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_intent_policy(struct ldlm_namespace *ns,
  60                              struct ldlm_lock **lockp, void *req_cookie,
  61                              ldlm_mode_t mode, int flags, void *data);
  62 static int mds_postsetup(struct obd_device *obd);
  63 static int mds_cleanup(struct obd_device *obd, int flags);
  64
  65 /* Assumes caller has already pushed into the kernel filesystem context */
  66 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  67                         loff_t offset, int count)
  68 {
  69         struct ptlrpc_bulk_desc *desc;
  70         struct l_wait_info lwi;
  71         struct page **pages;
  72         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  73         ENTRY;
  74
  75         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  76
  77         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  78         OBD_ALLOC(pages, sizeof(*pages) * npages);
  79         if (!pages)
  80                 GOTO(out, rc = -ENOMEM);
  81
  82         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  83                                     MDS_BULK_PORTAL);
  84         if (desc == NULL)
  85                 GOTO(out_free, rc = -ENOMEM);
  86
  87         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  88                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  89
  90                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  91                 if (pages[i] == NULL)
  92                         GOTO(cleanup_buf, rc = -ENOMEM);
  93
  94                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  95         }
  96
  97         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  98                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  99                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 100                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 101                        file->f_dentry->d_inode->i_size);
 102
 103                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 104                                      page_address(pages[i]), tmpsize, &offset);
 105
 106                 if (rc != tmpsize)
 107                         GOTO(cleanup_buf, rc = -EIO);
 108         }
 109
 110         LASSERT(desc->bd_nob == count);
 111
 112         rc = ptlrpc_start_bulk_transfer(desc);
 113         if (rc)
 114                 GOTO(cleanup_buf, rc);
 115
 116         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 117                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 118                        OBD_FAIL_MDS_SENDPAGE, rc);
 119                 GOTO(abort_bulk, rc);
 120         }
 121
 122         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 123         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 124         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 125
 126         if (rc == 0) {
 127                 if (desc->bd_success &&
 128                     desc->bd_nob_transferred == count)
 129                         GOTO(cleanup_buf, rc);
 130
 131                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 132         }
 133
 134         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 135                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 136                   desc->bd_nob_transferred, count,
 137                   req->rq_export->exp_client_uuid.uuid,
 138                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 139
 140         ptlrpc_fail_export(req->rq_export);
 141
 142         EXIT;
 143  abort_bulk:
 144         ptlrpc_abort_bulk (desc);
 145  cleanup_buf:
 146         for (i = 0; i < npages; i++)
 147                 if (pages[i])
 148                         __free_pages(pages[i], 0);
 149
 150         ptlrpc_free_bulk(desc);
 151  out_free:
 152         OBD_FREE(pages, sizeof(*pages) * npages);
 153  out:
 154         return rc;
 155 }
 156
 157 /* only valid locked dentries or errors should be returned */
 158 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 159                                      struct vfsmount **mnt, int lock_mode,
 160                                      struct lustre_handle *lockh,
 161                                      char *name, int namelen, __u64 lockpart)
 162 {
 163         struct mds_obd *mds = &obd->u.mds;
 164         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 165         struct ldlm_res_id res_id = { .name = {0} };
 166         int flags = 0, rc;
 167         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 168
 169         ENTRY;
 170
 171         if (IS_ERR(de))
 172                 RETURN(de);
 173
 174         res_id.name[0] = de->d_inode->i_ino;
 175         res_id.name[1] = de->d_inode->i_generation;
 176 #ifdef S_PDIROPS
 177         lockh[1].cookie = 0;
 178         if (name && IS_PDIROPS(de->d_inode)) {
 179                 ldlm_policy_data_t cpolicy =
 180                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 181                 /* lock just dir { ino, generation } to flush client cache */
 182                 if (lock_mode == LCK_PW) {
 183                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 184                                               res_id, LDLM_IBITS,
 185                                               &cpolicy, LCK_CW, &flags,
 186                                               mds_blocking_ast,
 187                                               ldlm_completion_ast, NULL, NULL,
 188                                               NULL, 0, NULL, lockh + 1);
 189                         if (rc != ELDLM_OK) {
 190                                 l_dput(de);
 191                                 RETURN(ERR_PTR(-ENOLCK));
 192                         }
 193                        flags = 0;
 194                 }
 195
 196                 res_id.name[2] = full_name_hash(name, namelen);
 197                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
 198                        de->d_inode->i_ino, de->d_inode->i_generation,
 199                        res_id.name[2]);
 200         }
 201 #endif
 202         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 203                               LDLM_IBITS, &policy, lock_mode, &flags,
 204                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 205                               NULL, 0, NULL, lockh);
 206         if (rc != ELDLM_OK) {
 207                 l_dput(de);
 208                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 209 #ifdef S_PDIROPS
 210                 if (lockh[1].cookie)
 211                         ldlm_lock_decref(lockh + 1, LCK_CW);
 212 #endif
 213         }
 214
 215         RETURN(retval);
 216 }
 217
 218 #ifndef DCACHE_DISCONNECTED
 219 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 220 #endif
 221
 222
 223 /* Look up an entry by inode number. */
 224 /* this function ONLY returns valid dget'd dentries with an initialized inode
 225    or errors */
 226 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 227                               struct vfsmount **mnt)
 228 {
 229         char fid_name[32];
 230         unsigned long ino = fid->id;
 231         __u32 generation = fid->generation;
 232         struct inode *inode;
 233         struct dentry *result;
 234
 235         if (ino == 0)
 236                 RETURN(ERR_PTR(-ESTALE));
 237
 238         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 239
 240         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 241                ino, generation, mds->mds_sb);
 242
 243         /* under ext3 this is neither supposed to return bad inodes
 244            nor NULL inodes. */
 245         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 246         if (IS_ERR(result))
 247                 RETURN(result);
 248
 249         inode = result->d_inode;
 250         if (!inode)
 251                 RETURN(ERR_PTR(-ENOENT));
 252
 253         if (generation && inode->i_generation != generation) {
 254                 /* we didn't find the right inode.. */
 255                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 256                        inode->i_ino, (unsigned long)inode->i_nlink,
 257                        atomic_read(&inode->i_count), inode->i_generation,
 258                        generation);
 259                 dput(result);
 260                 RETURN(ERR_PTR(-ENOENT));
 261         }
 262
 263         if (mnt) {
 264                 *mnt = mds->mds_vfsmnt;
 265                 mntget(*mnt);
 266         }
 267
 268         RETURN(result);
 269 }
 270
 271
 272 /* Establish a connection to the MDS.
 273  *
 274  * This will set up an export structure for the client to hold state data
 275  * about that client, like open files, the last operation number it did
 276  * on the server, etc.
 277  */
 278 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 279                        struct obd_uuid *cluuid)
 280 {
 281         struct obd_export *exp;
 282         struct mds_export_data *med; /*  */
 283         struct mds_client_data *mcd;
 284         int rc, abort_recovery;
 285         ENTRY;
 286
 287         if (!conn || !obd || !cluuid)
 288                 RETURN(-EINVAL);
 289
 290         /* Check for aborted recovery. */
 291         spin_lock_bh(&obd->obd_processing_task_lock);
 292         abort_recovery = obd->obd_abort_recovery;
 293         spin_unlock_bh(&obd->obd_processing_task_lock);
 294         if (abort_recovery)
 295                 target_abort_recovery(obd);
 296
 297         /* XXX There is a small race between checking the list and adding a
 298          * new connection for the same UUID, but the real threat (list
 299          * corruption when multiple different clients connect) is solved.
 300          *
 301          * There is a second race between adding the export to the list,
 302          * and filling in the client data below.  Hence skipping the case
 303          * of NULL mcd above.  We should already be controlling multiple
 304          * connects at the client, and we can't hold the spinlock over
 305          * memory allocations without risk of deadlocking.
