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) 2001-2005 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 the Lustre file system, http://www.lustre.org
  14  *   Lustre is a trademark of Cluster File Systems, Inc.
  15  *
  16  *   You may have signed or agreed to another license before downloading
  17  *   this software.  If so, you are bound by the terms and conditions
  18  *   of that agreement, and the following does not apply to you.  See the
  19  *   LICENSE file included with this distribution for more information.
  20  *
  21  *   If you did not agree to a different license, then this copy of Lustre
  22  *   is open source software; you can redistribute it and/or modify it
  23  *   under the terms of version 2 of the GNU General Public License as
  24  *   published by the Free Software Foundation.
  25  *
  26  *   In either case, Lustre is distributed in the hope that it will be
  27  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  28  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  29  *   license text for more details.
  30  */
  31
  32 #ifndef EXPORT_SYMTAB
  33 # define EXPORT_SYMTAB
  34 #endif
  35 #define DEBUG_SUBSYSTEM S_MDS
  36
  37 #include <lustre_mds.h>
  38 #include <linux/module.h>
  39 #include <linux/init.h>
  40 #include <linux/random.h>
  41 #include <linux/fs.h>
  42 #include <linux/jbd.h>
  43 #include <linux/ext3_fs.h>
  44 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  45 # include <linux/smp_lock.h>
  46 # include <linux/buffer_head.h>
  47 # include <linux/workqueue.h>
  48 # include <linux/mount.h>
  49 #else
  50 # include <linux/locks.h>
  51 #endif
  52
  53 #include <obd_class.h>
  54 #include <lustre_dlm.h>
  55 #include <obd_lov.h>
  56 #include <lustre_fsfilt.h>
  57 #include <lprocfs_status.h>
  58 #include <lustre_commit_confd.h>
  59 #include <lustre_quota.h>
  60 #include <lustre_disk.h>
  61 #include <lustre_param.h>
  62
  63 #include "mds_internal.h"
  64
  65 int mds_num_threads;
  66 CFS_MODULE_PARM(mds_num_threads, "i", int, 0444,
  67                 "number of MDS service threads to start");
  68
  69 static int mds_intent_policy(struct ldlm_namespace *ns,
  70                              struct ldlm_lock **lockp, void *req_cookie,
  71                              ldlm_mode_t mode, int flags, void *data);
  72 static int mds_postsetup(struct obd_device *obd);
  73 static int mds_cleanup(struct obd_device *obd);
  74
  75 /* Assumes caller has already pushed into the kernel filesystem context */
  76 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  77                         loff_t offset, int count)
  78 {
  79         struct ptlrpc_bulk_desc *desc;
  80         struct l_wait_info lwi;
  81         struct page **pages;
  82         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  83         ENTRY;
  84
  85         LASSERT((offset & ~CFS_PAGE_MASK) == 0); /* I'm dubious about this */
  86
  87         npages = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
  88         OBD_ALLOC(pages, sizeof(*pages) * npages);
  89         if (!pages)
  90                 GOTO(out, rc = -ENOMEM);
  91
  92         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  93                                     MDS_BULK_PORTAL);
  94         if (desc == NULL)
  95                 GOTO(out_free, rc = -ENOMEM);
  96
  97         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  98                 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
  99
 100                 pages[i] = alloc_pages(GFP_KERNEL, 0);
 101                 if (pages[i] == NULL)
 102                         GOTO(cleanup_buf, rc = -ENOMEM);
 103
 104                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
 105         }
 106
 107         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 108                 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
 109                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 110                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 111                        file->f_dentry->d_inode->i_size);
 112
 113                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 114                                      kmap(pages[i]), tmpsize, &offset);
 115                 kunmap(pages[i]);
 116
 117                 if (rc != tmpsize)
 118                         GOTO(cleanup_buf, rc = -EIO);
 119         }
 120
 121         LASSERT(desc->bd_nob == count);
 122
 123         rc = ptlrpc_start_bulk_transfer(desc);
 124         if (rc)
 125                 GOTO(cleanup_buf, rc);
 126
 127         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 128                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 129                        OBD_FAIL_MDS_SENDPAGE, rc);
 130                 GOTO(abort_bulk, rc);
 131         }
 132
 133         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 134         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 135         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 136
 137         if (rc == 0) {
 138                 if (desc->bd_success &&
 139                     desc->bd_nob_transferred == count)
 140                         GOTO(cleanup_buf, rc);
 141
 142                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 143         }
 144
 145         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 146                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 147                   desc->bd_nob_transferred, count,
 148                   req->rq_export->exp_client_uuid.uuid,
 149                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 150
 151         class_fail_export(req->rq_export);
 152
 153         EXIT;
 154  abort_bulk:
 155         ptlrpc_abort_bulk (desc);
 156  cleanup_buf:
 157         for (i = 0; i < npages; i++)
 158                 if (pages[i])
 159                         __free_pages(pages[i], 0);
 160
 161         ptlrpc_free_bulk(desc);
 162  out_free:
 163         OBD_FREE(pages, sizeof(*pages) * npages);
 164  out:
 165         return rc;
 166 }
 167
 168 /* only valid locked dentries or errors should be returned */
 169 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 170                                      struct vfsmount **mnt, int lock_mode,
 171                                      struct lustre_handle *lockh,
 172                                      char *name, int namelen, __u64 lockpart)
 173 {
 174         struct mds_obd *mds = &obd->u.mds;
 175         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 176         struct ldlm_res_id res_id = { .name = {0} };
 177         int flags = LDLM_FL_ATOMIC_CB, rc;
 178         ldlm_policy_data_t policy = { .l_inodebits = { lockpart} };
 179         ENTRY;
 180
 181         if (IS_ERR(de))
 182                 RETURN(de);
 183
 184         res_id.name[0] = de->d_inode->i_ino;
 185         res_id.name[1] = de->d_inode->i_generation;
 186         rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id,
 187                                     LDLM_IBITS, &policy, lock_mode, &flags,
 188                                     ldlm_blocking_ast, ldlm_completion_ast,
 189                                     NULL, NULL, 0, NULL, lockh);
 190         if (rc != ELDLM_OK) {
 191                 l_dput(de);
 192                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 193         }
 194
 195         RETURN(retval);
 196 }
 197
 198 /* Look up an entry by inode number. */
 199 /* this function ONLY returns valid dget'd dentries with an initialized inode
 200    or errors */
 201 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 202                               struct vfsmount **mnt)
 203 {
 204         char fid_name[32];
 205         unsigned long ino = fid->id;
 206         __u32 generation = fid->generation;
 207         struct inode *inode;
 208         struct dentry *result;
 209
 210         if (ino == 0)
 211                 RETURN(ERR_PTR(-ESTALE));
 212
 213         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 214
 215         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 216                ino, generation, mds->mds_obt.obt_sb);
 217
 218         /* under ext3 this is neither supposed to return bad inodes
 219            nor NULL inodes. */
 220         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 221         if (IS_ERR(result))
 222                 RETURN(result);
 223
 224         inode = result->d_inode;
 225         if (!inode)
 226                 RETURN(ERR_PTR(-ENOENT));
 227
 228         if (inode->i_generation == 0 || inode->i_nlink == 0) {
 229                 LCONSOLE_WARN("Found inode with zero generation or link -- this"
 230                               " may indicate disk corruption (inode: %lu/%u, "
 231                               "link %lu, count %d)\n", inode->i_ino,
 232                               inode->i_generation,(unsigned long)inode->i_nlink,
 233                               atomic_read(&inode->i_count));
 234                 dput(result);
 235                 RETURN(ERR_PTR(-ENOENT));
 236         }
 237
 238         if (generation && inode->i_generation != generation) {
 239                 /* we didn't find the right inode.. */
 240                 CDEBUG(D_INODE, "found wrong generation: inode %lu, link: %lu, "
 241                        "count: %d, generation %u/%u\n", inode->i_ino,
 242                        (unsigned long)inode->i_nlink,
 243                        atomic_read(&inode->i_count), inode->i_generation,
 244                        generation);
 245                 dput(result);
 246                 RETURN(ERR_PTR(-ENOENT));
 247         }
 248
 249         if (mnt) {
 250                 *mnt = mds->mds_vfsmnt;
 251                 mntget(*mnt);
 252         }
 253
 254         RETURN(result);
 255 }
 256
 257 static int mds_connect_internal(struct obd_export *exp,
 258                                 struct obd_connect_data *data)
 259 {
 260         struct obd_device *obd = exp->exp_obd;
 261         if (data != NULL) {
 262                 data->ocd_connect_flags &= MDS_CONNECT_SUPPORTED;
 263                 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
 264
 265                 /* If no known bits (which should not happen, probably,
 266                    as everybody should support LOOKUP and UPDATE bits at least)
 267                    revert to compat mode with plain locks. */
 268                 if (!data->ocd_ibits_known &&
 269                     data->ocd_connect_flags & OBD_CONNECT_IBITS)
 270                         data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
 271
 272                 if (!obd->u.mds.mds_fl_acl)
 273                         data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
 274
 275                 if (!obd->u.mds.mds_fl_user_xattr)
 276                         data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
 277
 278                 exp->exp_connect_flags = data->ocd_connect_flags;
 279                 data->ocd_version = LUSTRE_VERSION_CODE;
 280                 exp->exp_mds_data.med_ibits_known = data->ocd_ibits_known;
 281         }
 282
 283         if (obd->u.mds.mds_fl_acl &&
 284             ((exp->exp_connect_flags & OBD_CONNECT_ACL) == 0)) {
 285                 CWARN("%s: MDS requires ACL support but client does not\n",
 286                       obd->obd_name);
 287                 return -EBADE;
 288         }
 289         return 0;
 290 }
 291
 292 static int mds_reconnect(struct obd_export *exp, struct obd_device *obd,
 293                          struct obd_uuid *cluuid,
 294                          struct obd_connect_data *data)
 295 {
 296         int rc;
 297         ENTRY;
 298
 299         if (exp == NULL || obd == NULL || cluuid == NULL)
 300                 RETURN(-EINVAL);
 301
 302         rc = mds_connect_internal(exp, data);
 303
 304         RETURN(rc);
 305 }
 306
 307 /* Establish a connection to the MDS.
 308  *
 309  * This will set up an export structure for the client to hold state data
 310  * about that client, like open files, the last operation number it did
 311  * on the server, etc.
 312  */
 313 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 314                        struct obd_uuid *cluuid, struct obd_connect_data *data)
 315 {
 316         struct obd_export *exp;
 317         struct mds_export_data *med;
 318         struct mds_client_data *mcd = NULL;
 319         int rc, abort_recovery;
 320         ENTRY;
 321
 322         if (!conn || !obd || !cluuid)
 323                 RETURN(-EINVAL);
 324
 325         /* Check for aborted recovery. */
 326         spin_lock_bh(&obd->obd_processing_task_lock);
 327         abort_recovery = obd->obd_abort_recovery;
 328         spin_unlock_bh(&obd->obd_processing_task_lock);
 329         if (abort_recovery)
 330                 target_abort_recovery(obd);
 331
 332         /* XXX There is a small race between checking the list and adding a
 333          * new connection for the same UUID, but the real threat (list
 334          * corruption when multiple different clients connect) is solved.
 335          *
 336          * There is a second race between adding the export to the list,
 337          * and filling in the client data below.  Hence skipping the case
 338          * of NULL mcd above.  We should already be controlling multiple
 339          * connects at the client, and we can't hold the spinlock over
 340          * memory allocations without risk of deadlocking.
