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 3
  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_HA;
  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         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_fl_oss_capa = 0;
1945         INIT_LIST_HEAD(&filter->fo_capa_keys);
1946         filter->fo_capa_hash = init_capa_hash();
1947         if (filter->fo_capa_hash == NULL)
1948                 GOTO(err_ops, rc = -ENOMEM);
1949
1950         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1951         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1952                                                 LDLM_NAMESPACE_GREEDY);
1953         if (obd->obd_namespace == NULL)
1954                 GOTO(err_post, rc = -ENOMEM);
1955         obd->obd_namespace->ns_lvbp = obd;
1956         obd->obd_namespace->ns_lvbo = &filter_lvbo;
1957         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1958
1959         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1960                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1961
1962         rc = llog_cat_initialize(obd, NULL, 1, NULL);
1963         if (rc) {
1964                 CERROR("failed to setup llogging subsystems\n");
1965                 GOTO(err_post, rc);
1966         }
1967
1968         rc = lquota_setup(filter_quota_interface_ref, obd);
1969         if (rc)
1970                 GOTO(err_post, rc);
1971
1972         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1973         if (uuid_ptr != NULL) {
1974                 class_uuid_unparse(uuid_ptr, &uuid);
1975                 str = uuid.uuid;
1976         } else {
1977                 str = "no UUID";
1978         }
1979
1980         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1981
1982         if (obd->obd_recovering) {
1983                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1984                               "recovery until %d %s reconnect, or if no clients"
1985                               " reconnect for %d:%.02d; during that time new "
1986                               "clients will not be allowed to connect. "
1987                               "Recovery progress can be monitored by watching "
1988                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1989                               obd->obd_name, lustre_cfg_string(lcfg, 1),
1990                               label ?: "", label ? "/" : "", str,
1991                               obd->obd_max_recoverable_clients,
1992                               (obd->obd_max_recoverable_clients == 1)
1993                               ? "client" : "clients",
1994                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
1995                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
1996                               obd->obd_name);
1997         } else {
1998                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
1999                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2000                               label ?: "", label ? "/" : "", str,
2001                               obd->obd_replayable ? "enabled" : "disabled");
2002         }
2003
2004         RETURN(0);
2005
2006 err_post:
2007         filter_post(obd);
2008 err_ops:
2009         fsfilt_put_ops(obd->obd_fsops);
2010         filter_iobuf_pool_done(filter);
2011 err_mntput:
2012         server_put_mount(obd->obd_name, mnt);
2013         obd->u.obt.obt_sb = 0;
2014         return rc;
2015 }
2016
2017 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2018 {
2019         struct lprocfs_static_vars lvars;
2020         unsigned long addr;
2021         struct page *page;
2022         int rc;
2023
2024         CLASSERT(offsetof(struct obd_device, u.obt) ==
2025                  offsetof(struct obd_device, u.filter.fo_obt));
2026
2027         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2028                 RETURN(-EINVAL);
2029
2030         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2031         OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
2032         if (!page)
2033                 RETURN(-ENOMEM);
2034         addr = (unsigned long)cfs_page_address(page);
2035         clear_page((void *)addr);
2036
2037         /* lprocfs must be setup before the filter so state can be safely added
2038          * to /proc incrementally as the filter is setup */
2039         lprocfs_filter_init_vars(&lvars);
2040         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2041             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2042                 /* Init obdfilter private stats here */
2043                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2044                                      LPROCFS_CNTR_AVGMINMAX,
2045                                      "read_bytes", "bytes");
2046                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2047                                      LPROCFS_CNTR_AVGMINMAX,
2048                                      "write_bytes", "bytes");
2049
2050                 lproc_filter_attach_seqstat(obd);
2051                 obd->obd_proc_exports = proc_mkdir("exports",
2052                                                    obd->obd_proc_entry);
2053         }
2054
2055         memcpy((void *)addr, lustre_cfg_buf(lcfg, 4),
2056                LUSTRE_CFG_BUFLEN(lcfg, 4));
2057         rc = filter_common_setup(obd, lcfg, (void *)addr);
2058         OBD_PAGE_FREE(page);
2059
2060         if (rc) {
2061                 lprocfs_obd_cleanup(obd);
2062                 lprocfs_free_obd_stats(obd);
2063         }
2064
2065         return rc;
2066 }
2067
2068 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2069 static struct llog_operations filter_size_orig_logops = {
2070         lop_setup: llog_obd_origin_setup,
2071         lop_cleanup: llog_obd_origin_cleanup,
2072         lop_add: llog_obd_origin_add
2073 };
2074
2075 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2076                             struct obd_device *tgt, int count,
2077                             struct llog_catid *catid,
2078                             struct obd_uuid *uuid)
2079 {
2080         struct llog_ctxt *ctxt;
2081         int rc;
2082         ENTRY;
2083
2084         filter_mds_ost_repl_logops = llog_client_ops;
2085         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2086         filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2087         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2088
2089         rc = llog_setup(obd, llogs, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2090                         &filter_mds_ost_repl_logops);
2091         if (rc)
2092                 RETURN(rc);
2093
2094         /* FIXME - assign unlink_cb for filter's recovery */
2095         if (!llogs)
2096                 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2097         else
2098                 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2099
2100         LASSERT(ctxt != NULL);
2101         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2102
2103         rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2104                         &filter_size_orig_logops);
2105         RETURN(rc);
2106 }
2107
2108 static int filter_group_llog_cleanup(struct llog_ctxt *ctxt)
2109 {
2110         int rc = 0;
2111         ENTRY;
2112
2113         if (CTXTP(ctxt, cleanup))
2114                 rc = CTXTP(ctxt, cleanup)(ctxt);
2115
2116         if (ctxt->loc_exp)
2117                 class_export_put(ctxt->loc_exp);
2118         OBD_FREE(ctxt, sizeof(*ctxt));
2119
2120         RETURN(rc);
2121 }
2122
2123 static int filter_group_llog_finish(struct obd_llogs *llogs)
2124 {
2125         struct llog_ctxt *ctxt;
2126         int rc = 0, rc2 = 0;
2127         ENTRY;
2128
2129         ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2130         if (ctxt)
2131                 rc = filter_group_llog_cleanup(ctxt);
2132
2133         ctxt = llog_get_context_from_llogs(llogs, LLOG_SIZE_ORIG_CTXT);
2134         if (ctxt)
2135                 rc2 = filter_group_llog_cleanup(ctxt);
2136         if (!rc)
2137                 rc = rc2;
2138
2139         RETURN(rc);
2140 }
2141
2142 static int filter_llog_finish(struct obd_device *obd, int count)
2143 {
2144         struct llog_ctxt *ctxt;
2145         int rc = 0, rc2 = 0;
2146         ENTRY;
2147
2148         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2149         if (ctxt)
2150                 rc = llog_cleanup(ctxt);
2151
2152         ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
2153         if (ctxt)
2154                 rc2 = llog_cleanup(ctxt);
2155         if (!rc)
2156                 rc = rc2;
2157
2158         RETURN(rc);
2159 }
2160
2161 struct obd_llogs *filter_grab_llog_for_group(struct obd_device *obd, int group,
2162                                              struct obd_export *export)
2163 {
2164         struct filter_group_llog *fglog, *nlog;
2165         struct filter_obd *filter;
2166         struct llog_ctxt *ctxt;
2167         struct list_head *cur;
2168         int rc;
2169
2170         filter = &obd->u.filter;
2171
2172         spin_lock(&filter->fo_llog_list_lock);
