lustre/obdfilter/filter.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/obdfilter/filter.c
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *   Author: Peter Braam <braam@clusterfs.com>
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 /*
  27  * Invariant: Get O/R i_mutex for lookup, if needed, before any journal ops
  28  *            (which need to get journal_lock, may block if journal full).
  29  *
  30  * Invariant: Call filter_start_transno() before any journal ops to avoid the
  31  *            same deadlock problem.  We can (and want) to get rid of the
  32  *            transno sem in favour of the dir/inode i_mutex to avoid single
  33  *            threaded operation on the OST.
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_FILTER
  37
  38 #include <linux/config.h>
  39 #include <linux/module.h>
  40 #include <linux/fs.h>
  41 #include <linux/dcache.h>
  42 #include <linux/init.h>
  43 #include <linux/version.h>
  44 #include <linux/sched.h>
  45 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
  46 # include <linux/mount.h>
  47 # include <linux/buffer_head.h>
  48 #endif
  49
  50 #include <obd_class.h>
  51 #include <obd_lov.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_fsfilt.h>
  54 #include <lprocfs_status.h>
  55 #include <lustre_log.h>
  56 #include <lustre_commit_confd.h>
  57 #include <libcfs/list.h>
  58 #include <lustre_disk.h>
  59 #include <lustre_quota.h>
  60 #include <linux/slab.h>
  61 #include <lustre_param.h>
  62
  63 #include "filter_internal.h"
  64
  65 static struct lvfs_callback_ops filter_lvfs_ops;
  66 kmem_cache_t *ll_fmd_cachep;
  67
  68 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
  69                              void *cb_data, int error)
  70 {
  71         obd_transno_commit_cb(obd, transno, error);
  72 }
  73
  74 /* Assumes caller has already pushed us into the kernel context. */
  75 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
  76                           int rc)
  77 {
  78         struct filter_obd *filter = &exp->exp_obd->u.filter;
  79         struct filter_export_data *fed = &exp->exp_filter_data;
  80         struct filter_client_data *fcd = fed->fed_fcd;
  81         __u64 last_rcvd;
  82         loff_t off;
  83         int err, log_pri = D_HA;
  84
  85         /* Propagate error code. */
  86         if (rc)
  87                 RETURN(rc);
  88
  89         if (!exp->exp_obd->obd_replayable || oti == NULL)
  90                 RETURN(rc);
  91
  92         /* we don't allocate new transnos for replayed requests */
  93         if (oti->oti_transno == 0) {
  94                 spin_lock(&filter->fo_translock);
  95                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
  96                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
  97                 spin_unlock(&filter->fo_translock);
  98                 oti->oti_transno = last_rcvd;
  99         } else {
 100                 spin_lock(&filter->fo_translock);
 101                 last_rcvd = oti->oti_transno;
 102                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
 103                         filter->fo_fsd->lsd_last_transno =
 104                                 cpu_to_le64(last_rcvd);
 105                 spin_unlock(&filter->fo_translock);
 106         }
 107         fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
 108
 109         /* could get xid from oti, if it's ever needed */
 110         fcd->fcd_last_xid = 0;
 111
 112         off = fed->fed_lr_off;
 113         if (off <= 0) {
 114                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
 115                        fed->fed_lr_idx, fed->fed_lr_off);
 116                 err = -EINVAL;
 117         } else {
 118                 fsfilt_add_journal_cb(exp->exp_obd, last_rcvd, oti->oti_handle,
 119                                       filter_commit_cb, NULL);
 120                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
 121                                           fcd, sizeof(*fcd), &off,
 122                                           exp->exp_need_sync);
 123         }
 124         if (err) {
 125                 log_pri = D_ERROR;
 126                 if (rc == 0)
 127                         rc = err;
 128         }
 129
 130         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
 131                last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
 132
 133         RETURN(rc);
 134 }
 135
 136 void f_dput(struct dentry *dentry)
 137 {
 138         /* Can't go inside filter_ddelete because it can block */
 139         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
 140                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
 141         LASSERT(atomic_read(&dentry->d_count) > 0);
 142
 143         dput(dentry);
 144 }
 145
 146 static void init_brw_stats(struct brw_stats *brw_stats)
 147 {
 148         int i;
 149         for (i = 0; i < BRW_LAST; i++)
 150                 spin_lock_init(&brw_stats->hist[i].oh_lock);
 151 }
 152
 153 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
 154    plus the procfs overhead :( */
 155 static int filter_export_stats_init(struct obd_device *obd,
 156                                     struct obd_export *exp)
 157 {
 158         struct filter_export_data *fed = &exp->exp_filter_data;
 159         struct proc_dir_entry *brw_entry;
 160         int rc, num_stats;
 161         ENTRY;
 162
 163         init_brw_stats(&fed->fed_brw_stats);
 164
 165         rc = lprocfs_exp_setup(exp);
 166         if (rc)
 167                 RETURN(rc);
 168
 169         /* Create a per export proc entry for brw_stats */
 170         brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
 171         if (brw_entry == NULL)
 172                RETURN(-ENOMEM);
 173         brw_entry->proc_fops = &filter_per_export_stats_fops;
 174         brw_entry->data = fed;
 175
 176         /* Create a per export proc entry for ops stats */
 177         num_stats = (sizeof(*obd->obd_type->typ_ops) / sizeof(void *)) +
 178                      LPROC_FILTER_LAST - 1;
 179         exp->exp_ops_stats = lprocfs_alloc_stats(num_stats);
 180         if (exp->exp_ops_stats == NULL)
 181               RETURN(-ENOMEM);
 182         lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
 183         lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
 184                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
 185         lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
 186                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
 187         lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
 188         RETURN(0);
 189 }
 190
 191 /* Add client data to the FILTER.  We use a bitmap to locate a free space
 192  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
 193  * Otherwise, we have just read the data from the last_rcvd file and
 194  * we know its offset. */
 195 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
 196                              int cl_idx)
 197 {
 198         struct filter_obd *filter = &obd->u.filter;
 199         struct filter_export_data *fed = &exp->exp_filter_data;
 200         unsigned long *bitmap = filter->fo_last_rcvd_slots;
 201         int new_client = (cl_idx == -1);
 202
 203         ENTRY;
 204
 205         LASSERT(bitmap != NULL);
 206         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
 207
 208         /* Self-export */
 209         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0) {
 210                 init_brw_stats(&fed->fed_brw_stats);
 211                 RETURN(0);
 212         }
 213
 214         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 215          * there's no need for extra complication here
 216          */
 217         if (new_client) {
 218                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 219         repeat:
 220                 if (cl_idx >= LR_MAX_CLIENTS) {
 221                         CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
 222                                cl_idx);
 223                         RETURN(-EOVERFLOW);
 224                 }
 225                 if (test_and_set_bit(cl_idx, bitmap)) {
 226                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
 227                                                     cl_idx);
 228                         goto repeat;
 229                 }
 230         } else {
 231                 if (test_and_set_bit(cl_idx, bitmap)) {
 232                         CERROR("FILTER client %d: bit already set in bitmap!\n",
 233                                cl_idx);
 234                         LBUG();
 235                 }
 236         }
 237
 238         fed->fed_lr_idx = cl_idx;
 239         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
 240                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
 241         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
 242         filter_export_stats_init(obd, exp);
 243
 244         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
 245                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
 246
 247         if (new_client) {
 248                 struct lvfs_run_ctxt saved;
 249                 loff_t off = fed->fed_lr_off;
 250                 void *handle;
 251                 int rc;
 252
 253                 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
 254                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
 255
 256                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 257                 /* Transaction needed to fix bug 1403 */
 258                 handle = fsfilt_start(obd,
 259                                       filter->fo_rcvd_filp->f_dentry->d_inode,
 260                                       FSFILT_OP_SETATTR, NULL);
 261                 if (IS_ERR(handle)) {
 262                         rc = PTR_ERR(handle);
 263                         CERROR("unable to start transaction: rc %d\n", rc);
 264                 } else {
 265                         rc = fsfilt_add_journal_cb(obd, 0, handle,
 266                                                    target_client_add_cb, exp);
 267                         if (rc == 0)
 268                                 exp->exp_need_sync = 1;
 269                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
 270                                                  fed->fed_fcd,
 271                                                  sizeof(*fed->fed_fcd),
 272                                                  &off, rc /* sync if no cb */);
 273                         fsfilt_commit(obd,
 274                                       filter->fo_rcvd_filp->f_dentry->d_inode,
 275                                       handle, 0);
 276                 }
 277                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 278
 279                 if (rc) {
 280                         CERROR("error writing %s client idx %u: rc %d\n",
 281                                LAST_RCVD, fed->fed_lr_idx, rc);
 282                         RETURN(rc);
 283                 }
 284         }
 285         RETURN(0);
 286 }
 287
 288 static int filter_client_free(struct obd_export *exp)
 289 {
 290         struct filter_export_data *fed = &exp->exp_filter_data;
 291         struct filter_obd *filter = &exp->exp_obd->u.filter;
 292         struct obd_device *obd = exp->exp_obd;
 293         struct filter_client_data zero_fcd;
 294         struct lvfs_run_ctxt saved;
 295         int rc;
 296         loff_t off;
 297         ENTRY;
 298
 299         if (fed->fed_fcd == NULL)
 300                 RETURN(0);
 301
 302         /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 303         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
 304                 GOTO(free, 0);
 305
 306         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
 307                fed->fed_lr_idx, off, fed->fed_fcd->fcd_uuid);
 308
 309         lprocfs_exp_cleanup(exp);
 310
 311         LASSERT(filter->fo_last_rcvd_slots != NULL);
 312
 313         off = fed->fed_lr_off;
 314
 315         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
 316          * we leak a client slot that will be cleaned on the next recovery. */
 317         if (off <= 0) {
 318                 CERROR("%s: client idx %d has med_off %lld\n",
 319                        obd->obd_name, fed->fed_lr_idx, off);
 320                 GOTO(free, rc = -EINVAL);
 321         }
 322
 323         /* Clear the bit _after_ zeroing out the client so we don't
 324            race with filter_client_add and zero out new clients.*/
 325         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
 326                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
 327                        fed->fed_lr_idx);
 328                 LBUG();
 329         }
 330
 331         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
 332                 memset(&zero_fcd, 0, sizeof zero_fcd);
 333                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 334                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
 335                                          sizeof(zero_fcd), &off,
 336                                          !exp->exp_libclient);
 337
 338                 if (rc == 0)
 339                         /* update server's transno */
 340                         filter_update_server_data(obd, filter->fo_rcvd_filp,
 341                                                   filter->fo_fsd,
 342                                                   !exp->exp_libclient);
 343                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 344
 345                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 346                        "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
 347                        fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
 348                        LAST_RCVD, rc);
 349         }
 350
 351         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
 352                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
 353                        fed->fed_lr_idx);
 354                 LBUG();
 355         }
 356
 357         EXIT;
 358 free:
 359         OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
 360         fed->fed_fcd = NULL;
 361
 362         return 0;
 363 }
 364
 365 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
 366 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
 367                                          struct filter_mod_data *fmd)
 368 {
 369         LASSERT_SPIN_LOCKED(&fed->fed_lock);
 370         if (--fmd->fmd_refcount == 0) {
 371                 /* XXX when we have persistent reservations and the handle
 372                  * is stored herein we need to drop it here. */
 373                 fed->fed_mod_count--;
 374                 list_del(&fmd->fmd_list);
 375                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
 376         }
 377 }
 378
 379 /* drop fmd reference, free it if last ref */
 380 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
 381 {
 382         struct filter_export_data *fed;
 383
 384         if (fmd == NULL)
 385                 return;
 386
 387         fed = &exp->exp_filter_data;
 388         spin_lock(&fed->fed_lock);
 389         filter_fmd_put_nolock(fed, fmd); /* caller reference */
 390         spin_unlock(&fed->fed_lock);
 391 }
 392
 393 /* expire entries from the end of the list if there are too many
 394  * or they are too old */
 395 static void filter_fmd_expire_nolock(struct filter_obd *filter,
 396                                      struct filter_export_data *fed,
 397                                      struct filter_mod_data *keep)
 398 {
 399         struct filter_mod_data *fmd, *tmp;
 400
 401         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
 402                 if (fmd == keep)
 403                         break;
 404
 405                 if (time_before(jiffies, fmd->fmd_expire) &&
 406                     fed->fed_mod_count < filter->fo_fmd_max_num)
 407                         break;
 408
 409                 list_del_init(&fmd->fmd_list);
 410                 filter_fmd_put_nolock(fed, fmd); /* list reference */
 411         }
 412 }
 413
 414 void filter_fmd_expire(struct obd_export *exp)
 415 {
 416         spin_lock(&exp->exp_filter_data.fed_lock);
 417         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
 418                                  &exp->exp_filter_data, NULL);
 419         spin_unlock(&exp->exp_filter_data.fed_lock);
 420 }
 421
 422 /* find specified objid, group in export fmd list.
 423  * caller must hold fed_lock and take fmd reference itself */
 424 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
 425                                                 struct filter_export_data *fed,
 426                                                 obd_id objid, obd_gr group)
 427 {
 428         struct filter_mod_data *found = NULL, *fmd;
 429
 430         LASSERT_SPIN_LOCKED(&fed->fed_lock);
 431
 432         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
 433                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
 434                         found = fmd;
 435                         list_del(&fmd->fmd_list);
 436                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
 437                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
 438                         break;
 439                 }
 440         }
 441
 442         filter_fmd_expire_nolock(filter, fed, found);
 443
 444         return found;
 445 }
 446
 447 /* Find fmd based on objid and group, or return NULL if not found. */
 448 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
 449                                         obd_id objid, obd_gr group)
 450 {
 451         struct filter_mod_data *fmd;
 452
 453         spin_lock(&exp->exp_filter_data.fed_lock);
 454         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
 455                                      &exp->exp_filter_data, objid, group);
 456         if (fmd)
 457                 fmd->fmd_refcount++;    /* caller reference */
 458         spin_unlock(&exp->exp_filter_data.fed_lock);
 459
 460         return fmd;
 461 }
 462
 463 /* Find fmd based on objid and group, or create a new one if none is found.
 464  * It is possible for this function to return NULL under memory pressure,
 465  * or if objid = 0 is passed (which will only cause old entries to expire).