 306          */
 307         rc = class_connect(conn, obd, cluuid);
 308         if (rc)
 309                 RETURN(rc);
 310         exp = class_conn2export(conn);
 311         LASSERT(exp);
 312         med = &exp->exp_mds_data;
 313
 314         OBD_ALLOC(mcd, sizeof(*mcd));
 315         if (!mcd) {
 316                 CERROR("mds: out of memory for client data\n");
 317                 GOTO(out, rc = -ENOMEM);
 318         }
 319
 320         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 321         med->med_mcd = mcd;
 322
 323         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 324         if (rc == 0)
 325                 EXIT;
 326 out:
 327         if (rc) {
 328                 OBD_FREE(mcd, sizeof(*mcd));
 329                 class_disconnect(exp, 0);
 330         }
 331         class_export_put(exp);
 332
 333         return rc;
 334 }
 335
 336 static int mds_init_export(struct obd_export *exp)
 337 {
 338         struct mds_export_data *med = &exp->exp_mds_data;
 339
 340         INIT_LIST_HEAD(&med->med_open_head);
 341         spin_lock_init(&med->med_open_lock);
 342         RETURN(0);
 343 }
 344
 345 static int mds_destroy_export(struct obd_export *export)
 346 {
 347         struct mds_export_data *med;
 348         struct obd_device *obd = export->exp_obd;
 349         struct obd_run_ctxt saved;
 350         int rc = 0;
 351         ENTRY;
 352
 353         med = &export->exp_mds_data;
 354         target_destroy_export(export);
 355
 356         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 357                 GOTO(out, 0);
 358
 359         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 360         /* Close any open files (which may also cause orphan unlinking). */
 361         spin_lock(&med->med_open_lock);
 362         while (!list_empty(&med->med_open_head)) {
 363                 struct list_head *tmp = med->med_open_head.next;
 364                 struct mds_file_data *mfd =
 365                         list_entry(tmp, struct mds_file_data, mfd_list);
 366                 BDEVNAME_DECLARE_STORAGE(btmp);
 367
 368                 /* bug 1579: fix force-closing for 2.5 */
 369                 struct dentry *dentry = mfd->mfd_dentry;
 370
 371                 list_del(&mfd->mfd_list);
 372                 spin_unlock(&med->med_open_lock);
 373
 374                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 375                        dentry->d_name.len, dentry->d_name.name,
 376                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 377                        dentry->d_inode->i_ino);
 378                 rc = mds_mfd_close(NULL, obd, mfd,
 379                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 380
 381                 if (rc)
 382                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 383                 spin_lock(&med->med_open_lock);
 384         }
 385         spin_unlock(&med->med_open_lock);
 386         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 387
 388 out:
 389         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 390
 391         RETURN(rc);
 392 }
 393
 394 static int mds_disconnect(struct obd_export *exp, int flags)
 395 {
 396         struct obd_device *obd;
 397         struct mds_obd *mds;
 398         unsigned long irqflags;
 399         int rc;
 400         ENTRY;
 401
 402         LASSERT(exp);
 403         class_export_get(exp);
 404
 405         obd = class_exp2obd(exp);
 406         if (obd == NULL) {
 407                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 408                        exp->exp_handle.h_cookie);
 409                 RETURN(-EINVAL);
 410         }
 411         mds = &obd->u.mds;
 412
 413         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 414                         && !atomic_read(&mds->mds_real_clients)) {
 415                 /* there was no client at all */
 416                 mds_lmv_disconnect(obd, flags);
 417         }
 418
 419         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 420                         && atomic_dec_and_test(&mds->mds_real_clients)) {
 421                 /* time to drop LMV connections */
 422                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 423                        "Disconnnect from LMV now\n",
 424                        obd->obd_name, exp->exp_client_uuid.uuid);
 425                 mds_lmv_disconnect(obd, flags);
 426         }
 427
 428         spin_lock_irqsave(&exp->exp_lock, irqflags);
 429         exp->exp_flags = flags;
 430         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 431
 432         /* Disconnect early so that clients can't keep using export */
 433         rc = class_disconnect(exp, flags);
 434         ldlm_cancel_locks_for_export(exp);
 435
 436         /* complete all outstanding replies */
 437         spin_lock_irqsave(&exp->exp_lock, irqflags);
 438         while (!list_empty(&exp->exp_outstanding_replies)) {
 439                 struct ptlrpc_reply_state *rs =
 440                         list_entry(exp->exp_outstanding_replies.next,
 441                                    struct ptlrpc_reply_state, rs_exp_list);
 442                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 443
 444                 spin_lock(&svc->srv_lock);
 445                 list_del_init(&rs->rs_exp_list);
 446                 ptlrpc_schedule_difficult_reply(rs);
 447                 spin_unlock(&svc->srv_lock);
 448         }
 449         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 450
 451         class_export_put(exp);
 452         RETURN(rc);
 453 }
 454
 455 static int mds_getstatus(struct ptlrpc_request *req)
 456 {
 457         struct mds_obd *mds = mds_req2mds(req);
 458         struct mds_body *body;
 459         int rc, size = sizeof(*body);
 460         ENTRY;
 461
 462         rc = lustre_pack_reply(req, 1, &size, NULL);
 463         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 464                 CERROR("mds: out of memory for message: size=%d\n", size);
 465                 req->rq_status = -ENOMEM;       /* superfluous? */
 466                 RETURN(-ENOMEM);
 467         }
 468
 469         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 470         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 471
 472         /* the last_committed and last_xid fields are filled in for all
 473          * replies already - no need to do so here also.
 474          */
 475         RETURN(0);
 476 }
 477
 478 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 479                      void *data, int flag)
 480 {
 481         int do_ast;
 482         ENTRY;
 483
 484         if (flag == LDLM_CB_CANCELING) {
 485                 /* Don't need to do anything here. */
 486                 RETURN(0);
 487         }
 488
 489         /* XXX layering violation!  -phil */
 490         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 491         /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
 492          * such that mds_blocking_ast is called just before l_i_p takes the
 493          * ns_lock, then by the time we get the lock, we might not be the
 494          * correct blocking function anymore.  So check, and return early, if
 495          * so. */
 496         if (lock->l_blocking_ast != mds_blocking_ast) {
 497                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 498                 RETURN(0);
 499         }
 500
 501         lock->l_flags |= LDLM_FL_CBPENDING;
 502         do_ast = (!lock->l_readers && !lock->l_writers);
 503         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 504
 505         if (do_ast) {
 506                 struct lustre_handle lockh;
 507                 int rc;
 508
 509                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 510                 ldlm_lock2handle(lock, &lockh);
 511                 rc = ldlm_cli_cancel(&lockh);
 512                 if (rc < 0)
 513                         CERROR("ldlm_cli_cancel: %d\n", rc);
 514         } else {
 515                 LDLM_DEBUG(lock, "Lock still has references, will be "
 516                            "cancelled later");
 517         }
 518         RETURN(0);
 519 }
 520
 521 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 522  * Call with lock=0 if the caller has already taken the i_sem. */
 523 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 524                 struct mds_body *body, struct inode *inode, int lock)
 525 {
 526         struct mds_obd *mds = &obd->u.mds;
 527         void *lmm;
 528         int lmm_size;
 529         int rc;
 530         ENTRY;
 531
 532         lmm = lustre_msg_buf(msg, offset, 0);
 533         if (lmm == NULL) {
 534                 /* Some problem with getting eadata when I sized the reply
 535                  * buffer... */
 536                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 537                        inode->i_ino);
 538                 RETURN(0);
 539         }
 540         lmm_size = msg->buflens[offset];
 541
 542         /* I don't really like this, but it is a sanity check on the client
 543          * MD request.  However, if the client doesn't know how much space
 544          * to reserve for the MD, it shouldn't be bad to have too much space.
 545          */
 546         if (lmm_size > mds->mds_max_mdsize) {
 547                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 548                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 549                 // RETURN(-EINVAL);
 550         }
 551
 552         if (lock)
 553                 down(&inode->i_sem);
 554         rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
 555         if (lock)
 556                 up(&inode->i_sem);
 557         if (rc < 0) {
 558                 CERROR("Error %d reading eadata for ino %lu\n",
 559                        rc, inode->i_ino);
 560         } else if (rc > 0) {
 561                 lmm_size = rc;
 562                 if (S_ISREG(inode->i_mode))
 563                         rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
 564                 if (rc > 0)
 565                         lmm_size = rc;
 566                 body->valid |= OBD_MD_FLEASIZE;
 567                 body->eadatasize = lmm_size;
 568                 rc = 0;
 569         }
 570
 571         RETURN(rc);
 572 }
 573
 574 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 575                                 struct ptlrpc_request *req,
 576                                 struct mds_body *reqbody, int reply_off)
 577 {
 578         struct mds_body *body;
 579         struct inode *inode = dentry->d_inode;
 580         int rc = 0;
 581         ENTRY;
 582
 583         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
 584                 RETURN(-ENOENT);
 585
 586         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 587         LASSERT(body != NULL);                 /* caller prepped reply */
 588
 589         if (dentry->d_flags & DCACHE_CROSS_REF) {
 590                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 591                        (unsigned long) dentry->d_mdsnum,
 592                        (unsigned long) dentry->d_inum,
 593                        (unsigned long) dentry->d_generation);
 594                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 595                 body->fid1.id = dentry->d_inum;
 596                 body->fid1.mds = dentry->d_mdsnum;
 597                 body->fid1.generation = dentry->d_generation;
 598                 RETURN(0);
 599         }
 600         mds_pack_inode2fid(obd, &body->fid1, inode);
 601         mds_pack_inode2body(obd, body, inode);
 602
 603         if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
 604                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 605                                  inode, 1);
 606
 607                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 608                 if (!(body->valid & OBD_MD_FLEASIZE))
 609                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 610                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 611         } else if (S_ISLNK(inode->i_mode) &&
 612                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 613                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 614                 int len;
 615
 616                 LASSERT (symname != NULL);       /* caller prepped reply */
 617                 len = req->rq_repmsg->buflens[reply_off + 1];
 618
 619                 rc = inode->i_op->readlink(dentry, symname, len);
 620                 if (rc < 0) {
 621                         CERROR("readlink failed: %d\n", rc);
 622                 } else if (rc != len - 1) {
 623                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 624                                 rc, len - 1);
 625                         rc = -EINVAL;
 626                 } else {
 627                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 628                         body->valid |= OBD_MD_LINKNAME;
 629                         body->eadatasize = rc + 1;
 630                         symname[rc] = 0;        /* NULL terminate */
 631                         rc = 0;
 632                 }
 633         } else if (S_ISDIR(inode->i_mode) &&
 634                         (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
 635                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 636                                 inode, 1);
 637         }
 638
 639         RETURN(rc);
 640 }
 641
 642 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
 643                                         struct dentry *dentry,
 644                                         int offset)
 645 {
 646         int rc = 0, size[1] = {sizeof(struct mds_body)};
 647         ENTRY;
 648
 649         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 650                 CERROR("failed MDS_GETATTR_PACK test\n");
 651                 req->rq_status = -ENOMEM;
 652                 GOTO(out, rc = -ENOMEM);
 653         }
 654
 655         rc = lustre_pack_reply(req, 1, size, NULL);
 656         if (rc) {
 657                 CERROR("out of memory\n");
 658                 GOTO(out, req->rq_status = rc);
 659         }
 660
 661         EXIT;
 662  out:
 663         return(rc);
 664 }
 665
 666 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 667                                 int offset)
 668 {
 669         struct mds_obd *mds = mds_req2mds(req);
 670         struct mds_body *body;
 671         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 672         ENTRY;
 673