 341          */
 342         rc = class_connect(conn, obd, cluuid);
 343         if (rc)
 344                 RETURN(rc);
 345         exp = class_conn2export(conn);
 346         LASSERT(exp);
 347         med = &exp->exp_mds_data;
 348
 349         rc = mds_connect_internal(exp, data);
 350         if (rc)
 351                 GOTO(out, rc);
 352
 353         OBD_ALLOC(mcd, sizeof(*mcd));
 354         if (!mcd)
 355                 GOTO(out, rc = -ENOMEM);
 356
 357         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 358         med->med_mcd = mcd;
 359
 360         rc = mds_client_add(obd, exp, -1);
 361         GOTO(out, rc);
 362
 363 out:
 364         if (rc) {
 365                 if (mcd) {
 366                         OBD_FREE(mcd, sizeof(*mcd));
 367                         med->med_mcd = NULL;
 368                 }
 369                 class_disconnect(exp);
 370         } else {
 371                 class_export_put(exp);
 372         }
 373
 374         RETURN(rc);
 375 }
 376
 377 int mds_init_export(struct obd_export *exp)
 378 {
 379         struct mds_export_data *med = &exp->exp_mds_data;
 380
 381         INIT_LIST_HEAD(&med->med_open_head);
 382         spin_lock_init(&med->med_open_lock);
 383         exp->exp_connecting = 1;
 384         RETURN(0);
 385 }
 386
 387 static int mds_destroy_export(struct obd_export *export)
 388 {
 389         struct mds_export_data *med;
 390         struct obd_device *obd = export->exp_obd;
 391         struct mds_obd *mds = &obd->u.mds;
 392         struct lvfs_run_ctxt saved;
 393         struct lov_mds_md *lmm;
 394         struct llog_cookie *logcookies;
 395         int rc = 0;
 396         ENTRY;
 397
 398         med = &export->exp_mds_data;
 399         target_destroy_export(export);
 400
 401         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 402                 RETURN(0);
 403
 404         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 405         /* Close any open files (which may also cause orphan unlinking). */
 406
 407         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 408         if (lmm == NULL) {
 409                 CWARN("%s: allocation failure during cleanup; can not force "
 410                       "close file handles on this service.\n", obd->obd_name);
 411                 GOTO(out, rc = -ENOMEM);
 412         }
 413
 414         OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
 415         if (logcookies == NULL) {
 416                 CWARN("%s: allocation failure during cleanup; can not force "
 417                       "close file handles on this service.\n", obd->obd_name);
 418                 OBD_FREE(lmm, mds->mds_max_mdsize);
 419                 GOTO(out, rc = -ENOMEM);
 420         }
 421
 422         spin_lock(&med->med_open_lock);
 423         while (!list_empty(&med->med_open_head)) {
 424                 struct list_head *tmp = med->med_open_head.next;
 425                 struct mds_file_data *mfd =
 426                         list_entry(tmp, struct mds_file_data, mfd_list);
 427                 int lmm_size = mds->mds_max_mdsize;
 428                 umode_t mode = mfd->mfd_dentry->d_inode->i_mode;
 429                 __u64 valid = 0;
 430
 431                 /* Remove mfd handle so it can't be found again.
 432                  * We are consuming the mfd_list reference here. */
 433                 mds_mfd_unlink(mfd, 0);
 434                 spin_unlock(&med->med_open_lock);
 435
 436                 /* If you change this message, be sure to update
 437                  * replay_single:test_46 */
 438                 CDEBUG(D_INODE|D_IOCTL, "%s: force closing file handle for "
 439                        "%.*s (ino %lu)\n", obd->obd_name,
 440                        mfd->mfd_dentry->d_name.len,mfd->mfd_dentry->d_name.name,
 441                        mfd->mfd_dentry->d_inode->i_ino);
 442
 443                 rc = mds_get_md(obd, mfd->mfd_dentry->d_inode, lmm,&lmm_size,1);
 444                 if (rc < 0)
 445                         CWARN("mds_get_md failure, rc=%d\n", rc);
 446                 else
 447                         valid |= OBD_MD_FLEASIZE;
 448
 449                 /* child orphan sem protects orphan_dec_test and
 450                  * is_orphan race, mds_mfd_close drops it */
 451                 MDS_DOWN_WRITE_ORPHAN_SEM(mfd->mfd_dentry->d_inode);
 452
 453                 rc = mds_mfd_close(NULL, REQ_REC_OFF, obd, mfd,
 454                                    !(export->exp_flags & OBD_OPT_FAILOVER),
 455                                    lmm, lmm_size, logcookies,
 456                                    mds->mds_max_cookiesize,
 457                                    &valid);
 458
 459                 if (rc)
 460                         CDEBUG(D_INODE|D_IOCTL, "Error closing file: %d\n", rc);
 461
 462                 if (valid & OBD_MD_FLCOOKIE) {
 463                         rc = mds_osc_destroy_orphan(obd, mode, lmm,
 464                                                     lmm_size, logcookies, 1);
 465                         if (rc < 0) {
 466                                 CDEBUG(D_INODE, "%s: destroy of orphan failed,"
 467                                        " rc = %d\n", obd->obd_name, rc);
 468                                 rc = 0;
 469                         }
 470                         valid &= ~OBD_MD_FLCOOKIE;
 471                 }
 472
 473                 spin_lock(&med->med_open_lock);
 474         }
 475
 476         OBD_FREE(logcookies, mds->mds_max_cookiesize);
 477         OBD_FREE(lmm, mds->mds_max_mdsize);
 478
 479         spin_unlock(&med->med_open_lock);
 480
 481         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 482         mds_client_free(export);
 483
 484  out:
 485         RETURN(rc);
 486 }
 487
 488 static int mds_disconnect(struct obd_export *exp)
 489 {
 490         int rc;
 491         ENTRY;
 492
 493         LASSERT(exp);
 494         class_export_get(exp);
 495
 496         /* Disconnect early so that clients can't keep using export */
 497         rc = class_disconnect(exp);
 498         if (exp->exp_obd->obd_namespace != NULL)
 499                 ldlm_cancel_locks_for_export(exp);
 500
 501         /* complete all outstanding replies */
 502         spin_lock(&exp->exp_lock);
 503         while (!list_empty(&exp->exp_outstanding_replies)) {
 504                 struct ptlrpc_reply_state *rs =
 505                         list_entry(exp->exp_outstanding_replies.next,
 506                                    struct ptlrpc_reply_state, rs_exp_list);
 507                 struct ptlrpc_service *svc = rs->rs_service;
 508
 509                 spin_lock(&svc->srv_lock);
 510                 list_del_init(&rs->rs_exp_list);
 511                 ptlrpc_schedule_difficult_reply(rs);
 512                 spin_unlock(&svc->srv_lock);
 513         }
 514         spin_unlock(&exp->exp_lock);
 515
 516         class_export_put(exp);
 517         RETURN(rc);
 518 }
 519
 520 static int mds_getstatus(struct ptlrpc_request *req)
 521 {
 522         struct mds_obd *mds = mds_req2mds(req);
 523         struct mds_body *body;
 524         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 525         ENTRY;
 526
 527         rc = lustre_pack_reply(req, 2, size, NULL);
 528         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 529                 CERROR("mds: out of memory for message\n");
 530                 req->rq_status = -ENOMEM;       /* superfluous? */
 531                 RETURN(-ENOMEM);
 532         }
 533
 534         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 535         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 536
 537         /* the last_committed and last_xid fields are filled in for all
 538          * replies already - no need to do so here also.
 539          */
 540         RETURN(0);
 541 }
 542
 543 /* get the LOV EA from @inode and store it into @md.  It can be at most
 544  * @size bytes, and @size is updated with the actual EA size.
 545  * The EA size is also returned on success, and -ve errno on failure.
 546  * If there is no EA then 0 is returned. */
 547 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
 548                int *size, int lock)
 549 {
 550         int rc = 0;
 551         int lmm_size;
 552
 553         if (lock)
 554                 LOCK_INODE_MUTEX(inode);
 555         rc = fsfilt_get_md(obd, inode, md, *size, "lov");
 556
 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                 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
 563
 564                 if (rc == 0) {
 565                         *size = lmm_size;
 566                         rc = lmm_size;
 567                 } else if (rc > 0) {
 568                         *size = rc;
 569                 }
 570         } else {
 571                 *size = 0;
 572         }
 573         if (lock)
 574                 UNLOCK_INODE_MUTEX(inode);
 575
 576         RETURN (rc);
 577 }
 578
 579
 580 /* Call with lock=1 if you want mds_pack_md to take the i_mutex.
 581  * Call with lock=0 if the caller has already taken the i_mutex. */
 582 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 583                 struct mds_body *body, struct inode *inode, int lock)
 584 {
 585         struct mds_obd *mds = &obd->u.mds;
 586         void *lmm;
 587         int lmm_size;
 588         int rc;
 589         ENTRY;
 590
 591         lmm = lustre_msg_buf(msg, offset, 0);
 592         if (lmm == NULL) {
 593                 /* Some problem with getting eadata when I sized the reply
 594                  * buffer... */
 595                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 596                        inode->i_ino);
 597                 RETURN(0);
 598         }
 599         lmm_size = lustre_msg_buflen(msg, offset);
 600
 601         /* I don't really like this, but it is a sanity check on the client
 602          * MD request.  However, if the client doesn't know how much space
 603          * to reserve for the MD, it shouldn't be bad to have too much space.
