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_postsetup(struct obd_device *obd);
  60 static int mds_cleanup(struct obd_device *obd, int flags);
  61
  62 /* Assumes caller has already pushed into the kernel filesystem context */
  63 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  64                         loff_t offset, int count)
  65 {
  66         struct ptlrpc_bulk_desc *desc;
  67         struct l_wait_info lwi;
  68         struct page **pages;
  69         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  70         ENTRY;
  71
  72         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  73
  74         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  75         OBD_ALLOC(pages, sizeof(*pages) * npages);
  76         if (!pages)
  77                 GOTO(out, rc = -ENOMEM);
  78
  79         desc = ptlrpc_prep_bulk_exp (req, 1, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
  80         if (desc == NULL)
  81                 GOTO(out_free, rc = -ENOMEM);
  82
  83         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  84                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  85
  86                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  87                 if (pages[i] == NULL)
  88                         GOTO(cleanup_buf, rc = -ENOMEM);
  89
  90                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  91         }
  92
  93         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  94                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  95                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
  96                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
  97                        file->f_dentry->d_inode->i_size);
  98
  99                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 100                                      page_address(pages[i]), tmpsize, &offset);
 101
 102                 if (rc != tmpsize)
 103                         GOTO(cleanup_buf, rc = -EIO);
 104         }
 105
 106         LASSERT(desc->bd_nob == count);
 107
 108         rc = ptlrpc_start_bulk_transfer(desc);
 109         if (rc)
 110                 GOTO(cleanup_buf, rc);
 111
 112         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 113                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 114                        OBD_FAIL_MDS_SENDPAGE, rc);
 115                 GOTO(abort_bulk, rc);
 116         }
 117
 118         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 119         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 120         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 121
 122         if (rc == 0) {
 123                 if (desc->bd_success &&
 124                     desc->bd_nob_transferred == count)
 125                         GOTO(cleanup_buf, rc);
 126
 127                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 128         }
 129
 130         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 131                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 132                   desc->bd_nob_transferred, count,
 133                   req->rq_export->exp_client_uuid.uuid,
 134                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 135
 136         ptlrpc_fail_export(req->rq_export);
 137
 138         EXIT;
 139  abort_bulk:
 140         ptlrpc_abort_bulk (desc);
 141  cleanup_buf:
 142         for (i = 0; i < npages; i++)
 143                 if (pages[i])
 144                         __free_pages(pages[i], 0);
 145
 146         ptlrpc_free_bulk(desc);
 147  out_free:
 148         OBD_FREE(pages, sizeof(*pages) * npages);
 149  out:
 150         return rc;
 151 }
 152
 153 /* only valid locked dentries or errors should be returned */
 154 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 155                                      struct vfsmount **mnt, int lock_mode,
 156                                      struct lustre_handle *lockh,
 157                                      char *name, int namelen)
 158 {
 159         struct mds_obd *mds = &obd->u.mds;
 160         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 161         struct ldlm_res_id res_id = { .name = {0} };
 162         int flags = 0, rc;
 163         ENTRY;
 164
 165         if (IS_ERR(de))
 166                 RETURN(de);
 167
 168         res_id.name[0] = de->d_inode->i_ino;
 169         res_id.name[1] = de->d_inode->i_generation;
 170         rc = ldlm_cli_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, lockh);
 174         if (rc != ELDLM_OK) {
 175                 l_dput(de);
 176                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 177         }
 178
 179         RETURN(retval);
 180 }
 181
 182 #ifndef DCACHE_DISCONNECTED
 183 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 184 #endif
 185
 186
 187 /* Look up an entry by inode number. */
 188 /* this function ONLY returns valid dget'd dentries with an initialized inode
 189    or errors */
 190 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 191                               struct vfsmount **mnt)
 192 {
 193         char fid_name[32];
 194         unsigned long ino = fid->id;
 195         __u32 generation = fid->generation;
 196         struct inode *inode;
 197         struct dentry *result;
 198
 199         if (ino == 0)
 200                 RETURN(ERR_PTR(-ESTALE));
 201
 202         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 203
 204         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 205                ino, generation, mds->mds_sb);
 206
 207         /* under ext3 this is neither supposed to return bad inodes
 208            nor NULL inodes. */
 209         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 210         if (IS_ERR(result))
 211                 RETURN(result);
 212
 213         inode = result->d_inode;
 214         if (!inode)
 215                 RETURN(ERR_PTR(-ENOENT));
 216
 217         if (generation && inode->i_generation != generation) {
 218                 /* we didn't find the right inode.. */
 219                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 220                        inode->i_ino, (unsigned long)inode->i_nlink,
 221                        atomic_read(&inode->i_count), inode->i_generation,
 222                        generation);
 223                 dput(result);
 224                 RETURN(ERR_PTR(-ENOENT));
 225         }
 226
 227         if (mnt) {
 228                 *mnt = mds->mds_vfsmnt;
 229                 mntget(*mnt);
 230         }
 231
 232         RETURN(result);
 233 }
 234
 235
 236 /* Establish a connection to the MDS.
 237  *
 238  * This will set up an export structure for the client to hold state data
 239  * about that client, like open files, the last operation number it did
 240  * on the server, etc.
 241  */
 242 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 243                        struct obd_uuid *cluuid)
 244 {
 245         struct obd_export *exp;
 246         struct mds_export_data *med; /*  */
 247         struct mds_client_data *mcd;
 248         int rc, abort_recovery;
 249         ENTRY;
 250
 251         if (!conn || !obd || !cluuid)
 252                 RETURN(-EINVAL);
 253
 254         /* Check for aborted recovery. */
 255         spin_lock_bh(&obd->obd_processing_task_lock);
 256         abort_recovery = obd->obd_abort_recovery;
 257         spin_unlock_bh(&obd->obd_processing_task_lock);
 258         if (abort_recovery)
 259                 target_abort_recovery(obd);
 260
 261         /* XXX There is a small race between checking the list and adding a
 262          * new connection for the same UUID, but the real threat (list
 263          * corruption when multiple different clients connect) is solved.
 264          *
 265          * There is a second race between adding the export to the list,
 266          * and filling in the client data below.  Hence skipping the case
 267          * of NULL mcd above.  We should already be controlling multiple
 268          * connects at the client, and we can't hold the spinlock over
 269          * memory allocations without risk of deadlocking.