 466  * Currently this is not fatal because any fmd state is transient and
 467  * may also be freed when it gets sufficiently old. */
 468 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
 469                                        obd_id objid, obd_gr group)
 470 {
 471         struct filter_export_data *fed = &exp->exp_filter_data;
 472         struct filter_mod_data *found = NULL, *fmd_new = NULL;
 473
 474         OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, SLAB_NOFS, sizeof(*fmd_new));
 475
 476         spin_lock(&fed->fed_lock);
 477         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
 478         if (fmd_new) {
 479                 if (found == NULL) {
 480                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
 481                         fmd_new->fmd_id = objid;
 482                         fmd_new->fmd_gr = group;
 483                         fmd_new->fmd_refcount++;   /* list reference */
 484                         found = fmd_new;
 485                         fed->fed_mod_count++;
 486                 } else {
 487                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
 488                 }
 489         }
 490         if (found) {
 491                 found->fmd_refcount++;          /* caller reference */
 492                 found->fmd_expire = jiffies +
 493                         exp->exp_obd->u.filter.fo_fmd_max_age;
 494         }
 495
 496         spin_unlock(&fed->fed_lock);
 497
 498         return found;
 499 }
 500
 501 #ifdef DO_FMD_DROP
 502 /* drop fmd list reference so it will disappear when last reference is put.
 503  * This isn't so critical because it would in fact only affect the one client
 504  * that is doing the unlink and at worst we have an stale entry referencing
 505  * an object that should never be used again. */
 506 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
 507 {
 508         struct filter_mod_data *found = NULL;
 509
 510         spin_lock(&exp->exp_filter_data.fed_lock);
 511         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
 512         if (found) {
 513                 list_del_init(&found->fmd_list);
 514                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
 515         }
 516         spin_unlock(&exp->exp_filter_data.fed_lock);
 517 }
 518 #else
 519 #define filter_fmd_drop(exp, objid, group)
 520 #endif
 521
 522 /* remove all entries from fmd list */
 523 static void filter_fmd_cleanup(struct obd_export *exp)
 524 {
 525         struct filter_export_data *fed = &exp->exp_filter_data;
 526         struct filter_mod_data *fmd = NULL, *tmp;
 527
 528         spin_lock(&fed->fed_lock);
 529         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
 530                 list_del_init(&fmd->fmd_list);
 531                 filter_fmd_put_nolock(fed, fmd);
 532         }
 533         spin_unlock(&fed->fed_lock);
 534 }
 535
 536 static int filter_init_export(struct obd_export *exp)
 537 {
 538         spin_lock_init(&exp->exp_filter_data.fed_lock);
 539         INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
 540         exp->exp_connecting = 1;
 541
 542         return 0;
 543 }
 544
 545 static int filter_free_server_data(struct filter_obd *filter)
 546 {
 547         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
 548         filter->fo_fsd = NULL;
 549         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
 550         filter->fo_last_rcvd_slots = NULL;
 551         return 0;
 552 }
 553
 554 /* assumes caller is already in kernel ctxt */
 555 int filter_update_server_data(struct obd_device *obd, struct file *filp,
 556                               struct lr_server_data *fsd, int force_sync)
 557 {
 558         loff_t off = 0;
 559         int rc;
 560         ENTRY;
 561
 562         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
 563         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
 564                le64_to_cpu(fsd->lsd_last_transno));
 565         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
 566                le64_to_cpu(fsd->lsd_mount_count));
 567
 568         fsd->lsd_compat14 = fsd->lsd_last_transno;
 569         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off,force_sync);
 570         if (rc)
 571                 CERROR("error writing lr_server_data: rc = %d\n", rc);
 572
 573         RETURN(rc);
 574 }
 575
 576 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
 577                              int force_sync)
 578 {
 579         struct filter_obd *filter = &obd->u.filter;
 580         __u64 tmp;
 581         loff_t off = 0;
 582         int rc;
 583         ENTRY;
 584
 585         if (filter->fo_last_objid_files[group] == NULL) {
 586                 CERROR("Object group "LPU64" not fully setup; not updating "
 587                        "last_objid\n", group);
 588                 RETURN(-EINVAL);
 589         }
 590
 591         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
 592                obd->obd_name, group, filter->fo_last_objids[group]);
 593
 594         tmp = cpu_to_le64(filter->fo_last_objids[group]);
 595         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
 596                                  &tmp, sizeof(tmp), &off, force_sync);
 597         if (rc)
 598                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
 599                        group, rc);
 600         RETURN(rc);
 601 }
 602
 603 /* assumes caller has already in kernel ctxt */
 604 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
 605 {
 606         struct filter_obd *filter = &obd->u.filter;
 607         struct lr_server_data *fsd;
 608         struct filter_client_data *fcd = NULL;
 609         struct inode *inode = filp->f_dentry->d_inode;
 610         unsigned long last_rcvd_size = inode->i_size;
 611         __u64 mount_count;
 612         int cl_idx;
 613         loff_t off = 0;
 614         int rc;
 615
 616         /* ensure padding in the struct is the correct size */
 617         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
 618                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
 619         CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
 620                  sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
 621
 622         OBD_ALLOC(fsd, sizeof(*fsd));
 623         if (!fsd)
 624                 RETURN(-ENOMEM);
 625         filter->fo_fsd = fsd;
 626
 627         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
 628         if (filter->fo_last_rcvd_slots == NULL) {
 629                 OBD_FREE(fsd, sizeof(*fsd));
 630                 RETURN(-ENOMEM);
 631         }
 632
 633         if (last_rcvd_size == 0) {
 634                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 635
 636                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
 637                 fsd->lsd_last_transno = 0;
 638                 mount_count = fsd->lsd_mount_count = 0;
 639                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
 640                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
 641                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
 642                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
 643                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
 644                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
 645         } else {
 646                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
 647                 if (rc) {
 648                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
 649                                LAST_RCVD, rc);
 650                         GOTO(err_fsd, rc);
 651                 }
 652                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 653                         LCONSOLE_ERROR("Trying to start OBD %s using the wrong"
 654                                        " disk %s. Were the /dev/ assignments "
 655                                        "rearranged?\n",
 656                                        obd->obd_uuid.uuid, fsd->lsd_uuid);
 657                         GOTO(err_fsd, rc = -EINVAL);
 658                 }
 659                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
 660                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
 661                 /* COMPAT_146 */
 662                 /* Assume old last_rcvd format unless I_C_LR is set */
 663                 if (!(fsd->lsd_feature_incompat &
 664                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
 665                         fsd->lsd_last_transno = fsd->lsd_compat14;
 666                 /* end COMPAT_146 */
 667         }
 668
 669         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
 670                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 671                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
 672                        ~FILTER_INCOMPAT_SUPP);
 673                 GOTO(err_fsd, rc = -EINVAL);
 674         }
 675         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
 676                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 677                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
 678                        ~FILTER_ROCOMPAT_SUPP);
 679                 /* Do something like remount filesystem read-only */
 680                 GOTO(err_fsd, rc = -EINVAL);
 681         }
 682
 683         CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
 684                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
 685         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 686                obd->obd_name, mount_count + 1);
 687         CDEBUG(D_INODE, "%s: server data size: %u\n",
 688                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
 689         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 690                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
 691         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 692                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
 693         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
 694                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
 695         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 696                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
 697                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
 698                 le16_to_cpu(fsd->lsd_client_size));
 699
 700         if (!obd->obd_replayable) {
 701                 CWARN("%s: recovery support OFF\n", obd->obd_name);
 702                 GOTO(out, rc = 0);
 703         }
 704
 705         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
 706              off < last_rcvd_size; cl_idx++) {
 707                 __u64 last_rcvd;
 708                 struct obd_export *exp;
 709                 struct filter_export_data *fed;
 710
 711                 if (!fcd) {
 712                         OBD_ALLOC(fcd, sizeof(*fcd));
 713                         if (!fcd)
 714                                 GOTO(err_client, rc = -ENOMEM);
 715                 }
 716
 717                 /* Don't assume off is incremented properly by
 718                  * fsfilt_read_record(), in case sizeof(*fcd)
 719                  * isn't the same as fsd->lsd_client_size.  */
 720                 off = le32_to_cpu(fsd->lsd_client_start) +
 721                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
 722                 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
 723                 if (rc) {
 724                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
 725                                LAST_RCVD, cl_idx, off, rc);
 726                         break; /* read error shouldn't cause startup to fail */
 727                 }
 728
 729                 if (fcd->fcd_uuid[0] == '\0') {
 730                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 731                                cl_idx);
 732                         continue;
 733                 }
 734
 735                 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
 736
 737                 /* These exports are cleaned up by filter_disconnect(), so they
 738                  * need to be set up like real exports as filter_connect() does.
 739                  */
 740                 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
 741                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 742                        " srv lr: "LPU64"\n", fcd->fcd_uuid, cl_idx,
 743                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno));
 744                 if (IS_ERR(exp)) {
 745                         if (PTR_ERR(exp) == -EALREADY) {
 746                                 /* export already exists, zero out this one */
 747                                 CERROR("Zeroing out duplicate export due to "
 748                                        "bug 10479.\n");
 749                                 fcd->fcd_uuid[0] = '\0';
 750                         } else {
 751                                 GOTO(err_client, rc = PTR_ERR(exp));
 752                         }
 753                 } else {
 754                         fed = &exp->exp_filter_data;
 755                         fed->fed_fcd = fcd;
 756                         rc = filter_client_add(obd, exp, cl_idx);
 757                         LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
 758
 759                         fcd = NULL;
 760                         exp->exp_replay_needed = 1;
 761                         exp->exp_connecting = 0;
 762                         obd->obd_recoverable_clients++;
 763                         obd->obd_max_recoverable_clients++;
 764                         class_export_put(exp);
 765                 }
 766
 767                 /* Need to check last_rcvd even for duplicated exports. */
 768                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
 769                        cl_idx, last_rcvd);
 770
 771                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
 772                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
 773         }
 774
 775         if (fcd)
 776                 OBD_FREE(fcd, sizeof(*fcd));
 777
 778         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
 779
 780         if (obd->obd_recoverable_clients) {
 781                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 782                       "last_rcvd "LPU64"\n", obd->obd_name,
 783                       obd->obd_recoverable_clients,
 784                       le64_to_cpu(fsd->lsd_last_transno));
 785                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 786                 obd->obd_recovering = 1;
 787                 obd->obd_recovery_start = CURRENT_SECONDS;
 788                 /* Only used for lprocfs_status */
 789                 obd->obd_recovery_end = obd->obd_recovery_start +
 790                         OBD_RECOVERY_TIMEOUT;
 791         }
 792
 793 out:
 794         filter->fo_mount_count = mount_count + 1;
 795         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
 796
 797         /* save it, so mount count and last_transno is current */
 798         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
 799         if (rc)
 800                 GOTO(err_client, rc);
 801
 802         RETURN(0);
 803
 804 err_client:
 805         class_disconnect_exports(obd);
 806 err_fsd:
 807         filter_free_server_data(filter);
 808         RETURN(rc);
 809 }
 810
 811 static int filter_cleanup_groups(struct obd_device *obd)
 812 {
 813         struct filter_obd *filter = &obd->u.filter;
 814         struct file *filp;
 815         struct dentry *dentry;
 816         int i;
 817         ENTRY;
 818
 819         if (filter->fo_dentry_O_groups != NULL) {
 820                 for (i = 0; i < FILTER_GROUPS; i++) {
 821                         dentry = filter->fo_dentry_O_groups[i];
 822                         if (dentry != NULL)
 823                                 f_dput(dentry);
 824                 }
 825                 OBD_FREE(filter->fo_dentry_O_groups,
 826                          FILTER_GROUPS * sizeof(*filter->fo_dentry_O_groups));
 827                 filter->fo_dentry_O_groups = NULL;
 828         }
 829         if (filter->fo_last_objid_files != NULL) {
 830                 for (i = 0; i < FILTER_GROUPS; i++) {
 831                         filp = filter->fo_last_objid_files[i];
 832                         if (filp != NULL)
 833                                 filp_close(filp, 0);
 834                 }
 835                 OBD_FREE(filter->fo_last_objid_files,
 836                          FILTER_GROUPS * sizeof(*filter->fo_last_objid_files));
 837                 filter->fo_last_objid_files = NULL;
 838         }
 839         if (filter->fo_dentry_O_sub != NULL) {
 840                 for (i = 0; i < filter->fo_subdir_count; i++) {
 841                         dentry = filter->fo_dentry_O_sub[i];
 842                         if (dentry != NULL)
 843                                 f_dput(dentry);
 844                 }
 845                 OBD_FREE(filter->fo_dentry_O_sub,
 846                          filter->fo_subdir_count *
 847                          sizeof(*filter->fo_dentry_O_sub));
 848                 filter->fo_dentry_O_sub = NULL;
 849         }
 850         if (filter->fo_last_objids != NULL) {
 851                 OBD_FREE(filter->fo_last_objids,
 852                          FILTER_GROUPS * sizeof(*filter->fo_last_objids));
 853                 filter->fo_last_objids = NULL;
 854         }
 855         if (filter->fo_dentry_O != NULL) {
 856                 f_dput(filter->fo_dentry_O);
 857                 filter->fo_dentry_O = NULL;
