lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *   Author: Mike Shaver <shaver@clusterfs.com>
  12  *
  13  *   This file is part of Lustre, http://www.lustre.org.
  14  *
  15  *   Lustre is free software; you can redistribute it and/or
  16  *   modify it under the terms of version 2 of the GNU General Public
  17  *   License as published by the Free Software Foundation.
  18  *
  19  *   Lustre is distributed in the hope that it will be useful,
  20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   GNU General Public License for more details.
  23  *
  24  *   You should have received a copy of the GNU General Public License
  25  *   along with Lustre; if not, write to the Free Software
  26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/lustre_mds.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/init.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/random.h>
  40 #include <linux/fs.h>
  41 #include <linux/jbd.h>
  42 #include <linux/ext3_fs.h>
  43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  44 # include <linux/smp_lock.h>
  45 # include <linux/buffer_head.h>
  46 # include <linux/workqueue.h>
  47 # include <linux/mount.h>
  48 #else
  49 # include <linux/locks.h>
  50 #endif
  51 #include <linux/obd_lov.h>
  52 #include <linux/lustre_mds.h>
  53 #include <linux/lustre_fsfilt.h>
  54 #include <linux/lprocfs_status.h>
  55 #include <linux/lustre_commit_confd.h>
  56
  57 #include "mds_internal.h"
  58
  59 static int mds_cleanup(struct obd_device *obd, int flags);
  60
  61 static int mds_bulk_timeout(void *data)
  62 {
  63         struct ptlrpc_bulk_desc *desc = data;
  64         struct obd_export *exp = desc->bd_export;
  65
  66         DEBUG_REQ(D_ERROR, desc->bd_req,"bulk send timed out: evicting %s@%s\n",
  67                   exp->exp_client_uuid.uuid,
  68                   exp->exp_connection->c_remote_uuid.uuid);
  69         ptlrpc_fail_export(exp);
  70         ptlrpc_abort_bulk (desc);
  71         RETURN(1);
  72 }
  73
  74 /* Assumes caller has already pushed into the kernel filesystem context */
  75 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  76                         loff_t offset, int count)
  77 {
  78         struct ptlrpc_bulk_desc *desc;
  79         struct l_wait_info lwi;
  80         struct page **pages;
  81         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  82         ENTRY;
  83
  84         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  85
  86         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  87         OBD_ALLOC(pages, sizeof(*pages) * npages);
  88         if (!pages)
  89                 GOTO(out, rc = -ENOMEM);
  90
  91         desc = ptlrpc_prep_bulk_exp (req, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
  92         if (desc == NULL)
  93                 GOTO(out_free, rc = -ENOMEM);
  94
  95         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  96                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  97
  98                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  99                 if (pages[i] == NULL)
 100                         GOTO(cleanup_buf, rc = -ENOMEM);
 101
 102                 rc = ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 103                 if (rc != 0)
 104                         GOTO(cleanup_buf, rc);
 105         }
 106
 107         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 108                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 109                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 110                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 111                        file->f_dentry->d_inode->i_size);
 112
 113                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 114                                      page_address(pages[i]), tmpsize, &offset);
 115
 116                 if (rc != tmpsize)
 117                         GOTO(cleanup_buf, rc = -EIO);
 118         }
 119
 120         rc = ptlrpc_bulk_put(desc);
 121         if (rc)
 122                 GOTO(cleanup_buf, rc);
 123
 124         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 125                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 126                        OBD_FAIL_MDS_SENDPAGE, rc);
 127                 ptlrpc_abort_bulk(desc);
 128                 GOTO(cleanup_buf, rc);
 129         }
 130
 131         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, mds_bulk_timeout, desc);
 132         rc = l_wait_event(desc->bd_waitq, ptlrpc_bulk_complete (desc), &lwi);
 133         if (rc) {
 134                 LASSERT (rc == -ETIMEDOUT);
 135                 GOTO(cleanup_buf, rc);
 136         }
 137
 138         EXIT;
 139  cleanup_buf:
 140         for (i = 0; i < npages; i++)
 141                 if (pages[i])
 142                         __free_pages(pages[i], 0);
 143
 144         ptlrpc_free_bulk(desc);
 145  out_free:
 146         OBD_FREE(pages, sizeof(*pages) * npages);
 147  out:
 148         return rc;
 149 }
 150
 151 /* only valid locked dentries or errors should be returned */
 152 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 153                                      struct vfsmount **mnt, int lock_mode,
 154                                      struct lustre_handle *lockh,
 155                                      char *name, int namelen)
 156 {
 157         struct mds_obd *mds = &obd->u.mds;
 158         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 159         struct ldlm_res_id res_id = { .name = {0} };
 160         int flags = 0, rc;
 161         ENTRY;
 162
 163         if (IS_ERR(de))
 164                 RETURN(de);
 165
 166         res_id.name[0] = de->d_inode->i_ino;
 167         res_id.name[1] = de->d_inode->i_generation;
 168         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 169                               res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 170                               &flags, ldlm_completion_ast,
 171                               mds_blocking_ast, NULL, lockh);
 172         if (rc != ELDLM_OK) {
 173                 l_dput(de);
 174                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 175         }
 176
 177         RETURN(retval);
 178 }
 179
 180 #ifndef DCACHE_DISCONNECTED
 181 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 182 #endif
 183
 184
 185 /* Look up an entry by inode number. */
 186 /* this function ONLY returns valid dget'd dentries with an initialized inode
 187    or errors */
 188 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 189                               struct vfsmount **mnt)
 190 {
 191         char fid_name[32];
 192         unsigned long ino = fid->id;
 193         __u32 generation = fid->generation;
 194         struct inode *inode;
 195         struct dentry *result;
 196
 197         if (ino == 0)
 198                 RETURN(ERR_PTR(-ESTALE));
 199
 200         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 201
 202         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 203                ino, generation, mds->mds_sb);
 204
 205         /* under ext3 this is neither supposed to return bad inodes
 206            nor NULL inodes. */
 207         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 208         if (IS_ERR(result))
 209                 RETURN(result);
 210
 211         inode = result->d_inode;
 212         if (!inode)
 213                 RETURN(ERR_PTR(-ENOENT));
 214
 215         if (generation && inode->i_generation != generation) {
 216                 /* we didn't find the right inode.. */
 217                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 218                        inode->i_ino, (unsigned long)inode->i_nlink,
 219                        atomic_read(&inode->i_count), inode->i_generation,
 220                        generation);
 221                 dput(result);
 222                 RETURN(ERR_PTR(-ENOENT));
 223         }
 224
 225         if (mnt) {
 226                 *mnt = mds->mds_vfsmnt;
 227                 mntget(*mnt);
 228         }
 229
 230         RETURN(result);
 231 }
 232
 233
 234 /* Establish a connection to the MDS.
 235  *
 236  * This will set up an export structure for the client to hold state data
 237  * about that client, like open files, the last operation number it did
 238  * on the server, etc.
 239  */
 240 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 241                        struct obd_uuid *cluuid)
 242 {
 243         struct obd_export *exp;
 244         struct mds_export_data *med; /*  */
 245         struct mds_client_data *mcd;
 246         int rc, abort_recovery;
 247         ENTRY;
 248
 249         if (!conn || !obd || !cluuid)
 250                 RETURN(-EINVAL);
 251
 252         /* Check for aborted recovery. */
 253         spin_lock_bh(&obd->obd_processing_task_lock);
 254         abort_recovery = obd->obd_abort_recovery;
 255         spin_unlock_bh(&obd->obd_processing_task_lock);
 256         if (abort_recovery)
 257                 target_abort_recovery(obd);
 258
 259         /* XXX There is a small race between checking the list and adding a
 260          * new connection for the same UUID, but the real threat (list
 261          * corruption when multiple different clients connect) is solved.
