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         unsigned int tmpname = ll_insecure_random_int();
1175         struct ldlm_res_id res_id = { .name = {0} };
1176         struct obd_export *exp = req->rq_export;
1177         struct obd_device *obd = exp->exp_obd;
1178         struct mds_obd *mds = &obd->u.mds;
1179         struct ost_body *body, *repbody;
1180         int rc, size = sizeof(*repbody);
1181         char fidname[LL_FID_NAMELEN];
1182         struct inode *parent_inode;
1183         struct lustre_handle lockh;
1184         struct obd_run_ctxt saved;
1185         ldlm_policy_data_t policy;
1186         int mealen, flags = 0;
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         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1211         LASSERT(!IS_ERR(handle));
1212
1213         sprintf(fidname, "%u", tmpname);
1214         new = simple_mkdir(mds->mds_objects_dir, fidname,
1215                         body->oa.o_mode, 1);
1216         LASSERT(!IS_ERR(new));
1217         LASSERT(new->d_inode != NULL);
1218
1219         if (body->oa.o_valid & OBD_MD_FLID) {
1220                 /* this is new object for splitted dir. we have to
1221                  * prevent recursive splitting on it -bzzz */
1222                 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1223                 OBD_ALLOC(mea, mealen);
1224                 LASSERT(mea != NULL);
1225                 mea->mea_count = 0;
1226                 down(&new->d_inode->i_sem);
1227                 handle = fsfilt_start(obd, new->d_inode, FSFILT_OP_SETATTR, NULL);
1228                 LASSERT(!IS_ERR(handle));
1229                 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1230                 LASSERT(rc == 0);
1231                 fsfilt_commit(obd, new->d_inode, handle, 0);
1232                 LASSERT(rc == 0);
1233                 up(&new->d_inode->i_sem);
1234                 OBD_FREE(mea, mealen);
1235         }
1236         obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1237         repbody->oa.o_id = new->d_inode->i_ino;
1238         repbody->oa.o_generation = new->d_inode->i_generation;
1239         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1240
1241         down(&parent_inode->i_sem);
1242         rc = fsfilt_del_dir_entry(obd, new);
1243         up(&parent_inode->i_sem);
1244         LASSERT(rc == 0);
1245
1246         rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
1247         LASSERT(rc == 0);
1248
1249         res_id.name[0] = new->d_inode->i_ino;
1250         res_id.name[1] = new->d_inode->i_generation;
1251         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1252         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1253                         res_id, LDLM_IBITS, &policy,
1254                         LCK_EX, &flags, mds_blocking_ast,
1255                         ldlm_completion_ast, NULL, NULL,
1256                         NULL, 0, NULL, &lockh);
1257         LASSERT(rc == ELDLM_OK);
1258
1259         CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1260                         (unsigned long) new->d_inode->i_ino,
1261                         (unsigned long) new->d_inode->i_generation,
1262                         (unsigned) new->d_inode->i_mode);
1263
1264         l_dput(new);
1265         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
1266         ptlrpc_save_lock(req, &lockh, LCK_EX);
1267         RETURN(0);
1268 }
1269
1270 static int mds_get_info(struct obd_export *exp, __u32 keylen,
1271                            void *key, __u32 *vallen, void *val)
1272 {
1273         struct obd_device *obd;
1274         struct mds_obd *mds;
1275         ENTRY;
1276
1277         obd = class_exp2obd(exp);
1278         if (obd == NULL) {
1279                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1280                        exp->exp_handle.h_cookie);
1281                 RETURN(-EINVAL);
1282         }
1283
1284         mds = &obd->u.mds;
1285         keylen == strlen("mdsize");
1286         if (keylen && memcmp(key, "mdsize", keylen) == 0) {
1287                 __u32 *mdsize = val;
1288                 *vallen = sizeof(*mdsize);
1289                 *mdsize = mds->mds_max_mdsize;
1290                 RETURN(0);
1291         }
1292
1293         CDEBUG(D_IOCTL, "invalid key\n");
1294         RETURN(-EINVAL);
1295 }
1296
1297 static int mdt_get_info(struct ptlrpc_request *req)
1298 {
1299         char *key;
1300         struct obd_export *exp = req->rq_export;
1301         int keylen, rc = 0, size = sizeof(obd_id);
1302         obd_id *reply;
1303         ENTRY;
1304
1305         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1306         if (key == NULL) {
1307                 DEBUG_REQ(D_HA, req, "no get_info key");
1308                 RETURN(-EFAULT);
1309         }
1310         keylen = req->rq_reqmsg->buflens[0];
1311
1312         if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1313                 RETURN(-EPROTO);
1314
1315         rc = lustre_pack_reply(req, 1, &size, NULL);
1316         if (rc)
1317                 RETURN(rc);
1318
1319         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1320         rc = obd_get_info(exp, keylen, key, &size, reply);
1321         req->rq_repmsg->status = 0;
1322         RETURN(rc);
1323 }
1324
1325 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1326                            void *key, __u32 vallen, void *val)
1327 {
1328         struct obd_device *obd;
1329         struct mds_obd *mds;
1330         int rc;
1331         ENTRY;
1332
1333         obd = class_exp2obd(exp);
