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 static int mds_bulk_timeout(void *data)
  63 {
  64         struct ptlrpc_bulk_desc *desc = data;
  65         struct obd_export *exp = desc->bd_export;
  66
  67         DEBUG_REQ(D_ERROR, desc->bd_req,"bulk send timed out: evicting %s@%s\n",
  68                   exp->exp_client_uuid.uuid,
  69                   exp->exp_connection->c_remote_uuid.uuid);
  70         ptlrpc_fail_export(exp);
  71         ptlrpc_abort_bulk (desc);
  72         RETURN(1);
  73 }
  74
  75 /* Assumes caller has already pushed into the kernel filesystem context */
  76 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  77                         loff_t offset, int count)
  78 {
  79         struct ptlrpc_bulk_desc *desc;
  80         struct l_wait_info lwi;
  81         struct page **pages;
  82         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  83         ENTRY;
  84
  85         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  86
  87         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  88         OBD_ALLOC(pages, sizeof(*pages) * npages);
  89         if (!pages)
  90                 GOTO(out, rc = -ENOMEM);
  91
  92         desc = ptlrpc_prep_bulk_exp (req, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
  93         if (desc == NULL)
  94                 GOTO(out_free, rc = -ENOMEM);
  95
  96         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  97                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  98
  99                 pages[i] = alloc_pages(GFP_KERNEL, 0);
 100                 if (pages[i] == NULL)
 101                         GOTO(cleanup_buf, rc = -ENOMEM);
 102
 103                 rc = ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 104                 if (rc != 0)
 105                         GOTO(cleanup_buf, rc);
 106         }
 107
 108         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 109                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 110                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 111                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 112                        file->f_dentry->d_inode->i_size);
 113
 114                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 115                                      page_address(pages[i]), tmpsize, &offset);
 116
 117                 if (rc != tmpsize)
 118                         GOTO(cleanup_buf, rc = -EIO);
 119         }
 120
 121         rc = ptlrpc_bulk_put(desc);
 122         if (rc)
 123                 GOTO(cleanup_buf, rc);
 124
 125         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 126                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 127                        OBD_FAIL_MDS_SENDPAGE, rc);
 128                 ptlrpc_abort_bulk(desc);
 129                 GOTO(cleanup_buf, rc);
 130         }
 131
 132         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, mds_bulk_timeout, desc);
 133         rc = l_wait_event(desc->bd_waitq, ptlrpc_bulk_complete (desc), &lwi);
 134         if (rc) {
 135                 LASSERT (rc == -ETIMEDOUT);
 136                 GOTO(cleanup_buf, rc);
 137         }
 138
 139         EXIT;
 140  cleanup_buf:
 141         for (i = 0; i < npages; i++)
 142                 if (pages[i])
 143                         __free_pages(pages[i], 0);
 144
 145         ptlrpc_free_bulk(desc);
 146  out_free:
 147         OBD_FREE(pages, sizeof(*pages) * npages);
 148  out:
 149         return rc;
 150 }
 151
 152 /* only valid locked dentries or errors should be returned */
 153 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 154                                      struct vfsmount **mnt, int lock_mode,
 155                                      struct lustre_handle *lockh,
 156                                      char *name, int namelen)
 157 {
 158         struct mds_obd *mds = &obd->u.mds;
 159         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 160         struct ldlm_res_id res_id = { .name = {0} };
 161         int flags = 0, rc;
 162         ENTRY;
 163
 164         if (IS_ERR(de))
 165                 RETURN(de);
 166
 167         res_id.name[0] = de->d_inode->i_ino;
 168         res_id.name[1] = de->d_inode->i_generation;
 169         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 170                               res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 171                               &flags, ldlm_completion_ast,
 172                               mds_blocking_ast, NULL, lockh);
 173         if (rc != ELDLM_OK) {
 174                 l_dput(de);
 175                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 176         }
 177
 178         RETURN(retval);
 179 }
 180
 181 #ifndef DCACHE_DISCONNECTED
 182 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 183 #endif
 184
 185
 186 /* Look up an entry by inode number. */
 187 /* this function ONLY returns valid dget'd dentries with an initialized inode
 188    or errors */
 189 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 190                               struct vfsmount **mnt)
 191 {
 192         char fid_name[32];
 193         unsigned long ino = fid->id;
 194         __u32 generation = fid->generation;
 195         struct inode *inode;
 196         struct dentry *result;
 197
 198         if (ino == 0)
 199                 RETURN(ERR_PTR(-ESTALE));
 200
 201         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 202
 203         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 204                ino, generation, mds->mds_sb);
 205
 206         /* under ext3 this is neither supposed to return bad inodes
 207            nor NULL inodes. */
 208         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 209         if (IS_ERR(result))
 210                 RETURN(result);
 211
 212         inode = result->d_inode;
 213         if (!inode)
 214                 RETURN(ERR_PTR(-ENOENT));
 215
 216         if (generation && inode->i_generation != generation) {
 217                 /* we didn't find the right inode.. */
 218                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 219                        inode->i_ino, (unsigned long)inode->i_nlink,
 220                        atomic_read(&inode->i_count), inode->i_generation,
 221                        generation);
 222                 dput(result);
 223                 RETURN(ERR_PTR(-ENOENT));
 224         }
 225
 226         if (mnt) {
 227                 *mnt = mds->mds_vfsmnt;
 228                 mntget(*mnt);
 229         }
 230
 231         RETURN(result);
 232 }
 233
 234
 235 /* Establish a connection to the MDS.
 236  *
 237  * This will set up an export structure for the client to hold state data
 238  * about that client, like open files, the last operation number it did
 239  * on the server, etc.