 262          *
 263          * There is a second race between adding the export to the list,
 264          * and filling in the client data below.  Hence skipping the case
 265          * of NULL mcd above.  We should already be controlling multiple
 266          * connects at the client, and we can't hold the spinlock over
 267          * memory allocations without risk of deadlocking.
 268          */
 269         rc = class_connect(conn, obd, cluuid);
 270         if (rc)
 271                 RETURN(rc);
 272         exp = class_conn2export(conn);
 273         LASSERT(exp);
 274         med = &exp->exp_mds_data;
 275
 276         OBD_ALLOC(mcd, sizeof(*mcd));
 277         if (!mcd) {
 278                 CERROR("mds: out of memory for client data\n");
 279                 GOTO(out, rc = -ENOMEM);
 280         }
 281
 282         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 283         med->med_mcd = mcd;
 284
 285         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 286         if (rc == 0)
 287                 EXIT;
 288 out:
 289         if (rc) {
 290                 OBD_FREE(mcd, sizeof(*mcd));
 291                 class_disconnect(exp, 0);
 292         }
 293         class_export_put(exp);
 294
 295         return rc;
 296 }
 297
 298 static int mds_init_export(struct obd_export *exp)
 299 {
 300         struct mds_export_data *med = &exp->exp_mds_data;
 301
 302         INIT_LIST_HEAD(&med->med_open_head);
 303         spin_lock_init(&med->med_open_lock);
 304         RETURN(0);
 305 }
 306
 307 static int mds_destroy_export(struct obd_export *export)
 308 {
 309         struct mds_export_data *med;
 310         struct obd_device *obd = export->exp_obd;
 311         struct obd_run_ctxt saved;
 312         int rc = 0;
 313         ENTRY;
 314
 315         med = &export->exp_mds_data;
 316         target_destroy_export(export);
 317
 318         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 319         /* Close any open files (which may also cause orphan unlinking). */
 320         spin_lock(&med->med_open_lock);
 321         while (!list_empty(&med->med_open_head)) {
 322                 struct list_head *tmp = med->med_open_head.next;
 323                 struct mds_file_data *mfd =
 324                         list_entry(tmp, struct mds_file_data, mfd_list);
 325                 BDEVNAME_DECLARE_STORAGE(btmp);
 326
 327                 /* bug 1579: fix force-closing for 2.5 */
 328                 struct dentry *dentry = mfd->mfd_dentry;
 329
 330                 list_del(&mfd->mfd_list);
 331                 spin_unlock(&med->med_open_lock);
 332
 333                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 334                        dentry->d_name.len, dentry->d_name.name,
 335                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 336                        dentry->d_inode->i_ino);
 337                 rc = mds_mfd_close(NULL, obd, mfd,
 338                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 339
 340                 if (rc)
 341                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 342                 spin_lock(&med->med_open_lock);
 343         }
 344         spin_unlock(&med->med_open_lock);
 345         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 346
 347         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 348
 349         RETURN(rc);
 350 }
 351
 352 static int mds_disconnect(struct obd_export *export, int flags)
 353 {
 354         unsigned long irqflags;
 355         int rc;
 356         ENTRY;
 357
 358         ldlm_cancel_locks_for_export(export);
 359
 360         spin_lock_irqsave(&export->exp_lock, irqflags);
 361         export->exp_flags = flags;
 362         spin_unlock_irqrestore(&export->exp_lock, irqflags);
 363
 364         rc = class_disconnect(export, flags);
 365         RETURN(rc);
 366 }
 367
 368 static int mds_getstatus(struct ptlrpc_request *req)
 369 {
 370         struct mds_obd *mds = mds_req2mds(req);
 371         struct mds_body *body;
 372         int rc, size = sizeof(*body);
 373         ENTRY;
 374
 375         rc = lustre_pack_reply(req, 1, &size, NULL);
 376         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 377                 CERROR("mds: out of memory for message: size=%d\n", size);
 378                 req->rq_status = -ENOMEM;       /* superfluous? */
 379                 RETURN(-ENOMEM);
 380         }
 381
 382         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 383         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 384
 385         /* the last_committed and last_xid fields are filled in for all
 386          * replies already - no need to do so here also.
 387          */
 388         RETURN(0);
 389 }
 390
 391 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 392                      void *data, int flag)
 393 {
 394         int do_ast;
 395         ENTRY;
 396
 397         if (flag == LDLM_CB_CANCELING) {
 398                 /* Don't need to do anything here. */
 399                 RETURN(0);
 400         }
 401
 402         /* XXX layering violation!  -phil */
 403         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 404         /* Get this: if mds_blocking_ast is racing with ldlm_intent_policy,
 405          * such that mds_blocking_ast is called just before l_i_p takes the
 406          * ns_lock, then by the time we get the lock, we might not be the
 407          * correct blocking function anymore.  So check, and return early, if
 408          * so. */
 409         if (lock->l_blocking_ast != mds_blocking_ast) {
 410                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 411                 RETURN(0);
 412         }
 413
 414         lock->l_flags |= LDLM_FL_CBPENDING;
 415         do_ast = (!lock->l_readers && !lock->l_writers);
 416         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 417
 418         if (do_ast) {
 419                 struct lustre_handle lockh;
 420                 int rc;
 421
 422                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 423                 ldlm_lock2handle(lock, &lockh);
 424                 rc = ldlm_cli_cancel(&lockh);
 425                 if (rc < 0)
 426                         CERROR("ldlm_cli_cancel: %d\n", rc);
 427         } else {
 428                 LDLM_DEBUG(lock, "Lock still has references, will be "
 429                            "cancelled later");
 430         }
 431         RETURN(0);
 432 }
 433
 434 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 435  * Call with lock=0 if the caller has already taken the i_sem. */
 436 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 437                 struct mds_body *body, struct inode *inode, int lock)
 438 {
 439         struct mds_obd *mds = &obd->u.mds;
 440         void *lmm;
 441         int lmm_size;
 442         int rc;
 443         ENTRY;
 444
 445         lmm = lustre_msg_buf(msg, offset, 0);
 446         if (lmm == NULL) {
 447                 /* Some problem with getting eadata when I sized the reply
 448                  * buffer... */
 449                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 450                        inode->i_ino);
 451                 RETURN(0);
 452         }
 453         lmm_size = msg->buflens[offset];
 454
 455         /* I don't really like this, but it is a sanity check on the client
 456          * MD request.  However, if the client doesn't know how much space
 457          * to reserve for the MD, it shouldn't be bad to have too much space.