 270          */
 271         rc = class_connect(conn, obd, cluuid);
 272         if (rc)
 273                 RETURN(rc);
 274         exp = class_conn2export(conn);
 275         LASSERT(exp);
 276         med = &exp->exp_mds_data;
 277
 278         OBD_ALLOC(mcd, sizeof(*mcd));
 279         if (!mcd) {
 280                 CERROR("mds: out of memory for client data\n");
 281                 GOTO(out, rc = -ENOMEM);
 282         }
 283
 284         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 285         med->med_mcd = mcd;
 286
 287         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 288         if (rc == 0)
 289                 EXIT;
 290 out:
 291         if (rc) {
 292                 OBD_FREE(mcd, sizeof(*mcd));
 293                 class_disconnect(exp, 0);
 294         }
 295         class_export_put(exp);
 296
 297         return rc;
 298 }
 299
 300 static int mds_init_export(struct obd_export *exp)
 301 {
 302         struct mds_export_data *med = &exp->exp_mds_data;
 303
 304         INIT_LIST_HEAD(&med->med_open_head);
 305         spin_lock_init(&med->med_open_lock);
 306         RETURN(0);
 307 }
 308
 309 static int mds_destroy_export(struct obd_export *export)
 310 {
 311         struct mds_export_data *med;
 312         struct obd_device *obd = export->exp_obd;
 313         struct obd_run_ctxt saved;
 314         int rc = 0;
 315         ENTRY;
 316
 317         med = &export->exp_mds_data;
 318         target_destroy_export(export);
 319
 320         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 321         /* Close any open files (which may also cause orphan unlinking). */
 322         spin_lock(&med->med_open_lock);
 323         while (!list_empty(&med->med_open_head)) {
 324                 struct list_head *tmp = med->med_open_head.next;
 325                 struct mds_file_data *mfd =
 326                         list_entry(tmp, struct mds_file_data, mfd_list);
 327                 BDEVNAME_DECLARE_STORAGE(btmp);
 328
 329                 /* bug 1579: fix force-closing for 2.5 */
 330                 struct dentry *dentry = mfd->mfd_dentry;
 331
 332                 list_del(&mfd->mfd_list);
 333                 spin_unlock(&med->med_open_lock);
 334
 335                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 336                        dentry->d_name.len, dentry->d_name.name,
 337                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 338                        dentry->d_inode->i_ino);
 339                 rc = mds_mfd_close(NULL, obd, mfd,
 340                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 341
 342                 if (rc)
 343                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 344                 spin_lock(&med->med_open_lock);
 345         }
 346         spin_unlock(&med->med_open_lock);
 347         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 348
 349         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 350
 351         RETURN(rc);
 352 }
 353
 354 static int mds_disconnect(struct obd_export *export, int flags)
 355 {
 356         unsigned long irqflags;
 357         int rc;
 358         ENTRY;
 359
 360         ldlm_cancel_locks_for_export(export);
 361
 362         /* complete all outstanding replies */
 363         spin_lock_irqsave (&export->exp_lock, irqflags);
 364         while (!list_empty (&export->exp_outstanding_replies)) {
 365                 struct ptlrpc_reply_state *rs =
 366                         list_entry (export->exp_outstanding_replies.next,
 367                                     struct ptlrpc_reply_state, rs_exp_list);
 368                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 369
 370                 spin_lock (&svc->srv_lock);
 371                 list_del_init (&rs->rs_exp_list);
 372                 ptlrpc_schedule_difficult_reply (rs);
 373                 spin_unlock (&svc->srv_lock);
 374         }
 375         spin_unlock_irqrestore (&export->exp_lock, irqflags);
 376
 377         spin_lock_irqsave(&export->exp_lock, irqflags);
 378         export->exp_flags = flags;
 379         spin_unlock_irqrestore(&export->exp_lock, irqflags);
 380
 381         rc = class_disconnect(export, flags);
 382         RETURN(rc);
 383 }
 384
 385 static int mds_getstatus(struct ptlrpc_request *req)
 386 {
 387         struct mds_obd *mds = mds_req2mds(req);
 388         struct mds_body *body;
 389         int rc, size = sizeof(*body);
 390         ENTRY;
 391
 392         rc = lustre_pack_reply(req, 1, &size, NULL);
 393         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 394                 CERROR("mds: out of memory for message: size=%d\n", size);
 395                 req->rq_status = -ENOMEM;       /* superfluous? */
 396                 RETURN(-ENOMEM);
 397         }
 398
 399         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 400         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 401
 402         /* the last_committed and last_xid fields are filled in for all
 403          * replies already - no need to do so here also.
 404          */
 405         RETURN(0);
 406 }
 407
 408 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 409                      void *data, int flag)
 410 {
 411         int do_ast;
 412         ENTRY;
 413
 414         if (flag == LDLM_CB_CANCELING) {
 415                 /* Don't need to do anything here. */
 416                 RETURN(0);
 417         }
 418
 419         /* XXX layering violation!  -phil */
 420         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 421         /* Get this: if mds_blocking_ast is racing with ldlm_intent_policy,
 422          * such that mds_blocking_ast is called just before l_i_p takes the
 423          * ns_lock, then by the time we get the lock, we might not be the
 424          * correct blocking function anymore.  So check, and return early, if
 425          * so. */
 426         if (lock->l_blocking_ast != mds_blocking_ast) {
 427                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 428                 RETURN(0);
 429         }
 430
 431         lock->l_flags |= LDLM_FL_CBPENDING;
 432         do_ast = (!lock->l_readers && !lock->l_writers);
 433         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 434
 435         if (do_ast) {
 436                 struct lustre_handle lockh;
 437                 int rc;
 438
 439                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 440                 ldlm_lock2handle(lock, &lockh);
 441                 rc = ldlm_cli_cancel(&lockh);
 442                 if (rc < 0)
 443                         CERROR("ldlm_cli_cancel: %d\n", rc);
 444         } else {
 445                 LDLM_DEBUG(lock, "Lock still has references, will be "
 446                            "cancelled later");
 447         }
 448         RETURN(0);
 449 }
 450
 451 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 452  * Call with lock=0 if the caller has already taken the i_sem. */
 453 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 454                 struct mds_body *body, struct inode *inode, int lock)
 455 {
 456         struct mds_obd *mds = &obd->u.mds;
 457         void *lmm;
 458         int lmm_size;
 459         int rc;
 460         ENTRY;
 461
 462         lmm = lustre_msg_buf(msg, offset, 0);
 463         if (lmm == NULL) {
 464                 /* Some problem with getting eadata when I sized the reply
 465                  * buffer... */
 466                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 467                        inode->i_ino);
 468                 RETURN(0);
 469         }
 470         lmm_size = msg->buflens[offset];
 471
 472         /* I don't really like this, but it is a sanity check on the client
 473          * MD request.  However, if the client doesn't know how much space
 474          * to reserve for the MD, it shouldn't be bad to have too much space.