 240  */
 241 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 242                        struct obd_uuid *cluuid)
 243 {
 244         struct obd_export *exp;
 245         struct mds_export_data *med; /*  */
 246         struct mds_client_data *mcd;
 247         int rc, abort_recovery;
 248         ENTRY;
 249
 250         if (!conn || !obd || !cluuid)
 251                 RETURN(-EINVAL);
 252
 253         /* Check for aborted recovery. */
 254         spin_lock_bh(&obd->obd_processing_task_lock);
 255         abort_recovery = obd->obd_abort_recovery;
 256         spin_unlock_bh(&obd->obd_processing_task_lock);
 257         if (abort_recovery)
 258                 target_abort_recovery(obd);
 259
 260         /* XXX There is a small race between checking the list and adding a
 261          * new connection for the same UUID, but the real threat (list
 262          * corruption when multiple different clients connect) is solved.
 263          *
 264          * There is a second race between adding the export to the list,
 265          * and filling in the client data below.  Hence skipping the case
 266          * of NULL mcd above.  We should already be controlling multiple
 267          * connects at the client, and we can't hold the spinlock over
 268          * memory allocations without risk of deadlocking.
 269          */
 270         rc = class_connect(conn, obd, cluuid);
 271         if (rc)
 272                 RETURN(rc);
 273         exp = class_conn2export(conn);
 274         LASSERT(exp);
 275         med = &exp->exp_mds_data;
 276
 277         OBD_ALLOC(mcd, sizeof(*mcd));
 278         if (!mcd) {
 279                 CERROR("mds: out of memory for client data\n");
 280                 GOTO(out, rc = -ENOMEM);
 281         }
 282
 283         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 284         med->med_mcd = mcd;
 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         return fsfilt_statfs(obd, obd->u.mds.mds_sb, osfs);
 788 }
 789
 790 static int mds_statfs(struct ptlrpc_request *req)
 791 {
 792         struct obd_device *obd = req->rq_export->exp_obd;
 793         int rc, size = sizeof(struct obd_statfs);
 794         ENTRY;
 795
 796         rc = lustre_pack_reply(req, 1, &size, NULL);
 797         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 798                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
 799                 GOTO(out, rc);
 800         }
 801
 802         /* We call this so that we can cache a bit - 1 jiffie worth */
 803         rc = obd_statfs(obd, lustre_msg_buf(req->rq_repmsg,0,size),jiffies-HZ);
 804         if (rc) {
 805                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
 806                 GOTO(out, rc);
 807         }
 808
 809         EXIT;
 810 out:
 811         req->rq_status = rc;
 812         return 0;
 813 }
 814
 815 static int mds_sync(struct ptlrpc_request *req)
 816 {
 817         struct obd_device *obd = req->rq_export->exp_obd;
 818         struct mds_obd *mds = &obd->u.mds;
 819         struct mds_body *body;
 820         int rc, size = sizeof(*body);
 821         ENTRY;
 822
 823         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 824         if (body == NULL)
 825                 GOTO(out, rc = -EPROTO);
 826
 827         rc = lustre_pack_reply(req, 1, &size, NULL);
 828         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
 829                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
 830                 GOTO(out, rc);
 831         }
 832
 833         if (body->fid1.id == 0) {
 834                 /* a fid of zero is taken to mean "sync whole filesystem" */
 835                 rc = fsfilt_sync(obd, mds->mds_sb);
 836                 if (rc)
 837                         GOTO(out, rc);
 838         } else {
 839                 /* just any file to grab fsync method - "file" arg unused */
 840                 struct file *file = mds->mds_rcvd_filp;
 841                 struct dentry *de;
 842
 843                 de = mds_fid2dentry(mds, &body->fid1, NULL);
 844                 if (IS_ERR(de))
 845                         GOTO(out, rc = PTR_ERR(de));
 846
 847                 rc = file->f_op->fsync(NULL, de, 1);
 848                 l_dput(de);
 849                 if (rc)
 850                         GOTO(out, rc);
 851
 852                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 853                 mds_pack_inode2fid(&body->fid1, de->d_inode);
 854                 mds_pack_inode2body(body, de->d_inode);
 855         }
 856 out:
 857         req->rq_status = rc;
 858         return 0;
 859 }
 860
 861 /* mds_readpage does not take a DLM lock on the inode, because the client must
 862  * already have a PR lock.