 458          */
 459         if (lmm_size > mds->mds_max_mdsize) {
 460                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 461                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 462                 // RETURN(-EINVAL);
 463         }
 464
 465         if (lock)
 466                 down(&inode->i_sem);
 467         rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
 468         if (lock)
 469                 up(&inode->i_sem);
 470         if (rc < 0) {
 471                 CERROR("Error %d reading eadata for ino %lu\n",
 472                        rc, inode->i_ino);
 473         } else if (rc > 0) {
 474                 lmm_size = rc;
 475                 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
 476
 477                 if (rc > 0)
 478                         lmm_size = rc;
 479                 body->valid |= OBD_MD_FLEASIZE;
 480                 body->eadatasize = lmm_size;
 481                 rc = 0;
 482         }
 483
 484         RETURN(rc);
 485 }
 486
 487 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 488                                 struct ptlrpc_request *req,
 489                                 struct mds_body *reqbody, int reply_off)
 490 {
 491         struct mds_body *body;
 492         struct inode *inode = dentry->d_inode;
 493         int rc = 0;
 494         ENTRY;
 495
 496         if (inode == NULL)
 497                 RETURN(-ENOENT);
 498
 499         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 500         LASSERT(body != NULL);                 /* caller prepped reply */
 501
 502         mds_pack_inode2fid(&body->fid1, inode);
 503         mds_pack_inode2body(body, inode);
 504
 505         if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
 506                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 507                                  inode, 1);
 508
 509                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 510                 if (!(body->valid & OBD_MD_FLEASIZE))
 511                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 512                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 513         } else if (S_ISLNK(inode->i_mode) &&
 514                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 515                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 516                 int len;
 517
 518                 LASSERT (symname != NULL);       /* caller prepped reply */
 519                 len = req->rq_repmsg->buflens[reply_off + 1];
 520
 521                 rc = inode->i_op->readlink(dentry, symname, len);
 522                 if (rc < 0) {
 523                         CERROR("readlink failed: %d\n", rc);
 524                 } else if (rc != len - 1) {
 525                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 526                                 rc, len - 1);
 527                         rc = -EINVAL;
 528                 } else {
 529                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 530                         body->valid |= OBD_MD_LINKNAME;
 531                         body->eadatasize = rc + 1;
 532                         symname[rc] = 0;        /* NULL terminate */
 533                         rc = 0;
 534                 }
 535         }
 536
 537         RETURN(rc);
 538 }
 539
 540 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 541                                 int offset)
 542 {
 543         struct mds_obd *mds = mds_req2mds(req);
 544         struct mds_body *body;
 545         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 546         ENTRY;
 547
 548         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 549         LASSERT(body != NULL);                 /* checked by caller */
 550         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 551
 552         if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
 553                 int rc;
 554                 down(&inode->i_sem);
 555                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 556                 up(&inode->i_sem);
 557                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 558                        rc, inode->i_ino);
 559                 if (rc < 0) {
 560                         if (rc != -ENODATA)
 561                                 CERROR("error getting inode %lu MD: rc = %d\n",
 562                                        inode->i_ino, rc);
 563                         size[bufcount] = 0;
 564                 } else if (rc > mds->mds_max_mdsize) {
 565                         size[bufcount] = 0;
 566                         CERROR("MD size %d larger than maximum possible %u\n",
 567                                rc, mds->mds_max_mdsize);
 568                 } else {
 569                         size[bufcount] = rc;
 570                 }
 571                 bufcount++;
 572         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 573                 if (inode->i_size + 1 != body->eadatasize)
 574                         CERROR("symlink size: %Lu, reply space: %d\n",
 575                                inode->i_size + 1, body->eadatasize);
 576                 size[bufcount] = MIN(inode->i_size + 1, body->eadatasize);
 577                 bufcount++;
 578                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 579                        inode->i_size + 1, body->eadatasize);
 580         }
 581
 582         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 583                 CERROR("failed MDS_GETATTR_PACK test\n");
 584                 req->rq_status = -ENOMEM;
 585                 GOTO(out, rc = -ENOMEM);
 586         }
 587
 588         rc = lustre_pack_reply(req, bufcount, size, NULL);
 589         if (rc) {
 590                 CERROR("out of memory\n");
 591                 GOTO(out, req->rq_status = rc);
 592         }
 593
 594         EXIT;
 595  out:
 596         return(rc);
 597 }
 598
 599 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
 600                             struct lustre_handle *child_lockh)
 601 {
 602         struct obd_device *obd = req->rq_export->exp_obd;
 603         struct ldlm_reply *rep = NULL;
 604         struct obd_run_ctxt saved;
 605         struct mds_body *body;
 606         struct dentry *dparent = NULL, *dchild = NULL;
 607         struct obd_ucred uc;
 608         struct lustre_handle parent_lockh;
 609         int namesize;
 610         int rc = 0, cleanup_phase = 0;
 611         char *name;
 612         ENTRY;
 613
 614         LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
 615
 616         /* Swab now, before anyone looks inside the request */
 617
 618         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 619                                   lustre_swab_mds_body);
 620         if (body == NULL) {
 621                 CERROR("Can't swab mds_body\n");
 622                 GOTO(cleanup, rc = -EFAULT);
 623         }
 624
 625         LASSERT_REQSWAB(req, offset + 1);
 626         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 627         if (name == NULL) {
 628                 CERROR("Can't unpack name\n");
 629                 GOTO(cleanup, rc = -EFAULT);
 630         }
 631         namesize = req->rq_reqmsg->buflens[offset + 1];
 632
 633         LASSERT (offset == 0 || offset == 2);
 634         /* if requests were at offset 2, the getattr reply goes back at 1 */
 635         if (offset) {
 636                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 637                 offset = 1;
 638         }
 639
 640         uc.ouc_fsuid = body->fsuid;
 641         uc.ouc_fsgid = body->fsgid;
 642         uc.ouc_cap = body->capability;
 643         uc.ouc_suppgid1 = body->suppgid;
 644         uc.ouc_suppgid2 = -1;
 645         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 646         cleanup_phase = 1; /* kernel context */
 647         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 648
 649         /* FIXME: handle raw lookup */
 650 #if 0
 651         if (body->valid == OBD_MD_FLID) {
 652                 struct mds_body *mds_reply;
 653                 int size = sizeof(*mds_reply);
 654                 ino_t inum;
 655                 // The user requested ONLY the inode number, so do a raw lookup
 656                 rc = lustre_pack_reply(req, 1, &size, NULL);
 657                 if (rc) {
 658                         CERROR("out of memory\n");
 659                         GOTO(cleanup, rc);
 660                 }
 661
 662                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 663
 664                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 665                                            sizeof(*mds_reply));
 666                 mds_reply->fid1.id = inum;
 667                 mds_reply->valid = OBD_MD_FLID;
 668                 GOTO(cleanup, rc);
 669         }
 670 #endif
 671
 672         rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
 673                                          &parent_lockh, &dparent, LCK_PR,
 674                                          name, namesize, child_lockh,
 675                                          &dchild, LCK_PR);
 676         if (rc)
 677                 GOTO(cleanup, rc);
 678
 679         cleanup_phase = 2; /* dchild, dparent, locks */
 680
 681         if (dchild->d_inode == NULL) {
 682                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 683                 /* in the intent case, the policy clears this error:
 684                    the disposition is enough */
 685                 GOTO(cleanup, rc = -ENOENT);
 686         } else {
 687                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 688         }
 689
 690         if (req->rq_repmsg == NULL) {
 691                 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 692                 if (rc != 0) {
 693                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 694                         GOTO (cleanup, rc);
 695                 }
 696         }
 697
 698         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 699         GOTO(cleanup, rc); /* returns the lock to the client */
 700
 701  cleanup:
 702         switch (cleanup_phase) {
 703         case 2:
 704                 if (rc && dchild->d_inode)
 705                         ldlm_lock_decref(child_lockh, LCK_PR);
 706                 ldlm_lock_decref(&parent_lockh, LCK_PR);
 707                 l_dput(dchild);
 708                 l_dput(dparent);
 709         case 1:
 710                 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 711         default: ;
 712         }