 858         }
 859         RETURN(0);
 860 }
 861
 862 /* FIXME: object groups */
 863 static int filter_prep_groups(struct obd_device *obd)
 864 {
 865         struct filter_obd *filter = &obd->u.filter;
 866         struct dentry *dentry, *O_dentry;
 867         struct file *filp;
 868         int i, rc = 0, cleanup_phase = 0;
 869         ENTRY;
 870
 871         O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
 872         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
 873         if (IS_ERR(O_dentry)) {
 874                 rc = PTR_ERR(O_dentry);
 875                 CERROR("cannot open/create O: rc = %d\n", rc);
 876                 GOTO(cleanup, rc);
 877         }
 878         filter->fo_dentry_O = O_dentry;
 879         cleanup_phase = 1; /* O_dentry */
 880
 881         OBD_ALLOC(filter->fo_last_objids, FILTER_GROUPS * sizeof(__u64));
 882         if (filter->fo_last_objids == NULL)
 883                 GOTO(cleanup, rc = -ENOMEM);
 884         cleanup_phase = 2; /* groups */
 885
 886         OBD_ALLOC(filter->fo_dentry_O_groups, FILTER_GROUPS * sizeof(dentry));
 887         if (filter->fo_dentry_O_groups == NULL)
 888                 GOTO(cleanup, rc = -ENOMEM);
 889         OBD_ALLOC(filter->fo_last_objid_files, FILTER_GROUPS * sizeof(filp));
 890         if (filter->fo_last_objid_files == NULL)
 891                 GOTO(cleanup, rc = -ENOMEM);
 892
 893         for (i = 0; i < FILTER_GROUPS; i++) {
 894                 char name[25];
 895                 loff_t off = 0;
 896
 897                 sprintf(name, "%d", i);
 898                 dentry = simple_mkdir(O_dentry, name, 0700, 1);
 899                 CDEBUG(D_INODE, "got/created O/%s: %p\n", name, dentry);
 900                 if (IS_ERR(dentry)) {
 901                         rc = PTR_ERR(dentry);
 902                         CERROR("cannot lookup/create O/%s: rc = %d\n",
 903                                name, rc);
 904                         GOTO(cleanup, rc);
 905                 }
 906                 filter->fo_dentry_O_groups[i] = dentry;
 907
 908                 sprintf(name, "O/%d/LAST_ID", i);
 909                 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
 910                 if (IS_ERR(filp)) {
 911                         rc = PTR_ERR(filp);
 912                         CERROR("cannot create %s: rc = %d\n", name, rc);
 913                         GOTO(cleanup, rc);
 914                 }
 915                 filter->fo_last_objid_files[i] = filp;
 916
 917                 if (filp->f_dentry->d_inode->i_size == 0) {
 918                         filter->fo_last_objids[i] = FILTER_INIT_OBJID;
 919                         rc = filter_update_last_objid(obd, i, 1);
 920                         if (rc)
 921                                 GOTO(cleanup, rc);
 922                         continue;
 923                 }
 924
 925                 rc = fsfilt_read_record(obd, filp, &filter->fo_last_objids[i],
 926                                         sizeof(__u64), &off);
 927                 if (rc) {
 928                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
 929                                name, rc);
 930                         GOTO(cleanup, rc);
 931                 }
 932                 filter->fo_last_objids[i] =
 933                         le64_to_cpu(filter->fo_last_objids[i]);
 934                 CDEBUG(D_HA, "%s: server last_objid group %d: "LPU64"\n",
 935                        obd->obd_name, i, filter->fo_last_objids[i]);
 936         }
 937
 938         if (filter->fo_subdir_count) {
 939                 O_dentry = filter->fo_dentry_O_groups[0];
 940                 OBD_ALLOC(filter->fo_dentry_O_sub,
 941                           filter->fo_subdir_count * sizeof(dentry));
 942                 if (filter->fo_dentry_O_sub == NULL)
 943                         GOTO(cleanup, rc = -ENOMEM);
 944
 945                 for (i = 0; i < filter->fo_subdir_count; i++) {
 946                         char dir[20];
 947                         snprintf(dir, sizeof(dir), "d%u", i);
 948
 949                         dentry = simple_mkdir(O_dentry, dir, 0700, 1);
 950                         CDEBUG(D_INODE, "got/created O/0/%s: %p\n", dir,dentry);
 951                         if (IS_ERR(dentry)) {
 952                                 rc = PTR_ERR(dentry);
 953                                 CERROR("can't lookup/create O/0/%s: rc = %d\n",
 954                                        dir, rc);
 955                                 GOTO(cleanup, rc);
 956                         }
 957                         filter->fo_dentry_O_sub[i] = dentry;
 958                 }
 959         }
 960         RETURN(0);
 961
 962  cleanup:
 963         filter_cleanup_groups(obd);
 964         return rc;
 965 }
 966
 967 /* setup the object store with correct subdirectories */
 968 static int filter_prep(struct obd_device *obd)
 969 {
 970         struct lvfs_run_ctxt saved;
 971         struct filter_obd *filter = &obd->u.filter;
 972         struct file *file;
 973         struct inode *inode;
 974         int rc = 0;
 975         ENTRY;
 976
 977         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 978         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
 979         if (!file || IS_ERR(file)) {
 980                 rc = PTR_ERR(file);
 981                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
 982                        LAST_RCVD, rc);
 983                 GOTO(out, rc);
 984         }
 985         filter->fo_rcvd_filp = file;
 986         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 987                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 988                        file->f_dentry->d_inode->i_mode);
 989                 GOTO(err_filp, rc = -ENOENT);
 990         }
 991
 992         inode = file->f_dentry->d_parent->d_inode;
 993         /* We use i_op->unlink directly in filter_vfs_unlink() */
 994         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
 995                 CERROR("%s: filesystem does not support create/unlink ops\n",
 996                        obd->obd_name);
 997                 GOTO(err_filp, rc = -EOPNOTSUPP);
 998         }
 999
1000         rc = filter_init_server_data(obd, file);
1001         if (rc) {
1002                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1003                 GOTO(err_filp, rc);
1004         }
1005         /* open and create health check io file*/
1006         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1007         if (IS_ERR(file)) {
1008                 rc = PTR_ERR(file);
1009                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1010                        HEALTH_CHECK, rc);
1011                 GOTO(err_filp, rc);
1012         }
1013         filter->fo_health_check_filp = file;
1014         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1015                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1016                        file->f_dentry->d_inode->i_mode);
1017                 GOTO(err_health_check, rc = -ENOENT);
1018         }
1019         rc = lvfs_check_io_health(obd, file);
1020         if (rc)
1021                 GOTO(err_health_check, rc);
1022
1023         rc = filter_prep_groups(obd);
1024         if (rc)
1025                 GOTO(err_server_data, rc);
1026  out:
1027         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1028
1029         return(rc);
1030
1031  err_server_data:
1032         //class_disconnect_exports(obd, 0);
1033         filter_free_server_data(filter);
1034  err_health_check:
1035         if (filp_close(filter->fo_health_check_filp, 0))
1036                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1037         filter->fo_health_check_filp = NULL;
1038  err_filp:
1039         if (filp_close(filter->fo_rcvd_filp, 0))
1040                 CERROR("can't close %s after error\n", LAST_RCVD);
1041         filter->fo_rcvd_filp = NULL;
1042         goto out;
1043 }
1044
1045 /* cleanup the filter: write last used object id to status file */
1046 static void filter_post(struct obd_device *obd)
1047 {
1048         struct lvfs_run_ctxt saved;
1049         struct filter_obd *filter = &obd->u.filter;
1050         int rc, i;
1051
1052         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1053          * best to start a transaction with h_sync, because we removed this
1054          * from lastobjid */
1055
1056         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1057         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1058                                        filter->fo_fsd, 0);
1059         if (rc)
1060                 CERROR("error writing server data: rc = %d\n", rc);
1061
1062         for (i = 0; i < FILTER_GROUPS; i++) {
1063                 rc = filter_update_last_objid(obd, i, (i == FILTER_GROUPS - 1));
1064                 if (rc)
1065                         CERROR("error writing group %d lastobjid: rc = %d\n",
1066                                i, rc);
1067         }
1068
1069         rc = filp_close(filter->fo_rcvd_filp, 0);
1070         filter->fo_rcvd_filp = NULL;
1071         if (rc)
1072                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1073
1074         rc = filp_close(filter->fo_health_check_filp, 0);
1075         filter->fo_health_check_filp = NULL;
1076         if (rc)
1077                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1078
1079         filter_cleanup_groups(obd);
1080         filter_free_server_data(filter);
1081         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1082 }
1083
1084 static void filter_set_last_id(struct filter_obd *filter,
1085                                obd_id id, obd_gr group)
1086 {
1087         LASSERT(filter->fo_fsd != NULL);
1088         LASSERT(group <= FILTER_GROUPS);
1089
1090         spin_lock(&filter->fo_objidlock);
1091         filter->fo_last_objids[group] = id;
1092         spin_unlock(&filter->fo_objidlock);
1093 }
1094
1095 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1096 {
1097         obd_id id;
1098         LASSERT(filter->fo_fsd != NULL);
1099         LASSERT(group <= FILTER_GROUPS);
1100
1101         /* FIXME: object groups */
1102         spin_lock(&filter->fo_objidlock);
1103         id = filter->fo_last_objids[group];
1104         spin_unlock(&filter->fo_objidlock);
1105
1106         return id;
1107 }
1108
1109 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1110 {
1111         LOCK_INODE_MUTEX(dparent->d_inode);
1112         return 0;
1113 }
1114
1115 /* We never dget the object parent, so DON'T dput it either */
1116 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1117 {
1118         struct filter_obd *filter = &obd->u.filter;
1119         LASSERT(group < FILTER_GROUPS); /* FIXME: object groups */
1120
1121         if (group > 0 || filter->fo_subdir_count == 0)
1122                 return filter->fo_dentry_O_groups[group];
1123
1124         return filter->fo_dentry_O_sub[objid & (filter->fo_subdir_count - 1)];
1125 }
1126
1127 /* We never dget the object parent, so DON'T dput it either */
1128 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1129                                   obd_id objid)
1130 {
1131         unsigned long now = jiffies;
1132         struct dentry *dparent = filter_parent(obd, group, objid);
1133         int rc;
1134
1135         if (IS_ERR(dparent))
1136                 return dparent;
1137
1138         rc = filter_lock_dentry(obd, dparent);
1139         fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1140         return rc ? ERR_PTR(rc) : dparent;
1141 }
1142
1143 /* We never dget the object parent, so DON'T dput it either */
1144 static void filter_parent_unlock(struct dentry *dparent)
1145 {
1146         UNLOCK_INODE_MUTEX(dparent->d_inode);
1147 }
1148
1149 /* How to get files, dentries, inodes from object id's.
1150  *
1151  * If dir_dentry is passed, the caller has already locked the parent
1152  * appropriately for this operation (normally a write lock).  If
1153  * dir_dentry is NULL, we do a read lock while we do the lookup to
1154  * avoid races with create/destroy and such changing the directory
1155  * internal to the filesystem code. */
1156 struct dentry *filter_fid2dentry(struct obd_device *obd,
1157                                  struct dentry *dir_dentry,
1158                                  obd_gr group, obd_id id)
1159 {
1160         struct dentry *dparent = dir_dentry;
1161         struct dentry *dchild;
1162         char name[32];
1163         int len;
1164         ENTRY;
1165
1166         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1167             !obd->u.filter.fo_destroy_in_progress) {
1168                 /* don't fail lookups for orphan recovery, it causes
1169                  * later LBUGs when objects still exist during precreate */
1170                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1171                 RETURN(ERR_PTR(-ENOENT));
1172         }
1173
1174         if (id == 0) {
1175                 CERROR("fatal: invalid object id 0\n");
1176                 RETURN(ERR_PTR(-ESTALE));
1177         }
1178
1179         len = sprintf(name, LPU64, id);
1180         if (dir_dentry == NULL) {
1181                 dparent = filter_parent_lock(obd, group, id);
1182                 if (IS_ERR(dparent)) {
1183                         CERROR("%s: error getting object "LPU64":"LPU64
1184                                " parent: rc %ld\n", obd->obd_name,
1185                                id, group, PTR_ERR(dparent));
1186                         RETURN(dparent);
1187                 }
1188         }
1189         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1190                dparent->d_name.len, dparent->d_name.name, name);
1191         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1192         if (dir_dentry == NULL)
1193                 filter_parent_unlock(dparent);
1194         if (IS_ERR(dchild)) {
1195                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1196                        PTR_ERR(dchild));
1197                 RETURN(dchild);
1198         }
1199
1200         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1201                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1202                        obd->obd_name, id, dchild->d_inode->i_ino);
1203                 f_dput(dchild);
1204                 RETURN(ERR_PTR(-ENOENT));
1205         }
1206
1207         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1208                name, dchild, atomic_read(&dchild->d_count));
1209
1210         LASSERT(atomic_read(&dchild->d_count) > 0);
1211
1212         RETURN(dchild);
1213 }
1214
1215 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid)
1216 {
1217         struct lustre_handle lockh;
1218         int flags = LDLM_AST_DISCARD_DATA, rc;
1219         struct ldlm_res_id res_id = { .name = { objid } };
1220         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1221
1222         ENTRY;
1223         /* Tell the clients that the object is gone now and that they should
1224          * throw away any cached pages. */
1225         rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id, LDLM_EXTENT,
1226                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1227                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1228                                     &lockh);
1229
1230         /* We only care about the side-effects, just drop the lock. */
1231         if (rc == ELDLM_OK)
1232                 ldlm_lock_decref(&lockh, LCK_PW);
1233
1234         RETURN(rc);
1235 }
1236
1237 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1238  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1239  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1240  * reverse.  Caller must take i_sem before starting the transaction and we
1241  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1242 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1243 {
1244         int rc;
1245         ENTRY;
1246
1247         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1248          * itself we already hold dir->i_mutex for child create/unlink ops */
1249         LASSERT(dentry->d_inode != NULL);
1250         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1251         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1252
1253
1254         /* may_delete() */
1255         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1256                 GOTO(out, rc = -ENOENT);
1257
1258         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1259         if (rc)
1260                 GOTO(out, rc);
1261
1262         if (IS_APPEND(dir))
1263                 GOTO(out, rc = -EPERM);
1264
1265         /* check_sticky() */
1266         if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1267             IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1268                 GOTO(out, rc = -EPERM);
1269
1270         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1271          *       that was done because of journal_start (which is already done
1272          *       here) or some other ordering issue. */
1273         DQUOT_INIT(dir);
1274
1275 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1276         rc = security_inode_unlink(dir, dentry);
1277         if (rc)
1278                 GOTO(out, rc);
1279 #endif
1280
1281         rc = dir->i_op->unlink(dir, dentry);
1282 out:
1283         /* need to drop i_mutex before we lose inode reference */
1284         UNLOCK_INODE_MUTEX(dentry->d_inode);
1285         if (rc == 0)
1286                 d_delete(dentry);
1287
1288         RETURN(rc);
1289 }
1290
1291 /* Caller must hold LCK_PW on parent and push us into kernel context.
1292  * Caller must hold child i_mutex, we drop it always.