 863  *
 864  * If we were to take another one here, a deadlock will result, if another
 865  * thread is already waiting for a PW lock. */
 866 static int mds_readpage(struct ptlrpc_request *req)
 867 {
 868         struct obd_device *obd = req->rq_export->exp_obd;
 869         struct vfsmount *mnt;
 870         struct dentry *de;
 871         struct file *file;
 872         struct mds_body *body, *repbody;
 873         struct obd_run_ctxt saved;
 874         int rc, size = sizeof(*repbody);
 875         struct obd_ucred uc;
 876         ENTRY;
 877
 878         rc = lustre_pack_reply(req, 1, &size, NULL);
 879         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 880                 CERROR("mds: out of memory\n");
 881                 GOTO(out, rc = -ENOMEM);
 882         }
 883
 884         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
 885         if (body == NULL)
 886                 GOTO (out, rc = -EFAULT);
 887
 888         uc.ouc_fsuid = body->fsuid;
 889         uc.ouc_fsgid = body->fsgid;
 890         uc.ouc_cap = body->capability;
 891         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 892         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
 893         if (IS_ERR(de))
 894                 GOTO(out_pop, rc = PTR_ERR(de));
 895
 896         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
 897
 898         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 899         /* note: in case of an error, dentry_open puts dentry */
 900         if (IS_ERR(file))
 901                 GOTO(out_pop, rc = PTR_ERR(file));
 902
 903         /* body->size is actually the offset -eeb */
 904         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
 905                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
 906                        body->size, de->d_inode->i_blksize);
 907                 GOTO(out_file, rc = -EFAULT);
 908         }
 909
 910         /* body->nlink is actually the #bytes to read -eeb */
 911         if (body->nlink & (de->d_inode->i_blksize - 1)) {
 912                 CERROR("size %u is not multiple of blocksize %lu\n",
 913                        body->nlink, de->d_inode->i_blksize);
 914                 GOTO(out_file, rc = -EFAULT);
 915         }
 916
 917         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
 918         repbody->size = file->f_dentry->d_inode->i_size;
 919         repbody->valid = OBD_MD_FLSIZE;
 920
 921         /* to make this asynchronous make sure that the handling function
 922            doesn't send a reply when this function completes. Instead a
 923            callback function would send the reply */
 924         /* body->size is actually the offset -eeb */
 925         rc = mds_sendpage(req, file, body->size, body->nlink);
 926
 927 out_file:
 928         filp_close(file, 0);
 929 out_pop:
 930         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 931 out:
 932         req->rq_status = rc;
 933         RETURN(0);
 934 }
 935
 936 int mds_reint(struct ptlrpc_request *req, int offset,
 937               struct lustre_handle *lockh)
 938 {
 939         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
 940         int rc;
 941
 942         OBD_ALLOC(rec, sizeof(*rec));
 943         if (rec == NULL)
 944                 RETURN(-ENOMEM);
 945
 946         rc = mds_update_unpack(req, offset, rec);
 947         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 948                 CERROR("invalid record\n");
 949                 GOTO(out, req->rq_status = -EINVAL);
 950         }
 951         /* rc will be used to interrupt a for loop over multiple records */
 952         rc = mds_reint_rec(rec, offset, req, lockh);
 953  out:
 954         OBD_FREE(rec, sizeof(*rec));
 955         return rc;
 956 }
 957
 958 static int mds_filter_recovery_request(struct ptlrpc_request *req,
 959                                        struct obd_device *obd, int *process)
 960 {
 961         switch (req->rq_reqmsg->opc) {
 962         case MDS_CONNECT: /* This will never get here, but for completeness. */
 963         case OST_CONNECT: /* This will never get here, but for completeness. */
 964         case MDS_DISCONNECT:
 965         case OST_DISCONNECT:
 966                *process = 1;
 967                RETURN(0);
 968
 969         case MDS_CLOSE:
 970         case MDS_SYNC: /* used in unmounting */
 971         case OBD_PING:
 972         case MDS_REINT:
 973         case LDLM_ENQUEUE:
 974                 *process = target_queue_recovery_request(req, obd);
 975                 RETURN(0);
 976
 977         default:
 978                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 979                 *process = 0;
 980                 /* XXX what should we set rq_status to here? */
 981                 req->rq_status = -EAGAIN;
 982                 RETURN(ptlrpc_error(req));
 983         }
 984 }
 985
 986 static char *reint_names[] = {
 987         [REINT_SETATTR] "setattr",
 988         [REINT_CREATE]  "create",
 989         [REINT_LINK]    "link",
 990         [REINT_UNLINK]  "unlink",
 991         [REINT_RENAME]  "rename",
 992         [REINT_OPEN]    "open",
 993 };
 994
 995 int mds_handle(struct ptlrpc_request *req)
 996 {
 997         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
 998         int rc = 0;
 999         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1000         struct obd_device *obd = NULL;
1001         ENTRY;
1002
1003         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1004
1005         LASSERT(current->journal_info == NULL);
1006         /* XXX identical to OST */
1007         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1008                 struct mds_export_data *med;
1009                 int recovering, abort_recovery;
1010
1011                 if (req->rq_export == NULL) {
1012                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1013                                req->rq_reqmsg->opc);
1014                         req->rq_status = -ENOTCONN;
1015                         GOTO(out, rc = -ENOTCONN);
1016                 }
1017
1018                 med = &req->rq_export->exp_mds_data;
1019                 obd = req->rq_export->exp_obd;
1020                 mds = &obd->u.mds;
1021
1022                 /* Check for aborted recovery. */
1023                 spin_lock_bh(&obd->obd_processing_task_lock);
1024                 abort_recovery = obd->obd_abort_recovery;
1025                 recovering = obd->obd_recovering;
1026                 spin_unlock_bh(&obd->obd_processing_task_lock);
1027                 if (abort_recovery) {
1028                         target_abort_recovery(obd);
1029                 } else if (recovering) {
1030                         rc = mds_filter_recovery_request(req, obd,
1031                                                          &should_process);
1032                         if (rc || !should_process)
1033                                 RETURN(rc);
1034                 }
1035         }
1036
1037         switch (req->rq_reqmsg->opc) {
1038         case MDS_CONNECT:
1039                 DEBUG_REQ(D_INODE, req, "connect");
1040                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1041                 rc = target_handle_connect(req, mds_handle);
1042                 if (!rc)
1043                         /* Now that we have an export, set mds. */
1044                         mds = mds_req2mds(req);
1045                 break;
1046
1047         case MDS_DISCONNECT:
1048                 DEBUG_REQ(D_INODE, req, "disconnect");
1049                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1050                 rc = target_handle_disconnect(req);
1051                 req->rq_status = rc;            /* superfluous? */
1052                 break;
1053
1054         case MDS_GETSTATUS:
1055                 DEBUG_REQ(D_INODE, req, "getstatus");
1056                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1057                 rc = mds_getstatus(req);
1058                 break;
1059
1060         case MDS_GETATTR:
1061                 DEBUG_REQ(D_INODE, req, "getattr");
1062                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1063                 rc = mds_getattr(0, req);
1064                 break;
1065
1066         case MDS_GETATTR_NAME: {
1067                 struct lustre_handle lockh;
1068                 DEBUG_REQ(D_INODE, req, "getattr_name");
1069                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1070
1071                 /* If this request gets a reconstructed reply, we won't be
1072                  * acquiring any new locks in mds_getattr_name, so we don't
1073                  * want to cancel.