2173         list_for_each(cur, &filter->fo_llog_list) {
2174                 fglog = list_entry(cur, struct filter_group_llog, list);
2175                 if (fglog->group == group) {
2176                         if (!(fglog->exp == NULL || fglog->exp == export || export == NULL))
2177                                 CWARN("%s: export for group %d changes: 0x%p -> 0x%p\n",
2178                                       obd->obd_name, group, fglog->exp, export);
2179                         spin_unlock(&filter->fo_llog_list_lock);
2180                         goto init;
2181                 }
2182         }
2183         spin_unlock(&filter->fo_llog_list_lock);
2184
2185         if (export == NULL)
2186                 RETURN(NULL);
2187
2188         OBD_ALLOC_PTR(fglog);
2189         if (fglog == NULL)
2190                 RETURN(NULL);
2191         fglog->group = group;
2192
2193         OBD_ALLOC_PTR(fglog->llogs);
2194         if (fglog->llogs == NULL) {
2195                 OBD_FREE_PTR(fglog);
2196                 RETURN(NULL);
2197         }
2198
2199         spin_lock(&filter->fo_llog_list_lock);
2200         list_for_each(cur, &filter->fo_llog_list) {
2201                 nlog = list_entry(cur, struct filter_group_llog, list);
2202                 LASSERT(nlog->group != group);
2203         }
2204         list_add(&fglog->list, &filter->fo_llog_list);
2205         spin_unlock(&filter->fo_llog_list_lock);
2206
2207         rc = llog_cat_initialize(obd, fglog->llogs, 1, NULL);
2208         if (rc) {
2209                 OBD_FREE_PTR(fglog->llogs);
2210                 OBD_FREE_PTR(fglog);
2211                 RETURN(NULL);
2212         }
2213
2214 init:
2215         if (export) {
2216                 fglog->exp = export;
2217                 ctxt = llog_get_context_from_llogs(fglog->llogs,
2218                                                LLOG_MDS_OST_REPL_CTXT);
2219                 LASSERT(ctxt != NULL);
2220
2221                 llog_receptor_accept(ctxt, export->exp_imp_reverse);
2222         }
2223         CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n",
2224                obd->obd_name, fglog->llogs, group);
2225
2226         RETURN(fglog->llogs);
2227 }
2228
2229 static int filter_llog_connect(struct obd_export *exp,
2230                                struct llogd_conn_body *body)
2231 {
2232         struct obd_device *obd = exp->exp_obd;
2233         struct llog_ctxt *ctxt;
2234         struct obd_llogs *llog;
2235         int rc;
2236         ENTRY;
2237
2238         CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2239                 (unsigned) body->lgdc_logid.lgl_ogr,
2240                 (unsigned) body->lgdc_logid.lgl_oid,
2241                 (unsigned) body->lgdc_logid.lgl_ogen);
2242
2243         llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr, exp);
2244         LASSERT(llog != NULL);
2245         ctxt = llog_get_context_from_llogs(llog, body->lgdc_ctxt_idx);
2246         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2247                  body->lgdc_ctxt_idx);
2248         rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2249                           &body->lgdc_gen, NULL);
2250         if (rc != 0)
2251                 CERROR("failed to connect rc %d idx %d\n", rc,
2252                                 body->lgdc_ctxt_idx);
2253
2254         RETURN(rc);
2255 }
2256
2257 static int filter_llog_preclean (struct obd_device *obd)
2258 {
2259         struct filter_group_llog *log;
2260         struct filter_obd *filter;
2261         int rc = 0;
2262         ENTRY;
2263
2264         filter = &obd->u.filter;
2265         spin_lock(&filter->fo_llog_list_lock);
2266         while (!list_empty(&filter->fo_llog_list)) {
2267                 log = list_entry(filter->fo_llog_list.next,
2268                                  struct filter_group_llog, list);
2269                 list_del(&log->list);
2270                 spin_unlock(&filter->fo_llog_list_lock);
2271
2272                 rc = filter_group_llog_finish(log->llogs);
2273                 if (rc)
2274                         CERROR("failed to cleanup llogging subsystem for %u\n",
2275                                 log->group);
2276                 OBD_FREE_PTR(log->llogs);
2277                 OBD_FREE_PTR(log);
2278                 spin_lock(&filter->fo_llog_list_lock);
2279         }
2280         spin_unlock(&filter->fo_llog_list_lock);
2281
2282         rc = obd_llog_finish(obd, 0);
2283         if (rc)
2284                 CERROR("failed to cleanup llogging subsystem\n");
2285
2286         RETURN(rc);
2287 }
2288
2289 static int filter_precleanup(struct obd_device *obd,
2290                              enum obd_cleanup_stage stage)
2291 {
2292         int rc = 0;
2293         ENTRY;
2294
2295         switch(stage) {
2296         case OBD_CLEANUP_EARLY:
2297                 break;
2298         case OBD_CLEANUP_EXPORTS:
2299                 target_cleanup_recovery(obd);
2300                 break;
2301         case OBD_CLEANUP_SELF_EXP:
2302                 rc = filter_llog_preclean(obd);
2303                 break;
2304         case OBD_CLEANUP_OBD:
2305                 break;
2306         }
2307         RETURN(rc);
2308 }
2309
2310 static int filter_cleanup(struct obd_device *obd)
2311 {
2312         struct filter_obd *filter = &obd->u.filter;
2313         ENTRY;
2314
2315         if (obd->obd_fail)
2316                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2317                               "will be preserved.\n", obd->obd_name);
2318
2319         if (!list_empty(&obd->obd_exports)) {
2320                 CERROR("%s: still has clients!\n", obd->obd_name);
2321                 class_disconnect_exports(obd);
2322                 if (!list_empty(&obd->obd_exports)) {
2323                         CERROR("still has exports after forced cleanup?\n");
2324                         RETURN(-EBUSY);
2325                 }
2326         }
2327
2328         lprocfs_obd_cleanup(obd);
2329         lprocfs_free_obd_stats(obd);
2330         lquota_cleanup(filter_quota_interface_ref, obd);
2331
2332         /* Stop recovery before namespace cleanup. */
2333         target_stop_recovery_thread(obd);
2334         target_cleanup_recovery(obd);
2335
2336         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2337
2338         if (obd->u.obt.obt_sb == NULL)
2339                 RETURN(0);
2340
2341         filter_post(obd);
2342
2343         shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2344
2345         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2346
2347         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2348         obd->u.obt.obt_sb = NULL;
2349
2350         fsfilt_put_ops(obd->obd_fsops);
2351
2352         filter_iobuf_pool_done(filter);
2353
2354         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2355
2356         RETURN(0);
2357 }
2358
2359 static int filter_connect_internal(struct obd_export *exp,
2360                                    struct obd_connect_data *data)
2361 {
2362         if (!data)
2363                 RETURN(0);
2364
2365         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2366                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2367                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2368                data->ocd_connect_flags, data->ocd_version,
2369                data->ocd_grant, data->ocd_index);
2370
2371         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2372         exp->exp_connect_flags = data->ocd_connect_flags;
2373         data->ocd_version = LUSTRE_VERSION_CODE;
2374
2375         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2376                 struct filter_export_data *fed = &exp->exp_filter_data;
2377                 obd_size left, want;
2378
2379                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2380                 left = filter_grant_space_left(exp);
2381                 want = data->ocd_grant;
2382                 filter_grant(exp, fed->fed_grant, want, left);
2383                 data->ocd_grant = fed->fed_grant;
2384                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2385
2386                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2387                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2388                        exp->exp_client_uuid.uuid, exp,
2389                        data->ocd_grant, want, left);
2390         }
2391
2392         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2393                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2394                 struct lr_server_data *lsd = filter->fo_fsd;
2395                 int index = le32_to_cpu(lsd->lsd_ost_index);
2396
2397                 if (!(lsd->lsd_feature_compat &
2398                       cpu_to_le32(OBD_COMPAT_OST))) {
2399                         /* this will only happen on the first connect */
2400                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2401                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2402                         filter_update_server_data(exp->exp_obd,