1293  * Caller is also required to ensure that dchild->d_inode exists. */
1294 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1295                                    obd_gr group, struct dentry *dparent,
1296                                    struct dentry *dchild)
1297 {
1298         struct inode *inode = dchild->d_inode;
1299         int rc;
1300
1301         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1302                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1303                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1304                        (unsigned long)inode->i_nlink,
1305                        atomic_read(&inode->i_count));
1306         }
1307
1308         rc = filter_vfs_unlink(dparent->d_inode, dchild);
1309         if (rc)
1310                 CERROR("error unlinking objid %.*s: rc %d\n",
1311                        dchild->d_name.len, dchild->d_name.name, rc);
1312         return(rc);
1313 }
1314
1315 static int filter_intent_policy(struct ldlm_namespace *ns,
1316                                 struct ldlm_lock **lockp, void *req_cookie,
1317                                 ldlm_mode_t mode, int flags, void *data)
1318 {
1319         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1320         struct ptlrpc_request *req = req_cookie;
1321         struct ldlm_lock *lock = *lockp, *l = NULL;
1322         struct ldlm_resource *res = lock->l_resource;
1323         ldlm_processing_policy policy;
1324         struct ost_lvb *res_lvb, *reply_lvb;
1325         struct ldlm_reply *rep;
1326         struct list_head *tmp;
1327         ldlm_error_t err;
1328         int rc, tmpflags = 0, only_liblustre = 0;
1329         int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1330                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1331                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1332         ENTRY;
1333
1334         policy = ldlm_get_processing_policy(res);
1335         LASSERT(policy != NULL);
1336         LASSERT(req != NULL);
1337
1338         rc = lustre_pack_reply(req, 3, repsize, NULL);
1339         if (rc)
1340                 RETURN(req->rq_status = rc);
1341
1342         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1343         LASSERT(rep != NULL);
1344
1345         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1346                                    sizeof(*reply_lvb));
1347         LASSERT(reply_lvb != NULL);
1348
1349         //fixup_handle_for_resent_req(req, lock, &lockh);
1350
1351         /* If we grant any lock at all, it will be a whole-file read lock.
1352          * Call the extent policy function to see if our request can be
1353          * granted, or is blocked. */
1354         lock->l_policy_data.l_extent.start = 0;
1355         lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1356         lock->l_req_mode = LCK_PR;
1357
1358         LASSERT(ns == res->lr_namespace);
1359         lock_res(res);
1360         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1361         check_res_locked(res);
1362
1363         /* FIXME: we should change the policy function slightly, to not make
1364          * this list at all, since we just turn around and free it */
1365         while (!list_empty(&rpc_list)) {
1366                 struct ldlm_lock *wlock =
1367                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1368                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1369                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1370                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1371                 list_del_init(&wlock->l_cp_ast);
1372                 LDLM_LOCK_PUT(wlock);
1373         }
1374
1375         /* The lock met with no resistance; we're finished. */
1376         if (rc == LDLM_ITER_CONTINUE) {
1377                 /* do not grant locks to the liblustre clients: they cannot
1378                  * handle ASTs robustly.  We need to do this while still
1379                  * holding ns_lock to avoid the lock remaining on the res_link
1380                  * list (and potentially being added to l_pending_list by an
1381                  * AST) when we are going to drop this lock ASAP. */
1382                 if (lock->l_export->exp_libclient ||
1383                     OBD_FAIL_CHECK(OBD_FAIL_LDLM_GLIMPSE)) {
1384                         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
1385                         ldlm_resource_unlink_lock(lock);
1386                         err = ELDLM_LOCK_ABORTED;
1387                 } else {
1388                         err = ELDLM_LOCK_REPLACED;
1389                 }
1390                 unlock_res(res);
1391                 RETURN(err);
1392         }
1393
1394         /* Do not grant any lock, but instead send GL callbacks.  The extent
1395          * policy nicely created a list of all PW locks for us.  We will choose
1396          * the highest of those which are larger than the size in the LVB, if
1397          * any, and perform a glimpse callback. */
1398         res_lvb = res->lr_lvb_data;
1399         LASSERT(res_lvb != NULL);
1400         *reply_lvb = *res_lvb;
1401
1402         list_for_each(tmp, &res->lr_granted) {
1403                 struct ldlm_lock *tmplock =
1404                         list_entry(tmp, struct ldlm_lock, l_res_link);
1405
1406                 if (tmplock->l_granted_mode == LCK_PR)
1407                         continue;
1408                 /*
1409                  * ->ns_lock guarantees that no new locks are granted, and,
1410                  * therefore, that res->lr_lvb_data cannot increase beyond the
1411                  * end of already granted lock. As a result, it is safe to
1412                  * check against "stale" reply_lvb->lvb_size value without
1413                  * res->lr_lvb_sem.
1414                  */
1415                 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1416                         continue;
1417
1418                 /* Don't send glimpse ASTs to liblustre clients.  They aren't
1419                  * listening for them, and they do entirely synchronous I/O
1420                  * anyways. */
1421                 if (tmplock->l_export == NULL ||
1422                     tmplock->l_export->exp_libclient == 1) {
1423                         only_liblustre = 1;
1424                         continue;
1425                 }
1426
1427                 if (l == NULL) {
1428                         l = LDLM_LOCK_GET(tmplock);
1429                         continue;
1430                 }
1431
1432                 if (l->l_policy_data.l_extent.start >
1433                     tmplock->l_policy_data.l_extent.start)
1434                         continue;
1435
1436                 LDLM_LOCK_PUT(l);
1437                 l = LDLM_LOCK_GET(tmplock);
1438         }
1439         unlock_res(res);
1440
1441         /* There were no PW locks beyond the size in the LVB; finished. */
1442         if (l == NULL) {
1443                 if (only_liblustre) {
1444                         /* If we discovered a liblustre client with a PW lock,
1445                          * however, the LVB may be out of date!  The LVB is
1446                          * updated only on glimpse (which we don't do for
1447                          * liblustre clients) and cancel (which the client
1448                          * obviously has not yet done).  So if it has written
1449                          * data but kept the lock, the LVB is stale and needs
1450                          * to be updated from disk.
1451                          *
1452                          * Of course, this will all disappear when we switch to
1453                          * taking liblustre locks on the OST. */
1454                         if (ns->ns_lvbo && ns->ns_lvbo->lvbo_update)
1455                                 ns->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1456                 }
1457                 RETURN(ELDLM_LOCK_ABORTED);
1458         }
1459
1460         /*
1461          * This check is for lock taken in filter_prepare_destroy() that does
1462          * not have l_glimpse_ast set. So the logic is: if there is a lock
1463          * with no l_glimpse_ast set, this object is being destroyed already.
1464          *
1465          * Hence, if you are grabbing DLM locks on the server, always set
1466          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1467          */
1468         if (l->l_glimpse_ast == NULL) {
1469                 /* We are racing with unlink(); just return -ENOENT */
1470                 rep->lock_policy_res1 = -ENOENT;
1471                 goto out;
1472         }
1473
1474         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1475         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1476         /* Update the LVB from disk if the AST failed (this is a legal race) */
1477         /*
1478          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1479          * sending ast is not handled. This can result in lost client writes.
1480          */
1481         if (rc != 0 && ns->ns_lvbo && ns->ns_lvbo->lvbo_update)
1482                 ns->ns_lvbo->lvbo_update(res, NULL, 0, 1);
1483
1484         lock_res(res);
1485         *reply_lvb = *res_lvb;
1486         unlock_res(res);
1487
1488  out:
1489         LDLM_LOCK_PUT(l);
1490
1491         RETURN(ELDLM_LOCK_ABORTED);
1492 }
1493
1494 /*
1495  * per-obd_device iobuf pool.
1496  *
1497  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1498  * allocations in write-path has to be minimized (see bug 5137).
1499  *
1500  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1501  * OST threads (see ost_thread_{init,done}()).
1502  *
1503  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1504  * at the OST layer there are only (potentially) multiple obd_device of type
1505  * unknown at the time of OST thread creation.
1506  *
1507  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1508  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1509  * it's very own iobuf.
1510  *
1511  * Functions below
1512  *
1513  *     filter_kiobuf_pool_init()
1514  *
1515  *     filter_kiobuf_pool_done()
1516  *
1517  *     filter_iobuf_get()
1518  *
1519  * operate on this array. They are "generic" in a sense that they don't depend
1520  * on actual type of iobuf's (the latter depending on Linux kernel version).
1521  */
1522
1523 /*
1524  * destroy pool created by filter_iobuf_pool_init
1525  */
1526 static void filter_iobuf_pool_done(struct filter_obd *filter)
1527 {
1528         struct filter_iobuf **pool;
1529         int i;
1530
1531         ENTRY;
1532
1533         pool = filter->fo_iobuf_pool;
1534         if (pool != NULL) {
1535                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1536                         if (pool[i] != NULL)
1537                                 filter_free_iobuf(pool[i]);
1538                 }
1539                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1540                 filter->fo_iobuf_pool = NULL;
1541         }
1542         EXIT;
1543 }
1544
1545 /*
1546  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1547  */
1548 static int filter_iobuf_pool_init(struct filter_obd *filter)
1549 {
1550         void **pool;
1551
1552         ENTRY;
1553
1554         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1555                       GFP_KERNEL);
1556         if (filter->fo_iobuf_pool == NULL)
1557                 RETURN(-ENOMEM);
1558
1559         filter->fo_iobuf_count = OSS_THREADS_MAX;
1560
1561         RETURN(0);
1562 }
1563
1564 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1565  * many threads there will be so we allocate a large empty array and only
1566  * fill in those slots that are actually in use.
1567  * If we haven't allocated a pool entry for this thread before, do so now. */
1568 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1569 {
1570         int thread_id                    = oti ? oti->oti_thread_id : -1;
1571         struct filter_iobuf  *pool       = NULL;
1572         struct filter_iobuf **pool_place = NULL;
1573
1574         if (thread_id >= 0) {
1575                 LASSERT(thread_id < filter->fo_iobuf_count);
1576                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1577         }
1578
1579         if (unlikely(pool == NULL)) {
1580                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1581                                           PTLRPC_MAX_BRW_PAGES);
1582                 if (pool_place != NULL)
1583                         *pool_place = pool;
1584         }
1585
1586         return pool;
1587 }
1588
1589 /* mount the file system (secretly).  lustre_cfg parameters are:
1590  * 1 = device
1591  * 2 = fstype
1592  * 3 = flags: failover=f, failout=n
1593  * 4 = mount options
1594  */
1595 int filter_common_setup(struct obd_device *obd, obd_count len, void *buf,
1596                         void *option)
1597 {
1598         struct lustre_cfg* lcfg = buf;
1599         struct filter_obd *filter = &obd->u.filter;
1600         struct vfsmount *mnt;
1601         struct lustre_mount_info *lmi;
1602         struct obd_uuid uuid;
1603         __u8 *uuid_ptr;
1604         char *str, *label;
1605         char ns_name[48];
1606         int rc;
1607         ENTRY;
1608
1609         if (lcfg->lcfg_bufcount < 3 ||
1610             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1611             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1612                 RETURN(-EINVAL);
1613
1614         lmi = server_get_mount(obd->obd_name);
1615         if (lmi) {
1616                 /* We already mounted in lustre_fill_super.