1074                  */
1075                 lockh.cookie = 0;
1076                 rc = mds_getattr_name(0, req, &lockh);
1077                 /* this non-intent call (from an ioctl) is special */
1078                 req->rq_status = rc;
1079                 if (rc == 0 && lockh.cookie)
1080                         ldlm_lock_decref(&lockh, LCK_PR);
1081                 break;
1082         }
1083         case MDS_STATFS:
1084                 DEBUG_REQ(D_INODE, req, "statfs");
1085                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1086                 rc = mds_statfs(req);
1087                 break;
1088
1089         case MDS_READPAGE:
1090                 DEBUG_REQ(D_INODE, req, "readpage");
1091                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1092                 rc = mds_readpage(req);
1093
1094                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SENDPAGE, 0);
1095
1096                 break;
1097
1098         case MDS_REINT: {
1099                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1100                 __u32  opc;
1101                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1102                                mds->mds_max_cookiesize};
1103                 int bufcount;
1104
1105                 /* NB only peek inside req now; mds_reint() will swab it */
1106                 if (opcp == NULL) {
1107                         CERROR ("Can't inspect opcode\n");
1108                         rc = -EINVAL;
1109                         break;
1110                 }
1111                 opc = *opcp;
1112                 if (lustre_msg_swabbed (req->rq_reqmsg))
1113                         __swab32s(&opc);
1114
1115                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1116                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1117                            reint_names[opc] == NULL) ? reint_names[opc] :
1118                                                        "unknown opcode");
1119
1120                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1121
1122                 if (opc == REINT_UNLINK)
1123                         bufcount = 3;
1124                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1125                         bufcount = 2;
1126                 else
1127                         bufcount = 1;
1128
1129                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1130                 if (rc)
1131                         break;
1132
1133                 rc = mds_reint(req, 0, NULL);
1134                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1135                 break;
1136         }
1137
1138         case MDS_CLOSE:
1139                 DEBUG_REQ(D_INODE, req, "close");
1140                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1141                 rc = mds_close(req);
1142                 break;
1143
1144         case MDS_DONE_WRITING:
1145                 DEBUG_REQ(D_INODE, req, "done_writing");
1146                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1147                 rc = mds_done_writing(req);
1148                 break;
1149
1150         case MDS_PIN:
1151                 DEBUG_REQ(D_INODE, req, "pin");
1152                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1153                 rc = mds_pin(req);
1154                 break;
1155
1156         case MDS_SYNC:
1157                 DEBUG_REQ(D_INODE, req, "sync");
1158                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1159                 rc = mds_sync(req);
1160                 break;
1161
1162         case OBD_PING:
1163                 DEBUG_REQ(D_INODE, req, "ping");
1164                 rc = target_handle_ping(req);
1165                 break;
1166
1167         case OBD_LOG_CANCEL:
1168                 CDEBUG(D_INODE, "log cancel\n");
1169                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1170                 rc = -ENOTSUPP; /* la la la */
1171                 break;
1172
1173         case LDLM_ENQUEUE:
1174                 DEBUG_REQ(D_INODE, req, "enqueue");
1175                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1176                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1177                                          ldlm_server_blocking_ast);
1178                 break;
1179         case LDLM_CONVERT:
1180                 DEBUG_REQ(D_INODE, req, "convert");
1181                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1182                 rc = ldlm_handle_convert(req);
1183                 break;
1184         case LDLM_BL_CALLBACK:
1185         case LDLM_CP_CALLBACK:
1186                 DEBUG_REQ(D_INODE, req, "callback");
1187                 CERROR("callbacks should not happen on MDS\n");
1188                 LBUG();
1189                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1190                 break;
1191         case LLOG_ORIGIN_HANDLE_CREATE:
1192                 DEBUG_REQ(D_INODE, req, "llog_init");
1193                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1194                 rc = llog_origin_handle_create(req);
1195                 break;
1196         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1197                 DEBUG_REQ(D_INODE, req, "llog next block");
1198                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1199                 rc = llog_origin_handle_next_block(req);
1200                 break;
1201         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1202                 DEBUG_REQ(D_INODE, req, "llog read header");
1203                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1204                 rc = llog_origin_handle_read_header(req);
1205                 break;
1206         case LLOG_ORIGIN_HANDLE_CLOSE:
1207                 DEBUG_REQ(D_INODE, req, "llog close");
1208                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1209                 rc = llog_origin_handle_close(req);
1210                 break;
1211         case LLOG_CATINFO:
1212                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1213                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1214                 rc = llog_catinfo(req);
1215                 break;
1216         default:
1217                 req->rq_status = -ENOTSUPP;
1218                 rc = ptlrpc_error(req);
1219                 RETURN(rc);
1220         }
1221
1222         LASSERT(current->journal_info == NULL);
1223
1224         EXIT;
1225
1226         /* If we're DISCONNECTing, the mds_export_data is already freed */
1227         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1228                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1229                 struct obd_device *obd = list_entry(mds, struct obd_device,
1230                                                     u.mds);
1231                 req->rq_repmsg->last_xid =
1232                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1233
1234                 if (!obd->obd_no_transno) {
1235                         req->rq_repmsg->last_committed =
1236                                 obd->obd_last_committed;
1237                 } else {
1238                         DEBUG_REQ(D_IOCTL, req,
1239                                   "not sending last_committed update");
1240                 }
1241                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1242                        ", xid "LPU64"\n",
1243                        mds->mds_last_transno, obd->obd_last_committed,
1244                        req->rq_xid);
1245         }
1246  out:
1247
1248         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1249                 if (obd && obd->obd_recovering) {
1250                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1251                         return target_queue_final_reply(req, rc);
1252                 }
1253                 /* Lost a race with recovery; let the error path DTRT. */
1254                 rc = req->rq_status = -ENOTCONN;
1255         }
1256
1257         target_send_reply(req, rc, fail);
1258         return 0;
1259 }
1260
1261 /* Update the server data on disk.  This stores the new mount_count and
1262  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1263  * then the server last_rcvd value may be less than that of the clients.