1334         if (obd == NULL) {
1335                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1336                        exp->exp_handle.h_cookie);
1337                 RETURN(-EINVAL);
1338         }
1339
1340         mds = &obd->u.mds;
1341         keylen == strlen("client");
1342         if (keylen && memcmp(key, "client", keylen) == 0) {
1343                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
1344                         atomic_inc(&mds->mds_real_clients);
1345                         CDEBUG(D_OTHER, "%s: peer from %s is real client (%d)\n",
1346                                         obd->obd_name,
1347                                         exp->exp_client_uuid.uuid,
1348                                         atomic_read(&mds->mds_real_clients));
1349                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
1350                 }
1351                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1352                 LASSERT(rc == 0);
1353                 RETURN(0);
1354         }
1355
1356         CDEBUG(D_IOCTL, "invalid key\n");
1357         RETURN(-EINVAL);
1358 }
1359
1360 static int mdt_set_info(struct ptlrpc_request *req)
1361 {
1362         char *key;
1363         struct obd_export *exp = req->rq_export;
1364         int keylen, rc = 0, size = sizeof(obd_id);
1365         ENTRY;
1366
1367         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1368         if (key == NULL) {
1369                 DEBUG_REQ(D_HA, req, "no get_info key");
1370                 RETURN(-EFAULT);
1371         }
1372         keylen = req->rq_reqmsg->buflens[0];
1373
1374         if ((keylen < strlen("client") || memcmp(key, "client", 6) != 0))
1375                 RETURN(-EPROTO);
1376
1377         rc = lustre_pack_reply(req, 0, NULL, NULL);
1378         if (rc)
1379                 RETURN(rc);
1380         rc = obd_set_info(exp, keylen, key, size, NULL);
1381         req->rq_repmsg->status = 0;
1382         RETURN(rc);
1383 }
1384
1385 extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *);
1386
1387 int mds_handle(struct ptlrpc_request *req)
1388 {
1389         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1390         int rc = 0;
1391         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1392         struct obd_device *obd = NULL;
1393         ENTRY;
1394
1395         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1396
1397         LASSERT(current->journal_info == NULL);
1398         /* XXX identical to OST */
1399         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1400                 struct mds_export_data *med;
1401                 int recovering, abort_recovery;
1402
1403                 if (req->rq_export == NULL) {
1404                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1405                                req->rq_reqmsg->opc);
1406                         req->rq_status = -ENOTCONN;
1407                         GOTO(out, rc = -ENOTCONN);
1408                 }
1409
1410                 med = &req->rq_export->exp_mds_data;
1411                 obd = req->rq_export->exp_obd;
1412                 mds = &obd->u.mds;
1413
1414                 /* Check for aborted recovery. */
1415                 spin_lock_bh(&obd->obd_processing_task_lock);
1416                 abort_recovery = obd->obd_abort_recovery;
1417                 recovering = obd->obd_recovering;
1418                 spin_unlock_bh(&obd->obd_processing_task_lock);
1419                 if (abort_recovery) {
1420                         target_abort_recovery(obd);
1421                 } else if (recovering) {
1422                         rc = mds_filter_recovery_request(req, obd,
1423                                                          &should_process);
1424                         if (rc || !should_process)
1425                                 RETURN(rc);
1426                 }
1427         }
1428
1429         switch (req->rq_reqmsg->opc) {
1430         case MDS_CONNECT:
1431                 DEBUG_REQ(D_INODE, req, "connect");
1432                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1433                 rc = target_handle_connect(req, mds_handle);
1434                 if (!rc)
1435                         /* Now that we have an export, set mds. */
1436                         mds = mds_req2mds(req);
1437                 break;
1438
1439         case MDS_DISCONNECT:
1440                 DEBUG_REQ(D_INODE, req, "disconnect");
1441                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1442                 rc = target_handle_disconnect(req);
1443                 req->rq_status = rc;            /* superfluous? */
1444                 break;
1445
1446         case MDS_GETSTATUS:
1447                 DEBUG_REQ(D_INODE, req, "getstatus");
1448                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1449                 rc = mds_getstatus(req);
1450                 break;
1451
1452         case MDS_GETATTR:
1453                 DEBUG_REQ(D_INODE, req, "getattr");
1454                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1455                 rc = mds_getattr(0, req);
1456                 break;
1457
1458         case MDS_GETATTR_NAME: {
1459                 struct lustre_handle lockh;
1460                 DEBUG_REQ(D_INODE, req, "getattr_name");
1461                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1462
1463                 /* If this request gets a reconstructed reply, we won't be
1464                  * acquiring any new locks in mds_getattr_name, so we don't
1465                  * want to cancel.