 713         return rc;
 714 }
 715
 716 static int mds_getattr(int offset, struct ptlrpc_request *req)
 717 {
 718         struct mds_obd *mds = mds_req2mds(req);
 719         struct obd_device *obd = req->rq_export->exp_obd;
 720         struct obd_run_ctxt saved;
 721         struct dentry *de;
 722         struct mds_body *body;
 723         struct obd_ucred uc;
 724         int rc = 0;
 725         ENTRY;
 726
 727         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
 728                                    lustre_swab_mds_body);
 729         if (body == NULL) {
 730                 CERROR ("Can't unpack body\n");
 731                 RETURN (-EFAULT);
 732         }
 733
 734         uc.ouc_fsuid = body->fsuid;
 735         uc.ouc_fsgid = body->fsgid;
 736         uc.ouc_cap = body->capability;
 737         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 738         de = mds_fid2dentry(mds, &body->fid1, NULL);
 739         if (IS_ERR(de)) {
 740                 rc = req->rq_status = -ENOENT;
 741                 GOTO(out_pop, PTR_ERR(de));
 742         }
 743
 744         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
 745         if (rc != 0) {
 746                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
 747                 GOTO (out_pop, rc);
 748         }
 749
 750         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
 751
 752         l_dput(de);
 753         GOTO(out_pop, rc);
 754 out_pop:
 755         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 756         return rc;
 757 }
 758
 759
 760 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 761                           unsigned long max_age)
 762 {
 763         return fsfilt_statfs(obd, obd->u.mds.mds_sb, osfs);
 764 }
 765
 766 static int mds_statfs(struct ptlrpc_request *req)
 767 {
 768         struct obd_device *obd = req->rq_export->exp_obd;
 769         int rc, size = sizeof(struct obd_statfs);
 770         ENTRY;
 771
 772         rc = lustre_pack_reply(req, 1, &size, NULL);
 773         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 774                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
 775                 GOTO(out, rc);
 776         }
 777
 778         /* We call this so that we can cache a bit - 1 jiffie worth */
 779         rc = obd_statfs(obd, lustre_msg_buf(req->rq_repmsg,0,size),jiffies-HZ);
 780         if (rc) {
 781                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
 782                 GOTO(out, rc);
 783         }
 784
 785         EXIT;
 786 out:
 787         req->rq_status = rc;
 788         return 0;
 789 }
 790
 791 static int mds_sync(struct ptlrpc_request *req)
 792 {
 793         struct obd_device *obd = req->rq_export->exp_obd;
 794         struct mds_obd *mds = &obd->u.mds;
 795         struct mds_body *body;
 796         int rc, size = sizeof(*body);
 797         ENTRY;
 798
 799         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 800         if (body == NULL)
 801                 GOTO(out, rc = -EPROTO);
 802
 803         rc = lustre_pack_reply(req, 1, &size, NULL);
 804         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
 805                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
 806                 GOTO(out, rc);
 807         }
 808
 809         if (body->fid1.id == 0) {
 810                 /* a fid of zero is taken to mean "sync whole filesystem" */
 811                 rc = fsfilt_sync(obd, mds->mds_sb);
 812                 if (rc)
 813                         GOTO(out, rc);
 814         } else {
 815                 /* just any file to grab fsync method - "file" arg unused */
 816                 struct file *file = mds->mds_rcvd_filp;
 817                 struct dentry *de;
 818
 819                 de = mds_fid2dentry(mds, &body->fid1, NULL);
 820                 if (IS_ERR(de))
 821                         GOTO(out, rc = PTR_ERR(de));
 822
 823                 rc = file->f_op->fsync(NULL, de, 1);
 824                 l_dput(de);
 825                 if (rc)
 826                         GOTO(out, rc);
 827
 828                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 829                 mds_pack_inode2fid(&body->fid1, de->d_inode);
 830                 mds_pack_inode2body(body, de->d_inode);
 831         }
 832 out:
 833         req->rq_status = rc;
 834         return 0;
 835 }
 836
 837 static int mds_readpage(struct ptlrpc_request *req)
 838 {
 839         struct obd_device *obd = req->rq_export->exp_obd;
 840         struct vfsmount *mnt;
 841         struct dentry *de;
 842         struct file *file;
 843         struct mds_body *body, *repbody;
 844         struct lustre_handle lockh;
 845         struct obd_run_ctxt saved;
 846         int rc, size = sizeof(*repbody);
 847         struct obd_ucred uc;
 848         ENTRY;
 849
 850         rc = lustre_pack_reply(req, 1, &size, NULL);
 851         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 852                 CERROR("mds: out of memory\n");
 853                 GOTO(out, rc = -ENOMEM);
 854         }
 855
 856         body = lustre_swab_reqbuf (req, 0, sizeof (*body),
 857                                    lustre_swab_mds_body);
 858         if (body == NULL)
 859                 GOTO (out, rc = -EFAULT);
 860
 861         uc.ouc_fsuid = body->fsuid;
 862         uc.ouc_fsgid = body->fsgid;
 863         uc.ouc_cap = body->capability;
 864         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 865         de = mds_fid2locked_dentry(obd, &body->fid1, &mnt, LCK_PR,
 866                                    &lockh, NULL, 0);
 867         if (IS_ERR(de))
 868                 GOTO(out_pop, rc = PTR_ERR(de));
 869
 870         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
 871
 872         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 873         /* note: in case of an error, dentry_open puts dentry */
 874         if (IS_ERR(file))
 875                 GOTO(out_lock, rc = PTR_ERR(file));
 876
 877         /* body->size is actually the offset -eeb */
 878         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
 879                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
 880                        body->size, de->d_inode->i_blksize);
 881                 GOTO(out_file, rc = -EFAULT);
 882         }
 883
 884         /* body->nlink is actually the #bytes to read -eeb */
 885         if (body->nlink & (de->d_inode->i_blksize - 1)) {
 886                 CERROR("size %u is not multiple of blocksize %lu\n",
 887                        body->nlink, de->d_inode->i_blksize);
 888                 GOTO(out_file, rc = -EFAULT);
 889         }
 890
 891         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
 892         repbody->size = file->f_dentry->d_inode->i_size;
 893         repbody->valid = OBD_MD_FLSIZE;
 894
 895         /* to make this asynchronous make sure that the handling function
 896            doesn't send a reply when this function completes. Instead a
 897            callback function would send the reply */
 898         /* body->size is actually the offset -eeb */
 899         rc = mds_sendpage(req, file, body->size, body->nlink);
 900
 901 out_file:
 902         filp_close(file, 0);
 903 out_lock:
 904         ldlm_lock_decref(&lockh, LCK_PR);
 905 out_pop:
 906         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 907 out:
 908         req->rq_status = rc;
 909         RETURN(0);
 910 }
 911
 912 int mds_reint(struct ptlrpc_request *req, int offset,
 913               struct lustre_handle *lockh)
 914 {
 915         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
 916         int rc;
 917
 918         OBD_ALLOC(rec, sizeof(*rec));
 919         if (rec == NULL)
 920                 RETURN(-ENOMEM);
 921
 922         rc = mds_update_unpack(req, offset, rec);
 923         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 924                 CERROR("invalid record\n");
 925                 GOTO(out, req->rq_status = -EINVAL);
 926         }
 927         /* rc will be used to interrupt a for loop over multiple records */
 928         rc = mds_reint_rec(rec, offset, req, lockh);
 929  out:
 930         OBD_FREE(rec, sizeof(*rec));
 931         return rc;
 932 }
 933
 934 static int mds_filter_recovery_request(struct ptlrpc_request *req,
 935                                        struct obd_device *obd, int *process)
 936 {
 937         switch (req->rq_reqmsg->opc) {
 938         case MDS_CONNECT: /* This will never get here, but for completeness. */
 939         case OST_CONNECT: /* This will never get here, but for completeness. */
 940         case MDS_DISCONNECT:
 941         case OST_DISCONNECT:
 942                *process = 1;
 943                RETURN(0);
 944
 945         case MDS_CLOSE:
 946         case MDS_SYNC: /* used in unmounting */
 947         case OBD_PING:
 948         case MDS_REINT:
 949         case LDLM_ENQUEUE:
 950                 *process = target_queue_recovery_request(req, obd);
 951                 RETURN(0);
 952
 953         default:
 954                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
 955                 *process = 0;
 956                 /* XXX what should we set rq_status to here? */
 957                 req->rq_status = -EAGAIN;
 958                 RETURN(ptlrpc_error(req));
 959         }
 960 }
 961
 962 static char *reint_names[] = {
 963         [REINT_SETATTR] "setattr",
 964         [REINT_CREATE]  "create",
 965         [REINT_LINK]    "link",
 966         [REINT_UNLINK]  "unlink",
 967         [REINT_RENAME]  "rename",
 968         [REINT_OPEN]    "open",
 969 };
 970
 971 int mds_handle(struct ptlrpc_request *req)
 972 {
 973         int should_process;
 974         int rc = 0;
 975         struct mds_obd *mds = NULL; /* quell gcc overwarning */
 976         struct obd_device *obd = NULL;
 977         ENTRY;
 978
 979         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
 980
 981         LASSERT(current->journal_info == NULL);
 982         /* XXX identical to OST */
 983         if (req->rq_reqmsg->opc != MDS_CONNECT) {
 984                 struct mds_export_data *med;
 985                 int recovering, abort_recovery;
 986
 987                 if (req->rq_export == NULL) {
 988                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
 989                                req->rq_reqmsg->opc);
 990                         req->rq_status = -ENOTCONN;
 991                         GOTO(out, rc = -ENOTCONN);
 992                 }
 993
 994                 med = &req->rq_export->exp_mds_data;
 995                 obd = req->rq_export->exp_obd;
