
 883         int rc, size = sizeof(*repbody);
 884         struct obd_ucred uc;
 885         ENTRY;
 886
 887         rc = lustre_pack_reply(req, 1, &size, NULL);
 888         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 889                 CERROR("mds: out of memory\n");
 890                 GOTO(out, rc = -ENOMEM);
 891         }
 892
 893         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
 894         if (body == NULL)
 895                 GOTO (out, rc = -EFAULT);
 896
 897         uc.ouc_fsuid = body->fsuid;
 898         uc.ouc_fsgid = body->fsgid;
 899         uc.ouc_cap = body->capability;
 900         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 901         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
 902         if (IS_ERR(de))
 903                 GOTO(out_pop, rc = PTR_ERR(de));
 904
 905         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
 906
 907         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 908         /* note: in case of an error, dentry_open puts dentry */
 909         if (IS_ERR(file))
 910                 GOTO(out_pop, rc = PTR_ERR(file));
 911
 912         /* body->size is actually the offset -eeb */
 913         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
 914                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
 915                        body->size, de->d_inode->i_blksize);
 916                 GOTO(out_file, rc = -EFAULT);
 917         }
 918
 919         /* body->nlink is actually the #bytes to read -eeb */
 920         if (body->nlink & (de->d_inode->i_blksize - 1)) {
 921                 CERROR("size %u is not multiple of blocksize %lu\n",
 922                        body->nlink, de->d_inode->i_blksize);
 923                 GOTO(out_file, rc = -EFAULT);
 924         }
 925
 926         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
 927         repbody->size = file->f_dentry->d_inode->i_size;
 928         repbody->valid = OBD_MD_FLSIZE;
 929
 930         /* to make this asynchronous make sure that the handling function
 931            doesn't send a reply when this function completes. Instead a
 932            callback function would send the reply */
 933         /* body->size is actually the offset -eeb */
 934         rc = mds_sendpage(req, file, body->size, body->nlink);
 935
 936 out_file:
 937         filp_close(file, 0);
 938 out_pop:
 939         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 940 out:
 941         req->rq_status = rc;
 942         RETURN(0);
 943 }
 944
 945 int mds_reint(struct ptlrpc_request *req, int offset,
 946               struct lustre_handle *lockh)
 947 {
 948         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
 949         int rc;
 950
 951         OBD_ALLOC(rec, sizeof(*rec));
 952         if (rec == NULL)
 953                 RETURN(-ENOMEM);
 954
 955         rc = mds_update_unpack(req, offset, rec);
 956         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 957                 CERROR("invalid record\n");
 958                 GOTO(out, req->rq_status = -EINVAL);
 959         }
 960         /* rc will be used to interrupt a for loop over multiple records */
 961         rc = mds_reint_rec(rec, offset, req, lockh);
 962  out:
 963         OBD_FREE(rec, sizeof(*rec));
 964         return rc;
 965 }
 966
 967 static int mds_filter_recovery_request(struct ptlrpc_request *req,
 968                                        struct obd_device *obd, int *process)
 969 {
 970         switch (req->rq_reqmsg->opc) {
 971         case MDS_CONNECT: /* This will never get here, but for completeness. */
 972         case OST_CONNECT: /* This will never get here, but for completeness. */
 973         case MDS_DISCONNECT:
 974         case OST_DISCONNECT:
 975                *process = 1;
 976                RETURN(0);
 977
 978         case MDS_CLOSE:
 979         case MDS_SYNC: /* used in unmounting */
 980         case OBD_PING:
 981         case MDS_REINT:
 982         case LDLM_ENQUEUE:
 983                 *process = target_queue_recovery_request(req, obd);
 984                 RETURN(0);
 985
 986         default:
 987                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 988                 *process = 0;
 989                 /* XXX what should we set rq_status to here? */
 990                 req->rq_status = -EAGAIN;
 991                 RETURN(ptlrpc_error(req));
 992         }
 993 }
 994
 995 static char *reint_names[] = {
 996         [REINT_SETATTR] "setattr",
 997         [REINT_CREATE]  "create",
 998         [REINT_LINK]    "link",
 999         [REINT_UNLINK]  "unlink",
1000         [REINT_RENAME]  "rename",
1001         [REINT_OPEN]    "open",
1002 };
1003
1004 int mds_handle(struct ptlrpc_request *req)
1005 {
1006         int should_process;
1007         int rc = 0;
1008         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1009         struct obd_device *obd = NULL;
1010         ENTRY;
1011
1012         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1013
1014         LASSERT(current->journal_info == NULL);
1015         /* XXX identical to OST */
1016         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1017                 struct mds_export_data *med;
1018                 int recovering, abort_recovery;
1019
1020                 if (req->rq_export == NULL) {
1021                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1022                                req->rq_reqmsg->opc);
1023                         req->rq_status = -ENOTCONN;
