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 #ifndef AUTOCONF_INCLUDED
  39 #include <linux/config.h>
  40 #endif
  41 #include <linux/module.h>
  42 #include <linux/fs.h>
  43 #include <linux/dcache.h>
  44 #include <linux/init.h>
  45 #include <linux/version.h>
  46 #include <linux/sched.h>
  47 #include <linux/mount.h>
  48 #include <linux/buffer_head.h>
  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 /* Group 0 is no longer a legal group, to catch uninitialized IDs */
  66 #define FILTER_MIN_GROUPS FILTER_GROUP_MDS0
  67 static struct lvfs_callback_ops filter_lvfs_ops;
  68 cfs_mem_cache_t *ll_fmd_cachep;
  69
  70 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
  71                              void *cb_data, int error)
  72 {
  73         obd_transno_commit_cb(obd, transno, error);
  74 }
  75
  76 /* Assumes caller has already pushed us into the kernel context. */
  77 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
  78                           int rc, int force_sync)
  79 {
  80         struct filter_obd *filter = &exp->exp_obd->u.filter;
  81         struct filter_export_data *fed = &exp->exp_filter_data;
  82         struct filter_client_data *fcd = fed->fed_fcd;
  83         __u64 last_rcvd;
  84         loff_t off;
  85         int err, log_pri = D_RPCTRACE;
  86
  87         /* Propagate error code. */
  88         if (rc)
  89                 RETURN(rc);
  90
  91         if (!exp->exp_obd->obd_replayable || oti == NULL)
  92                 RETURN(rc);
  93
  94         /* we don't allocate new transnos for replayed requests */
  95         if (oti->oti_transno == 0) {
  96                 spin_lock(&filter->fo_translock);
  97                 last_rcvd = le64_to_cpu(filter->fo_fsd->lsd_last_transno) + 1;
  98                 filter->fo_fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
  99                 spin_unlock(&filter->fo_translock);
 100                 oti->oti_transno = last_rcvd;
 101         } else {
 102                 spin_lock(&filter->fo_translock);
 103                 last_rcvd = oti->oti_transno;
 104                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->lsd_last_transno))
 105                         filter->fo_fsd->lsd_last_transno =
 106                                 cpu_to_le64(last_rcvd);
 107                 spin_unlock(&filter->fo_translock);
 108         }
 109         fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
 110
 111         /* could get xid from oti, if it's ever needed */
 112         fcd->fcd_last_xid = 0;
 113
 114         off = fed->fed_lr_off;
 115         if (off <= 0) {
 116                 CERROR("%s: client idx %d is %lld\n", exp->exp_obd->obd_name,
 117                        fed->fed_lr_idx, fed->fed_lr_off);
 118                 err = -EINVAL;
 119         } else {
 120                 if (!force_sync)
 121                         force_sync = fsfilt_add_journal_cb(exp->exp_obd,
 122                                                            last_rcvd,
 123                                                            oti->oti_handle,
 124                                                            filter_commit_cb,
 125                                                            NULL);
 126
 127                 err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp,
 128                                           fcd, sizeof(*fcd), &off,
 129                                           force_sync | exp->exp_need_sync);
 130                 if (force_sync)
 131                         filter_commit_cb(exp->exp_obd, last_rcvd, NULL, err);
 132         }
 133         if (err) {
 134                 log_pri = D_ERROR;
 135                 if (rc == 0)
 136                         rc = err;
 137         }
 138
 139         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
 140                last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
 141
 142         RETURN(rc);
 143 }
 144
 145 void f_dput(struct dentry *dentry)
 146 {
 147         /* Can't go inside filter_ddelete because it can block */
 148         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
 149                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
 150         LASSERT(atomic_read(&dentry->d_count) > 0);
 151
 152         dput(dentry);
 153 }
 154
 155 static void init_brw_stats(struct brw_stats *brw_stats)
 156 {
 157         int i;
 158         for (i = 0; i < BRW_LAST; i++)
 159                 spin_lock_init(&brw_stats->hist[i].oh_lock);
 160 }
 161
 162 /* brw_stats are 2128, ops are 3916, ldlm are 204, so 6248 bytes per client,
 163    plus the procfs overhead :( */
 164 static int filter_export_stats_init(struct obd_device *obd,
 165                                     struct obd_export *exp)
 166 {
 167         struct filter_export_data *fed = &exp->exp_filter_data;
 168         struct proc_dir_entry *brw_entry;
 169         int rc, num_stats;
 170         ENTRY;
 171
 172         init_brw_stats(&fed->fed_brw_stats);
 173
 174         if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
 175                 /* Self-export gets no proc entry */
 176                 RETURN(0);
 177
 178         rc = lprocfs_exp_setup(exp);
 179         if (rc)
 180                 RETURN(rc);
 181
 182         /* Create a per export proc entry for brw_stats */
 183         brw_entry = create_proc_entry("brw_stats", 0644, exp->exp_proc);
 184         if (brw_entry == NULL)
 185                RETURN(-ENOMEM);
 186         brw_entry->proc_fops = &filter_per_export_stats_fops;
 187         brw_entry->data = fed;
 188
 189         /* Create a per export proc entry for ops stats */
 190         num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
 191                      LPROC_FILTER_LAST - 1;
 192         exp->exp_ops_stats = lprocfs_alloc_stats(num_stats,
 193                                                  LPROCFS_STATS_FLAG_NOPERCPU);
 194         if (exp->exp_ops_stats == NULL)
 195               RETURN(-ENOMEM);
 196         lprocfs_init_ops_stats(LPROC_FILTER_LAST, exp->exp_ops_stats);
 197         lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_READ_BYTES,
 198                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
 199         lprocfs_counter_init(exp->exp_ops_stats, LPROC_FILTER_WRITE_BYTES,
 200                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
 201         lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
 202         RETURN(0);
 203 }
 204
 205 /* Add client data to the FILTER.  We use a bitmap to locate a free space
 206  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
 207  * Otherwise, we have just read the data from the last_rcvd file and
 208  * we know its offset. */
 209 static int filter_client_add(struct obd_device *obd, struct obd_export *exp,
 210                              int cl_idx)
 211 {
 212         struct filter_obd *filter = &obd->u.filter;
 213         struct filter_export_data *fed = &exp->exp_filter_data;
 214         unsigned long *bitmap = filter->fo_last_rcvd_slots;
 215         int new_client = (cl_idx == -1);
 216
 217         ENTRY;
 218
 219         LASSERT(bitmap != NULL);
 220         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
 221
 222         /* Self-export */
 223         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid) == 0)
 224                 RETURN(0);
 225
 226         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 227          * there's no need for extra complication here
 228          */
 229         if (new_client) {
 230                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 231         repeat:
 232                 if (cl_idx >= LR_MAX_CLIENTS) {
 233                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
 234                                cl_idx);
 235                         RETURN(-EOVERFLOW);
 236                 }
 237                 if (test_and_set_bit(cl_idx, bitmap)) {
 238                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
 239                                                     cl_idx);
 240                         goto repeat;
 241                 }
 242         } else {
 243                 if (test_and_set_bit(cl_idx, bitmap)) {
 244                         CERROR("FILTER client %d: bit already set in bitmap!\n",
 245                                cl_idx);
 246                         LBUG();
 247                 }
 248         }
 249
 250         fed->fed_lr_idx = cl_idx;
 251         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->lsd_client_start) +
 252                 cl_idx * le16_to_cpu(filter->fo_fsd->lsd_client_size);
 253         LASSERTF(fed->fed_lr_off > 0, "fed_lr_off = %llu\n", fed->fed_lr_off);
 254
 255         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
 256                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
 257
 258         if (new_client) {
 259                 struct lvfs_run_ctxt saved;
 260                 loff_t off = fed->fed_lr_off;
 261                 int rc;
 262                 void *handle;
 263
 264                 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
 265                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
 266
 267                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 268                 /* Transaction needed to fix bug 1403 */
 269                 handle = fsfilt_start(obd,
 270                                       filter->fo_rcvd_filp->f_dentry->d_inode,
 271                                       FSFILT_OP_SETATTR, NULL);
 272                 if (IS_ERR(handle)) {
 273                         rc = PTR_ERR(handle);
 274                         CERROR("unable to start transaction: rc %d\n", rc);
 275                 } else {
 276                         rc = fsfilt_add_journal_cb(obd, 0, handle,
 277                                                    target_client_add_cb, exp);
 278                         if (rc == 0) {
 279                                 spin_lock(&exp->exp_lock);
 280                                 exp->exp_need_sync = 1;
 281                                 spin_unlock(&exp->exp_lock);
 282                         }
 283                         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp,
 284                                                  fed->fed_fcd,
 285                                                  sizeof(*fed->fed_fcd),
 286                                                  &off, rc /* sync if no cb */);
 287                         fsfilt_commit(obd,
 288                                       filter->fo_rcvd_filp->f_dentry->d_inode,
 289                                       handle, 0);
 290                 }
 291                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 292
 293                 if (rc) {
 294                         CERROR("error writing %s client idx %u: rc %d\n",
 295                                LAST_RCVD, fed->fed_lr_idx, rc);
 296                         RETURN(rc);
 297                 }
 298         }
 299         RETURN(0);
 300 }
 301
 302 static int filter_client_free(struct obd_export *exp)
 303 {
 304         struct filter_export_data *fed = &exp->exp_filter_data;
 305         struct filter_obd *filter = &exp->exp_obd->u.filter;
 306         struct obd_device *obd = exp->exp_obd;
 307         struct filter_client_data zero_fcd;
 308         struct lvfs_run_ctxt saved;
 309         int rc;
 310         loff_t off;
 311         ENTRY;
 312
 313         if (fed->fed_fcd == NULL)
 314                 RETURN(0);
 315
 316         /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 317         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
 318                 GOTO(free, 0);
 319
 320         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
 321                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
 322
 323         LASSERT(filter->fo_last_rcvd_slots != NULL);
 324
 325         off = fed->fed_lr_off;
 326
 327         /* Don't clear fed_lr_idx here as it is likely also unset.  At worst
 328          * we leak a client slot that will be cleaned on the next recovery. */
 329         if (off <= 0) {
 330                 CERROR("%s: client idx %d has med_off %lld\n",
 331                        obd->obd_name, fed->fed_lr_idx, off);
 332                 GOTO(free, rc = -EINVAL);
 333         }
 334
 335         /* Clear the bit _after_ zeroing out the client so we don't
 336            race with filter_client_add and zero out new clients.*/
 337         if (!test_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
 338                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
 339                        fed->fed_lr_idx);
 340                 LBUG();
 341         }
 342
 343         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
 344                 memset(&zero_fcd, 0, sizeof zero_fcd);
 345                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 346                 rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
 347                                          sizeof(zero_fcd), &off,
 348                                          (!exp->exp_libclient ||
 349                                           exp->exp_need_sync));
 350                 if (rc == 0)
 351                         /* update server's transno */
 352                         filter_update_server_data(obd, filter->fo_rcvd_filp,
 353                                                   filter->fo_fsd,
 354                                                   !exp->exp_libclient);
 355                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 356
 357                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 358                        "zeroing out client %s at idx %u (%llu) in %s rc %d\n",
 359                        fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
 360                        LAST_RCVD, rc);
 361         }
 362
 363         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
 364                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
 365                        fed->fed_lr_idx);
 366                 LBUG();
 367         }
 368
 369         EXIT;
 370 free:
 371         OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
 372         fed->fed_fcd = NULL;
 373
 374         return 0;
 375 }
 376
 377 /* drop fmd reference, free it if last ref. must be called with fed_lock held.*/
 378 static inline void filter_fmd_put_nolock(struct filter_export_data *fed,
 379                                          struct filter_mod_data *fmd)
 380 {
 381         LASSERT_SPIN_LOCKED(&fed->fed_lock);
 382         if (--fmd->fmd_refcount == 0) {
 383                 /* XXX when we have persistent reservations and the handle
 384                  * is stored herein we need to drop it here. */
 385                 fed->fed_mod_count--;
 386                 list_del(&fmd->fmd_list);
 387                 OBD_SLAB_FREE(fmd, ll_fmd_cachep, sizeof(*fmd));
 388         }
 389 }
 390
 391 /* drop fmd reference, free it if last ref */
 392 void filter_fmd_put(struct obd_export *exp, struct filter_mod_data *fmd)
 393 {
 394         struct filter_export_data *fed;
 395
 396         if (fmd == NULL)
 397                 return;
 398
 399         fed = &exp->exp_filter_data;
 400         spin_lock(&fed->fed_lock);
 401         filter_fmd_put_nolock(fed, fmd); /* caller reference */
 402         spin_unlock(&fed->fed_lock);
 403 }
 404
 405 /* expire entries from the end of the list if there are too many
 406  * or they are too old */
 407 static void filter_fmd_expire_nolock(struct filter_obd *filter,
 408                                      struct filter_export_data *fed,
 409                                      struct filter_mod_data *keep)
 410 {
 411         struct filter_mod_data *fmd, *tmp;
 412
 413         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
 414                 if (fmd == keep)
 415                         break;
 416
 417                 if (time_before(jiffies, fmd->fmd_expire) &&
 418                     fed->fed_mod_count < filter->fo_fmd_max_num)
 419                         break;
 420
 421                 list_del_init(&fmd->fmd_list);
 422                 filter_fmd_put_nolock(fed, fmd); /* list reference */
 423         }
 424 }
 425
 426 void filter_fmd_expire(struct obd_export *exp)
 427 {
 428         spin_lock(&exp->exp_filter_data.fed_lock);
 429         filter_fmd_expire_nolock(&exp->exp_obd->u.filter,
 430                                  &exp->exp_filter_data, NULL);
 431         spin_unlock(&exp->exp_filter_data.fed_lock);
 432 }
 433
 434 /* find specified objid, group in export fmd list.
 435  * caller must hold fed_lock and take fmd reference itself */
 436 static struct filter_mod_data *filter_fmd_find_nolock(struct filter_obd *filter,
 437                                                 struct filter_export_data *fed,
 438                                                 obd_id objid, obd_gr group)
 439 {
 440         struct filter_mod_data *found = NULL, *fmd;
 441
 442         LASSERT_SPIN_LOCKED(&fed->fed_lock);
 443
 444         list_for_each_entry_reverse(fmd, &fed->fed_mod_list, fmd_list) {
 445                 if (fmd->fmd_id == objid && fmd->fmd_gr == group) {
 446                         found = fmd;
 447                         list_del(&fmd->fmd_list);
 448                         list_add_tail(&fmd->fmd_list, &fed->fed_mod_list);
 449                         fmd->fmd_expire = jiffies + filter->fo_fmd_max_age;
 450                         break;
 451                 }
 452         }
 453
 454         filter_fmd_expire_nolock(filter, fed, found);
 455
 456         return found;
 457 }
 458
 459 /* Find fmd based on objid and group, or return NULL if not found. */
 460 struct filter_mod_data *filter_fmd_find(struct obd_export *exp,
 461                                         obd_id objid, obd_gr group)
 462 {
 463         struct filter_mod_data *fmd;
 464
 465         spin_lock(&exp->exp_filter_data.fed_lock);
 466         fmd = filter_fmd_find_nolock(&exp->exp_obd->u.filter,
 467                                      &exp->exp_filter_data, objid, group);
 468         if (fmd)
 469                 fmd->fmd_refcount++;    /* caller reference */
 470         spin_unlock(&exp->exp_filter_data.fed_lock);
 471
 472         return fmd;
 473 }
 474
 475 /* Find fmd based on objid and group, or create a new one if none is found.
 476  * It is possible for this function to return NULL under memory pressure,
 477  * or if objid = 0 is passed (which will only cause old entries to expire).
 478  * Currently this is not fatal because any fmd state is transient and
 479  * may also be freed when it gets sufficiently old. */
 480 struct filter_mod_data *filter_fmd_get(struct obd_export *exp,
 481                                        obd_id objid, obd_gr group)
 482 {
 483         struct filter_export_data *fed = &exp->exp_filter_data;
 484         struct filter_mod_data *found = NULL, *fmd_new = NULL;
 485
 486         OBD_SLAB_ALLOC(fmd_new, ll_fmd_cachep, CFS_ALLOC_IO, sizeof(*fmd_new));
 487
 488         spin_lock(&fed->fed_lock);
 489         found = filter_fmd_find_nolock(&exp->exp_obd->u.filter,fed,objid,group);
 490         if (fmd_new) {
 491                 if (found == NULL) {
 492                         list_add_tail(&fmd_new->fmd_list, &fed->fed_mod_list);
 493                         fmd_new->fmd_id = objid;
 494                         fmd_new->fmd_gr = group;
 495                         fmd_new->fmd_refcount++;   /* list reference */
 496                         found = fmd_new;
 497                         fed->fed_mod_count++;
 498                 } else {
 499                         OBD_SLAB_FREE(fmd_new, ll_fmd_cachep, sizeof(*fmd_new));
 500                 }
 501         }
 502         if (found) {
 503                 found->fmd_refcount++;          /* caller reference */
 504                 found->fmd_expire = jiffies +
 505                         exp->exp_obd->u.filter.fo_fmd_max_age;
 506         }
 507
 508         spin_unlock(&fed->fed_lock);
 509
 510         return found;
 511 }
 512
 513 #ifdef DO_FMD_DROP
 514 /* drop fmd list reference so it will disappear when last reference is put.