1617                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1618                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1619                 mnt = lmi->lmi_mnt;
1620                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1621         } else {
1622                 /* old path - used by lctl */
1623                 CERROR("Using old MDS mount method\n");
1624                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1625                                     MS_NOATIME|MS_NODIRATIME,
1626                                     lustre_cfg_string(lcfg, 1), option);
1627                 if (IS_ERR(mnt)) {
1628                         rc = PTR_ERR(mnt);
1629                         LCONSOLE_ERROR("Can't mount disk %s (%d)\n",
1630                                        lustre_cfg_string(lcfg, 1), rc);
1631                         RETURN(rc);
1632                 }
1633
1634                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1635         }
1636         if (IS_ERR(obd->obd_fsops))
1637                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1638
1639         rc = filter_iobuf_pool_init(filter);
1640         if (rc != 0)
1641                 GOTO(err_ops, rc);
1642
1643         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1644
1645         /* failover is the default */
1646         obd->obd_replayable = 1;
1647
1648         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1649                 str = lustre_cfg_string(lcfg, 3);
1650                 if (strchr(str, 'n')) {
1651                         CWARN("%s: recovery disabled\n", obd->obd_name);
1652                         obd->obd_replayable = 0;
1653                 }
1654         }
1655
1656         filter->fo_vfsmnt = mnt;
1657         obd->u.obt.obt_sb = mnt->mnt_sb;
1658         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1659         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1660
1661         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1662         obd->obd_lvfs_ctxt.pwdmnt = mnt;
1663         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1664         obd->obd_lvfs_ctxt.fs = get_ds();
1665         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1666
1667         rc = filter_prep(obd);
1668         if (rc)
1669                 GOTO(err_ops, rc);
1670
1671         filter->fo_destroy_in_progress = 0;
1672         sema_init(&filter->fo_create_lock, 1);
1673         spin_lock_init(&filter->fo_translock);
1674         spin_lock_init(&filter->fo_objidlock);
1675         INIT_LIST_HEAD(&filter->fo_export_list);
1676         sema_init(&filter->fo_alloc_lock, 1);
1677         init_brw_stats(&filter->fo_filter_stats);
1678         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1679         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1680         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1681
1682         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1683         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
1684         if (obd->obd_namespace == NULL)
1685                 GOTO(err_post, rc = -ENOMEM);
1686         obd->obd_namespace->ns_lvbp = obd;
1687         obd->obd_namespace->ns_lvbo = &filter_lvbo;
1688         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1689
1690         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1691                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1692
1693         rc = llog_cat_initialize(obd, 1, NULL);
1694         if (rc) {
1695                 CERROR("failed to setup llogging subsystems\n");
1696                 GOTO(err_post, rc);
1697         }
1698
1699         rc = lquota_setup(filter_quota_interface_ref, obd);
1700         if (rc)
1701                 GOTO(err_post, rc);
1702
1703         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1704         if (uuid_ptr != NULL) {
1705                 class_uuid_unparse(uuid_ptr, &uuid);
1706                 str = uuid.uuid;
1707         } else {
1708                 str = "no UUID";
1709         }
1710
1711         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1712
1713         if (obd->obd_recovering) {
1714                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1715                               "recovery until %d %s reconnect, or if no clients"
1716                               " reconnect for %d:%.02d; during that time new "
1717                               "clients will not be allowed to connect. "
1718                               "Recovery progress can be monitored by watching "
1719                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1720                               obd->obd_name, lustre_cfg_string(lcfg, 1),
1721                               label ?: "", label ? "/" : "", str,
1722                               obd->obd_recoverable_clients,
1723                               (obd->obd_recoverable_clients == 1)
1724                               ? "client" : "clients",
1725                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
1726                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
1727                               obd->obd_name);
1728         } else {
1729                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
1730                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
1731                               label ?: "", label ? "/" : "", str,
1732                               obd->obd_replayable ? "enabled" : "disabled");
1733         }
1734
1735         RETURN(0);
1736
1737 err_post:
1738         filter_post(obd);
1739 err_ops:
1740         fsfilt_put_ops(obd->obd_fsops);
1741         filter_iobuf_pool_done(filter);
1742 err_mntput:
1743         server_put_mount(obd->obd_name, mnt);
1744         obd->u.obt.obt_sb = 0;
1745         return rc;
1746 }
1747
1748 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
1749 {
1750         struct lprocfs_static_vars lvars;
1751         struct lustre_cfg* lcfg = buf;
1752         unsigned long page;
1753         int rc;
1754
1755         CLASSERT(offsetof(struct obd_device, u.obt) ==
1756                  offsetof(struct obd_device, u.filter.fo_obt));
1757
1758         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
1759                 RETURN(-EINVAL);
1760
1761         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
1762         page = get_zeroed_page(GFP_KERNEL);
1763         if (!page)
1764                 RETURN(-ENOMEM);
1765
1766         memcpy((void *)page, lustre_cfg_buf(lcfg, 4),
1767                LUSTRE_CFG_BUFLEN(lcfg, 4));
1768         rc = filter_common_setup(obd, len, buf, (void *)page);
1769         free_page(page);
1770
1771         lprocfs_init_vars(filter, &lvars);
1772         if (rc == 0 && lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1773             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
1774                 /* Init obdfilter private stats here */
1775                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1776                                      LPROCFS_CNTR_AVGMINMAX,
1777                                      "read_bytes", "bytes");
1778                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1779                                      LPROCFS_CNTR_AVGMINMAX,
1780                                      "write_bytes", "bytes");
1781                 lproc_filter_attach_seqstat(obd);
1782                 obd->obd_proc_exports = proc_mkdir("exports",
1783                                                    obd->obd_proc_entry);
1784         }
1785
1786         return rc;
1787 }
1788
1789 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
1790 static struct llog_operations filter_size_orig_logops = {
1791         lop_setup: llog_obd_origin_setup,
1792         lop_cleanup: llog_obd_origin_cleanup,
1793         lop_add: llog_obd_origin_add
1794 };
1795
1796 static int filter_llog_init(struct obd_device *obd, struct obd_device *tgt,
1797                             int count, struct llog_catid *catid,
1798                             struct obd_uuid *uuid)
1799 {
1800         struct llog_ctxt *ctxt;
1801         int rc;
1802         ENTRY;
1803
1804         filter_mds_ost_repl_logops = llog_client_ops;
1805         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
1806         filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
1807         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
1808
1809         rc = llog_setup(obd, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
1810                         &filter_mds_ost_repl_logops);
1811         if (rc)
1812                 RETURN(rc);
1813
1814         /* FIXME - assign unlink_cb for filter's recovery */
1815         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
1816         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
1817
1818         rc = llog_setup(obd, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
1819                         &filter_size_orig_logops);
1820         RETURN(rc);
1821 }
1822
1823 static int filter_llog_finish(struct obd_device *obd, int count)
1824 {
1825         struct llog_ctxt *ctxt;
1826         int rc = 0, rc2 = 0;
1827         ENTRY;
1828
1829         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
1830         if (ctxt)
1831                 rc = llog_cleanup(ctxt);
1832
1833         ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
1834         if (ctxt)
1835                 rc2 = llog_cleanup(ctxt);
1836         if (!rc)
1837                 rc = rc2;
1838
1839         RETURN(rc);
1840 }
1841
1842 static int filter_precleanup(struct obd_device *obd,
1843                              enum obd_cleanup_stage stage)
1844 {
1845         int rc = 0;
1846         ENTRY;
1847
1848         switch(stage) {
1849         case OBD_CLEANUP_EARLY:
1850                 break;
1851         case OBD_CLEANUP_EXPORTS:
1852                 target_cleanup_recovery(obd);
1853                 break;
1854         case OBD_CLEANUP_SELF_EXP:
1855                 rc = filter_llog_finish(obd, 0);
1856                 break;
1857         case OBD_CLEANUP_OBD:
1858                 break;
1859         }
1860         RETURN(rc);
1861 }
1862
1863 static int filter_cleanup(struct obd_device *obd)
1864 {
1865         struct filter_obd *filter = &obd->u.filter;
1866         ENTRY;
1867
1868         if (obd->obd_fail)
1869                 LCONSOLE_WARN("%s: shutting down for failover; client state "
1870                               "will be preserved.\n", obd->obd_name);
1871
1872         if (!list_empty(&obd->obd_exports)) {
1873                 CERROR("%s: still has clients!\n", obd->obd_name);
1874                 class_disconnect_exports(obd);
1875                 if (!list_empty(&obd->obd_exports)) {
1876                         CERROR("still has exports after forced cleanup?\n");
1877                         RETURN(-EBUSY);
1878                 }
1879         }
1880
1881         lprocfs_obd_cleanup(obd);
1882         lprocfs_free_obd_stats(obd);
1883
1884         lquota_cleanup(filter_quota_interface_ref, obd);
1885
1886         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
1887
1888         if (obd->u.obt.obt_sb == NULL)
1889                 RETURN(0);
1890
1891         filter_post(obd);
1892
1893         shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
1894
1895         LL_DQUOT_OFF(obd->u.obt.obt_sb);
1896
1897         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
1898         obd->u.obt.obt_sb = NULL;
1899
1900         fsfilt_put_ops(obd->obd_fsops);
1901
1902         filter_iobuf_pool_done(filter);
1903
1904         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
1905
1906         RETURN(0);
1907 }
1908
1909 static int filter_connect_internal(struct obd_export *exp,
1910                                    struct obd_connect_data *data)
1911 {
1912         if (!data)
1913                 RETURN(0);
1914
1915         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
1916                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
1917                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
1918                data->ocd_connect_flags, data->ocd_version,
1919                data->ocd_grant, data->ocd_index);
1920
1921         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
1922         exp->exp_connect_flags = data->ocd_connect_flags;
1923         data->ocd_version = LUSTRE_VERSION_CODE;
1924
1925         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
1926                 obd_size left, want;
1927
1928                 spin_lock(&exp->exp_obd->obd_osfs_lock);
1929                 left = filter_grant_space_left(exp);
1930                 want = data->ocd_grant;
1931                 data->ocd_grant = filter_grant(exp, 0, want, left);
1932                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
1933
1934                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
1935                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
1936                        exp->exp_client_uuid.uuid, exp,
1937                        data->ocd_grant, want, left);
1938         }
1939
1940         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
1941                 struct filter_obd *filter = &exp->exp_obd->u.filter;
1942                 struct lr_server_data *lsd = filter->fo_fsd;
1943                 int index = le32_to_cpu(lsd->lsd_ost_index);
1944
1945                 if (!(lsd->lsd_feature_compat &
1946                       cpu_to_le32(OBD_COMPAT_OST))) {
1947                         /* this will only happen on the first connect */
1948                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
1949                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
1950                         filter_update_server_data(exp->exp_obd,
1951                                                   filter->fo_rcvd_filp, lsd, 1);
1952                 } else if (index != data->ocd_index) {
1953                         LCONSOLE_ERROR("Connection from %s to index %u doesn't "
1954                                        "match actual OST index %u in last_rcvd "
1955                                        "file, bad configuration?\n",
1956                                        obd_export_nid2str(exp), index,
1957                                        data->ocd_index);
1958                         RETURN(-EBADF);
1959                 }
1960         }
1961
1962         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
1963                 data->ocd_brw_size = 65536;
1964         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
1965                 data->ocd_brw_size = min(data->ocd_brw_size,
1966                                          (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
1967                 LASSERT(data->ocd_brw_size);
1968         }
1969
1970         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
1971          * FIXME: We don't strictly need the COMPAT flag for that,
1972          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
1973          * FIXME: We needed it for the index, as index 0 is valid. */
1974
1975         RETURN(0);
1976 }
1977
1978 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
1979                             struct obd_uuid *cluuid,
1980                             struct obd_connect_data *data)
1981 {
1982         int rc;
1983         ENTRY;
1984
1985         if (exp == NULL || obd == NULL || cluuid == NULL)
1986                 RETURN(-EINVAL);
1987
1988         rc = filter_connect_internal(exp, data);
1989
1990         RETURN(rc);
1991 }
1992
1993 /* nearly identical to mds_connect */
1994 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
1995                           struct obd_uuid *cluuid,
1996                           struct obd_connect_data *data)
1997 {
1998         struct obd_export *exp;
1999         struct filter_export_data *fed;
2000         struct filter_client_data *fcd = NULL;
2001         int rc;
2002         ENTRY;
2003
2004         if (conn == NULL || obd == NULL || cluuid == NULL)
2005                 RETURN(-EINVAL);
2006
2007         rc = class_connect(conn, obd, cluuid);
2008         if (rc)
2009                 RETURN(rc);
2010         exp = class_conn2export(conn);
2011         LASSERT(exp != NULL);
2012
2013         fed = &exp->exp_filter_data;
2014
2015         rc = filter_connect_internal(exp, data);
2016         if (rc)
2017                 GOTO(cleanup, rc);
2018
2019         if (!obd->obd_replayable)
2020                 GOTO(cleanup, rc = 0);
2021
2022         OBD_ALLOC(fcd, sizeof(*fcd));
2023         if (!fcd) {
2024                 CERROR("filter: out of memory for client data\n");
2025                 GOTO(cleanup, rc = -ENOMEM);
2026         }
2027
2028         memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2029         fed->fed_fcd = fcd;
2030
2031         rc = filter_client_add(obd, exp, -1);
2032
2033         GOTO(cleanup, rc);
2034
2035 cleanup:
2036         if (rc) {
2037                 if (fcd) {
2038                         OBD_FREE(fcd, sizeof(*fcd));
2039                         fed->fed_fcd = NULL;
2040                 }
2041                 class_disconnect(exp);
2042         } else {
2043                 class_export_put(exp);
2044         }
2045
2046         RETURN(rc);
2047 }
2048
2049 /* Do extra sanity checks for grant accounting.  We do this at connect,
2050  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2051  * always get rid of it or turn it off when we know accounting is good. */
2052 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2053 {
2054         struct filter_export_data *fed;
2055         struct obd_export *exp;
2056         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2057         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2058         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2059
2060         if (list_empty(&obd->obd_exports))
2061                 return;
2062
2063         spin_lock(&obd->obd_osfs_lock);
2064         spin_lock(&obd->obd_dev_lock);
2065         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2066                 int error = 0;
2067                 fed = &exp->exp_filter_data;
2068                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2069                     fed->fed_dirty < 0)
2070                         error = 1;
2071                 if (maxsize > 0) { /* we may not have done a statfs yet */
2072                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2073                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2074                                  exp->exp_client_uuid.uuid, exp,
2075                                  fed->fed_grant, fed->fed_pending, maxsize);
2076                         LASSERTF(fed->fed_dirty <= maxsize,
2077                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2078                                  exp->exp_client_uuid.uuid, exp,
2079                                  fed->fed_dirty, maxsize);
2080                 }
2081                 if (error)
2082                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2083                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2084                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2085                 else
2086                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2087                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2088                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2089                 tot_granted += fed->fed_grant + fed->fed_pending;
2090                 tot_pending += fed->fed_pending;
2091                 tot_dirty += fed->fed_dirty;
2092         }
2093         fo_tot_granted = obd->u.filter.fo_tot_granted;
2094         fo_tot_pending = obd->u.filter.fo_tot_pending;
2095         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2096         spin_unlock(&obd->obd_dev_lock);
2097         spin_unlock(&obd->obd_osfs_lock);
2098
2099         /* Do these assertions outside the spinlocks so we don't kill system */
2100         if (tot_granted != fo_tot_granted)
2101                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2102                        func, tot_granted, fo_tot_granted);
2103         if (tot_pending != fo_tot_pending)
2104                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2105                        func, tot_pending, fo_tot_pending);
2106         if (tot_dirty != fo_tot_dirty)
2107                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2108                        func, tot_dirty, fo_tot_dirty);
2109         if (tot_pending > tot_granted)
2110                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2111                        func, tot_pending, tot_granted);
2112         if (tot_granted > maxsize)
2113                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2114                        func, tot_granted, maxsize);
2115         if (tot_dirty > maxsize)
2116                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2117                        func, tot_dirty, maxsize);
2118 }
2119
2120 /* Remove this client from the grant accounting totals.  We also remove
2121  * the export from the obd device under the osfs and dev locks to ensure
2122  * that the filter_grant_sanity_check() calculations are always valid.