 604          */
 605         if (lmm_size > mds->mds_max_mdsize) {
 606                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 607                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 608                 // RETURN(-EINVAL);
 609         }
 610
 611         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
 612         if (rc > 0) {
 613                 if (S_ISDIR(inode->i_mode))
 614                         body->valid |= OBD_MD_FLDIREA;
 615                 else
 616                         body->valid |= OBD_MD_FLEASIZE;
 617                 body->eadatasize = lmm_size;
 618                 rc = 0;
 619         }
 620
 621         RETURN(rc);
 622 }
 623
 624 #ifdef CONFIG_FS_POSIX_ACL
 625 static
 626 int mds_pack_posix_acl(struct inode *inode, struct lustre_msg *repmsg,
 627                        struct mds_body *repbody, int repoff)
 628 {
 629         struct dentry de = { .d_inode = inode };
 630         int buflen, rc;
 631         ENTRY;
 632
 633         LASSERT(repbody->aclsize == 0);
 634         LASSERT(lustre_msg_bufcount(repmsg) > repoff);
 635
 636         buflen = lustre_msg_buflen(repmsg, repoff);
 637         if (!buflen)
 638                 GOTO(out, 0);
 639
 640         if (!inode->i_op || !inode->i_op->getxattr)
 641                 GOTO(out, 0);
 642
 643         lock_24kernel();
 644         rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
 645                                    lustre_msg_buf(repmsg, repoff, buflen),
 646                                    buflen);
 647         unlock_24kernel();
 648
 649         if (rc >= 0)
 650                 repbody->aclsize = rc;
 651         else if (rc != -ENODATA) {
 652                 CERROR("buflen %d, get acl: %d\n", buflen, rc);
 653                 RETURN(rc);
 654         }
 655         EXIT;
 656 out:
 657         repbody->valid |= OBD_MD_FLACL;
 658         return 0;
 659 }
 660 #else
 661 #define mds_pack_posix_acl(inode, repmsg, repbody, repoff) 0
 662 #endif
 663
 664 int mds_pack_acl(struct mds_export_data *med, struct inode *inode,
 665                  struct lustre_msg *repmsg, struct mds_body *repbody,
 666                  int repoff)
 667 {
 668         return mds_pack_posix_acl(inode, repmsg, repbody, repoff);
 669 }
 670
 671 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 672                                 struct ptlrpc_request *req,
 673                                 struct mds_body *reqbody, int reply_off)
 674 {
 675         struct mds_body *body;
 676         struct inode *inode = dentry->d_inode;
 677         int rc = 0;
 678         ENTRY;
 679
 680         if (inode == NULL)
 681                 RETURN(-ENOENT);
 682
 683         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 684         LASSERT(body != NULL);                 /* caller prepped reply */
 685
 686         mds_pack_inode2fid(&body->fid1, inode);
 687         body->flags = reqbody->flags; /* copy MDS_BFLAG_EXT_FLAGS if present */
 688         mds_pack_inode2body(body, inode);
 689         reply_off++;
 690
 691         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
 692             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
 693                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off, body,
 694                                  inode, 1);
 695
 696                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 697                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 698                     !(body->valid & OBD_MD_FLDIREA))
 699                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 700                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 701
 702                 lustre_shrink_reply(req, reply_off, body->eadatasize, 0);
 703                 if (body->eadatasize)
 704                         reply_off++;
 705         } else if (S_ISLNK(inode->i_mode) &&
 706                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 707                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off, 0);
 708                 int len;
 709
 710                 LASSERT (symname != NULL);       /* caller prepped reply */
 711                 len = lustre_msg_buflen(req->rq_repmsg, reply_off);
 712
 713                 rc = inode->i_op->readlink(dentry, symname, len);
 714                 if (rc < 0) {
 715                         CERROR("readlink failed: %d\n", rc);
 716                 } else if (rc != len - 1) {
 717                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 718                                 rc, len - 1);
 719                         rc = -EINVAL;
 720                 } else {
 721                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 722                         body->valid |= OBD_MD_LINKNAME;
 723                         body->eadatasize = rc + 1;
 724                         symname[rc] = 0;        /* NULL terminate */
 725                         rc = 0;
 726                 }
 727                 reply_off++;
 728         } else if (reqbody->valid == OBD_MD_FLFLAGS &&
 729                    reqbody->flags & MDS_BFLAG_EXT_FLAGS) {
 730                 int flags;
 731
 732                 /* We only return the full set of flags on ioctl, otherwise we
 733                  * get enough flags from the inode in mds_pack_inode2body(). */
 734                 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_GETFLAGS,
 735                                       (long)&flags);
 736                 if (rc == 0)
 737                         body->flags = flags | MDS_BFLAG_EXT_FLAGS;
 738         }
 739
 740         if (reqbody->valid & OBD_MD_FLMODEASIZE) {
 741                 struct mds_obd *mds = mds_req2mds(req);
 742                 body->max_cookiesize = mds->mds_max_cookiesize;
 743                 body->max_mdsize = mds->mds_max_mdsize;
 744                 body->valid |= OBD_MD_FLMODEASIZE;
 745         }
 746
 747         if (rc)
 748                 RETURN(rc);
 749
 750 #ifdef CONFIG_FS_POSIX_ACL
 751         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
 752             (reqbody->valid & OBD_MD_FLACL)) {
 753                 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
 754                                   inode, req->rq_repmsg,
 755                                   body, reply_off);
 756
 757                 lustre_shrink_reply(req, reply_off, body->aclsize, 0);
 758                 if (body->aclsize)
 759                         reply_off++;
 760         }
 761 #endif
 762
 763         RETURN(rc);
 764 }
 765
 766 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 767                                 int offset)
 768 {
 769         struct mds_obd *mds = mds_req2mds(req);
 770         struct mds_body *body;
 771         int rc, bufcount = 2;
 772         int size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 773         ENTRY;
 774
 775         LASSERT(offset == REQ_REC_OFF); /* non-intent */
 776
 777         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
 778         LASSERT(body != NULL);                 /* checked by caller */
 779         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 780
 781         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 782             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
 783                 LOCK_INODE_MUTEX(inode);
 784                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
 785                                    "lov");
 786                 UNLOCK_INODE_MUTEX(inode);
 787                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 788                        rc, inode->i_ino);
 789                 if (rc < 0) {
 790                         if (rc != -ENODATA) {
 791                                 CERROR("error getting inode %lu MD: rc = %d\n",
 792                                        inode->i_ino, rc);
 793                                 RETURN(rc);
 794                         }
 795                         size[bufcount] = 0;
 796                 } else if (rc > mds->mds_max_mdsize) {
 797                         size[bufcount] = 0;
 798                         CERROR("MD size %d larger than maximum possible %u\n",
 799                                rc, mds->mds_max_mdsize);
 800                 } else {
 801                         size[bufcount] = rc;
 802                 }
 803                 bufcount++;
 804         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 805                 if (inode->i_size + 1 != body->eadatasize)
 806                         CERROR("symlink size: %Lu, reply space: %d\n",
 807                                inode->i_size + 1, body->eadatasize);
 808                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 809                 bufcount++;
 810                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 811                        inode->i_size + 1, body->eadatasize);
 812         }
 813
 814 #ifdef CONFIG_FS_POSIX_ACL
 815         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
 816             (body->valid & OBD_MD_FLACL)) {
 817                 struct dentry de = { .d_inode = inode };
 818
 819                 size[bufcount] = 0;
 820                 if (inode->i_op && inode->i_op->getxattr) {
 821                         lock_24kernel();
 822                         rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
 823                                                    NULL, 0);
 824                         unlock_24kernel();
 825
 826                         if (rc < 0) {
 827                                 if (rc != -ENODATA) {
 828                                         CERROR("got acl size: %d\n", rc);
 829                                         RETURN(rc);
 830                                 }
 831                         } else
 832                                 size[bufcount] = rc;
 833                 }
 834                 bufcount++;
 835         }
 836 #endif
 837
 838         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 839                 CERROR("failed MDS_GETATTR_PACK test\n");
 840                 req->rq_status = -ENOMEM;
 841                 RETURN(-ENOMEM);
 842         }
 843
 844         rc = lustre_pack_reply(req, bufcount, size, NULL);
 845         if (rc) {
 846                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
 847                 req->rq_status = rc;
 848                 RETURN(rc);
 849         }
 850
 851         RETURN(0);
 852 }
 853
 854 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
 855                             int child_part, struct lustre_handle *child_lockh)
 856 {
 857         struct obd_device *obd = req->rq_export->exp_obd;
 858         struct mds_obd *mds = &obd->u.mds;
 859         struct ldlm_reply *rep = NULL;
 860         struct lvfs_run_ctxt saved;
 861         struct mds_body *body;
 862         struct dentry *dparent = NULL, *dchild = NULL;
 863         struct lvfs_ucred uc = {NULL,};
 864         struct lustre_handle parent_lockh;
 865         int namesize;
 866         int rc = 0, cleanup_phase = 0, resent_req = 0;
 867         char *name;
 868         ENTRY;
 869
 870         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 871
 872         /* Swab now, before anyone looks inside the request */
 873         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 874                                   lustre_swab_mds_body);
 875         if (body == NULL) {
 876                 CERROR("Can't swab mds_body\n");
 877                 RETURN(-EFAULT);
 878         }
 879
 880         LASSERT_REQSWAB(req, offset + 1);
 881         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 882         if (name == NULL) {
 883                 CERROR("Can't unpack name\n");
 884                 RETURN(-EFAULT);
 885         }
 886         namesize = lustre_msg_buflen(req->rq_reqmsg, offset + 1);
 887         /* namesize less than 2 means we have empty name, probably came from
 888            revalidate by cfid, so no point in having name to be set */
 889         if (namesize <= 1)
 890                 name = NULL;
 891
 892         rc = mds_init_ucred(&uc, req, offset);
 893         if (rc)
 894                 GOTO(cleanup, rc);
 895
 896         LASSERT(offset == REQ_REC_OFF || offset == DLM_INTENT_REC_OFF);
 897         /* if requests were at offset 2, the getattr reply goes back at 1 */
 898         if (offset == DLM_INTENT_REC_OFF) {
 899                 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 900                                      sizeof(*rep));
 901                 offset = DLM_REPLY_REC_OFF;
 902         }
 903
 904         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 905         cleanup_phase = 1; /* kernel context */
 906         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 907
 908         /* FIXME: handle raw lookup */
 909 #if 0
 910         if (body->valid == OBD_MD_FLID) {
 911                 struct mds_body *mds_reply;
 912                 int size = sizeof(*mds_reply);
 913                 ino_t inum;
 914                 // The user requested ONLY the inode number, so do a raw lookup
 915                 rc = lustre_pack_reply(req, 1, &size, NULL);
 916                 if (rc) {
 917                         CERROR("out of memory\n");
 918                         GOTO(cleanup, rc);
 919                 }
 920
 921                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 922
 923                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 924                                            sizeof(*mds_reply));
 925                 mds_reply->fid1.id = inum;
 926                 mds_reply->valid = OBD_MD_FLID;
 927                 GOTO(cleanup, rc);
 928         }
 929 #endif
 930
 931         if (lustre_handle_is_used(child_lockh)) {
 932                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 933                 resent_req = 1;
 934         }
 935
 936         if (resent_req == 0) {
 937                 if (name) {
 938                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout*2);
 939                         rc = mds_get_parent_child_locked(obd, &obd->u.mds,
 940                                                          &body->fid1,
 941                                                          &parent_lockh,
 942                                                          &dparent, LCK_CR,
 943                                                          MDS_INODELOCK_UPDATE,
 944                                                          name, namesize,
 945                                                          child_lockh, &dchild,
 946                                                          LCK_CR, child_part);
 947                 } else {
 948                         /* For revalidate by fid we always take UPDATE lock */
 949                         dchild = mds_fid2locked_dentry(obd, &body->fid2, NULL,
 950                                                        LCK_CR, child_lockh,
 951                                                        NULL, 0, child_part);
 952                         LASSERT(dchild);
 953                         if (IS_ERR(dchild))
 954                                 rc = PTR_ERR(dchild);
 955                 }
 956                 if (rc)
 957                         GOTO(cleanup, rc);
 958         } else {
 959                 struct ldlm_lock *granted_lock;
 960                 struct ll_fid child_fid;
 961                 struct ldlm_resource *res;
 962                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 963                 granted_lock = ldlm_handle2lock(child_lockh);
 964                 LASSERTF(granted_lock != NULL, LPU64"/%u lockh "LPX64"\n",
 965                          body->fid1.id, body->fid1.generation,
 966                          child_lockh->cookie);
 967
 968
 969                 res = granted_lock->l_resource;
 970                 child_fid.id = res->lr_name.name[0];
 971                 child_fid.generation = res->lr_name.name[1];
 972                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 973                 LASSERT(!IS_ERR(dchild));
 974                 LDLM_LOCK_PUT(granted_lock);
 975         }
 976
 977         cleanup_phase = 2; /* dchild, dparent, locks */
 978
 979         if (dchild->d_inode == NULL) {
 980                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 981                 /* in the intent case, the policy clears this error:
 982                    the disposition is enough */
 983                 GOTO(cleanup, rc = -ENOENT);
 984         } else {
 985                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 986         }
 987