 515  * This isn't so critical because it would in fact only affect the one client
 516  * that is doing the unlink and at worst we have an stale entry referencing
 517  * an object that should never be used again. */
 518 static void filter_fmd_drop(struct obd_export *exp, obd_id objid, obd_gr group)
 519 {
 520         struct filter_mod_data *found = NULL;
 521
 522         spin_lock(&exp->exp_filter_data.fed_lock);
 523         found = filter_fmd_find_nolock(&exp->exp_filter_data, objid, group);
 524         if (found) {
 525                 list_del_init(&found->fmd_list);
 526                 filter_fmd_put_nolock(&exp->exp_filter_data, found);
 527         }
 528         spin_unlock(&exp->exp_filter_data.fed_lock);
 529 }
 530 #else
 531 #define filter_fmd_drop(exp, objid, group)
 532 #endif
 533
 534 /* remove all entries from fmd list */
 535 static void filter_fmd_cleanup(struct obd_export *exp)
 536 {
 537         struct filter_export_data *fed = &exp->exp_filter_data;
 538         struct filter_mod_data *fmd = NULL, *tmp;
 539
 540         spin_lock(&fed->fed_lock);
 541         list_for_each_entry_safe(fmd, tmp, &fed->fed_mod_list, fmd_list) {
 542                 list_del_init(&fmd->fmd_list);
 543                 filter_fmd_put_nolock(fed, fmd);
 544         }
 545         spin_unlock(&fed->fed_lock);
 546 }
 547
 548 static int filter_init_export(struct obd_export *exp)
 549 {
 550         spin_lock_init(&exp->exp_filter_data.fed_lock);
 551         INIT_LIST_HEAD(&exp->exp_filter_data.fed_mod_list);
 552
 553         spin_lock(&exp->exp_lock);
 554         exp->exp_connecting = 1;
 555         spin_unlock(&exp->exp_lock);
 556
 557         return 0;
 558 }
 559
 560 static int filter_free_server_data(struct filter_obd *filter)
 561 {
 562         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
 563         filter->fo_fsd = NULL;
 564         OBD_FREE(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
 565         filter->fo_last_rcvd_slots = NULL;
 566         return 0;
 567 }
 568
 569 /* assumes caller is already in kernel ctxt */
 570 int filter_update_server_data(struct obd_device *obd, struct file *filp,
 571                               struct lr_server_data *fsd, int force_sync)
 572 {
 573         loff_t off = 0;
 574         int rc;
 575         ENTRY;
 576
 577         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->lsd_uuid);
 578         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
 579                le64_to_cpu(fsd->lsd_last_transno));
 580         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
 581                le64_to_cpu(fsd->lsd_mount_count));
 582
 583         fsd->lsd_compat14 = fsd->lsd_last_transno;
 584         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off, force_sync);
 585         if (rc)
 586                 CERROR("error writing lr_server_data: rc = %d\n", rc);
 587
 588         RETURN(rc);
 589 }
 590
 591 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
 592                              int force_sync)
 593 {
 594         struct filter_obd *filter = &obd->u.filter;
 595         __u64 tmp;
 596         loff_t off = 0;
 597         int rc;
 598         ENTRY;
 599
 600         if (filter->fo_last_objid_files[group] == NULL) {
 601                 CERROR("Object group "LPU64" not fully setup; not updating "
 602                        "last_objid\n", group);
 603                 RETURN(-EINVAL);
 604         }
 605
 606         CDEBUG(D_INODE, "%s: server last_objid for group "LPU64": "LPU64"\n",
 607                obd->obd_name, group, filter->fo_last_objids[group]);
 608
 609         tmp = cpu_to_le64(filter->fo_last_objids[group]);
 610         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
 611                                  &tmp, sizeof(tmp), &off, force_sync);
 612         if (rc)
 613                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
 614                        group, rc);
 615         RETURN(rc);
 616 }
 617 extern int ost_handle(struct ptlrpc_request *req);
 618 /* assumes caller has already in kernel ctxt */
 619 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
 620 {
 621         struct filter_obd *filter = &obd->u.filter;
 622         struct lr_server_data *fsd;
 623         struct filter_client_data *fcd = NULL;
 624         struct inode *inode = filp->f_dentry->d_inode;
 625         unsigned long last_rcvd_size = i_size_read(inode);
 626         __u64 mount_count;
 627         int cl_idx;
 628         loff_t off = 0;
 629         int rc;
 630
 631         /* ensure padding in the struct is the correct size */
 632         CLASSERT (offsetof(struct lr_server_data, lsd_padding) +
 633                  sizeof(fsd->lsd_padding) == LR_SERVER_SIZE);
 634         CLASSERT (offsetof(struct filter_client_data, fcd_padding) +
 635                  sizeof(fcd->fcd_padding) == LR_CLIENT_SIZE);
 636
 637         OBD_ALLOC(fsd, sizeof(*fsd));
 638         if (!fsd)
 639                 RETURN(-ENOMEM);
 640         filter->fo_fsd = fsd;
 641
 642         OBD_ALLOC(filter->fo_last_rcvd_slots, LR_MAX_CLIENTS / 8);
 643         if (filter->fo_last_rcvd_slots == NULL) {
 644                 OBD_FREE(fsd, sizeof(*fsd));
 645                 RETURN(-ENOMEM);
 646         }
 647
 648         if (last_rcvd_size == 0) {
 649                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 650
 651                 memcpy(fsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->lsd_uuid));
 652                 fsd->lsd_last_transno = 0;
 653                 mount_count = fsd->lsd_mount_count = 0;
 654                 fsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
 655                 fsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
 656                 fsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
 657                 fsd->lsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
 658                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
 659                 fsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_OST);
 660         } else {
 661                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
 662                 if (rc) {
 663                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
 664                                LAST_RCVD, rc);
 665                         GOTO(err_fsd, rc);
 666                 }
 667                 if (strcmp(fsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 668                         LCONSOLE_ERROR_MSG(0x134, "Trying to start OBD %s "
 669                                            "using the wrong disk %s. Were the "
 670                                            "/dev/ assignments rearranged?\n",
 671                                            obd->obd_uuid.uuid, fsd->lsd_uuid);
 672                         GOTO(err_fsd, rc = -EINVAL);
 673                 }
 674                 mount_count = le64_to_cpu(fsd->lsd_mount_count);
 675                 filter->fo_subdir_count = le16_to_cpu(fsd->lsd_subdir_count);
 676                 /* COMPAT_146 */
 677                 /* Assume old last_rcvd format unless I_C_LR is set */
 678                 if (!(fsd->lsd_feature_incompat &
 679                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
 680                         fsd->lsd_last_transno = fsd->lsd_compat14;
 681                 /* end COMPAT_146 */
 682         }
 683
 684         if (fsd->lsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
 685                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 686                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_incompat) &
 687                        ~FILTER_INCOMPAT_SUPP);
 688                 GOTO(err_fsd, rc = -EINVAL);
 689         }
 690         if (fsd->lsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
 691                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 692                        obd->obd_name, le32_to_cpu(fsd->lsd_feature_rocompat) &
 693                        ~FILTER_ROCOMPAT_SUPP);
 694                 /* Do something like remount filesystem read-only */
 695                 GOTO(err_fsd, rc = -EINVAL);
 696         }
 697
 698         CDEBUG(D_INODE, "%s: server last_transno : "LPU64"\n",
 699                obd->obd_name, le64_to_cpu(fsd->lsd_last_transno));
 700         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 701                obd->obd_name, mount_count + 1);
 702         CDEBUG(D_INODE, "%s: server data size: %u\n",
 703                obd->obd_name, le32_to_cpu(fsd->lsd_server_size));
 704         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 705                obd->obd_name, le32_to_cpu(fsd->lsd_client_start));
 706         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 707                obd->obd_name, le32_to_cpu(fsd->lsd_client_size));
 708         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
 709                obd->obd_name, le16_to_cpu(fsd->lsd_subdir_count));
 710         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 711                last_rcvd_size <= le32_to_cpu(fsd->lsd_client_start) ? 0 :
 712                (last_rcvd_size - le32_to_cpu(fsd->lsd_client_start)) /
 713                 le16_to_cpu(fsd->lsd_client_size));
 714
 715         if (!obd->obd_replayable) {
 716                 CWARN("%s: recovery support OFF\n", obd->obd_name);
 717                 GOTO(out, rc = 0);
 718         }
 719
 720         for (cl_idx = 0, off = le32_to_cpu(fsd->lsd_client_start);
 721              off < last_rcvd_size; cl_idx++) {
 722                 __u64 last_rcvd;
 723                 struct obd_export *exp;
 724                 struct filter_export_data *fed;
 725
 726                 if (!fcd) {
 727                         OBD_ALLOC(fcd, sizeof(*fcd));
 728                         if (!fcd)
 729                                 GOTO(err_client, rc = -ENOMEM);
 730                 }
 731
 732                 /* Don't assume off is incremented properly by
 733                  * fsfilt_read_record(), in case sizeof(*fcd)
 734                  * isn't the same as fsd->lsd_client_size.  */
 735                 off = le32_to_cpu(fsd->lsd_client_start) +
 736                         cl_idx * le16_to_cpu(fsd->lsd_client_size);
 737                 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
 738                 if (rc) {
 739                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
 740                                LAST_RCVD, cl_idx, off, rc);
 741                         break; /* read error shouldn't cause startup to fail */
 742                 }
 743
 744                 if (fcd->fcd_uuid[0] == '\0') {
 745                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 746                                cl_idx);
 747                         continue;
 748                 }
 749
 750                 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
 751
 752                 /* These exports are cleaned up by filter_disconnect(), so they
 753                  * need to be set up like real exports as filter_connect() does.
 754                  */
 755                 exp = class_new_export(obd, (struct obd_uuid *)fcd->fcd_uuid);
 756
 757                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 758                        " srv lr: "LPU64" fcd_group %d\n", fcd->fcd_uuid, cl_idx,
 759                        last_rcvd, le64_to_cpu(fsd->lsd_last_transno),
 760                        le32_to_cpu(fcd->fcd_group));
 761                 if (IS_ERR(exp)) {
 762                         if (PTR_ERR(exp) == -EALREADY) {
 763                                 /* export already exists, zero out this one */
 764                                 CERROR("Zeroing out duplicate export due to "
 765                                        "bug 10479.\n");
 766                                 fcd->fcd_uuid[0] = '\0';
 767                         } else {
 768                                 GOTO(err_client, rc = PTR_ERR(exp));
 769                         }
 770                 } else {
 771                         fed = &exp->exp_filter_data;
 772                         fed->fed_fcd = fcd;
 773                         fed->fed_group = le32_to_cpu(fcd->fcd_group);
 774                         filter_export_stats_init(obd, exp);
 775                         rc = filter_client_add(obd, exp, cl_idx);
 776                         /* can't fail for existing client */
 777                         LASSERTF(rc == 0, "rc = %d\n", rc);
 778
 779                         fcd = NULL;
 780                         spin_lock(&exp->exp_lock);
 781                         exp->exp_connecting = 0;
 782                         exp->exp_in_recovery = 0;
 783                         spin_unlock(&exp->exp_lock);
 784                         obd->obd_max_recoverable_clients++;
 785                         class_export_put(exp);
 786                 }
 787
 788                 /* Need to check last_rcvd even for duplicated exports. */
 789                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
 790                        cl_idx, last_rcvd);
 791
 792                 if (last_rcvd > le64_to_cpu(fsd->lsd_last_transno))
 793                         fsd->lsd_last_transno = cpu_to_le64(last_rcvd);
 794         }
 795
 796         if (fcd)
 797                 OBD_FREE(fcd, sizeof(*fcd));
 798
 799         obd->obd_last_committed = le64_to_cpu(fsd->lsd_last_transno);
 800
 801         target_recovery_init(obd, ost_handle);
 802
 803 out:
 804         filter->fo_mount_count = mount_count + 1;
 805         fsd->lsd_mount_count = cpu_to_le64(filter->fo_mount_count);
 806
 807         /* save it, so mount count and last_transno is current */
 808         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
 809         if (rc)
 810                 GOTO(err_client, rc);
 811
 812         RETURN(0);
 813
 814 err_client:
 815         target_recovery_fini(obd);
 816 err_fsd:
 817         filter_free_server_data(filter);
 818         RETURN(rc);
 819 }
 820
 821 static int filter_cleanup_groups(struct obd_device *obd)
 822 {
 823         struct filter_obd *filter = &obd->u.filter;
 824         struct file *filp;
 825         struct dentry *dentry;
 826         int i, j;
 827         ENTRY;
 828
 829         if (filter->fo_dentry_O_groups != NULL) {
 830                 for (i = 0; i < filter->fo_group_count; i++) {
 831                         dentry = filter->fo_dentry_O_groups[i];
 832                         if (dentry != NULL)
 833                                 f_dput(dentry);
 834                 }
 835                 OBD_FREE(filter->fo_dentry_O_groups,
 836                          filter->fo_group_count *
 837                          sizeof(*filter->fo_dentry_O_groups));
 838                 filter->fo_dentry_O_groups = NULL;
 839         }
 840         if (filter->fo_last_objid_files != NULL) {
 841                 for (i = 0; i < filter->fo_group_count; i++) {
 842                         filp = filter->fo_last_objid_files[i];
 843                         if (filp != NULL)
 844                                 filp_close(filp, 0);
 845                 }
 846                 OBD_FREE(filter->fo_last_objid_files,
 847                          filter->fo_group_count *
 848                          sizeof(*filter->fo_last_objid_files));
 849                 filter->fo_last_objid_files = NULL;
 850         }
 851         if (filter->fo_dentry_O_sub != NULL) {
 852                 for (i = 0; i < filter->fo_group_count; i++) {
 853                         for (j = 0; j < filter->fo_subdir_count; j++) {
 854                                 dentry = filter->fo_dentry_O_sub[i].dentry[j];
 855                                 if (dentry != NULL)
 856                                         f_dput(dentry);
 857                         }
 858                 }
 859                 OBD_FREE(filter->fo_dentry_O_sub,
 860                          filter->fo_group_count *
 861                          sizeof(*filter->fo_dentry_O_sub));
 862                 filter->fo_dentry_O_sub = NULL;
 863         }
 864         if (filter->fo_last_objids != NULL) {
 865                 OBD_FREE(filter->fo_last_objids,
 866                          filter->fo_group_count *
 867                          sizeof(*filter->fo_last_objids));
 868                 filter->fo_last_objids = NULL;
 869         }
 870         if (filter->fo_dentry_O != NULL) {
 871                 f_dput(filter->fo_dentry_O);
 872                 filter->fo_dentry_O = NULL;
 873         }
 874         RETURN(0);
 875 }
 876
 877 static int filter_update_last_group(struct obd_device *obd, int group)
 878 {
 879         struct filter_obd *filter = &obd->u.filter;
 880         struct file *filp = NULL;
 881         int last_group = 0, rc;
 882         loff_t off = 0;
 883         ENTRY;
 884
 885         if (group <= filter->fo_committed_group)
 886                 RETURN(0);
 887
 888         filp = filp_open("LAST_GROUP", O_RDWR, 0700);
 889         if (IS_ERR(filp)) {
 890                 rc = PTR_ERR(filp);
 891                 filp = NULL;
 892                 CERROR("cannot open LAST_GROUP: rc = %d\n", rc);
 893                 GOTO(cleanup, rc);
 894         }
 895
 896         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
 897         if (rc) {
 898                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n",rc);
 899                 GOTO(cleanup, rc);
 900         }
 901         LASSERT(off == 0 || last_group >= FILTER_MIN_GROUPS);
 902         CDEBUG(D_INODE, "%s: previous %d, new %d\n",
 903                obd->obd_name, last_group, group);
 904
 905         off = 0;
 906         last_group = group;
 907         /* must be sync: bXXXX */
 908         rc = fsfilt_write_record(obd, filp, &last_group, sizeof(__u32), &off, 1);
 909         if (rc) {
 910                 CDEBUG(D_INODE, "error updating LAST_GROUP: rc %d\n", rc);
 911                 GOTO(cleanup, rc);
 912         }
 913
 914         filter->fo_committed_group = group;
 915 cleanup:
 916         if (filp)
 917                 filp_close(filp, 0);
 918         RETURN(rc);
 919 }
 920
 921 static int filter_read_group_internal(struct obd_device *obd, int group,
 922                                       int create)
 923 {
 924         struct filter_obd *filter = &obd->u.filter;
 925         __u64 *new_objids = NULL;
 926         struct filter_subdirs *new_subdirs = NULL, *tmp_subdirs = NULL;
 927         struct dentry **new_groups = NULL;
 928         struct file **new_files = NULL;
 929         struct dentry *dentry;
 930         struct file *filp;
 931         int old_count = filter->fo_group_count, rc, stage = 0, i;
 932         char name[25];
 933         __u64 last_objid;
 934         loff_t off = 0;
 935         int len = group + 1;
 936
 937         snprintf(name, 24, "%d", group);
 938         name[24] = '\0';
 939
 940         if (!create) {
 941                 dentry = ll_lookup_one_len(name, filter->fo_dentry_O,
 942                                            strlen(name));
 943                 if (IS_ERR(dentry)) {
 944                         CERROR("Cannot lookup expected object group %d: %ld\n",