1024                         GOTO(out, rc = -ENOTCONN);
1025                 }
1026
1027                 med = &req->rq_export->exp_mds_data;
1028                 obd = req->rq_export->exp_obd;
1029                 mds = &obd->u.mds;
1030
1031                 /* Check for aborted recovery. */
1032                 spin_lock_bh(&obd->obd_processing_task_lock);
1033                 abort_recovery = obd->obd_abort_recovery;
1034                 recovering = obd->obd_recovering;
1035                 spin_unlock_bh(&obd->obd_processing_task_lock);
1036                 if (abort_recovery) {
1037                         target_abort_recovery(obd);
1038                 } else if (recovering) {
1039                         rc = mds_filter_recovery_request(req, obd,
1040                                                          &should_process);
1041                         if (rc || !should_process)
1042                                 RETURN(rc);
1043                 }
1044         }
1045
1046         switch (req->rq_reqmsg->opc) {
1047         case MDS_CONNECT:
1048                 DEBUG_REQ(D_INODE, req, "connect");
1049                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1050                 rc = target_handle_connect(req, mds_handle);
1051                 if (!rc)
1052                         /* Now that we have an export, set mds. */
1053                         mds = mds_req2mds(req);
1054                 break;
1055
1056         case MDS_DISCONNECT:
1057                 DEBUG_REQ(D_INODE, req, "disconnect");
1058                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1059                 rc = target_handle_disconnect(req);
1060                 req->rq_status = rc;            /* superfluous? */
1061                 break;
1062
1063         case MDS_GETSTATUS:
1064                 DEBUG_REQ(D_INODE, req, "getstatus");
1065                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1066                 rc = mds_getstatus(req);
1067                 break;
1068
1069         case MDS_GETATTR:
1070                 DEBUG_REQ(D_INODE, req, "getattr");
1071                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1072                 rc = mds_getattr(0, req);
1073                 break;
1074
1075         case MDS_GETATTR_NAME: {
1076                 struct lustre_handle lockh;
1077                 DEBUG_REQ(D_INODE, req, "getattr_name");
1078                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1079
1080                 /* If this request gets a reconstructed reply, we won't be
1081                  * acquiring any new locks in mds_getattr_name, so we don't
1082                  * want to cancel.
1083                  */
1084                 lockh.cookie = 0;
1085                 rc = mds_getattr_name(0, req, &lockh);
1086                 /* this non-intent call (from an ioctl) is special */
1087                 req->rq_status = rc;
1088                 if (rc == 0 && lockh.cookie)
1089                         ldlm_lock_decref(&lockh, LCK_PR);
1090                 break;
1091         }
1092         case MDS_STATFS:
1093                 DEBUG_REQ(D_INODE, req, "statfs");
1094                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1095                 rc = mds_statfs(req);
1096                 break;
1097
1098         case MDS_READPAGE:
1099                 DEBUG_REQ(D_INODE, req, "readpage");
1100                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1101                 rc = mds_readpage(req);
1102
1103                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1104                         return 0;
1105                 break;
1106
1107         case MDS_REINT: {
1108                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1109                 __u32  opc;
1110                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1111                                mds->mds_max_cookiesize};
1112                 int bufcount;
1113
1114                 /* NB only peek inside req now; mds_reint() will swab it */
1115                 if (opcp == NULL) {
1116                         CERROR ("Can't inspect opcode\n");
1117                         rc = -EINVAL;
1118                         break;
1119                 }
1120                 opc = *opcp;
1121                 if (lustre_msg_swabbed (req->rq_reqmsg))
1122                         __swab32s(&opc);
1123
1124                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1125                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1126                            reint_names[opc] == NULL) ? reint_names[opc] :
1127                                                        "unknown opcode");
1128
1129                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1130
1131                 if (opc == REINT_UNLINK)
1132                         bufcount = 3;
1133                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1134                         bufcount = 2;
1135                 else
1136                         bufcount = 1;
1137
1138                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1139                 if (rc)
1140                         break;
1141
1142                 rc = mds_reint(req, 0, NULL);
1143                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1144                 break;
1145         }
1146
1147         case MDS_CLOSE:
1148                 DEBUG_REQ(D_INODE, req, "close");
1149                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1150                 rc = mds_close(req);