 674         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 675         LASSERT(body != NULL);                 /* checked by caller */
 676         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 677
 678         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode))
 679                         && (body->valid & OBD_MD_FLEASIZE)) {
 680                 int rc;
 681                 down(&inode->i_sem);
 682                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 683                 up(&inode->i_sem);
 684                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 685                        rc, inode->i_ino);
 686                 if (rc < 0) {
 687                         if (rc != -ENODATA)
 688                                 CERROR("error getting inode %lu MD: rc = %d\n",
 689                                        inode->i_ino, rc);
 690                         size[bufcount] = 0;
 691                 } else if (rc > mds->mds_max_mdsize) {
 692                         size[bufcount] = 0;
 693                         CERROR("MD size %d larger than maximum possible %u\n",
 694                                rc, mds->mds_max_mdsize);
 695                 } else {
 696                         size[bufcount] = rc;
 697                 }
 698                 bufcount++;
 699         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 700                 if (inode->i_size + 1 != body->eadatasize)
 701                         CERROR("symlink size: %Lu, reply space: %d\n",
 702                                inode->i_size + 1, body->eadatasize);
 703                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 704                 bufcount++;
 705                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 706                        inode->i_size + 1, body->eadatasize);
 707         }
 708
 709         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 710                 CERROR("failed MDS_GETATTR_PACK test\n");
 711                 req->rq_status = -ENOMEM;
 712                 GOTO(out, rc = -ENOMEM);
 713         }
 714
 715         rc = lustre_pack_reply(req, bufcount, size, NULL);
 716         if (rc) {
 717                 CERROR("out of memory\n");
 718                 GOTO(out, req->rq_status = rc);
 719         }
 720
 721         EXIT;
 722  out:
 723         return(rc);
 724 }
 725
 726 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
 727                             struct lustre_handle *child_lockh, int child_part)
 728 {
 729         struct obd_device *obd = req->rq_export->exp_obd;
 730         struct ldlm_reply *rep = NULL;
 731         struct obd_run_ctxt saved;
 732         struct mds_body *body;
 733         struct dentry *dparent = NULL, *dchild = NULL;
 734         struct obd_ucred uc;
 735         struct lustre_handle parent_lockh[2];
 736         int namesize;
 737         int rc = 0, cleanup_phase = 0, resent_req = 0;
 738         char *name;
 739         ENTRY;
 740
 741         LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
 742
 743         /* Swab now, before anyone looks inside the request */
 744
 745         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 746                                   lustre_swab_mds_body);
 747         if (body == NULL) {
 748                 CERROR("Can't swab mds_body\n");
 749                 GOTO(cleanup, rc = -EFAULT);
 750         }
 751
 752         LASSERT_REQSWAB(req, offset + 1);
 753         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 754         if (name == NULL) {
 755                 CERROR("Can't unpack name\n");
 756                 GOTO(cleanup, rc = -EFAULT);
 757         }
 758         namesize = req->rq_reqmsg->buflens[offset + 1];
 759
 760         LASSERT (offset == 0 || offset == 2);
 761         /* if requests were at offset 2, the getattr reply goes back at 1 */
 762         if (offset) {
 763                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 764                 offset = 1;
 765         }
 766
 767         uc.ouc_fsuid = body->fsuid;
 768         uc.ouc_fsgid = body->fsgid;
 769         uc.ouc_cap = body->capability;
 770         uc.ouc_suppgid1 = body->suppgid;
 771         uc.ouc_suppgid2 = -1;
 772         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 773         cleanup_phase = 1; /* kernel context */
 774         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 775
 776         LASSERT(namesize > 0);
 777         if (namesize == 1) {
 778                 /* we have no dentry here, drop LOOKUP bit */
 779                 child_part &= ~MDS_INODELOCK_LOOKUP;
 780                 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
 781                        obd->obd_name, (unsigned long) body->fid1.id,
 782                        (unsigned long) body->fid1.generation);
 783                 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
 784                                                parent_lockh, NULL, 0, child_part);
 785                 if (IS_ERR(dchild)) {
 786                         CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
 787                         GOTO(cleanup, rc = PTR_ERR(dchild));
 788                 }
 789                 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
 790 #ifdef S_PDIROPS
 791                 if (parent_lockh[1].cookie)
 792                         ldlm_lock_decref(parent_lockh + 1, LCK_CW);
 793 #endif
 794                 cleanup_phase = 2;
 795                 goto fill_inode;
 796         }
 797
 798         /* FIXME: handle raw lookup */
 799 #if 0
 800         if (body->valid == OBD_MD_FLID) {
 801                 struct mds_body *mds_reply;
 802                 int size = sizeof(*mds_reply);
 803                 ino_t inum;
 804                 // The user requested ONLY the inode number, so do a raw lookup
 805                 rc = lustre_pack_reply(req, 1, &size, NULL);
 806                 if (rc) {
 807                         CERROR("out of memory\n");
 808                         GOTO(cleanup, rc);
 809                 }
 810
 811                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 812
 813                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 814                                            sizeof(*mds_reply));
 815                 mds_reply->fid1.id = inum;
 816                 mds_reply->valid = OBD_MD_FLID;
 817                 GOTO(cleanup, rc);
 818         }
 819 #endif
 820
 821         if (child_lockh->cookie != 0) {
 822                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 823                 resent_req = 1;
 824         }
 825
 826         if (resent_req == 0) {
 827                 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
 828                                                  parent_lockh, &dparent,
 829                                                  LCK_PR, MDS_INODELOCK_LOOKUP,
 830                                                  name, namesize,
 831                                                  child_lockh, &dchild, LCK_PR,
 832                                                  child_part);
 833                 if (rc)
 834                         GOTO(cleanup, rc);
 835         } else {
 836                 struct ldlm_lock *granted_lock;
 837                 struct ll_fid child_fid;
 838                 struct ldlm_resource *res;
 839                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 840                 granted_lock = ldlm_handle2lock(child_lockh);
 841                 LASSERT(granted_lock);
 842
 843                 res = granted_lock->l_resource;
 844                 child_fid.id = res->lr_name.name[0];
 845                 child_fid.generation = res->lr_name.name[1];
 846                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 847                 LASSERT(dchild);
 848                 LDLM_LOCK_PUT(granted_lock);
 849         }
 850
 851         cleanup_phase = 2; /* dchild, dparent, locks */
 852
 853 fill_inode:
 854         if (!DENTRY_VALID(dchild)) {
 855                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 856                 /* in the intent case, the policy clears this error:
 857                    the disposition is enough */
 858                 GOTO(cleanup, rc = -ENOENT);
 859         } else {
 860                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 861         }
 862
 863         if (req->rq_repmsg == NULL) {
 864                 if (dchild->d_flags & DCACHE_CROSS_REF)
 865                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
 866                 else
 867                         rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 868                 if (rc != 0) {
 869                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 870                         GOTO (cleanup, rc);
 871                 }
 872         }
 873
 874         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 875         GOTO(cleanup, rc); /* returns the lock to the client */
 876
 877  cleanup:
 878         switch (cleanup_phase) {
 879         case 2:
 880                 if (resent_req == 0) {
 881                         if (rc && DENTRY_VALID(dchild))
 882                                 ldlm_lock_decref(child_lockh, LCK_PR);
 883                         if (dparent) {
 884                                 ldlm_lock_decref(parent_lockh, LCK_PR);
 885 #ifdef S_PDIROPS
 886                                 if (parent_lockh[1].cookie != 0)
 887                                         ldlm_lock_decref(parent_lockh + 1,
 888                                                         LCK_CW);
 889 #endif
 890                         }
 891                         if (dparent)
 892                                 l_dput(dparent);
 893                 }
 894                 l_dput(dchild);
 895         case 1:
 896                 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 897         default: ;
 898         }
 899         return rc;
 900 }
 901
 902 static int mds_getattr(int offset, struct ptlrpc_request *req)
 903 {
 904         struct mds_obd *mds = mds_req2mds(req);
 905         struct obd_device *obd = req->rq_export->exp_obd;
 906         struct obd_run_ctxt saved;
 907         struct dentry *de;
 908         struct mds_body *body;
 909         struct obd_ucred uc;
 910         int rc = 0;
 911         ENTRY;
 912
 913         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
 914                                    lustre_swab_mds_body);
 915         if (body == NULL) {
 916                 CERROR ("Can't unpack body\n");
 917                 RETURN (-EFAULT);
 918         }
 919
 920         uc.ouc_fsuid = body->fsuid;
 921         uc.ouc_fsgid = body->fsgid;
 922         uc.ouc_cap = body->capability;
 923         push_ctxt(&saved, &obd->obd_ctxt, &uc);
 924         de = mds_fid2dentry(mds, &body->fid1, NULL);
 925         if (IS_ERR(de)) {
 926                 rc = req->rq_status = -ENOENT;
 927                 GOTO(out_pop, PTR_ERR(de));
 928         }
 929
 930         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
 931         if (rc != 0) {
 932                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
 933                 GOTO (out_pop, rc);
 934         }
 935
 936         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
 937
 938         l_dput(de);
 939         GOTO(out_pop, rc);
 940 out_pop:
 941         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
 942         return rc;
 943 }
 944
 945
 946 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 947                           unsigned long max_age)
 948 {
 949         int rc;
 950
 951         spin_lock(&obd->obd_osfs_lock);
 952         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
 953         if (rc == 0)
 954                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
 955         spin_unlock(&obd->obd_osfs_lock);
 956
 957         return rc;
 958 }
 959
 960 static int mds_statfs(struct ptlrpc_request *req)
 961 {
 962         struct obd_device *obd = req->rq_export->exp_obd;
 963         int rc, size = sizeof(struct obd_statfs);
 964         ENTRY;
 965
 966         rc = lustre_pack_reply(req, 1, &size, NULL);
 967         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 968                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
 969                 GOTO(out, rc);
 970         }
 971
 972         /* We call this so that we can cache a bit - 1 jiffie worth */
 973         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
 974                             jiffies - HZ);
 975         if (rc) {
 976                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
 977                 GOTO(out, rc);
 978         }
 979
 980         EXIT;
 981 out:
 982         req->rq_status = rc;
 983         return 0;
 984 }
 985
 986 static int mds_sync(struct ptlrpc_request *req)
 987 {
 988         struct obd_device *obd = req->rq_export->exp_obd;
 989         struct mds_obd *mds = &obd->u.mds;
 990         struct mds_body *body;
 991         int rc, size = sizeof(*body);
 992         ENTRY;
 993
 994         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 995         if (body == NULL)
 996                 GOTO(out, rc = -EPROTO);
 997
 998         rc = lustre_pack_reply(req, 1, &size, NULL);
 999         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1000                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1001                 GOTO(out, rc);
1002         }
1003
1004         if (body->fid1.id == 0) {
1005                 /* a fid of zero is taken to mean "sync whole filesystem" */
1006                 rc = fsfilt_sync(obd, mds->mds_sb);
1007                 if (rc)
1008                         GOTO(out, rc);
1009         } else {
1010                 /* just any file to grab fsync method - "file" arg unused */
1011                 struct file *file = mds->mds_rcvd_filp;
1012                 struct dentry *de;
1013
1014                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1015                 if (IS_ERR(de))
1016                         GOTO(out, rc = PTR_ERR(de));
1017
1018                 rc = file->f_op->fsync(NULL, de, 1);
1019                 l_dput(de);
1020                 if (rc)
1021                         GOTO(out, rc);
1022
1023                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1024                 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1025                 mds_pack_inode2body(obd, body, de->d_inode);
1026         }
1027 out:
1028         req->rq_status = rc;
1029         return 0;
1030 }
1031
1032 /* mds_readpage does not take a DLM lock on the inode, because the client must
1033  * already have a PR lock.