2403                                                   filter->fo_rcvd_filp, lsd, 1);
2404                 } else if (index != data->ocd_index) {
2405                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2406                                            " %u doesn't match actual OST index"
2407                                            " %u in last_rcvd file, bad "
2408                                            "configuration?\n",
2409                                            obd_export_nid2str(exp), index,
2410                                            data->ocd_index);
2411                         RETURN(-EBADF);
2412                 }
2413         }
2414
2415         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2416                 data->ocd_brw_size = 65536;
2417         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2418                 data->ocd_brw_size = min(data->ocd_brw_size,
2419                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2420                 LASSERT(data->ocd_brw_size);
2421         }
2422
2423         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2424          * FIXME: We don't strictly need the COMPAT flag for that,
2425          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2426          * FIXME: We needed it for the index, as index 0 is valid. */
2427
2428         RETURN(0);
2429 }
2430
2431 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2432                             struct obd_uuid *cluuid,
2433                             struct obd_connect_data *data)
2434 {
2435         int rc;
2436         ENTRY;
2437
2438         if (exp == NULL || obd == NULL || cluuid == NULL)
2439                 RETURN(-EINVAL);
2440
2441         rc = filter_connect_internal(exp, data);
2442
2443         RETURN(rc);
2444 }
2445
2446 /* nearly identical to mds_connect */
2447 static int filter_connect(const struct lu_env *env,
2448                           struct lustre_handle *conn, struct obd_device *obd,
2449                           struct obd_uuid *cluuid,
2450                           struct obd_connect_data *data)
2451 {
2452         struct lvfs_run_ctxt saved;
2453         struct obd_export *exp;
2454         struct filter_export_data *fed;
2455         struct filter_client_data *fcd = NULL;
2456         __u32 group;
2457         int rc;
2458         ENTRY;
2459
2460         if (conn == NULL || obd == NULL || cluuid == NULL)
2461                 RETURN(-EINVAL);
2462
2463         rc = class_connect(conn, obd, cluuid);
2464         if (rc)
2465                 RETURN(rc);
2466         exp = class_conn2export(conn);
2467         LASSERT(exp != NULL);
2468
2469         fed = &exp->exp_filter_data;
2470
2471         rc = filter_connect_internal(exp, data);
2472         if (rc)
2473                 GOTO(cleanup, rc);
2474
2475         filter_export_stats_init(obd, exp);
2476         group = data->ocd_group;
2477         if (obd->obd_replayable) {
2478                 OBD_ALLOC(fcd, sizeof(*fcd));
2479                 if (!fcd) {
2480                         CERROR("filter: out of memory for client data\n");
2481                         GOTO(cleanup, rc = -ENOMEM);
2482                 }
2483
2484                 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2485                 fed->fed_fcd = fcd;
2486                 fed->fed_fcd->fcd_group = group;
2487                 rc = filter_client_add(obd, exp, -1);
2488                 if (rc)
2489                         GOTO(cleanup, rc);
2490         }
2491         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2492                obd->obd_name, exp->exp_handle.h_cookie, group);
2493         if (group == 0)
2494                 GOTO(cleanup, rc);
2495
2496         if (fed->fed_group != 0 && fed->fed_group != group) {
2497                 CERROR("!!! This export (nid %s) used object group %d "
2498                        "earlier; now it's trying to use group %d!  This could "
2499                        "be a bug in the MDS.  Tell CFS.\n",
2500                        obd_export_nid2str(exp), fed->fed_group, group);
2501                 GOTO(cleanup, rc = -EPROTO);
2502         }
2503         fed->fed_group = group;
2504
2505         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2506         rc = filter_read_groups(obd, group, 1);
2507         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2508         if (rc != 0) {
2509                 CERROR("can't read group %u\n", group);
2510                 GOTO(cleanup, rc);
2511         }
2512
2513         GOTO(cleanup, rc);
2514
2515 cleanup:
2516         if (rc) {
2517                 if (fcd) {
2518                         OBD_FREE(fcd, sizeof(*fcd));
2519                         fed->fed_fcd = NULL;
2520                 }
2521                 class_disconnect(exp);
2522         } else {
2523                 class_export_put(exp);
2524         }
2525
2526         RETURN(rc);
2527 }
2528
2529 /* Do extra sanity checks for grant accounting.  We do this at connect,
2530  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2531  * always get rid of it or turn it off when we know accounting is good. */
2532 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2533 {
2534         struct filter_export_data *fed;
2535         struct obd_export *exp;
2536         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2537         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2538         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2539
2540         if (list_empty(&obd->obd_exports))
2541                 return;
2542
2543         /* We don't want to do this for large machines that do lots of
2544            mounts or unmounts.  It burns... */
2545         if (obd->obd_num_exports > 100)
2546                 return;
2547
2548         spin_lock(&obd->obd_osfs_lock);
2549         spin_lock(&obd->obd_dev_lock);
2550         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2551                 int error = 0;
2552                 fed = &exp->exp_filter_data;
2553                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2554                     fed->fed_dirty < 0)
2555                         error = 1;
2556                 if (maxsize > 0) { /* we may not have done a statfs yet */
2557                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2558                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2559                                  exp->exp_client_uuid.uuid, exp,
2560                                  fed->fed_grant, fed->fed_pending, maxsize);
2561                         LASSERTF(fed->fed_dirty <= maxsize,
2562                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2563                                  exp->exp_client_uuid.uuid, exp,
2564                                  fed->fed_dirty, maxsize);
2565                 }
2566                 if (error)
2567                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2568                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2569                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2570                 else
2571                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2572                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2573                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2574                 tot_granted += fed->fed_grant + fed->fed_pending;
2575                 tot_pending += fed->fed_pending;
2576                 tot_dirty += fed->fed_dirty;
2577         }
2578         fo_tot_granted = obd->u.filter.fo_tot_granted;
2579         fo_tot_pending = obd->u.filter.fo_tot_pending;
2580         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2581         spin_unlock(&obd->obd_dev_lock);
2582         spin_unlock(&obd->obd_osfs_lock);
2583
2584         /* Do these assertions outside the spinlocks so we don't kill system */
2585         if (tot_granted != fo_tot_granted)
2586                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2587                        func, tot_granted, fo_tot_granted);
2588         if (tot_pending != fo_tot_pending)
2589                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2590                        func, tot_pending, fo_tot_pending);
2591         if (tot_dirty != fo_tot_dirty)
2592                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2593                        func, tot_dirty, fo_tot_dirty);
2594         if (tot_pending > tot_granted)
2595                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2596                        func, tot_pending, tot_granted);
2597         if (tot_granted > maxsize)
2598                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2599                        func, tot_granted, maxsize);
2600         if (tot_dirty > maxsize)
2601                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2602                        func, tot_dirty, maxsize);
2603 }
2604
2605 /* Remove this client from the grant accounting totals.  We also remove
2606  * the export from the obd device under the osfs and dev locks to ensure
2607  * that the filter_grant_sanity_check() calculations are always valid.