 996                 mds = &obd->u.mds;
 997
 998                 /* Check for aborted recovery. */
 999                 spin_lock_bh(&obd->obd_processing_task_lock);
1000                 abort_recovery = obd->obd_abort_recovery;
1001                 recovering = obd->obd_recovering;
1002                 spin_unlock_bh(&obd->obd_processing_task_lock);
1003                 if (abort_recovery) {
1004                         target_abort_recovery(obd);
1005                 } else if (recovering) {
1006                         rc = mds_filter_recovery_request(req, obd,
1007                                                          &should_process);
1008                         if (rc || !should_process)
1009                                 RETURN(rc);
1010                 }
1011         }
1012
1013         switch (req->rq_reqmsg->opc) {
1014         case MDS_CONNECT:
1015                 DEBUG_REQ(D_INODE, req, "connect");
1016                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1017                 rc = target_handle_connect(req, mds_handle);
1018                 if (!rc)
1019                         /* Now that we have an export, set mds. */
1020                         mds = mds_req2mds(req);
1021                 break;
1022
1023         case MDS_DISCONNECT:
1024                 DEBUG_REQ(D_INODE, req, "disconnect");
1025                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1026                 rc = target_handle_disconnect(req);
1027                 req->rq_status = rc;            /* superfluous? */
1028                 break;
1029
1030         case MDS_GETSTATUS:
1031                 DEBUG_REQ(D_INODE, req, "getstatus");
1032                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1033                 rc = mds_getstatus(req);
1034                 break;
1035
1036         case MDS_GETATTR:
1037                 DEBUG_REQ(D_INODE, req, "getattr");
1038                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1039                 rc = mds_getattr(0, req);
1040                 break;
1041
1042         case MDS_GETATTR_NAME: {
1043                 struct lustre_handle lockh;
1044                 DEBUG_REQ(D_INODE, req, "getattr_name");
1045                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1046
1047                 /* If this request gets a reconstructed reply, we won't be
1048                  * acquiring any new locks in mds_getattr_name, so we don't
1049                  * want to cancel.
1050                  */
1051                 lockh.cookie = 0;
1052                 rc = mds_getattr_name(0, req, &lockh);
1053                 /* this non-intent call (from an ioctl) is special */
1054                 req->rq_status = rc;
1055                 if (rc == 0 && lockh.cookie)
1056                         ldlm_lock_decref(&lockh, LCK_PR);
1057                 break;
1058         }
1059         case MDS_STATFS:
1060                 DEBUG_REQ(D_INODE, req, "statfs");
1061                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1062                 rc = mds_statfs(req);
1063                 break;
1064
1065         case MDS_READPAGE:
1066                 DEBUG_REQ(D_INODE, req, "readpage");
1067                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1068                 rc = mds_readpage(req);
1069
1070                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1071                         return 0;
1072                 break;
1073
1074         case MDS_REINT: {
1075                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1076                 __u32  opc;
1077                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1078                                mds->mds_max_cookiesize};
1079                 int bufcount;
1080
1081                 /* NB only peek inside req now; mds_reint() will swab it */
1082                 if (opcp == NULL) {
1083                         CERROR ("Can't inspect opcode\n");
1084                         rc = -EINVAL;
1085                         break;
1086                 }
1087                 opc = *opcp;
1088                 if (lustre_msg_swabbed (req->rq_reqmsg))
1089                         __swab32s(&opc);
1090
1091                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1092                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1093                            reint_names[opc] == NULL) ? reint_names[opc] :
1094                                                        "unknown opcode");
1095
1096                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1097
1098                 if (opc == REINT_UNLINK)
1099                         bufcount = 3;
1100                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1101                         bufcount = 2;
1102                 else
1103                         bufcount = 1;
1104
1105                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1106                 if (rc)
1107                         break;
1108
1109                 rc = mds_reint(req, 0, NULL);
1110                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1111                 break;
1112         }
1113
1114         case MDS_CLOSE:
1115                 DEBUG_REQ(D_INODE, req, "close");
1116                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1117                 rc = mds_close(req);
1118                 break;
1119
1120         case MDS_DONE_WRITING:
1121                 DEBUG_REQ(D_INODE, req, "done_writing");
1122                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1123                 rc = mds_done_writing(req);
1124                 break;
1125
1126         case MDS_PIN:
1127                 DEBUG_REQ(D_INODE, req, "pin");
1128                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1129                 rc = mds_pin(req);
1130                 break;
1131
1132         case MDS_SYNC:
1133                 DEBUG_REQ(D_INODE, req, "sync");
1134                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1135                 rc = mds_sync(req);
1136                 break;
1137
1138         case OBD_PING:
1139                 DEBUG_REQ(D_INODE, req, "ping");
1140                 rc = target_handle_ping(req);
1141                 break;
1142
1143         case OBD_LOG_CANCEL:
1144                 CDEBUG(D_INODE, "log cancel\n");
1145                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1146                 rc = -ENOTSUPP; /* la la la */
1147                 break;
1148
1149         case LDLM_ENQUEUE:
1150                 DEBUG_REQ(D_INODE, req, "enqueue");
1151                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1152                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1153                                          ldlm_server_blocking_ast);
1154                 break;
1155         case LDLM_CONVERT:
1156                 DEBUG_REQ(D_INODE, req, "convert");
1157                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1158                 rc = ldlm_handle_convert(req);
1159                 break;
1160         case LDLM_BL_CALLBACK:
1161         case LDLM_CP_CALLBACK:
1162                 DEBUG_REQ(D_INODE, req, "callback");
1163                 CERROR("callbacks should not happen on MDS\n");
1164                 LBUG();
1165                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1166                 break;
1167         case LLOG_ORIGIN_HANDLE_CREATE:
1168                 DEBUG_REQ(D_INODE, req, "llog_init");
1169                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1170                 rc = llog_origin_handle_create(req);
1171                 break;
1172         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1173                 DEBUG_REQ(D_INODE, req, "llog next block");
1174                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1175                 rc = llog_origin_handle_next_block(req);
1176                 break;
1177         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1178                 DEBUG_REQ(D_INODE, req, "llog read header");
1179                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1180                 rc = llog_origin_handle_read_header(req);
1181                 break;
1182         case LLOG_ORIGIN_HANDLE_CLOSE:
1183                 DEBUG_REQ(D_INODE, req, "llog close");
1184                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1185                 rc = llog_origin_handle_close(req);
1186                 break;
1187         default:
1188                 req->rq_status = -ENOTSUPP;
1189                 rc = ptlrpc_error(req);
1190                 RETURN(rc);
1191         }
1192
1193         LASSERT(current->journal_info == NULL);
1194
1195         EXIT;
1196
1197         /* If we're DISCONNECTing, the mds_export_data is already freed */
1198         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1199                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1200                 struct obd_device *obd = list_entry(mds, struct obd_device,
1201                                                     u.mds);
1202                 req->rq_repmsg->last_xid =
1203                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1204
1205                 if (!obd->obd_no_transno) {
1206                         req->rq_repmsg->last_committed =
1207                                 obd->obd_last_committed;
1208                 } else {
1209                         DEBUG_REQ(D_IOCTL, req,
1210                                   "not sending last_committed update");
1211                 }
1212                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1213                        ", xid "LPU64"\n",
1214                        mds->mds_last_transno, obd->obd_last_committed,
1215                        req->rq_xid);
1216         }
1217  out:
1218
1219         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1220                 if (obd && obd->obd_recovering) {
1221                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1222                         return target_queue_final_reply(req, rc);
1223                 }
1224                 /* Lost a race with recovery; let the error path DTRT. */
1225                 rc = req->rq_status = -ENOTCONN;
1226         }
1227
1228         target_send_reply(req, rc, OBD_FAIL_MDS_ALL_REPLY_NET);
1229         return 0;
1230 }
1231
1232 /* Update the server data on disk.  This stores the new mount_count and
1233  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1234  * then the server last_rcvd value may be less than that of the clients.