1466                  */
1467                 lockh.cookie = 0;
1468                 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1469                 /* this non-intent call (from an ioctl) is special */
1470                 req->rq_status = rc;
1471                 if (rc == 0 && lockh.cookie)
1472                         ldlm_lock_decref(&lockh, LCK_PR);
1473                 break;
1474         }
1475         case MDS_STATFS:
1476                 DEBUG_REQ(D_INODE, req, "statfs");
1477                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1478                 rc = mds_statfs(req);
1479                 break;
1480
1481         case MDS_READPAGE:
1482                 DEBUG_REQ(D_INODE, req, "readpage");
1483                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1484                 rc = mds_readpage(req);
1485
1486                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1487                         if (req->rq_reply_state) {
1488                                 lustre_free_reply_state (req->rq_reply_state);
1489                                 req->rq_reply_state = NULL;
1490                         }
1491                         RETURN(0);
1492                 }
1493
1494                 break;
1495
1496         case MDS_REINT: {
1497                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1498                 __u32  opc;
1499                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1500                                mds->mds_max_cookiesize};
1501                 int bufcount;
1502
1503                 /* NB only peek inside req now; mds_reint() will swab it */
1504                 if (opcp == NULL) {
1505                         CERROR ("Can't inspect opcode\n");
1506                         rc = -EINVAL;
1507                         break;
1508                 }
1509                 opc = *opcp;
1510                 if (lustre_msg_swabbed (req->rq_reqmsg))
1511                         __swab32s(&opc);
1512
1513                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1514                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1515                            reint_names[opc] == NULL) ? reint_names[opc] :
1516                                                        "unknown opcode");
1517
1518                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1519
1520                 if (opc == REINT_UNLINK)
1521                         bufcount = 3;
1522                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1523                         bufcount = 2;
1524                 else
1525                         bufcount = 1;
1526
1527                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1528                 if (rc)
1529                         break;
1530
1531                 rc = mds_reint(req, 0, NULL);
1532                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1533                 break;
1534         }
1535
1536         case MDS_CLOSE:
1537                 DEBUG_REQ(D_INODE, req, "close");
1538                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1539                 rc = mds_close(req);
1540                 break;
1541
1542         case MDS_DONE_WRITING:
1543                 DEBUG_REQ(D_INODE, req, "done_writing");
1544                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1545                 rc = mds_done_writing(req);
1546                 break;
1547
1548         case MDS_PIN:
1549                 DEBUG_REQ(D_INODE, req, "pin");
1550                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1551                 rc = mds_pin(req);
1552                 break;
1553
1554         case MDS_SYNC:
1555                 DEBUG_REQ(D_INODE, req, "sync");
1556                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1557                 rc = mds_sync(req);
1558                 break;
1559
1560         case OBD_PING:
1561                 DEBUG_REQ(D_INODE, req, "ping");
1562                 rc = target_handle_ping(req);
1563                 break;
1564
1565         case OBD_LOG_CANCEL:
1566                 CDEBUG(D_INODE, "log cancel\n");
1567                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1568                 rc = -ENOTSUPP; /* la la la */
1569                 break;
1570
1571         case LDLM_ENQUEUE:
1572                 DEBUG_REQ(D_INODE, req, "enqueue");
1573                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1574                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1575                                          ldlm_server_blocking_ast, NULL);
1576                 break;
1577         case LDLM_CONVERT:
1578                 DEBUG_REQ(D_INODE, req, "convert");
1579                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1580                 rc = ldlm_handle_convert(req);
1581                 break;
1582         case LDLM_BL_CALLBACK:
1583         case LDLM_CP_CALLBACK:
1584                 DEBUG_REQ(D_INODE, req, "callback");
1585                 CERROR("callbacks should not happen on MDS\n");
1586                 LBUG();
1587                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1588                 break;
1589         case LLOG_ORIGIN_HANDLE_CREATE:
1590                 DEBUG_REQ(D_INODE, req, "llog_init");
1591                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1592                 rc = llog_origin_handle_create(req);
1593                 break;
1594         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1595                 DEBUG_REQ(D_INODE, req, "llog next block");
1596                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1597                 rc = llog_origin_handle_next_block(req);
1598                 break;
1599         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1600                 DEBUG_REQ(D_INODE, req, "llog read header");
1601                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1602                 rc = llog_origin_handle_read_header(req);
1603                 break;
1604         case LLOG_ORIGIN_HANDLE_CLOSE:
1605                 DEBUG_REQ(D_INODE, req, "llog close");
1606                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1607                 rc = llog_origin_handle_close(req);
1608                 break;
1609         case OST_CREATE:
1610                 DEBUG_REQ(D_INODE, req, "ost_create");
1611                 rc = mdt_obj_create(req);
1612                 break;
1613         case OST_GET_INFO:
1614                 DEBUG_REQ(D_INODE, req, "get_info");
1615                 rc = mdt_get_info(req);
1616                 break;
1617         case OST_SET_INFO:
1618                 DEBUG_REQ(D_INODE, req, "set_info");