 945                                group, PTR_ERR(dentry));
 946                         RETURN(PTR_ERR(dentry));
 947                 }
 948         } else {
 949                 dentry = simple_mkdir(filter->fo_dentry_O, name, 0700, 1);
 950                 if (IS_ERR(dentry)) {
 951                         CERROR("cannot lookup/create O/%s: rc = %ld\n", name,
 952                                PTR_ERR(dentry));
 953                         RETURN(PTR_ERR(dentry));
 954                 }
 955         }
 956         stage = 1;
 957
 958         snprintf(name, 24, "O/%d/LAST_ID", group);
 959         name[24] = '\0';
 960         filp = filp_open(name, O_CREAT | O_RDWR, 0700);
 961         if (IS_ERR(filp)) {
 962                 CERROR("cannot create %s: rc = %ld\n", name, PTR_ERR(filp));
 963                 GOTO(cleanup, rc = PTR_ERR(filp));
 964         }
 965         stage = 2;
 966
 967         rc = fsfilt_read_record(obd, filp, &last_objid, sizeof(__u64), &off);
 968         if (rc) {
 969                 CDEBUG(D_INODE, "error reading %s: rc %d\n", name, rc);
 970                 GOTO(cleanup, rc);
 971         }
 972
 973         if (filter->fo_subdir_count) {
 974                 OBD_ALLOC(tmp_subdirs, sizeof(*tmp_subdirs));
 975                 if (tmp_subdirs == NULL)
 976                         GOTO(cleanup, rc = -ENOMEM);
 977                 stage = 3;
 978
 979                 for (i = 0; i < filter->fo_subdir_count; i++) {
 980                         char dir[20];
 981                         snprintf(dir, sizeof(dir), "d%u", i);
 982
 983                         tmp_subdirs->dentry[i] = simple_mkdir(dentry, dir, 0700, 1);
 984                         if (IS_ERR(tmp_subdirs->dentry[i])) {
 985                                 rc = PTR_ERR(tmp_subdirs->dentry[i]);
 986                                 CERROR("can't lookup/create O/%d/%s: rc = %d\n",
 987                                        group, dir, rc);
 988                                 GOTO(cleanup, rc);
 989                         }
 990
 991                         CDEBUG(D_INODE, "got/created O/%d/%s: %p\n", group, dir,
 992                                tmp_subdirs->dentry[i]);
 993                 }
 994         }
 995
 996         /* 'group' is an index; we need an array of length 'group + 1' */
 997         if (group + 1 > old_count) {
 998                 OBD_ALLOC(new_objids, len * sizeof(*new_objids));
 999                 OBD_ALLOC(new_subdirs, len * sizeof(*new_subdirs));
1000                 OBD_ALLOC(new_groups, len * sizeof(*new_groups));
1001                 OBD_ALLOC(new_files, len * sizeof(*new_files));
1002                 stage = 4;
1003                 if (new_objids == NULL || new_subdirs == NULL ||
1004                     new_groups == NULL || new_files == NULL)
1005                         GOTO(cleanup, rc = -ENOMEM);
1006
1007                 if (old_count) {
1008                         memcpy(new_objids, filter->fo_last_objids,
1009                                old_count * sizeof(*new_objids));
1010                         memcpy(new_subdirs, filter->fo_dentry_O_sub,
1011                                old_count * sizeof(*new_subdirs));
1012                         memcpy(new_groups, filter->fo_dentry_O_groups,
1013                                old_count * sizeof(*new_groups));
1014                         memcpy(new_files, filter->fo_last_objid_files,
1015                                old_count * sizeof(*new_files));
1016
1017                         OBD_FREE(filter->fo_last_objids,
1018                                  old_count * sizeof(*new_objids));
1019                         OBD_FREE(filter->fo_dentry_O_sub,
1020                                  old_count * sizeof(*new_subdirs));
1021                         OBD_FREE(filter->fo_dentry_O_groups,
1022                                  old_count * sizeof(*new_groups));
1023                         OBD_FREE(filter->fo_last_objid_files,
1024                                  old_count * sizeof(*new_files));
1025                 }
1026                 filter->fo_last_objids = new_objids;
1027                 filter->fo_dentry_O_sub = new_subdirs;
1028                 filter->fo_dentry_O_groups = new_groups;
1029                 filter->fo_last_objid_files = new_files;
1030                 filter->fo_group_count = len;
1031         }
1032
1033         filter->fo_dentry_O_groups[group] = dentry;
1034         filter->fo_last_objid_files[group] = filp;
1035         if (filter->fo_subdir_count) {
1036                 filter->fo_dentry_O_sub[group] = *tmp_subdirs;
1037                 OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1038         }
1039
1040         filter_update_last_group(obd, group);
1041
1042         if (i_size_read(filp->f_dentry->d_inode) == 0) {
1043                 filter->fo_last_objids[group] = FILTER_INIT_OBJID;
1044                 rc = filter_update_last_objid(obd, group, 1);
1045                 RETURN(rc);
1046         }
1047
1048         filter->fo_last_objids[group] = le64_to_cpu(last_objid);
1049         CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
1050                obd->obd_name, group, last_objid);
1051         RETURN(0);
1052  cleanup:
1053         switch (stage) {
1054         case 4:
1055                 if (new_objids != NULL)
1056                         OBD_FREE(new_objids, len * sizeof(*new_objids));
1057                 if (new_subdirs != NULL)
1058                         OBD_FREE(new_subdirs, len * sizeof(*new_subdirs));
1059                 if (new_groups != NULL)
1060                         OBD_FREE(new_groups, len * sizeof(*new_groups));
1061                 if (new_files != NULL)
1062                         OBD_FREE(new_files, len * sizeof(*new_files));
1063         case 3:
1064                 if (filter->fo_subdir_count) {
1065                         for (i = 0; i < filter->fo_subdir_count; i++) {
1066                                 if (tmp_subdirs->dentry[i] != NULL)
1067                                         dput(tmp_subdirs->dentry[i]);
1068                         }
1069                         OBD_FREE(tmp_subdirs, sizeof(*tmp_subdirs));
1070                 }
1071         case 2:
1072                 filp_close(filp, 0);
1073         case 1:
1074                 dput(dentry);
1075         }
1076         RETURN(rc);
1077 }
1078
1079 static int filter_read_groups(struct obd_device *obd, int last_group,
1080                               int create)
1081 {
1082         struct filter_obd *filter = &obd->u.filter;
1083         int old_count, group, rc = 0;
1084
1085         down(&filter->fo_init_lock);
1086         old_count = filter->fo_group_count;
1087         for (group = old_count; group <= last_group; group++) {
1088                 if (group == 0)
1089                         continue; /* no group zero */
1090
1091                 rc = filter_read_group_internal(obd, group, create);
1092                 if (rc != 0)
1093                         break;
1094         }
1095         up(&filter->fo_init_lock);
1096         return rc;
1097 }
1098
1099 /* FIXME: object groups */
1100 static int filter_prep_groups(struct obd_device *obd)
1101 {
1102         struct filter_obd *filter = &obd->u.filter;
1103         struct dentry *dentry, *O_dentry;
1104         struct file *filp;
1105         int    last_group, rc = 0, cleanup_phase = 0;
1106         loff_t off = 0;
1107         ENTRY;
1108
1109         O_dentry = simple_mkdir(current->fs->pwd, "O", 0700, 1);
1110         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
1111         if (IS_ERR(O_dentry)) {
1112                 rc = PTR_ERR(O_dentry);
1113                 CERROR("cannot open/create O: rc = %d\n", rc);
1114                 GOTO(cleanup, rc);
1115         }
1116         filter->fo_dentry_O = O_dentry;
1117         cleanup_phase = 1; /* O_dentry */
1118
1119         /* Lookup "R" to tell if we're on an old OST FS and need to convert
1120          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
1121          * some time post 1.0 when all old-style OSTs have converted along
1122          * with the init_objid hack. */
1123         dentry = ll_lookup_one_len("R", O_dentry, 1);
1124         if (IS_ERR(dentry))
1125                 GOTO(cleanup, rc = PTR_ERR(dentry));
1126         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
1127                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
1128                 ENTRY;
1129
1130                 CWARN("converting OST to new object layout\n");
1131                 if (IS_ERR(O0_dentry)) {
1132                         rc = PTR_ERR(O0_dentry);
1133                         CERROR("error looking up O/0: rc %d\n", rc);
1134                         GOTO(cleanup_R, rc);
1135                 }
1136
1137                 if (O0_dentry->d_inode) {
1138                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
1139                         GOTO(cleanup_O0, rc = -EEXIST);
1140                 }
1141
1142                 LOCK_INODE_MUTEX(O_dentry->d_inode);
1143                 rc = vfs_rename(O_dentry->d_inode, dentry,
1144                                 O_dentry->d_inode, O0_dentry);
1145                 UNLOCK_INODE_MUTEX(O_dentry->d_inode);
1146
1147                 if (rc) {
1148                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
1149                         GOTO(cleanup_O0, rc);
1150                 }
1151                 filter->fo_fsd->lsd_feature_incompat |=
1152                         cpu_to_le32(OBD_INCOMPAT_GROUPS);
1153                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1154                                                filter->fo_fsd, 1);
1155                 GOTO(cleanup_O0, rc);
1156
1157         cleanup_O0:
1158                 f_dput(O0_dentry);
1159         cleanup_R:
1160                 f_dput(dentry);
1161                 if (rc)
1162                         GOTO(cleanup, rc);
1163         } else {
1164                 f_dput(dentry);
1165         }
1166
1167         cleanup_phase = 2; /* groups */
1168
1169         /* we have to initialize all groups before first connections from
1170          * clients because they may send create/destroy for any group -bzzz */
1171         filp = filp_open("LAST_GROUP", O_CREAT | O_RDWR, 0700);
1172         if (IS_ERR(filp)) {
1173                 CERROR("cannot create LAST_GROUP: rc = %ld\n", PTR_ERR(filp));
1174                 GOTO(cleanup, rc = PTR_ERR(filp));
1175         }
1176         cleanup_phase = 3; /* filp */
1177
1178         rc = fsfilt_read_record(obd, filp, &last_group, sizeof(__u32), &off);
1179         if (rc) {
1180                 CDEBUG(D_INODE, "error reading LAST_GROUP: rc %d\n", rc);
1181                 GOTO(cleanup, rc);
1182         }
1183         if (off == 0) {
1184                 last_group = FILTER_MIN_GROUPS;
1185         } else {
1186                 LASSERT(last_group >= FILTER_MIN_GROUPS);
1187         }
1188
1189         CWARN("%s: initialize groups [%d,%d]\n", obd->obd_name,
1190               FILTER_MIN_GROUPS, last_group);
1191         filter->fo_committed_group = last_group;
1192         rc = filter_read_groups(obd, last_group, 1);
1193         if (rc)
1194                 GOTO(cleanup, rc);
1195
1196         filp_close(filp, 0);
1197         RETURN(0);
1198
1199  cleanup:
1200         switch (cleanup_phase) {
1201         case 3:
1202                 filp_close(filp, 0);
1203         case 2:
1204                 filter_cleanup_groups(obd);
1205         case 1:
1206                 f_dput(filter->fo_dentry_O);
1207                 filter->fo_dentry_O = NULL;
1208         default:
1209                 break;
1210         }
1211         return rc;
1212
1213 }
1214
1215 /* setup the object store with correct subdirectories */
1216 static int filter_prep(struct obd_device *obd)
1217 {
1218         struct lvfs_run_ctxt saved;
1219         struct filter_obd *filter = &obd->u.filter;
1220         struct file *file;
1221         struct inode *inode;
1222         int rc = 0;
1223         ENTRY;
1224
1225         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1226         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
1227         if (!file || IS_ERR(file)) {
1228                 rc = PTR_ERR(file);
1229                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
1230                        LAST_RCVD, rc);
1231                 GOTO(out, rc);
1232         }
1233         filter->fo_rcvd_filp = file;
1234         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1235                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
1236                        file->f_dentry->d_inode->i_mode);
1237                 GOTO(err_filp, rc = -ENOENT);
1238         }
1239
1240         inode = file->f_dentry->d_parent->d_inode;
1241         /* We use i_op->unlink directly in filter_vfs_unlink() */
1242         if (!inode->i_op || !inode->i_op->create || !inode->i_op->unlink) {
1243                 CERROR("%s: filesystem does not support create/unlink ops\n",
1244                        obd->obd_name);
1245                 GOTO(err_filp, rc = -EOPNOTSUPP);
1246         }
1247
1248         rc = filter_init_server_data(obd, file);
1249         if (rc) {
1250                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
1251                 GOTO(err_filp, rc);
1252         }
1253         /* open and create health check io file*/
1254         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
1255         if (IS_ERR(file)) {
1256                 rc = PTR_ERR(file);
1257                 CERROR("OBD filter: cannot open/create %s rc = %d\n",
1258                        HEALTH_CHECK, rc);
1259                 GOTO(err_server_data, rc);
1260         }
1261         filter->fo_health_check_filp = file;
1262         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
1263                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
1264                        file->f_dentry->d_inode->i_mode);
1265                 GOTO(err_health_check, rc = -ENOENT);
1266         }
1267         rc = lvfs_check_io_health(obd, file);
1268         if (rc)
1269                 GOTO(err_health_check, rc);
1270
1271         rc = filter_prep_groups(obd);
1272         if (rc)
1273                 GOTO(err_health_check, rc);
1274 out:
1275         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1276
1277         return(rc);
1278
1279 err_health_check:
1280         if (filp_close(filter->fo_health_check_filp, 0))
1281                 CERROR("can't close %s after error\n", HEALTH_CHECK);
1282         filter->fo_health_check_filp = NULL;
1283 err_server_data:
1284         target_recovery_fini(obd);
1285         filter_free_server_data(filter);
1286 err_filp:
1287         if (filp_close(filter->fo_rcvd_filp, 0))
1288                 CERROR("can't close %s after error\n", LAST_RCVD);
1289         filter->fo_rcvd_filp = NULL;
1290         goto out;
1291 }
1292
1293 /* cleanup the filter: write last used object id to status file */
1294 static void filter_post(struct obd_device *obd)
1295 {
1296         struct lvfs_run_ctxt saved;
1297         struct filter_obd *filter = &obd->u.filter;
1298         int rc, i;
1299
1300         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
1301          * best to start a transaction with h_sync, because we removed this
1302          * from lastobjid */
1303
1304         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1305         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
1306                                        filter->fo_fsd, 0);
1307         if (rc)
1308                 CERROR("error writing server data: rc = %d\n", rc);
1309
1310         for (i = 1; i < filter->fo_group_count; i++) {
1311                 rc = filter_update_last_objid(obd, i,
1312                                 (i == filter->fo_group_count - 1));
1313                 if (rc)
1314                         CERROR("error writing group %d lastobjid: rc = %d\n",
1315                                i, rc);
1316         }
1317
1318         rc = filp_close(filter->fo_rcvd_filp, 0);
1319         filter->fo_rcvd_filp = NULL;
1320         if (rc)
1321                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
1322
1323         rc = filp_close(filter->fo_health_check_filp, 0);
1324         filter->fo_health_check_filp = NULL;
1325         if (rc)
1326                 CERROR("error closing %s: rc = %d\n", HEALTH_CHECK, rc);
1327
1328         filter_cleanup_groups(obd);
1329         filter_free_server_data(filter);
1330         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1331
1332         filter_free_capa_keys(filter);
1333         cleanup_capa_hash(filter->fo_capa_hash);
1334 }
1335
1336 static void filter_set_last_id(struct filter_obd *filter,
1337                                obd_id id, obd_gr group)
1338 {
1339         LASSERT(filter->fo_fsd != NULL);
1340         LASSERT(group <= filter->fo_group_count);
1341
1342         spin_lock(&filter->fo_objidlock);
1343         filter->fo_last_objids[group] = id;
1344         spin_unlock(&filter->fo_objidlock);
1345 }
1346
1347 obd_id filter_last_id(struct filter_obd *filter, obd_gr group)
1348 {
1349         obd_id id;
1350         LASSERT(filter->fo_fsd != NULL);
1351         LASSERT(group <= filter->fo_group_count);
1352
1353         /* FIXME: object groups */
1354         spin_lock(&filter->fo_objidlock);
1355         id = filter->fo_last_objids[group];
1356         spin_unlock(&filter->fo_objidlock);
1357
1358         return id;
1359 }
1360
1361 static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
1362 {
1363         LOCK_INODE_MUTEX(dparent->d_inode);
1364         return 0;
1365 }
1366
1367 /* We never dget the object parent, so DON'T dput it either */
1368 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
1369 {
1370         struct filter_obd *filter = &obd->u.filter;
1371         struct filter_subdirs *subdirs;
1372         LASSERT(group < filter->fo_group_count); /* FIXME: object groups */
1373
1374         if (group > 0 || filter->fo_subdir_count == 0)
1375                 return filter->fo_dentry_O_groups[group];
1376
1377         subdirs = &filter->fo_dentry_O_sub[group];
1378         return subdirs->dentry[objid & (filter->fo_subdir_count - 1)];
1379 }
1380
1381 /* We never dget the object parent, so DON'T dput it either */
1382 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
1383                                   obd_id objid)
1384 {
1385         unsigned long now = jiffies;
1386         struct dentry *dparent = filter_parent(obd, group, objid);
1387         int rc;
1388
1389         if (IS_ERR(dparent))
1390                 return dparent;
1391         if (dparent == NULL)
1392                 return ERR_PTR(-ENOENT);
1393
1394         rc = filter_lock_dentry(obd, dparent);
1395         fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1396         return rc ? ERR_PTR(rc) : dparent;
1397 }
1398
1399 /* We never dget the object parent, so DON'T dput it either */
1400 static void filter_parent_unlock(struct dentry *dparent)
1401 {
1402         UNLOCK_INODE_MUTEX(dparent->d_inode);
1403 }
1404
1405 /* How to get files, dentries, inodes from object id's.