2608  * The client should do something similar when it invalidates its import. */
2609 static void filter_grant_discard(struct obd_export *exp)
2610 {
2611         struct obd_device *obd = exp->exp_obd;
2612         struct filter_obd *filter = &obd->u.filter;
2613         struct filter_export_data *fed = &exp->exp_filter_data;
2614
2615         spin_lock(&obd->obd_osfs_lock);
2616         spin_lock(&obd->obd_dev_lock);
2617         list_del_init(&exp->exp_obd_chain);
2618         spin_unlock(&obd->obd_dev_lock);
2619
2620         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2621                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2622                  obd->obd_name, filter->fo_tot_granted,
2623                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2624         filter->fo_tot_granted -= fed->fed_grant;
2625         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2626                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2627                  obd->obd_name, filter->fo_tot_pending,
2628                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2629         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2630         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2631                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2632                  obd->obd_name, filter->fo_tot_dirty,
2633                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2634         filter->fo_tot_dirty -= fed->fed_dirty;
2635         fed->fed_dirty = 0;
2636         fed->fed_grant = 0;
2637
2638         spin_unlock(&obd->obd_osfs_lock);
2639 }
2640
2641 static int filter_destroy_export(struct obd_export *exp)
2642 {
2643         ENTRY;
2644
2645         if (exp->exp_filter_data.fed_pending)
2646                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2647                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2648                        exp, exp->exp_filter_data.fed_pending);
2649
2650         target_destroy_export(exp);
2651
2652         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2653                 RETURN(0);
2654
2655         lprocfs_exp_cleanup(exp);
2656
2657         if (exp->exp_obd->obd_replayable)
2658                 filter_client_free(exp);
2659         else
2660                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2661
2662         filter_grant_discard(exp);
2663         filter_fmd_cleanup(exp);
2664
2665         if (!(exp->exp_flags & OBD_OPT_FORCE))
2666                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2667
2668         RETURN(0);
2669 }
2670
2671 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2672 {
2673         struct filter_group_llog *fglog, *nlog;
2674         struct filter_obd *filter;
2675         int worked = 0, group;
2676         struct llog_ctxt *ctxt;
2677         ENTRY;
2678
2679         filter = &obd->u.filter;
2680
2681         /* we can't sync log holding spinlock. also, we do not want to get
2682          * into livelock. so we do following: loop over MDS's exports in
2683          * group order and skip already synced llogs -bzzz */
2684         do {
2685                 /* look for group with min. number, but > worked */
2686                 fglog = NULL;
2687                 group = 1 << 30;
2688                 spin_lock(&filter->fo_llog_list_lock);
2689                 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
2690                         if (nlog->group <= worked) {
2691                                 /* this group is already synced */
2692                                 continue;
2693                         }
2694                         if (group < nlog->group) {
2695                                 /* we have group with smaller number to sync */
2696                                 continue;
2697                         }
2698                         /* store current minimal group */
2699                         fglog = nlog;
2700                         group = nlog->group;
2701                 }
2702                 spin_unlock(&filter->fo_llog_list_lock);
2703
2704                 if (fglog == NULL)
2705                         break;
2706
2707                 worked = fglog->group;
2708                 if (fglog->exp && (dexp == fglog->exp || dexp == NULL)) {
2709                         ctxt = llog_get_context_from_llogs(fglog->llogs,
2710                                                 LLOG_MDS_OST_REPL_CTXT);
2711                         LASSERT(ctxt != NULL);
2712                         llog_sync(ctxt, fglog->exp);
2713                 }
2714         } while (fglog != NULL);
2715 }
2716
2717 /* also incredibly similar to mds_disconnect */
2718 static int filter_disconnect(struct obd_export *exp)
2719 {
2720         struct obd_device *obd = exp->exp_obd;
2721         int rc;
2722         ENTRY;
2723
2724         LASSERT(exp);
2725         class_export_get(exp);
2726
2727         if (!(exp->exp_flags & OBD_OPT_FORCE))
2728                 filter_grant_sanity_check(obd, __FUNCTION__);
2729         filter_grant_discard(exp);
2730
2731         /* Disconnect early so that clients can't keep using export */
2732         rc = class_disconnect(exp);
2733         if (exp->exp_obd->obd_namespace != NULL)
2734                 ldlm_cancel_locks_for_export(exp);
2735
2736         fsfilt_sync(obd, obd->u.obt.obt_sb);
2737
2738         /* flush any remaining cancel messages out to the target */
2739         filter_sync_llogs(obd, exp);
2740         class_export_put(exp);
2741         RETURN(rc);
2742 }
2743
2744 /* reverse import is changed, sync all cancels */
2745 static void filter_revimp_update(struct obd_export *exp)
2746 {
2747         ENTRY;
2748
2749         LASSERT(exp);
2750         class_export_get(exp);
2751
2752         /* flush any remaining cancel messages out to the target */
2753         filter_sync_llogs(exp->exp_obd, exp);
2754         class_export_put(exp);
2755         EXIT;
2756 }
2757
2758 static int filter_ping(struct obd_export *exp)
2759 {
2760         filter_fmd_expire(exp);
2761
2762         return 0;
2763 }
2764
2765 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2766                                   const char *what, int quiet)
2767 {
2768         struct dentry *dchild = NULL;
2769         obd_gr group = 0;
2770
2771         if (oa->o_valid & OBD_MD_FLGROUP)
2772                 group = oa->o_gr;
2773
2774         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2775
2776         if (IS_ERR(dchild)) {
2777                 CERROR("%s error looking up object: "LPU64":"LPU64"\n",
2778                        what, group, oa->o_id);
2779                 RETURN(dchild);
2780         }
2781
2782         if (dchild->d_inode == NULL) {
2783                 if (!quiet)
2784                         CERROR("%s: %s on non-existent object: "LPU64"\n",
2785                                obd->obd_name, what, oa->o_id);
2786                 f_dput(dchild);
2787                 RETURN(ERR_PTR(-ENOENT));
2788         }
2789
2790         return dchild;
2791 }
2792
2793 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2794 {
2795         struct dentry *dentry = NULL;
2796         struct obd_device *obd;
2797         int rc = 0;
2798         ENTRY;
2799
2800         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2801                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
2802         if (rc)
2803                 RETURN(rc);
2804
2805         obd = class_exp2obd(exp);
2806         if (obd == NULL) {
2807                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2808                 RETURN(-EINVAL);
2809         }
2810
2811         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2812         if (IS_ERR(dentry))
2813                 RETURN(PTR_ERR(dentry));
2814
2815         /* Limit the valid bits in the return data to what we actually use */
2816         oinfo->oi_oa->o_valid = OBD_MD_FLID;
2817         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2818
2819         f_dput(dentry);
2820         RETURN(rc);
2821 }
2822
2823 /* this should be enabled/disabled in condition to enabled/disabled large
2824  * inodes (fast EAs) in backing store FS. */
2825 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2826                         void *handle, struct obdo *oa)
2827 {
2828         struct obd_device *obd = exp->exp_obd;
2829         int rc = 0;
2830         ENTRY;
2831
2832         if (oa->o_valid & OBD_MD_FLFID) {
2833                 struct filter_fid ff;
2834
2835                 if (!(oa->o_valid & OBD_MD_FLGROUP))
2836                         oa->o_gr = 0;
2837                 /* packing fid and converting it to LE for storing into EA.
2838                  * Here ->o_stripe_idx should be filled by LOV and rest of
2839                  * fields - by client. */
2840                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2841                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2842                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2843                 ff.ff_objid = cpu_to_le64(oa->o_id);
2844                 ff.ff_group = cpu_to_le64(oa->o_gr);
2845
2846                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2847                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2848                        oa->o_generation, oa->o_id, oa->o_gr);
2849
2850                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2851                 if (rc)
2852                         CERROR("store fid in object failed! rc: %d\n", rc);
2853         } else {
2854                 CDEBUG(D_HA, "OSS object without fid info!\n");
2855         }
2856
2857         RETURN(rc);
2858 }
2859
2860 /* this is called from filter_truncate() until we have filter_punch() */
2861 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2862                             struct obdo *oa, struct obd_trans_info *oti)
2863 {
2864         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2865         struct llog_cookie *fcc = NULL;
2866         struct filter_obd *filter;
2867         int rc, err, locked = 0, sync = 0;
2868         unsigned int ia_valid;
2869         struct inode *inode;
2870         struct iattr iattr;
2871         void *handle;
2872         ENTRY;
2873
2874         LASSERT(dentry != NULL);
2875         LASSERT(!IS_ERR(dentry));
2876
2877         inode = dentry->d_inode;
2878         LASSERT(inode != NULL);
2879
2880         filter = &exp->exp_obd->u.filter;
2881         iattr_from_obdo(&iattr, oa, oa->o_valid);
2882         ia_valid = iattr.ia_valid;
2883
2884         if (oa->o_valid & OBD_MD_FLCOOKIE) {
2885                 OBD_ALLOC(fcc, sizeof(*fcc));
2886                 if (fcc != NULL)
2887                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2888         }
2889
2890         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2891                 DQUOT_INIT(inode);
2892                 LOCK_INODE_MUTEX(inode);
2893                 locked = 1;
2894         }
2895
2896         /* If the inode still has SUID+SGID bits set (see filter_precreate())
2897          * then we will accept the UID+GID sent by the client during write for
2898          * initializing the ownership of this inode.  We only allow this to
2899          * happen once so clear these bits in setattr. In 2.6 kernels it is
2900          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2901          * the flags that are actually being set. */