1619                 rc = mdt_set_info(req);
1620                 break;
1621         case OST_WRITE:
1622                 CDEBUG(D_INODE, "write\n");
1623                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1624                 rc = ost_brw_write(req, NULL);
1625                 LASSERT(current->journal_info == NULL);
1626                 /* mdt_brw sends its own replies */
1627                 RETURN(rc);
1628                 break;
1629         case LLOG_CATINFO:
1630                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1631                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1632                 rc = llog_catinfo(req);
1633                 break;
1634         default:
1635                 req->rq_status = -ENOTSUPP;
1636                 rc = ptlrpc_error(req);
1637                 RETURN(rc);
1638         }
1639
1640         LASSERT(current->journal_info == NULL);
1641
1642         EXIT;
1643
1644         /* If we're DISCONNECTing, the mds_export_data is already freed */
1645         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1646                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1647                 struct obd_device *obd = list_entry(mds, struct obd_device,
1648                                                     u.mds);
1649                 req->rq_repmsg->last_xid =
1650                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1651
1652                 if (!obd->obd_no_transno) {
1653                         req->rq_repmsg->last_committed =
1654                                 obd->obd_last_committed;
1655                 } else {
1656                         DEBUG_REQ(D_IOCTL, req,
1657                                   "not sending last_committed update");
1658                 }
1659                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1660                        ", xid "LPU64"\n",
1661                        mds->mds_last_transno, obd->obd_last_committed,
1662                        req->rq_xid);
1663         }
1664  out:
1665
1666         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1667                 if (obd && obd->obd_recovering) {
1668                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1669                         return target_queue_final_reply(req, rc);
1670                 }
1671                 /* Lost a race with recovery; let the error path DTRT. */
1672                 rc = req->rq_status = -ENOTCONN;
1673         }
1674
1675         target_send_reply(req, rc, fail);
1676         return 0;
1677 }
1678
1679 /* Update the server data on disk.  This stores the new mount_count and
1680  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1681  * then the server last_rcvd value may be less than that of the clients.
1682  * This will alert us that we may need to do client recovery.
1683  *
1684  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1685  */
1686 int mds_update_server_data(struct obd_device *obd, int force_sync)
1687 {
1688         struct mds_obd *mds = &obd->u.mds;
1689         struct mds_server_data *msd = mds->mds_server_data;
1690         struct file *filp = mds->mds_rcvd_filp;
1691         struct obd_run_ctxt saved;
1692         loff_t off = 0;
1693         int rc;
1694         ENTRY;
1695
1696         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1697         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1698
1699         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1700                mds->mds_mount_count, mds->mds_last_transno);
1701         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1702         if (rc)
1703                 CERROR("error writing MDS server data: rc = %d\n", rc);
1704         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1705
1706         RETURN(rc);
1707 }
1708
1709
1710 /* mount the file system (secretly) */
1711 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1712 {
1713         struct lustre_cfg* lcfg = buf;
1714         struct mds_obd *mds = &obd->u.mds;
1715         struct vfsmount *mnt;
1716         int rc = 0;
1717         unsigned long page;
1718         ENTRY;
1719
1720         dev_clear_rdonly(2);
1721
1722         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1723                 RETURN(rc = -EINVAL);
1724
1725         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1726         if (IS_ERR(obd->obd_fsops))
1727                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1728
1729         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1730
1731         /* we have to know mdsnum before touching underlying fs -bzzz */
1732         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4) {
1733                 class_uuid_t uuid;
1734
1735                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
1736                        obd->obd_name, lcfg->lcfg_inlbuf4);
1737
1738                 generate_random_uuid(uuid);
1739                 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
1740
1741                 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen4);
1742                 if (mds->mds_lmv_name == NULL)
1743                         RETURN(rc = -ENOMEM);
1744
1745                 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf4,
1746                                 lcfg->lcfg_inllen4);
1747                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1748                 if (rc) {
1749                         OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen4);
1750                         GOTO(err_ops, rc);
1751                 }
1752         }
1753
1754         if (!(page = __get_free_page(GFP_KERNEL)))
1755                 RETURN(-ENOMEM);
1756
1757         memset((void *)page, 0, PAGE_SIZE);
1758         sprintf((char *)page, "iopen_nopriv");
1759
1760         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1761                             lcfg->lcfg_inlbuf1, (void *)page);
1762         free_page(page);
1763         if (IS_ERR(mnt)) {
1764                 rc = PTR_ERR(mnt);
1765                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1766                 GOTO(err_ops, rc);
1767         }
1768
1769         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1770
1771         sema_init(&mds->mds_orphan_recovery_sem, 1);
1772         sema_init(&mds->mds_epoch_sem, 1);
1773         spin_lock_init(&mds->mds_transno_lock);
1774         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1775         atomic_set(&mds->mds_open_count, 0);
1776         atomic_set(&mds->mds_real_clients, 0);
1777
1778         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
1779                                                 LDLM_NAMESPACE_SERVER);
1780         if (obd->obd_namespace == NULL) {