2123  * The client should do something similar when it invalidates its import. */
2124 static void filter_grant_discard(struct obd_export *exp)
2125 {
2126         struct obd_device *obd = exp->exp_obd;
2127         struct filter_obd *filter = &obd->u.filter;
2128         struct filter_export_data *fed = &exp->exp_filter_data;
2129
2130         spin_lock(&obd->obd_osfs_lock);
2131         spin_lock(&obd->obd_dev_lock);
2132         list_del_init(&exp->exp_obd_chain);
2133         spin_unlock(&obd->obd_dev_lock);
2134
2135         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2136                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2137                  obd->obd_name, filter->fo_tot_granted,
2138                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2139         filter->fo_tot_granted -= fed->fed_grant;
2140         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2141                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2142                  obd->obd_name, filter->fo_tot_pending,
2143                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2144         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2145         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2146                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2147                  obd->obd_name, filter->fo_tot_dirty,
2148                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2149         filter->fo_tot_dirty -= fed->fed_dirty;
2150         fed->fed_dirty = 0;
2151         fed->fed_grant = 0;
2152
2153         spin_unlock(&obd->obd_osfs_lock);
2154 }
2155
2156 static int filter_destroy_export(struct obd_export *exp)
2157 {
2158         ENTRY;
2159
2160         if (exp->exp_filter_data.fed_pending)
2161                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2162                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2163                        exp, exp->exp_filter_data.fed_pending);
2164
2165         target_destroy_export(exp);
2166
2167         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2168                 RETURN(0);
2169
2170         if (exp->exp_obd->obd_replayable)
2171                 filter_client_free(exp);
2172         else
2173                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2174
2175         filter_grant_discard(exp);
2176         filter_fmd_cleanup(exp);
2177
2178         if (!(exp->exp_flags & OBD_OPT_FORCE))
2179                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2180
2181         RETURN(0);
2182 }
2183
2184 /* also incredibly similar to mds_disconnect */
2185 static int filter_disconnect(struct obd_export *exp)
2186 {
2187         struct obd_device *obd = exp->exp_obd;
2188         struct llog_ctxt *ctxt;
2189         int rc, err;
2190         ENTRY;
2191
2192         LASSERT(exp);
2193         class_export_get(exp);
2194
2195         if (!(exp->exp_flags & OBD_OPT_FORCE))
2196                 filter_grant_sanity_check(obd, __FUNCTION__);
2197         filter_grant_discard(exp);
2198
2199         /* Disconnect early so that clients can't keep using export */
2200         rc = class_disconnect(exp);
2201         if (exp->exp_obd->obd_namespace != NULL)
2202                 ldlm_cancel_locks_for_export(exp);
2203
2204         /* flush any remaining cancel messages out to the target */
2205         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2206         err = llog_sync(ctxt, exp);
2207         if (err)
2208                 CERROR("error flushing logs to MDS: rc %d\n", err);
2209
2210         class_export_put(exp);
2211         RETURN(rc);
2212 }
2213
2214 static int filter_ping(struct obd_export *exp)
2215 {
2216         filter_fmd_expire(exp);
2217
2218         return 0;
2219 }
2220
2221 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2222                                   const char *what, int quiet)
2223 {
2224         struct dentry *dchild = NULL;
2225
2226         if (!(oa->o_valid & OBD_MD_FLGROUP))
2227                 oa->o_gr = 0;
2228
2229         dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
2230
2231         if (IS_ERR(dchild)) {
2232                 CERROR("%s error looking up object: "LPU64"\n",
2233                        what, oa->o_id);
2234                 RETURN(dchild);
2235         }
2236
2237         if (dchild->d_inode == NULL) {
2238                 if (!quiet)
2239                         CERROR("%s: %s on non-existent object: "LPU64"\n",
2240                                obd->obd_name, what, oa->o_id);
2241                 f_dput(dchild);
2242                 RETURN(ERR_PTR(-ENOENT));
2243         }
2244
2245         return dchild;
2246 }
2247
2248 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2249 {
2250         struct dentry *dentry = NULL;
2251         struct obd_device *obd;
2252         int rc = 0;
2253         ENTRY;
2254
2255         obd = class_exp2obd(exp);
2256         if (obd == NULL) {
2257                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2258                 RETURN(-EINVAL);
2259         }
2260
2261         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2262         if (IS_ERR(dentry))
2263                 RETURN(PTR_ERR(dentry));
2264
2265         /* Limit the valid bits in the return data to what we actually use */
2266         oinfo->oi_oa->o_valid = OBD_MD_FLID;
2267         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2268
2269         f_dput(dentry);
2270         RETURN(rc);
2271 }
2272
2273 /* this should be enabled/disabled in condition to enabled/disabled large
2274  * inodes (fast EAs) in backing store FS. */
2275 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2276                         void *handle, struct obdo *oa)
2277 {
2278         struct obd_device *obd = exp->exp_obd;
2279         int rc = 0;
2280         ENTRY;
2281
2282         if (oa->o_valid & OBD_MD_FLFID) {
2283                 struct filter_fid ff;
2284
2285                 if (!(oa->o_valid & OBD_MD_FLGROUP))
2286                         oa->o_gr = 0;
2287
2288                 /* packing fid and converting it to LE for storing into EA.
2289                  * Here ->o_stripe_idx should be filled by LOV and rest of
2290                  * fields - by client. */
2291                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2292                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2293                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2294                 ff.ff_objid = cpu_to_le64(oa->o_id);
2295                 ff.ff_group = cpu_to_le64(oa->o_gr);
2296
2297                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2298                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2299                        oa->o_generation, oa->o_id, oa->o_gr);
2300
2301                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2302                 if (rc)
2303                         CERROR("store fid in object failed! rc: %d\n", rc);
2304         } else {
2305                 CDEBUG(D_HA, "OSS object without fid info!\n");
2306         }
2307
2308         RETURN(rc);
2309 }
2310
2311 /* this is called from filter_truncate() until we have filter_punch() */
2312 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2313                             struct obdo *oa, struct obd_trans_info *oti)
2314 {
2315         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2316         struct llog_cookie *fcc = NULL;
2317         struct filter_obd *filter;
2318         int rc, err, locked = 0;
2319         unsigned int ia_valid;
2320         struct inode *inode;
2321         struct iattr iattr;
2322         void *handle;
2323         ENTRY;
2324
2325         LASSERT(dentry != NULL);
2326         LASSERT(!IS_ERR(dentry));
2327
2328         inode = dentry->d_inode;
2329         LASSERT(inode != NULL);
2330
2331         filter = &exp->exp_obd->u.filter;
2332         iattr_from_obdo(&iattr, oa, oa->o_valid);
2333         ia_valid = iattr.ia_valid;
2334
2335         if (oa->o_valid & OBD_MD_FLCOOKIE) {
2336                 OBD_ALLOC(fcc, sizeof(*fcc));
2337                 if (fcc != NULL)
2338                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2339         }
2340
2341         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2342                 DQUOT_INIT(inode);
2343                 LOCK_INODE_MUTEX(inode);
2344                 locked = 1;
2345         }
2346
2347         /* If the inode still has SUID+SGID bits set (see filter_precreate())
2348          * then we will accept the UID+GID sent by the client during write for
2349          * initializing the ownership of this inode.  We only allow this to
2350          * happen once so clear these bits in setattr. In 2.6 kernels it is
2351          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2352          * the flags that are actually being set. */
2353         if (ia_valid & (ATTR_UID | ATTR_GID)) {
2354                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2355                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2356
2357                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2358                         if (!(ia_valid & ATTR_MODE)) {
2359                                 iattr.ia_mode = inode->i_mode;
2360                                 iattr.ia_valid |= ATTR_MODE;
2361                         }
2362                         iattr.ia_mode &= ~S_ISUID;
2363                 }
2364                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2365                         if (!(iattr.ia_valid & ATTR_MODE)) {
2366                                 iattr.ia_mode = inode->i_mode;
2367                                 iattr.ia_valid |= ATTR_MODE;
2368                         }
2369                         iattr.ia_mode &= ~S_ISGID;
2370                 }
2371
2372                 orig_ids[USRQUOTA] = inode->i_uid;
2373                 orig_ids[GRPQUOTA] = inode->i_gid;
2374                 handle = fsfilt_start_log(exp->exp_obd, inode,
2375                                           FSFILT_OP_SETATTR, oti, 1);
2376
2377                 /* update inode EA only once when inode is suid bit marked. As
2378                  * on 2.6.x UID and GID may be set separately, we check here
2379                  * only one of them to avoid double setting. */
2380                 if (inode->i_mode & S_ISUID)
2381                         filter_update_fidea(exp, inode, handle, oa);
2382         } else {
2383                 handle = fsfilt_start(exp->exp_obd, inode,
2384                                       FSFILT_OP_SETATTR, oti);
2385         }
2386
2387         if (IS_ERR(handle))
2388                 GOTO(out_unlock, rc = PTR_ERR(handle));
2389
2390         if (oa->o_valid & OBD_MD_FLFLAGS) {
2391                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2392                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2393         } else {
2394                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2395                 if (fcc != NULL) {
2396                         /* set cancel cookie callback function */
2397                         if (fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2398                                                   filter_cancel_cookies_cb,
2399                                                   fcc))
2400                                 exp->exp_need_sync = 1;
2401                         else
2402                                 fcc = NULL;
2403                 }
2404         }
2405
2406         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2407                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2408
2409         /* The truncate might have used up our transaction credits.  Make
2410          * sure we have one left for the last_rcvd update. */
2411         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2412         rc = filter_finish_transno(exp, oti, rc);
2413
2414         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2415         if (err) {
2416                 CERROR("error on commit, err = %d\n", err);
2417                 if (!rc)
2418                         rc = err;
2419         }
2420
2421         if (locked) {
2422                 /* Let's flush truncated page on disk immediately, then we can
2423                  * avoid need to search for page aliases before directio writes
2424                  * and this sort of stuff at expense of somewhat slower
2425                  * truncates not on a page boundary. I believe this is the only
2426                  * place in filter code that can lead to pages getting to
2427                  * pagecache so far. */
2428                 filter_clear_truncated_page(inode);
2429                 UNLOCK_INODE_MUTEX(inode);
2430                 locked = 0;
2431         }
2432
2433         EXIT;
2434 out_unlock:
2435         if (locked)
2436                 UNLOCK_INODE_MUTEX(inode);
2437
2438         if (fcc)
2439                 OBD_FREE(fcc, sizeof(*fcc));
2440
2441         /* trigger quota release */
2442         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2443                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2444                 int rc2 = lquota_adjust(filter_quota_interface_ref, exp->exp_obd, cur_ids,
2445                                         orig_ids, rc, FSFILT_OP_SETATTR);
2446                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
2447                        "filter adjust qunit. (rc:%d)\n", rc2);
2448         }
2449         return rc;
2450 }
2451
2452 /* this is called from filter_truncate() until we have filter_punch() */
2453 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
2454                    struct obd_trans_info *oti)
2455 {
2456         struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id } };
2457         struct ldlm_valblock_ops *ns_lvbo;
2458         struct filter_mod_data *fmd;
2459         struct lvfs_run_ctxt saved;
2460         struct filter_obd *filter;
2461         struct ldlm_resource *res;
2462         struct dentry *dentry;
2463         int rc;
2464         ENTRY;
2465
2466         dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
2467                                     __FUNCTION__, 1);
2468         if (IS_ERR(dentry))
2469                 RETURN(PTR_ERR(dentry));
2470
2471         filter = &exp->exp_obd->u.filter;
2472         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
2473         lock_kernel();
2474
2475         if (oinfo->oi_oa->o_valid &
2476             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
2477                 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
2478                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
2479                         fmd->fmd_mactime_xid = oti->oti_xid;
2480                 filter_fmd_put(exp, fmd);
2481         }
2482
2483         /* setting objects attributes (including owner/group) */
2484         rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
2485         if (rc)
2486                 GOTO(out_unlock, rc);
2487
2488         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
2489                                 res_id, LDLM_EXTENT, 0);
2490
2491         if (res != NULL) {
2492                 ns_lvbo = res->lr_namespace->ns_lvbo;
2493                 if (ns_lvbo && ns_lvbo->lvbo_update)
2494                         rc = ns_lvbo->lvbo_update(res, NULL, 0, 0);
2495                 ldlm_resource_putref(res);
2496         }
2497
2498         oinfo->oi_oa->o_valid = OBD_MD_FLID;
2499
2500         /* Quota release need uid/gid info */
2501         obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
2502                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
2503
2504         EXIT;
2505 out_unlock:
2506         unlock_kernel();
2507         f_dput(dentry);
2508         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
2509         return rc;
2510 }
2511
2512 /* XXX identical to osc_unpackmd */
2513 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2514                            struct lov_mds_md *lmm, int lmm_bytes)
2515 {
2516         int lsm_size;
2517         ENTRY;
2518
2519         if (lmm != NULL) {
2520                 if (lmm_bytes < sizeof (*lmm)) {
2521                         CERROR("lov_mds_md too small: %d, need %d\n",
2522                                lmm_bytes, (int)sizeof(*lmm));
2523                         RETURN(-EINVAL);
2524                 }
2525                 /* XXX LOV_MAGIC etc check? */
2526
2527                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
2528                         CERROR("lov_mds_md: zero lmm_object_id\n");
2529                         RETURN(-EINVAL);
2530                 }
2531         }
2532
2533         lsm_size = lov_stripe_md_size(1);
2534         if (lsmp == NULL)
2535                 RETURN(lsm_size);
2536
2537         if (*lsmp != NULL && lmm == NULL) {
2538                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
2539                 OBD_FREE(*lsmp, lsm_size);
2540                 *lsmp = NULL;
2541                 RETURN(0);
2542         }
2543
2544         if (*lsmp == NULL) {
2545                 OBD_ALLOC(*lsmp, lsm_size);
2546                 if (*lsmp == NULL)
2547                         RETURN(-ENOMEM);
2548                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
2549                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
2550                         OBD_FREE(*lsmp, lsm_size);
2551                         RETURN(-ENOMEM);
2552                 }
2553                 loi_init((*lsmp)->lsm_oinfo[0]);
2554         }
2555
2556         if (lmm != NULL) {
2557                 /* XXX zero *lsmp? */
2558                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
2559                 LASSERT((*lsmp)->lsm_object_id);
2560         }
2561
2562         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
2563
2564         RETURN(lsm_size);
2565 }
2566
2567 /* caller must hold fo_create_lock */
2568 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
2569                                       struct filter_obd *filter)
2570 {
2571         struct obdo doa; /* XXX obdo on stack */
2572         obd_id last, id;
2573         int rc;
2574         ENTRY;
2575
2576         LASSERT(oa);
2577         LASSERT(down_trylock(&filter->fo_create_lock) != 0);
2578
2579         memset(&doa, 0, sizeof(doa));
2580         if (oa->o_valid & OBD_MD_FLGROUP) {
2581                 doa.o_valid |= OBD_MD_FLGROUP;
2582                 doa.o_gr = oa->o_gr;
2583         } else {
2584                 doa.o_gr = 0;
2585         }
2586         doa.o_mode = S_IFREG;
2587
2588         if (!filter->fo_destroy_in_progress) {
2589                 CERROR("%s: destroy_in_progress already cleared\n",
2590                         exp->exp_obd->obd_name);
2591                 RETURN(0);
2592         }
2593
2594         last = filter_last_id(filter, doa.o_gr);
2595         CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
2596                exp->exp_obd->obd_name, oa->o_id + 1, last);
2597         for (id = last; id > oa->o_id; id--) {
2598                 doa.o_id = id;
2599                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
2600                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
2601                         CEMERG("error destroying precreate objid "LPU64": %d\n",
2602                                id, rc);
2603                 filter_set_last_id(filter, id - 1, doa.o_gr);
2604                 /* update last_id on disk periodically so that if we restart
2605                  * we don't need to re-scan all of the just-deleted objects. */
2606                 if ((id & 511) == 0)
2607                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
2608         }
2609
2610         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
2611                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
2612
2613         rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