2902         if (ia_valid & (ATTR_UID | ATTR_GID)) {
2903                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2904                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2905
2906                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2907                         if (!(ia_valid & ATTR_MODE)) {
2908                                 iattr.ia_mode = inode->i_mode;
2909                                 iattr.ia_valid |= ATTR_MODE;
2910                         }
2911                         iattr.ia_mode &= ~S_ISUID;
2912                 }
2913                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2914                         if (!(iattr.ia_valid & ATTR_MODE)) {
2915                                 iattr.ia_mode = inode->i_mode;
2916                                 iattr.ia_valid |= ATTR_MODE;
2917                         }
2918                         iattr.ia_mode &= ~S_ISGID;
2919                 }
2920
2921                 orig_ids[USRQUOTA] = inode->i_uid;
2922                 orig_ids[GRPQUOTA] = inode->i_gid;
2923                 handle = fsfilt_start_log(exp->exp_obd, inode,
2924                                           FSFILT_OP_SETATTR, oti, 1);
2925                 if (IS_ERR(handle))
2926                         GOTO(out_unlock, rc = PTR_ERR(handle));
2927
2928                 /* update inode EA only once when inode is suid bit marked. As
2929                  * on 2.6.x UID and GID may be set separately, we check here
2930                  * only one of them to avoid double setting. */
2931                 if (inode->i_mode & S_ISUID)
2932                         filter_update_fidea(exp, inode, handle, oa);
2933         } else {
2934                 handle = fsfilt_start(exp->exp_obd, inode,
2935                                       FSFILT_OP_SETATTR, oti);
2936                 if (IS_ERR(handle))
2937                         GOTO(out_unlock, rc = PTR_ERR(handle));
2938         }
2939         if (oa->o_valid & OBD_MD_FLFLAGS) {
2940                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2941                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2942         } else {
2943                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2944                 if (fcc != NULL)
2945                         /* set cancel cookie callback function */
2946                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2947                                                      filter_cancel_cookies_cb,
2948                                                      fcc);
2949         }
2950
2951         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2952                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2953
2954         /* The truncate might have used up our transaction credits.  Make
2955          * sure we have one left for the last_rcvd update. */
2956         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2957
2958         rc = filter_finish_transno(exp, oti, rc, sync);
2959         if (sync) {
2960                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
2961                 fcc = NULL;
2962         }
2963
2964         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2965         if (err) {
2966                 CERROR("error on commit, err = %d\n", err);
2967                 if (!rc)
2968                         rc = err;
2969         } else {
2970                 fcc = NULL;
2971         }
2972
2973         if (locked) {
2974                 /* Let's flush truncated page on disk immediately, then we can
2975                  * avoid need to search for page aliases before directio writes
2976                  * and this sort of stuff at expense of somewhat slower
2977                  * truncates not on a page boundary. I believe this is the only
2978                  * place in filter code that can lead to pages getting to
2979                  * pagecache so far. */
2980                 filter_clear_truncated_page(inode);
2981                 UNLOCK_INODE_MUTEX(inode);
2982                 locked = 0;
2983         }
2984
2985         EXIT;
2986 out_unlock:
2987         if (locked)
2988                 UNLOCK_INODE_MUTEX(inode);
2989
2990         if (fcc)
2991                 OBD_FREE(fcc, sizeof(*fcc));
2992
2993         /* trigger quota release */
2994         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2995                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2996                 int rc2 = lquota_adjust(filter_quota_interface_ref,
2997                                         exp->exp_obd, cur_ids,
2998                                         orig_ids, rc, FSFILT_OP_SETATTR);
2999                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
3000                        "filter adjust qunit. (rc:%d)\n", rc2);
3001         }
3002         return rc;
3003 }
3004
3005 /* this is called from filter_truncate() until we have filter_punch() */
3006 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3007                    struct obd_trans_info *oti)
3008 {
3009         struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3010                                                 oinfo->oi_oa->o_gr, 0 } };
3011         struct filter_mod_data *fmd;
3012         struct lvfs_run_ctxt saved;
3013         struct filter_obd *filter;
3014         struct ldlm_resource *res;
3015         struct dentry *dentry;
3016         int rc;
3017         ENTRY;
3018
3019         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3020                               oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3021         if (rc)
3022                 RETURN(rc);
3023
3024         dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3025                                     __FUNCTION__, 1);
3026         if (IS_ERR(dentry))
3027                 RETURN(PTR_ERR(dentry));
3028
3029         filter = &exp->exp_obd->u.filter;
3030         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3031         lock_kernel();
3032
3033         if (oinfo->oi_oa->o_valid &
3034             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3035                 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3036                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3037                         fmd->fmd_mactime_xid = oti->oti_xid;
3038                 filter_fmd_put(exp, fmd);
3039         }
3040
3041         /* setting objects attributes (including owner/group) */
3042         rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3043         if (rc)
3044                 GOTO(out_unlock, rc);
3045
3046         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3047                                 &res_id, LDLM_EXTENT, 0);
3048
3049         if (res != NULL) {
3050                 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3051                 ldlm_resource_putref(res);
3052         }
3053
3054         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3055
3056         /* Quota release need uid/gid info */
3057         obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3058                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3059
3060         EXIT;
3061 out_unlock:
3062         unlock_kernel();
3063         f_dput(dentry);
3064         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3065         return rc;
3066 }
3067
3068 /* XXX identical to osc_unpackmd */
3069 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3070                            struct lov_mds_md *lmm, int lmm_bytes)
3071 {
3072         int lsm_size;
3073         ENTRY;
3074
3075         if (lmm != NULL) {
3076                 if (lmm_bytes < sizeof (*lmm)) {
3077                         CERROR("lov_mds_md too small: %d, need %d\n",
3078                                lmm_bytes, (int)sizeof(*lmm));
3079                         RETURN(-EINVAL);
3080                 }
3081                 /* XXX LOV_MAGIC etc check? */
3082
3083                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3084                         CERROR("lov_mds_md: zero lmm_object_id\n");
3085                         RETURN(-EINVAL);
3086                 }
3087         }
3088
3089         lsm_size = lov_stripe_md_size(1);
3090         if (lsmp == NULL)
3091                 RETURN(lsm_size);
3092
3093         if (*lsmp != NULL && lmm == NULL) {
3094                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3095                 OBD_FREE(*lsmp, lsm_size);
3096                 *lsmp = NULL;
3097                 RETURN(0);
3098         }
3099
3100         if (*lsmp == NULL) {
3101                 OBD_ALLOC(*lsmp, lsm_size);
3102                 if (*lsmp == NULL)
3103                         RETURN(-ENOMEM);
3104
3105                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3106                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3107                         OBD_FREE(*lsmp, lsm_size);
3108                         RETURN(-ENOMEM);
3109                 }
3110                 loi_init((*lsmp)->lsm_oinfo[0]);
3111         }
3112
3113         if (lmm != NULL) {
3114                 /* XXX zero *lsmp? */
3115                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3116                 LASSERT((*lsmp)->lsm_object_id);
3117         }
3118
3119         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3120
3121         RETURN(lsm_size);
3122 }
3123
3124 /* caller must hold fo_create_locks[oa->o_gr] */
3125 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3126                                       struct filter_obd *filter)
3127 {
3128         struct obdo doa; /* XXX obdo on stack */
3129         obd_id last, id;
3130         int rc;
3131         ENTRY;
3132
3133         LASSERT(oa);
3134         LASSERT(oa->o_gr != 0);
3135         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3136         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3137
3138         memset(&doa, 0, sizeof(doa));
3139
3140         doa.o_valid |= OBD_MD_FLGROUP;
3141         doa.o_gr = oa->o_gr;
3142         doa.o_mode = S_IFREG;
3143
3144         if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3145                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3146                        exp->exp_obd->obd_name, doa.o_gr);
3147                 RETURN(0);
3148         }
3149
3150         last = filter_last_id(filter, doa.o_gr);
3151
3152         CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3153                exp->exp_obd->obd_name, oa->o_id + 1, last);
3154
3155         for (id = last; id > oa->o_id; id--) {
3156                 doa.o_id = id;
3157                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3158                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3159                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3160                                id, rc);
3161                 filter_set_last_id(filter, id - 1, doa.o_gr);
3162                 /* update last_id on disk periodically so that if we restart
3163                  * we don't need to re-scan all of the just-deleted objects. */
3164                 if ((id & 511) == 0)
3165                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3166         }
3167
3168         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3169                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3170