1781                 mds_cleanup(obd, 0);
1782                 GOTO(err_put, rc = -ENOMEM);
1783         }
1784         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1785
1786         rc = mds_fs_setup(obd, mnt);
1787         if (rc) {
1788                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1789                 GOTO(err_ns, rc);
1790         }
1791
1792         rc = llog_start_commit_thread();
1793         if (rc < 0)
1794                 GOTO(err_fs, rc);
1795
1796
1797         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1798                 class_uuid_t uuid;
1799
1800                 generate_random_uuid(uuid);
1801                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1802
1803                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1804                 if (mds->mds_profile == NULL)
1805                         GOTO(err_fs, rc = -ENOMEM);
1806
1807                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1808                        lcfg->lcfg_inllen3);
1809         }
1810
1811         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1812                            "mds_ldlm_client", &obd->obd_ldlm_client);
1813         obd->obd_replayable = 1;
1814
1815         rc = mds_postsetup(obd);
1816         if (rc)
1817                 GOTO(err_fs, rc);
1818         RETURN(0);
1819
1820 err_fs:
1821         /* No extra cleanup needed for llog_init_commit_thread() */
1822         mds_fs_cleanup(obd, 0);
1823 err_ns:
1824         ldlm_namespace_free(obd->obd_namespace, 0);
1825         obd->obd_namespace = NULL;
1826 err_put:
1827         unlock_kernel();
1828         mntput(mds->mds_vfsmnt);
1829         mds->mds_sb = 0;
1830         lock_kernel();
1831 err_ops:
1832         fsfilt_put_ops(obd->obd_fsops);
1833         return rc;
1834 }
1835
1836 static int mds_postsetup(struct obd_device *obd)
1837 {
1838         struct mds_obd *mds = &obd->u.mds;
1839         struct llog_ctxt *ctxt;
1840         int rc = 0;
1841         ENTRY;
1842
1843         rc = llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
1844                         obd, 0, NULL, &llog_lvfs_ops);
1845         if (rc)
1846                 RETURN(rc);
1847
1848         if (mds->mds_profile) {
1849                 struct obd_run_ctxt saved;
1850                 struct lustre_profile *lprof;
1851                 struct config_llog_instance cfg;
1852
1853                 cfg.cfg_instance = NULL;
1854                 cfg.cfg_uuid = mds->mds_lov_uuid;
1855                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1856                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1857                 rc = class_config_parse_llog(ctxt, mds->mds_profile, &cfg);
1858                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1859                 if (rc)
1860                         GOTO(err_llog, rc);
1861
1862                 lprof = class_get_profile(mds->mds_profile);
1863                 if (lprof == NULL) {
1864                         CERROR("No profile found: %s\n", mds->mds_profile);
1865                         GOTO(err_cleanup, rc = -ENOENT);
1866                 }
1867                 rc = mds_lov_connect(obd, lprof->lp_osc);
1868                 if (rc)
1869                         GOTO(err_cleanup, rc);
1870
1871                 rc = mds_lmv_postsetup(obd);
1872                 if (rc)
1873                         GOTO(err_cleanup, rc);
1874         }
1875
1876         RETURN(rc);
1877
1878 err_cleanup:
1879         mds_lov_clean(obd);
1880 err_llog:
1881         ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1882         llog_cleanup(ctxt);
1883         RETURN(rc);
1884 }
1885
1886 static int mds_postrecov(struct obd_device *obd)
1887
1888 {
1889         struct llog_ctxt *ctxt;
1890         int rc, rc2;
1891         ENTRY;
1892
1893         LASSERT(!obd->obd_recovering);
1894         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
1895         LASSERT(ctxt != NULL);
1896
1897         rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
1898                           NULL, NULL, NULL);
1899         if (rc != 0) {
1900                 CERROR("faild at llog_origin_connect: %d\n", rc);
1901         }
1902
1903         rc = mds_cleanup_orphans(obd);
1904
1905         rc2 = mds_lov_set_nextid(obd);
1906         if (rc2 == 0)
1907                 rc2 = rc;
1908         RETURN(rc2);
1909 }
1910
1911 int mds_lov_clean(struct obd_device *obd)
1912 {
1913         struct mds_obd *mds = &obd->u.mds;
1914         struct llog_ctxt *ctxt;
1915
1916         if (mds->mds_profile) {
1917                 char * cln_prof;
1918                 struct config_llog_instance cfg;
1919                 struct obd_run_ctxt saved;
1920                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1921
1922                 OBD_ALLOC(cln_prof, len);
1923                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1924
1925                 cfg.cfg_instance = NULL;
1926                 cfg.cfg_uuid = mds->mds_lov_uuid;
1927
1928                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1929                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1930                 class_config_parse_llog(ctxt, cln_prof, &cfg);
1931                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1932
1933                 OBD_FREE(cln_prof, len);
1934                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1935                 mds->mds_profile = NULL;
1936         }
1937         RETURN(0);
1938 }
1939
1940 int mds_lmv_clean(struct obd_device *obd)
1941 {
1942         struct mds_obd *mds = &obd->u.mds;
1943
1944         if (mds->mds_lmv_name) {
1945                 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
1946                 mds->mds_lmv_name = NULL;
1947         }
1948         RETURN(0);
1949 }
1950
1951 static int mds_precleanup(struct obd_device *obd, int flags)
1952 {
1953         int rc = 0;
1954         ENTRY;
1955
1956         mds_lmv_clean(obd);
1957         mds_lov_disconnect(obd, flags);
1958         mds_lov_clean(obd);
1959         llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
1960         RETURN(rc);
1961 }
1962
1963 static int mds_cleanup(struct obd_device *obd, int flags)
1964 {
1965         struct mds_obd *mds = &obd->u.mds;
1966         ENTRY;
1967
1968         if (mds->mds_sb == NULL)
1969                 RETURN(0);
1970
1971         mds_update_server_data(obd, 1);
1972         if (mds->mds_lov_objids != NULL) {
1973                 OBD_FREE(mds->mds_lov_objids,
1974                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1975         }
1976         mds_fs_cleanup(obd, flags);
1977
1978         unlock_kernel();
1979
1980         /* 2 seems normal on mds, (may_umount() also expects 2