1034  *
1035  * If we were to take another one here, a deadlock will result, if another
1036  * thread is already waiting for a PW lock. */
1037 static int mds_readpage(struct ptlrpc_request *req)
1038 {
1039         struct obd_device *obd = req->rq_export->exp_obd;
1040         struct vfsmount *mnt;
1041         struct dentry *de;
1042         struct file *file;
1043         struct mds_body *body, *repbody;
1044         struct obd_run_ctxt saved;
1045         int rc, size = sizeof(*repbody);
1046         struct obd_ucred uc;
1047         ENTRY;
1048
1049         rc = lustre_pack_reply(req, 1, &size, NULL);
1050         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1051                 CERROR("mds: out of memory\n");
1052                 GOTO(out, rc = -ENOMEM);
1053         }
1054
1055         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1056         if (body == NULL)
1057                 GOTO (out, rc = -EFAULT);
1058
1059         uc.ouc_fsuid = body->fsuid;
1060         uc.ouc_fsgid = body->fsgid;
1061         uc.ouc_cap = body->capability;
1062         push_ctxt(&saved, &obd->obd_ctxt, &uc);
1063         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1064         if (IS_ERR(de))
1065                 GOTO(out_pop, rc = PTR_ERR(de));
1066
1067         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1068
1069         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1070         /* note: in case of an error, dentry_open puts dentry */
1071         if (IS_ERR(file))
1072                 GOTO(out_pop, rc = PTR_ERR(file));
1073
1074         /* body->size is actually the offset -eeb */
1075         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1076                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1077                        body->size, de->d_inode->i_blksize);
1078                 GOTO(out_file, rc = -EFAULT);
1079         }
1080
1081         /* body->nlink is actually the #bytes to read -eeb */
1082         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1083                 CERROR("size %u is not multiple of blocksize %lu\n",
1084                        body->nlink, de->d_inode->i_blksize);
1085                 GOTO(out_file, rc = -EFAULT);
1086         }
1087
1088         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1089         repbody->size = file->f_dentry->d_inode->i_size;
1090         repbody->valid = OBD_MD_FLSIZE;
1091
1092         /* to make this asynchronous make sure that the handling function
1093            doesn't send a reply when this function completes. Instead a
1094            callback function would send the reply */
1095         /* body->size is actually the offset -eeb */
1096         rc = mds_sendpage(req, file, body->size, body->nlink);
1097
1098 out_file:
1099         filp_close(file, 0);
1100 out_pop:
1101         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
1102 out:
1103         req->rq_status = rc;
1104         RETURN(0);
1105 }
1106
1107 int mds_reint(struct ptlrpc_request *req, int offset,
1108               struct lustre_handle *lockh)
1109 {
1110         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1111         int rc;
1112         ENTRY;
1113
1114         OBD_ALLOC(rec, sizeof(*rec));
1115         if (rec == NULL)
1116                 RETURN(-ENOMEM);
1117
1118         rc = mds_update_unpack(req, offset, rec);
1119         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1120                 CERROR("invalid record\n");
1121                 GOTO(out, req->rq_status = -EINVAL);
1122         }
1123         /* rc will be used to interrupt a for loop over multiple records */
1124         rc = mds_reint_rec(rec, offset, req, lockh);
1125  out:
1126         OBD_FREE(rec, sizeof(*rec));
1127         RETURN(rc);
1128 }
1129
1130 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1131                                        struct obd_device *obd, int *process)
1132 {
1133         switch (req->rq_reqmsg->opc) {
1134         case MDS_CONNECT: /* This will never get here, but for completeness. */
1135         case OST_CONNECT: /* This will never get here, but for completeness. */
1136         case MDS_DISCONNECT:
1137         case OST_DISCONNECT:
1138                *process = 1;
1139                RETURN(0);
1140
1141         case MDS_CLOSE:
1142         case MDS_SYNC: /* used in unmounting */
1143         case OBD_PING:
1144         case MDS_REINT:
1145         case LDLM_ENQUEUE:
1146                 *process = target_queue_recovery_request(req, obd);
1147                 RETURN(0);
1148
1149         default:
1150                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1151                 *process = 0;
1152                 /* XXX what should we set rq_status to here? */
1153                 req->rq_status = -EAGAIN;
1154                 RETURN(ptlrpc_error(req));
1155         }
1156 }
1157
1158 static char *reint_names[] = {
1159         [REINT_SETATTR] "setattr",
1160         [REINT_CREATE]  "create",
1161         [REINT_LINK]    "link",
1162         [REINT_UNLINK]  "unlink",
1163         [REINT_RENAME]  "rename",
1164         [REINT_OPEN]    "open",
1165 };
1166
1167 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  |\
1168                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1169                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1170                             OBD_MD_FLID)
1171
1172 static int mdt_obj_create(struct ptlrpc_request *req)
1173 {
1174         struct ldlm_res_id res_id = { .name = {0} };
1175         struct obd_export *exp = req->rq_export;
1176         struct obd_device *obd = exp->exp_obd;
1177         struct mds_obd *mds = &obd->u.mds;
1178         struct ost_body *body, *repbody;
1179         int rc, size = sizeof(*repbody);
1180         char fidname[LL_FID_NAMELEN];
1181         struct inode *parent_inode;
1182         struct lustre_handle lockh;
1183         struct obd_run_ctxt saved;
1184         ldlm_policy_data_t policy;
1185         int mealen, flags = 0;
1186         unsigned int tmpname;
1187         struct obd_ucred uc;
1188         struct dentry *new;
1189         struct mea *mea;
1190         void *handle;
1191         ENTRY;
1192
1193         parent_inode = mds->mds_objects_dir->d_inode;
1194
1195         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
1196         if (body == NULL)
1197                 RETURN(-EFAULT);
1198
1199         uc.ouc_fsuid = body->oa.o_uid;
1200         uc.ouc_fsgid = body->oa.o_gid;
1201
1202         push_ctxt(&saved, &obd->obd_ctxt, &uc);
1203
1204         rc = lustre_pack_reply(req, 1, &size, NULL);
1205         if (rc)
1206                 RETURN(rc);
1207
1208         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1209
1210 repeat:
1211         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1212         LASSERT(!IS_ERR(handle));
1213
1214         tmpname = ll_insecure_random_int();
1215         sprintf(fidname, "%u", tmpname);
1216         new = simple_mkdir(mds->mds_objects_dir, fidname,
1217                         body->oa.o_mode, 1);
1218         if (IS_ERR(new)) {
1219                 CERROR("%s: can't create new inode %s) for mkdir: %d\n",
1220                        obd->obd_name, fidname, (int) PTR_ERR(new));
1221                 if (PTR_ERR(new) == -EEXIST) {
1222                         fsfilt_commit(obd, parent_inode, handle, 0);
1223                         goto repeat;
1224                 }
1225         }
1226         LASSERT(!IS_ERR(new));
1227         LASSERT(new->d_inode != NULL);
1228
1229         if (body->oa.o_valid & OBD_MD_FLID) {
1230                 /* this is new object for splitted dir. we have to
1231                  * prevent recursive splitting on it -bzzz */
1232                 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1233                 OBD_ALLOC(mea, mealen);
1234                 LASSERT(mea != NULL);
1235                 mea->mea_count = 0;
1236                 down(&new->d_inode->i_sem);
1237                 handle = fsfilt_start(obd, new->d_inode, FSFILT_OP_SETATTR, NULL);
1238                 LASSERT(!IS_ERR(handle));
1239                 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1240                 LASSERT(rc == 0);
1241                 fsfilt_commit(obd, new->d_inode, handle, 0);
1242                 LASSERT(rc == 0);
1243                 up(&new->d_inode->i_sem);
1244                 OBD_FREE(mea, mealen);
1245         }
1246         obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1247         repbody->oa.o_id = new->d_inode->i_ino;
1248         repbody->oa.o_generation = new->d_inode->i_generation;
1249         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1250
1251         down(&parent_inode->i_sem);
1252         rc = fsfilt_del_dir_entry(obd, new);
1253         up(&parent_inode->i_sem);