3171         rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3172         clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3173
3174         RETURN(rc);
3175 }
3176
3177 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3178                             obd_gr group, int *num);
3179 /* returns a negative error or a nonnegative number of files to create */
3180 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3181                                    obd_gr group, struct obd_trans_info *oti)
3182 {
3183         struct obd_device *obd = exp->exp_obd;
3184         struct filter_obd *filter = &obd->u.filter;
3185         int diff, rc;
3186         ENTRY;
3187
3188         /* delete orphans request */
3189         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3190                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3191                         CERROR("%s: dropping old orphan cleanup request\n",
3192                                obd->obd_name);
3193                         RETURN(0);
3194                 }
3195                 /* This causes inflight precreates to abort and drop lock */
3196                 set_bit(group, &filter->fo_destroys_in_progress);
3197                 down(&filter->fo_create_locks[group]);
3198                 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3199                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3200                                exp->exp_obd->obd_name, group);
3201                         up(&filter->fo_create_locks[group]);
3202                         RETURN(0);
3203                 }
3204                 diff = oa->o_id - filter_last_id(filter, group);
3205                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3206                        filter_last_id(filter, group), diff);
3207
3208                 if (-diff > OST_MAX_PRECREATE) {
3209                         CERROR("%s: ignoring bogus orphan destroy request: "
3210                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3211                                oa->o_id, filter_last_id(filter, group));
3212                         /* FIXME: should reset precreate_next_id on MDS */
3213                         GOTO(out, rc = -EINVAL);
3214                 }
3215                 if (diff < 0) {
3216                         rc = filter_destroy_precreated(exp, oa, filter);
3217                         if (rc)
3218                                 CERROR("%s: unable to write lastobjid, but "
3219                                        "orphans were deleted\n", obd->obd_name);
3220                         GOTO(out, rc);
3221                 } else {
3222                         /* XXX: Used by MDS for the first time! */
3223                         clear_bit(group, &filter->fo_destroys_in_progress);
3224                 }
3225         } else {
3226                 down(&filter->fo_create_locks[group]);
3227                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3228                         CERROR("%s: dropping old precreate request\n",
3229                                obd->obd_name);
3230                         GOTO(out, rc = 0);
3231                 }
3232                 /* only precreate if group == 0 and o_id is specfied */
3233                 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3234                         diff = 1;
3235                 else
3236                         diff = oa->o_id - filter_last_id(filter, group);
3237                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3238                        filter_last_id(filter, group), diff);
3239
3240                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3241                          oa->o_id, filter_last_id(filter, group), diff);
3242         }
3243
3244         if (diff > 0) {
3245                 oa->o_id = filter_last_id(&obd->u.filter, group);
3246                 rc = filter_precreate(obd, oa, group, &diff);
3247                 oa->o_id = filter_last_id(&obd->u.filter, group);
3248                 oa->o_gr = group;
3249                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3250                 GOTO(out, rc);
3251         }
3252         /* else diff == 0 */
3253         GOTO(out, rc = 0);
3254 out:
3255         up(&filter->fo_create_locks[group]);
3256         return rc;
3257 }
3258
3259 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3260                          __u64 max_age)
3261 {
3262         struct filter_obd *filter = &obd->u.filter;
3263         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3264         int rc;
3265         ENTRY;
3266
3267         /* at least try to account for cached pages.  its still racey and
3268          * might be under-reporting if clients haven't announced their
3269          * caches with brw recently */
3270         spin_lock(&obd->obd_osfs_lock);
3271         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3272         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3273         spin_unlock(&obd->obd_osfs_lock);
3274
3275         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3276                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3277                filter->fo_tot_dirty, filter->fo_tot_granted,
3278                filter->fo_tot_pending,
3279                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3280
3281         filter_grant_sanity_check(obd, __FUNCTION__);
3282
3283         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3284                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
3285                                  osfs->os_bsize - 1) >> blockbits));
3286
3287         /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3288          * stop creating files on MDS if OST is not good shape to create
3289          * objects.*/
3290         osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3291                 EROFS : 0;
3292         RETURN(rc);
3293 }
3294
3295 static int filter_use_existing_obj(struct obd_device *obd,
3296                                    struct dentry *dchild, void **handle,
3297                                    int *cleanup_phase)
3298 {
3299         struct inode *inode = dchild->d_inode;
3300         struct iattr iattr;
3301         int rc;
3302
3303         if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3304                 return 0;
3305
3306         *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3307         if (IS_ERR(*handle))
3308                 return PTR_ERR(*handle);
3309
3310         iattr.ia_valid = ATTR_MODE;
3311         iattr.ia_mode = S_ISUID | S_ISGID |0666;
3312         rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3313         if (rc == 0)
3314                 *cleanup_phase = 3;
3315
3316         return rc;
3317 }
3318
3319
3320 /* We rely on the fact that only one thread will be creating files in a given
3321  * group at a time, which is why we don't need an atomic filter_get_new_id.
3322  * Even if we had that atomic function, the following race would exist:
3323  *
3324  * thread 1: gets id x from filter_next_id
3325  * thread 2: gets id (x + 1) from filter_next_id
3326  * thread 2: creates object (x + 1)
3327  * thread 1: tries to create object x, gets -ENOSPC
3328  *
3329  * Caller must hold fo_create_locks[group]
3330  */
3331 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3332                             obd_gr group, int *num)
3333 {
3334         struct dentry *dchild = NULL, *dparent = NULL;
3335         struct filter_obd *filter;
3336         struct obd_statfs *osfs;
3337         int err = 0, rc = 0, recreate_obj = 0, i;
3338         unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3339         obd_id next_id;
3340         void *handle = NULL;
3341         ENTRY;
3342
3343         filter = &obd->u.filter;
3344
3345         LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3346
3347         OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3348
3349         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3350             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3351                 recreate_obj = 1;
3352         } else {
3353                 OBD_ALLOC(osfs, sizeof(*osfs));
3354                 if (osfs == NULL)
3355                         RETURN(-ENOMEM);
3356                 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3357                 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3358                         CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3359                                LPU64"\n", obd->obd_name, osfs->os_bavail <<
3360                                filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3361                         *num = 0;
3362                         rc = -ENOSPC;
3363                 }
3364                 OBD_FREE(osfs, sizeof(*osfs));
3365                 if (rc)
3366                         RETURN(rc);
3367         }
3368
3369         CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3370                " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3371
3372         for (i = 0; i < *num && err == 0; i++) {
3373                 int cleanup_phase = 0;
3374
3375                 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3376                         CWARN("%s: create aborted by destroy\n",
3377                               obd->obd_name);
3378                         rc = -EAGAIN;
3379                         break;
3380                 }
3381
3382                 if (recreate_obj) {
3383                         __u64 last_id;
3384                         next_id = oa->o_id;
3385                         last_id = filter_last_id(filter, group);
3386                         if (next_id > last_id) {
3387                                 CERROR("Error: Trying to recreate obj greater"
3388                                        "than last id "LPD64" > "LPD64"\n",
3389                                        next_id, last_id);
3390                                 GOTO(cleanup, rc = -EINVAL);
3391                         }
3392                 } else
3393                         next_id = filter_last_id(filter, group) + 1;
3394
3395                 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3396
3397                 dparent = filter_parent_lock(obd, group, next_id);
3398                 if (IS_ERR(dparent))
3399                         GOTO(cleanup, rc = PTR_ERR(dparent));
3400                 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3401
3402                 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3403                 if (IS_ERR(dchild))
3404                         GOTO(cleanup, rc = PTR_ERR(dchild));
3405                 cleanup_phase = 2;  /* f_dput(dchild) */
3406
3407                 if (dchild->d_inode != NULL) {
3408                         /* This would only happen if lastobjid was bad on disk*/
3409                         /* Could also happen if recreating missing obj but it
3410                          * already exists. */
3411                         if (recreate_obj) {
3412                                 CERROR("%s: recreating existing object %.*s?\n",
3413                                        obd->obd_name, dchild->d_name.len,
3414                                        dchild->d_name.name);
3415                         } else {
3416                                 /* Use these existing objects if they are
3417                                  * zero length. */
3418                                 if (dchild->d_inode->i_size == 0) {