 988         if (req->rq_repmsg == NULL) {
 989                 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 990                 if (rc != 0) {
 991                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 992                         GOTO (cleanup, rc);
 993                 }
 994         }
 995
 996         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 997         GOTO(cleanup, rc); /* returns the lock to the client */
 998
 999  cleanup:
1000         switch (cleanup_phase) {
1001         case 2:
1002                 if (resent_req == 0) {
1003                         if (rc && dchild->d_inode)
1004                                 ldlm_lock_decref(child_lockh, LCK_CR);
1005                         if (name) {
1006                                 ldlm_lock_decref(&parent_lockh, LCK_CR);
1007                                 l_dput(dparent);
1008                         }
1009                 }
1010                 l_dput(dchild);
1011         case 1:
1012                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1013         default:
1014                 mds_exit_ucred(&uc, mds);
1015                 if (req->rq_reply_state == NULL) {
1016                         req->rq_status = rc;
1017                         lustre_pack_reply(req, 1, NULL, NULL);
1018                 }
1019         }
1020         return rc;
1021 }
1022
1023 static int mds_getattr(struct ptlrpc_request *req, int offset)
1024 {
1025         struct mds_obd *mds = mds_req2mds(req);
1026         struct obd_device *obd = req->rq_export->exp_obd;
1027         struct lvfs_run_ctxt saved;
1028         struct dentry *de;
1029         struct mds_body *body;
1030         struct lvfs_ucred uc = { NULL, };
1031         int rc = 0;
1032         ENTRY;
1033
1034         OBD_COUNTER_INCREMENT(obd, getattr);
1035
1036         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1037                                   lustre_swab_mds_body);
1038         if (body == NULL)
1039                 RETURN(-EFAULT);
1040
1041         rc = mds_init_ucred(&uc, req, offset);
1042         if (rc)
1043                 GOTO(out_ucred, rc);
1044
1045         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1046         de = mds_fid2dentry(mds, &body->fid1, NULL);
1047         if (IS_ERR(de)) {
1048                 rc = req->rq_status = PTR_ERR(de);
1049                 GOTO(out_pop, rc);
1050         }
1051
1052         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1053         if (rc != 0) {
1054                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1055                 GOTO(out_pop, rc);
1056         }
1057
1058         req->rq_status = mds_getattr_internal(obd, de, req, body,REPLY_REC_OFF);
1059
1060         l_dput(de);
1061         GOTO(out_pop, rc);
1062 out_pop:
1063         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1064 out_ucred:
1065         if (req->rq_reply_state == NULL) {
1066                 req->rq_status = rc;
1067                 lustre_pack_reply(req, 1, NULL, NULL);
1068         }
1069         mds_exit_ucred(&uc, mds);
1070         return rc;
1071 }
1072
1073 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1074                           __u64 max_age)
1075 {
1076         int rc;
1077
1078         spin_lock(&obd->obd_osfs_lock);
1079         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
1080         if (rc == 0)
1081                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1082         spin_unlock(&obd->obd_osfs_lock);
1083
1084         return rc;
1085 }
1086
1087 static int mds_statfs(struct ptlrpc_request *req)
1088 {
1089         struct obd_device *obd = req->rq_export->exp_obd;
1090         int rc, size[2] = { sizeof(struct ptlrpc_body),
1091                             sizeof(struct obd_statfs) };
1092         ENTRY;
1093
1094         /* This will trigger a watchdog timeout */
1095         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1096                          (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1097         OBD_COUNTER_INCREMENT(obd, statfs);
1098
1099         rc = lustre_pack_reply(req, 2, size, NULL);
1100         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1101                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1102                 GOTO(out, rc);
1103         }
1104
1105         /* We call this so that we can cache a bit - 1 jiffie worth */
1106         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1107                                                 size[REPLY_REC_OFF]),
1108                             cfs_time_current_64() - HZ);
1109         if (rc) {
1110                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1111                 GOTO(out, rc);
1112         }
1113
1114         EXIT;
1115 out:
1116         req->rq_status = rc;
1117         return 0;
1118 }
1119
1120 static int mds_sync(struct ptlrpc_request *req, int offset)
1121 {
1122         struct obd_device *obd = req->rq_export->exp_obd;
1123         struct mds_obd *mds = &obd->u.mds;
1124         struct mds_body *body;
1125         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
1126         ENTRY;
1127
1128         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1129                                   lustre_swab_mds_body);
1130         if (body == NULL)
1131                 GOTO(out, rc = -EFAULT);
1132
1133         rc = lustre_pack_reply(req, 2, size, NULL);
1134         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1135                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1136                 GOTO(out, rc);
1137         }
1138
1139         if (body->fid1.id == 0) {
1140                 /* a fid of zero is taken to mean "sync whole filesystem" */
1141                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
1142                 GOTO(out, rc);
1143         } else {
1144                 struct dentry *de;
1145
1146                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1147                 if (IS_ERR(de))
1148                         GOTO(out, rc = PTR_ERR(de));
1149
1150                 /* The file parameter isn't used for anything */
1151                 if (de->d_inode->i_fop && de->d_inode->i_fop->fsync)
1152                         rc = de->d_inode->i_fop->fsync(NULL, de, 1);
1153                 if (rc == 0) {
1154                         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1155                                               sizeof(*body));
1156                         mds_pack_inode2fid(&body->fid1, de->d_inode);
1157                         mds_pack_inode2body(body, de->d_inode);
1158                 }
1159
1160                 l_dput(de);
1161                 GOTO(out, rc);
1162         }
1163 out:
1164         req->rq_status = rc;
1165         return 0;
1166 }
1167
1168 /* mds_readpage does not take a DLM lock on the inode, because the client must
1169  * already have a PR lock.
1170  *
1171  * If we were to take another one here, a deadlock will result, if another
1172  * thread is already waiting for a PW lock. */
1173 static int mds_readpage(struct ptlrpc_request *req, int offset)
1174 {
1175         struct obd_device *obd = req->rq_export->exp_obd;
1176         struct mds_obd *mds = &obd->u.mds;
1177         struct vfsmount *mnt;
1178         struct dentry *de;
1179         struct file *file;
1180         struct mds_body *body, *repbody;
1181         struct lvfs_run_ctxt saved;
1182         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
1183         struct lvfs_ucred uc = {NULL,};
1184         ENTRY;
1185
1186         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1187                 RETURN(-ENOMEM);
1188
1189         rc = lustre_pack_reply(req, 2, size, NULL);
1190         if (rc) {
1191                 CERROR("error packing readpage reply: rc %d\n", rc);
1192                 GOTO(out, rc);
1193         }
1194
1195         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1196                                   lustre_swab_mds_body);
1197         if (body == NULL)
1198                 GOTO (out, rc = -EFAULT);
1199
1200         rc = mds_init_ucred(&uc, req, offset);
1201         if (rc)
1202                 GOTO(out, rc);
1203
1204         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1205         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1206         if (IS_ERR(de))
1207                 GOTO(out_pop, rc = PTR_ERR(de));
1208
1209         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1210
1211         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1212         /* note: in case of an error, dentry_open puts dentry */
1213         if (IS_ERR(file))
1214                 GOTO(out_pop, rc = PTR_ERR(file));
1215
1216         /* body->size is actually the offset -eeb */
1217         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1218                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1219                        body->size, de->d_inode->i_blksize);
1220                 GOTO(out_file, rc = -EFAULT);
1221         }
1222
1223         /* body->nlink is actually the #bytes to read -eeb */
1224         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1225                 CERROR("size %u is not multiple of blocksize %lu\n",
1226                        body->nlink, de->d_inode->i_blksize);
1227                 GOTO(out_file, rc = -EFAULT);
1228         }
1229
1230         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1231                                  sizeof(*repbody));
1232         repbody->size = file->f_dentry->d_inode->i_size;
1233         repbody->valid = OBD_MD_FLSIZE;
1234
1235         /* to make this asynchronous make sure that the handling function
1236            doesn't send a reply when this function completes. Instead a
1237            callback function would send the reply */
1238         /* body->size is actually the offset -eeb */
1239         rc = mds_sendpage(req, file, body->size, body->nlink);
1240
1241 out_file:
1242         filp_close(file, 0);
1243 out_pop:
1244         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1245 out:
1246         mds_exit_ucred(&uc, mds);
1247         req->rq_status = rc;
1248         RETURN(0);
1249 }
1250
1251 int mds_reint(struct ptlrpc_request *req, int offset,
1252               struct lustre_handle *lockh)
1253 {
1254         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1255         int rc;
1256
1257         OBD_ALLOC(rec, sizeof(*rec));
1258         if (rec == NULL)
1259                 RETURN(-ENOMEM);
1260
1261         rc = mds_update_unpack(req, offset, rec);
1262         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1263                 CERROR("invalid record\n");
1264                 GOTO(out, req->rq_status = -EINVAL);
1265         }
1266
1267         /* rc will be used to interrupt a for loop over multiple records */
1268         rc = mds_reint_rec(rec, offset, req, lockh);
1269  out:
1270         OBD_FREE(rec, sizeof(*rec));
1271         return rc;
1272 }
1273
1274 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1275                                        struct obd_device *obd, int *process)
1276 {
1277         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1278         case MDS_CONNECT: /* This will never get here, but for completeness. */
1279         case OST_CONNECT: /* This will never get here, but for completeness. */
1280         case MDS_DISCONNECT:
1281         case OST_DISCONNECT:
1282                *process = 1;
1283                RETURN(0);
1284
1285         case MDS_CLOSE:
1286         case MDS_SYNC: /* used in unmounting */
1287         case OBD_PING:
1288         case MDS_REINT:
1289         case LDLM_ENQUEUE:
1290                 *process = target_queue_recovery_request(req, obd);
1291                 RETURN(0);
1292
1293         default:
1294                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1295                 *process = 0;
1296                 /* XXX what should we set rq_status to here? */
1297                 req->rq_status = -EAGAIN;
1298                 RETURN(ptlrpc_error(req));
1299         }
1300 }
1301
1302 static char *reint_names[] = {
1303         [REINT_SETATTR] "setattr",
1304         [REINT_CREATE]  "create",
1305         [REINT_LINK]    "link",
1306         [REINT_UNLINK]  "unlink",
1307         [REINT_RENAME]  "rename",
1308         [REINT_OPEN]    "open",
1309 };
1310
1311 static int mds_set_info_rpc(struct obd_export *exp, struct ptlrpc_request *req)
1312 {
1313         char *key;
1314         __u32 *val;
1315         int keylen, rc = 0;
1316         ENTRY;
1317
1318         key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1319         if (key == NULL) {
1320                 DEBUG_REQ(D_HA, req, "no set_info key");
1321                 RETURN(-EFAULT);
1322         }
1323         keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1324
1325         val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, sizeof(*val));
1326         if (val == NULL) {
1327                 DEBUG_REQ(D_HA, req, "no set_info val");
1328                 RETURN(-EFAULT);
1329         }
1330
1331         rc = lustre_pack_reply(req, 1, NULL, NULL);
1332         if (rc)
1333                 RETURN(rc);
1334         lustre_msg_set_status(req->rq_repmsg, 0);
1335
1336         if (keylen < strlen("read-only") ||
1337             memcmp(key, "read-only", keylen) != 0)
1338                 RETURN(-EINVAL);
1339
1340         if (*val)
1341                 exp->exp_connect_flags |= OBD_CONNECT_RDONLY;
1342         else
1343                 exp->exp_connect_flags &= ~OBD_CONNECT_RDONLY;
1344
1345         RETURN(0);
1346 }
1347
1348 static int mds_handle_quotacheck(struct ptlrpc_request *req)
1349 {
1350         struct obd_quotactl *oqctl;
1351         int rc;
1352         ENTRY;
1353
1354         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1355                                    lustre_swab_obd_quotactl);
1356         if (oqctl == NULL)
1357                 RETURN(-EPROTO);
1358
1359         rc = lustre_pack_reply(req, 1, NULL, NULL);
1360         if (rc) {
1361                 CERROR("mds: out of memory while packing quotacheck reply\n");
1362                 RETURN(rc);
1363         }
1364
1365         req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1366         RETURN(0);
1367 }
1368
1369 static int mds_handle_quotactl(struct ptlrpc_request *req)
1370 {
1371         struct obd_quotactl *oqctl, *repoqc;
1372         int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1373         ENTRY;
1374
1375         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1376                                    lustre_swab_obd_quotactl);
1377         if (oqctl == NULL)
1378                 RETURN(-EPROTO);
1379
1380         rc = lustre_pack_reply(req, 2, size, NULL);
1381         if (rc)
1382                 RETURN(rc);
1383
1384         repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1385
1386         req->rq_status = obd_quotactl(req->rq_export, oqctl);
1387         *repoqc = *oqctl;
1388         RETURN(0);
1389 }
1390
1391 static int mds_msg_check_version(struct lustre_msg *msg)
1392 {
1393         int rc;
1394
1395         switch (lustre_msg_get_opc(msg)) {
1396         case MDS_CONNECT:
1397         case MDS_DISCONNECT:
1398         case OBD_PING:
1399                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1400                 if (rc)
1401                         CERROR("bad opc %u version %08x, expecting %08x\n",
1402                                lustre_msg_get_opc(msg),
1403                                lustre_msg_get_version(msg),
1404                                LUSTRE_OBD_VERSION);
1405                 break;
1406         case MDS_GETSTATUS:
1407         case MDS_GETATTR:
1408         case MDS_GETATTR_NAME:
1409         case MDS_STATFS:
1410         case MDS_READPAGE:
1411         case MDS_REINT:
1412         case MDS_CLOSE:
1413         case MDS_DONE_WRITING:
1414         case MDS_PIN:
1415         case MDS_SYNC:
1416         case MDS_GETXATTR:
1417         case MDS_SETXATTR:
1418         case MDS_SET_INFO:
1419         case MDS_QUOTACHECK:
1420         case MDS_QUOTACTL:
1421         case QUOTA_DQACQ:
1422         case QUOTA_DQREL:
1423                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1424                 if (rc)
1425                         CERROR("bad opc %u version %08x, expecting %08x\n",
1426                                lustre_msg_get_opc(msg),
1427                                lustre_msg_get_version(msg),
1428                                LUSTRE_MDS_VERSION);
1429                 break;
1430         case LDLM_ENQUEUE:
1431         case LDLM_CONVERT:
1432         case LDLM_BL_CALLBACK:
1433         case LDLM_CP_CALLBACK:
1434                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1435                 if (rc)
1436                         CERROR("bad opc %u version %08x, expecting %08x\n",
1437                                lustre_msg_get_opc(msg),
1438                                lustre_msg_get_version(msg),
1439                                LUSTRE_DLM_VERSION);
1440                 break;
1441         case OBD_LOG_CANCEL:
1442         case LLOG_ORIGIN_HANDLE_CREATE:
1443         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1444         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1445         case LLOG_ORIGIN_HANDLE_CLOSE:
1446         case LLOG_ORIGIN_HANDLE_DESTROY:
1447         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1448         case LLOG_CATINFO:
1449                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1450                 if (rc)
1451                         CERROR("bad opc %u version %08x, expecting %08x\n",
1452                                lustre_msg_get_opc(msg),
1453                                lustre_msg_get_version(msg),
1454                                LUSTRE_LOG_VERSION);
1455                 break;
1456         default:
1457                 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
1458                 rc = -ENOTSUPP;
1459         }
1460         return rc;
1461 }
1462
1463 int mds_handle(struct ptlrpc_request *req)
1464 {
1465         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1466         int rc = 0;
1467         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1468         struct obd_device *obd = NULL;
1469         ENTRY;
1470
1471         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1472
1473         LASSERT(current->journal_info == NULL);
1474
1475         rc = mds_msg_check_version(req->rq_reqmsg);
1476         if (rc) {
1477                 CERROR("MDS drop mal-formed request\n");
1478                 RETURN(rc);
1479         }
1480
1481         /* XXX identical to OST */
1482         if (lustre_msg_get_opc(req->rq_reqmsg) != MDS_CONNECT) {
1483                 struct mds_export_data *med;
1484                 int recovering, abort_recovery;
1485
1486                 if (req->rq_export == NULL) {
1487                         CERROR("operation %d on unconnected MDS from %s\n",
1488                                lustre_msg_get_opc(req->rq_reqmsg),
1489                                libcfs_id2str(req->rq_peer));
1490                         req->rq_status = -ENOTCONN;
1491                         GOTO(out, rc = -ENOTCONN);
1492                 }
1493
1494                 med = &req->rq_export->exp_mds_data;
1495                 obd = req->rq_export->exp_obd;
1496                 mds = &obd->u.mds;
1497
1498                 /* sanity check: if the xid matches, the request must
1499                  * be marked as a resent or replayed */
1500                 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
1501                    req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid))
1502                         if (!(lustre_msg_get_flags(req->rq_reqmsg) &
1503                                  (MSG_RESENT | MSG_REPLAY))) {
1504                                 CERROR("rq_xid "LPU64" matches last_xid, "
1505                                        "expected RESENT flag\n",
1506                                         req->rq_xid);
1507                                 req->rq_status = -ENOTCONN;
1508                                 GOTO(out, rc = -EFAULT);
1509                         }
1510                 /* else: note the opposite is not always true; a
1511                  * RESENT req after a failover will usually not match
1512                  * the last_xid, since it was likely never
1513                  * committed. A REPLAYed request will almost never
1514                  * match the last xid, however it could for a
1515                  * committed, but still retained, open. */
1516
1517                 /* Check for aborted recovery. */
1518                 spin_lock_bh(&obd->obd_processing_task_lock);
1519                 abort_recovery = obd->obd_abort_recovery;
1520                 recovering = obd->obd_recovering;
1521                 spin_unlock_bh(&obd->obd_processing_task_lock);
1522                 if (abort_recovery) {
1523                         target_abort_recovery(obd);
1524                 } else if (recovering) {
1525                         rc = mds_filter_recovery_request(req, obd,
1526                                                          &should_process);
1527                         if (rc || !should_process)
1528                                 RETURN(rc);
1529                 }
1530         }
1531
1532         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1533         case MDS_CONNECT:
1534                 DEBUG_REQ(D_INODE, req, "connect");
1535                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1536                 rc = target_handle_connect(req, mds_handle);
1537                 if (!rc) {
1538                         /* Now that we have an export, set mds. */
1539                         obd = req->rq_export->exp_obd;
1540                         mds = mds_req2mds(req);
1541                 }
1542                 break;
1543
1544         case MDS_DISCONNECT:
1545                 DEBUG_REQ(D_INODE, req, "disconnect");
1546                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1547                 rc = target_handle_disconnect(req);
1548                 req->rq_status = rc;            /* superfluous? */
1549                 break;
1550
1551         case MDS_GETSTATUS:
1552                 DEBUG_REQ(D_INODE, req, "getstatus");
1553                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1554                 rc = mds_getstatus(req);
1555                 break;
1556
1557         case MDS_GETATTR:
1558                 DEBUG_REQ(D_INODE, req, "getattr");
1559                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1560                 rc = mds_getattr(req, REQ_REC_OFF);
1561                 break;
1562
1563         case MDS_SETXATTR:
1564                 DEBUG_REQ(D_INODE, req, "setxattr");
1565                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SETXATTR_NET, 0);
1566                 rc = mds_setxattr(req);
1567                 break;
1568
1569         case MDS_GETXATTR:
1570                 DEBUG_REQ(D_INODE, req, "getxattr");
1571                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETXATTR_NET, 0);
1572                 rc = mds_getxattr(req);
1573                 break;
1574
1575         case MDS_GETATTR_NAME: {
1576                 struct lustre_handle lockh = { 0 };
1577                 DEBUG_REQ(D_INODE, req, "getattr_name");
1578                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1579
1580                 /* If this request gets a reconstructed reply, we won't be
1581                  * acquiring any new locks in mds_getattr_lock, so we don't
1582                  * want to cancel.
1583                  */
1584                 rc = mds_getattr_lock(req, REQ_REC_OFF, MDS_INODELOCK_UPDATE,
1585                                       &lockh);
1586                 /* this non-intent call (from an ioctl) is special */
1587                 req->rq_status = rc;
1588                 if (rc == 0 && lustre_handle_is_used(&lockh))
1589                         ldlm_lock_decref(&lockh, LCK_CR);
1590                 break;
1591         }
1592         case MDS_STATFS:
1593                 DEBUG_REQ(D_INODE, req, "statfs");
1594                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1595                 rc = mds_statfs(req);
1596                 break;
1597
1598         case MDS_READPAGE:
1599                 DEBUG_REQ(D_INODE, req, "readpage");
1600                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1601                 rc = mds_readpage(req, REQ_REC_OFF);
1602
1603                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1604                         RETURN(0);
1605                 }
1606
1607                 break;
1608
1609         case MDS_REINT: {
1610                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
1611                                              sizeof(*opcp));
1612                 __u32  opc;
1613                 int size[4] = { sizeof(struct ptlrpc_body),
1614                                 sizeof(struct mds_body),
1615                                 mds->mds_max_mdsize,
1616                                 mds->mds_max_cookiesize };
1617                 int bufcount;
1618
1619                 /* NB only peek inside req now; mds_reint() will swab it */
1620                 if (opcp == NULL) {
1621                         CERROR ("Can't inspect opcode\n");
1622                         rc = -EINVAL;
1623                         break;
1624                 }
1625                 opc = *opcp;
1626                 if (lustre_msg_swabbed(req->rq_reqmsg))
1627                         __swab32s(&opc);
1628
1629                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1630                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1631                            reint_names[opc] == NULL) ? reint_names[opc] :
1632                                                        "unknown opcode");
1633
1634                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1635
1636                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1637                         bufcount = 4;
1638                 else if (opc == REINT_OPEN)
1639                         bufcount = 3;
1640                 else
1641                         bufcount = 2;
1642
1643                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1644                 if (rc)
1645                         break;
1646
1647                 rc = mds_reint(req, REQ_REC_OFF, NULL);
1648                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1649                 break;
1650         }
1651
1652         case MDS_CLOSE:
1653                 DEBUG_REQ(D_INODE, req, "close");
1654                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1655                 rc = mds_close(req, REQ_REC_OFF);
1656                 break;
1657
1658         case MDS_DONE_WRITING:
1659                 DEBUG_REQ(D_INODE, req, "done_writing");
1660                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1661                 rc = mds_done_writing(req, REQ_REC_OFF);
1662                 break;
1663
1664         case MDS_PIN:
1665                 DEBUG_REQ(D_INODE, req, "pin");
1666                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1667                 rc = mds_pin(req, REQ_REC_OFF);
1668                 break;
1669
1670         case MDS_SYNC:
1671                 DEBUG_REQ(D_INODE, req, "sync");
1672                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1673                 rc = mds_sync(req, REQ_REC_OFF);
1674                 break;
1675
1676         case MDS_SET_INFO:
1677                 DEBUG_REQ(D_INODE, req, "set_info");
1678                 rc = mds_set_info_rpc(req->rq_export, req);
1679                 break;
1680
1681         case MDS_QUOTACHECK:
1682                 DEBUG_REQ(D_INODE, req, "quotacheck");
1683                 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACHECK_NET, 0);
1684                 rc = mds_handle_quotacheck(req);
1685                 break;
1686
1687         case MDS_QUOTACTL:
1688                 DEBUG_REQ(D_INODE, req, "quotactl");
1689                 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACTL_NET, 0);
1690                 rc = mds_handle_quotactl(req);
1691                 break;
1692
1693         case OBD_PING:
1694                 DEBUG_REQ(D_INODE, req, "ping");
1695                 rc = target_handle_ping(req);
1696                 break;
1697
1698         case OBD_LOG_CANCEL:
1699                 CDEBUG(D_INODE, "log cancel\n");
1700                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1701                 rc = -ENOTSUPP; /* la la la */
1702                 break;
1703
1704         case LDLM_ENQUEUE:
1705                 DEBUG_REQ(D_INODE, req, "enqueue");
1706                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1707                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1708                                          ldlm_server_blocking_ast, NULL);
1709                 fail = OBD_FAIL_LDLM_REPLY;
1710                 break;
1711         case LDLM_CONVERT:
1712                 DEBUG_REQ(D_INODE, req, "convert");
1713                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1714                 rc = ldlm_handle_convert(req);
1715                 break;
1716         case LDLM_BL_CALLBACK:
1717         case LDLM_CP_CALLBACK:
1718                 DEBUG_REQ(D_INODE, req, "callback");
1719                 CERROR("callbacks should not happen on MDS\n");
1720                 LBUG();
1721                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1722                 break;
1723         case LLOG_ORIGIN_HANDLE_CREATE:
1724                 DEBUG_REQ(D_INODE, req, "llog_init");
1725                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1726                 rc = llog_origin_handle_create(req);
1727                 break;
1728         case LLOG_ORIGIN_HANDLE_DESTROY:
1729                 DEBUG_REQ(D_INODE, req, "llog_init");
1730                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1731                 rc = llog_origin_handle_destroy(req);
1732                 break;
1733         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1734                 DEBUG_REQ(D_INODE, req, "llog next block");
1735                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1736                 rc = llog_origin_handle_next_block(req);
1737                 break;
1738         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1739                 DEBUG_REQ(D_INODE, req, "llog prev block");
1740                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1741                 rc = llog_origin_handle_prev_block(req);
1742                 break;
1743         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1744                 DEBUG_REQ(D_INODE, req, "llog read header");
1745                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1746                 rc = llog_origin_handle_read_header(req);
1747                 break;
1748         case LLOG_ORIGIN_HANDLE_CLOSE:
1749                 DEBUG_REQ(D_INODE, req, "llog close");
1750                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1751                 rc = llog_origin_handle_close(req);
1752                 break;
1753         case LLOG_CATINFO:
1754                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1755                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1756                 rc = llog_catinfo(req);
1757                 break;
1758         default:
1759                 req->rq_status = -ENOTSUPP;
1760                 rc = ptlrpc_error(req);
1761                 RETURN(rc);
1762         }
1763
1764         LASSERT(current->journal_info == NULL);
1765
1766         /* If we're DISCONNECTing, the mds_export_data is already freed */
1767         if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != MDS_DISCONNECT) {
1768                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1769
1770                 /* I don't think last_xid is used for anyway, so I'm not sure
1771                    if we need to care about last_close_xid here.*/
1772                 lustre_msg_set_last_xid(req->rq_repmsg,
1773                                        le64_to_cpu(med->med_mcd->mcd_last_xid));
1774
1775                 target_committed_to_req(req);
1776         }
1777
1778         EXIT;
1779  out:
1780
1781         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1782                 if (obd && obd->obd_recovering) {
1783                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1784                         return target_queue_final_reply(req, rc);
1785                 }
1786                 /* Lost a race with recovery; let the error path DTRT. */
1787                 rc = req->rq_status = -ENOTCONN;
1788         }
1789
1790         target_send_reply(req, rc, fail);
1791         return 0;
1792 }
1793
1794 /* Update the server data on disk.  This stores the new mount_count and
1795  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1796  * then the server last_rcvd value may be less than that of the clients.