1981           fwiw), but we only see 1 at this point in obdfilter. */
1982         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1983                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1984                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1985
1986         mntput(mds->mds_vfsmnt);
1987
1988         mds->mds_sb = 0;
1989
1990         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1991
1992         spin_lock_bh(&obd->obd_processing_task_lock);
1993         if (obd->obd_recovering) {
1994                 target_cancel_recovery_timer(obd);
1995                 obd->obd_recovering = 0;
1996         }
1997         spin_unlock_bh(&obd->obd_processing_task_lock);
1998
1999         lock_kernel();
2000         dev_clear_rdonly(2);
2001         fsfilt_put_ops(obd->obd_fsops);
2002
2003         RETURN(0);
2004 }
2005
2006 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2007                                         struct ldlm_lock *new_lock,
2008                                         struct lustre_handle *lockh)
2009 {
2010         struct obd_export *exp = req->rq_export;
2011         struct obd_device *obd = exp->exp_obd;
2012         struct ldlm_request *dlmreq =
2013                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2014         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2015         struct list_head *iter;
2016
2017         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2018                 return;
2019
2020         l_lock(&obd->obd_namespace->ns_lock);
2021         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2022                 struct ldlm_lock *lock;
2023                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2024                 if (lock == new_lock)
2025                         continue;
2026                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2027                         lockh->cookie = lock->l_handle.h_cookie;
2028                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2029                                   lockh->cookie);
2030                         l_unlock(&obd->obd_namespace->ns_lock);
2031                         return;
2032                 }
2033         }
2034         l_unlock(&obd->obd_namespace->ns_lock);
2035
2036         /* This remote handle isn't enqueued, so we never received or
2037          * processed this request.  Clear MSG_RESENT, because it can
2038          * be handled like any normal request now. */
2039
2040         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2041
2042         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2043                   remote_hdl.cookie);
2044 }
2045
2046 int intent_disposition(struct ldlm_reply *rep, int flag)
2047 {
2048         if (!rep)
2049                 return 0;
2050         return (rep->lock_policy_res1 & flag);
2051 }
2052
2053 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2054 {
2055         if (!rep)
2056                 return;
2057         rep->lock_policy_res1 |= flag;
2058 }
2059
2060 static int mds_intent_policy(struct ldlm_namespace *ns,
2061                              struct ldlm_lock **lockp, void *req_cookie,
2062                              ldlm_mode_t mode, int flags, void *data)
2063 {
2064         struct ptlrpc_request *req = req_cookie;
2065         struct ldlm_lock *lock = *lockp;
2066         struct ldlm_intent *it;
2067         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2068         struct ldlm_reply *rep;
2069         struct lustre_handle lockh = { 0 };
2070         struct ldlm_lock *new_lock;
2071         int getattr_part = MDS_INODELOCK_UPDATE;
2072         int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2073                                           sizeof(struct mds_body),
2074                                           mds->mds_max_mdsize,
2075                                           mds->mds_max_cookiesize};
2076         ENTRY;
2077
2078         LASSERT(req != NULL);
2079
2080         if (req->rq_reqmsg->bufcount <= 1) {
2081                 /* No intent was provided */
2082                 int size = sizeof(struct ldlm_reply);
2083                 rc = lustre_pack_reply(req, 1, &size, NULL);
2084                 LASSERT(rc == 0);
2085                 RETURN(0);
2086         }
2087
2088         it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2089         if (it == NULL) {
2090                 CERROR("Intent missing\n");
2091                 RETURN(req->rq_status = -EFAULT);
2092         }
2093
2094         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2095
2096         rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
2097                                NULL);
2098         if (rc)
2099                 RETURN(req->rq_status = rc);
2100
2101         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2102         intent_set_disposition(rep, DISP_IT_EXECD);
2103
2104         fixup_handle_for_resent_req(req, lock, &lockh);
2105
2106         /* execute policy */
2107         switch ((long)it->opc) {
2108         case IT_OPEN:
2109         case IT_CREAT|IT_OPEN:
2110                 /* XXX swab here to assert that an mds_open reint
2111                  * packet is following */
2112                 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2113 #if 0
2114                 /* We abort the lock if the lookup was negative and
2115                  * we did not make it to the OPEN portion */
2116                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2117                         RETURN(ELDLM_LOCK_ABORTED);
2118                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2119                     !intent_disposition(rep, DISP_OPEN_OPEN))
2120 #endif
2121                         if (rep->lock_policy_res2) {
2122                                 /* mds_open returns ENOLCK where it
2123                                  * should return zero, but it has no
2124                                  * lock to return */
2125                                 if (rep->lock_policy_res2 == ENOLCK)
2126                                         rep->lock_policy_res2 = 0;
2127                                 RETURN(ELDLM_LOCK_ABORTED);
2128                         }
2129                 break;
2130         case IT_LOOKUP:
2131                 getattr_part = MDS_INODELOCK_LOOKUP;
2132         case IT_GETATTR:
2133                 getattr_part |= MDS_INODELOCK_LOOKUP;
2134         case IT_READDIR:
2135                 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2136                                                          getattr_part);
2137                 /* FIXME: LDLM can set req->rq_status. MDS sets
2138                    policy_res{1,2} with disposition and status.