1264  * This will alert us that we may need to do client recovery.
1265  *
1266  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1267  */
1268 int mds_update_server_data(struct obd_device *obd, int force_sync)
1269 {
1270         struct mds_obd *mds = &obd->u.mds;
1271         struct mds_server_data *msd = mds->mds_server_data;
1272         struct file *filp = mds->mds_rcvd_filp;
1273         struct obd_run_ctxt saved;
1274         loff_t off = 0;
1275         int rc;
1276         ENTRY;
1277
1278         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1279         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1280
1281         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1282                mds->mds_mount_count, mds->mds_last_transno);
1283         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1284         if (rc)
1285                 CERROR("error writing MDS server data: rc = %d\n", rc);
1286         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1287
1288         RETURN(rc);
1289 }
1290
1291
1292 /* mount the file system (secretly) */
1293 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1294 {
1295         struct lustre_cfg* lcfg = buf;
1296         struct mds_obd *mds = &obd->u.mds;
1297         struct vfsmount *mnt;
1298         int rc = 0;
1299         unsigned long page;
1300         ENTRY;
1301
1302         dev_clear_rdonly(2);
1303
1304         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1305                 RETURN(rc = -EINVAL);
1306
1307         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1308         if (IS_ERR(obd->obd_fsops))
1309                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1310
1311         if (!(page = __get_free_page(GFP_KERNEL)))
1312                 RETURN(-ENOMEM);
1313
1314         memset((void *)page, 0, PAGE_SIZE);
1315         sprintf((char *)page, "iopen_nopriv");
1316
1317         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1318                             lcfg->lcfg_inlbuf1, (void *)page);
1319         free_page(page);
1320         if (IS_ERR(mnt)) {
1321                 rc = PTR_ERR(mnt);
1322                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1323                 GOTO(err_ops, rc);
1324         }
1325
1326         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1327
1328         sema_init(&mds->mds_orphan_recovery_sem, 1);
1329         sema_init(&mds->mds_epoch_sem, 1);
1330         spin_lock_init(&mds->mds_transno_lock);
1331         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1332         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1333         atomic_set(&mds->mds_open_count, 0);
1334
1335         obd->obd_namespace = ldlm_namespace_new("mds_server",
1336                                                 LDLM_NAMESPACE_SERVER);
1337         if (obd->obd_namespace == NULL) {
1338                 mds_cleanup(obd, 0);
1339                 GOTO(err_put, rc = -ENOMEM);
1340         }
1341
1342         rc = mds_fs_setup(obd, mnt);
1343         if (rc) {
1344                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1345                 GOTO(err_ns, rc);
1346         }
1347
1348         rc = llog_start_commit_thread();
1349         if (rc < 0)
1350                 GOTO(err_fs, rc);
1351
1352
1353         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1354                 class_uuid_t uuid;
1355
1356                 generate_random_uuid(uuid);
1357                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1358
1359                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1360                 if (mds->mds_profile == NULL)
1361                         GOTO(err_fs, rc = -ENOMEM);
1362
1363                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1364                        lcfg->lcfg_inllen3);
1365
1366         }
1367
1368         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1369                            "mds_ldlm_client", &obd->obd_ldlm_client);
1370         obd->obd_replayable = 1;
1371
1372         rc = mds_postsetup(obd);
1373         if (rc)
1374                 GOTO(err_fs, rc);
1375         RETURN(0);
1376
1377 err_fs:
1378         /* No extra cleanup needed for llog_init_commit_thread() */
1379         mds_fs_cleanup(obd, 0);
1380 err_ns:
1381         ldlm_namespace_free(obd->obd_namespace, 0);
1382         obd->obd_namespace = NULL;
1383 err_put:
1384         unlock_kernel();
1385         mntput(mds->mds_vfsmnt);
1386         mds->mds_sb = 0;
1387         lock_kernel();
1388 err_ops:
1389         fsfilt_put_ops(obd->obd_fsops);
1390         return rc;
1391 }
1392
1393 static int mds_postsetup(struct obd_device *obd)
1394 {
1395         struct mds_obd *mds = &obd->u.mds;
1396         int rc = 0;
1397         ENTRY;
1398
1399
1400         rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1401                         &llog_lvfs_ops);
1402         if (rc)
1403                 RETURN(rc);
1404
1405         if (mds->mds_profile) {
1406                 struct obd_run_ctxt saved;
1407                 struct lustre_profile *lprof;
1408                 struct config_llog_instance cfg;
1409
1410                 cfg.cfg_instance = NULL;
1411                 cfg.cfg_uuid = mds->mds_lov_uuid;
1412                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1413                 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1414                                              mds->mds_profile, &cfg);
1415                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1416                 if (rc)
1417                         GOTO(err_llog, rc);
1418
1419                 lprof = class_get_profile(mds->mds_profile);
1420                 if (lprof == NULL) {
1421                         CERROR("No profile found: %s\n", mds->mds_profile);
1422                         GOTO(err_cleanup, rc = -ENOENT);
1423                 }
1424                 rc = mds_lov_connect(obd, lprof->lp_osc);
1425                 if (rc)
1426                         GOTO(err_cleanup, rc);
1427         }
1428
1429         RETURN(rc);
1430
1431 err_cleanup:
1432         mds_lov_clean(obd);
1433 err_llog:
1434         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1435         RETURN(rc);
1436 }
1437
1438 static int mds_postrecov(struct obd_device *obd)
1439
1440 {
1441         int rc, rc2;
1442
1443         LASSERT(!obd->obd_recovering);
1444         LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1445