2614         filter->fo_destroy_in_progress = 0;
2615
2616         RETURN(rc);
2617 }
2618
2619 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
2620                             obd_gr group, int *num);
2621 /* returns a negative error or a nonnegative number of files to create */
2622 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
2623                                    obd_gr group, struct obd_trans_info *oti)
2624 {
2625         struct obd_device *obd = exp->exp_obd;
2626         struct filter_obd *filter = &obd->u.filter;
2627         int diff, rc;
2628         ENTRY;
2629
2630         /* delete orphans request */
2631         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
2632                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
2633                         CERROR("%s: dropping old orphan cleanup request\n",
2634                                obd->obd_name);
2635                         RETURN(0);
2636                 }
2637
2638                 /* This causes inflight precreates to abort and drop lock */
2639                 filter->fo_destroy_in_progress = 1;
2640                 down(&filter->fo_create_lock);
2641                 diff = oa->o_id - filter_last_id(filter, group);
2642                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
2643                        filter_last_id(filter, group), diff);
2644
2645                 if (-diff > OST_MAX_PRECREATE) {
2646                         CERROR("%s: ignoring bogus orphan destroy request: "
2647                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
2648                                oa->o_id, filter_last_id(filter, group));
2649                         /* FIXME: should reset precreate_next_id on MDS */
2650                         GOTO(out, rc = -EINVAL);
2651                 }
2652                 if (diff < 0) {
2653                         rc = filter_destroy_precreated(exp, oa, filter);
2654                         if (rc)
2655                                 CERROR("%s: unable to write lastobjid, but "
2656                                        "orphans were deleted\n", obd->obd_name);
2657                         GOTO(out, rc);
2658                 } else {
2659                         /*XXX used by MDS for the first time! */
2660                         filter->fo_destroy_in_progress = 0;
2661                 }
2662         } else {
2663                 down(&filter->fo_create_lock);
2664                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
2665                         CERROR("%s: dropping old precreate request\n",
2666                                obd->obd_name);
2667                         GOTO(out, rc = 0);
2668                 }
2669                 /* only precreate if group == 0 and o_id is specfied */
2670                 if (group != 0 || oa->o_id == 0)
2671                         diff = 1;
2672                 else
2673                         diff = oa->o_id - filter_last_id(filter, group);
2674                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
2675                        filter_last_id(filter, group), diff);
2676
2677                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
2678                          oa->o_id, filter_last_id(filter, group), diff);
2679         }
2680
2681         if (diff > 0) {
2682                 oa->o_id = filter_last_id(&obd->u.filter, group);
2683                 rc = filter_precreate(obd, oa, group, &diff);
2684                 oa->o_id = filter_last_id(&obd->u.filter, group);
2685                 oa->o_valid = OBD_MD_FLID;
2686                 GOTO(out, rc);
2687         }
2688         /* else diff == 0 */
2689         GOTO(out, rc = 0);
2690 out:
2691         up(&filter->fo_create_lock);
2692         return rc;
2693 }
2694
2695 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2696                          __u64 max_age)
2697 {
2698         struct filter_obd *filter = &obd->u.filter;
2699         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
2700         int rc;
2701         ENTRY;
2702
2703         /* at least try to account for cached pages.  its still racey and
2704          * might be under-reporting if clients haven't announced their
2705          * caches with brw recently */
2706         spin_lock(&obd->obd_osfs_lock);
2707         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
2708         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
2709         spin_unlock(&obd->obd_osfs_lock);
2710
2711         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
2712                " pending "LPU64" free "LPU64" avail "LPU64"\n",
2713                filter->fo_tot_dirty, filter->fo_tot_granted,
2714                filter->fo_tot_pending,
2715                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
2716
2717         filter_grant_sanity_check(obd, __FUNCTION__);
2718
2719         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
2720                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
2721                                  osfs->os_bsize - 1) >> blockbits));
2722
2723         /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
2724          * stop creating files on MDS if OST is not good shape to create
2725          * objects.*/
2726         osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
2727                 EROFS : 0;
2728         RETURN(rc);
2729 }
2730
2731 /* We rely on the fact that only one thread will be creating files in a given
2732  * group at a time, which is why we don't need an atomic filter_get_new_id.
2733  * Even if we had that atomic function, the following race would exist:
2734  *
2735  * thread 1: gets id x from filter_next_id
2736  * thread 2: gets id (x + 1) from filter_next_id
2737  * thread 2: creates object (x + 1)
2738  * thread 1: tries to create object x, gets -ENOSPC
2739  *
2740  * Caller must hold fo_create_lock
2741  */
2742 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
2743                             obd_gr group, int *num)
2744 {
2745         struct dentry *dchild = NULL, *dparent = NULL;
2746         struct filter_obd *filter;
2747         int err = 0, rc = 0, recreate_obj = 0, i;
2748         unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
2749         obd_id next_id;
2750         void *handle = NULL;
2751         ENTRY;
2752
2753         filter = &obd->u.filter;
2754
2755         LASSERT(down_trylock(&filter->fo_create_lock) != 0);
2756
2757         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
2758             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
2759                 recreate_obj = 1;
2760         } else {
2761                 struct obd_statfs *osfs;
2762
2763                 OBD_ALLOC(osfs, sizeof(*osfs));
2764                 if (osfs == NULL)
2765                         RETURN(-ENOMEM);
2766                 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
2767                 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
2768                         CDEBUG(D_HA,"%s: not enough space for create "LPU64"\n",
2769                                obd->obd_name, osfs->os_bavail <<
2770                                filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
2771                         *num = 0;
2772                         rc = -ENOSPC;
2773                 }
2774                 OBD_FREE(osfs, sizeof(*osfs));
2775                 if (rc)
2776                         RETURN(rc);
2777         }
2778
2779         CDEBUG(D_HA, "%s: precreating %d objects in group "LPU64" at "LPU64"\n",
2780                obd->obd_name, *num, group, oa->o_id);
2781
2782         for (i = 0; i < *num && err == 0; i++) {
2783                 int cleanup_phase = 0;
2784
2785                 if (filter->fo_destroy_in_progress) {
2786                         CWARN("%s: precreate aborted by destroy\n",
2787                               obd->obd_name);
2788                         rc = -EAGAIN;
2789                         break;
2790                 }
2791
2792                 if (recreate_obj) {
2793                         __u64 last_id;
2794                         next_id = oa->o_id;
2795                         last_id = filter_last_id(filter, group);
2796                         if (next_id > last_id) {
2797                                 CERROR("Error: Trying to recreate obj greater"
2798                                        "than last id "LPD64" > "LPD64"\n",
2799                                        next_id, last_id);
2800                                 GOTO(cleanup, rc = -EINVAL);
2801                         }
2802                 } else
2803                         next_id = filter_last_id(filter, group) + 1;
2804
2805                 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
2806
2807                 dparent = filter_parent_lock(obd, group, next_id);
2808                 if (IS_ERR(dparent))
2809                         GOTO(cleanup, rc = PTR_ERR(dparent));
2810                 cleanup_phase = 1;
2811
2812                 dchild = filter_fid2dentry(obd, dparent, group, next_id);
2813                 if (IS_ERR(dchild))
2814                         GOTO(cleanup, rc = PTR_ERR(dchild));
2815                 cleanup_phase = 2;
2816
2817                 if (dchild->d_inode != NULL) {
2818                         /* This would only happen if lastobjid was bad on disk*/
2819                         /* Could also happen if recreating missing obj but
2820                          * already exists
2821                          */
2822                         if (recreate_obj) {
2823                                 CERROR("%s: recreating existing object %.*s?\n",
2824                                        obd->obd_name, dchild->d_name.len,
2825                                        dchild->d_name.name);
2826                         } else {
2827                                 CERROR("%s: Serious error: objid %.*s already "
2828                                        "exists; is this filesystem corrupt?\n",
2829                                        obd->obd_name, dchild->d_name.len,
2830                                        dchild->d_name.name);
2831                                 LBUG();
2832                         }
2833                         GOTO(cleanup, rc = -EEXIST);
2834                 }
2835
2836                 handle = fsfilt_start_log(obd, dparent->d_inode,
2837                                           FSFILT_OP_CREATE, NULL, 1);
2838                 if (IS_ERR(handle))
2839                         GOTO(cleanup, rc = PTR_ERR(handle));
2840                 cleanup_phase = 3;
2841
2842                 /* We mark object SUID+SGID to flag it for accepting UID+GID
2843                  * from client on first write.  Currently the permission bits
2844                  * on the OST are never used, so this is OK. */
2845                 rc = ll_vfs_create(dparent->d_inode, dchild,
2846                                    S_IFREG |  S_ISUID | S_ISGID | 0666, NULL);
2847                 if (rc) {
2848                         CERROR("create failed rc = %d\n", rc);
2849                         GOTO(cleanup, rc);
2850                 }
2851
2852                 if (!recreate_obj) {
2853                         filter_set_last_id(filter, next_id, group);
2854                         err = filter_update_last_objid(obd, group, 0);
2855                         if (err)
2856                                 CERROR("unable to write lastobjid "
2857                                        "but file created\n");
2858                 }
2859
2860         cleanup:
2861                 switch(cleanup_phase) {
2862                 case 3:
2863                         err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
2864                         if (err) {
2865                                 CERROR("error on commit, err = %d\n", err);
2866                                 if (!rc)
2867                                         rc = err;
2868                         }
2869                 case 2:
2870                         f_dput(dchild);
2871                 case 1:
2872                         filter_parent_unlock(dparent);
2873                 case 0:
2874                         break;
2875                 }
2876
2877                 if (rc)
2878                         break;
2879                 if (time_after(jiffies, enough_time)) {
2880                         CDEBUG(D_HA, "%s: precreate slow - want %d got %d \n",
2881                                obd->obd_name, *num, i);
2882                         break;
2883                 }
2884         }
2885         *num = i;
2886
2887         CDEBUG(D_HA, "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
2888                obd->obd_name, i, group, filter->fo_last_objids[group], rc);
2889
2890         RETURN(rc);
2891 }
2892
2893 static int filter_create(struct obd_export *exp, struct obdo *oa,
2894                          struct lov_stripe_md **ea, struct obd_trans_info *oti)
2895 {
2896         struct obd_device *obd = NULL;
2897         struct lvfs_run_ctxt saved;
2898         struct lov_stripe_md *lsm = NULL;
2899         int rc = 0;
2900         ENTRY;
2901
2902         if (!(oa->o_valid & OBD_MD_FLGROUP))
2903                 oa->o_gr = 0;
2904
2905         CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
2906                oa->o_gr, oa->o_id);
2907         if (ea != NULL) {
2908                 lsm = *ea;
2909                 if (lsm == NULL) {
2910                         rc = obd_alloc_memmd(exp, &lsm);
2911                         if (rc < 0)
2912                                 RETURN(rc);
2913                 }
2914         }
2915
2916         obd = exp->exp_obd;
2917         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2918
2919         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
2920             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
2921                 if (oa->o_id > filter_last_id(&obd->u.filter, oa->o_gr)) {
2922                         CERROR("recreate objid "LPU64" > last id "LPU64"\n",
2923                                oa->o_id, filter_last_id(&obd->u.filter,
2924                                                         oa->o_gr));
2925                         rc = -EINVAL;
2926                 } else {
2927                         struct filter_obd *filter = &obd->u.filter;
2928                         int diff = 1;
2929
2930                         down(&filter->fo_create_lock);
2931                         rc = filter_precreate(obd, oa, oa->o_gr, &diff);
2932                         up(&filter->fo_create_lock);
2933                 }
2934         } else {
2935                 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
2936         }
2937
2938         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2939         if (rc && ea != NULL && *ea != lsm) {
2940                 obd_free_memmd(exp, &lsm);
2941         } else if (rc == 0 && ea != NULL) {
2942                 /* XXX LOV STACKING: the lsm that is passed to us from
2943                  * LOV does not have valid lsm_oinfo data structs, so
2944                  * don't go touching that.  This needs to be fixed in a
2945                  * big way. */
2946                 lsm->lsm_object_id = oa->o_id;
2947                 *ea = lsm;
2948         }
2949
2950         RETURN(rc);
2951 }
2952
2953 int filter_destroy(struct obd_export *exp, struct obdo *oa,
2954                    struct lov_stripe_md *md, struct obd_trans_info *oti,
2955                    struct obd_export *md_exp)
2956 {
2957         unsigned int qcids[MAXQUOTAS] = {0, 0};
2958         struct obd_device *obd;
2959         struct filter_obd *filter;
2960         struct dentry *dchild = NULL, *dparent = NULL;
2961         struct lvfs_run_ctxt saved;
2962         void *handle = NULL;
2963         struct llog_cookie *fcc = NULL;
2964         int rc, rc2, cleanup_phase = 0;
2965         struct iattr iattr;
2966         ENTRY;
2967
2968         if (!(oa->o_valid & OBD_MD_FLGROUP))
2969                 oa->o_gr = 0;
2970
2971         obd = exp->exp_obd;
2972         filter = &obd->u.filter;
2973
2974         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2975         cleanup_phase = 1;
2976
2977         dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
2978         if (IS_ERR(dchild))
2979                 GOTO(cleanup, rc = PTR_ERR(dchild));
2980         cleanup_phase = 2;
2981
2982         if (dchild->d_inode == NULL) {
2983                 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
2984                        oa->o_id);
2985                 /* If object already gone, cancel cookie right now */
2986                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
2987                         fcc = obdo_logcookie(oa);
2988                         llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
2989                                     NULL, 1, fcc, 0);
2990                         fcc = NULL; /* we didn't allocate fcc, don't free it */
2991                 }
2992                 GOTO(cleanup, rc = -ENOENT);
2993         }
2994
2995         filter_prepare_destroy(obd, oa->o_id);
2996
2997         /* Our MDC connection is established by the MDS to us */
2998         if (oa->o_valid & OBD_MD_FLCOOKIE) {
2999                 OBD_ALLOC(fcc, sizeof(*fcc));
3000                 if (fcc != NULL)
3001                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3002         }
3003         DQUOT_INIT(dchild->d_inode);
3004
3005         /* we're gonna truncate it first in order to avoid possible deadlock:
3006          *      P1                      P2
3007          * open trasaction      open transaction
3008          * down(i_zombie)       down(i_zombie)
3009          *                      restart transaction
3010          * (see BUG 4180) -bzzz
3011          */
3012         LOCK_INODE_MUTEX(dchild->d_inode);
3013         handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3014                                   NULL, 1);
3015         if (IS_ERR(handle)) {
3016                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3017                 GOTO(cleanup, rc = PTR_ERR(handle));
3018         }
3019
3020         iattr.ia_valid = ATTR_SIZE;
3021         iattr.ia_size = 0;
3022         rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3023         rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3024         UNLOCK_INODE_MUTEX(dchild->d_inode);
3025         if (rc)
3026                 GOTO(cleanup, rc);
3027         if (rc2)
3028                 GOTO(cleanup, rc = rc2);
3029
3030         /* We don't actually need to lock the parent until we are unlinking
3031          * here, and not while truncating above.  That avoids holding the
3032          * parent lock for a long time during truncate, which can block other
3033          * threads from doing anything to objects in that directory. bug 7171 */
3034         dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3035         if (IS_ERR(dparent))
3036                 GOTO(cleanup, rc = PTR_ERR(dparent));
3037         cleanup_phase = 3; /* filter_parent_unlock */
3038
3039         LOCK_INODE_MUTEX(dchild->d_inode);
3040         handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3041         if (IS_ERR(handle)) {
3042                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3043                 GOTO(cleanup, rc = PTR_ERR(handle));
3044         }
3045         cleanup_phase = 4; /* fsfilt_commit */
3046
3047         /* Quota release need uid/gid of inode */
3048         obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3049
3050         filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3051
3052         /* this drops dchild->d_inode->i_mutex unconditionally */
3053         rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3054
3055         EXIT;
3056 cleanup:
3057         switch(cleanup_phase) {
3058         case 4:
3059                 if (fcc != NULL) {
3060                         if (fsfilt_add_journal_cb(obd, 0, oti ?