2139                    - replay: returns 0 & req->status is old status
2140                    - otherwise: returns req->status */
2141                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2142                         rep->lock_policy_res2 = 0;
2143                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2144                     rep->lock_policy_res2)
2145                         RETURN(ELDLM_LOCK_ABORTED);
2146                 if (req->rq_status != 0) {
2147                         LBUG();
2148                         rep->lock_policy_res2 = req->rq_status;
2149                         RETURN(ELDLM_LOCK_ABORTED);
2150                 }
2151                 break;
2152         default:
2153                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2154                 LBUG();
2155         }
2156
2157         /* By this point, whatever function we called above must have either
2158          * filled in 'lockh', been an intent replay, or returned an error.  We
2159          * want to allow replayed RPCs to not get a lock, since we would just
2160          * drop it below anyways because lock replay is done separately by the
2161          * client afterwards.  For regular RPCs we want to give the new lock to
2162          * the client instead of whatever lock it was about to get. */
2163         new_lock = ldlm_handle2lock(&lockh);
2164         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2165                 RETURN(0);
2166
2167         LASSERT(new_lock != NULL);
2168
2169         /* If we've already given this lock to a client once, then we should
2170          * have no readers or writers.  Otherwise, we should have one reader
2171          * _or_ writer ref (which will be zeroed below) before returning the
2172          * lock to a client. */
2173         if (new_lock->l_export == req->rq_export) {
2174                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2175         } else {
2176                 LASSERT(new_lock->l_export == NULL);
2177                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2178         }
2179
2180         *lockp = new_lock;
2181
2182         if (new_lock->l_export == req->rq_export) {
2183                 /* Already gave this to the client, which means that we
2184                  * reconstructed a reply. */
2185                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2186                         MSG_RESENT);
2187                 RETURN(ELDLM_LOCK_REPLACED);
2188         }
2189
2190         /* Fixup the lock to be given to the client */
2191         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2192         new_lock->l_readers = 0;
2193         new_lock->l_writers = 0;
2194
2195         new_lock->l_export = class_export_get(req->rq_export);
2196         list_add(&new_lock->l_export_chain,
2197                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2198
2199         new_lock->l_blocking_ast = lock->l_blocking_ast;
2200         new_lock->l_completion_ast = lock->l_completion_ast;
2201
2202         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2203                sizeof(lock->l_remote_handle));
2204
2205         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2206
2207         LDLM_LOCK_PUT(new_lock);
2208         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2209
2210         RETURN(ELDLM_LOCK_REPLACED);
2211 }
2212
2213 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2214 {
2215         struct lprocfs_static_vars lvars;
2216
2217         lprocfs_init_multi_vars(0, &lvars);
2218         return lprocfs_obd_attach(dev, lvars.obd_vars);
2219 }
2220
2221 int mds_detach(struct obd_device *dev)
2222 {
2223         return lprocfs_obd_detach(dev);
2224 }
2225
2226 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2227 {
2228         struct lprocfs_static_vars lvars;
2229
2230         lprocfs_init_multi_vars(1, &lvars);
2231         return lprocfs_obd_attach(dev, lvars.obd_vars);
2232 }
2233
2234 int mdt_detach(struct obd_device *dev)
2235 {
2236         return lprocfs_obd_detach(dev);
2237 }
2238
2239 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
2240 {
2241         struct mds_obd *mds = &obddev->u.mds;
2242         int rc = 0;
2243         ENTRY;
2244
2245         mds->mds_service =
2246                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2247                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2248                                 mds_handle, "mds",
2249                                 obddev->obd_proc_entry);
2250
2251         if (!mds->mds_service) {
2252                 CERROR("failed to start service\n");
2253                 RETURN(rc = -ENOMEM);
2254         }
2255
2256         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
2257                                     "ll_mdt");
2258         if (rc)
2259                 GOTO(err_thread, rc);
2260
2261         mds->mds_setattr_service =
2262                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2263                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2264                                 mds_handle, "mds_setattr",