1797  * This will alert us that we may need to do client recovery.
1798  *
1799  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1800  */
1801 int mds_update_server_data(struct obd_device *obd, int force_sync)
1802 {
1803         struct mds_obd *mds = &obd->u.mds;
1804         struct lr_server_data *lsd = mds->mds_server_data;
1805         struct file *filp = mds->mds_rcvd_filp;
1806         struct lvfs_run_ctxt saved;
1807         loff_t off = 0;
1808         int rc;
1809         ENTRY;
1810
1811         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1812                mds->mds_mount_count, mds->mds_last_transno);
1813
1814         spin_lock(&mds->mds_transno_lock);
1815         lsd->lsd_last_transno = cpu_to_le64(mds->mds_last_transno);
1816         spin_unlock(&mds->mds_transno_lock);
1817
1818         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1819         rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off,force_sync);
1820         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1821         if (rc)
1822                 CERROR("error writing MDS server data: rc = %d\n", rc);
1823
1824         RETURN(rc);
1825 }
1826
1827 static void fsoptions_to_mds_flags(struct mds_obd *mds, char *options)
1828 {
1829         char *p = options;
1830
1831         if (!options)
1832                 return;
1833
1834         while (*options) {
1835                 int len;
1836
1837                 while (*p && *p != ',')
1838                         p++;
1839
1840                 len = p - options;
1841                 if (len == sizeof("user_xattr") - 1 &&
1842                     memcmp(options, "user_xattr", len) == 0) {
1843                         mds->mds_fl_user_xattr = 1;
1844                         LCONSOLE_INFO("Enabling user_xattr\n");
1845                 } else if (len == sizeof("nouser_xattr") - 1 &&
1846                            memcmp(options, "nouser_xattr", len) == 0) {
1847                         mds->mds_fl_user_xattr = 0;
1848                         LCONSOLE_INFO("Disabling user_xattr\n");
1849                 } else if (len == sizeof("acl") - 1 &&
1850                            memcmp(options, "acl", len) == 0) {
1851 #ifdef CONFIG_FS_POSIX_ACL
1852                         mds->mds_fl_acl = 1;
1853                         LCONSOLE_INFO("Enabling ACL\n");
1854 #else
1855                         CWARN("ignoring unsupported acl mount option\n");
1856 #endif
1857                 } else if (len == sizeof("noacl") - 1 &&
1858                            memcmp(options, "noacl", len) == 0) {
1859 #ifdef CONFIG_FS_POSIX_ACL
1860                         mds->mds_fl_acl = 0;
1861                         LCONSOLE_INFO("Disabling ACL\n");
1862 #endif
1863                 }
1864
1865                 options = ++p;
1866         }
1867 }
1868
1869 /* mount the file system (secretly).  lustre_cfg parameters are:
1870  * 1 = device
1871  * 2 = fstype
1872  * 3 = config name
1873  * 4 = mount options
1874  */
1875 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1876 {
1877         struct lprocfs_static_vars lvars;
1878         struct lustre_cfg* lcfg = buf;
1879         struct mds_obd *mds = &obd->u.mds;
1880         struct lustre_mount_info *lmi;
1881         struct vfsmount *mnt;
1882         struct obd_uuid uuid;
1883         __u8 *uuid_ptr;
1884         char *options, *str, *label;
1885         char ns_name[48];
1886         unsigned long page;
1887         int rc = 0;
1888         ENTRY;
1889
1890         /* setup 1:/dev/loop/0 2:ext3 3:mdsA 4:errors=remount-ro,iopen_nopriv */
1891
1892         CLASSERT(offsetof(struct obd_device, u.obt) ==
1893                  offsetof(struct obd_device, u.mds.mds_obt));
1894
1895         if (lcfg->lcfg_bufcount < 3)
1896                 RETURN(rc = -EINVAL);
1897
1898         if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
1899                 RETURN(rc = -EINVAL);
1900
1901         lmi = server_get_mount(obd->obd_name);
1902         if (lmi) {
1903                 /* We already mounted in lustre_fill_super.
1904                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1905                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1906                 fsoptions_to_mds_flags(mds, lsi->lsi_ldd->ldd_mount_opts);
1907                 fsoptions_to_mds_flags(mds, lsi->lsi_lmd->lmd_opts);
1908                 mnt = lmi->lmi_mnt;
1909                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1910         } else {
1911                 /* old path - used by lctl */
1912                 CERROR("Using old MDS mount method\n");
1913                 page = __get_free_page(GFP_KERNEL);
1914                 if (!page)
1915                         RETURN(-ENOMEM);
1916
1917                 options = (char *)page;
1918                 memset(options, 0, CFS_PAGE_SIZE);
1919
1920                 /* here we use "iopen_nopriv" hardcoded, because it affects
1921                  * MDS utility and the rest of options are passed by mount
1922                  * options. Probably this should be moved to somewhere else
1923                  * like startup scripts or lconf. */
1924                 strcpy(options, "iopen_nopriv");
1925
1926                 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4)) {
1927                         sprintf(options + strlen(options), ",%s",
1928                                 lustre_cfg_string(lcfg, 4));
1929                         fsoptions_to_mds_flags(mds, options);
1930                 }
1931
1932                 mnt = do_kern_mount(lustre_cfg_string(lcfg, 2), 0,
1933                                     lustre_cfg_string(lcfg, 1),
1934                                     (void *)options);
1935                 free_page(page);
1936                 if (IS_ERR(mnt)) {
1937                         rc = PTR_ERR(mnt);
1938                         LCONSOLE_ERROR("Can't mount disk %s (%d)\n",
1939                                        lustre_cfg_string(lcfg, 1), rc);
1940                         RETURN(rc);
1941                 }
1942
1943                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1944         }
1945         if (IS_ERR(obd->obd_fsops))
1946                 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
1947
1948         CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
1949
1950         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1951
1952         sema_init(&mds->mds_epoch_sem, 1);
1953         spin_lock_init(&mds->mds_transno_lock);
1954         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1955         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1956         mds->mds_atime_diff = MAX_ATIME_DIFF;
1957
1958         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1959         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
1960         if (obd->obd_namespace == NULL) {
1961                 mds_cleanup(obd);
1962                 GOTO(err_ops, rc = -ENOMEM);
1963         }
1964         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1965
1966         rc = mds_fs_setup(obd, mnt);
1967         if (rc) {
1968                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
1969                        obd->obd_name, rc);
1970                 GOTO(err_ns, rc);
1971         }
1972
1973         rc = llog_start_commit_thread();
1974         if (rc < 0)
1975                 GOTO(err_fs, rc);
1976
1977         if (lcfg->lcfg_bufcount >= 4 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1978                 class_uuid_t uuid;
1979
1980                 generate_random_uuid(uuid);
1981                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1982
1983                 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
1984                 if (mds->mds_profile == NULL)
1985                         GOTO(err_fs, rc = -ENOMEM);
1986
1987                 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
1988                         LUSTRE_CFG_BUFLEN(lcfg, 3));
1989         }
1990
1991         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1992                            "mds_ldlm_client", &obd->obd_ldlm_client);
1993         obd->obd_replayable = 1;
1994
1995         rc = lquota_setup(mds_quota_interface_ref, obd);
1996         if (rc)
1997                 GOTO(err_fs, rc);
1998
1999         mds->mds_group_hash = upcall_cache_init(obd->obd_name);
2000         if (IS_ERR(mds->mds_group_hash)) {
2001                 rc = PTR_ERR(mds->mds_group_hash);
2002                 mds->mds_group_hash = NULL;
2003                 GOTO(err_qctxt, rc);
2004         }
2005
2006         /* Don't wait for mds_postrecov trying to clear orphans */
2007         obd->obd_async_recov = 1;
2008         rc = mds_postsetup(obd);
2009         /* Bug 11557 - allow async abort_recov start
2010            FIXME can remove most of this obd_async_recov plumbing
2011         obd->obd_async_recov = 0;
2012         */
2013         if (rc)
2014                 GOTO(err_qctxt, rc);
2015
2016         lprocfs_init_vars(mds, &lvars);
2017         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2018             lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
2019                 /* Init private stats here */
2020                 mds_stats_counter_init(obd->obd_stats);
2021                 obd->obd_proc_exports = proc_mkdir("exports",
2022                                                    obd->obd_proc_entry);
2023         }
2024
2025         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2026         if (uuid_ptr != NULL) {
2027                 class_uuid_unparse(uuid_ptr, &uuid);
2028                 str = uuid.uuid;
2029         } else {
2030                 str = "no UUID";
2031         }
2032
2033         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2034         if (obd->obd_recovering) {
2035                 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2036                               "recovery until %d %s reconnect, or if no clients"
2037                               " reconnect for %d:%.02d; during that time new "
2038                               "clients will not be allowed to connect. "
2039                               "Recovery progress can be monitored by watching "
2040                               "/proc/fs/lustre/mds/%s/recovery_status.\n",
2041                               obd->obd_name, lustre_cfg_string(lcfg, 1),
2042                               label ?: "", label ? "/" : "", str,
2043                               obd->obd_recoverable_clients,
2044                               (obd->obd_recoverable_clients == 1) ?