1235  * This will alert us that we may need to do client recovery.
1236  *
1237  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1238  */
1239 int mds_update_server_data(struct obd_device *obd, int force_sync)
1240 {
1241         struct mds_obd *mds = &obd->u.mds;
1242         struct mds_server_data *msd = mds->mds_server_data;
1243         struct file *filp = mds->mds_rcvd_filp;
1244         struct obd_run_ctxt saved;
1245         loff_t off = 0;
1246         int rc;
1247         ENTRY;
1248
1249         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1250         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1251         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1252
1253         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1254                mds->mds_mount_count, mds->mds_last_transno);
1255         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1256         if (rc)
1257                 CERROR("error writing MDS server data: rc = %d\n", rc);
1258         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1259         RETURN(rc);
1260 }
1261
1262
1263 /* mount the file system (secretly) */
1264 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1265 {
1266         struct lustre_cfg* lcfg = buf;
1267         struct mds_obd *mds = &obd->u.mds;
1268         struct vfsmount *mnt;
1269         int rc = 0;
1270         unsigned long page;
1271         ENTRY;
1272
1273         dev_clear_rdonly(2);
1274
1275         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1276                 RETURN(rc = -EINVAL);
1277
1278         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1279         if (IS_ERR(obd->obd_fsops))
1280                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1281
1282         if (!(page = __get_free_page(GFP_KERNEL)))
1283                 RETURN(-ENOMEM);
1284
1285         memset((void *)page, 0, PAGE_SIZE);
1286         sprintf((char *)page, "iopen_nopriv");
1287
1288         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1289                             lcfg->lcfg_inlbuf1, (void *)page);
1290         free_page(page);
1291         if (IS_ERR(mnt)) {
1292                 rc = PTR_ERR(mnt);
1293                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1294                 GOTO(err_ops, rc);
1295         }
1296
1297         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1298
1299         sema_init(&mds->mds_orphan_recovery_sem, 1);
1300         sema_init(&mds->mds_epoch_sem, 1);
1301         spin_lock_init(&mds->mds_transno_lock);
1302         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1303         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1304         atomic_set(&mds->mds_open_count, 0);
1305
1306         obd->obd_namespace = ldlm_namespace_new("mds_server",
1307                                                 LDLM_NAMESPACE_SERVER);
1308         if (obd->obd_namespace == NULL) {
1309                 mds_cleanup(obd, 0);
1310                 GOTO(err_put, rc = -ENOMEM);
1311         }
1312
1313         rc = mds_fs_setup(obd, mnt);
1314         if (rc) {
1315                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1316                 GOTO(err_ns, rc);
1317         }
1318
1319         rc = llog_start_commit_thread();
1320         if (rc < 0)
1321                 GOTO(err_fs, rc);
1322
1323
1324         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1325                 class_uuid_t uuid;
1326
1327                 generate_random_uuid(uuid);
1328                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1329
1330                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1331                 if (mds->mds_profile == NULL)
1332                         GOTO(err_fs, rc = -ENOMEM);
1333
1334                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1335                        lcfg->lcfg_inllen3);
1336
1337         }
1338
1339         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1340                            "mds_ldlm_client", &obd->obd_ldlm_client);
1341         obd->obd_replayable = 1;
1342
1343         RETURN(0);
1344
1345 err_fs:
1346         /* No extra cleanup needed for llog_init_commit_thread() */
1347         mds_fs_cleanup(obd, 0);
1348 err_ns:
1349         ldlm_namespace_free(obd->obd_namespace, 0);
1350         obd->obd_namespace = NULL;
1351 err_put:
1352         unlock_kernel();
1353         mntput(mds->mds_vfsmnt);
1354         mds->mds_sb = 0;
1355         lock_kernel();
1356 err_ops:
1357         fsfilt_put_ops(obd->obd_fsops);
1358         return rc;
1359 }
1360
1361 static int mds_postsetup(struct obd_device *obd)
1362 {
1363         struct mds_obd *mds = &obd->u.mds;
1364         int rc = 0;
1365         ENTRY;
1366
1367
1368         rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1369                         &llog_lvfs_ops);
1370         if (rc)
1371                 RETURN(rc);
1372
1373         if (mds->mds_profile) {
1374                 struct obd_run_ctxt saved;
1375                 struct lustre_profile *lprof;
1376                 struct config_llog_instance cfg;
1377
1378                 cfg.cfg_instance = NULL;
1379                 cfg.cfg_uuid = mds->mds_lov_uuid;
1380                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1381                 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1382                                              mds->mds_profile, &cfg);
1383                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1384                 if (rc)
1385                         GOTO(err_llog, rc);
1386
1387                 lprof = class_get_profile(mds->mds_profile);
1388                 if (lprof == NULL) {
1389                         CERROR("No profile found: %s\n", mds->mds_profile);
1390                         GOTO(err_cleanup, rc = -ENOENT);
1391                 }
1392                 rc = mds_lov_connect(obd, lprof->lp_osc);
1393                 if (rc)
1394                         GOTO(err_cleanup, rc);
1395         }
1396
1397         RETURN(rc);
1398
1399 err_cleanup:
1400         mds_lov_clean(obd);
1401 err_llog:
1402         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1403         RETURN(rc);
1404 }
1405
1406 static int mds_postrecov(struct obd_device *obd)
1407
1408 {
1409         int rc, rc2;
1410
1411         LASSERT(!obd->obd_recovering);
1412
1413 #ifdef ENABLE_ORPHANS
1414         rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1415                           obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1416         if (rc != 0) {
1417                 CERROR("faild at llog_origin_connect: %d\n", rc);
1418         }
1419 #endif
1420         rc = mds_cleanup_orphans(obd);
1421
1422         rc2 = mds_lov_set_nextid(obd);
1423         if (rc2 == 0)
1424                 rc2 = rc;
1425         RETURN(rc2);
1426 }
1427
1428 int mds_lov_clean(struct obd_device *obd)
1429 {
1430         struct mds_obd *mds = &obd->u.mds;
1431
1432         if (mds->mds_profile) {
1433                 char * cln_prof;
1434                 struct config_llog_instance cfg;
1435                 struct obd_run_ctxt saved;
1436                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1437
1438                 OBD_ALLOC(cln_prof, len);
1439                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1440
1441                 cfg.cfg_instance = NULL;
1442                 cfg.cfg_uuid = mds->mds_lov_uuid;
1443
1444                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1445                 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1446                                         cln_prof, &cfg);
1447                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1448
1449                 OBD_FREE(cln_prof, len);
1450                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1451                 mds->mds_profile = NULL;
1452         }
1453         RETURN(0);
1454 }
1455
1456 static int mds_precleanup(struct obd_device *obd, int flags)
1457 {
1458         int rc = 0;
1459         ENTRY;
1460
1461         mds_lov_disconnect(obd, flags);
1462         mds_lov_clean(obd);
1463         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1464         RETURN(rc);
1465 }
1466
1467 static int mds_cleanup(struct obd_device *obd, int flags)
1468 {
1469         struct mds_obd *mds = &obd->u.mds;
1470         ENTRY;
1471
1472         if (mds->mds_sb == NULL)
1473                 RETURN(0);
1474
1475         mds_update_server_data(obd, 1);
1476         if (mds->mds_lov_objids != NULL) {
1477                 OBD_FREE(mds->mds_lov_objids,