 475          */
 476         if (lmm_size > mds->mds_max_mdsize) {
 477                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 478                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 479                 // RETURN(-EINVAL);
 480         }
 481
 482         if (lock)
 483                 down(&inode->i_sem);
 484         rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
 485         if (lock)
 486                 up(&inode->i_sem);
 487         if (rc < 0) {
 488                 CERROR("Error %d reading eadata for ino %lu\n",
 489                        rc, inode->i_ino);
 490         } else if (rc > 0) {
 491                 lmm_size = rc;
 492                 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
 493
 494                 if (rc > 0)
 495                         lmm_size = rc;
 496                 body->valid |= OBD_MD_FLEASIZE;
 497                 body->eadatasize = lmm_size;
 498                 rc = 0;
 499         }
 500
 501         RETURN(rc);
 502 }
 503
 504 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 505                                 struct ptlrpc_request *req,
 506                                 struct mds_body *reqbody, int reply_off)
 507 {
 508         struct mds_body *body;
 509         struct inode *inode = dentry->d_inode;
 510         int rc = 0;
 511         ENTRY;
 512
 513         if (inode == NULL)
 514                 RETURN(-ENOENT);
 515
 516         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 517         LASSERT(body != NULL);                 /* caller prepped reply */
 518
 519         mds_pack_inode2fid(&body->fid1, inode);
 520         mds_pack_inode2body(body, inode);
 521
 522         if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
 523                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 524                                  inode, 1);
 525
 526                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 527                 if (!(body->valid & OBD_MD_FLEASIZE))
 528                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 529                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 530         } else if (S_ISLNK(inode->i_mode) &&
 531                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 532                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 533                 int len;
 534
 535                 LASSERT (symname != NULL);       /* caller prepped reply */
 536                 len = req->rq_repmsg->buflens[reply_off + 1];
 537
 538                 rc = inode->i_op->readlink(dentry, symname, len);
 539                 if (rc < 0) {
 540                         CERROR("readlink failed: %d\n", rc);
 541                 } else if (rc != len - 1) {
 542                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 543                                 rc, len - 1);
 544                         rc = -EINVAL;
 545                 } else {
 546                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 547                         body->valid |= OBD_MD_LINKNAME;
 548                         body->eadatasize = rc + 1;
 549                         symname[rc] = 0;        /* NULL terminate */
 550                         rc = 0;
 551                 }
 552         }
 553
 554         RETURN(rc);
 555 }
 556
 557 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 558                                 int offset)
 559 {
 560         struct mds_obd *mds = mds_req2mds(req);
 561         struct mds_body *body;
 562         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 563         ENTRY;
 564
 565         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 566         LASSERT(body != NULL);                 /* checked by caller */
 567         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 568
 569         if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
 570                 int rc;
 571                 down(&inode->i_sem);
 572                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 573                 up(&inode->i_sem);
 574                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 575                        rc, inode->i_ino);
 576                 if (rc < 0) {
 577                         if (rc != -ENODATA)
 578                                 CERROR("error getting inode %lu MD: rc = %d\n",
 579                                        inode->i_ino, rc);
 580                         size[bufcount] = 0;
 581                 } else if (rc > mds->mds_max_mdsize) {
 582                         size[bufcount] = 0;
 583                         CERROR("MD size %d larger than maximum possible %u\n",
 584                                rc, mds->mds_max_mdsize);
 585                 } else {
 586                         size[bufcount] = rc;
 587                 }
 588                 bufcount++;
 589         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 590                 if (inode->i_size + 1 != body->eadatasize)
 591                         CERROR("symlink size: %Lu, reply space: %d\n",
 592                                inode->i_size + 1, body->eadatasize);
 593                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 594                 bufcount++;
 595                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 596                        inode->i_size + 1, body->eadatasize);
 597         }
 598
 599         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 600                 CERROR("failed MDS_GETATTR_PACK test\n");
 601                 req->rq_status = -ENOMEM;
 602                 GOTO(out, rc = -ENOMEM);
 603         }
 604
 605         rc = lustre_pack_reply(req, bufcount, size, NULL);
 606         if (rc) {
 607                 CERROR("out of memory\n");
 608                 GOTO(out, req->rq_status = rc);
 609         }
 610
 611         EXIT;
 612  out:
 613         return(rc);
 614 }
 615
 616 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
 617                             struct lustre_handle *child_lockh)
 618 {
 619         struct obd_device *obd = req->rq_export->exp_obd;
 620         struct ldlm_reply *rep = NULL;
 621         struct obd_run_ctxt saved;
 622         struct mds_body *body;
 623         struct dentry *dparent = NULL, *dchild = NULL;
 624         struct obd_ucred uc;
 625         struct lustre_handle parent_lockh;
 626         int namesize;
 627         int rc = 0, cleanup_phase = 0, resent_req = 0;
 628         char *name;
 629         ENTRY;
 630
 631         LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
 632
 633         /* Swab now, before anyone looks inside the request */
 634
 635         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 636                                   lustre_swab_mds_body);
 637         if (body == NULL) {
 638                 CERROR("Can't swab mds_body\n");
 639                 GOTO(cleanup, rc = -EFAULT);
 640         }
 641
 642         LASSERT_REQSWAB(req, offset + 1);
 643         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 644         if (name == NULL) {
 645                 CERROR("Can't unpack name\n");
 646                 GOTO(cleanup, rc = -EFAULT);
 647         }
 648         namesize = req->rq_reqmsg->buflens[offset + 1];
 649
 650         LASSERT (offset == 0 || offset == 2);
 651         /* if requests were at offset 2, the getattr reply goes back at 1 */
 652         if (offset) {
 653                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 654                 offset = 1;
 655         }
 656
 657         uc.ouc_fsuid = body->fsuid;
 658         uc.ouc_fsgid = body->fsgid;
 659         uc.ouc_cap = body->capability;
 660         uc.ouc_suppgid1 = body->suppgid;
 661         uc.ouc_suppgid2 = -1;
 662         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 663         cleanup_phase = 1; /* kernel context */
 664         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 665
 666         /* FIXME: handle raw lookup */
 667 #if 0
 668         if (body->valid == OBD_MD_FLID) {
 669                 struct mds_body *mds_reply;
 670                 int size = sizeof(*mds_reply);
 671                 ino_t inum;
 672                 // The user requested ONLY the inode number, so do a raw lookup
 673                 rc = lustre_pack_reply(req, 1, &size, NULL);
 674                 if (rc) {
 675                         CERROR("out of memory\n");
 676                         GOTO(cleanup, rc);
 677                 }
 678
 679                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 680
 681                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 682                                            sizeof(*mds_reply));
 683                 mds_reply->fid1.id = inum;
 684                 mds_reply->valid = OBD_MD_FLID;
 685                 GOTO(cleanup, rc);
 686         }
 687 #endif
 688
 689         if (child_lockh->cookie != 0) {
 690                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 691                 resent_req = 1;
 692         }
 693
 694         if (resent_req == 0) {
 695                 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
 696                                                  &parent_lockh, &dparent,
 697                                                  LCK_PR, name, namesize,
 698                                                  child_lockh, &dchild, LCK_PR);
 699                 if (rc)
 700                         GOTO(cleanup, rc);
 701         } else {
 702                 struct ldlm_lock *granted_lock;
 703                 struct ll_fid child_fid;
 704                 struct ldlm_resource *res;
 705                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 706                 granted_lock = ldlm_handle2lock(child_lockh);
 707                 LASSERT(granted_lock);
 708
 709                 res = granted_lock->l_resource;
 710                 child_fid.id = res->lr_name.name[0];
 711                 child_fid.generation = res->lr_name.name[1];
 712                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 713                 LASSERT(dchild);
 714                 LDLM_LOCK_PUT(granted_lock);
 715         }
 716
 717         cleanup_phase = 2; /* dchild, dparent, locks */
 718
 719         if (dchild->d_inode == NULL) {
 720                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 721                 /* in the intent case, the policy clears this error:
 722                    the disposition is enough */
 723                 GOTO(cleanup, rc = -ENOENT);
 724         } else {
 725                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 726         }
 727
 728         if (req->rq_repmsg == NULL) {
 729                 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 730                 if (rc != 0) {
 731                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 732                         GOTO (cleanup, rc);
 733                 }
 734         }
 735
 736         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 737         GOTO(cleanup, rc); /* returns the lock to the client */
 738
 739  cleanup:
 740         switch (cleanup_phase) {
 741         case 2:
 742                 if (resent_req == 0) {