1151                 break;
1152
1153         case MDS_DONE_WRITING:
1154                 DEBUG_REQ(D_INODE, req, "done_writing");
1155                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1156                 rc = mds_done_writing(req);
1157                 break;
1158
1159         case MDS_PIN:
1160                 DEBUG_REQ(D_INODE, req, "pin");
1161                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1162                 rc = mds_pin(req);
1163                 break;
1164
1165         case MDS_SYNC:
1166                 DEBUG_REQ(D_INODE, req, "sync");
1167                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1168                 rc = mds_sync(req);
1169                 break;
1170
1171         case OBD_PING:
1172                 DEBUG_REQ(D_INODE, req, "ping");
1173                 rc = target_handle_ping(req);
1174                 break;
1175
1176         case OBD_LOG_CANCEL:
1177                 CDEBUG(D_INODE, "log cancel\n");
1178                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1179                 rc = -ENOTSUPP; /* la la la */
1180                 break;
1181
1182         case LDLM_ENQUEUE:
1183                 DEBUG_REQ(D_INODE, req, "enqueue");
1184                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1185                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1186                                          ldlm_server_blocking_ast);
1187                 break;
1188         case LDLM_CONVERT:
1189                 DEBUG_REQ(D_INODE, req, "convert");
1190                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1191                 rc = ldlm_handle_convert(req);
1192                 break;
1193         case LDLM_BL_CALLBACK:
1194         case LDLM_CP_CALLBACK:
1195                 DEBUG_REQ(D_INODE, req, "callback");
1196                 CERROR("callbacks should not happen on MDS\n");
1197                 LBUG();
1198                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1199                 break;
1200         case LLOG_ORIGIN_HANDLE_CREATE:
1201                 DEBUG_REQ(D_INODE, req, "llog_init");
1202                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1203                 rc = llog_origin_handle_create(req);
1204                 break;
1205         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1206                 DEBUG_REQ(D_INODE, req, "llog next block");
1207                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1208                 rc = llog_origin_handle_next_block(req);
1209                 break;
1210         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1211                 DEBUG_REQ(D_INODE, req, "llog read header");
1212                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1213                 rc = llog_origin_handle_read_header(req);
1214                 break;
1215         case LLOG_ORIGIN_HANDLE_CLOSE:
1216                 DEBUG_REQ(D_INODE, req, "llog close");
1217                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1218                 rc = llog_origin_handle_close(req);
1219                 break;
1220         case LLOG_CATINFO:
1221                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1222                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1223                 rc = llog_catinfo(req);
1224                 break;
1225         default:
1226                 req->rq_status = -ENOTSUPP;
1227                 rc = ptlrpc_error(req);
1228                 RETURN(rc);
1229         }
1230
1231         LASSERT(current->journal_info == NULL);
1232
1233         EXIT;
1234
1235         /* If we're DISCONNECTing, the mds_export_data is already freed */
1236         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1237                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1238                 struct obd_device *obd = list_entry(mds, struct obd_device,
1239                                                     u.mds);
1240                 req->rq_repmsg->last_xid =
1241                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1242
1243                 if (!obd->obd_no_transno) {
1244                         req->rq_repmsg->last_committed =
1245                                 obd->obd_last_committed;
1246                 } else {
1247                         DEBUG_REQ(D_IOCTL, req,
1248                                   "not sending last_committed update");
1249                 }
1250                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1251                        ", xid "LPU64"\n",
1252                        mds->mds_last_transno, obd->obd_last_committed,
1253                        req->rq_xid);
1254         }
1255  out:
1256
1257         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1258                 if (obd && obd->obd_recovering) {
1259                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1260                         return target_queue_final_reply(req, rc);
1261                 }
1262                 /* Lost a race with recovery; let the error path DTRT. */
1263                 rc = req->rq_status = -ENOTCONN;
1264         }
1265
1266         target_send_reply(req, rc, OBD_FAIL_MDS_ALL_REPLY_NET);
1267         return 0;
1268 }
1269
1270 /* Update the server data on disk.  This stores the new mount_count and
1271  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1272  * then the server last_rcvd value may be less than that of the clients.