1406  *
1407  * If dir_dentry is passed, the caller has already locked the parent
1408  * appropriately for this operation (normally a write lock).  If
1409  * dir_dentry is NULL, we do a read lock while we do the lookup to
1410  * avoid races with create/destroy and such changing the directory
1411  * internal to the filesystem code. */
1412 struct dentry *filter_fid2dentry(struct obd_device *obd,
1413                                  struct dentry *dir_dentry,
1414                                  obd_gr group, obd_id id)
1415 {
1416         struct dentry *dparent = dir_dentry;
1417         struct dentry *dchild;
1418         char name[32];
1419         int len;
1420         ENTRY;
1421
1422         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1423             obd->u.filter.fo_destroys_in_progress == 0) {
1424                 /* don't fail lookups for orphan recovery, it causes
1425                  * later LBUGs when objects still exist during precreate */
1426                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1427                 RETURN(ERR_PTR(-ENOENT));
1428         }
1429         if (id == 0) {
1430                 CERROR("fatal: invalid object id 0\n");
1431                 RETURN(ERR_PTR(-ESTALE));
1432         }
1433
1434         len = sprintf(name, LPU64, id);
1435         if (dir_dentry == NULL) {
1436                 dparent = filter_parent_lock(obd, group, id);
1437                 if (IS_ERR(dparent)) {
1438                         CERROR("%s: error getting object "LPU64":"LPU64
1439                                " parent: rc %ld\n", obd->obd_name,
1440                                id, group, PTR_ERR(dparent));
1441                         RETURN(dparent);
1442                 }
1443         }
1444         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1445                dparent->d_name.len, dparent->d_name.name, name);
1446         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1447         if (dir_dentry == NULL)
1448                 filter_parent_unlock(dparent);
1449         if (IS_ERR(dchild)) {
1450                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1451                        PTR_ERR(dchild));
1452                 RETURN(dchild);
1453         }
1454
1455         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1456                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1457                        obd->obd_name, id, dchild->d_inode->i_ino);
1458                 f_dput(dchild);
1459                 RETURN(ERR_PTR(-ENOENT));
1460         }
1461
1462         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1463                name, dchild, atomic_read(&dchild->d_count));
1464
1465         LASSERT(atomic_read(&dchild->d_count) > 0);
1466
1467         RETURN(dchild);
1468 }
1469
1470 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1471                                   obd_id group)
1472 {
1473         struct lustre_handle lockh;
1474         int flags = LDLM_AST_DISCARD_DATA, rc;
1475         struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1476         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1477
1478         ENTRY;
1479         /* Tell the clients that the object is gone now and that they should
1480          * throw away any cached pages. */
1481         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1482                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1483                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1484                                     &lockh);
1485
1486         /* We only care about the side-effects, just drop the lock. */
1487         if (rc == ELDLM_OK)
1488                 ldlm_lock_decref(&lockh, LCK_PW);
1489
1490         RETURN(rc);
1491 }
1492
1493 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1494  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1495  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1496  * reverse.  Caller must take i_sem before starting the transaction and we
1497  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1498 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1499 {
1500         int rc;
1501         ENTRY;
1502
1503         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1504          * itself we already hold dir->i_mutex for child create/unlink ops */
1505         LASSERT(dentry->d_inode != NULL);
1506         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1507         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1508
1509
1510         /* may_delete() */
1511         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1512                 GOTO(out, rc = -ENOENT);
1513
1514         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1515         if (rc)
1516                 GOTO(out, rc);
1517
1518         if (IS_APPEND(dir))
1519                 GOTO(out, rc = -EPERM);
1520
1521         /* check_sticky() */
1522         if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1523             IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1524                 GOTO(out, rc = -EPERM);
1525
1526         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1527          *       that was done because of journal_start (which is already done
1528          *       here) or some other ordering issue. */
1529         DQUOT_INIT(dir);
1530
1531         rc = security_inode_unlink(dir, dentry);
1532         if (rc)
1533                 GOTO(out, rc);
1534
1535         rc = dir->i_op->unlink(dir, dentry);
1536 out:
1537         /* need to drop i_mutex before we lose inode reference */
1538         UNLOCK_INODE_MUTEX(dentry->d_inode);
1539         if (rc == 0)
1540                 d_delete(dentry);
1541
1542         RETURN(rc);
1543 }
1544
1545 /* Caller must hold LCK_PW on parent and push us into kernel context.
1546  * Caller must hold child i_mutex, we drop it always.
1547  * Caller is also required to ensure that dchild->d_inode exists. */
1548 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1549                                    obd_gr group, struct dentry *dparent,
1550                                    struct dentry *dchild)
1551 {
1552         struct inode *inode = dchild->d_inode;
1553         int rc;
1554
1555         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1556                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1557                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1558                        (unsigned long)inode->i_nlink,
1559                        atomic_read(&inode->i_count));
1560         }
1561
1562         rc = filter_vfs_unlink(dparent->d_inode, dchild);
1563         if (rc)
1564                 CERROR("error unlinking objid %.*s: rc %d\n",
1565                        dchild->d_name.len, dchild->d_name.name, rc);
1566         return(rc);
1567 }
1568
1569 static int filter_intent_policy(struct ldlm_namespace *ns,
1570                                 struct ldlm_lock **lockp, void *req_cookie,
1571                                 ldlm_mode_t mode, int flags, void *data)
1572 {
1573         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1574         struct ptlrpc_request *req = req_cookie;
1575         struct ldlm_lock *lock = *lockp, *l = NULL;
1576         struct ldlm_resource *res = lock->l_resource;
1577         ldlm_processing_policy policy;
1578         struct ost_lvb *res_lvb, *reply_lvb;
1579         struct ldlm_reply *rep;
1580         struct list_head *tmp;
1581         ldlm_error_t err;
1582         int rc, tmpflags = 0, only_liblustre = 0;
1583         int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1584                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1585                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1586         ENTRY;
1587
1588         policy = ldlm_get_processing_policy(res);
1589         LASSERT(policy != NULL);
1590         LASSERT(req != NULL);
1591
1592         rc = lustre_pack_reply(req, 3, repsize, NULL);
1593         if (rc)
1594                 RETURN(req->rq_status = rc);
1595
1596         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1597         LASSERT(rep != NULL);
1598
1599         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1600                                    sizeof(*reply_lvb));
1601         LASSERT(reply_lvb != NULL);
1602
1603         //fixup_handle_for_resent_req(req, lock, &lockh);
1604
1605         /* If we grant any lock at all, it will be a whole-file read lock.
1606          * Call the extent policy function to see if our request can be
1607          * granted, or is blocked. */
1608         lock->l_policy_data.l_extent.start = 0;
1609         lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1610         lock->l_req_mode = LCK_PR;
1611
1612         LASSERT(ns == res->lr_namespace);
1613         lock_res(res);
1614         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1615         check_res_locked(res);
1616
1617         /* FIXME: we should change the policy function slightly, to not make
1618          * this list at all, since we just turn around and free it */
1619         while (!list_empty(&rpc_list)) {
1620                 struct ldlm_lock *wlock =
1621                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1622                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1623                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1624                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1625                 list_del_init(&wlock->l_cp_ast);
1626                 LDLM_LOCK_PUT(wlock);
1627         }
1628
1629         /* The lock met with no resistance; we're finished. */
1630         if (rc == LDLM_ITER_CONTINUE) {
1631                 /* do not grant locks to the liblustre clients: they cannot
1632                  * handle ASTs robustly.  We need to do this while still
1633                  * holding ns_lock to avoid the lock remaining on the res_link
1634                  * list (and potentially being added to l_pending_list by an
1635                  * AST) when we are going to drop this lock ASAP. */
1636                 if (lock->l_export->exp_libclient ||
1637                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1638                         ldlm_resource_unlink_lock(lock);
1639                         err = ELDLM_LOCK_ABORTED;
1640                 } else {
1641                         err = ELDLM_LOCK_REPLACED;
1642                 }
1643                 unlock_res(res);
1644                 RETURN(err);
1645         }
1646
1647         /* Do not grant any lock, but instead send GL callbacks.  The extent
1648          * policy nicely created a list of all PW locks for us.  We will choose
1649          * the highest of those which are larger than the size in the LVB, if
1650          * any, and perform a glimpse callback. */
1651         res_lvb = res->lr_lvb_data;
1652         LASSERT(res_lvb != NULL);
1653         *reply_lvb = *res_lvb;
1654
1655         list_for_each(tmp, &res->lr_granted) {
1656                 struct ldlm_lock *tmplock =
1657                         list_entry(tmp, struct ldlm_lock, l_res_link);
1658
1659                 if (tmplock->l_granted_mode == LCK_PR)
1660                         continue;
1661                 /*
1662                  * ->ns_lock guarantees that no new locks are granted, and,
1663                  * therefore, that res->lr_lvb_data cannot increase beyond the
1664                  * end of already granted lock. As a result, it is safe to
1665                  * check against "stale" reply_lvb->lvb_size value without
1666                  * res->lr_lvb_sem.
1667                  */
1668                 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1669                         continue;
1670
1671                 /* Don't send glimpse ASTs to liblustre clients.  They aren't
1672                  * listening for them, and they do entirely synchronous I/O
1673                  * anyways. */
1674                 if (tmplock->l_export == NULL ||
1675                     tmplock->l_export->exp_libclient == 1) {
1676                         only_liblustre = 1;
1677                         continue;
1678                 }
1679
1680                 if (l == NULL) {
1681                         l = LDLM_LOCK_GET(tmplock);
1682                         continue;
1683                 }
1684
1685                 if (l->l_policy_data.l_extent.start >
1686                     tmplock->l_policy_data.l_extent.start)
1687                         continue;
1688
1689                 LDLM_LOCK_PUT(l);
1690                 l = LDLM_LOCK_GET(tmplock);
1691         }
1692         unlock_res(res);
1693
1694         /* There were no PW locks beyond the size in the LVB; finished. */
1695         if (l == NULL) {
1696                 if (only_liblustre) {
1697                         /* If we discovered a liblustre client with a PW lock,
1698                          * however, the LVB may be out of date!  The LVB is
1699                          * updated only on glimpse (which we don't do for
1700                          * liblustre clients) and cancel (which the client
1701                          * obviously has not yet done).  So if it has written
1702                          * data but kept the lock, the LVB is stale and needs
1703                          * to be updated from disk.
1704                          *
1705                          * Of course, this will all disappear when we switch to
1706                          * taking liblustre locks on the OST. */
1707                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1708                 }
1709                 RETURN(ELDLM_LOCK_ABORTED);
1710         }
1711
1712         /*
1713          * This check is for lock taken in filter_prepare_destroy() that does
1714          * not have l_glimpse_ast set. So the logic is: if there is a lock
1715          * with no l_glimpse_ast set, this object is being destroyed already.
1716          *
1717          * Hence, if you are grabbing DLM locks on the server, always set
1718          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1719          */
1720         if (l->l_glimpse_ast == NULL) {
1721                 /* We are racing with unlink(); just return -ENOENT */
1722                 rep->lock_policy_res1 = -ENOENT;
1723                 goto out;
1724         }
1725
1726         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1727         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1728         /* Update the LVB from disk if the AST failed (this is a legal race) */
1729         /*
1730          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1731          * sending ast is not handled. This can result in lost client writes.
1732          */
1733         if (rc != 0)
1734                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1735
1736         lock_res(res);
1737         *reply_lvb = *res_lvb;
1738         unlock_res(res);
1739
1740  out:
1741         LDLM_LOCK_PUT(l);
1742
1743         RETURN(ELDLM_LOCK_ABORTED);
1744 }
1745
1746 /*
1747  * per-obd_device iobuf pool.
1748  *
1749  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1750  * allocations in write-path has to be minimized (see bug 5137).
1751  *
1752  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1753  * OST threads (see ost_thread_{init,done}()).
1754  *
1755  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1756  * at the OST layer there are only (potentially) multiple obd_device of type
1757  * unknown at the time of OST thread creation.
1758  *
1759  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1760  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1761  * it's very own iobuf.
1762  *
1763  * Functions below
1764  *
1765  *     filter_kiobuf_pool_init()
1766  *
1767  *     filter_kiobuf_pool_done()
1768  *
1769  *     filter_iobuf_get()
1770  *
1771  * operate on this array. They are "generic" in a sense that they don't depend
1772  * on actual type of iobuf's (the latter depending on Linux kernel version).
1773  */
1774
1775 /*
1776  * destroy pool created by filter_iobuf_pool_init
1777  */
1778 static void filter_iobuf_pool_done(struct filter_obd *filter)
1779 {
1780         struct filter_iobuf **pool;
1781         int i;
1782
1783         ENTRY;
1784
1785         pool = filter->fo_iobuf_pool;
1786         if (pool != NULL) {
1787                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1788                         if (pool[i] != NULL)
1789                                 filter_free_iobuf(pool[i]);
1790                 }
1791                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1792                 filter->fo_iobuf_pool = NULL;
1793         }
1794         EXIT;
1795 }
1796
1797 /*
1798  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1799  */
1800 static int filter_iobuf_pool_init(struct filter_obd *filter)
1801 {
1802         void **pool;
1803
1804         ENTRY;
1805
1806
1807         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1808                       GFP_KERNEL);
1809         if (filter->fo_iobuf_pool == NULL)
1810                 RETURN(-ENOMEM);
1811
1812         filter->fo_iobuf_count = OSS_THREADS_MAX;
1813
1814         RETURN(0);
1815 }
1816
1817 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1818  * many threads there will be so we allocate a large empty array and only
1819  * fill in those slots that are actually in use.
1820  * If we haven't allocated a pool entry for this thread before, do so now. */
1821 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1822 {
1823         int thread_id                    = oti ? oti->oti_thread_id : -1;
1824         struct filter_iobuf  *pool       = NULL;
1825         struct filter_iobuf **pool_place = NULL;
1826
1827         if (thread_id >= 0) {
1828                 LASSERT(thread_id < filter->fo_iobuf_count);
1829                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1830         }
1831
1832         if (unlikely(pool == NULL)) {
1833                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1834                                           PTLRPC_MAX_BRW_PAGES);
1835                 if (pool_place != NULL)
1836                         *pool_place = pool;
1837         }
1838
1839         return pool;
1840 }
1841
1842 /* mount the file system (secretly).  lustre_cfg parameters are:
1843  * 1 = device
1844  * 2 = fstype
1845  * 3 = flags: failover=f, failout=n
1846  * 4 = mount options
1847  */
1848 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1849                         void *option)
1850 {
1851         struct filter_obd *filter = &obd->u.filter;
1852         struct vfsmount *mnt;
1853         struct lustre_mount_info *lmi;
1854         struct obd_uuid uuid;
1855         __u8 *uuid_ptr;
1856         char *str, *label;
1857         char ns_name[48];
1858         int rc, i;
1859         ENTRY;
1860
1861         if (lcfg->lcfg_bufcount < 3 ||
1862             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1863             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1864                 RETURN(-EINVAL);
1865
1866         lmi = server_get_mount(obd->obd_name);
1867         if (lmi) {
1868                 /* We already mounted in lustre_fill_super.