1446         rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1447                           obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1448         if (rc != 0) {
1449                 CERROR("faild at llog_origin_connect: %d\n", rc);
1450         }
1451
1452         rc = mds_cleanup_orphans(obd);
1453
1454         rc2 = mds_lov_set_nextid(obd);
1455         if (rc2 == 0)
1456                 rc2 = rc;
1457         RETURN(rc2);
1458 }
1459
1460 int mds_lov_clean(struct obd_device *obd)
1461 {
1462         struct mds_obd *mds = &obd->u.mds;
1463
1464         if (mds->mds_profile) {
1465                 char * cln_prof;
1466                 struct config_llog_instance cfg;
1467                 struct obd_run_ctxt saved;
1468                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1469
1470                 OBD_ALLOC(cln_prof, len);
1471                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1472
1473                 cfg.cfg_instance = NULL;
1474                 cfg.cfg_uuid = mds->mds_lov_uuid;
1475
1476                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1477                 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1478                                         cln_prof, &cfg);
1479                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1480
1481                 OBD_FREE(cln_prof, len);
1482                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1483                 mds->mds_profile = NULL;
1484         }
1485         RETURN(0);
1486 }
1487
1488 static int mds_precleanup(struct obd_device *obd, int flags)
1489 {
1490         int rc = 0;
1491         ENTRY;
1492
1493         mds_lov_disconnect(obd, flags);
1494         mds_lov_clean(obd);
1495         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1496         RETURN(rc);
1497 }
1498
1499 static int mds_cleanup(struct obd_device *obd, int flags)
1500 {
1501         struct mds_obd *mds = &obd->u.mds;
1502         ENTRY;
1503
1504         if (mds->mds_sb == NULL)
1505                 RETURN(0);
1506
1507         mds_update_server_data(obd, 1);
1508         if (mds->mds_lov_objids != NULL) {
1509                 OBD_FREE(mds->mds_lov_objids,
1510                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1511         }
1512         mds_fs_cleanup(obd, flags);
1513
1514         unlock_kernel();
1515
1516         /* 2 seems normal on mds, (may_umount() also expects 2
1517           fwiw), but we only see 1 at this point in obdfilter. */
1518         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1519                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1520                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1521
1522         mntput(mds->mds_vfsmnt);
1523
1524         mds->mds_sb = 0;
1525
1526         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1527
1528         spin_lock_bh(&obd->obd_processing_task_lock);
1529         if (obd->obd_recovering) {
1530                 target_cancel_recovery_timer(obd);
1531                 obd->obd_recovering = 0;
1532         }
1533         spin_unlock_bh(&obd->obd_processing_task_lock);
1534
1535         lock_kernel();
1536         dev_clear_rdonly(2);
1537         fsfilt_put_ops(obd->obd_fsops);
1538
1539         RETURN(0);
1540 }
1541
1542 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1543                                         struct ldlm_lock *new_lock,
1544                                         struct lustre_handle *lockh)
1545 {
1546         struct obd_export *exp = req->rq_export;
1547         struct obd_device *obd = exp->exp_obd;
1548         struct ldlm_request *dlmreq =
1549                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1550         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1551         struct list_head *iter;
1552
1553         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1554                 return;
1555
1556         l_lock(&obd->obd_namespace->ns_lock);
1557         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1558                 struct ldlm_lock *lock;
1559                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1560                 if (lock == new_lock)
1561                         continue;
1562                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1563                         lockh->cookie = lock->l_handle.h_cookie;
1564                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1565                                   lockh->cookie);
1566                         l_unlock(&obd->obd_namespace->ns_lock);
1567                         return;
1568                 }
1569         }
1570         l_unlock(&obd->obd_namespace->ns_lock);
1571
1572         /* This remote handle isn't enqueued, so we never received or
1573          * processed this request.  Clear MSG_RESENT, because it can
1574          * be handled like any normal request now. */
1575
1576         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1577
1578         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1579                   remote_hdl.cookie);
1580 }
1581
1582 int intent_disposition(struct ldlm_reply *rep, int flag)
1583 {
1584         if (!rep)
1585                 return 0;
1586         return (rep->lock_policy_res1 & flag);
1587 }
1588
1589 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1590 {
1591         if (!rep)
1592                 return;
1593         rep->lock_policy_res1 |= flag;
1594 }
1595
1596 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1597                               struct ldlm_lock **lockp, void *req_cookie,
1598                               ldlm_mode_t mode, int flags, void *data)
1599 {
1600         struct ptlrpc_request *req = req_cookie;
1601         struct ldlm_lock *lock = *lockp;
1602         int rc;
1603         ENTRY;
1604
1605         if (!req_cookie)
1606                 RETURN(0);
1607
1608         if (req->rq_reqmsg->bufcount > 1) {
1609                 /* an intent needs to be considered */
1610                 struct ldlm_intent *it;
1611                 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1612                 struct ldlm_reply *rep;
1613                 struct lustre_handle lockh = { 0 };
1614                 struct ldlm_lock *new_lock;
1615                 int offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1616                                               sizeof(struct mds_body),
1617                                               mds->mds_max_mdsize,