1254         LASSERT(rc == 0);
1255
1256         rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
1257         LASSERT(rc == 0);
1258
1259         res_id.name[0] = new->d_inode->i_ino;
1260         res_id.name[1] = new->d_inode->i_generation;
1261         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1262         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1263                         res_id, LDLM_IBITS, &policy,
1264                         LCK_EX, &flags, mds_blocking_ast,
1265                         ldlm_completion_ast, NULL, NULL,
1266                         NULL, 0, NULL, &lockh);
1267         LASSERT(rc == ELDLM_OK);
1268
1269         CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1270                         (unsigned long) new->d_inode->i_ino,
1271                         (unsigned long) new->d_inode->i_generation,
1272                         (unsigned) new->d_inode->i_mode);
1273
1274         l_dput(new);
1275         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
1276         ptlrpc_save_lock(req, &lockh, LCK_EX);
1277         RETURN(0);
1278 }
1279
1280 static int mds_get_info(struct obd_export *exp, __u32 keylen,
1281                            void *key, __u32 *vallen, void *val)
1282 {
1283         struct obd_device *obd;
1284         struct mds_obd *mds;
1285         ENTRY;
1286
1287         obd = class_exp2obd(exp);
1288         if (obd == NULL) {
1289                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1290                        exp->exp_handle.h_cookie);
1291                 RETURN(-EINVAL);
1292         }
1293
1294         mds = &obd->u.mds;
1295         keylen == strlen("mdsize");
1296         if (keylen && memcmp(key, "mdsize", keylen) == 0) {
1297                 __u32 *mdsize = val;
1298                 *vallen = sizeof(*mdsize);
1299                 *mdsize = mds->mds_max_mdsize;
1300                 RETURN(0);
1301         }
1302
1303         CDEBUG(D_IOCTL, "invalid key\n");
1304         RETURN(-EINVAL);
1305 }
1306
1307 static int mdt_get_info(struct ptlrpc_request *req)
1308 {
1309         char *key;
1310         struct obd_export *exp = req->rq_export;
1311         int keylen, rc = 0, size = sizeof(obd_id);
1312         obd_id *reply;
1313         ENTRY;
1314
1315         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1316         if (key == NULL) {
1317                 DEBUG_REQ(D_HA, req, "no get_info key");
1318                 RETURN(-EFAULT);
1319         }
1320         keylen = req->rq_reqmsg->buflens[0];
1321
1322         if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1323                 RETURN(-EPROTO);
1324
1325         rc = lustre_pack_reply(req, 1, &size, NULL);
1326         if (rc)
1327                 RETURN(rc);
1328
1329         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1330         rc = obd_get_info(exp, keylen, key, &size, reply);
1331         req->rq_repmsg->status = 0;
1332         RETURN(rc);
1333 }
1334
1335 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1336                            void *key, __u32 vallen, void *val)
1337 {
1338         struct obd_device *obd;
1339         struct mds_obd *mds;
1340         int rc;
1341         ENTRY;
1342
1343         obd = class_exp2obd(exp);
1344         if (obd == NULL) {
1345                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1346                        exp->exp_handle.h_cookie);
1347                 RETURN(-EINVAL);
1348         }
1349
1350         mds = &obd->u.mds;
1351         keylen == strlen("client");
1352         if (keylen && memcmp(key, "client", keylen) == 0) {
1353                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
1354                         atomic_inc(&mds->mds_real_clients);
1355                         CDEBUG(D_OTHER, "%s: peer from %s is real client (%d)\n",
1356                                         obd->obd_name,
1357                                         exp->exp_client_uuid.uuid,
1358                                         atomic_read(&mds->mds_real_clients));
1359                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
1360                 }
1361                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1362                 LASSERT(rc == 0);
1363                 RETURN(0);
1364         }
1365
1366         CDEBUG(D_IOCTL, "invalid key\n");
1367         RETURN(-EINVAL);
1368 }
1369
1370 static int mdt_set_info(struct ptlrpc_request *req)
1371 {
1372         char *key;
1373         struct obd_export *exp = req->rq_export;
1374         int keylen, rc = 0, size = sizeof(obd_id);
1375         ENTRY;
1376
1377         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1378         if (key == NULL) {
1379                 DEBUG_REQ(D_HA, req, "no get_info key");
1380                 RETURN(-EFAULT);
1381         }
1382         keylen = req->rq_reqmsg->buflens[0];
1383
1384         if ((keylen < strlen("client") || memcmp(key, "client", 6) != 0))
1385                 RETURN(-EPROTO);
1386
1387         rc = lustre_pack_reply(req, 0, NULL, NULL);
1388         if (rc)
1389                 RETURN(rc);
1390         rc = obd_set_info(exp, keylen, key, size, NULL);
1391         req->rq_repmsg->status = 0;
1392         RETURN(rc);
1393 }
1394
1395 extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *);
1396
1397 int mds_handle(struct ptlrpc_request *req)
1398 {
1399         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1400         int rc = 0;
1401         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1402         struct obd_device *obd = NULL;
1403         ENTRY;
1404
1405         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1406
1407         LASSERT(current->journal_info == NULL);
1408         /* XXX identical to OST */
1409         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1410                 struct mds_export_data *med;
1411                 int recovering, abort_recovery;
1412
1413                 if (req->rq_export == NULL) {
1414                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1415                                req->rq_reqmsg->opc);
1416                         req->rq_status = -ENOTCONN;
1417                         GOTO(out, rc = -ENOTCONN);
1418                 }
1419
1420                 med = &req->rq_export->exp_mds_data;
1421                 obd = req->rq_export->exp_obd;
1422                 mds = &obd->u.mds;
1423
1424                 /* Check for aborted recovery. */
1425                 spin_lock_bh(&obd->obd_processing_task_lock);
1426                 abort_recovery = obd->obd_abort_recovery;
1427                 recovering = obd->obd_recovering;
1428                 spin_unlock_bh(&obd->obd_processing_task_lock);
1429                 if (abort_recovery) {
1430                         target_abort_recovery(obd);
1431                 } else if (recovering) {
1432                         rc = mds_filter_recovery_request(req, obd,
1433                                                          &should_process);
1434                         if (rc || !should_process)
1435                                 RETURN(rc);
1436                 }
1437         }
1438
1439         switch (req->rq_reqmsg->opc) {
1440         case MDS_CONNECT:
1441                 DEBUG_REQ(D_INODE, req, "connect");
1442                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1443                 rc = target_handle_connect(req, mds_handle);
1444                 if (!rc)
1445                         /* Now that we have an export, set mds. */
1446                         mds = mds_req2mds(req);
1447                 break;
1448
1449         case MDS_DISCONNECT:
1450                 DEBUG_REQ(D_INODE, req, "disconnect");
1451                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1452                 rc = target_handle_disconnect(req);
1453                 req->rq_status = rc;            /* superfluous? */
1454                 break;
1455
1456         case MDS_GETSTATUS:
1457                 DEBUG_REQ(D_INODE, req, "getstatus");
1458                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1459                 rc = mds_getstatus(req);
1460                 break;
1461
1462         case MDS_GETATTR:
1463                 DEBUG_REQ(D_INODE, req, "getattr");
1464                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1465                 rc = mds_getattr(0, req);
1466                 break;
1467
1468         case MDS_GETATTR_NAME: {
1469                 struct lustre_handle lockh;
1470                 DEBUG_REQ(D_INODE, req, "getattr_name");
1471                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1472
1473                 /* If this request gets a reconstructed reply, we won't be
1474                  * acquiring any new locks in mds_getattr_name, so we don't
1475                  * want to cancel.