3061                                                   oti->oti_handle : handle,
3062                                                   filter_cancel_cookies_cb,
3063                                                   fcc) == 0)
3064                                 fcc = NULL;
3065                 }
3066                 rc = filter_finish_transno(exp, oti, rc);
3067                 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3068                 if (rc2) {
3069                         CERROR("error on commit, err = %d\n", rc2);
3070                         if (!rc)
3071                                 rc = rc2;
3072                 }
3073         case 3:
3074                 filter_parent_unlock(dparent);
3075         case 2:
3076                 f_dput(dchild);
3077                 if (fcc != NULL)
3078                         OBD_FREE(fcc, sizeof(*fcc));
3079         case 1:
3080                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3081                 break;
3082         default:
3083                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3084                 LBUG();
3085         }
3086
3087         /* trigger quota release */
3088         qcids[USRQUOTA] = oa->o_uid;
3089         qcids[GRPQUOTA] = oa->o_gid;
3090         rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3091                             FSFILT_OP_UNLINK);
3092         CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3093                "filter adjust qunit! (rc:%d)\n", rc2);
3094         return rc;
3095 }
3096
3097 /* NB start and end are used for punch, but not truncate */
3098 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3099                            struct obd_trans_info *oti,
3100                            struct ptlrpc_request_set *rqset)
3101 {
3102         int rc;
3103         ENTRY;
3104
3105         if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3106                 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3107                        oinfo->oi_policy.l_extent.end);
3108                 RETURN(-EFAULT);
3109         }
3110
3111         CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3112                ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3113                oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3114
3115         oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3116         rc = filter_setattr(exp, oinfo, oti);
3117         RETURN(rc);
3118 }
3119
3120 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3121                        struct lov_stripe_md *lsm, obd_off start, obd_off end)
3122 {
3123         struct lvfs_run_ctxt saved;
3124         struct filter_obd *filter;
3125         struct dentry *dentry;
3126         struct llog_ctxt *ctxt;
3127         int rc, rc2;
3128         ENTRY;
3129
3130         filter = &exp->exp_obd->u.filter;
3131
3132         /* an objid of zero is taken to mean "sync whole filesystem" */
3133         if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3134                 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3135                 /* flush any remaining cancel messages out to the target */
3136                 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3137                 llog_sync(ctxt, exp);
3138                 RETURN(rc);
3139         }
3140
3141         dentry = filter_oa2dentry(exp->exp_obd, oa);
3142         if (IS_ERR(dentry))
3143                 RETURN(PTR_ERR(dentry));
3144
3145         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3146
3147         LOCK_INODE_MUTEX(dentry->d_inode);
3148
3149         rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3150         if (rc == 0) {
3151                 /* just any file to grab fsync method - "file" arg unused */
3152                 struct file *file = filter->fo_rcvd_filp;
3153
3154                 if (file->f_op && file->f_op->fsync)
3155                         rc = file->f_op->fsync(NULL, dentry, 1);
3156
3157                 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3158                 if (!rc)
3159                         rc = rc2;
3160         }
3161         UNLOCK_INODE_MUTEX(dentry->d_inode);
3162
3163         oa->o_valid = OBD_MD_FLID;
3164         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3165
3166         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3167
3168         f_dput(dentry);
3169         RETURN(rc);
3170 }
3171
3172 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3173                            void *key, __u32 *vallen, void *val)
3174 {
3175         struct obd_device *obd;
3176         ENTRY;
3177
3178         obd = class_exp2obd(exp);
3179         if (obd == NULL) {
3180                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3181                 RETURN(-EINVAL);
3182         }
3183
3184         if (keylen == strlen("blocksize") &&
3185             memcmp(key, "blocksize", keylen) == 0) {
3186                 __u32 *blocksize = val;
3187                 *vallen = sizeof(*blocksize);
3188                 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3189                 RETURN(0);
3190         }
3191
3192         if (keylen == strlen("blocksize_bits") &&
3193             memcmp(key, "blocksize_bits", keylen) == 0) {
3194                 __u32 *blocksize_bits = val;
3195                 *vallen = sizeof(*blocksize_bits);
3196                 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3197                 RETURN(0);
3198         }
3199
3200         if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
3201                 obd_id *last_id = val;
3202                 /* FIXME: object groups */
3203                 *last_id = filter_last_id(&obd->u.filter, 0);
3204                 RETURN(0);
3205         }
3206         CDEBUG(D_IOCTL, "invalid key\n");
3207         RETURN(-EINVAL);
3208 }
3209
3210 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3211                                  void *key, __u32 vallen, void *val,
3212                                  struct ptlrpc_request_set *set)
3213 {
3214         struct obd_device *obd;
3215         struct llog_ctxt *ctxt;
3216         int rc = 0;
3217         ENTRY;
3218
3219         obd = exp->exp_obd;
3220         if (obd == NULL) {
3221                 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3222                 RETURN(-EINVAL);
3223         }
3224
3225         if (keylen < strlen(KEY_MDS_CONN) ||
3226             memcmp(key, KEY_MDS_CONN, keylen) != 0)
3227                 RETURN(-EINVAL);
3228
3229         LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3230                       obd_export_nid2str(exp));
3231         obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3232
3233         /* setup llog imports */
3234         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
3235         rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3236
3237         lquota_setinfo(filter_quota_interface_ref, exp, obd);
3238
3239         RETURN(rc);
3240 }
3241
3242 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3243                      int len, void *karg, void *uarg)
3244 {
3245         struct obd_device *obd = exp->exp_obd;
3246         struct obd_ioctl_data *data = karg;
3247         int rc = 0;
3248
3249         switch (cmd) {
3250         case OBD_IOC_ABORT_RECOVERY: {
3251                 CERROR("aborting recovery for device %s\n", obd->obd_name);
3252                 target_abort_recovery(obd);
3253                 RETURN(0);
3254         }
3255
3256         case OBD_IOC_SYNC: {
3257                 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3258                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3259                 RETURN(rc);
3260         }
3261
3262         case OBD_IOC_SET_READONLY: {
3263                 void *handle;
3264                 struct super_block *sb = obd->u.obt.obt_sb;
3265                 struct inode *inode = sb->s_root->d_inode;
3266                 BDEVNAME_DECLARE_STORAGE(tmp);
3267                 CERROR("*** setting device %s read-only ***\n",
3268                        ll_bdevname(sb, tmp));
3269
3270                 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3271                 if (!IS_ERR(handle))
3272                         rc = fsfilt_commit(obd, inode, handle, 1);
3273
3274                 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3275                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3276
3277                 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
3278                 RETURN(0);
3279         }
3280
3281         case OBD_IOC_CATLOGLIST: {
3282                 rc = llog_catalog_list(obd, 1, data);
3283                 RETURN(rc);
3284         }
3285
3286         case OBD_IOC_LLOG_CANCEL:
3287         case OBD_IOC_LLOG_REMOVE:
3288         case OBD_IOC_LLOG_INFO:
3289         case OBD_IOC_LLOG_PRINT: {
3290                 /* FIXME to be finished */
3291                 RETURN(-EOPNOTSUPP);
3292 /*
3293                 struct llog_ctxt *ctxt = NULL;
3294
3295                 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3296                 rc = llog_ioctl(ctxt, cmd, data);
3297                 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3298
3299                 RETURN(rc);
3300 */
3301         }
3302
3303
3304         default:
3305                 RETURN(-EINVAL);
3306         }
3307         RETURN(0);
3308 }
3309
3310 static int filter_health_check(struct obd_device *obd)
3311 {
3312         struct filter_obd *filter = &obd->u.filter;
3313         int rc = 0;
3314
3315         /*
3316          * health_check to return 0 on healthy
3317          * and 1 on unhealthy.
3318          */
3319         if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3320                 rc = 1;
3321
3322         LASSERT(filter->fo_health_check_filp != NULL);
3323         rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3324
3325         return rc;
3326 }
3327
3328 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
3329                                              void *data)
3330 {
3331         return filter_fid2dentry(data, NULL, gr, id);
3332 }
3333
3334 static int filter_process_config(struct obd_device *obd, obd_count len, void *buf)
3335 {
3336         struct lustre_cfg *lcfg = buf;
3337         struct lprocfs_static_vars lvars;
3338         int rc = 0;
3339
3340         lprocfs_init_vars(filter, &lvars);
3341
3342         rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
3343
3344         return(rc);
3345 }
3346
3347 static struct lvfs_callback_ops filter_lvfs_ops = {
3348         l_fid2dentry:     filter_lvfs_fid2dentry,
3349 };
3350
3351 static struct obd_ops filter_obd_ops = {
3352         .o_owner          = THIS_MODULE,
3353         .o_get_info       = filter_get_info,
3354         .o_set_info_async = filter_set_info_async,
3355         .o_setup          = filter_setup,
3356         .o_precleanup     = filter_precleanup,
3357         .o_cleanup        = filter_cleanup,
3358         .o_connect        = filter_connect,
3359         .o_reconnect      = filter_reconnect,
3360         .o_disconnect     = filter_disconnect,
3361         .o_ping           = filter_ping,
3362         .o_init_export    = filter_init_export,
3363         .o_destroy_export = filter_destroy_export,
3364         .o_statfs         = filter_statfs,
3365         .o_getattr        = filter_getattr,
3366         .o_unpackmd       = filter_unpackmd,
3367         .o_create         = filter_create,
3368         .o_setattr        = filter_setattr,
3369         .o_destroy        = filter_destroy,
3370         .o_brw            = filter_brw,
3371         .o_punch          = filter_truncate,
3372         .o_sync           = filter_sync,
3373         .o_preprw         = filter_preprw,
3374         .o_commitrw       = filter_commitrw,
3375         .o_llog_init      = filter_llog_init,
3376         .o_llog_finish    = filter_llog_finish,
3377         .o_iocontrol      = filter_iocontrol,
3378         .o_health_check   = filter_health_check,
3379         .o_process_config = filter_process_config,
3380 };
3381
3382 quota_interface_t *filter_quota_interface_ref;
3383 extern quota_interface_t filter_quota_interface;
3384
3385 static int __init obdfilter_init(void)
3386 {
3387         struct lprocfs_static_vars lvars;
3388         int rc;
3389
3390         printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
3391
3392         lprocfs_init_vars(filter, &lvars);
3393
3394         request_module("lquota");
3395         OBD_ALLOC(obdfilter_created_scratchpad,
3396                   OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3397                   sizeof(*obdfilter_created_scratchpad));
3398         if (obdfilter_created_scratchpad == NULL)
3399                 return -ENOMEM;
3400
3401         ll_fmd_cachep = kmem_cache_create("ll_fmd_cache",
3402                                          sizeof(struct filter_mod_data),
3403                                          0, 0, NULL, NULL);
3404         if (!ll_fmd_cachep)
3405                 GOTO(out, rc = -ENOMEM);
3406
3407         filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
3408         init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
3409
3410         rc = class_register_type(&filter_obd_ops, lvars.module_vars,
3411                                  LUSTRE_OST_NAME);
3412         if (rc) {
3413                 int err;
3414
3415                 err = kmem_cache_destroy(ll_fmd_cachep);
3416                 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
3417                 ll_fmd_cachep = NULL;
3418 out:
3419                 if (filter_quota_interface_ref)
3420                         PORTAL_SYMBOL_PUT(filter_quota_interface);
3421
3422                 OBD_FREE(obdfilter_created_scratchpad,
3423                          OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3424                          sizeof(*obdfilter_created_scratchpad));
3425         }
3426
3427         return rc;
3428 }
3429
3430 static void __exit obdfilter_exit(void)
3431 {
3432         if (filter_quota_interface_ref)
3433                 PORTAL_SYMBOL_PUT(filter_quota_interface);
3434
3435         if (ll_fmd_cachep) {
3436                 int rc = kmem_cache_destroy(ll_fmd_cachep);
3437                 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
3438                 ll_fmd_cachep = NULL;
3439         }
3440
3441         class_unregister_type(LUSTRE_OST_NAME);
3442         OBD_FREE(obdfilter_created_scratchpad,
3443                  OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
3444                  sizeof(*obdfilter_created_scratchpad));
3445 }
3446
3447 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3448 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
3449 MODULE_LICENSE("GPL");
3450
3451 module_init(obdfilter_init);
3452 module_exit(obdfilter_exit);