 743                         if (rc && dchild->d_inode)
 744                                 ldlm_lock_decref(child_lockh, LCK_PR);
 745                         ldlm_lock_decref(&parent_lockh, LCK_PR);
 746                         l_dput(dparent);
 747                 }
 748                 l_dput(dchild);
 749         case 1:
 750                 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 751         default: ;
 752         }
 753         return rc;
 754 }
 755
 756 static int mds_getattr(int offset, struct ptlrpc_request *req)
 757 {
 758         struct mds_obd *mds = mds_req2mds(req);
 759         struct obd_device *obd = req->rq_export->exp_obd;
 760         struct obd_run_ctxt saved;
 761         struct dentry *de;
 762         struct mds_body *body;
 763         struct obd_ucred uc;
 764         int rc = 0;
 765         ENTRY;
 766
 767         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
 768                                    lustre_swab_mds_body);
 769         if (body == NULL) {
 770                 CERROR ("Can't unpack body\n");
 771                 RETURN (-EFAULT);
 772         }
 773
 774         uc.ouc_fsuid = body->fsuid;
 775         uc.ouc_fsgid = body->fsgid;
 776         uc.ouc_cap = body->capability;
 777         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 778         de = mds_fid2dentry(mds, &body->fid1, NULL);
 779         if (IS_ERR(de)) {
 780                 rc = req->rq_status = -ENOENT;
 781                 GOTO(out_pop, PTR_ERR(de));
 782         }
 783
 784         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
 785         if (rc != 0) {
 786                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
 787                 GOTO (out_pop, rc);
 788         }
 789
 790         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
 791
 792         l_dput(de);
 793         GOTO(out_pop, rc);
 794 out_pop:
 795         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 796         return rc;
 797 }
 798
 799
 800 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 801                           unsigned long max_age)
 802 {
 803         int rc;
 804
 805         spin_lock(&obd->obd_osfs_lock);
 806         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
 807         if (rc == 0)
 808                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
 809         spin_unlock(&obd->obd_osfs_lock);
 810
 811         return rc;
 812 }
 813
 814 static int mds_statfs(struct ptlrpc_request *req)
 815 {
 816         struct obd_device *obd = req->rq_export->exp_obd;
 817         int rc, size = sizeof(struct obd_statfs);
 818         ENTRY;
 819
 820         rc = lustre_pack_reply(req, 1, &size, NULL);
 821         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 822                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
 823                 GOTO(out, rc);
 824         }
 825
 826         /* We call this so that we can cache a bit - 1 jiffie worth */
 827         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
 828                             jiffies - HZ);
 829         if (rc) {
 830                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
 831                 GOTO(out, rc);
 832         }
 833
 834         EXIT;
 835 out:
 836         req->rq_status = rc;
 837         return 0;
 838 }
 839
 840 static int mds_sync(struct ptlrpc_request *req)
 841 {
 842         struct obd_device *obd = req->rq_export->exp_obd;
 843         struct mds_obd *mds = &obd->u.mds;
 844         struct mds_body *body;
 845         int rc, size = sizeof(*body);
 846         ENTRY;
 847
 848         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 849         if (body == NULL)
 850                 GOTO(out, rc = -EPROTO);
 851
 852         rc = lustre_pack_reply(req, 1, &size, NULL);
 853         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
 854                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
 855                 GOTO(out, rc);
 856         }
 857
 858         if (body->fid1.id == 0) {
 859                 /* a fid of zero is taken to mean "sync whole filesystem" */
 860                 rc = fsfilt_sync(obd, mds->mds_sb);
 861                 if (rc)
 862                         GOTO(out, rc);
 863         } else {
 864                 /* just any file to grab fsync method - "file" arg unused */
 865                 struct file *file = mds->mds_rcvd_filp;
 866                 struct dentry *de;
 867
 868                 de = mds_fid2dentry(mds, &body->fid1, NULL);
 869                 if (IS_ERR(de))
 870                         GOTO(out, rc = PTR_ERR(de));
 871
 872                 rc = file->f_op->fsync(NULL, de, 1);
 873                 l_dput(de);
 874                 if (rc)
 875                         GOTO(out, rc);
 876
 877                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 878                 mds_pack_inode2fid(&body->fid1, de->d_inode);
 879                 mds_pack_inode2body(body, de->d_inode);
 880         }
 881 out:
 882         req->rq_status = rc;
 883         return 0;
 884 }
 885
 886 /* mds_readpage does not take a DLM lock on the inode, because the client must
 887  * already have a PR lock.
 888  *
 889  * If we were to take another one here, a deadlock will result, if another
 890  * thread is already waiting for a PW lock. */
 891 static int mds_readpage(struct ptlrpc_request *req)
 892 {
 893         struct obd_device *obd = req->rq_export->exp_obd;
 894         struct vfsmount *mnt;
 895         struct dentry *de;
 896         struct file *file;
 897         struct mds_body *body, *repbody;
 898         struct obd_run_ctxt saved;
 899         int rc, size = sizeof(*repbody);
 900         struct obd_ucred uc;
 901         ENTRY;
 902
 903         rc = lustre_pack_reply(req, 1, &size, NULL);
 904         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 905                 CERROR("mds: out of memory\n");
 906                 GOTO(out, rc = -ENOMEM);
 907         }
 908
 909         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
 910         if (body == NULL)
 911                 GOTO (out, rc = -EFAULT);
 912
 913         uc.ouc_fsuid = body->fsuid;
 914         uc.ouc_fsgid = body->fsgid;
 915         uc.ouc_cap = body->capability;
 916         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 917         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
 918         if (IS_ERR(de))
 919                 GOTO(out_pop, rc = PTR_ERR(de));
 920
 921         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
 922
 923         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 924         /* note: in case of an error, dentry_open puts dentry */
 925         if (IS_ERR(file))
 926                 GOTO(out_pop, rc = PTR_ERR(file));
 927
 928         /* body->size is actually the offset -eeb */
 929         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
 930                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
 931                        body->size, de->d_inode->i_blksize);
 932                 GOTO(out_file, rc = -EFAULT);
 933         }
 934
 935         /* body->nlink is actually the #bytes to read -eeb */
 936         if (body->nlink & (de->d_inode->i_blksize - 1)) {
 937                 CERROR("size %u is not multiple of blocksize %lu\n",
 938                        body->nlink, de->d_inode->i_blksize);
 939                 GOTO(out_file, rc = -EFAULT);
 940         }
 941
 942         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
 943         repbody->size = file->f_dentry->d_inode->i_size;
 944         repbody->valid = OBD_MD_FLSIZE;
 945
 946         /* to make this asynchronous make sure that the handling function
 947            doesn't send a reply when this function completes. Instead a
 948            callback function would send the reply */
 949         /* body->size is actually the offset -eeb */
 950         rc = mds_sendpage(req, file, body->size, body->nlink);
 951
 952 out_file:
 953         filp_close(file, 0);
 954 out_pop:
 955         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 956 out:
 957         req->rq_status = rc;
 958         RETURN(0);
 959 }
 960
 961 int mds_reint(struct ptlrpc_request *req, int offset,
 962               struct lustre_handle *lockh)
 963 {
 964         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
 965         int rc;
 966
 967         OBD_ALLOC(rec, sizeof(*rec));
 968         if (rec == NULL)
 969                 RETURN(-ENOMEM);
 970
 971         rc = mds_update_unpack(req, offset, rec);
 972         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 973                 CERROR("invalid record\n");
 974                 GOTO(out, req->rq_status = -EINVAL);
 975         }
 976         /* rc will be used to interrupt a for loop over multiple records */
 977         rc = mds_reint_rec(rec, offset, req, lockh);
 978  out:
 979         OBD_FREE(rec, sizeof(*rec));
 980         return rc;
 981 }
 982
 983 static int mds_filter_recovery_request(struct ptlrpc_request *req,
 984                                        struct obd_device *obd, int *process)
 985 {
 986         switch (req->rq_reqmsg->opc) {
 987         case MDS_CONNECT: /* This will never get here, but for completeness. */
 988         case OST_CONNECT: /* This will never get here, but for completeness. */
 989         case MDS_DISCONNECT:
 990         case OST_DISCONNECT:
 991                *process = 1;
 992                RETURN(0);
 993
 994         case MDS_CLOSE:
 995         case MDS_SYNC: /* used in unmounting */
 996         case OBD_PING:
 997         case MDS_REINT:
 998         case LDLM_ENQUEUE:
 999                 *process = target_queue_recovery_request(req, obd);
1000                 RETURN(0);
1001
1002         default:
1003                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1004                 *process = 0;
1005                 /* XXX what should we set rq_status to here? */
1006                 req->rq_status = -EAGAIN;
1007                 RETURN(ptlrpc_error(req));
1008         }
1009 }
1010
1011 static char *reint_names[] = {
1012         [REINT_SETATTR] "setattr",
1013         [REINT_CREATE]  "create",
1014         [REINT_LINK]    "link",
1015         [REINT_UNLINK]  "unlink",
1016         [REINT_RENAME]  "rename",
1017         [REINT_OPEN]    "open",
1018 };
1019
1020 int mds_handle(struct ptlrpc_request *req)
1021 {
1022         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1023         int rc = 0;
1024         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1025         struct obd_device *obd = NULL;
1026         ENTRY;
1027
1028         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1029
1030         LASSERT(current->journal_info == NULL);
1031         /* XXX identical to OST */
1032         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1033                 struct mds_export_data *med;
1034                 int recovering, abort_recovery;
1035
1036                 if (req->rq_export == NULL) {
1037                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1038                                req->rq_reqmsg->opc);
1039                         req->rq_status = -ENOTCONN;
1040                         GOTO(out, rc = -ENOTCONN);