1869                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1870                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1871                 mnt = lmi->lmi_mnt;
1872                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1873         } else {
1874                 /* old path - used by lctl */
1875                 CERROR("Using old MDS mount method\n");
1876                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1877                                     MS_NOATIME|MS_NODIRATIME,
1878                                     lustre_cfg_string(lcfg, 1), option);
1879                 if (IS_ERR(mnt)) {
1880                         rc = PTR_ERR(mnt);
1881                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1882                                            lustre_cfg_string(lcfg, 1), rc);
1883                         RETURN(rc);
1884                 }
1885
1886                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1887         }
1888         if (IS_ERR(obd->obd_fsops))
1889                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1890
1891         rc = filter_iobuf_pool_init(filter);
1892         if (rc != 0)
1893                 GOTO(err_ops, rc);
1894
1895         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1896
1897         /* failover is the default */
1898         obd->obd_replayable = 1;
1899
1900         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1901                 str = lustre_cfg_string(lcfg, 3);
1902                 if (strchr(str, 'n')) {
1903                         CWARN("%s: recovery disabled\n", obd->obd_name);
1904                         obd->obd_replayable = 0;
1905                 }
1906         }
1907
1908         filter->fo_vfsmnt = mnt;
1909         obd->u.obt.obt_sb = mnt->mnt_sb;
1910         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1911         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1912
1913         fsfilt_setup(obd, obd->u.obt.obt_sb);
1914
1915         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1916         obd->obd_lvfs_ctxt.pwdmnt = mnt;
1917         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1918         obd->obd_lvfs_ctxt.fs = get_ds();
1919         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1920
1921         sema_init(&filter->fo_init_lock, 1);
1922         filter->fo_committed_group = 0;
1923
1924         rc = filter_prep(obd);
1925         if (rc)
1926                 GOTO(err_ops, rc);
1927
1928         filter->fo_destroys_in_progress = 0;
1929         for (i = 0; i < 32; i++)
1930                 sema_init(&filter->fo_create_locks[i], 1);
1931
1932         spin_lock_init(&filter->fo_translock);
1933         spin_lock_init(&filter->fo_objidlock);
1934         CFS_INIT_LIST_HEAD(&filter->fo_export_list);
1935         sema_init(&filter->fo_alloc_lock, 1);
1936         init_brw_stats(&filter->fo_filter_stats);
1937         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1938         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1939         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1940
1941         INIT_LIST_HEAD(&filter->fo_llog_list);
1942         spin_lock_init(&filter->fo_llog_list_lock);
1943
1944         filter->fo_sptlrpc_lock = RW_LOCK_UNLOCKED;
1945         sptlrpc_rule_set_init(&filter->fo_sptlrpc_rset);
1946
1947         filter->fo_fl_oss_capa = 0;
1948         INIT_LIST_HEAD(&filter->fo_capa_keys);
1949         filter->fo_capa_hash = init_capa_hash();
1950         if (filter->fo_capa_hash == NULL)
1951                 GOTO(err_ops, rc = -ENOMEM);
1952
1953         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1954         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1955                                                 LDLM_NAMESPACE_GREEDY);
1956         if (obd->obd_namespace == NULL)
1957                 GOTO(err_post, rc = -ENOMEM);
1958         obd->obd_namespace->ns_lvbp = obd;
1959         obd->obd_namespace->ns_lvbo = &filter_lvbo;
1960         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1961
1962         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1963                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1964
1965         rc = llog_cat_initialize(obd, &obd->obd_olg, 1, NULL);
1966         if (rc) {
1967                 CERROR("failed to setup llogging subsystems\n");
1968                 GOTO(err_post, rc);
1969         }
1970
1971         rc = lquota_setup(filter_quota_interface_ref, obd);
1972         if (rc)
1973                 GOTO(err_post, rc);
1974
1975         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1976         if (uuid_ptr != NULL) {
1977                 class_uuid_unparse(uuid_ptr, &uuid);
1978                 str = uuid.uuid;
1979         } else {
1980                 str = "no UUID";
1981         }
1982
1983         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1984
1985         if (obd->obd_recovering) {
1986                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1987                               "recovery until %d %s reconnect, or if no clients"
1988                               " reconnect for %d:%.02d; during that time new "
1989                               "clients will not be allowed to connect. "
1990                               "Recovery progress can be monitored by watching "
1991                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1992                               obd->obd_name, lustre_cfg_string(lcfg, 1),
1993                               label ?: "", label ? "/" : "", str,
1994                               obd->obd_max_recoverable_clients,
1995                               (obd->obd_max_recoverable_clients == 1)
1996                               ? "client" : "clients",
1997                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
1998                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
1999                               obd->obd_name);
2000         } else {
2001                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
2002                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2003                               label ?: "", label ? "/" : "", str,
2004                               obd->obd_replayable ? "enabled" : "disabled");
2005         }
2006
2007         RETURN(0);
2008
2009 err_post:
2010         filter_post(obd);
2011 err_ops:
2012         fsfilt_put_ops(obd->obd_fsops);
2013         filter_iobuf_pool_done(filter);
2014 err_mntput:
2015         server_put_mount(obd->obd_name, mnt);
2016         obd->u.obt.obt_sb = 0;
2017         return rc;
2018 }
2019
2020 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2021 {
2022         struct lprocfs_static_vars lvars;
2023         unsigned long addr;
2024         struct page *page;
2025         int rc;
2026
2027         CLASSERT(offsetof(struct obd_device, u.obt) ==
2028                  offsetof(struct obd_device, u.filter.fo_obt));
2029
2030         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2031                 RETURN(-EINVAL);
2032
2033         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2034         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2035         if (!page)
2036                 RETURN(-ENOMEM);
2037         addr = (unsigned long)cfs_page_address(page);
2038         clear_page((void *)addr);
2039
2040         /* lprocfs must be setup before the filter so state can be safely added
2041          * to /proc incrementally as the filter is setup */
2042         lprocfs_filter_init_vars(&lvars);
2043         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2044             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2045                 /* Init obdfilter private stats here */
2046                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2047                                      LPROCFS_CNTR_AVGMINMAX,
2048                                      "read_bytes", "bytes");
2049                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2050                                      LPROCFS_CNTR_AVGMINMAX,
2051                                      "write_bytes", "bytes");
2052
2053                 lproc_filter_attach_seqstat(obd);
2054                 obd->obd_proc_exports = proc_mkdir("exports",
2055                                                    obd->obd_proc_entry);
2056         }
2057
2058         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2059                LUSTRE_CFG_BUFLEN(lcfg, 4));
2060         rc = filter_common_setup(obd, lcfg, (void *)addr);
2061         OBD_PAGE_FREE(page);
2062
2063         if (rc) {
2064                 lprocfs_obd_cleanup(obd);
2065                 lprocfs_free_obd_stats(obd);
2066         }
2067
2068         return rc;
2069 }
2070
2071 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2072 static struct llog_operations filter_size_orig_logops = {
2073         lop_setup: llog_obd_origin_setup,
2074         lop_cleanup: llog_obd_origin_cleanup,
2075         lop_add: llog_obd_origin_add
2076 };
2077
2078 static int filter_llog_init(struct obd_device *obd, int group,
2079                             struct obd_device *tgt, int count,
2080                             struct llog_catid *catid,
2081                             struct obd_uuid *uuid)
2082 {
2083         struct filter_obd *filter = &obd->u.filter;
2084         struct obd_llog_group *olg;
2085         struct llog_ctxt *ctxt;
2086         int rc;
2087         ENTRY;
2088
2089         olg = filter_find_olg(obd, group);
2090         if (IS_ERR(olg))
2091                 RETURN(PTR_ERR(olg));
2092
2093         if (group == OBD_LLOG_GROUP) {
2094                 LASSERT(filter->fo_lcm == NULL);
2095                 OBD_ALLOC(filter->fo_lcm, sizeof(struct llog_commit_master));
2096                 if (!filter->fo_lcm)
2097                         RETURN(-ENOMEM);
2098
2099                 rc = llog_init_commit_master((struct llog_commit_master *)
2100                                              filter->fo_lcm);
2101                 if (rc)
2102                         GOTO(cleanup, rc);
2103
2104         filter_mds_ost_repl_logops = llog_client_ops;
2105         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2106         filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2107         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2108         } else {
2109                 LASSERT(filter->fo_lcm != NULL);
2110         }
2111         rc = llog_setup(obd, olg, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2112                         &filter_mds_ost_repl_logops);
2113         if (rc)
2114                 GOTO(cleanup, rc);
2115
2116         /* FIXME - assign unlink_cb for filter's recovery */
2117         LASSERT(olg);
2118         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2119
2120         LASSERT(ctxt != NULL);
2121         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2122         ctxt->loc_lcm = obd->u.filter.fo_lcm;
2123         rc = llog_start_commit_thread(ctxt->loc_lcm);
2124         llog_ctxt_put(ctxt);
2125         if (rc)
2126                 GOTO(cleanup, rc);
2127
2128         rc = llog_setup(obd, olg, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2129                         &filter_size_orig_logops);
2130
2131 cleanup:
2132         if (rc) {
2133                 llog_cleanup_commit_master(filter->fo_lcm, 0);
2134                 OBD_FREE(filter->fo_lcm, sizeof(struct llog_commit_master));
2135                 filter->fo_lcm = NULL;
2136         }
2137         RETURN(rc);
2138 }
2139
2140 static int filter_group_llog_finish(struct obd_llog_group *olg)
2141 {
2142         struct llog_ctxt *ctxt;
2143         int rc = 0, rc2 = 0;
2144         ENTRY;
2145
2146         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
2147         if (ctxt)
2148                 rc = llog_cleanup(ctxt);
2149
2150         ctxt = llog_group_get_ctxt(olg, LLOG_SIZE_ORIG_CTXT);
2151         if (ctxt)
2152                 rc2 = llog_cleanup(ctxt);
2153         if (!rc)
2154                 rc = rc2;
2155
2156         RETURN(rc);
2157 }
2158
2159 static int filter_llog_finish(struct obd_device *obd, int count)
2160 {
2161         int rc;
2162         ENTRY;
2163
2164         if (obd->u.filter.fo_lcm) {
2165                 llog_cleanup_commit_master((struct llog_commit_master *)
2166                                            obd->u.filter.fo_lcm, 0);
2167                 OBD_FREE(obd->u.filter.fo_lcm,
2168                          sizeof(struct llog_commit_master));
2169                 obd->u.filter.fo_lcm = NULL;
2170         }
2171         /* finish obd llog group */
2172         rc = filter_group_llog_finish(&obd->obd_olg);
2173
2174         RETURN(rc);
2175 }
2176
2177 struct obd_llog_group *filter_find_olg(struct obd_device *obd, int group)
2178 {
2179         struct obd_llog_group *olg, *nolg;
2180         struct filter_obd *filter;
2181         int rc;
2182
2183         filter = &obd->u.filter;
2184
2185         if (group == OBD_LLOG_GROUP)
2186                 RETURN(&obd->obd_olg);
2187
2188         spin_lock(&filter->fo_llog_list_lock);
2189         list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2190                 if (olg->olg_group == group) {
2191                         spin_unlock(&filter->fo_llog_list_lock);
2192                         RETURN(olg);
2193                 }
2194         }
2195         spin_unlock(&filter->fo_llog_list_lock);
2196
2197         OBD_ALLOC_PTR(olg);
2198         if (olg == NULL)
2199                 RETURN(ERR_PTR(-ENOMEM));
2200
2201         llog_group_init(olg, group);
2202         spin_lock(&filter->fo_llog_list_lock);
2203         list_for_each_entry(nolg, &filter->fo_llog_list, olg_list) {
2204                 LASSERT(nolg->olg_group != group);
2205         }
2206         list_add(&olg->olg_list, &filter->fo_llog_list);
2207         spin_unlock(&filter->fo_llog_list_lock);
2208
2209         rc = llog_cat_initialize(obd, olg, 1, NULL);
2210         if (rc) {
2211                 spin_lock(&filter->fo_llog_list_lock);
2212                 list_del(&olg->olg_list);
2213                 spin_unlock(&filter->fo_llog_list_lock);
2214                 OBD_FREE_PTR(olg);
2215                 RETURN(ERR_PTR(rc));
2216         }
2217         CDEBUG(D_OTHER, "%s: new llog group %u (0x%p)\n",
2218                obd->obd_name, group, olg);
2219
2220         RETURN(olg);
2221 }
2222
2223 static int filter_llog_connect(struct obd_export *exp,
2224                                struct llogd_conn_body *body)
2225 {
2226         struct obd_device *obd = exp->exp_obd;
2227         struct llog_ctxt *ctxt;
2228         struct obd_llog_group *olg;
2229         int rc;
2230         ENTRY;
2231
2232         CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2233                 (unsigned) body->lgdc_logid.lgl_ogr,
2234                 (unsigned) body->lgdc_logid.lgl_oid,
2235                 (unsigned) body->lgdc_logid.lgl_ogen);
2236
2237         olg = filter_find_olg(obd, body->lgdc_logid.lgl_ogr);
2238         if (IS_ERR(olg))
2239                 RETURN(PTR_ERR(olg));
2240         llog_group_set_export(olg, exp);
2241
2242         ctxt = llog_group_get_ctxt(olg, body->lgdc_ctxt_idx);
2243         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2244                  body->lgdc_ctxt_idx);
2245         rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2246                           &body->lgdc_gen, NULL);
2247         llog_ctxt_put(ctxt);
2248         if (rc != 0)
2249                 CERROR("failed to connect rc %d idx %d\n", rc,
2250                                 body->lgdc_ctxt_idx);
2251
2252         RETURN(rc);
2253 }
2254
2255 static int filter_llog_preclean (struct obd_device *obd)
2256 {
2257         struct obd_llog_group *olg;
2258         struct filter_obd *filter;
2259         int rc = 0;
2260         ENTRY;
2261
2262         rc = obd_llog_finish(obd, 0);
2263         if (rc)
2264                 CERROR("failed to cleanup llogging subsystem\n");
2265
2266         filter = &obd->u.filter;
2267         spin_lock(&filter->fo_llog_list_lock);
2268         while (!list_empty(&filter->fo_llog_list)) {
2269                 olg = list_entry(filter->fo_llog_list.next,
2270                                  struct obd_llog_group, olg_list);
2271                 list_del(&olg->olg_list);
2272                 spin_unlock(&filter->fo_llog_list_lock);
2273
2274                 rc = filter_group_llog_finish(olg);
2275                 if (rc)
2276                         CERROR("failed to cleanup llogging subsystem for %u\n",
2277                                olg->olg_group);
2278                 OBD_FREE_PTR(olg);
2279                 spin_lock(&filter->fo_llog_list_lock);
2280         }
2281         spin_unlock(&filter->fo_llog_list_lock);
2282
2283         RETURN(rc);
2284 }
2285
2286 static int filter_precleanup(struct obd_device *obd,
2287                              enum obd_cleanup_stage stage)
2288 {
2289         int rc = 0;
2290         ENTRY;
2291
2292         switch(stage) {
2293         case OBD_CLEANUP_EARLY:
2294                 break;
2295         case OBD_CLEANUP_EXPORTS:
2296                 target_cleanup_recovery(obd);
2297                 rc = filter_llog_preclean(obd);
2298                 break;
2299         case OBD_CLEANUP_SELF_EXP:
2300                 break;
2301         case OBD_CLEANUP_OBD:
2302                 break;
2303         }
2304         RETURN(rc);
2305 }
2306
2307 static int filter_cleanup(struct obd_device *obd)
2308 {
2309         struct filter_obd *filter = &obd->u.filter;
2310         ENTRY;
2311
2312         if (obd->obd_fail)
2313                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2314                               "will be preserved.\n", obd->obd_name);
2315
2316         if (!list_empty(&obd->obd_exports)) {
2317                 CERROR("%s: still has clients!\n", obd->obd_name);
2318                 class_disconnect_exports(obd);
2319                 if (!list_empty(&obd->obd_exports)) {
2320                         CERROR("still has exports after forced cleanup?\n");
2321                         RETURN(-EBUSY);
2322                 }
2323         }
2324
2325         lprocfs_obd_cleanup(obd);
2326         lprocfs_free_obd_stats(obd);
2327         lquota_cleanup(filter_quota_interface_ref, obd);
2328
2329         /* Stop recovery before namespace cleanup. */
2330         target_stop_recovery_thread(obd);
2331         target_cleanup_recovery(obd);
2332
2333         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2334
2335         sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
2336
2337         if (obd->u.obt.obt_sb == NULL)
2338                 RETURN(0);
2339
2340         filter_post(obd);
2341
2342         shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2343
2344         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2345
2346         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2347         obd->u.obt.obt_sb = NULL;
2348
2349         fsfilt_put_ops(obd->obd_fsops);
2350
2351         filter_iobuf_pool_done(filter);
2352
2353         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2354
2355         RETURN(0);
2356 }
2357
2358 static int filter_connect_internal(struct obd_export *exp,
2359                                    struct obd_connect_data *data)
2360 {
2361         if (!data)
2362                 RETURN(0);
2363
2364         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2365                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2366                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2367                data->ocd_connect_flags, data->ocd_version,
2368                data->ocd_grant, data->ocd_index);
2369
2370         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2371         exp->exp_connect_flags = data->ocd_connect_flags;
2372         data->ocd_version = LUSTRE_VERSION_CODE;
2373
2374         if ((exp->exp_connect_flags & OBD_CONNECT_FID) == 0) {
2375                 CWARN("%s: OST requires FID support (flag="LPX64
2376                       "), but client not\n",
2377                       exp->exp_obd->obd_name,
2378                       exp->exp_connect_flags);
2379                 RETURN(-EBADF);
2380         }
2381
2382         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2383                 struct filter_export_data *fed = &exp->exp_filter_data;
2384                 obd_size left, want;
2385
2386                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2387                 left = filter_grant_space_left(exp);
2388                 want = data->ocd_grant;
2389                 filter_grant(exp, fed->fed_grant, want, left);
2390                 data->ocd_grant = fed->fed_grant;
2391                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2392
2393                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2394                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2395                        exp->exp_client_uuid.uuid, exp,
2396                        data->ocd_grant, want, left);
2397         }
2398
2399         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2400                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2401                 struct lr_server_data *lsd = filter->fo_fsd;
2402                 int index = le32_to_cpu(lsd->lsd_ost_index);
2403
2404                 if (!(lsd->lsd_feature_compat &
2405                       cpu_to_le32(OBD_COMPAT_OST))) {
2406                         /* this will only happen on the first connect */
2407                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2408                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2409                         filter_update_server_data(exp->exp_obd,
2410                                                   filter->fo_rcvd_filp, lsd, 1);
2411                 } else if (index != data->ocd_index) {
2412                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2413                                            " %u doesn't match actual OST index"
2414                                            " %u in last_rcvd file, bad "
2415                                            "configuration?\n",
2416                                            obd_export_nid2str(exp), index,
2417                                            data->ocd_index);
2418                         RETURN(-EBADF);
2419                 }
2420         }
2421
2422         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2423                 data->ocd_brw_size = 65536;
2424         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2425                 data->ocd_brw_size = min(data->ocd_brw_size,
2426                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2427                 LASSERT(data->ocd_brw_size);
2428         }
2429
2430         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2431          * FIXME: We don't strictly need the COMPAT flag for that,
2432          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2433          * FIXME: We needed it for the index, as index 0 is valid. */
2434
2435         RETURN(0);
2436 }
2437
2438 static int filter_reconnect(const struct lu_env *env,
2439                             struct obd_export *exp, struct obd_device *obd,
2440                             struct obd_uuid *cluuid,
2441                             struct obd_connect_data *data)
2442 {
2443         int rc;
2444         ENTRY;
2445
2446         if (exp == NULL || obd == NULL || cluuid == NULL)
2447                 RETURN(-EINVAL);
2448
2449         rc = filter_connect_internal(exp, data);
2450
2451         RETURN(rc);
2452 }
2453
2454 /* nearly identical to mds_connect */
2455 static int filter_connect(const struct lu_env *env,
2456                           struct lustre_handle *conn, struct obd_device *obd,
2457                           struct obd_uuid *cluuid,
2458                           struct obd_connect_data *data)
2459 {
2460         struct lvfs_run_ctxt saved;
2461         struct obd_export *exp;
2462         struct filter_export_data *fed;
2463         struct filter_client_data *fcd = NULL;
2464         __u32 group;
2465         int rc;
2466         ENTRY;
2467
2468         if (conn == NULL || obd == NULL || cluuid == NULL)
2469                 RETURN(-EINVAL);
2470
2471         rc = class_connect(conn, obd, cluuid);
2472         if (rc)
2473                 RETURN(rc);
2474         exp = class_conn2export(conn);
2475         LASSERT(exp != NULL);
2476
2477         fed = &exp->exp_filter_data;
2478
2479         rc = filter_connect_internal(exp, data);
2480         if (rc)
2481                 GOTO(cleanup, rc);
2482
2483         filter_export_stats_init(obd, exp);
2484         group = data->ocd_group;
2485         if (obd->obd_replayable) {
2486                 OBD_ALLOC(fcd, sizeof(*fcd));
2487                 if (!fcd) {
2488                         CERROR("filter: out of memory for client data\n");
2489                         GOTO(cleanup, rc = -ENOMEM);
2490                 }
2491
2492                 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2493                 fed->fed_fcd = fcd;
2494                 fed->fed_fcd->fcd_group = group;
2495                 rc = filter_client_add(obd, exp, -1);
2496                 if (rc)
2497                         GOTO(cleanup, rc);
2498         }
2499         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2500                obd->obd_name, exp->exp_handle.h_cookie, group);
2501         if (group == 0)
2502                 GOTO(cleanup, rc);
2503
2504         if (fed->fed_group != 0 && fed->fed_group != group) {
2505                 CERROR("!!! This export (nid %s) used object group %d "
2506                        "earlier; now it's trying to use group %d!  This could "
2507                        "be a bug in the MDS.  Tell CFS.\n",
2508                        obd_export_nid2str(exp), fed->fed_group, group);
2509                 GOTO(cleanup, rc = -EPROTO);
2510         }
2511         fed->fed_group = group;
2512
2513         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2514         rc = filter_read_groups(obd, group, 1);
2515         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2516         if (rc != 0) {
2517                 CERROR("can't read group %u\n", group);
2518                 GOTO(cleanup, rc);
2519         }
2520
2521         GOTO(cleanup, rc);
2522
2523 cleanup:
2524         if (rc) {
2525                 if (fcd) {
2526                         OBD_FREE(fcd, sizeof(*fcd));
2527                         fed->fed_fcd = NULL;
2528                 }
2529                 class_disconnect(exp);
2530         } else {
2531                 class_export_put(exp);
2532         }
2533
2534         RETURN(rc);
2535 }
2536
2537 /* Do extra sanity checks for grant accounting.  We do this at connect,
2538  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2539  * always get rid of it or turn it off when we know accounting is good. */
2540 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2541 {
2542         struct filter_export_data *fed;
2543         struct obd_export *exp;
2544         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2545         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2546         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2547
2548         if (list_empty(&obd->obd_exports))
2549                 return;
2550
2551         /* We don't want to do this for large machines that do lots of
2552            mounts or unmounts.  It burns... */
2553         if (obd->obd_num_exports > 100)
2554                 return;
2555
2556         spin_lock(&obd->obd_osfs_lock);
2557         spin_lock(&obd->obd_dev_lock);
2558         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2559                 int error = 0;
2560                 fed = &exp->exp_filter_data;
2561                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2562                     fed->fed_dirty < 0)
2563                         error = 1;
2564                 if (maxsize > 0) { /* we may not have done a statfs yet */
2565                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2566                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2567                                  exp->exp_client_uuid.uuid, exp,
2568                                  fed->fed_grant, fed->fed_pending, maxsize);
2569                         LASSERTF(fed->fed_dirty <= maxsize,
2570                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2571                                  exp->exp_client_uuid.uuid, exp,
2572                                  fed->fed_dirty, maxsize);
2573                 }
2574                 if (error)
2575                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2576                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2577                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2578                 else
2579                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2580                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2581                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2582                 tot_granted += fed->fed_grant + fed->fed_pending;
2583                 tot_pending += fed->fed_pending;
2584                 tot_dirty += fed->fed_dirty;
2585         }
2586         fo_tot_granted = obd->u.filter.fo_tot_granted;
2587         fo_tot_pending = obd->u.filter.fo_tot_pending;
2588         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2589         spin_unlock(&obd->obd_dev_lock);
2590         spin_unlock(&obd->obd_osfs_lock);
2591
2592         /* Do these assertions outside the spinlocks so we don't kill system */
2593         if (tot_granted != fo_tot_granted)
2594                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2595                        func, tot_granted, fo_tot_granted);
2596         if (tot_pending != fo_tot_pending)
2597                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2598                        func, tot_pending, fo_tot_pending);
2599         if (tot_dirty != fo_tot_dirty)
2600                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2601                        func, tot_dirty, fo_tot_dirty);
2602         if (tot_pending > tot_granted)
2603                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2604                        func, tot_pending, tot_granted);
2605         if (tot_granted > maxsize)
2606                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2607                        func, tot_granted, maxsize);
2608         if (tot_dirty > maxsize)
2609                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2610                        func, tot_dirty, maxsize);
2611 }
2612
2613 /* Remove this client from the grant accounting totals.  We also remove
2614  * the export from the obd device under the osfs and dev locks to ensure
2615  * that the filter_grant_sanity_check() calculations are always valid.