3419                                         rc = filter_use_existing_obj(obd,dchild,
3420                                                       &handle, &cleanup_phase);
3421                                         if (rc == 0)
3422                                                 goto set_last_id;
3423                                         else
3424                                                 GOTO(cleanup, rc);
3425                                 }
3426
3427                                 CERROR("%s: Serious error: objid %.*s already "
3428                                        "exists; is this filesystem corrupt?\n",
3429                                        obd->obd_name, dchild->d_name.len,
3430                                        dchild->d_name.name);
3431                                 LBUG();
3432                         }
3433                         GOTO(cleanup, rc = -EEXIST);
3434                 }
3435
3436                 handle = fsfilt_start_log(obd, dparent->d_inode,
3437                                           FSFILT_OP_CREATE, NULL, 1);
3438                 if (IS_ERR(handle))
3439                         GOTO(cleanup, rc = PTR_ERR(handle));
3440                 cleanup_phase = 3;
3441
3442                 /* We mark object SUID+SGID to flag it for accepting UID+GID
3443                  * from client on first write.  Currently the permission bits
3444                  * on the OST are never used, so this is OK. */
3445                 rc = ll_vfs_create(dparent->d_inode, dchild,
3446                                    S_IFREG |  S_ISUID | S_ISGID | 0666, NULL);
3447                 if (rc) {
3448                         CERROR("create failed rc = %d\n", rc);
3449                         GOTO(cleanup, rc);
3450                 }
3451
3452 set_last_id:
3453                 if (!recreate_obj) {
3454                         filter_set_last_id(filter, next_id, group);
3455                         err = filter_update_last_objid(obd, group, 0);
3456                         if (err)
3457                                 CERROR("unable to write lastobjid "
3458                                        "but file created\n");
3459                 }
3460
3461         cleanup:
3462                 switch(cleanup_phase) {
3463                 case 3:
3464                         err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3465                         if (err) {
3466                                 CERROR("error on commit, err = %d\n", err);
3467                                 if (!rc)
3468                                         rc = err;
3469                         }
3470                 case 2:
3471                         f_dput(dchild);
3472                 case 1:
3473                         filter_parent_unlock(dparent);
3474                 case 0:
3475                         break;
3476                 }
3477
3478                 if (rc)
3479                         break;
3480                 if (time_after(jiffies, enough_time)) {
3481                         CDEBUG(D_RPCTRACE,
3482                                "%s: precreate slow - want %d got %d \n",
3483                                obd->obd_name, *num, i);
3484                         break;
3485                 }
3486         }
3487         *num = i;
3488
3489         CDEBUG(D_RPCTRACE,
3490                "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3491                obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3492
3493         RETURN(rc);
3494 }
3495
3496 static int filter_create(struct obd_export *exp, struct obdo *oa,
3497                          struct lov_stripe_md **ea, struct obd_trans_info *oti)
3498 {
3499         struct filter_export_data *fed;
3500         struct obd_device *obd = NULL;
3501         struct filter_obd *filter;
3502         struct lvfs_run_ctxt saved;
3503         struct lov_stripe_md *lsm = NULL;
3504         int rc = 0, diff, group = oa->o_gr;
3505         ENTRY;
3506
3507         if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3508                 CERROR("!!! nid %s sent invalid object group %d\n",
3509                         obd_export_nid2str(exp), group);
3510                 RETURN(-EINVAL);
3511         }
3512
3513         obd = exp->exp_obd;
3514         fed = &exp->exp_filter_data;
3515         filter = &obd->u.filter;
3516
3517         if (fed->fed_group != group) {
3518                 CERROR("!!! this export (nid %s) used object group %d "
3519                         "earlier; now it's trying to use group %d!  This could "
3520                         "be a bug in the MDS.  Tell CFS.\n",
3521                         obd_export_nid2str(exp), fed->fed_group, group);
3522                 RETURN(-ENOTUNIQ);
3523         }
3524
3525         CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3526                oa->o_gr, oa->o_id);
3527         if (ea != NULL) {
3528                 lsm = *ea;
3529                 if (lsm == NULL) {
3530                         rc = obd_alloc_memmd(exp, &lsm);
3531                         if (rc < 0)
3532                                 RETURN(rc);
3533                 }
3534         }
3535
3536         obd = exp->exp_obd;
3537         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3538
3539         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3540             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3541                 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3542                         CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3543                                oa->o_id, filter_last_id(filter,
3544                                                         oa->o_gr));
3545                         rc = -EINVAL;
3546                 } else {
3547                         diff = 1;
3548                         down(&filter->fo_create_locks[oa->o_gr]);
3549                         rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3550                         up(&filter->fo_create_locks[oa->o_gr]);
3551                 }
3552         } else {
3553                 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3554         }
3555
3556         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3557         if (rc && ea != NULL && *ea != lsm) {
3558                 obd_free_memmd(exp, &lsm);
3559         } else if (rc == 0 && ea != NULL) {
3560                 /* XXX LOV STACKING: the lsm that is passed to us from
3561                  * LOV does not have valid lsm_oinfo data structs, so
3562                  * don't go touching that.  This needs to be fixed in a
3563                  * big way. */
3564                 lsm->lsm_object_id = oa->o_id;
3565                 *ea = lsm;
3566         }
3567
3568         RETURN(rc);
3569 }
3570
3571 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3572                    struct lov_stripe_md *md, struct obd_trans_info *oti,
3573                    struct obd_export *md_exp)
3574 {
3575         unsigned int qcids[MAXQUOTAS] = {0, 0};
3576         struct obd_device *obd;
3577         struct filter_obd *filter;
3578         struct dentry *dchild = NULL, *dparent = NULL;
3579         struct lvfs_run_ctxt saved;
3580         void *handle = NULL;
3581         struct llog_cookie *fcc = NULL;
3582         int rc, rc2, cleanup_phase = 0, sync = 0;
3583         struct iattr iattr;
3584         ENTRY;
3585
3586         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3587
3588         obd = exp->exp_obd;
3589         filter = &obd->u.filter;
3590
3591         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3592         cleanup_phase = 1;
3593
3594         dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3595         if (IS_ERR(dchild))
3596                 GOTO(cleanup, rc = PTR_ERR(dchild));
3597         cleanup_phase = 2;
3598
3599         if (dchild->d_inode == NULL) {
3600                 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3601                        oa->o_id);
3602                 /* If object already gone, cancel cookie right now */
3603                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3604                         fcc = obdo_logcookie(oa);
3605                         llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3606                                     NULL, 1, fcc, 0);
3607                         fcc = NULL; /* we didn't allocate fcc, don't free it */
3608                 }
3609                 GOTO(cleanup, rc = -ENOENT);
3610         }
3611
3612         filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3613
3614         /* Our MDC connection is established by the MDS to us */
3615         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3616                 OBD_ALLOC(fcc, sizeof(*fcc));
3617                 if (fcc != NULL)
3618                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3619         }
3620         DQUOT_INIT(dchild->d_inode);
3621
3622         /* we're gonna truncate it first in order to avoid possible deadlock:
3623          *      P1                      P2
3624          * open trasaction      open transaction
3625          * down(i_zombie)       down(i_zombie)
3626          *                      restart transaction
3627          * (see BUG 4180) -bzzz
3628          */
3629         LOCK_INODE_MUTEX(dchild->d_inode);
3630         handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3631                                   NULL, 1);
3632         if (IS_ERR(handle)) {
3633                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3634                 GOTO(cleanup, rc = PTR_ERR(handle));
3635         }
3636
3637         iattr.ia_valid = ATTR_SIZE;
3638         iattr.ia_size = 0;
3639         rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3640         rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3641         UNLOCK_INODE_MUTEX(dchild->d_inode);
3642         if (rc)
3643                 GOTO(cleanup, rc);
3644         if (rc2)
3645                 GOTO(cleanup, rc = rc2);
3646
3647         /* We don't actually need to lock the parent until we are unlinking
3648          * here, and not while truncating above.  That avoids holding the
3649          * parent lock for a long time during truncate, which can block other
3650          * threads from doing anything to objects in that directory. bug 7171 */
3651         dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3652         if (IS_ERR(dparent))
3653                 GOTO(cleanup, rc = PTR_ERR(dparent));
3654         cleanup_phase = 3; /* filter_parent_unlock */
3655
3656         LOCK_INODE_MUTEX(dchild->d_inode);
3657         handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3658         if (IS_ERR(handle)) {
3659                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3660                 GOTO(cleanup, rc = PTR_ERR(handle));
3661         }
3662         cleanup_phase = 4; /* fsfilt_commit */
3663
3664         /* Quota release need uid/gid of inode */
3665         obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3666
3667         filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3668
3669         /* this drops dchild->d_inode->i_mutex unconditionally */
3670         rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3671
3672         EXIT;
3673 cleanup:
3674         switch(cleanup_phase) {
3675         case 4:
3676                 if (fcc != NULL)
3677                         sync = fsfilt_add_journal_cb(obd, 0, oti ?