1041                 }
1042
1043                 med = &req->rq_export->exp_mds_data;
1044                 obd = req->rq_export->exp_obd;
1045                 mds = &obd->u.mds;
1046
1047                 /* Check for aborted recovery. */
1048                 spin_lock_bh(&obd->obd_processing_task_lock);
1049                 abort_recovery = obd->obd_abort_recovery;
1050                 recovering = obd->obd_recovering;
1051                 spin_unlock_bh(&obd->obd_processing_task_lock);
1052                 if (abort_recovery) {
1053                         target_abort_recovery(obd);
1054                 } else if (recovering) {
1055                         rc = mds_filter_recovery_request(req, obd,
1056                                                          &should_process);
1057                         if (rc || !should_process)
1058                                 RETURN(rc);
1059                 }
1060         }
1061
1062         switch (req->rq_reqmsg->opc) {
1063         case MDS_CONNECT:
1064                 DEBUG_REQ(D_INODE, req, "connect");
1065                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1066                 rc = target_handle_connect(req, mds_handle);
1067                 if (!rc)
1068                         /* Now that we have an export, set mds. */
1069                         mds = mds_req2mds(req);
1070                 break;
1071
1072         case MDS_DISCONNECT:
1073                 DEBUG_REQ(D_INODE, req, "disconnect");
1074                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1075                 rc = target_handle_disconnect(req);
1076                 req->rq_status = rc;            /* superfluous? */
1077                 break;
1078
1079         case MDS_GETSTATUS:
1080                 DEBUG_REQ(D_INODE, req, "getstatus");
1081                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1082                 rc = mds_getstatus(req);
1083                 break;
1084
1085         case MDS_GETATTR:
1086                 DEBUG_REQ(D_INODE, req, "getattr");
1087                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1088                 rc = mds_getattr(0, req);
1089                 break;
1090
1091         case MDS_GETATTR_NAME: {
1092                 struct lustre_handle lockh;
1093                 DEBUG_REQ(D_INODE, req, "getattr_name");
1094                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1095
1096                 /* If this request gets a reconstructed reply, we won't be
1097                  * acquiring any new locks in mds_getattr_name, so we don't
1098                  * want to cancel.
1099                  */
1100                 lockh.cookie = 0;
1101                 rc = mds_getattr_name(0, req, &lockh);
1102                 /* this non-intent call (from an ioctl) is special */
1103                 req->rq_status = rc;
1104                 if (rc == 0 && lockh.cookie)
1105                         ldlm_lock_decref(&lockh, LCK_PR);
1106                 break;
1107         }
1108         case MDS_STATFS:
1109                 DEBUG_REQ(D_INODE, req, "statfs");
1110                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1111                 rc = mds_statfs(req);
1112                 break;
1113
1114         case MDS_READPAGE:
1115                 DEBUG_REQ(D_INODE, req, "readpage");
1116                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1117                 rc = mds_readpage(req);
1118
1119                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1120                         if (req->rq_reply_state) {
1121                                 lustre_free_reply_state (req->rq_reply_state);
1122                                 req->rq_reply_state = NULL;
1123                         }
1124                         RETURN(0);
1125                 }
1126
1127                 break;
1128
1129         case MDS_REINT: {
1130                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1131                 __u32  opc;
1132                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1133                                mds->mds_max_cookiesize};
1134                 int bufcount;
1135
1136                 /* NB only peek inside req now; mds_reint() will swab it */
1137                 if (opcp == NULL) {
1138                         CERROR ("Can't inspect opcode\n");
1139                         rc = -EINVAL;
1140                         break;
1141                 }
1142                 opc = *opcp;
1143                 if (lustre_msg_swabbed (req->rq_reqmsg))
1144                         __swab32s(&opc);
1145
1146                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1147                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1148                            reint_names[opc] == NULL) ? reint_names[opc] :
1149                                                        "unknown opcode");
1150
1151                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1152
1153                 if (opc == REINT_UNLINK)
1154                         bufcount = 3;
1155                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1156                         bufcount = 2;
1157                 else
1158                         bufcount = 1;
1159
1160                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1161                 if (rc)
1162                         break;
1163
1164                 rc = mds_reint(req, 0, NULL);
1165                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1166                 break;
1167         }
1168
1169         case MDS_CLOSE:
1170                 DEBUG_REQ(D_INODE, req, "close");
1171                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1172                 rc = mds_close(req);
1173                 break;
1174
1175         case MDS_DONE_WRITING:
1176                 DEBUG_REQ(D_INODE, req, "done_writing");
1177                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1178                 rc = mds_done_writing(req);
1179                 break;
1180
1181         case MDS_PIN:
1182                 DEBUG_REQ(D_INODE, req, "pin");
1183                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1184                 rc = mds_pin(req);
1185                 break;
1186
1187         case MDS_SYNC:
1188                 DEBUG_REQ(D_INODE, req, "sync");
1189                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1190                 rc = mds_sync(req);
1191                 break;
1192
1193         case OBD_PING:
1194                 DEBUG_REQ(D_INODE, req, "ping");
1195                 rc = target_handle_ping(req);
1196                 break;
1197
1198         case OBD_LOG_CANCEL:
1199                 CDEBUG(D_INODE, "log cancel\n");
1200                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1201                 rc = -ENOTSUPP; /* la la la */
1202                 break;
1203
1204         case LDLM_ENQUEUE:
1205                 DEBUG_REQ(D_INODE, req, "enqueue");
1206                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1207                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1208                                          ldlm_server_blocking_ast);
1209                 break;
1210         case LDLM_CONVERT:
1211                 DEBUG_REQ(D_INODE, req, "convert");
1212                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1213                 rc = ldlm_handle_convert(req);
1214                 break;
1215         case LDLM_BL_CALLBACK:
1216         case LDLM_CP_CALLBACK:
1217                 DEBUG_REQ(D_INODE, req, "callback");
1218                 CERROR("callbacks should not happen on MDS\n");
1219                 LBUG();
1220                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1221                 break;
1222         case LLOG_ORIGIN_HANDLE_CREATE:
1223                 DEBUG_REQ(D_INODE, req, "llog_init");
1224                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1225                 rc = llog_origin_handle_create(req);
1226                 break;
1227         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1228                 DEBUG_REQ(D_INODE, req, "llog next block");
1229                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1230                 rc = llog_origin_handle_next_block(req);
1231                 break;
1232         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1233                 DEBUG_REQ(D_INODE, req, "llog read header");
1234                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1235                 rc = llog_origin_handle_read_header(req);
1236                 break;
1237         case LLOG_ORIGIN_HANDLE_CLOSE:
1238                 DEBUG_REQ(D_INODE, req, "llog close");
1239                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1240                 rc = llog_origin_handle_close(req);
1241                 break;
1242         case LLOG_CATINFO:
1243                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1244                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1245                 rc = llog_catinfo(req);
1246                 break;
1247         default:
1248                 req->rq_status = -ENOTSUPP;
1249                 rc = ptlrpc_error(req);
1250                 RETURN(rc);
1251         }
1252
1253         LASSERT(current->journal_info == NULL);
1254
1255         EXIT;
1256
1257         /* If we're DISCONNECTing, the mds_export_data is already freed */
1258         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1259                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1260                 struct obd_device *obd = list_entry(mds, struct obd_device,
1261                                                     u.mds);
1262                 req->rq_repmsg->last_xid =
1263                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1264
1265                 if (!obd->obd_no_transno) {
1266                         req->rq_repmsg->last_committed =
1267                                 obd->obd_last_committed;
1268                 } else {
1269                         DEBUG_REQ(D_IOCTL, req,
1270                                   "not sending last_committed update");
1271                 }
1272                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1273                        ", xid "LPU64"\n",
1274                        mds->mds_last_transno, obd->obd_last_committed,
1275                        req->rq_xid);
1276         }
1277  out:
1278
1279         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1280                 if (obd && obd->obd_recovering) {
1281                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1282                         return target_queue_final_reply(req, rc);
1283                 }
1284                 /* Lost a race with recovery; let the error path DTRT. */
1285                 rc = req->rq_status = -ENOTCONN;
1286         }
1287
1288         target_send_reply(req, rc, fail);
1289         return 0;
1290 }
1291
1292 /* Update the server data on disk.  This stores the new mount_count and
1293  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1294  * then the server last_rcvd value may be less than that of the clients.