1273  * This will alert us that we may need to do client recovery.
1274  *
1275  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1276  */
1277 int mds_update_server_data(struct obd_device *obd, int force_sync)
1278 {
1279         struct mds_obd *mds = &obd->u.mds;
1280         struct mds_server_data *msd = mds->mds_server_data;
1281         struct file *filp = mds->mds_rcvd_filp;
1282         struct obd_run_ctxt saved;
1283         loff_t off = 0;
1284         int rc;
1285         ENTRY;
1286
1287         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1288         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1289
1290         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1291                mds->mds_mount_count, mds->mds_last_transno);
1292         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1293         if (rc)
1294                 CERROR("error writing MDS server data: rc = %d\n", rc);
1295         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1296
1297         RETURN(rc);
1298 }
1299
1300
1301 /* mount the file system (secretly) */
1302 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1303 {
1304         struct lustre_cfg* lcfg = buf;
1305         struct mds_obd *mds = &obd->u.mds;
1306         struct vfsmount *mnt;
1307         int rc = 0;
1308         unsigned long page;
1309         ENTRY;
1310
1311         dev_clear_rdonly(2);
1312
1313         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1314                 RETURN(rc = -EINVAL);
1315
1316         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1317         if (IS_ERR(obd->obd_fsops))
1318                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1319
1320         if (!(page = __get_free_page(GFP_KERNEL)))
1321                 RETURN(-ENOMEM);
1322
1323         memset((void *)page, 0, PAGE_SIZE);
1324         sprintf((char *)page, "iopen_nopriv");
1325
1326         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1327                             lcfg->lcfg_inlbuf1, (void *)page);
1328         free_page(page);
1329         if (IS_ERR(mnt)) {
1330                 rc = PTR_ERR(mnt);
1331                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1332                 GOTO(err_ops, rc);
1333         }
1334
1335         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1336
1337         sema_init(&mds->mds_orphan_recovery_sem, 1);
1338         sema_init(&mds->mds_epoch_sem, 1);
1339         spin_lock_init(&mds->mds_transno_lock);
1340         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1341         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1342         atomic_set(&mds->mds_open_count, 0);
1343
1344         obd->obd_namespace = ldlm_namespace_new("mds_server",
1345                                                 LDLM_NAMESPACE_SERVER);
1346         if (obd->obd_namespace == NULL) {
1347                 mds_cleanup(obd, 0);
1348                 GOTO(err_put, rc = -ENOMEM);
1349         }
1350
1351         rc = mds_fs_setup(obd, mnt);
1352         if (rc) {
1353                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1354                 GOTO(err_ns, rc);
1355         }
1356
1357         rc = llog_start_commit_thread();
1358         if (rc < 0)
1359                 GOTO(err_fs, rc);
1360
1361
1362         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1363                 class_uuid_t uuid;
1364
1365                 generate_random_uuid(uuid);
1366                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1367
1368                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1369                 if (mds->mds_profile == NULL)
1370                         GOTO(err_fs, rc = -ENOMEM);
1371
1372                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1373                        lcfg->lcfg_inllen3);
1374
1375         }
1376
1377         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1378                            "mds_ldlm_client", &obd->obd_ldlm_client);
1379         obd->obd_replayable = 1;
1380
1381         RETURN(0);
1382
1383 err_fs:
1384         /* No extra cleanup needed for llog_init_commit_thread() */
1385         mds_fs_cleanup(obd, 0);
1386 err_ns:
1387         ldlm_namespace_free(obd->obd_namespace, 0);
1388         obd->obd_namespace = NULL;
1389 err_put:
1390         unlock_kernel();
1391         mntput(mds->mds_vfsmnt);
1392         mds->mds_sb = 0;
1393         lock_kernel();
1394 err_ops:
1395         fsfilt_put_ops(obd->obd_fsops);
1396         return rc;
1397 }
1398
1399 static int mds_postsetup(struct obd_device *obd)
1400 {
1401         struct mds_obd *mds = &obd->u.mds;
1402         int rc = 0;
1403         ENTRY;
1404
1405
1406         rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1407                         &llog_lvfs_ops);
1408         if (rc)
1409                 RETURN(rc);
1410
1411         if (mds->mds_profile) {
1412                 struct obd_run_ctxt saved;
1413                 struct lustre_profile *lprof;
1414                 struct config_llog_instance cfg;