2045                               "client" : "clients",
2046                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
2047                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
2048                               obd->obd_name);
2049         } else {
2050                 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2051                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2052                               label ?: "", label ? "/" : "", str,
2053                               obd->obd_replayable ? "enabled" : "disabled");
2054         }
2055
2056         ldlm_timeout = 6;
2057
2058         RETURN(0);
2059
2060 err_qctxt:
2061         lquota_cleanup(mds_quota_interface_ref, obd);
2062 err_fs:
2063         /* No extra cleanup needed for llog_init_commit_thread() */
2064         mds_fs_cleanup(obd);
2065         upcall_cache_cleanup(mds->mds_group_hash);
2066         mds->mds_group_hash = NULL;
2067 err_ns:
2068         ldlm_namespace_free(obd->obd_namespace, 0);
2069         obd->obd_namespace = NULL;
2070 err_ops:
2071         fsfilt_put_ops(obd->obd_fsops);
2072 err_put:
2073         server_put_mount(obd->obd_name, mnt);
2074         obd->u.obt.obt_sb = NULL;
2075         return rc;
2076 }
2077
2078 static int mds_lov_clean(struct obd_device *obd)
2079 {
2080         struct mds_obd *mds = &obd->u.mds;
2081         struct obd_device *osc = mds->mds_osc_obd;
2082         ENTRY;
2083
2084         if (mds->mds_profile) {
2085                 class_del_profile(mds->mds_profile);
2086                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2087                 mds->mds_profile = NULL;
2088         }
2089
2090         /* There better be a lov */
2091         if (!osc)
2092                 RETURN(0);
2093         if (IS_ERR(osc))
2094                 RETURN(PTR_ERR(osc));
2095
2096         obd_register_observer(osc, NULL);
2097
2098         /* Give lov our same shutdown flags */
2099         osc->obd_force = obd->obd_force;
2100         osc->obd_fail = obd->obd_fail;
2101
2102         /* Cleanup the lov */
2103         obd_disconnect(mds->mds_osc_exp);
2104         class_manual_cleanup(osc);
2105         mds->mds_osc_exp = NULL;
2106
2107         RETURN(0);
2108 }
2109
2110 static int mds_postsetup(struct obd_device *obd)
2111 {
2112         struct mds_obd *mds = &obd->u.mds;
2113         int rc = 0;
2114         ENTRY;
2115
2116         rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2117                         &llog_lvfs_ops);
2118         if (rc)
2119                 RETURN(rc);
2120
2121         rc = llog_setup(obd, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2122                         &llog_lvfs_ops);
2123         if (rc)
2124                 RETURN(rc);
2125
2126         if (mds->mds_profile) {
2127                 struct lustre_profile *lprof;
2128                 /* The profile defines which osc and mdc to connect to, for a
2129                    client.  We reuse that here to figure out the name of the
2130                    lov to use (and ignore lprof->lp_mdc).
2131                    The profile was set in the config log with
2132                    LCFG_MOUNTOPT profilenm oscnm mdcnm */
2133                 lprof = class_get_profile(mds->mds_profile);
2134                 if (lprof == NULL) {
2135                         CERROR("No profile found: %s\n", mds->mds_profile);
2136                         GOTO(err_cleanup, rc = -ENOENT);
2137                 }
2138                 rc = mds_lov_connect(obd, lprof->lp_osc);
2139                 if (rc)
2140                         GOTO(err_cleanup, rc);
2141         }
2142
2143         RETURN(rc);
2144
2145 err_cleanup:
2146         mds_lov_clean(obd);
2147         llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2148         llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2149         RETURN(rc);
2150 }
2151
2152 int mds_postrecov(struct obd_device *obd)
2153 {
2154         int rc;
2155         ENTRY;
2156
2157         if (obd->obd_fail)
2158                 RETURN(0);
2159
2160         LASSERT(!obd->obd_recovering);
2161         LASSERT(llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT) != NULL);
2162
2163         /* FIXME why not put this in the synchronize? */
2164         /* set nextid first, so we are sure it happens */
2165         rc = mds_lov_set_nextid(obd);
2166         if (rc) {
2167                 CERROR("%s: mds_lov_set_nextid failed %d\n",
2168                        obd->obd_name, rc);
2169                 GOTO(out, rc);
2170         }
2171
2172         /* clean PENDING dir */
2173         rc = mds_cleanup_pending(obd);
2174         if (rc < 0)
2175                 GOTO(out, rc);
2176
2177         /* FIXME Does target_finish_recovery really need this to block? */
2178         /* Notify the LOV, which will in turn call mds_notify for each tgt */
2179         /* This means that we have to hack obd_notify to think we're obd_set_up
2180            during mds_lov_connect. */
2181         obd_notify(obd->u.mds.mds_osc_obd, NULL,
2182                    obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
2183                    OBD_NOTIFY_SYNC, NULL);
2184
2185         /* quota recovery */
2186         lquota_recovery(mds_quota_interface_ref, obd);
2187
2188 out:
2189         RETURN(rc);
2190 }
2191
2192 /* We need to be able to stop an mds_lov_synchronize */
2193 static int mds_lov_early_clean(struct obd_device *obd)
2194 {
2195         struct mds_obd *mds = &obd->u.mds;
2196         struct obd_device *osc = mds->mds_osc_obd;
2197
2198         if (!osc || (!obd->obd_force && !obd->obd_fail))
2199                 return(0);
2200
2201         CDEBUG(D_HA, "abort inflight\n");
2202         return (obd_precleanup(osc, OBD_CLEANUP_EARLY));
2203 }
2204
2205 static int mds_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2206 {
2207         int rc = 0;
2208         ENTRY;
2209
2210         switch (stage) {
2211         case OBD_CLEANUP_EARLY:
2212                 break;
2213         case OBD_CLEANUP_EXPORTS:
2214                 target_cleanup_recovery(obd);
2215                 mds_lov_early_clean(obd);
2216                 break;
2217         case OBD_CLEANUP_SELF_EXP:
2218                 mds_lov_disconnect(obd);
2219                 mds_lov_clean(obd);
2220                 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2221                 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2222                 rc = obd_llog_finish(obd, 0);
2223                 break;
2224         case OBD_CLEANUP_OBD:
2225                 break;
2226         }
2227         RETURN(rc);
2228 }
2229
2230 static int mds_cleanup(struct obd_device *obd)
2231 {
2232         struct mds_obd *mds = &obd->u.mds;
2233         lvfs_sbdev_type save_dev;
2234         ENTRY;
2235
2236         if (obd->u.obt.obt_sb == NULL)
2237                 RETURN(0);
2238         save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2239
2240         if (mds->mds_osc_exp)
2241                 /* lov export was disconnected by mds_lov_clean;
2242                    we just need to drop our ref */
2243                 class_export_put(mds->mds_osc_exp);
2244
2245         lprocfs_obd_cleanup(obd);
2246         lprocfs_free_obd_stats(obd);
2247
2248         lquota_cleanup(mds_quota_interface_ref, obd);
2249
2250         mds_update_server_data(obd, 1);
2251         if (mds->mds_lov_objids != NULL)
2252                 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2253         mds_fs_cleanup(obd);
2254
2255         upcall_cache_cleanup(mds->mds_group_hash);
2256         mds->mds_group_hash = NULL;
2257
2258         server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2259         obd->u.obt.obt_sb = NULL;
2260
2261         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2262
2263         spin_lock_bh(&obd->obd_processing_task_lock);
2264         if (obd->obd_recovering) {
2265                 target_cancel_recovery_timer(obd);
2266                 obd->obd_recovering = 0;
2267         }
2268         spin_unlock_bh(&obd->obd_processing_task_lock);
2269
2270         fsfilt_put_ops(obd->obd_fsops);
2271
2272         LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2273
2274         RETURN(0);
2275 }
2276
2277 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2278                                         struct ldlm_lock *new_lock,
2279                                         struct ldlm_lock **old_lock,
2280                                         struct lustre_handle *lockh)
2281 {
2282         struct obd_export *exp = req->rq_export;
2283         struct ldlm_request *dlmreq =
2284                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2285         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2286         struct list_head *iter;
2287
2288         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2289                 return;
2290
2291         spin_lock(&exp->exp_ldlm_data.led_lock);
2292         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2293                 struct ldlm_lock *lock;
2294                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2295                 if (lock == new_lock)
2296                         continue;
2297                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2298                         lockh->cookie = lock->l_handle.h_cookie;
2299                         LDLM_DEBUG(lock, "restoring lock cookie");
2300                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2301                                   lockh->cookie);
2302                         if (old_lock)
2303                                 *old_lock = LDLM_LOCK_GET(lock);
2304                         spin_unlock(&exp->exp_ldlm_data.led_lock);
2305                         return;
2306                 }
2307         }
2308         spin_unlock(&exp->exp_ldlm_data.led_lock);
2309
2310         /* If the xid matches, then we know this is a resent request,
2311          * and allow it. (It's probably an OPEN, for which we don't
2312          * send a lock */
2313         if (req->rq_xid ==
2314             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2315                 return;
2316
2317         if (req->rq_xid ==
2318             le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2319                 return;
2320
2321         /* This remote handle isn't enqueued, so we never received or
2322          * processed this request.  Clear MSG_RESENT, because it can
2323          * be handled like any normal request now. */
2324
2325         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2326
2327         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2328                   remote_hdl.cookie);
2329 }
2330
2331 int intent_disposition(struct ldlm_reply *rep, int flag)
2332 {
2333         if (!rep)
2334                 return 0;
2335         return (rep->lock_policy_res1 & flag);
2336 }
2337
2338 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2339 {
2340         if (!rep)
2341                 return;
2342         rep->lock_policy_res1 |= flag;
2343 }
2344
2345 static int mds_intent_policy(struct ldlm_namespace *ns,
2346                              struct ldlm_lock **lockp, void *req_cookie,
2347                              ldlm_mode_t mode, int flags, void *data)
2348 {
2349         struct ptlrpc_request *req = req_cookie;
2350         struct ldlm_lock *lock = *lockp;
2351         struct ldlm_intent *it;
2352         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2353         struct ldlm_reply *rep;
2354         struct lustre_handle lockh = { 0 };
2355         struct ldlm_lock *new_lock = NULL;
2356         int getattr_part = MDS_INODELOCK_UPDATE;
2357         int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2358                            [DLM_LOCKREPLY_OFF]   = sizeof(struct ldlm_reply),
2359                            [DLM_REPLY_REC_OFF]   = sizeof(struct mds_body),
2360                            [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2361         int repbufcnt = 4, rc;
2362         ENTRY;
2363
2364         LASSERT(req != NULL);
2365
2366         if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2367                 /* No intent was provided */
2368                 rc = lustre_pack_reply(req, 2, repsize, NULL);
2369                 LASSERT(rc == 0);
2370                 RETURN(0);
2371         }
2372
2373         it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2374                                 lustre_swab_ldlm_intent);
2375         if (it == NULL) {
2376                 CERROR("Intent missing\n");
2377                 RETURN(req->rq_status = -EFAULT);
2378         }
2379
2380         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2381
2382         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2383             (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2384                 /* we should never allow OBD_CONNECT_ACL if not configured */
2385                 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2386         else if (it->opc & IT_UNLINK)
2387                 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2388
2389         rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2390         if (rc)
2391                 RETURN(req->rq_status = rc);
2392
2393         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2394         intent_set_disposition(rep, DISP_IT_EXECD);
2395
2396
2397         /* execute policy */
2398         switch ((long)it->opc) {
2399         case IT_OPEN:
2400         case IT_CREAT|IT_OPEN:
2401                 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2402                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2403                                             &lockh);
2404                 /* XXX swab here to assert that an mds_open reint
2405                  * packet is following */
2406                 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2407                                                   &lockh);
2408 #if 0
2409                 /* We abort the lock if the lookup was negative and
2410                  * we did not make it to the OPEN portion */
2411                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2412                         RETURN(ELDLM_LOCK_ABORTED);
2413                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2414                     !intent_disposition(rep, DISP_OPEN_OPEN))
2415 #endif
2416
2417                 /* If there was an error of some sort or if we are not
2418                  * returning any locks */
2419                 if (rep->lock_policy_res2 ||
2420                     !intent_disposition(rep, DISP_OPEN_LOCK))
2421                         RETURN(ELDLM_LOCK_ABORTED);
2422                 break;
2423         case IT_LOOKUP:
2424                         getattr_part = MDS_INODELOCK_LOOKUP;
2425         case IT_GETATTR:
2426                         getattr_part |= MDS_INODELOCK_LOOKUP;
2427                         OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2428         case IT_READDIR:
2429                 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2430                                             &new_lock, &lockh);
2431
2432                 /* INODEBITS_INTEROP: if this lock was converted from a
2433                  * plain lock (client does not support inodebits), then
2434                  * child lock must be taken with both lookup and update
2435                  * bits set for all operations.