1295  * This will alert us that we may need to do client recovery.
1296  *
1297  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1298  */
1299 int mds_update_server_data(struct obd_device *obd, int force_sync)
1300 {
1301         struct mds_obd *mds = &obd->u.mds;
1302         struct mds_server_data *msd = mds->mds_server_data;
1303         struct file *filp = mds->mds_rcvd_filp;
1304         struct obd_run_ctxt saved;
1305         loff_t off = 0;
1306         int rc;
1307         ENTRY;
1308
1309         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1310         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1311
1312         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1313                mds->mds_mount_count, mds->mds_last_transno);
1314         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1315         if (rc)
1316                 CERROR("error writing MDS server data: rc = %d\n", rc);
1317         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1318
1319         RETURN(rc);
1320 }
1321
1322
1323 /* mount the file system (secretly) */
1324 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1325 {
1326         struct lustre_cfg* lcfg = buf;
1327         struct mds_obd *mds = &obd->u.mds;
1328         struct vfsmount *mnt;
1329         int rc = 0;
1330         unsigned long page;
1331         ENTRY;
1332
1333         dev_clear_rdonly(2);
1334
1335         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1336                 RETURN(rc = -EINVAL);
1337
1338         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1339         if (IS_ERR(obd->obd_fsops))
1340                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1341
1342         if (!(page = __get_free_page(GFP_KERNEL)))
1343                 RETURN(-ENOMEM);
1344
1345         memset((void *)page, 0, PAGE_SIZE);
1346         sprintf((char *)page, "iopen_nopriv");
1347
1348         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1349                             lcfg->lcfg_inlbuf1, (void *)page);
1350         free_page(page);
1351         if (IS_ERR(mnt)) {
1352                 rc = PTR_ERR(mnt);
1353                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1354                 GOTO(err_ops, rc);
1355         }
1356
1357         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1358
1359         sema_init(&mds->mds_orphan_recovery_sem, 1);
1360         sema_init(&mds->mds_epoch_sem, 1);
1361         spin_lock_init(&mds->mds_transno_lock);
1362         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1363         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1364         atomic_set(&mds->mds_open_count, 0);
1365
1366         obd->obd_namespace = ldlm_namespace_new("mds_server",
1367                                                 LDLM_NAMESPACE_SERVER);
1368         if (obd->obd_namespace == NULL) {
1369                 mds_cleanup(obd, 0);
1370                 GOTO(err_put, rc = -ENOMEM);
1371         }
1372
1373         rc = mds_fs_setup(obd, mnt);
1374         if (rc) {
1375                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1376                 GOTO(err_ns, rc);
1377         }
1378
1379         rc = llog_start_commit_thread();
1380         if (rc < 0)
1381                 GOTO(err_fs, rc);
1382
1383
1384         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1385                 class_uuid_t uuid;
1386
1387                 generate_random_uuid(uuid);
1388                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1389
1390                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1391                 if (mds->mds_profile == NULL)
1392                         GOTO(err_fs, rc = -ENOMEM);
1393
1394                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1395                        lcfg->lcfg_inllen3);
1396
1397         }
1398
1399         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1400                            "mds_ldlm_client", &obd->obd_ldlm_client);
1401         obd->obd_replayable = 1;
1402
1403         rc = mds_postsetup(obd);
1404         if (rc)
1405                 GOTO(err_fs, rc);
1406         RETURN(0);
1407
1408 err_fs:
1409         /* No extra cleanup needed for llog_init_commit_thread() */
1410         mds_fs_cleanup(obd, 0);
1411 err_ns:
1412         ldlm_namespace_free(obd->obd_namespace, 0);
1413         obd->obd_namespace = NULL;
1414 err_put:
1415         unlock_kernel();
1416         mntput(mds->mds_vfsmnt);
1417         mds->mds_sb = 0;
1418         lock_kernel();
1419 err_ops:
1420         fsfilt_put_ops(obd->obd_fsops);
1421         return rc;
1422 }
1423
1424 static int mds_postsetup(struct obd_device *obd)
1425 {
1426         struct mds_obd *mds = &obd->u.mds;
1427         int rc = 0;
1428         ENTRY;
1429
1430
1431         rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1432                         &llog_lvfs_ops);
1433         if (rc)
1434                 RETURN(rc);
1435
1436         if (mds->mds_profile) {
1437                 struct obd_run_ctxt saved;
1438                 struct lustre_profile *lprof;
1439                 struct config_llog_instance cfg;
1440
1441                 cfg.cfg_instance = NULL;
1442                 cfg.cfg_uuid = mds->mds_lov_uuid;
1443                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1444                 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1445                                              mds->mds_profile, &cfg);
1446                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1447                 if (rc)
1448                         GOTO(err_llog, rc);
1449
1450                 lprof = class_get_profile(mds->mds_profile);
1451                 if (lprof == NULL) {
1452                         CERROR("No profile found: %s\n", mds->mds_profile);
1453                         GOTO(err_cleanup, rc = -ENOENT);
1454                 }
1455                 rc = mds_lov_connect(obd, lprof->lp_osc);
1456                 if (rc)
1457                         GOTO(err_cleanup, rc);
1458         }
1459
1460         RETURN(rc);
1461
1462 err_cleanup:
1463         mds_lov_clean(obd);
1464 err_llog:
1465         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1466         RETURN(rc);
1467 }
1468
1469 static int mds_postrecov(struct obd_device *obd)
1470
1471 {
1472         int rc, rc2;
1473
1474         LASSERT(!obd->obd_recovering);
1475         LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1476
1477         rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1478                           obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1479         if (rc != 0) {
1480                 CERROR("faild at llog_origin_connect: %d\n", rc);
1481         }
1482
1483         rc = mds_cleanup_orphans(obd);
1484
1485         rc2 = mds_lov_set_nextid(obd);
1486         if (rc2 == 0)
1487                 rc2 = rc;
1488         RETURN(rc2);
1489 }
1490
1491 int mds_lov_clean(struct obd_device *obd)
1492 {
1493         struct mds_obd *mds = &obd->u.mds;
1494
1495         if (mds->mds_profile) {
1496                 char * cln_prof;
1497                 struct config_llog_instance cfg;
1498                 struct obd_run_ctxt saved;
1499                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1500
1501                 OBD_ALLOC(cln_prof, len);
1502                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1503
1504                 cfg.cfg_instance = NULL;
1505                 cfg.cfg_uuid = mds->mds_lov_uuid;
1506
1507                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1508                 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1509                                         cln_prof, &cfg);
1510                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1511
1512                 OBD_FREE(cln_prof, len);
1513                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1514                 mds->mds_profile = NULL;
1515         }
1516         RETURN(0);
1517 }
1518
1519 static int mds_precleanup(struct obd_device *obd, int flags)
1520 {
1521         int rc = 0;
1522         ENTRY;
1523
1524         mds_lov_disconnect(obd, flags);
1525         mds_lov_clean(obd);
1526         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1527         RETURN(rc);
1528 }
1529
1530 static int mds_cleanup(struct obd_device *obd, int flags)
1531 {
1532         struct mds_obd *mds = &obd->u.mds;
1533         ENTRY;
1534