1415
1416                 cfg.cfg_instance = NULL;
1417                 cfg.cfg_uuid = mds->mds_lov_uuid;
1418                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1419                 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1420                                              mds->mds_profile, &cfg);
1421                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1422                 if (rc)
1423                         GOTO(err_llog, rc);
1424
1425                 lprof = class_get_profile(mds->mds_profile);
1426                 if (lprof == NULL) {
1427                         CERROR("No profile found: %s\n", mds->mds_profile);
1428                         GOTO(err_cleanup, rc = -ENOENT);
1429                 }
1430                 rc = mds_lov_connect(obd, lprof->lp_osc);
1431                 if (rc)
1432                         GOTO(err_cleanup, rc);
1433         }
1434
1435         RETURN(rc);
1436
1437 err_cleanup:
1438         mds_lov_clean(obd);
1439 err_llog:
1440         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1441         RETURN(rc);
1442 }
1443
1444 static int mds_postrecov(struct obd_device *obd)
1445
1446 {
1447         int rc, rc2;
1448
1449         LASSERT(!obd->obd_recovering);
1450
1451 #ifdef ENABLE_ORPHANS
1452         rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1453                           obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1454         if (rc != 0) {
1455                 CERROR("faild at llog_origin_connect: %d\n", rc);
1456         }
1457 #endif
1458         rc = mds_cleanup_orphans(obd);
1459
1460         rc2 = mds_lov_set_nextid(obd);
1461         if (rc2 == 0)
1462                 rc2 = rc;
1463         RETURN(rc2);
1464 }
1465
1466 int mds_lov_clean(struct obd_device *obd)
1467 {
1468         struct mds_obd *mds = &obd->u.mds;
1469
1470         if (mds->mds_profile) {
1471                 char * cln_prof;
1472                 struct config_llog_instance cfg;
1473                 struct obd_run_ctxt saved;
1474                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1475
1476                 OBD_ALLOC(cln_prof, len);
1477                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1478
1479                 cfg.cfg_instance = NULL;
1480                 cfg.cfg_uuid = mds->mds_lov_uuid;
1481
1482                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1483                 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1484                                         cln_prof, &cfg);
1485                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1486
1487                 OBD_FREE(cln_prof, len);
1488                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1489                 mds->mds_profile = NULL;
1490         }
1491         RETURN(0);
1492 }
1493
1494 static int mds_precleanup(struct obd_device *obd, int flags)
1495 {
1496         int rc = 0;
1497         ENTRY;
1498
1499         mds_lov_disconnect(obd, flags);
1500         mds_lov_clean(obd);
1501         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1502         RETURN(rc);
1503 }
1504
1505 static int mds_cleanup(struct obd_device *obd, int flags)
1506 {
1507         struct mds_obd *mds = &obd->u.mds;
1508         ENTRY;
1509
1510         if (mds->mds_sb == NULL)
1511                 RETURN(0);
1512
1513         mds_update_server_data(obd, 1);
1514         if (mds->mds_lov_objids != NULL) {
1515                 OBD_FREE(mds->mds_lov_objids,
1516                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1517         }
1518         mds_fs_cleanup(obd, flags);
1519
1520         unlock_kernel();
1521
1522         /* 2 seems normal on mds, (may_umount() also expects 2
1523           fwiw), but we only see 1 at this point in obdfilter. */
1524         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1525                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1526                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1527
1528         mntput(mds->mds_vfsmnt);
1529
1530         mds->mds_sb = 0;
1531
1532         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1533
1534         if (obd->obd_recovering)
1535                 target_cancel_recovery_timer(obd);
1536         lock_kernel();
1537         dev_clear_rdonly(2);
1538         fsfilt_put_ops(obd->obd_fsops);
1539
1540         RETURN(0);
1541 }
1542
1543 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1544                                         struct ldlm_lock *new_lock,
1545                                         struct lustre_handle *lockh)
1546 {
1547         struct obd_export *exp = req->rq_export;
1548         struct obd_device *obd = exp->exp_obd;
1549         struct ldlm_request *dlmreq =
1550                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1551         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1552         struct list_head *iter;
1553
1554         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1555                 return;