1476                  */
1477                 lockh.cookie = 0;
1478                 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1479                 /* this non-intent call (from an ioctl) is special */
1480                 req->rq_status = rc;
1481                 if (rc == 0 && lockh.cookie)
1482                         ldlm_lock_decref(&lockh, LCK_PR);
1483                 break;
1484         }
1485         case MDS_STATFS:
1486                 DEBUG_REQ(D_INODE, req, "statfs");
1487                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1488                 rc = mds_statfs(req);
1489                 break;
1490
1491         case MDS_READPAGE:
1492                 DEBUG_REQ(D_INODE, req, "readpage");
1493                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1494                 rc = mds_readpage(req);
1495
1496                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1497                         if (req->rq_reply_state) {
1498                                 lustre_free_reply_state (req->rq_reply_state);
1499                                 req->rq_reply_state = NULL;
1500                         }
1501                         RETURN(0);
1502                 }
1503
1504                 break;
1505
1506         case MDS_REINT: {
1507                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1508                 __u32  opc;
1509                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1510                                mds->mds_max_cookiesize};
1511                 int bufcount;
1512
1513                 /* NB only peek inside req now; mds_reint() will swab it */
1514                 if (opcp == NULL) {
1515                         CERROR ("Can't inspect opcode\n");
1516                         rc = -EINVAL;
1517                         break;
1518                 }
1519                 opc = *opcp;
1520                 if (lustre_msg_swabbed (req->rq_reqmsg))
1521                         __swab32s(&opc);
1522
1523                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1524                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1525                            reint_names[opc] == NULL) ? reint_names[opc] :
1526                                                        "unknown opcode");
1527
1528                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1529
1530                 if (opc == REINT_UNLINK)
1531                         bufcount = 3;
1532                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1533                         bufcount = 2;
1534                 else
1535                         bufcount = 1;
1536
1537                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1538                 if (rc)
1539                         break;
1540
1541                 rc = mds_reint(req, 0, NULL);
1542                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1543                 break;
1544         }
1545
1546         case MDS_CLOSE:
1547                 DEBUG_REQ(D_INODE, req, "close");
1548                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1549                 rc = mds_close(req);
1550                 break;
1551
1552         case MDS_DONE_WRITING:
1553                 DEBUG_REQ(D_INODE, req, "done_writing");
1554                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1555                 rc = mds_done_writing(req);
1556                 break;
1557
1558         case MDS_PIN:
1559                 DEBUG_REQ(D_INODE, req, "pin");
1560                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1561                 rc = mds_pin(req);
1562                 break;
1563
1564         case MDS_SYNC:
1565                 DEBUG_REQ(D_INODE, req, "sync");
1566                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1567                 rc = mds_sync(req);
1568                 break;
1569
1570         case OBD_PING:
1571                 DEBUG_REQ(D_INODE, req, "ping");
1572                 rc = target_handle_ping(req);
1573                 break;
1574
1575         case OBD_LOG_CANCEL:
1576                 CDEBUG(D_INODE, "log cancel\n");
1577                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1578                 rc = -ENOTSUPP; /* la la la */
1579                 break;
1580
1581         case LDLM_ENQUEUE:
1582                 DEBUG_REQ(D_INODE, req, "enqueue");
1583                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1584                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1585                                          ldlm_server_blocking_ast, NULL);
1586                 break;
1587         case LDLM_CONVERT:
1588                 DEBUG_REQ(D_INODE, req, "convert");
1589                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1590                 rc = ldlm_handle_convert(req);
1591                 break;
1592         case LDLM_BL_CALLBACK:
1593         case LDLM_CP_CALLBACK:
1594                 DEBUG_REQ(D_INODE, req, "callback");
1595                 CERROR("callbacks should not happen on MDS\n");
1596                 LBUG();
1597                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1598                 break;
1599         case LLOG_ORIGIN_HANDLE_CREATE:
1600                 DEBUG_REQ(D_INODE, req, "llog_init");
1601                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1602                 rc = llog_origin_handle_create(req);
1603                 break;
1604         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1605                 DEBUG_REQ(D_INODE, req, "llog next block");
1606                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1607                 rc = llog_origin_handle_next_block(req);
1608                 break;
1609         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1610                 DEBUG_REQ(D_INODE, req, "llog read header");
1611                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1612                 rc = llog_origin_handle_read_header(req);
1613                 break;
1614         case LLOG_ORIGIN_HANDLE_CLOSE:
1615                 DEBUG_REQ(D_INODE, req, "llog close");
1616                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1617                 rc = llog_origin_handle_close(req);
1618                 break;
1619         case OST_CREATE:
1620                 DEBUG_REQ(D_INODE, req, "ost_create");
1621                 rc = mdt_obj_create(req);
1622                 break;
1623         case OST_GET_INFO:
1624                 DEBUG_REQ(D_INODE, req, "get_info");
1625                 rc = mdt_get_info(req);
1626                 break;
1627         case OST_SET_INFO:
1628                 DEBUG_REQ(D_INODE, req, "set_info");
1629                 rc = mdt_set_info(req);
1630                 break;
1631         case OST_WRITE:
1632                 CDEBUG(D_INODE, "write\n");
1633                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1634                 rc = ost_brw_write(req, NULL);
1635                 LASSERT(current->journal_info == NULL);
1636                 /* mdt_brw sends its own replies */
1637                 RETURN(rc);
1638                 break;
1639         case LLOG_CATINFO:
1640                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1641                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1642                 rc = llog_catinfo(req);
1643                 break;
1644         default:
1645                 req->rq_status = -ENOTSUPP;
1646                 rc = ptlrpc_error(req);
1647                 RETURN(rc);
1648         }
1649
1650         LASSERT(current->journal_info == NULL);
1651
1652         EXIT;
1653
1654         /* If we're DISCONNECTing, the mds_export_data is already freed */
1655         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1656                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1657                 struct obd_device *obd = list_entry(mds, struct obd_device,
1658                                                     u.mds);
1659                 req->rq_repmsg->last_xid =
1660                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1661
1662                 if (!obd->obd_no_transno) {
1663                         req->rq_repmsg->last_committed =
1664                                 obd->obd_last_committed;
1665                 } else {
1666                         DEBUG_REQ(D_IOCTL, req,
1667                                   "not sending last_committed update");
1668                 }
1669                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1670                        ", xid "LPU64"\n",
1671                        mds->mds_last_transno, obd->obd_last_committed,
1672                        req->rq_xid);
1673         }
1674  out:
1675
1676         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1677                 if (obd && obd->obd_recovering) {
1678                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1679                         return target_queue_final_reply(req, rc);
1680                 }
1681                 /* Lost a race with recovery; let the error path DTRT. */
1682                 rc = req->rq_status = -ENOTCONN;
1683         }
1684
1685         target_send_reply(req, rc, fail);
1686         return 0;
1687 }
1688
1689 /* Update the server data on disk.  This stores the new mount_count and
1690  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1691  * then the server last_rcvd value may be less than that of the clients.
1692  * This will alert us that we may need to do client recovery.
1693  *
1694  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1695  */
1696 int mds_update_server_data(struct obd_device *obd, int force_sync)
1697 {
1698         struct mds_obd *mds = &obd->u.mds;
1699         struct mds_server_data *msd = mds->mds_server_data;
1700         struct file *filp = mds->mds_rcvd_filp;
1701         struct obd_run_ctxt saved;
1702         loff_t off = 0;
1703         int rc;
1704         ENTRY;
1705
1706         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1707         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1708
1709         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1710                mds->mds_mount_count, mds->mds_last_transno);
1711         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1712         if (rc)
1713                 CERROR("error writing MDS server data: rc = %d\n", rc);
1714         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1715
1716         RETURN(rc);
1717 }
1718
1719
1720 /* mount the file system (secretly) */
1721 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1722 {
1723         struct lustre_cfg* lcfg = buf;
1724         struct mds_obd *mds = &obd->u.mds;
1725         struct vfsmount *mnt;
1726         int rc = 0;
1727         unsigned long page;
1728         ENTRY;
1729
1730         dev_clear_rdonly(2);
1731
1732         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1733                 RETURN(rc = -EINVAL);
1734
1735         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1736         if (IS_ERR(obd->obd_fsops))
1737                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1738
1739         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1740
1741         /* we have to know mdsnum before touching underlying fs -bzzz */
1742         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4) {
1743                 class_uuid_t uuid;
1744
1745                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
1746                        obd->obd_name, lcfg->lcfg_inlbuf4);
1747
1748                 generate_random_uuid(uuid);
1749                 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
1750
1751                 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen4);
1752                 if (mds->mds_lmv_name == NULL)
1753                         RETURN(rc = -ENOMEM);
1754
1755                 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf4,
1756                                 lcfg->lcfg_inllen4);
1757                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1758                 if (rc) {
1759                         OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen4);
1760                         GOTO(err_ops, rc);
1761                 }
1762         }
1763
1764         if (!(page = __get_free_page(GFP_KERNEL)))
1765                 RETURN(-ENOMEM);
1766
1767         memset((void *)page, 0, PAGE_SIZE);
1768         sprintf((char *)page, "iopen_nopriv");
1769
1770         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1771                             lcfg->lcfg_inlbuf1, (void *)page);
1772         free_page(page);
1773         if (IS_ERR(mnt)) {
1774                 rc = PTR_ERR(mnt);
1775                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1776                 GOTO(err_ops, rc);
1777         }
1778
1779         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1780
1781         sema_init(&mds->mds_orphan_recovery_sem, 1);
1782         sema_init(&mds->mds_epoch_sem, 1);
1783         spin_lock_init(&mds->mds_transno_lock);
1784         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1785         atomic_set(&mds->mds_open_count, 0);
1786         atomic_set(&mds->mds_real_clients, 0);
1787
1788         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
1789                                                 LDLM_NAMESPACE_SERVER);
1790         if (obd->obd_namespace == NULL) {