2616  * The client should do something similar when it invalidates its import. */
2617 static void filter_grant_discard(struct obd_export *exp)
2618 {
2619         struct obd_device *obd = exp->exp_obd;
2620         struct filter_obd *filter = &obd->u.filter;
2621         struct filter_export_data *fed = &exp->exp_filter_data;
2622
2623         spin_lock(&obd->obd_osfs_lock);
2624         spin_lock(&obd->obd_dev_lock);
2625         list_del_init(&exp->exp_obd_chain);
2626         spin_unlock(&obd->obd_dev_lock);
2627
2628         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2629                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2630                  obd->obd_name, filter->fo_tot_granted,
2631                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2632         filter->fo_tot_granted -= fed->fed_grant;
2633         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2634                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2635                  obd->obd_name, filter->fo_tot_pending,
2636                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2637         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2638         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2639                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2640                  obd->obd_name, filter->fo_tot_dirty,
2641                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2642         filter->fo_tot_dirty -= fed->fed_dirty;
2643         fed->fed_dirty = 0;
2644         fed->fed_grant = 0;
2645
2646         spin_unlock(&obd->obd_osfs_lock);
2647 }
2648
2649 static int filter_destroy_export(struct obd_export *exp)
2650 {
2651         ENTRY;
2652
2653         if (exp->exp_filter_data.fed_pending)
2654                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2655                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2656                        exp, exp->exp_filter_data.fed_pending);
2657
2658         /* Not ported yet the b1_6 quota functionality
2659          * lquota_clearinfo(filter_quota_interface_ref, exp, exp->exp_obd);
2660          */
2661
2662         target_destroy_export(exp);
2663
2664         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2665                 RETURN(0);
2666
2667         lprocfs_exp_cleanup(exp);
2668
2669         if (exp->exp_obd->obd_replayable)
2670                 filter_client_free(exp);
2671         else
2672                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2673
2674         filter_grant_discard(exp);
2675         filter_fmd_cleanup(exp);
2676
2677         if (!(exp->exp_flags & OBD_OPT_FORCE))
2678                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2679
2680         RETURN(0);
2681 }
2682
2683 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2684 {
2685         struct obd_llog_group *olg_min, *olg;
2686         struct filter_obd *filter;
2687         int worked = 0, group;
2688         struct llog_ctxt *ctxt;
2689         ENTRY;
2690
2691         filter = &obd->u.filter;
2692
2693         /* we can't sync log holding spinlock. also, we do not want to get
2694          * into livelock. so we do following: loop over MDS's exports in
2695          * group order and skip already synced llogs -bzzz */
2696         do {
2697                 /* look for group with min. number, but > worked */
2698                 olg_min = NULL;
2699                 group = 1 << 30;
2700                 spin_lock(&filter->fo_llog_list_lock);
2701                 list_for_each_entry(olg, &filter->fo_llog_list, olg_list) {
2702                         if (olg->olg_group <= worked) {
2703                                 /* this group is already synced */
2704                                 continue;
2705                         }
2706                         if (group < olg->olg_group) {
2707                                 /* we have group with smaller number to sync */
2708                                 continue;
2709                         }
2710                         /* store current minimal group */
2711                         olg_min = olg;
2712                         group = olg->olg_group;
2713                 }
2714                 spin_unlock(&filter->fo_llog_list_lock);
2715
2716                 if (olg_min == NULL)
2717                         break;
2718
2719                 worked = olg_min->olg_group;
2720                 if (olg_min->olg_exp &&
2721                     (dexp == olg_min->olg_exp || dexp == NULL)) {
2722                         int err;
2723                         ctxt = llog_group_get_ctxt(olg_min,
2724                                                 LLOG_MDS_OST_REPL_CTXT);
2725                         LASSERT(ctxt != NULL);
2726                         err = llog_sync(ctxt, olg_min->olg_exp);
2727                         llog_ctxt_put(ctxt);
2728                         if (err)
2729                                 CERROR("error flushing logs to MDS: rc %d\n",
2730                                        err);
2731                 }
2732         } while (olg_min != NULL);
2733 }
2734
2735 /* also incredibly similar to mds_disconnect */
2736 static int filter_disconnect(struct obd_export *exp)
2737 {
2738         struct obd_device *obd = exp->exp_obd;
2739         int rc;
2740         ENTRY;
2741
2742         LASSERT(exp);
2743         class_export_get(exp);
2744
2745         if (!(exp->exp_flags & OBD_OPT_FORCE))
2746                 filter_grant_sanity_check(obd, __FUNCTION__);
2747         filter_grant_discard(exp);
2748
2749         /* Disconnect early so that clients can't keep using export */
2750         rc = class_disconnect(exp);
2751         if (exp->exp_obd->obd_namespace != NULL)
2752                 ldlm_cancel_locks_for_export(exp);
2753
2754         fsfilt_sync(obd, obd->u.obt.obt_sb);
2755
2756         /* flush any remaining cancel messages out to the target */
2757         filter_sync_llogs(obd, exp);
2758         class_export_put(exp);
2759         RETURN(rc);
2760 }
2761
2762 /* reverse import is changed, sync all cancels */
2763 static void filter_revimp_update(struct obd_export *exp)
2764 {
2765         ENTRY;
2766
2767         LASSERT(exp);
2768         class_export_get(exp);
2769
2770         /* flush any remaining cancel messages out to the target */
2771         filter_sync_llogs(exp->exp_obd, exp);
2772         class_export_put(exp);
2773         EXIT;
2774 }
2775
2776 static int filter_ping(struct obd_export *exp)
2777 {
2778         filter_fmd_expire(exp);
2779
2780         return 0;
2781 }
2782
2783 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2784                                   const char *what, int quiet)
2785 {
2786         struct dentry *dchild = NULL;
2787         obd_gr group = 0;
2788
2789         if (oa->o_valid & OBD_MD_FLGROUP)
2790                 group = oa->o_gr;
2791
2792         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2793
2794         if (IS_ERR(dchild)) {
2795                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
2796                        what, group, oa->o_id);
2797                 RETURN(dchild);
2798         }
2799
2800         if (dchild->d_inode == NULL) {
2801                 if (!quiet)
2802                         CERROR("%s: %s on non-existent object: "LPU64"\n",
2803                                obd->obd_name, what, oa->o_id);
2804                 f_dput(dchild);
2805                 RETURN(ERR_PTR(-ENOENT));
2806         }
2807
2808         return dchild;
2809 }
2810
2811 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2812 {
2813         struct dentry *dentry = NULL;
2814         struct obd_device *obd;
2815         int rc = 0;
2816         ENTRY;
2817
2818         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2819                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
2820         if (rc)
2821                 RETURN(rc);
2822
2823         obd = class_exp2obd(exp);
2824         if (obd == NULL) {
2825                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2826                 RETURN(-EINVAL);
2827         }
2828
2829         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2830         if (IS_ERR(dentry))
2831                 RETURN(PTR_ERR(dentry));
2832
2833         /* Limit the valid bits in the return data to what we actually use */
2834         oinfo->oi_oa->o_valid = OBD_MD_FLID;
2835         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2836
2837         f_dput(dentry);
2838         RETURN(rc);
2839 }
2840
2841 /* this should be enabled/disabled in condition to enabled/disabled large
2842  * inodes (fast EAs) in backing store FS. */
2843 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2844                         void *handle, struct obdo *oa)
2845 {
2846         struct obd_device *obd = exp->exp_obd;
2847         int rc = 0;
2848         ENTRY;
2849
2850         if (oa->o_valid & OBD_MD_FLFID) {
2851                 struct filter_fid ff;
2852
2853                 if (!(oa->o_valid & OBD_MD_FLGROUP))
2854                         oa->o_gr = 0;
2855                 /* packing fid and converting it to LE for storing into EA.
2856                  * Here ->o_stripe_idx should be filled by LOV and rest of
2857                  * fields - by client. */
2858                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2859                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2860                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2861                 ff.ff_objid = cpu_to_le64(oa->o_id);
2862                 ff.ff_group = cpu_to_le64(oa->o_gr);
2863
2864                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2865                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2866                        oa->o_generation, oa->o_id, oa->o_gr);
2867
2868                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2869                 if (rc)
2870                         CERROR("store fid in object failed! rc: %d\n", rc);
2871         } else {
2872                 CDEBUG(D_HA, "OSS object without fid info!\n");
2873         }
2874
2875         RETURN(rc);
2876 }
2877
2878 /* this is called from filter_truncate() until we have filter_punch() */
2879 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2880                             struct obdo *oa, struct obd_trans_info *oti)
2881 {
2882         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2883         struct llog_cookie *fcc = NULL;
2884         struct filter_obd *filter;
2885         int rc, err, locked = 0, sync = 0;
2886         unsigned int ia_valid;
2887         struct inode *inode;
2888         struct iattr iattr;
2889         void *handle;
2890         ENTRY;
2891
2892         LASSERT(dentry != NULL);
2893         LASSERT(!IS_ERR(dentry));
2894
2895         inode = dentry->d_inode;
2896         LASSERT(inode != NULL);
2897
2898         filter = &exp->exp_obd->u.filter;
2899         iattr_from_obdo(&iattr, oa, oa->o_valid);
2900         ia_valid = iattr.ia_valid;
2901
2902         if (oa->o_valid & OBD_MD_FLCOOKIE) {
2903                 OBD_ALLOC(fcc, sizeof(*fcc));
2904                 if (fcc != NULL)
2905                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2906         }
2907
2908         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2909                 DQUOT_INIT(inode);
2910                 LOCK_INODE_MUTEX(inode);
2911                 locked = 1;
2912         }
2913
2914         /* If the inode still has SUID+SGID bits set (see filter_precreate())
2915          * then we will accept the UID+GID sent by the client during write for
2916          * initializing the ownership of this inode.  We only allow this to
2917          * happen once so clear these bits in setattr. In 2.6 kernels it is
2918          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2919          * the flags that are actually being set. */
2920         if (ia_valid & (ATTR_UID | ATTR_GID)) {
2921                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2922                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2923
2924                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2925                         if (!(ia_valid & ATTR_MODE)) {
2926                                 iattr.ia_mode = inode->i_mode;
2927                                 iattr.ia_valid |= ATTR_MODE;
2928                         }
2929                         iattr.ia_mode &= ~S_ISUID;
2930                 }
2931                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2932                         if (!(iattr.ia_valid & ATTR_MODE)) {
2933                                 iattr.ia_mode = inode->i_mode;
2934                                 iattr.ia_valid |= ATTR_MODE;
2935                         }
2936                         iattr.ia_mode &= ~S_ISGID;
2937                 }
2938
2939                 orig_ids[USRQUOTA] = inode->i_uid;
2940                 orig_ids[GRPQUOTA] = inode->i_gid;
2941                 handle = fsfilt_start_log(exp->exp_obd, inode,
2942                                           FSFILT_OP_SETATTR, oti, 1);
2943                 if (IS_ERR(handle))
2944                         GOTO(out_unlock, rc = PTR_ERR(handle));
2945
2946                 /* update inode EA only once when inode is suid bit marked. As
2947                  * on 2.6.x UID and GID may be set separately, we check here
2948                  * only one of them to avoid double setting. */
2949                 if (inode->i_mode & S_ISUID)
2950                         filter_update_fidea(exp, inode, handle, oa);
2951         } else {
2952                 handle = fsfilt_start(exp->exp_obd, inode,
2953                                       FSFILT_OP_SETATTR, oti);
2954                 if (IS_ERR(handle))
2955                         GOTO(out_unlock, rc = PTR_ERR(handle));
2956         }
2957         if (oa->o_valid & OBD_MD_FLFLAGS) {
2958                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2959                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2960         } else {
2961                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2962                 if (fcc != NULL)
2963                         /* set cancel cookie callback function */
2964                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2965                                                      filter_cancel_cookies_cb,
2966                                                      fcc);
2967         }
2968
2969         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2970                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2971
2972         /* The truncate might have used up our transaction credits.  Make
2973          * sure we have one left for the last_rcvd update. */
2974         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2975
2976         rc = filter_finish_transno(exp, oti, rc, sync);
2977         if (sync) {
2978                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
2979                 fcc = NULL;
2980         }
2981
2982         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2983         if (err) {
2984                 CERROR("error on commit, err = %d\n", err);
2985                 if (!rc)
2986                         rc = err;
2987         } else {
2988                 fcc = NULL;
2989         }
2990
2991         if (locked) {
2992                 /* Let's flush truncated page on disk immediately, then we can
2993                  * avoid need to search for page aliases before directio writes
2994                  * and this sort of stuff at expense of somewhat slower
2995                  * truncates not on a page boundary. I believe this is the only
2996                  * place in filter code that can lead to pages getting to
2997                  * pagecache so far. */
2998                 filter_clear_truncated_page(inode);
2999                 UNLOCK_INODE_MUTEX(inode);
3000                 locked = 0;
3001         }
3002
3003         EXIT;
3004 out_unlock:
3005         if (locked)
3006                 UNLOCK_INODE_MUTEX(inode);
3007
3008         if (fcc)
3009                 OBD_FREE(fcc, sizeof(*fcc));
3010
3011         /* trigger quota release */
3012         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
3013                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
3014                 int rc2 = lquota_adjust(filter_quota_interface_ref,
3015                                         exp->exp_obd, cur_ids,
3016                                         orig_ids, rc, FSFILT_OP_SETATTR);
3017                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3018                        "filter adjust qunit. (rc:%d)\n", rc2);
3019         }
3020         return rc;
3021 }
3022
3023 /* this is called from filter_truncate() until we have filter_punch() */
3024 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3025                    struct obd_trans_info *oti)
3026 {
3027         struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3028                                                 oinfo->oi_oa->o_gr, 0 } };
3029         struct filter_mod_data *fmd;
3030         struct lvfs_run_ctxt saved;
3031         struct filter_obd *filter;
3032         struct ldlm_resource *res;
3033         struct dentry *dentry;
3034         int rc;
3035         ENTRY;
3036
3037         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3038                               oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3039         if (rc)
3040                 RETURN(rc);
3041
3042         dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3043                                     __FUNCTION__, 1);
3044         if (IS_ERR(dentry))
3045                 RETURN(PTR_ERR(dentry));
3046
3047         filter = &exp->exp_obd->u.filter;
3048         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3049         lock_kernel();
3050
3051         if (oinfo->oi_oa->o_valid &
3052             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3053                 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3054                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3055                         fmd->fmd_mactime_xid = oti->oti_xid;
3056                 filter_fmd_put(exp, fmd);
3057         }
3058
3059         /* setting objects attributes (including owner/group) */
3060         rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3061         if (rc)
3062                 GOTO(out_unlock, rc);
3063
3064         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3065                                 &res_id, LDLM_EXTENT, 0);
3066
3067         if (res != NULL) {
3068                 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3069                 ldlm_resource_putref(res);
3070         }
3071
3072         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3073
3074         /* Quota release need uid/gid info */
3075         obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3076                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3077
3078         EXIT;
3079 out_unlock:
3080         unlock_kernel();
3081         f_dput(dentry);
3082         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3083         return rc;
3084 }
3085
3086 /* XXX identical to osc_unpackmd */
3087 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3088                            struct lov_mds_md *lmm, int lmm_bytes)
3089 {
3090         int lsm_size;
3091         ENTRY;
3092
3093         if (lmm != NULL) {
3094                 if (lmm_bytes < sizeof (*lmm)) {
3095                         CERROR("lov_mds_md too small: %d, need %d\n",
3096                                lmm_bytes, (int)sizeof(*lmm));
3097                         RETURN(-EINVAL);
3098                 }
3099                 /* XXX LOV_MAGIC etc check? */
3100
3101                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3102                         CERROR("lov_mds_md: zero lmm_object_id\n");
3103                         RETURN(-EINVAL);
3104                 }
3105         }
3106
3107         lsm_size = lov_stripe_md_size(1);
3108         if (lsmp == NULL)
3109                 RETURN(lsm_size);
3110
3111         if (*lsmp != NULL && lmm == NULL) {
3112                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3113                 OBD_FREE(*lsmp, lsm_size);
3114                 *lsmp = NULL;
3115                 RETURN(0);
3116         }
3117
3118         if (*lsmp == NULL) {
3119                 OBD_ALLOC(*lsmp, lsm_size);
3120                 if (*lsmp == NULL)
3121                         RETURN(-ENOMEM);
3122
3123                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3124                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3125                         OBD_FREE(*lsmp, lsm_size);
3126                         RETURN(-ENOMEM);
3127                 }
3128                 loi_init((*lsmp)->lsm_oinfo[0]);
3129         }
3130
3131         if (lmm != NULL) {
3132                 /* XXX zero *lsmp? */
3133                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3134                 LASSERT((*lsmp)->lsm_object_id);
3135         }
3136
3137         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3138
3139         RETURN(lsm_size);
3140 }
3141
3142 /* caller must hold fo_create_locks[oa->o_gr] */
3143 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3144                                       struct filter_obd *filter)
3145 {
3146         struct obdo doa; /* XXX obdo on stack */
3147         obd_id last, id;
3148         int rc;
3149         ENTRY;
3150
3151         LASSERT(oa);
3152         LASSERT(oa->o_gr != 0);
3153         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3154         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3155
3156         memset(&doa, 0, sizeof(doa));
3157
3158         doa.o_valid |= OBD_MD_FLGROUP;
3159         doa.o_gr = oa->o_gr;
3160         doa.o_mode = S_IFREG;
3161
3162         if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3163                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3164                        exp->exp_obd->obd_name, doa.o_gr);
3165                 RETURN(0);
3166         }
3167
3168         last = filter_last_id(filter, doa.o_gr);
3169
3170         CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3171                exp->exp_obd->obd_name, oa->o_id + 1, last);
3172
3173         for (id = last; id > oa->o_id; id--) {
3174                 doa.o_id = id;
3175                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3176                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3177                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3178                                id, rc);
3179                 filter_set_last_id(filter, id - 1, doa.o_gr);
3180                 /* update last_id on disk periodically so that if we restart
3181                  * we don't need to re-scan all of the just-deleted objects. */
3182                 if ((id & 511) == 0)
3183                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3184         }
3185
3186         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3187                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3188