1556
1557         l_lock(&obd->obd_namespace->ns_lock);
1558         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1559                 struct ldlm_lock *lock;
1560                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1561                 if (lock == new_lock)
1562                         continue;
1563                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1564                         lockh->cookie = lock->l_handle.h_cookie;
1565                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1566                                   lockh->cookie);
1567                         l_unlock(&obd->obd_namespace->ns_lock);
1568                         return;
1569                 }
1570         }
1571         l_unlock(&obd->obd_namespace->ns_lock);
1572
1573         /* This remote handle isn't enqueued, so we never received or
1574          * processed this request.  Clear MSG_RESENT, because it can
1575          * be handled like any normal request now. */
1576
1577         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1578
1579         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1580                   remote_hdl.cookie);
1581 }
1582
1583 int intent_disposition(struct ldlm_reply *rep, int flag)
1584 {
1585         if (!rep)
1586                 return 0;
1587         return (rep->lock_policy_res1 & flag);
1588 }
1589
1590 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1591 {
1592         if (!rep)
1593                 return;
1594         rep->lock_policy_res1 |= flag;
1595 }
1596
1597 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1598                               struct ldlm_lock **lockp, void *req_cookie,
1599                               ldlm_mode_t mode, int flags, void *data)
1600 {
1601         struct ptlrpc_request *req = req_cookie;
1602         struct ldlm_lock *lock = *lockp;
1603         int rc;
1604         ENTRY;
1605
1606         if (!req_cookie)
1607                 RETURN(0);
1608
1609         if (req->rq_reqmsg->bufcount > 1) {
1610                 /* an intent needs to be considered */
1611                 struct ldlm_intent *it;
1612                 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1613                 struct ldlm_reply *rep;
1614                 struct lustre_handle lockh = { 0 };
1615                 struct ldlm_lock *new_lock;
1616                 int offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1617                                               sizeof(struct mds_body),
1618                                               mds->mds_max_mdsize,
1619                                               mds->mds_max_cookiesize};
1620
1621                 it = lustre_swab_reqbuf(req, 1, sizeof (*it),
1622                                         lustre_swab_ldlm_intent);
1623                 if (it == NULL) {
1624                         CERROR ("Intent missing\n");
1625                         req->rq_status = -EFAULT;
1626                         RETURN(req->rq_status);
1627                 }
1628
1629                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1630                            ldlm_it2str(it->opc));
1631
1632                 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3,
1633                                        repsize, NULL);
1634                 if (rc)
1635                         RETURN(req->rq_status = rc);
1636
1637                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1638                 intent_set_disposition(rep, DISP_IT_EXECD);
1639
1640                 fixup_handle_for_resent_req(req, lock, &lockh);
1641
1642                 /* execute policy */
1643                 switch ((long)it->opc) {
1644                 case IT_OPEN:
1645                 case IT_CREAT|IT_OPEN:
1646                         /* XXX swab here to assert that an mds_open reint
1647                          * packet is following */
1648                         rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1649 #if 0
1650                         /* We abort the lock if the lookup was negative and
1651                          * we did not make it to the OPEN portion */
1652                         if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1653                                 RETURN(ELDLM_LOCK_ABORTED);
1654                         if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1655                             !intent_disposition(rep, DISP_OPEN_OPEN))
1656 #endif
1657                                 RETURN(ELDLM_LOCK_ABORTED);
1658                         break;
1659                 case IT_GETATTR:
1660                 case IT_LOOKUP:
1661                 case IT_READDIR:
1662                         rep->lock_policy_res2 = mds_getattr_name(offset, req,
1663                                                                  &lockh);
1664                         /* FIXME: LDLM can set req->rq_status. MDS sets
1665                            policy_res{1,2} with disposition and status.