1478                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1479         }
1480         mds_fs_cleanup(obd, flags);
1481
1482         unlock_kernel();
1483
1484         /* 2 seems normal on mds, (may_umount() also expects 2
1485           fwiw), but we only see 1 at this point in obdfilter. */
1486         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1487                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1488                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1489
1490         mntput(mds->mds_vfsmnt);
1491
1492         mds->mds_sb = 0;
1493
1494         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1495
1496         if (obd->obd_recovering)
1497                 target_cancel_recovery_timer(obd);
1498         lock_kernel();
1499         dev_clear_rdonly(2);
1500         fsfilt_put_ops(obd->obd_fsops);
1501
1502         RETURN(0);
1503 }
1504
1505 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1506                                         struct ldlm_lock *new_lock,
1507                                         struct lustre_handle *lockh)
1508 {
1509         struct obd_export *exp = req->rq_export;
1510         struct obd_device *obd = exp->exp_obd;
1511         struct ldlm_request *dlmreq =
1512                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1513         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1514         struct list_head *iter;
1515
1516         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1517                 return;
1518
1519         l_lock(&obd->obd_namespace->ns_lock);
1520         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1521                 struct ldlm_lock *lock;
1522                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1523                 if (lock == new_lock)
1524                         continue;
1525                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1526                         lockh->cookie = lock->l_handle.h_cookie;
1527                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1528                                   lockh->cookie);
1529                         l_unlock(&obd->obd_namespace->ns_lock);
1530                         return;
1531                 }
1532         }
1533         l_unlock(&obd->obd_namespace->ns_lock);
1534
1535         /* This remote handle isn't enqueued, so we never received or
1536          * processed this request.  Clear MSG_RESENT, because it can
1537          * be handled like any normal request now. */
1538
1539         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1540
1541         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1542                   remote_hdl.cookie);
1543 }
1544
1545 int intent_disposition(struct ldlm_reply *rep, int flag)
1546 {
1547         if (!rep)
1548                 return 0;
1549         return (rep->lock_policy_res1 & flag);
1550 }
1551
1552 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1553 {
1554         if (!rep)
1555                 return;
1556         rep->lock_policy_res1 |= flag;
1557 }
1558
1559 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1560                               struct ldlm_lock **lockp, void *req_cookie,
1561                               ldlm_mode_t mode, int flags, void *data)
1562 {
1563         struct ptlrpc_request *req = req_cookie;
1564         struct ldlm_lock *lock = *lockp;
1565         int rc;
1566         ENTRY;
1567
1568         if (!req_cookie)
1569                 RETURN(0);
1570
1571         if (req->rq_reqmsg->bufcount > 1) {
1572                 /* an intent needs to be considered */
1573                 struct ldlm_intent *it;
1574                 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1575                 struct ldlm_reply *rep;
1576                 struct lustre_handle lockh = { 0 };
1577                 struct ldlm_lock *new_lock;
1578                 int offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1579                                               sizeof(struct mds_body),
1580                                               mds->mds_max_mdsize,
1581                                               mds->mds_max_cookiesize};
1582
1583                 it = lustre_swab_reqbuf(req, 1, sizeof (*it),
1584                                         lustre_swab_ldlm_intent);
1585                 if (it == NULL) {
1586                         CERROR ("Intent missing\n");
1587                         req->rq_status = -EFAULT;
1588                         RETURN(req->rq_status);
1589                 }
1590
1591                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1592                            ldlm_it2str(it->opc));
1593
1594                 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3,
1595                                        repsize, NULL);
1596                 if (rc)
1597                         RETURN(req->rq_status = rc);
1598
1599                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1600                 intent_set_disposition(rep, DISP_IT_EXECD);
1601
1602                 fixup_handle_for_resent_req(req, lock, &lockh);
1603
1604                 /* execute policy */
1605                 switch ((long)it->opc) {
1606                 case IT_OPEN:
1607                 case IT_CREAT|IT_OPEN:
1608                         /* XXX swab here to assert that an mds_open reint
1609                          * packet is following */
1610                         rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1611 #if 0
1612                         /* We abort the lock if the lookup was negative and
1613                          * we did not make it to the OPEN portion */
1614                         if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1615                                 RETURN(ELDLM_LOCK_ABORTED);
1616                         if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1617                             !intent_disposition(rep, DISP_OPEN_OPEN))
1618 #endif
1619                                 RETURN(ELDLM_LOCK_ABORTED);
1620                         break;
1621                 case IT_GETATTR:
1622                 case IT_LOOKUP:
1623                 case IT_READDIR:
1624                         rep->lock_policy_res2 = mds_getattr_name(offset, req,
1625                                                                  &lockh);
1626                         /* FIXME: LDLM can set req->rq_status. MDS sets
1627                            policy_res{1,2} with disposition and status.
1628                            - replay: returns 0 & req->status is old status
1629                            - otherwise: returns req->status */
1630                         if (intent_disposition(rep, DISP_LOOKUP_NEG))
1631                                 rep->lock_policy_res2 = 0;
1632                         if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1633                             rep->lock_policy_res2)
1634                                 RETURN(ELDLM_LOCK_ABORTED);
1635                         if (req->rq_status != 0) {
1636                                 LBUG();
1637                                 rep->lock_policy_res2 = req->rq_status;
1638                                 RETURN(ELDLM_LOCK_ABORTED);
1639                         }
1640                         break;
1641                 default:
1642                         CERROR("Unhandled intent "LPD64"\n", it->opc);
1643                         LBUG();
1644                 }
1645
1646                 /* By this point, whatever function we called above must have
1647                  * either filled in 'lockh', been an intent replay, or returned
1648                  * an error.  We want to allow replayed RPCs to not get a lock,
1649                  * since we would just drop it below anyways because lock replay
1650                  * is done separately by the client afterwards.  For regular
1651                  * RPCs we want to give the new lock to the client instead of
1652                  * whatever lock it was about to get.
1653                  */
1654                 new_lock = ldlm_handle2lock(&lockh);
1655                 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1656                         RETURN(0);
1657
1658                 LASSERT(new_lock != NULL);
1659
1660                 /* If we've already given this lock to a client once, then we
1661                  * should have no readers or writers.  Otherwise, we should
1662                  * have one reader _or_ writer ref (which will be zeroed below)
1663                  * before returning the lock to a client.