2436                  */
2437                 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2438                         getattr_part = MDS_INODELOCK_LOOKUP |
2439                                        MDS_INODELOCK_UPDATE;
2440
2441                 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2442                                                          getattr_part, &lockh);
2443                 /* FIXME: LDLM can set req->rq_status. MDS sets
2444                    policy_res{1,2} with disposition and status.
2445                    - replay: returns 0 & req->status is old status
2446                    - otherwise: returns req->status */
2447                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2448                         rep->lock_policy_res2 = 0;
2449                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2450                     rep->lock_policy_res2)
2451                         RETURN(ELDLM_LOCK_ABORTED);
2452                 if (req->rq_status != 0) {
2453                         LBUG();
2454                         rep->lock_policy_res2 = req->rq_status;
2455                         RETURN(ELDLM_LOCK_ABORTED);
2456                 }
2457                 break;
2458         default:
2459                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2460                 RETURN(-EFAULT);
2461         }
2462
2463         /* By this point, whatever function we called above must have either
2464          * filled in 'lockh', been an intent replay, or returned an error.  We
2465          * want to allow replayed RPCs to not get a lock, since we would just
2466          * drop it below anyways because lock replay is done separately by the
2467          * client afterwards.  For regular RPCs we want to give the new lock to
2468          * the client instead of whatever lock it was about to get. */
2469         if (new_lock == NULL)
2470                 new_lock = ldlm_handle2lock(&lockh);
2471         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2472                 RETURN(0);
2473
2474         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2475                  it->opc, lockh.cookie);
2476
2477         /* If we've already given this lock to a client once, then we should
2478          * have no readers or writers.  Otherwise, we should have one reader
2479          * _or_ writer ref (which will be zeroed below) before returning the
2480          * lock to a client. */
2481         if (new_lock->l_export == req->rq_export) {
2482                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2483         } else {
2484                 LASSERT(new_lock->l_export == NULL);
2485                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2486         }
2487
2488         *lockp = new_lock;
2489
2490         if (new_lock->l_export == req->rq_export) {
2491                 /* Already gave this to the client, which means that we
2492                  * reconstructed a reply. */
2493                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2494                         MSG_RESENT);
2495                 RETURN(ELDLM_LOCK_REPLACED);
2496         }
2497
2498         /* Fixup the lock to be given to the client */
2499         lock_res_and_lock(new_lock);
2500         new_lock->l_readers = 0;
2501         new_lock->l_writers = 0;
2502
2503         new_lock->l_export = class_export_get(req->rq_export);
2504         spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2505         list_add(&new_lock->l_export_chain,
2506                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2507         spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2508
2509         new_lock->l_blocking_ast = lock->l_blocking_ast;
2510         new_lock->l_completion_ast = lock->l_completion_ast;
2511
2512         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2513                sizeof(lock->l_remote_handle));
2514
2515         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2516
2517         unlock_res_and_lock(new_lock);
2518         LDLM_LOCK_PUT(new_lock);
2519
2520         RETURN(ELDLM_LOCK_REPLACED);
2521 }
2522
2523 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2524 {
2525         struct mds_obd *mds = &obd->u.mds;
2526         struct lprocfs_static_vars lvars;
2527         int mds_min_threads = MDS_THREADS_AUTO_MIN;
2528         int mds_max_threads = MDS_THREADS_AUTO_MAX;
2529         int rc = 0;
2530         ENTRY;
2531
2532         lprocfs_init_vars(mdt, &lvars);
2533         lprocfs_obd_setup(obd, lvars.obd_vars);
2534
2535         sema_init(&mds->mds_health_sem, 1);
2536
2537         if (mds_num_threads) {
2538                 /* If mds_num_threads is set, it is the min and the max. */
2539                 if (mds_num_threads > MDS_THREADS_MAX)
2540                         mds_num_threads = MDS_THREADS_MAX;
2541                 if (mds_num_threads < MDS_THREADS_MIN)
2542                         mds_num_threads = MDS_THREADS_MIN;
2543                 mds_max_threads = mds_min_threads = mds_num_threads;
2544         }
2545
2546         mds->mds_service =
2547                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2548                                 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2549                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2550                                 mds_handle, LUSTRE_MDS_NAME,
2551                                 obd->obd_proc_entry, NULL,
2552                                 mds_min_threads, mds_max_threads, "ll_mdt");
2553
2554         if (!mds->mds_service) {
2555                 CERROR("failed to start service\n");
2556                 GOTO(err_lprocfs, rc = -ENOMEM);
2557         }
2558
2559         rc = ptlrpc_start_threads(obd, mds->mds_service);
2560         if (rc)
2561                 GOTO(err_thread, rc);
2562
2563         mds->mds_setattr_service =
2564                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2565                                 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2566                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2567                                 mds_handle, "mds_setattr",
2568                                 obd->obd_proc_entry, NULL,
2569                                 mds_min_threads, mds_max_threads,
2570                                 "ll_mdt_attr");
2571         if (!mds->mds_setattr_service) {
2572                 CERROR("failed to start getattr service\n");
2573                 GOTO(err_thread, rc = -ENOMEM);
2574         }
2575
2576         rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2577         if (rc)
2578                 GOTO(err_thread2, rc);
2579
2580         mds->mds_readpage_service =
2581                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2582                                 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2583                                 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2584                                 mds_handle, "mds_readpage",
2585                                 obd->obd_proc_entry, NULL,
2586                                 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2587                                 "ll_mdt_rdpg");
2588         if (!mds->mds_readpage_service) {
2589                 CERROR("failed to start readpage service\n");
2590                 GOTO(err_thread2, rc = -ENOMEM);
2591         }
2592
2593         rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2594
2595         if (rc)
2596                 GOTO(err_thread3, rc);
2597
2598         ping_evictor_start();
2599
2600         RETURN(0);
2601
2602 err_thread3:
2603         ptlrpc_unregister_service(mds->mds_readpage_service);
2604         mds->mds_readpage_service = NULL;
2605 err_thread2:
2606         ptlrpc_unregister_service(mds->mds_setattr_service);
2607         mds->mds_setattr_service = NULL;
2608 err_thread:
2609         ptlrpc_unregister_service(mds->mds_service);
2610         mds->mds_service = NULL;
2611 err_lprocfs:
2612         lprocfs_obd_cleanup(obd);
2613         return rc;
2614 }
2615
2616 static int mdt_cleanup(struct obd_device *obd)
2617 {
2618         struct mds_obd *mds = &obd->u.mds;
2619         ENTRY;
2620
2621         ping_evictor_stop();
2622
2623         down(&mds->mds_health_sem);
2624         ptlrpc_unregister_service(mds->mds_readpage_service);
2625         ptlrpc_unregister_service(mds->mds_setattr_service);
2626         ptlrpc_unregister_service(mds->mds_service);
2627         mds->mds_readpage_service = NULL;
2628         mds->mds_setattr_service = NULL;
2629         mds->mds_service = NULL;
2630         up(&mds->mds_health_sem);
2631
2632         lprocfs_obd_cleanup(obd);
2633
2634         RETURN(0);
2635 }
2636
2637 static int mdt_health_check(struct obd_device *obd)
2638 {
2639         struct mds_obd *mds = &obd->u.mds;
2640         int rc = 0;
2641
2642         down(&mds->mds_health_sem);
2643         rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2644         rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2645         rc |= ptlrpc_service_health_check(mds->mds_service);
2646         up(&mds->mds_health_sem);
2647
2648         /*
2649          * health_check to return 0 on healthy
2650          * and 1 on unhealthy.
2651          */
2652         if(rc != 0)
2653                 rc = 1;
2654
2655         return rc;
2656 }
2657
2658 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2659                                           void *data)
2660 {
2661         struct obd_device *obd = data;
2662         struct ll_fid fid;
2663         fid.id = id;
2664         fid.generation = gen;
2665         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2666 }
2667
2668 static int mds_health_check(struct obd_device *obd)
2669 {
2670         struct obd_device_target *odt = &obd->u.obt;
2671         struct mds_obd *mds = &obd->u.mds;
2672         int rc = 0;
2673
2674         if (odt->obt_sb->s_flags & MS_RDONLY)
2675                 rc = 1;
2676
2677         LASSERT(mds->mds_health_check_filp != NULL);
2678         rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2679
2680         return rc;
2681 }
2682
2683 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2684 {
2685         struct lustre_cfg *lcfg = buf;
2686         struct lprocfs_static_vars lvars;
2687         int rc;
2688
2689         lprocfs_init_vars(mds, &lvars);
2690
2691         rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2692
2693         return(rc);
2694 }
2695
2696 struct lvfs_callback_ops mds_lvfs_ops = {
2697         l_fid2dentry:     mds_lvfs_fid2dentry,
2698 };
2699
2700 /* use obd ops to offer management infrastructure */
2701 static struct obd_ops mds_obd_ops = {
2702         .o_owner           = THIS_MODULE,
2703         .o_connect         = mds_connect,
2704         .o_reconnect       = mds_reconnect,
2705         .o_init_export     = mds_init_export,
2706         .o_destroy_export  = mds_destroy_export,
2707         .o_disconnect      = mds_disconnect,
2708         .o_setup           = mds_setup,
2709         .o_precleanup      = mds_precleanup,
2710         .o_cleanup         = mds_cleanup,
2711         .o_postrecov       = mds_postrecov,
2712         .o_statfs          = mds_obd_statfs,
2713         .o_iocontrol       = mds_iocontrol,
2714         .o_create          = mds_obd_create,
2715         .o_destroy         = mds_obd_destroy,
2716         .o_llog_init       = mds_llog_init,
2717         .o_llog_finish     = mds_llog_finish,
2718         .o_notify          = mds_notify,
2719         .o_health_check    = mds_health_check,
2720         .o_process_config  = mds_process_config,
2721 };
2722
2723 static struct obd_ops mdt_obd_ops = {
2724         .o_owner           = THIS_MODULE,
2725         .o_setup           = mdt_setup,
2726         .o_cleanup         = mdt_cleanup,
2727         .o_health_check    = mdt_health_check,
2728 };
2729
2730 quota_interface_t *mds_quota_interface_ref;
2731 extern quota_interface_t mds_quota_interface;
2732
2733 static int __init mds_init(void)
2734 {
2735         int rc;
2736         struct lprocfs_static_vars lvars;
2737
2738         request_module("lquota");
2739         mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2740         rc = lquota_init(mds_quota_interface_ref);
2741         if (rc) {
2742                 if (mds_quota_interface_ref)
2743                         PORTAL_SYMBOL_PUT(mds_quota_interface);
2744                 return rc;
2745         }
2746         init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2747
2748         lprocfs_init_vars(mds, &lvars);
2749         class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
2750         lprocfs_init_vars(mdt, &lvars);
2751         class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
2752
2753         return 0;
2754 }
2755
2756 static void /*__exit*/ mds_exit(void)
2757 {
2758         lquota_exit(mds_quota_interface_ref);
2759         if (mds_quota_interface_ref)
2760                 PORTAL_SYMBOL_PUT(mds_quota_interface);
2761
2762         class_unregister_type(LUSTRE_MDS_NAME);
2763         class_unregister_type(LUSTRE_MDT_NAME);
2764 }
2765
2766 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2767 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2768 MODULE_LICENSE("GPL");
2769
2770 module_init(mds_init);
2771 module_exit(mds_exit);