1666                            - replay: returns 0 & req->status is old status
1667                            - otherwise: returns req->status */
1668                         if (intent_disposition(rep, DISP_LOOKUP_NEG))
1669                                 rep->lock_policy_res2 = 0;
1670                         if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1671                             rep->lock_policy_res2)
1672                                 RETURN(ELDLM_LOCK_ABORTED);
1673                         if (req->rq_status != 0) {
1674                                 LBUG();
1675                                 rep->lock_policy_res2 = req->rq_status;
1676                                 RETURN(ELDLM_LOCK_ABORTED);
1677                         }
1678                         break;
1679                 default:
1680                         CERROR("Unhandled intent "LPD64"\n", it->opc);
1681                         LBUG();
1682                 }
1683
1684                 /* By this point, whatever function we called above must have
1685                  * either filled in 'lockh', been an intent replay, or returned
1686                  * an error.  We want to allow replayed RPCs to not get a lock,
1687                  * since we would just drop it below anyways because lock replay
1688                  * is done separately by the client afterwards.  For regular
1689                  * RPCs we want to give the new lock to the client instead of
1690                  * whatever lock it was about to get.
1691                  */
1692                 new_lock = ldlm_handle2lock(&lockh);
1693                 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1694                         RETURN(0);
1695
1696                 LASSERT(new_lock != NULL);
1697
1698                 /* If we've already given this lock to a client once, then we
1699                  * should have no readers or writers.  Otherwise, we should
1700                  * have one reader _or_ writer ref (which will be zeroed below)
1701                  * before returning the lock to a client.
1702                  */
1703                 if (new_lock->l_export == req->rq_export) {
1704                         LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1705                 } else {
1706                         LASSERT(new_lock->l_export == NULL);
1707                         LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1708                 }
1709
1710                 *lockp = new_lock;
1711
1712                 if (new_lock->l_export == req->rq_export) {
1713                         /* Already gave this to the client, which means that we
1714                          * reconstructed a reply. */
1715                         LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1716                                 MSG_RESENT);
1717                         RETURN(ELDLM_LOCK_REPLACED);
1718                 }
1719
1720                 /* Fixup the lock to be given to the client */
1721                 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1722                 new_lock->l_readers = 0;
1723                 new_lock->l_writers = 0;
1724
1725                 new_lock->l_export = class_export_get(req->rq_export);
1726                 list_add(&new_lock->l_export_chain,
1727                          &new_lock->l_export->exp_ldlm_data.led_held_locks);
1728
1729                 new_lock->l_blocking_ast = lock->l_blocking_ast;
1730                 new_lock->l_completion_ast = lock->l_completion_ast;
1731
1732                 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1733                        sizeof(lock->l_remote_handle));
1734
1735                 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1736
1737                 LDLM_LOCK_PUT(new_lock);
1738                 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1739
1740                 RETURN(ELDLM_LOCK_REPLACED);
1741         } else {
1742                 int size = sizeof(struct ldlm_reply);
1743                 rc = lustre_pack_reply(req, 1, &size, NULL);
1744                 if (rc) {
1745                         LBUG();
1746                         RETURN(-ENOMEM);
1747                 }
1748         }
1749         RETURN(0);
1750 }
1751
1752 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1753 {
1754         struct lprocfs_static_vars lvars;
1755
1756         lprocfs_init_multi_vars(0, &lvars);
1757         return lprocfs_obd_attach(dev, lvars.obd_vars);
1758 }
1759
1760 int mds_detach(struct obd_device *dev)
1761 {
1762         return lprocfs_obd_detach(dev);
1763 }
1764
1765 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1766 {
1767         struct lprocfs_static_vars lvars;
1768
1769         lprocfs_init_multi_vars(1, &lvars);
1770         return lprocfs_obd_attach(dev, lvars.obd_vars);
1771 }
1772
1773 int mdt_detach(struct obd_device *dev)
1774 {
1775         return lprocfs_obd_detach(dev);
1776 }
1777
1778 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1779 {
1780         struct mds_obd *mds = &obddev->u.mds;
1781         int rc = 0;
1782         ENTRY;
1783
1784         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1785                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1786                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1787                                            mds_handle, "mds",