2265                                 obddev->obd_proc_entry);
2266         if (!mds->mds_setattr_service) {
2267                 CERROR("failed to start getattr service\n");
2268                 GOTO(err_thread, rc = -ENOMEM);
2269         }
2270
2271         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
2272                                  MDT_NUM_THREADS, "ll_mdt_attr");
2273         if (rc)
2274                 GOTO(err_thread2, rc);
2275
2276         mds->mds_readpage_service =
2277                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2278                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2279                                 mds_handle, "mds_readpage",
2280                                 obddev->obd_proc_entry);
2281         if (!mds->mds_readpage_service) {
2282                 CERROR("failed to start readpage service\n");
2283                 GOTO(err_thread2, rc = -ENOMEM);
2284         }
2285
2286         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
2287                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
2288
2289         if (rc)
2290                 GOTO(err_thread3, rc);
2291
2292         RETURN(0);
2293
2294 err_thread3:
2295         ptlrpc_unregister_service(mds->mds_readpage_service);
2296 err_thread2:
2297         ptlrpc_unregister_service(mds->mds_setattr_service);
2298 err_thread:
2299         ptlrpc_unregister_service(mds->mds_service);
2300         return rc;
2301 }
2302
2303
2304 static int mdt_cleanup(struct obd_device *obddev, int flags)
2305 {
2306         struct mds_obd *mds = &obddev->u.mds;
2307         ENTRY;
2308
2309         ptlrpc_stop_all_threads(mds->mds_readpage_service);
2310         ptlrpc_unregister_service(mds->mds_readpage_service);
2311
2312         ptlrpc_stop_all_threads(mds->mds_setattr_service);
2313         ptlrpc_unregister_service(mds->mds_setattr_service);
2314
2315         ptlrpc_stop_all_threads(mds->mds_service);
2316         ptlrpc_unregister_service(mds->mds_service);
2317
2318         RETURN(0);
2319 }
2320
2321 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
2322 {
2323         struct obd_device *obd = data;
2324         struct ll_fid fid;
2325         fid.id = id;
2326         fid.generation = gen;
2327         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2328 }
2329
2330 struct lvfs_callback_ops mds_lvfs_ops = {
2331         l_fid2dentry:     mds_lvfs_fid2dentry,
2332 };
2333
2334 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2335                 int objcount, struct obd_ioobj *obj,
2336                 int niocount, struct niobuf_remote *nb,
2337                 struct niobuf_local *res,
2338                 struct obd_trans_info *oti);
2339 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2340                  int objcount, struct obd_ioobj *obj, int niocount,
2341                  struct niobuf_local *res, struct obd_trans_info *oti,
2342                  int rc);
2343
2344 /* use obd ops to offer management infrastructure */
2345 static struct obd_ops mds_obd_ops = {
2346         o_owner:       THIS_MODULE,
2347         o_attach:      mds_attach,
2348         o_detach:      mds_detach,
2349         o_connect:     mds_connect,
2350         o_init_export:  mds_init_export,
2351         o_destroy_export:  mds_destroy_export,
2352         o_disconnect:  mds_disconnect,
2353         o_setup:       mds_setup,
2354         o_precleanup:  mds_precleanup,
2355         o_cleanup:     mds_cleanup,
2356         o_postrecov:   mds_postrecov,
2357         o_statfs:      mds_obd_statfs,
2358         o_iocontrol:   mds_iocontrol,
2359         o_create:      mds_obd_create,
2360         o_destroy:     mds_obd_destroy,
2361         o_llog_init:   mds_llog_init,
2362         o_llog_finish: mds_llog_finish,
2363         o_notify:      mds_notify,
2364         o_get_info:    mds_get_info,
2365         o_set_info:    mds_set_info,
2366         o_preprw:      mds_preprw,
2367         o_commitrw:    mds_commitrw,
2368 };
2369
2370 static struct obd_ops mdt_obd_ops = {
2371         o_owner:       THIS_MODULE,
2372         o_attach:      mdt_attach,
2373         o_detach:      mdt_detach,
2374         o_setup:       mdt_setup,
2375         o_cleanup:     mdt_cleanup,
2376 };
2377
2378 static int __init mds_init(void)
2379 {
2380         struct lprocfs_static_vars lvars;
2381
2382         lprocfs_init_multi_vars(0, &lvars);
2383         class_register_type(&mds_obd_ops, NULL, lvars.module_vars, LUSTRE_MDS_NAME);
2384         lprocfs_init_multi_vars(1, &lvars);
2385         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars, LUSTRE_MDT_NAME);
2386
2387         return 0;
2388 }
2389
2390 static void /*__exit*/ mds_exit(void)
2391 {
2392         class_unregister_type(LUSTRE_MDS_NAME);
2393         class_unregister_type(LUSTRE_MDT_NAME);
2394 }
2395
2396 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2397 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2398 MODULE_LICENSE("GPL");
2399
2400 module_init(mds_init);
2401 module_exit(mds_exit);