1535         if (mds->mds_sb == NULL)
1536                 RETURN(0);
1537
1538         mds_update_server_data(obd, 1);
1539         if (mds->mds_lov_objids != NULL) {
1540                 OBD_FREE(mds->mds_lov_objids,
1541                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1542         }
1543         mds_fs_cleanup(obd, flags);
1544
1545         unlock_kernel();
1546
1547         /* 2 seems normal on mds, (may_umount() also expects 2
1548           fwiw), but we only see 1 at this point in obdfilter. */
1549         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1550                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1551                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1552
1553         mntput(mds->mds_vfsmnt);
1554
1555         mds->mds_sb = 0;
1556
1557         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1558
1559         spin_lock_bh(&obd->obd_processing_task_lock);
1560         if (obd->obd_recovering) {
1561                 target_cancel_recovery_timer(obd);
1562                 obd->obd_recovering = 0;
1563         }
1564         spin_unlock_bh(&obd->obd_processing_task_lock);
1565
1566         lock_kernel();
1567         dev_clear_rdonly(2);
1568         fsfilt_put_ops(obd->obd_fsops);
1569
1570         RETURN(0);
1571 }
1572
1573 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1574                                         struct ldlm_lock *new_lock,
1575                                         struct lustre_handle *lockh)
1576 {
1577         struct obd_export *exp = req->rq_export;
1578         struct obd_device *obd = exp->exp_obd;
1579         struct ldlm_request *dlmreq =
1580                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1581         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1582         struct list_head *iter;
1583
1584         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1585                 return;
1586
1587         l_lock(&obd->obd_namespace->ns_lock);
1588         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1589                 struct ldlm_lock *lock;
1590                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1591                 if (lock == new_lock)
1592                         continue;
1593                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1594                         lockh->cookie = lock->l_handle.h_cookie;
1595                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1596                                   lockh->cookie);
1597                         l_unlock(&obd->obd_namespace->ns_lock);
1598                         return;
1599                 }
1600         }
1601         l_unlock(&obd->obd_namespace->ns_lock);
1602
1603         /* This remote handle isn't enqueued, so we never received or
1604          * processed this request.  Clear MSG_RESENT, because it can
1605          * be handled like any normal request now. */
1606
1607         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1608
1609         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1610                   remote_hdl.cookie);
1611 }
1612
1613 int intent_disposition(struct ldlm_reply *rep, int flag)
1614 {
1615         if (!rep)
1616                 return 0;
1617         return (rep->lock_policy_res1 & flag);
1618 }
1619
1620 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1621 {
1622         if (!rep)
1623                 return;
1624         rep->lock_policy_res1 |= flag;
1625 }
1626
1627 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1628                               struct ldlm_lock **lockp, void *req_cookie,
1629                               ldlm_mode_t mode, int flags, void *data)
1630 {
1631         struct ptlrpc_request *req = req_cookie;
1632         struct ldlm_lock *lock = *lockp;
1633         int rc;
1634         ENTRY;
1635
1636         if (!req_cookie)
1637                 RETURN(0);
1638
1639         if (req->rq_reqmsg->bufcount > 1) {
1640                 /* an intent needs to be considered */
1641                 struct ldlm_intent *it;
1642                 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1643                 struct ldlm_reply *rep;
1644                 struct lustre_handle lockh = { 0 };
1645                 struct ldlm_lock *new_lock;
1646                 int offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1647                                               sizeof(struct mds_body),
1648                                               mds->mds_max_mdsize,
1649                                               mds->mds_max_cookiesize};
1650
1651                 it = lustre_swab_reqbuf(req, 1, sizeof (*it),
1652                                         lustre_swab_ldlm_intent);
1653                 if (it == NULL) {
1654                         CERROR ("Intent missing\n");
1655                         req->rq_status = -EFAULT;
1656                         RETURN(req->rq_status);
1657                 }
1658
1659                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1660                            ldlm_it2str(it->opc));
1661
1662                 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3,
1663                                        repsize, NULL);
1664                 if (rc)
1665                         RETURN(req->rq_status = rc);
1666
1667                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1668                 intent_set_disposition(rep, DISP_IT_EXECD);
1669
1670                 fixup_handle_for_resent_req(req, lock, &lockh);
1671
1672                 /* execute policy */
1673                 switch ((long)it->opc) {
1674                 case IT_OPEN:
1675                 case IT_CREAT|IT_OPEN:
1676                         /* XXX swab here to assert that an mds_open reint
1677                          * packet is following */
1678                         rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1679 #if 0
1680                         /* We abort the lock if the lookup was negative and
1681                          * we did not make it to the OPEN portion */
1682                         if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1683                                 RETURN(ELDLM_LOCK_ABORTED);
1684                         if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1685                             !intent_disposition(rep, DISP_OPEN_OPEN))
1686 #endif
1687                                 RETURN(ELDLM_LOCK_ABORTED);
1688                         break;
1689                 case IT_GETATTR:
1690                 case IT_LOOKUP:
1691                 case IT_READDIR:
1692                         rep->lock_policy_res2 = mds_getattr_name(offset, req,
1693                                                                  &lockh);
1694                         /* FIXME: LDLM can set req->rq_status. MDS sets
1695                            policy_res{1,2} with disposition and status.
1696                            - replay: returns 0 & req->status is old status
1697                            - otherwise: returns req->status */
1698                         if (intent_disposition(rep, DISP_LOOKUP_NEG))
1699                                 rep->lock_policy_res2 = 0;
1700                         if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1701                             rep->lock_policy_res2)
1702                                 RETURN(ELDLM_LOCK_ABORTED);
1703                         if (req->rq_status != 0) {
1704                                 LBUG();
1705                                 rep->lock_policy_res2 = req->rq_status;
1706                                 RETURN(ELDLM_LOCK_ABORTED);
1707                         }
1708                         break;
1709                 default:
1710                         CERROR("Unhandled intent "LPD64"\n", it->opc);
1711                         LBUG();
1712                 }
1713
1714                 /* By this point, whatever function we called above must have
1715                  * either filled in 'lockh', been an intent replay, or returned
1716                  * an error.  We want to allow replayed RPCs to not get a lock,
1717                  * since we would just drop it below anyways because lock replay
1718                  * is done separately by the client afterwards.  For regular
1719                  * RPCs we want to give the new lock to the client instead of
1720                  * whatever lock it was about to get.