1788                                            obddev->obd_proc_entry);
1789
1790         if (!mds->mds_service) {
1791                 CERROR("failed to start service\n");
1792                 RETURN(rc = -ENOMEM);
1793         }
1794
1795         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1796                                     "ll_mdt");
1797         if (rc)
1798                 GOTO(err_thread, rc);
1799
1800         mds->mds_setattr_service =
1801                 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1802                                 MDS_BUFSIZE, MDS_MAXREQSIZE,
1803                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1804                                 mds_handle, "mds_setattr",
1805                                 obddev->obd_proc_entry);
1806         if (!mds->mds_setattr_service) {
1807                 CERROR("failed to start getattr service\n");
1808                 GOTO(err_thread, rc = -ENOMEM);
1809         }
1810
1811         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1812                                  MDT_NUM_THREADS, "ll_mdt_attr");
1813         if (rc)
1814                 GOTO(err_thread2, rc);
1815
1816         mds->mds_readpage_service =
1817                 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1818                                 MDS_BUFSIZE, MDS_MAXREQSIZE,
1819                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1820                                 mds_handle, "mds_readpage",
1821                                 obddev->obd_proc_entry);
1822         if (!mds->mds_readpage_service) {
1823                 CERROR("failed to start readpage service\n");
1824                 GOTO(err_thread2, rc = -ENOMEM);
1825         }
1826
1827         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1828                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
1829
1830         if (rc)
1831                 GOTO(err_thread3, rc);
1832
1833         RETURN(0);
1834
1835 err_thread3:
1836         ptlrpc_unregister_service(mds->mds_readpage_service);
1837 err_thread2:
1838         ptlrpc_unregister_service(mds->mds_setattr_service);
1839 err_thread:
1840         ptlrpc_unregister_service(mds->mds_service);
1841         return rc;
1842 }
1843
1844
1845 static int mdt_cleanup(struct obd_device *obddev, int flags)
1846 {
1847         struct mds_obd *mds = &obddev->u.mds;
1848         ENTRY;
1849
1850         ptlrpc_stop_all_threads(mds->mds_readpage_service);
1851         ptlrpc_unregister_service(mds->mds_readpage_service);
1852
1853         ptlrpc_stop_all_threads(mds->mds_setattr_service);
1854         ptlrpc_unregister_service(mds->mds_setattr_service);
1855
1856         ptlrpc_stop_all_threads(mds->mds_service);
1857         ptlrpc_unregister_service(mds->mds_service);
1858
1859         RETURN(0);
1860 }
1861
1862 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1863 {
1864         struct obd_device *obd = data;
1865         struct ll_fid fid;
1866         fid.id = id;
1867         fid.generation = gen;
1868         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1869 }
1870
1871 struct lvfs_callback_ops mds_lvfs_ops = {
1872         l_fid2dentry:     mds_lvfs_fid2dentry,
1873 };
1874
1875 /* use obd ops to offer management infrastructure */
1876 static struct obd_ops mds_obd_ops = {
1877         o_owner:       THIS_MODULE,
1878         o_attach:      mds_attach,
1879         o_detach:      mds_detach,
1880         o_connect:     mds_connect,
1881         o_init_export:  mds_init_export,
1882         o_destroy_export:  mds_destroy_export,
1883         o_disconnect:  mds_disconnect,
1884         o_setup:       mds_setup,
1885         o_postsetup:   mds_postsetup,
1886         o_precleanup:  mds_precleanup,
1887         o_cleanup:     mds_cleanup,
1888         o_postrecov:   mds_postrecov,
1889         o_statfs:      mds_obd_statfs,
1890         o_iocontrol:   mds_iocontrol,
1891         o_create:      mds_obd_create,
1892         o_destroy:     mds_obd_destroy,
1893         o_llog_init:   mds_llog_init,
1894         o_llog_finish: mds_llog_finish,
1895         o_notify:      mds_notify,
1896 };
1897
1898 static struct obd_ops mdt_obd_ops = {
1899         o_owner:       THIS_MODULE,
1900         o_attach:      mdt_attach,
1901         o_detach:      mdt_detach,
1902         o_setup:       mdt_setup,
1903         o_cleanup:     mdt_cleanup,
1904 };
1905
1906 static int __init mds_init(void)
1907 {
1908         struct lprocfs_static_vars lvars;
1909
1910         lprocfs_init_multi_vars(0, &lvars);
1911         class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1912         lprocfs_init_multi_vars(1, &lvars);
1913         class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1914         ldlm_register_intent(ldlm_intent_policy);
1915
1916         return 0;
1917 }
1918
1919 static void /*__exit*/ mds_exit(void)
1920 {
1921         ldlm_unregister_intent();
1922         class_unregister_type(LUSTRE_MDS_NAME);
1923         class_unregister_type(LUSTRE_MDT_NAME);
1924 }
1925
1926 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1927 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1928 MODULE_LICENSE("GPL");
1929
1930 module_init(mds_init);
1931 module_exit(mds_exit);