1791                 mds_cleanup(obd, 0);
1792                 GOTO(err_put, rc = -ENOMEM);
1793         }
1794         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1795
1796         rc = mds_fs_setup(obd, mnt);
1797         if (rc) {
1798                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1799                 GOTO(err_ns, rc);
1800         }
1801
1802         rc = llog_start_commit_thread();
1803         if (rc < 0)
1804                 GOTO(err_fs, rc);
1805
1806
1807         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1808                 class_uuid_t uuid;
1809
1810                 generate_random_uuid(uuid);
1811                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1812
1813                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1814                 if (mds->mds_profile == NULL)
1815                         GOTO(err_fs, rc = -ENOMEM);
1816
1817                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1818                        lcfg->lcfg_inllen3);
1819         }
1820
1821         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1822                            "mds_ldlm_client", &obd->obd_ldlm_client);
1823         obd->obd_replayable = 1;
1824
1825         rc = mds_postsetup(obd);
1826         if (rc)
1827                 GOTO(err_fs, rc);
1828         RETURN(0);
1829
1830 err_fs:
1831         /* No extra cleanup needed for llog_init_commit_thread() */
1832         mds_fs_cleanup(obd, 0);
1833 err_ns:
1834         ldlm_namespace_free(obd->obd_namespace, 0);
1835         obd->obd_namespace = NULL;
1836 err_put:
1837         unlock_kernel();
1838         mntput(mds->mds_vfsmnt);
1839         mds->mds_sb = 0;
1840         lock_kernel();
1841 err_ops:
1842         fsfilt_put_ops(obd->obd_fsops);
1843         return rc;
1844 }
1845
1846 static int mds_postsetup(struct obd_device *obd)
1847 {
1848         struct mds_obd *mds = &obd->u.mds;
1849         struct llog_ctxt *ctxt;
1850         int rc = 0;
1851         ENTRY;
1852
1853         rc = llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
1854                         obd, 0, NULL, &llog_lvfs_ops);
1855         if (rc)
1856                 RETURN(rc);
1857
1858         if (mds->mds_profile) {
1859                 struct obd_run_ctxt saved;
1860                 struct lustre_profile *lprof;
1861                 struct config_llog_instance cfg;
1862
1863                 cfg.cfg_instance = NULL;
1864                 cfg.cfg_uuid = mds->mds_lov_uuid;
1865                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1866                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1867                 rc = class_config_parse_llog(ctxt, mds->mds_profile, &cfg);
1868                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1869                 if (rc)
1870                         GOTO(err_llog, rc);
1871
1872                 lprof = class_get_profile(mds->mds_profile);
1873                 if (lprof == NULL) {
1874                         CERROR("No profile found: %s\n", mds->mds_profile);
1875                         GOTO(err_cleanup, rc = -ENOENT);
1876                 }
1877                 rc = mds_lov_connect(obd, lprof->lp_osc);
1878                 if (rc)
1879                         GOTO(err_cleanup, rc);
1880
1881                 rc = mds_lmv_postsetup(obd);
1882                 if (rc)
1883                         GOTO(err_cleanup, rc);
1884         }
1885
1886         RETURN(rc);
1887
1888 err_cleanup:
1889         mds_lov_clean(obd);
1890 err_llog:
1891         ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1892         llog_cleanup(ctxt);
1893         RETURN(rc);
1894 }
1895
1896 static int mds_postrecov(struct obd_device *obd)
1897
1898 {
1899         struct llog_ctxt *ctxt;
1900         int rc, rc2;
1901         ENTRY;
1902
1903         LASSERT(!obd->obd_recovering);
1904         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
1905         LASSERT(ctxt != NULL);
1906
1907         rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
1908                           NULL, NULL, NULL);
1909         if (rc != 0) {
1910                 CERROR("faild at llog_origin_connect: %d\n", rc);
1911         }
1912
1913         rc = mds_cleanup_orphans(obd);
1914
1915         rc2 = mds_lov_set_nextid(obd);
1916         if (rc2 == 0)
1917                 rc2 = rc;
1918         RETURN(rc2);
1919 }
1920
1921 int mds_lov_clean(struct obd_device *obd)
1922 {
1923         struct mds_obd *mds = &obd->u.mds;
1924         struct llog_ctxt *ctxt;
1925
1926         if (mds->mds_profile) {
1927                 char * cln_prof;
1928                 struct config_llog_instance cfg;
1929                 struct obd_run_ctxt saved;
1930                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1931
1932                 OBD_ALLOC(cln_prof, len);
1933                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1934
1935                 cfg.cfg_instance = NULL;
1936                 cfg.cfg_uuid = mds->mds_lov_uuid;
1937
1938                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1939                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1940                 class_config_parse_llog(ctxt, cln_prof, &cfg);
1941                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1942
1943                 OBD_FREE(cln_prof, len);
1944                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1945                 mds->mds_profile = NULL;
1946         }
1947         RETURN(0);
1948 }
1949
1950 int mds_lmv_clean(struct obd_device *obd)
1951 {
1952         struct mds_obd *mds = &obd->u.mds;
1953
1954         if (mds->mds_lmv_name) {
1955                 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
1956                 mds->mds_lmv_name = NULL;
1957         }
1958         RETURN(0);
1959 }
1960
1961 static int mds_precleanup(struct obd_device *obd, int flags)
1962 {
1963         int rc = 0;
1964         ENTRY;
1965
1966         mds_lmv_clean(obd);
1967         mds_lov_disconnect(obd, flags);
1968         mds_lov_clean(obd);
1969         llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
1970         RETURN(rc);
1971 }
1972
1973 static int mds_cleanup(struct obd_device *obd, int flags)
1974 {
1975         struct mds_obd *mds = &obd->u.mds;
1976         ENTRY;
1977
1978         if (mds->mds_sb == NULL)
1979                 RETURN(0);
1980
1981         mds_update_server_data(obd, 1);
1982         if (mds->mds_lov_objids != NULL) {
1983                 OBD_FREE(mds->mds_lov_objids,
1984                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1985         }
1986         mds_fs_cleanup(obd, flags);
1987
1988         unlock_kernel();
1989
1990         /* 2 seems normal on mds, (may_umount() also expects 2
1991           fwiw), but we only see 1 at this point in obdfilter. */
1992         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1993                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1994                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1995
1996         mntput(mds->mds_vfsmnt);
1997
1998         mds->mds_sb = 0;
1999
2000         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2001
2002         spin_lock_bh(&obd->obd_processing_task_lock);
2003         if (obd->obd_recovering) {
2004                 target_cancel_recovery_timer(obd);
2005                 obd->obd_recovering = 0;
2006         }
2007         spin_unlock_bh(&obd->obd_processing_task_lock);
2008
2009         lock_kernel();
2010         dev_clear_rdonly(2);
2011         fsfilt_put_ops(obd->obd_fsops);
2012
2013         RETURN(0);
2014 }
2015
2016 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2017                                         struct ldlm_lock *new_lock,
2018                                         struct lustre_handle *lockh)
2019 {
2020         struct obd_export *exp = req->rq_export;
2021         struct obd_device *obd = exp->exp_obd;
2022         struct ldlm_request *dlmreq =
2023                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2024         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2025         struct list_head *iter;
2026
2027         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2028                 return;
2029
2030         l_lock(&obd->obd_namespace->ns_lock);
2031         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2032                 struct ldlm_lock *lock;
2033                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2034                 if (lock == new_lock)
2035                         continue;
2036                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2037                         lockh->cookie = lock->l_handle.h_cookie;
2038                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2039                                   lockh->cookie);
2040                         l_unlock(&obd->obd_namespace->ns_lock);
2041                         return;
2042                 }
2043         }
2044         l_unlock(&obd->obd_namespace->ns_lock);
2045
2046         /* This remote handle isn't enqueued, so we never received or
2047          * processed this request.  Clear MSG_RESENT, because it can
2048          * be handled like any normal request now. */
2049
2050         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2051
2052         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2053                   remote_hdl.cookie);
2054 }
2055
2056 int intent_disposition(struct ldlm_reply *rep, int flag)
2057 {
2058         if (!rep)
2059                 return 0;
2060         return (rep->lock_policy_res1 & flag);
2061 }
2062
2063 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2064 {
2065         if (!rep)
2066                 return;
2067         rep->lock_policy_res1 |= flag;
2068 }
2069
2070 static int mds_intent_policy(struct ldlm_namespace *ns,
2071                              struct ldlm_lock **lockp, void *req_cookie,
2072                              ldlm_mode_t mode, int flags, void *data)
2073 {
2074         struct ptlrpc_request *req = req_cookie;
2075         struct ldlm_lock *lock = *lockp;
2076         struct ldlm_intent *it;
2077         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2078         struct ldlm_reply *rep;
2079         struct lustre_handle lockh = { 0 };
2080         struct ldlm_lock *new_lock;
2081         int getattr_part = MDS_INODELOCK_UPDATE;
2082         int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2083                                           sizeof(struct mds_body),
2084                                           mds->mds_max_mdsize,
2085                                           mds->mds_max_cookiesize};
2086         ENTRY;
2087
2088         LASSERT(req != NULL);
2089
2090         if (req->rq_reqmsg->bufcount <= 1) {
2091                 /* No intent was provided */
2092                 int size = sizeof(struct ldlm_reply);
2093                 rc = lustre_pack_reply(req, 1, &size, NULL);
2094                 LASSERT(rc == 0);
2095                 RETURN(0);
2096         }
2097
2098         it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2099         if (it == NULL) {
2100                 CERROR("Intent missing\n");
2101                 RETURN(req->rq_status = -EFAULT);
2102         }
2103
2104         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2105
2106         rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
2107                                NULL);
2108         if (rc)
2109                 RETURN(req->rq_status = rc);
2110
2111         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2112         intent_set_disposition(rep, DISP_IT_EXECD);
2113
2114         fixup_handle_for_resent_req(req, lock, &lockh);
2115
2116         /* execute policy */
2117         switch ((long)it->opc) {
2118         case IT_OPEN:
2119         case IT_CREAT|IT_OPEN:
2120                 /* XXX swab here to assert that an mds_open reint
2121                  * packet is following */
2122                 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2123 #if 0
2124                 /* We abort the lock if the lookup was negative and
2125                  * we did not make it to the OPEN portion */
2126                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2127                         RETURN(ELDLM_LOCK_ABORTED);
2128                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2129                     !intent_disposition(rep, DISP_OPEN_OPEN))
2130 #endif
2131                         if (rep->lock_policy_res2) {
2132                                 /* mds_open returns ENOLCK where it
2133                                  * should return zero, but it has no
2134                                  * lock to return */
2135                                 if (rep->lock_policy_res2 == ENOLCK)
2136                                         rep->lock_policy_res2 = 0;
2137                                 RETURN(ELDLM_LOCK_ABORTED);
2138                         }
2139                 break;
2140         case IT_LOOKUP:
2141                 getattr_part = MDS_INODELOCK_LOOKUP;
2142         case IT_GETATTR:
2143                 getattr_part |= MDS_INODELOCK_LOOKUP;
2144         case IT_READDIR:
2145                 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2146                                                          getattr_part);
2147                 /* FIXME: LDLM can set req->rq_status. MDS sets
2148                    policy_res{1,2} with disposition and status.