3189         rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3190         clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3191
3192         RETURN(rc);
3193 }
3194
3195 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3196                             obd_gr group, int *num);
3197 /* returns a negative error or a nonnegative number of files to create */
3198 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3199                                    obd_gr group, struct obd_trans_info *oti)
3200 {
3201         struct obd_device *obd = exp->exp_obd;
3202         struct filter_obd *filter = &obd->u.filter;
3203         int diff, rc;
3204         ENTRY;
3205
3206         /* delete orphans request */
3207         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3208                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3209                         CERROR("%s: dropping old orphan cleanup request\n",
3210                                obd->obd_name);
3211                         RETURN(0);
3212                 }
3213                 /* This causes inflight precreates to abort and drop lock */
3214                 set_bit(group, &filter->fo_destroys_in_progress);
3215                 down(&filter->fo_create_locks[group]);
3216                 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3217                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3218                                exp->exp_obd->obd_name, group);
3219                         up(&filter->fo_create_locks[group]);
3220                         RETURN(0);
3221                 }
3222                 diff = oa->o_id - filter_last_id(filter, group);
3223                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3224                        filter_last_id(filter, group), diff);
3225
3226                 if (-diff > OST_MAX_PRECREATE) {
3227                         CERROR("%s: ignoring bogus orphan destroy request: "
3228                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3229                                oa->o_id, filter_last_id(filter, group));
3230                         /* FIXME: should reset precreate_next_id on MDS */
3231                         GOTO(out, rc = -EINVAL);
3232                 }
3233                 if (diff < 0) {
3234                         rc = filter_destroy_precreated(exp, oa, filter);
3235                         if (rc)
3236                                 CERROR("%s: unable to write lastobjid, but "
3237                                        "orphans were deleted\n", obd->obd_name);
3238                         GOTO(out, rc);
3239                 } else {
3240                         /* XXX: Used by MDS for the first time! */
3241                         clear_bit(group, &filter->fo_destroys_in_progress);
3242                 }
3243         } else {
3244                 down(&filter->fo_create_locks[group]);
3245                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3246                         CERROR("%s: dropping old precreate request\n",
3247                                obd->obd_name);
3248                         GOTO(out, rc = 0);
3249                 }
3250                 /* only precreate if group == 0 and o_id is specfied */
3251                 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3252                         diff = 1;
3253                 else
3254                         diff = oa->o_id - filter_last_id(filter, group);
3255                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3256                        filter_last_id(filter, group), diff);
3257
3258                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3259                          oa->o_id, filter_last_id(filter, group), diff);
3260         }
3261
3262         if (diff > 0) {
3263                 oa->o_id = filter_last_id(&obd->u.filter, group);
3264                 rc = filter_precreate(obd, oa, group, &diff);
3265                 oa->o_id = filter_last_id(&obd->u.filter, group);
3266                 oa->o_gr = group;
3267                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3268                 GOTO(out, rc);
3269         }
3270         /* else diff == 0 */
3271         GOTO(out, rc = 0);
3272 out:
3273         up(&filter->fo_create_locks[group]);
3274         return rc;
3275 }
3276
3277 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3278                          __u64 max_age)
3279 {
3280         struct filter_obd *filter = &obd->u.filter;
3281         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3282         int rc;
3283         ENTRY;
3284
3285         /* at least try to account for cached pages.  its still racey and
3286          * might be under-reporting if clients haven't announced their
3287          * caches with brw recently */
3288         spin_lock(&obd->obd_osfs_lock);
3289         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3290         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3291         spin_unlock(&obd->obd_osfs_lock);
3292
3293         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3294                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3295                filter->fo_tot_dirty, filter->fo_tot_granted,
3296                filter->fo_tot_pending,
3297                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3298
3299         filter_grant_sanity_check(obd, __FUNCTION__);
3300
3301         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3302                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
3303                                  osfs->os_bsize - 1) >> blockbits));
3304
3305         /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3306          * stop creating files on MDS if OST is not good shape to create
3307          * objects.*/
3308         osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3309                 EROFS : 0;
3310         RETURN(rc);
3311 }
3312
3313 static int filter_use_existing_obj(struct obd_device *obd,
3314                                    struct dentry *dchild, void **handle,
3315                                    int *cleanup_phase)
3316 {
3317         struct inode *inode = dchild->d_inode;
3318         struct iattr iattr;
3319         int rc;
3320
3321         if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3322                 return 0;
3323
3324         *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3325         if (IS_ERR(*handle))
3326                 return PTR_ERR(*handle);
3327
3328         iattr.ia_valid = ATTR_MODE;
3329         iattr.ia_mode = S_ISUID | S_ISGID |0666;
3330         rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3331         if (rc == 0)
3332                 *cleanup_phase = 3;
3333
3334         return rc;
3335 }
3336
3337
3338 /* We rely on the fact that only one thread will be creating files in a given
3339  * group at a time, which is why we don't need an atomic filter_get_new_id.
3340  * Even if we had that atomic function, the following race would exist:
3341  *
3342  * thread 1: gets id x from filter_next_id
3343  * thread 2: gets id (x + 1) from filter_next_id
3344  * thread 2: creates object (x + 1)
3345  * thread 1: tries to create object x, gets -ENOSPC
3346  *
3347  * Caller must hold fo_create_locks[group]
3348  */
3349 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3350                             obd_gr group, int *num)
3351 {
3352         struct dentry *dchild = NULL, *dparent = NULL;
3353         struct filter_obd *filter;
3354         struct obd_statfs *osfs;
3355         int err = 0, rc = 0, recreate_obj = 0, i;
3356         unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3357         obd_id next_id;
3358         void *handle = NULL;
3359         ENTRY;
3360
3361         filter = &obd->u.filter;
3362
3363         LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3364
3365         OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3366
3367         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3368             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3369                 recreate_obj = 1;
3370         } else {
3371                 OBD_ALLOC(osfs, sizeof(*osfs));
3372                 if (osfs == NULL)
3373                         RETURN(-ENOMEM);
3374                 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3375                 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3376                         CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3377                                LPU64"\n", obd->obd_name, osfs->os_bavail <<
3378                                filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3379                         *num = 0;
3380                         rc = -ENOSPC;
3381                 }
3382                 OBD_FREE(osfs, sizeof(*osfs));
3383                 if (rc)
3384                         RETURN(rc);
3385         }
3386
3387         CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3388                " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3389
3390         for (i = 0; i < *num && err == 0; i++) {
3391                 int cleanup_phase = 0;
3392
3393                 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3394                         CWARN("%s: create aborted by destroy\n",
3395                               obd->obd_name);
3396                         rc = -EAGAIN;
3397                         break;
3398                 }
3399
3400                 if (recreate_obj) {
3401                         __u64 last_id;
3402                         next_id = oa->o_id;
3403                         last_id = filter_last_id(filter, group);
3404                         if (next_id > last_id) {
3405                                 CERROR("Error: Trying to recreate obj greater"
3406                                        "than last id "LPD64" > "LPD64"\n",
3407                                        next_id, last_id);
3408                                 GOTO(cleanup, rc = -EINVAL);
3409                         }
3410                 } else
3411                         next_id = filter_last_id(filter, group) + 1;
3412
3413                 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3414
3415                 dparent = filter_parent_lock(obd, group, next_id);
3416                 if (IS_ERR(dparent))
3417                         GOTO(cleanup, rc = PTR_ERR(dparent));
3418                 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3419
3420                 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3421                 if (IS_ERR(dchild))
3422                         GOTO(cleanup, rc = PTR_ERR(dchild));
3423                 cleanup_phase = 2;  /* f_dput(dchild) */
3424
3425                 if (dchild->d_inode != NULL) {
3426                         /* This would only happen if lastobjid was bad on disk*/
3427                         /* Could also happen if recreating missing obj but it
3428                          * already exists. */
3429                         if (recreate_obj) {
3430                                 CERROR("%s: recreating existing object %.*s?\n",
3431                                        obd->obd_name, dchild->d_name.len,
3432                                        dchild->d_name.name);
3433                         } else {
3434                                 /* Use these existing objects if they are
3435                                  * zero length. */
3436                                 if (dchild->d_inode->i_size == 0) {
3437                                         rc = filter_use_existing_obj(obd,dchild,
3438                                                       &handle, &cleanup_phase);
3439                                         if (rc == 0)
3440                                                 goto set_last_id;
3441                                         else
3442                                                 GOTO(cleanup, rc);
3443                                 }
3444
3445                                 CERROR("%s: Serious error: objid %.*s already "
3446                                        "exists; is this filesystem corrupt?\n",
3447                                        obd->obd_name, dchild->d_name.len,
3448                                        dchild->d_name.name);
3449                                 LBUG();
3450                         }
3451                         GOTO(cleanup, rc = -EEXIST);
3452                 }
3453
3454                 handle = fsfilt_start_log(obd, dparent->d_inode,
3455                                           FSFILT_OP_CREATE, NULL, 1);
3456                 if (IS_ERR(handle))
3457                         GOTO(cleanup, rc = PTR_ERR(handle));
3458                 cleanup_phase = 3;
3459
3460                 /* We mark object SUID+SGID to flag it for accepting UID+GID
3461                  * from client on first write.  Currently the permission bits
3462                  * on the OST are never used, so this is OK. */
3463                 rc = ll_vfs_create(dparent->d_inode, dchild,
3464                                    S_IFREG |  S_ISUID | S_ISGID | 0666, NULL);
3465                 if (rc) {
3466                         CERROR("create failed rc = %d\n", rc);
3467                         GOTO(cleanup, rc);
3468                 }
3469
3470 set_last_id:
3471                 if (!recreate_obj) {
3472                         filter_set_last_id(filter, next_id, group);
3473                         err = filter_update_last_objid(obd, group, 0);
3474                         if (err)
3475                                 CERROR("unable to write lastobjid "
3476                                        "but file created\n");
3477                 }
3478
3479         cleanup:
3480                 switch(cleanup_phase) {
3481                 case 3:
3482                         err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3483                         if (err) {
3484                                 CERROR("error on commit, err = %d\n", err);
3485                                 if (!rc)
3486                                         rc = err;
3487                         }
3488                 case 2:
3489                         f_dput(dchild);
3490                 case 1:
3491                         filter_parent_unlock(dparent);
3492                 case 0:
3493                         break;
3494                 }
3495
3496                 if (rc)
3497                         break;
3498                 if (time_after(jiffies, enough_time)) {
3499                         CDEBUG(D_RPCTRACE,
3500                                "%s: precreate slow - want %d got %d \n",
3501                                obd->obd_name, *num, i);
3502                         break;
3503                 }
3504         }
3505         *num = i;
3506
3507         CDEBUG(D_RPCTRACE,
3508                "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3509                obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3510
3511         RETURN(rc);
3512 }
3513
3514 static int filter_create(struct obd_export *exp, struct obdo *oa,
3515                          struct lov_stripe_md **ea, struct obd_trans_info *oti)
3516 {
3517         struct filter_export_data *fed;
3518         struct obd_device *obd = NULL;
3519         struct filter_obd *filter;
3520         struct lvfs_run_ctxt saved;
3521         struct lov_stripe_md *lsm = NULL;
3522         int rc = 0, diff, group = oa->o_gr;
3523         ENTRY;
3524
3525         if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3526                 CERROR("!!! nid %s sent invalid object group %d\n",
3527                         obd_export_nid2str(exp), group);
3528                 RETURN(-EINVAL);
3529         }
3530
3531         obd = exp->exp_obd;
3532         fed = &exp->exp_filter_data;
3533         filter = &obd->u.filter;
3534
3535         if (fed->fed_group != group) {
3536                 CERROR("!!! this export (nid %s) used object group %d "
3537                         "earlier; now it's trying to use group %d!  This could "
3538                         "be a bug in the MDS.  Tell CFS.\n",
3539                         obd_export_nid2str(exp), fed->fed_group, group);
3540                 RETURN(-ENOTUNIQ);
3541         }
3542
3543         CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3544                oa->o_gr, oa->o_id);
3545         if (ea != NULL) {
3546                 lsm = *ea;
3547                 if (lsm == NULL) {
3548                         rc = obd_alloc_memmd(exp, &lsm);
3549                         if (rc < 0)
3550                                 RETURN(rc);
3551                 }
3552         }
3553
3554         obd = exp->exp_obd;
3555         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3556
3557         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3558             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3559                 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3560                         CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3561                                oa->o_id, filter_last_id(filter,
3562                                                         oa->o_gr));
3563                         rc = -EINVAL;
3564                 } else {
3565                         diff = 1;
3566                         down(&filter->fo_create_locks[oa->o_gr]);
3567                         rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3568                         up(&filter->fo_create_locks[oa->o_gr]);
3569                 }
3570         } else {
3571                 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3572         }
3573
3574         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3575         if (rc && ea != NULL && *ea != lsm) {
3576                 obd_free_memmd(exp, &lsm);
3577         } else if (rc == 0 && ea != NULL) {
3578                 /* XXX LOV STACKING: the lsm that is passed to us from
3579                  * LOV does not have valid lsm_oinfo data structs, so
3580                  * don't go touching that.  This needs to be fixed in a
3581                  * big way. */
3582                 lsm->lsm_object_id = oa->o_id;
3583                 *ea = lsm;
3584         }
3585
3586         RETURN(rc);
3587 }
3588
3589 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3590                    struct lov_stripe_md *md, struct obd_trans_info *oti,
3591                    struct obd_export *md_exp)
3592 {
3593         unsigned int qcids[MAXQUOTAS] = {0, 0};
3594         struct obd_device *obd;
3595         struct filter_obd *filter;
3596         struct dentry *dchild = NULL, *dparent = NULL;
3597         struct lvfs_run_ctxt saved;
3598         void *handle = NULL;
3599         struct llog_cookie *fcc = NULL;
3600         int rc, rc2, cleanup_phase = 0, sync = 0;
3601         struct iattr iattr;
3602         ENTRY;
3603
3604         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3605
3606         obd = exp->exp_obd;
3607         filter = &obd->u.filter;
3608
3609         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3610         cleanup_phase = 1;
3611
3612         dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3613         if (IS_ERR(dchild))
3614                 GOTO(cleanup, rc = PTR_ERR(dchild));
3615         cleanup_phase = 2;
3616
3617         if (dchild->d_inode == NULL) {
3618                 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3619                        oa->o_id);
3620                 /* If object already gone, cancel cookie right now */
3621                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3622                         struct llog_ctxt *ctxt;
3623                         struct obd_llog_group *olg;
3624                         fcc = obdo_logcookie(oa);
3625                         olg = filter_find_olg(obd, oa->o_gr);
3626                         if (IS_ERR(olg))
3627                                 GOTO(cleanup, rc = PTR_ERR(olg));
3628                         llog_group_set_export(olg, exp);
3629
3630                         ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
3631                         llog_cancel(ctxt, NULL, 1, fcc, 0);
3632                         llog_ctxt_put(ctxt);
3633                         fcc = NULL; /* we didn't allocate fcc, don't free it */
3634                 }
3635                 GOTO(cleanup, rc = -ENOENT);
3636         }
3637
3638         filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3639
3640         /* Our MDC connection is established by the MDS to us */
3641         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3642                 OBD_ALLOC(fcc, sizeof(*fcc));
3643                 if (fcc != NULL)
3644                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3645         }
3646         DQUOT_INIT(dchild->d_inode);
3647
3648         /* we're gonna truncate it first in order to avoid possible deadlock:
3649          *      P1                      P2
3650          * open trasaction      open transaction
3651          * down(i_zombie)       down(i_zombie)
3652          *                      restart transaction
3653          * (see BUG 4180) -bzzz
3654          */
3655         LOCK_INODE_MUTEX(dchild->d_inode);
3656         handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3657                                   NULL, 1);
3658         if (IS_ERR(handle)) {
3659                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3660                 GOTO(cleanup, rc = PTR_ERR(handle));
3661         }
3662
3663         iattr.ia_valid = ATTR_SIZE;
3664         iattr.ia_size = 0;
3665         rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3666         rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3667         UNLOCK_INODE_MUTEX(dchild->d_inode);
3668         if (rc)
3669                 GOTO(cleanup, rc);
3670         if (rc2)
3671                 GOTO(cleanup, rc = rc2);
3672
3673         /* We don't actually need to lock the parent until we are unlinking
3674          * here, and not while truncating above.  That avoids holding the
3675          * parent lock for a long time during truncate, which can block other
3676          * threads from doing anything to objects in that directory. bug 7171 */
3677         dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3678         if (IS_ERR(dparent))
3679                 GOTO(cleanup, rc = PTR_ERR(dparent));
3680         cleanup_phase = 3; /* filter_parent_unlock */
3681
3682         LOCK_INODE_MUTEX(dchild->d_inode);
3683         handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3684         if (IS_ERR(handle)) {
3685                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3686                 GOTO(cleanup, rc = PTR_ERR(handle));
3687         }
3688         cleanup_phase = 4; /* fsfilt_commit */
3689
3690         /* Quota release need uid/gid of inode */
3691         obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3692
3693         filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3694
3695         /* this drops dchild->d_inode->i_mutex unconditionally */
3696         rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3697
3698         EXIT;
3699 cleanup:
3700         switch(cleanup_phase) {
3701         case 4:
3702                 if (fcc != NULL)
3703                         sync = fsfilt_add_journal_cb(obd, 0, oti ?