1664                  */
1665                 if (new_lock->l_export == req->rq_export) {
1666                         LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1667                 } else {
1668                         LASSERT(new_lock->l_export == NULL);
1669                         LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1670                 }
1671
1672                 *lockp = new_lock;
1673
1674                 if (new_lock->l_export == req->rq_export) {
1675                         /* Already gave this to the client, which means that we
1676                          * reconstructed a reply. */
1677                         LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1678                                 MSG_RESENT);
1679                         RETURN(ELDLM_LOCK_REPLACED);
1680                 }
1681
1682                 /* Fixup the lock to be given to the client */
1683                 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1684                 new_lock->l_readers = 0;
1685                 new_lock->l_writers = 0;
1686
1687                 new_lock->l_export = class_export_get(req->rq_export);
1688                 list_add(&new_lock->l_export_chain,
1689                          &new_lock->l_export->exp_ldlm_data.led_held_locks);
1690
1691                 new_lock->l_blocking_ast = lock->l_blocking_ast;
1692                 new_lock->l_completion_ast = lock->l_completion_ast;
1693
1694                 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1695                        sizeof(lock->l_remote_handle));
1696
1697                 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1698
1699                 LDLM_LOCK_PUT(new_lock);
1700                 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1701
1702                 RETURN(ELDLM_LOCK_REPLACED);
1703         } else {
1704                 int size = sizeof(struct ldlm_reply);
1705                 rc = lustre_pack_reply(req, 1, &size, NULL);
1706                 if (rc) {
1707                         LBUG();
1708                         RETURN(-ENOMEM);
1709                 }
1710         }
1711         RETURN(0);
1712 }
1713
1714 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1715 {
1716         struct lprocfs_static_vars lvars;
1717
1718         lprocfs_init_multi_vars(0, &lvars);
1719         return lprocfs_obd_attach(dev, lvars.obd_vars);
1720 }
1721
1722 int mds_detach(struct obd_device *dev)
1723 {
1724         return lprocfs_obd_detach(dev);
1725 }
1726
1727 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1728 {
1729         struct lprocfs_static_vars lvars;
1730
1731         lprocfs_init_multi_vars(1, &lvars);
1732         return lprocfs_obd_attach(dev, lvars.obd_vars);
1733 }
1734
1735 int mdt_detach(struct obd_device *dev)
1736 {
1737         return lprocfs_obd_detach(dev);
1738 }
1739
1740 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1741 {
1742         struct mds_obd *mds = &obddev->u.mds;
1743         int rc = 0;
1744         ENTRY;
1745
1746         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1747                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1748                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1749                                            mds_handle, "mds",
1750                                            obddev->obd_proc_entry);
1751
1752         if (!mds->mds_service) {
1753                 CERROR("failed to start service\n");
1754                 RETURN(rc = -ENOMEM);
1755         }
1756
1757         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1758                                     "ll_mdt");
1759         if (rc)
1760                 GOTO(err_thread, rc);
1761
1762         mds->mds_setattr_service =
1763                 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1764                                 MDS_BUFSIZE, MDS_MAXREQSIZE,
1765                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1766                                 mds_handle, "mds_setattr",
1767                                 obddev->obd_proc_entry);
1768         if (!mds->mds_setattr_service) {
1769                 CERROR("failed to start getattr service\n");
1770                 GOTO(err_thread, rc = -ENOMEM);
1771         }
1772
1773         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1774                                  MDT_NUM_THREADS, "ll_mdt_attr");
1775         if (rc)
1776                 GOTO(err_thread2, rc);
1777
1778         mds->mds_readpage_service =
1779                 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1780                                 MDS_BUFSIZE, MDS_MAXREQSIZE,
1781                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1782                                 mds_handle, "mds_readpage",
1783                                 obddev->obd_proc_entry);
1784         if (!mds->mds_readpage_service) {
1785                 CERROR("failed to start readpage service\n");
1786                 GOTO(err_thread2, rc = -ENOMEM);
1787         }
1788
1789         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1790                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
1791
1792         if (rc)
1793                 GOTO(err_thread3, rc);
1794
1795         RETURN(0);
1796
1797 err_thread3:
1798         ptlrpc_unregister_service(mds->mds_readpage_service);
1799 err_thread2:
1800         ptlrpc_unregister_service(mds->mds_setattr_service);
1801 err_thread:
1802         ptlrpc_unregister_service(mds->mds_service);
1803         return rc;
1804 }
1805
1806
1807 static int mdt_cleanup(struct obd_device *obddev, int flags)
1808 {
1809         struct mds_obd *mds = &obddev->u.mds;
1810         ENTRY;
1811
1812         ptlrpc_stop_all_threads(mds->mds_readpage_service);
1813         ptlrpc_unregister_service(mds->mds_readpage_service);
1814
1815         ptlrpc_stop_all_threads(mds->mds_setattr_service);
1816         ptlrpc_unregister_service(mds->mds_setattr_service);
1817
1818         ptlrpc_stop_all_threads(mds->mds_service);
1819         ptlrpc_unregister_service(mds->mds_service);
1820
1821         RETURN(0);
1822 }
1823
1824 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1825 {
1826         struct obd_device *obd = data;
1827         struct ll_fid fid;
1828         fid.id = id;
1829         fid.generation = gen;
1830         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1831 }
1832
1833 struct lvfs_callback_ops mds_lvfs_ops = {
1834         l_fid2dentry:     mds_lvfs_fid2dentry,
1835 };
1836
1837 /* use obd ops to offer management infrastructure */
1838 static struct obd_ops mds_obd_ops = {
1839         o_owner:       THIS_MODULE,
1840         o_attach:      mds_attach,
1841         o_detach:      mds_detach,
1842         o_connect:     mds_connect,
1843         o_init_export:  mds_init_export,
1844         o_destroy_export:  mds_destroy_export,
1845         o_disconnect:  mds_disconnect,
1846         o_setup:       mds_setup,
1847         o_postsetup:   mds_postsetup,
1848         o_precleanup:  mds_precleanup,
1849         o_cleanup:     mds_cleanup,
1850         o_postrecov:   mds_postrecov,
1851         o_statfs:      mds_obd_statfs,
1852         o_iocontrol:   mds_iocontrol,
1853         o_create:      mds_obd_create,
1854         o_destroy:     mds_obd_destroy,
1855         o_llog_init:   mds_llog_init,
1856         o_llog_finish: mds_llog_finish,
1857         o_notify:      mds_notify,
1858 };
1859
1860 static struct obd_ops mdt_obd_ops = {
1861         o_owner:       THIS_MODULE,
1862         o_attach:      mdt_attach,
1863         o_detach:      mdt_detach,
1864         o_setup:       mdt_setup,
1865         o_cleanup:     mdt_cleanup,
1866 };
1867
1868 static int __init mds_init(void)
1869 {
1870         struct lprocfs_static_vars lvars;
1871
1872         lprocfs_init_multi_vars(0, &lvars);
1873         class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1874         lprocfs_init_multi_vars(1, &lvars);
1875         class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1876         ldlm_register_intent(ldlm_intent_policy);
1877
1878         return 0;
1879 }
1880
1881 static void /*__exit*/ mds_exit(void)
1882 {
1883         ldlm_unregister_intent();
1884         class_unregister_type(LUSTRE_MDS_NAME);
1885         class_unregister_type(LUSTRE_MDT_NAME);
1886 }
1887
1888 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1889 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1890 MODULE_LICENSE("GPL");
1891
1892 module_init(mds_init);
1893 module_exit(mds_exit);