1618                                               mds->mds_max_cookiesize};
1619
1620                 it = lustre_swab_reqbuf(req, 1, sizeof (*it),
1621                                         lustre_swab_ldlm_intent);
1622                 if (it == NULL) {
1623                         CERROR ("Intent missing\n");
1624                         req->rq_status = -EFAULT;
1625                         RETURN(req->rq_status);
1626                 }
1627
1628                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1629                            ldlm_it2str(it->opc));
1630
1631                 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3,
1632                                        repsize, NULL);
1633                 if (rc)
1634                         RETURN(req->rq_status = rc);
1635
1636                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1637                 intent_set_disposition(rep, DISP_IT_EXECD);
1638
1639                 fixup_handle_for_resent_req(req, lock, &lockh);
1640
1641                 /* execute policy */
1642                 switch ((long)it->opc) {
1643                 case IT_OPEN:
1644                 case IT_CREAT|IT_OPEN:
1645                         /* XXX swab here to assert that an mds_open reint
1646                          * packet is following */
1647                         rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1648 #if 0
1649                         /* We abort the lock if the lookup was negative and
1650                          * we did not make it to the OPEN portion */
1651                         if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1652                                 RETURN(ELDLM_LOCK_ABORTED);
1653                         if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1654                             !intent_disposition(rep, DISP_OPEN_OPEN))
1655 #endif
1656                                 RETURN(ELDLM_LOCK_ABORTED);
1657                         break;
1658                 case IT_GETATTR:
1659                 case IT_LOOKUP:
1660                 case IT_READDIR:
1661                         rep->lock_policy_res2 = mds_getattr_name(offset, req,
1662                                                                  &lockh);
1663                         /* FIXME: LDLM can set req->rq_status. MDS sets
1664                            policy_res{1,2} with disposition and status.
1665                            - replay: returns 0 & req->status is old status
1666                            - otherwise: returns req->status */
1667                         if (intent_disposition(rep, DISP_LOOKUP_NEG))
1668                                 rep->lock_policy_res2 = 0;
1669                         if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1670                             rep->lock_policy_res2)
1671                                 RETURN(ELDLM_LOCK_ABORTED);
1672                         if (req->rq_status != 0) {
1673                                 LBUG();
1674                                 rep->lock_policy_res2 = req->rq_status;
1675                                 RETURN(ELDLM_LOCK_ABORTED);
1676                         }
1677                         break;
1678                 default:
1679                         CERROR("Unhandled intent "LPD64"\n", it->opc);
1680                         LBUG();
1681                 }
1682
1683                 /* By this point, whatever function we called above must have
1684                  * either filled in 'lockh', been an intent replay, or returned
1685                  * an error.  We want to allow replayed RPCs to not get a lock,
1686                  * since we would just drop it below anyways because lock replay
1687                  * is done separately by the client afterwards.  For regular
1688                  * RPCs we want to give the new lock to the client instead of
1689                  * whatever lock it was about to get.
1690                  */
1691                 new_lock = ldlm_handle2lock(&lockh);
1692                 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1693                         RETURN(0);
1694
1695                 LASSERT(new_lock != NULL);
1696
1697                 /* If we've already given this lock to a client once, then we
1698                  * should have no readers or writers.  Otherwise, we should
1699                  * have one reader _or_ writer ref (which will be zeroed below)
1700                  * before returning the lock to a client.
1701                  */
1702                 if (new_lock->l_export == req->rq_export) {
1703                         LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1704                 } else {
1705                         LASSERT(new_lock->l_export == NULL);
1706                         LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1707                 }
1708
1709                 *lockp = new_lock;
1710
1711                 if (new_lock->l_export == req->rq_export) {
1712                         /* Already gave this to the client, which means that we
1713                          * reconstructed a reply. */
1714                         LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1715                                 MSG_RESENT);
1716                         RETURN(ELDLM_LOCK_REPLACED);
1717                 }
1718
1719                 /* Fixup the lock to be given to the client */
1720                 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1721                 new_lock->l_readers = 0;
1722                 new_lock->l_writers = 0;
1723
1724                 new_lock->l_export = class_export_get(req->rq_export);
1725                 list_add(&new_lock->l_export_chain,
1726                          &new_lock->l_export->exp_ldlm_data.led_held_locks);
1727
1728                 new_lock->l_blocking_ast = lock->l_blocking_ast;
1729                 new_lock->l_completion_ast = lock->l_completion_ast;
1730
1731                 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1732                        sizeof(lock->l_remote_handle));
1733
1734                 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1735
1736                 LDLM_LOCK_PUT(new_lock);
1737                 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1738
1739                 RETURN(ELDLM_LOCK_REPLACED);
1740         } else {
1741                 int size = sizeof(struct ldlm_reply);
1742                 rc = lustre_pack_reply(req, 1, &size, NULL);
1743                 if (rc) {