1721                  */
1722                 new_lock = ldlm_handle2lock(&lockh);
1723                 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1724                         RETURN(0);
1725
1726                 LASSERT(new_lock != NULL);
1727
1728                 /* If we've already given this lock to a client once, then we
1729                  * should have no readers or writers.  Otherwise, we should
1730                  * have one reader _or_ writer ref (which will be zeroed below)
1731                  * before returning the lock to a client.
1732                  */
1733                 if (new_lock->l_export == req->rq_export) {
1734                         LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1735                 } else {
1736                         LASSERT(new_lock->l_export == NULL);
1737                         LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1738                 }
1739
1740                 *lockp = new_lock;
1741
1742                 if (new_lock->l_export == req->rq_export) {
1743                         /* Already gave this to the client, which means that we
1744                          * reconstructed a reply. */
1745                         LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1746                                 MSG_RESENT);
1747                         RETURN(ELDLM_LOCK_REPLACED);
1748                 }
1749
1750                 /* Fixup the lock to be given to the client */
1751                 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1752                 new_lock->l_readers = 0;
1753                 new_lock->l_writers = 0;
1754
1755                 new_lock->l_export = class_export_get(req->rq_export);
1756                 list_add(&new_lock->l_export_chain,
1757                          &new_lock->l_export->exp_ldlm_data.led_held_locks);
1758
1759                 new_lock->l_blocking_ast = lock->l_blocking_ast;
1760                 new_lock->l_completion_ast = lock->l_completion_ast;
1761
1762                 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1763                        sizeof(lock->l_remote_handle));
1764
1765                 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1766
1767                 LDLM_LOCK_PUT(new_lock);
1768                 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1769
1770                 RETURN(ELDLM_LOCK_REPLACED);
1771         } else {
1772                 int size = sizeof(struct ldlm_reply);
1773                 rc = lustre_pack_reply(req, 1, &size, NULL);
1774                 if (rc) {
1775                         LBUG();
1776                         RETURN(-ENOMEM);
1777                 }
1778         }
1779         RETURN(0);
1780 }
1781
1782 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1783 {
1784         struct lprocfs_static_vars lvars;
1785
1786         lprocfs_init_multi_vars(0, &lvars);
1787         return lprocfs_obd_attach(dev, lvars.obd_vars);
1788 }
1789
1790 int mds_detach(struct obd_device *dev)
1791 {
1792         return lprocfs_obd_detach(dev);
1793 }
1794
1795 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1796 {
1797         struct lprocfs_static_vars lvars;
1798
1799         lprocfs_init_multi_vars(1, &lvars);
1800         return lprocfs_obd_attach(dev, lvars.obd_vars);
1801 }
1802
1803 int mdt_detach(struct obd_device *dev)
1804 {
1805         return lprocfs_obd_detach(dev);
1806 }
1807
1808 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1809 {
1810         struct mds_obd *mds = &obddev->u.mds;
1811         int rc = 0;
1812         ENTRY;
1813
1814         mds->mds_service =
1815                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1816                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1817                                 mds_handle, "mds",
1818                                 obddev->obd_proc_entry);
1819
1820         if (!mds->mds_service) {
1821                 CERROR("failed to start service\n");
1822                 RETURN(rc = -ENOMEM);
1823         }
1824
1825         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1826                                     "ll_mdt");
1827         if (rc)
1828                 GOTO(err_thread, rc);
1829
1830         mds->mds_setattr_service =
1831                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1832                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1833                                 mds_handle, "mds_setattr",
1834                                 obddev->obd_proc_entry);
1835         if (!mds->mds_setattr_service) {
1836                 CERROR("failed to start getattr service\n");
1837                 GOTO(err_thread, rc = -ENOMEM);
1838         }
1839
1840         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1841                                  MDT_NUM_THREADS, "ll_mdt_attr");
1842         if (rc)
1843                 GOTO(err_thread2, rc);
1844
1845         mds->mds_readpage_service =
1846                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1847                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1848                                 mds_handle, "mds_readpage",
1849                                 obddev->obd_proc_entry);
1850         if (!mds->mds_readpage_service) {
1851                 CERROR("failed to start readpage service\n");
1852                 GOTO(err_thread2, rc = -ENOMEM);
1853         }
1854
1855         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1856                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
1857
1858         if (rc)
1859                 GOTO(err_thread3, rc);
1860
1861         RETURN(0);
1862
1863 err_thread3:
1864         ptlrpc_unregister_service(mds->mds_readpage_service);
1865 err_thread2:
1866         ptlrpc_unregister_service(mds->mds_setattr_service);
1867 err_thread:
1868         ptlrpc_unregister_service(mds->mds_service);
1869         return rc;
1870 }
1871
1872
1873 static int mdt_cleanup(struct obd_device *obddev, int flags)
1874 {
1875         struct mds_obd *mds = &obddev->u.mds;
1876         ENTRY;
1877
1878         ptlrpc_stop_all_threads(mds->mds_readpage_service);
1879         ptlrpc_unregister_service(mds->mds_readpage_service);
1880
1881         ptlrpc_stop_all_threads(mds->mds_setattr_service);
1882         ptlrpc_unregister_service(mds->mds_setattr_service);
1883
1884         ptlrpc_stop_all_threads(mds->mds_service);
1885         ptlrpc_unregister_service(mds->mds_service);
1886
1887         RETURN(0);
1888 }
1889
1890 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1891 {
1892         struct obd_device *obd = data;
1893         struct ll_fid fid;
1894         fid.id = id;
1895         fid.generation = gen;
1896         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1897 }
1898
1899 struct lvfs_callback_ops mds_lvfs_ops = {
1900         l_fid2dentry:     mds_lvfs_fid2dentry,
1901 };
1902
1903 /* use obd ops to offer management infrastructure */
1904 static struct obd_ops mds_obd_ops = {
1905         o_owner:       THIS_MODULE,
1906         o_attach:      mds_attach,
1907         o_detach:      mds_detach,
1908         o_connect:     mds_connect,
1909         o_init_export:  mds_init_export,
1910         o_destroy_export:  mds_destroy_export,
1911         o_disconnect:  mds_disconnect,
1912         o_setup:       mds_setup,
1913         o_precleanup:  mds_precleanup,
1914         o_cleanup:     mds_cleanup,
1915         o_postrecov:   mds_postrecov,
1916         o_statfs:      mds_obd_statfs,
1917         o_iocontrol:   mds_iocontrol,
1918         o_create:      mds_obd_create,
1919         o_destroy:     mds_obd_destroy,
1920         o_llog_init:   mds_llog_init,
1921         o_llog_finish: mds_llog_finish,
1922         o_notify:      mds_notify,
1923 };
1924
1925 static struct obd_ops mdt_obd_ops = {
1926         o_owner:       THIS_MODULE,
1927         o_attach:      mdt_attach,
1928         o_detach:      mdt_detach,
1929         o_setup:       mdt_setup,
1930         o_cleanup:     mdt_cleanup,
1931 };
1932
1933 static int __init mds_init(void)
1934 {
1935         struct lprocfs_static_vars lvars;
1936
1937         lprocfs_init_multi_vars(0, &lvars);
1938         class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1939         lprocfs_init_multi_vars(1, &lvars);
1940         class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1941         ldlm_register_intent(ldlm_intent_policy);
1942
1943         return 0;
1944 }
1945
1946 static void /*__exit*/ mds_exit(void)
1947 {
1948         ldlm_unregister_intent();
1949         class_unregister_type(LUSTRE_MDS_NAME);
1950         class_unregister_type(LUSTRE_MDT_NAME);
1951 }
1952
1953 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1954 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1955 MODULE_LICENSE("GPL");
1956
1957 module_init(mds_init);
1958 module_exit(mds_exit);