3704                                                      oti->oti_handle : handle,
3705                                                      filter_cancel_cookies_cb,
3706                                                      fcc);
3707                 /* If add_journal_cb failed, then filter_finish_transno
3708                  * will commit the handle and we will do a sync
3709                  * on commit. then we call callback directly to free
3710                  * the fcc.
3711                  */
3712                 rc = filter_finish_transno(exp, oti, rc, sync);
3713                 if (sync) {
3714                         filter_cancel_cookies_cb(obd, 0, fcc, rc);
3715                         fcc = NULL;
3716                 }
3717                 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3718                 if (rc2) {
3719                         CERROR("error on commit, err = %d\n", rc2);
3720                         if (!rc)
3721                                 rc = rc2;
3722                 } else {
3723                         fcc = NULL;
3724                 }
3725         case 3:
3726                 filter_parent_unlock(dparent);
3727         case 2:
3728                 f_dput(dchild);
3729                 if (fcc != NULL)
3730                         OBD_FREE(fcc, sizeof(*fcc));
3731         case 1:
3732                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3733                 break;
3734         default:
3735                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3736                 LBUG();
3737         }
3738
3739         /* trigger quota release */
3740         qcids[USRQUOTA] = oa->o_uid;
3741         qcids[GRPQUOTA] = oa->o_gid;
3742         rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3743                             FSFILT_OP_UNLINK);
3744
3745         if (rc2)
3746                 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3747         return rc;
3748 }
3749
3750 /* NB start and end are used for punch, but not truncate */
3751 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3752                            struct obd_trans_info *oti,
3753                            struct ptlrpc_request_set *rqset)
3754 {
3755         int rc;
3756         ENTRY;
3757
3758         if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3759                 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3760                        oinfo->oi_policy.l_extent.end);
3761                 RETURN(-EFAULT);
3762         }
3763
3764         CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3765                ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3766                oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3767
3768         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3769                               oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3770         if (rc)
3771                 RETURN(rc);
3772
3773         oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3774         rc = filter_setattr(exp, oinfo, oti);
3775         RETURN(rc);
3776 }
3777
3778 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3779                        struct lov_stripe_md *lsm, obd_off start, obd_off end,
3780                        void *capa)
3781 {
3782         struct lvfs_run_ctxt saved;
3783         struct filter_obd *filter;
3784         struct dentry *dentry;
3785         int rc, rc2;
3786         ENTRY;
3787
3788         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3789                               (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3790         if (rc)
3791                 RETURN(rc);
3792
3793         filter = &exp->exp_obd->u.filter;
3794
3795         /* an objid of zero is taken to mean "sync whole filesystem" */
3796         if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3797                 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3798                 /* flush any remaining cancel messages out to the target */
3799                 filter_sync_llogs(exp->exp_obd, exp);
3800                 RETURN(rc);
3801         }
3802
3803         dentry = filter_oa2dentry(exp->exp_obd, oa);
3804         if (IS_ERR(dentry))
3805                 RETURN(PTR_ERR(dentry));
3806
3807         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3808
3809         LOCK_INODE_MUTEX(dentry->d_inode);
3810
3811         rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3812         if (rc == 0) {
3813                 /* just any file to grab fsync method - "file" arg unused */
3814                 struct file *file = filter->fo_rcvd_filp;
3815
3816                 if (file->f_op && file->f_op->fsync)
3817                         rc = file->f_op->fsync(NULL, dentry, 1);
3818
3819                 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3820                 if (!rc)
3821                         rc = rc2;
3822         }
3823         UNLOCK_INODE_MUTEX(dentry->d_inode);
3824
3825         oa->o_valid = OBD_MD_FLID;
3826         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3827
3828         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3829
3830         f_dput(dentry);
3831         RETURN(rc);
3832 }
3833
3834 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3835                            void *key, __u32 *vallen, void *val)
3836 {
3837         struct obd_device *obd;
3838         ENTRY;
3839
3840         obd = class_exp2obd(exp);
3841         if (obd == NULL) {
3842                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3843                 RETURN(-EINVAL);
3844         }
3845
3846         if (KEY_IS("blocksize")) {
3847                 __u32 *blocksize = val;
3848                 if (blocksize) {
3849                         if (*vallen < sizeof(*blocksize))
3850                                 RETURN(-EOVERFLOW);
3851                         *blocksize = obd->u.obt.obt_sb->s_blocksize;
3852                 }
3853                 *vallen = sizeof(*blocksize);
3854                 RETURN(0);
3855         }
3856
3857         if (KEY_IS("blocksize_bits")) {
3858                 __u32 *blocksize_bits = val;
3859                 if (blocksize_bits) {
3860                         if (*vallen < sizeof(*blocksize_bits))
3861                                 RETURN(-EOVERFLOW);
3862                         *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3863                 }
3864                 *vallen = sizeof(*blocksize_bits);
3865                 RETURN(0);
3866         }
3867
3868         if (KEY_IS("last_id")) {
3869                 obd_id *last_id = val;
3870                 /* FIXME: object groups */
3871                 if (last_id) {
3872                         if (*vallen < sizeof(*last_id))
3873                                 RETURN(-EOVERFLOW);
3874                         *last_id = filter_last_id(&obd->u.filter,
3875                                                   exp->exp_filter_data.fed_group);
3876                 }
3877                 *vallen = sizeof(*last_id);
3878                 RETURN(0);
3879         }
3880
3881         CDEBUG(D_IOCTL, "invalid key\n");
3882         RETURN(-EINVAL);
3883 }
3884
3885 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3886                                  void *key, __u32 vallen, void *val,
3887                                  struct ptlrpc_request_set *set)
3888 {
3889         struct obd_device *obd;
3890         struct obd_llog_group *olg;
3891         struct llog_ctxt *ctxt;
3892         int rc = 0, group;
3893         ENTRY;
3894
3895         obd = exp->exp_obd;
3896         if (obd == NULL) {
3897                 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3898                 RETURN(-EINVAL);
3899         }
3900
3901         if (KEY_IS(KEY_CAPA_KEY)) {
3902                 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3903                 if (rc)
3904                         CERROR("filter update capability key failed: %d\n", rc);
3905                 RETURN(rc);
3906         }
3907
3908         if (KEY_IS(KEY_REVIMP_UPD)) {
3909                 filter_revimp_update(exp);
3910                 RETURN(0);
3911         }
3912
3913         if (!KEY_IS(KEY_MDS_CONN))
3914                 RETURN(-EINVAL);
3915
3916         LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3917                       obd_export_nid2str(exp));
3918         obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3919
3920         /* setup llog imports */
3921         LASSERT(val != NULL);
3922         group = (int)(*(__u32 *)val);
3923         LASSERT(group >= FILTER_GROUP_MDS0);
3924
3925         olg = filter_find_olg(obd, group);
3926         if (IS_ERR(olg))
3927                 RETURN(PTR_ERR(olg));
3928         llog_group_set_export(olg, exp);
3929
3930         ctxt = llog_group_get_ctxt(olg, LLOG_MDS_OST_REPL_CTXT);
3931         LASSERTF(ctxt != NULL, "ctxt is null\n"),
3932
3933         rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3934         llog_ctxt_put(ctxt);
3935
3936         lquota_setinfo(filter_quota_interface_ref, exp, obd);
3937
3938         RETURN(rc);
3939 }
3940
3941 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3942                      int len, void *karg, void *uarg)
3943 {
3944         struct obd_device *obd = exp->exp_obd;
3945         struct obd_ioctl_data *data = karg;
3946         int rc = 0;
3947
3948         switch (cmd) {
3949         case OBD_IOC_ABORT_RECOVERY: {
3950                 CERROR("aborting recovery for device %s\n", obd->obd_name);
3951                 target_stop_recovery_thread(obd);
3952                 RETURN(0);
3953         }
3954
3955         case OBD_IOC_SYNC: {
3956                 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
3957                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3958                 RETURN(rc);
3959         }
3960
3961         case OBD_IOC_SET_READONLY: {
3962                 void *handle;
3963                 struct super_block *sb = obd->u.obt.obt_sb;
3964                 struct inode *inode = sb->s_root->d_inode;
3965                 BDEVNAME_DECLARE_STORAGE(tmp);
3966                 CERROR("*** setting device %s read-only ***\n",
3967                        ll_bdevname(sb, tmp));
3968
3969                 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3970                 if (!IS_ERR(handle))
3971                         rc = fsfilt_commit(obd, inode, handle, 1);
3972
3973                 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3974                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3975
3976                 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
3977                 RETURN(0);
3978         }
3979
3980         case OBD_IOC_CATLOGLIST: {
3981                 rc = llog_catalog_list(obd, 1, data);
3982                 RETURN(rc);
3983         }
3984
3985         case OBD_IOC_LLOG_CANCEL:
3986         case OBD_IOC_LLOG_REMOVE:
3987         case OBD_IOC_LLOG_INFO:
3988         case OBD_IOC_LLOG_PRINT: {
3989                 /* FIXME to be finished */
3990                 RETURN(-EOPNOTSUPP);
3991 /*
3992                 struct llog_ctxt *ctxt = NULL;
3993
3994                 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3995                 rc = llog_ioctl(ctxt, cmd, data);
3996                 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3997
3998                 RETURN(rc);
3999 */
4000         }
4001
4002
4003         default:
4004                 RETURN(-EINVAL);
4005         }
4006         RETURN(0);
4007 }
4008
4009 static int filter_health_check(struct obd_device *obd)
4010 {
4011 #ifdef USE_HEALTH_CHECK_WRITE
4012         struct filter_obd *filter = &obd->u.filter;
4013 #endif
4014         int rc = 0;
4015
4016         /*
4017          * health_check to return 0 on healthy
4018          * and 1 on unhealthy.
4019          */
4020         if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
4021                 rc = 1;
4022
4023 #ifdef USE_HEALTH_CHECK_WRITE
4024         LASSERT(filter->fo_health_check_filp != NULL);
4025         rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
4026 #endif
4027         return rc;
4028 }
4029
4030 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4031                                              void *data)
4032 {
4033         return filter_fid2dentry(data, NULL, gr, id);
4034 }
4035
4036 static int filter_process_config(struct obd_device *obd, obd_count len,
4037                                  void *buf)
4038 {
4039         struct lustre_cfg *lcfg = buf;
4040         struct lprocfs_static_vars lvars;
4041         int rc = 0;
4042
4043         switch (lcfg->lcfg_command) {
4044         case LCFG_SPTLRPC_CONF: {
4045                 struct filter_obd       *filter = &obd->u.filter;
4046                 struct sptlrpc_conf_log *log;
4047                 struct sptlrpc_rule_set  tmp_rset;
4048
4049                 log = sptlrpc_conf_log_extract(lcfg);
4050                 if (IS_ERR(log)) {
4051                         rc = PTR_ERR(log);
4052                         break;
4053                 }
4054
4055                 sptlrpc_rule_set_init(&tmp_rset);
4056
4057                 rc = sptlrpc_rule_set_from_log(&tmp_rset, log);
4058                 if (rc) {
4059                         CERROR("obd %s: failed get sptlrpc rules: %d\n",
4060                                obd->obd_name, rc);
4061                         break;
4062                 }
4063
4064                 write_lock(&filter->fo_sptlrpc_lock);
4065                 sptlrpc_rule_set_free(&filter->fo_sptlrpc_rset);
4066                 filter->fo_sptlrpc_rset = tmp_rset;
4067                 write_unlock(&filter->fo_sptlrpc_lock);
4068
4069                 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4070                 break;
4071         }
4072         default:
4073                 lprocfs_filter_init_vars(&lvars);
4074
4075                 rc = class_process_proc_param(PARAM_OST, lvars.obd_vars,
4076                                               lcfg, obd);
4077                 break;
4078         }
4079
4080         return rc;
4081 }
4082
4083 static struct lvfs_callback_ops filter_lvfs_ops = {
4084         l_fid2dentry:     filter_lvfs_fid2dentry,
4085 };
4086
4087 static struct obd_ops filter_obd_ops = {
4088         .o_owner          = THIS_MODULE,
4089         .o_get_info       = filter_get_info,
4090         .o_set_info_async = filter_set_info_async,
4091         .o_setup          = filter_setup,
4092         .o_precleanup     = filter_precleanup,
4093         .o_cleanup        = filter_cleanup,
4094         .o_connect        = filter_connect,
4095         .o_reconnect      = filter_reconnect,
4096         .o_disconnect     = filter_disconnect,
4097         .o_ping           = filter_ping,
4098         .o_init_export    = filter_init_export,
4099         .o_destroy_export = filter_destroy_export,
4100         .o_statfs         = filter_statfs,
4101         .o_getattr        = filter_getattr,
4102         .o_unpackmd       = filter_unpackmd,
4103         .o_create         = filter_create,
4104         .o_setattr        = filter_setattr,
4105         .o_destroy        = filter_destroy,
4106         .o_brw            = filter_brw,
4107         .o_punch          = filter_truncate,
4108         .o_sync           = filter_sync,
4109         .o_preprw         = filter_preprw,
4110         .o_commitrw       = filter_commitrw,
4111         .o_llog_init      = filter_llog_init,
4112         .o_llog_connect   = filter_llog_connect,
4113         .o_llog_finish    = filter_llog_finish,
4114         .o_iocontrol      = filter_iocontrol,
4115         .o_health_check   = filter_health_check,
4116         .o_process_config = filter_process_config,
4117 };
4118
4119 quota_interface_t *filter_quota_interface_ref;
4120 extern quota_interface_t filter_quota_interface;
4121
4122 static int __init obdfilter_init(void)
4123 {
4124         struct lprocfs_static_vars lvars;
4125         int rc;
4126
4127         lprocfs_filter_init_vars(&lvars);
4128
4129         request_module("lquota");
4130         OBD_ALLOC(obdfilter_created_scratchpad,
4131                   OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4132                   sizeof(*obdfilter_created_scratchpad));
4133         if (obdfilter_created_scratchpad == NULL)
4134                 return -ENOMEM;
4135
4136         ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4137                                              sizeof(struct filter_mod_data),
4138                                              0, 0);
4139         if (!ll_fmd_cachep)
4140                 GOTO(out, rc = -ENOMEM);
4141
4142         filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4143         init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4144
4145         rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4146                                  LUSTRE_OST_NAME, NULL);
4147         if (rc) {
4148                 int err;
4149
4150                 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4151                 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4152                 ll_fmd_cachep = NULL;
4153 out:
4154                 if (filter_quota_interface_ref)
4155                         PORTAL_SYMBOL_PUT(filter_quota_interface);
4156
4157                 OBD_FREE(obdfilter_created_scratchpad,
4158                          OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4159                          sizeof(*obdfilter_created_scratchpad));
4160         }
4161
4162         return rc;
4163 }
4164
4165 static void __exit obdfilter_exit(void)
4166 {
4167         if (filter_quota_interface_ref)
4168                 PORTAL_SYMBOL_PUT(filter_quota_interface);
4169
4170         if (ll_fmd_cachep) {
4171                 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4172                 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4173                 ll_fmd_cachep = NULL;
4174         }
4175
4176         class_unregister_type(LUSTRE_OST_NAME);
4177         OBD_FREE(obdfilter_created_scratchpad,
4178                  OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4179                  sizeof(*obdfilter_created_scratchpad));
4180 }
4181
4182 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4183 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4184 MODULE_LICENSE("GPL");
4185
4186 module_init(obdfilter_init);
4187 module_exit(obdfilter_exit);