2149                    - replay: returns 0 & req->status is old status
2150                    - otherwise: returns req->status */
2151                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2152                         rep->lock_policy_res2 = 0;
2153                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2154                     rep->lock_policy_res2)
2155                         RETURN(ELDLM_LOCK_ABORTED);
2156                 if (req->rq_status != 0) {
2157                         LBUG();
2158                         rep->lock_policy_res2 = req->rq_status;
2159                         RETURN(ELDLM_LOCK_ABORTED);
2160                 }
2161                 break;
2162         default:
2163                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2164                 LBUG();
2165         }
2166
2167         /* By this point, whatever function we called above must have either
2168          * filled in 'lockh', been an intent replay, or returned an error.  We
2169          * want to allow replayed RPCs to not get a lock, since we would just
2170          * drop it below anyways because lock replay is done separately by the
2171          * client afterwards.  For regular RPCs we want to give the new lock to
2172          * the client instead of whatever lock it was about to get. */
2173         new_lock = ldlm_handle2lock(&lockh);
2174         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2175                 RETURN(0);
2176
2177         LASSERT(new_lock != NULL);
2178
2179         /* If we've already given this lock to a client once, then we should
2180          * have no readers or writers.  Otherwise, we should have one reader
2181          * _or_ writer ref (which will be zeroed below) before returning the
2182          * lock to a client. */
2183         if (new_lock->l_export == req->rq_export) {
2184                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2185         } else {
2186                 LASSERT(new_lock->l_export == NULL);
2187                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2188         }
2189
2190         *lockp = new_lock;
2191
2192         if (new_lock->l_export == req->rq_export) {
2193                 /* Already gave this to the client, which means that we
2194                  * reconstructed a reply. */
2195                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2196                         MSG_RESENT);
2197                 RETURN(ELDLM_LOCK_REPLACED);
2198         }
2199
2200         /* Fixup the lock to be given to the client */
2201         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2202         new_lock->l_readers = 0;
2203         new_lock->l_writers = 0;
2204
2205         new_lock->l_export = class_export_get(req->rq_export);
2206         list_add(&new_lock->l_export_chain,
2207                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2208
2209         new_lock->l_blocking_ast = lock->l_blocking_ast;
2210         new_lock->l_completion_ast = lock->l_completion_ast;
2211
2212         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2213                sizeof(lock->l_remote_handle));
2214
2215         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2216
2217         LDLM_LOCK_PUT(new_lock);
2218         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2219
2220         RETURN(ELDLM_LOCK_REPLACED);
2221 }
2222
2223 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2224 {
2225         struct lprocfs_static_vars lvars;
2226
2227         lprocfs_init_multi_vars(0, &lvars);
2228         return lprocfs_obd_attach(dev, lvars.obd_vars);
2229 }
2230
2231 int mds_detach(struct obd_device *dev)
2232 {
2233         return lprocfs_obd_detach(dev);
2234 }
2235
2236 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2237 {
2238         struct lprocfs_static_vars lvars;
2239
2240         lprocfs_init_multi_vars(1, &lvars);
2241         return lprocfs_obd_attach(dev, lvars.obd_vars);
2242 }
2243
2244 int mdt_detach(struct obd_device *dev)
2245 {
2246         return lprocfs_obd_detach(dev);
2247 }
2248
2249 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
2250 {
2251         struct mds_obd *mds = &obddev->u.mds;
2252         int rc = 0;
2253         ENTRY;
2254
2255         mds->mds_service =
2256                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2257                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2258                                 mds_handle, "mds",
2259                                 obddev->obd_proc_entry);
2260
2261         if (!mds->mds_service) {
2262                 CERROR("failed to start service\n");
2263                 RETURN(rc = -ENOMEM);
2264         }
2265
2266         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
2267                                     "ll_mdt");
2268         if (rc)
2269                 GOTO(err_thread, rc);
2270
2271         mds->mds_setattr_service =
2272                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2273                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2274                                 mds_handle, "mds_setattr",
2275                                 obddev->obd_proc_entry);
2276         if (!mds->mds_setattr_service) {
2277                 CERROR("failed to start getattr service\n");
2278                 GOTO(err_thread, rc = -ENOMEM);
2279         }
2280
2281         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
2282                                  MDT_NUM_THREADS, "ll_mdt_attr");
2283         if (rc)
2284                 GOTO(err_thread2, rc);
2285
2286         mds->mds_readpage_service =
2287                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2288                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2289                                 mds_handle, "mds_readpage",
2290                                 obddev->obd_proc_entry);
2291         if (!mds->mds_readpage_service) {
2292                 CERROR("failed to start readpage service\n");
2293                 GOTO(err_thread2, rc = -ENOMEM);
2294         }
2295
2296         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
2297                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
2298
2299         if (rc)
2300                 GOTO(err_thread3, rc);
2301
2302         RETURN(0);
2303
2304 err_thread3:
2305         ptlrpc_unregister_service(mds->mds_readpage_service);
2306 err_thread2:
2307         ptlrpc_unregister_service(mds->mds_setattr_service);
2308 err_thread:
2309         ptlrpc_unregister_service(mds->mds_service);
2310         return rc;
2311 }
2312
2313
2314 static int mdt_cleanup(struct obd_device *obddev, int flags)
2315 {
2316         struct mds_obd *mds = &obddev->u.mds;
2317         ENTRY;
2318
2319         ptlrpc_stop_all_threads(mds->mds_readpage_service);
2320         ptlrpc_unregister_service(mds->mds_readpage_service);
2321
2322         ptlrpc_stop_all_threads(mds->mds_setattr_service);
2323         ptlrpc_unregister_service(mds->mds_setattr_service);
2324
2325         ptlrpc_stop_all_threads(mds->mds_service);
2326         ptlrpc_unregister_service(mds->mds_service);
2327
2328         RETURN(0);
2329 }
2330
2331 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
2332 {
2333         struct obd_device *obd = data;
2334         struct ll_fid fid;
2335         fid.id = id;
2336         fid.generation = gen;
2337         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2338 }
2339
2340 struct lvfs_callback_ops mds_lvfs_ops = {
2341         l_fid2dentry:     mds_lvfs_fid2dentry,
2342 };
2343
2344 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2345                 int objcount, struct obd_ioobj *obj,
2346                 int niocount, struct niobuf_remote *nb,
2347                 struct niobuf_local *res,
2348                 struct obd_trans_info *oti);
2349 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2350                  int objcount, struct obd_ioobj *obj, int niocount,
2351                  struct niobuf_local *res, struct obd_trans_info *oti,
2352                  int rc);
2353
2354 /* use obd ops to offer management infrastructure */
2355 static struct obd_ops mds_obd_ops = {
2356         o_owner:       THIS_MODULE,
2357         o_attach:      mds_attach,
2358         o_detach:      mds_detach,
2359         o_connect:     mds_connect,
2360         o_init_export:  mds_init_export,
2361         o_destroy_export:  mds_destroy_export,
2362         o_disconnect:  mds_disconnect,
2363         o_setup:       mds_setup,
2364         o_precleanup:  mds_precleanup,
2365         o_cleanup:     mds_cleanup,
2366         o_postrecov:   mds_postrecov,
2367         o_statfs:      mds_obd_statfs,
2368         o_iocontrol:   mds_iocontrol,
2369         o_create:      mds_obd_create,
2370         o_destroy:     mds_obd_destroy,
2371         o_llog_init:   mds_llog_init,
2372         o_llog_finish: mds_llog_finish,
2373         o_notify:      mds_notify,
2374         o_get_info:    mds_get_info,
2375         o_set_info:    mds_set_info,
2376         o_preprw:      mds_preprw,
2377         o_commitrw:    mds_commitrw,
2378 };
2379
2380 static struct obd_ops mdt_obd_ops = {
2381         o_owner:       THIS_MODULE,
2382         o_attach:      mdt_attach,
2383         o_detach:      mdt_detach,
2384         o_setup:       mdt_setup,
2385         o_cleanup:     mdt_cleanup,
2386 };
2387
2388 static int __init mds_init(void)
2389 {
2390         struct lprocfs_static_vars lvars;
2391
2392         lprocfs_init_multi_vars(0, &lvars);
2393         class_register_type(&mds_obd_ops, NULL, lvars.module_vars, LUSTRE_MDS_NAME);
2394         lprocfs_init_multi_vars(1, &lvars);
2395         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars, LUSTRE_MDT_NAME);
2396
2397         return 0;
2398 }
2399
2400 static void /*__exit*/ mds_exit(void)
2401 {
2402         class_unregister_type(LUSTRE_MDS_NAME);
2403         class_unregister_type(LUSTRE_MDT_NAME);
2404 }
2405
2406 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2407 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2408 MODULE_LICENSE("GPL");
2409
2410 module_init(mds_init);
2411 module_exit(mds_exit);