1744                         LBUG();
1745                         RETURN(-ENOMEM);
1746                 }
1747         }
1748         RETURN(0);
1749 }
1750
1751 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1752 {
1753         struct lprocfs_static_vars lvars;
1754
1755         lprocfs_init_multi_vars(0, &lvars);
1756         return lprocfs_obd_attach(dev, lvars.obd_vars);
1757 }
1758
1759 int mds_detach(struct obd_device *dev)
1760 {
1761         return lprocfs_obd_detach(dev);
1762 }
1763
1764 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1765 {
1766         struct lprocfs_static_vars lvars;
1767
1768         lprocfs_init_multi_vars(1, &lvars);
1769         return lprocfs_obd_attach(dev, lvars.obd_vars);
1770 }
1771
1772 int mdt_detach(struct obd_device *dev)
1773 {
1774         return lprocfs_obd_detach(dev);
1775 }
1776
1777 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1778 {
1779         struct mds_obd *mds = &obddev->u.mds;
1780         int rc = 0;
1781         ENTRY;
1782
1783         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1784                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1785                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1786                                            mds_handle, "mds",
1787                                            obddev->obd_proc_entry);
1788
1789         if (!mds->mds_service) {
1790                 CERROR("failed to start service\n");
1791                 RETURN(rc = -ENOMEM);
1792         }
1793
1794         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1795                                     "ll_mdt");
1796         if (rc)
1797                 GOTO(err_thread, rc);
1798
1799         mds->mds_setattr_service =
1800                 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1801                                 MDS_BUFSIZE, MDS_MAXREQSIZE,
1802                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1803                                 mds_handle, "mds_setattr",
1804                                 obddev->obd_proc_entry);
1805         if (!mds->mds_setattr_service) {
1806                 CERROR("failed to start getattr service\n");
1807                 GOTO(err_thread, rc = -ENOMEM);
1808         }
1809
1810         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1811                                  MDT_NUM_THREADS, "ll_mdt_attr");
1812         if (rc)
1813                 GOTO(err_thread2, rc);
1814
1815         mds->mds_readpage_service =
1816                 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1817                                 MDS_BUFSIZE, MDS_MAXREQSIZE,
1818                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1819                                 mds_handle, "mds_readpage",
1820                                 obddev->obd_proc_entry);
1821         if (!mds->mds_readpage_service) {
1822                 CERROR("failed to start readpage service\n");
1823                 GOTO(err_thread2, rc = -ENOMEM);
1824         }
1825
1826         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1827                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
1828
1829         if (rc)
1830                 GOTO(err_thread3, rc);
1831
1832         RETURN(0);
1833
1834 err_thread3:
1835         ptlrpc_unregister_service(mds->mds_readpage_service);
1836 err_thread2:
1837         ptlrpc_unregister_service(mds->mds_setattr_service);
1838 err_thread:
1839         ptlrpc_unregister_service(mds->mds_service);
1840         return rc;
1841 }
1842
1843
1844 static int mdt_cleanup(struct obd_device *obddev, int flags)
1845 {
1846         struct mds_obd *mds = &obddev->u.mds;
1847         ENTRY;
1848
1849         ptlrpc_stop_all_threads(mds->mds_readpage_service);
1850         ptlrpc_unregister_service(mds->mds_readpage_service);
1851
1852         ptlrpc_stop_all_threads(mds->mds_setattr_service);
1853         ptlrpc_unregister_service(mds->mds_setattr_service);
1854
1855         ptlrpc_stop_all_threads(mds->mds_service);
1856         ptlrpc_unregister_service(mds->mds_service);
1857
1858         RETURN(0);
1859 }
1860
1861 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1862 {
1863         struct obd_device *obd = data;
1864         struct ll_fid fid;
1865         fid.id = id;
1866         fid.generation = gen;
1867         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1868 }
1869
1870 struct lvfs_callback_ops mds_lvfs_ops = {
1871         l_fid2dentry:     mds_lvfs_fid2dentry,
1872 };
1873
1874 /* use obd ops to offer management infrastructure */
1875 static struct obd_ops mds_obd_ops = {
1876         o_owner:       THIS_MODULE,
1877         o_attach:      mds_attach,
1878         o_detach:      mds_detach,
1879         o_connect:     mds_connect,
1880         o_init_export:  mds_init_export,
1881         o_destroy_export:  mds_destroy_export,
1882         o_disconnect:  mds_disconnect,
1883         o_setup:       mds_setup,
1884         o_precleanup:  mds_precleanup,
1885         o_cleanup:     mds_cleanup,
1886         o_postrecov:   mds_postrecov,
1887         o_statfs:      mds_obd_statfs,
1888         o_iocontrol:   mds_iocontrol,
1889         o_create:      mds_obd_create,
1890         o_destroy:     mds_obd_destroy,
1891         o_llog_init:   mds_llog_init,
1892         o_llog_finish: mds_llog_finish,
1893         o_notify:      mds_notify,
1894 };
1895
1896 static struct obd_ops mdt_obd_ops = {
1897         o_owner:       THIS_MODULE,
1898         o_attach:      mdt_attach,
1899         o_detach:      mdt_detach,
1900         o_setup:       mdt_setup,
1901         o_cleanup:     mdt_cleanup,
1902 };
1903
1904 static int __init mds_init(void)
1905 {
1906         struct lprocfs_static_vars lvars;
1907
1908         lprocfs_init_multi_vars(0, &lvars);
1909         class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1910         lprocfs_init_multi_vars(1, &lvars);
1911         class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1912         ldlm_register_intent(ldlm_intent_policy);
1913
1914         return 0;
1915 }
1916
1917 static void /*__exit*/ mds_exit(void)
1918 {
1919         ldlm_unregister_intent();
1920         class_unregister_type(LUSTRE_MDS_NAME);
1921         class_unregister_type(LUSTRE_MDT_NAME);
1922 }
1923
1924 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1925 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1926 MODULE_LICENSE("GPL");
1927
1928 module_init(mds_init);
1929 module_exit(mds_exit);