3678                                                      oti->oti_handle : handle,
3679                                                      filter_cancel_cookies_cb,
3680                                                      fcc);
3681                 /* If add_journal_cb failed, then filter_finish_transno
3682                  * will commit the handle and we will do a sync
3683                  * on commit. then we call callback directly to free
3684                  * the fcc.
3685                  */
3686                 rc = filter_finish_transno(exp, oti, rc, sync);
3687                 if (sync) {
3688                         filter_cancel_cookies_cb(obd, 0, fcc, rc);
3689                         fcc = NULL;
3690                 }
3691                 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3692                 if (rc2) {
3693                         CERROR("error on commit, err = %d\n", rc2);
3694                         if (!rc)
3695                                 rc = rc2;
3696                 } else {
3697                         fcc = NULL;
3698                 }
3699         case 3:
3700                 filter_parent_unlock(dparent);
3701         case 2:
3702                 f_dput(dchild);
3703                 if (fcc != NULL)
3704                         OBD_FREE(fcc, sizeof(*fcc));
3705         case 1:
3706                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3707                 break;
3708         default:
3709                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3710                 LBUG();
3711         }
3712
3713         /* trigger quota release */
3714         qcids[USRQUOTA] = oa->o_uid;
3715         qcids[GRPQUOTA] = oa->o_gid;
3716         rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3717                             FSFILT_OP_UNLINK);
3718
3719         if (rc2)
3720                 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3721         return rc;
3722 }
3723
3724 /* NB start and end are used for punch, but not truncate */
3725 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3726                            struct obd_trans_info *oti,
3727                            struct ptlrpc_request_set *rqset)
3728 {
3729         int rc;
3730         ENTRY;
3731
3732         if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3733                 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3734                        oinfo->oi_policy.l_extent.end);
3735                 RETURN(-EFAULT);
3736         }
3737
3738         CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3739                ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3740                oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3741
3742         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3743                               oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3744         if (rc)
3745                 RETURN(rc);
3746
3747         oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3748         rc = filter_setattr(exp, oinfo, oti);
3749         RETURN(rc);
3750 }
3751
3752 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3753                        struct lov_stripe_md *lsm, obd_off start, obd_off end,
3754                        void *capa)
3755 {
3756         struct lvfs_run_ctxt saved;
3757         struct filter_obd *filter;
3758         struct dentry *dentry;
3759         struct llog_ctxt *ctxt;
3760         int rc, rc2;
3761         ENTRY;
3762
3763         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3764                               (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3765         if (rc)
3766                 RETURN(rc);
3767
3768         filter = &exp->exp_obd->u.filter;
3769
3770         /* an objid of zero is taken to mean "sync whole filesystem" */
3771         if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3772                 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3773                 /* flush any remaining cancel messages out to the target */
3774                 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3775                 llog_sync(ctxt, exp);
3776                 RETURN(rc);
3777         }
3778
3779         dentry = filter_oa2dentry(exp->exp_obd, oa);
3780         if (IS_ERR(dentry))
3781                 RETURN(PTR_ERR(dentry));
3782
3783         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3784
3785         LOCK_INODE_MUTEX(dentry->d_inode);
3786
3787         rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3788         if (rc == 0) {
3789                 /* just any file to grab fsync method - "file" arg unused */
3790                 struct file *file = filter->fo_rcvd_filp;
3791
3792                 if (file->f_op && file->f_op->fsync)
3793                         rc = file->f_op->fsync(NULL, dentry, 1);
3794
3795                 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3796                 if (!rc)
3797                         rc = rc2;
3798         }
3799         UNLOCK_INODE_MUTEX(dentry->d_inode);
3800
3801         oa->o_valid = OBD_MD_FLID;
3802         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3803
3804         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3805
3806         f_dput(dentry);
3807         RETURN(rc);
3808 }
3809
3810 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3811                            void *key, __u32 *vallen, void *val)
3812 {
3813         struct obd_device *obd;
3814         ENTRY;
3815
3816         obd = class_exp2obd(exp);
3817         if (obd == NULL) {
3818                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3819                 RETURN(-EINVAL);
3820         }
3821
3822         if (KEY_IS("blocksize")) {
3823                 __u32 *blocksize = val;
3824                 if (blocksize) {
3825                         if (*vallen < sizeof(*blocksize))
3826                                 RETURN(-EOVERFLOW);
3827                         *blocksize = obd->u.obt.obt_sb->s_blocksize;
3828                 }
3829                 *vallen = sizeof(*blocksize);
3830                 RETURN(0);
3831         }
3832
3833         if (KEY_IS("blocksize_bits")) {
3834                 __u32 *blocksize_bits = val;
3835                 if (blocksize_bits) {
3836                         if (*vallen < sizeof(*blocksize_bits))
3837                                 RETURN(-EOVERFLOW);
3838                         *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3839                 }
3840                 *vallen = sizeof(*blocksize_bits);
3841                 RETURN(0);
3842         }
3843
3844         if (KEY_IS("last_id")) {
3845                 obd_id *last_id = val;
3846                 /* FIXME: object groups */
3847                 if (last_id) {
3848                         if (*vallen < sizeof(*last_id))
3849                                 RETURN(-EOVERFLOW);
3850                         *last_id = filter_last_id(&obd->u.filter, 0);
3851                 }
3852                 *vallen = sizeof(*last_id);
3853                 RETURN(0);
3854         }
3855
3856         CDEBUG(D_IOCTL, "invalid key\n");
3857         RETURN(-EINVAL);
3858 }
3859
3860 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3861                                  void *key, __u32 vallen, void *val,
3862                                  struct ptlrpc_request_set *set)
3863 {
3864         struct obd_device *obd;
3865         struct obd_llogs *llog;
3866         struct llog_ctxt *ctxt;
3867         int rc = 0, group;
3868         ENTRY;
3869
3870         obd = exp->exp_obd;
3871         if (obd == NULL) {
3872                 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3873                 RETURN(-EINVAL);
3874         }
3875
3876         if (KEY_IS(KEY_CAPA_KEY)) {
3877                 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3878                 if (rc)
3879                         CERROR("filter update capability key failed: %d\n", rc);
3880                 RETURN(rc);
3881         }
3882
3883         if (KEY_IS(KEY_REVIMP_UPD)) {
3884                 filter_revimp_update(exp);
3885                 RETURN(0);
3886         }
3887
3888         if (!KEY_IS(KEY_MDS_CONN))
3889                 RETURN(-EINVAL);
3890
3891         LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3892                       obd_export_nid2str(exp));
3893         obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3894
3895         /* setup llog imports */
3896         LASSERT(val != NULL);
3897         group = (int)(*(__u32 *)val);
3898         LASSERT(group >= FILTER_GROUP_MDS0);
3899
3900         llog = filter_grab_llog_for_group(obd, group, exp);
3901         LASSERT(llog != NULL);
3902         ctxt = llog_get_context_from_llogs(llog, LLOG_MDS_OST_REPL_CTXT);
3903         LASSERTF(ctxt != NULL, "ctxt is not null\n"),
3904
3905         rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3906
3907         lquota_setinfo(filter_quota_interface_ref, exp, obd);
3908
3909         RETURN(rc);
3910 }
3911
3912 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3913                      int len, void *karg, void *uarg)
3914 {
3915         struct obd_device *obd = exp->exp_obd;
3916         struct obd_ioctl_data *data = karg;
3917         int rc = 0;
3918
3919         switch (cmd) {
3920         case OBD_IOC_ABORT_RECOVERY: {
3921                 CERROR("aborting recovery for device %s\n", obd->obd_name);
3922                 target_stop_recovery_thread(obd);
3923                 RETURN(0);
3924         }
3925
3926         case OBD_IOC_SYNC: {
3927                 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
3928                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3929                 RETURN(rc);
3930         }
3931
3932         case OBD_IOC_SET_READONLY: {
3933                 void *handle;
3934                 struct super_block *sb = obd->u.obt.obt_sb;
3935                 struct inode *inode = sb->s_root->d_inode;
3936                 BDEVNAME_DECLARE_STORAGE(tmp);
3937                 CERROR("*** setting device %s read-only ***\n",
3938                        ll_bdevname(sb, tmp));
3939
3940                 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3941                 if (!IS_ERR(handle))
3942                         rc = fsfilt_commit(obd, inode, handle, 1);
3943
3944                 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3945                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3946
3947                 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
3948                 RETURN(0);
3949         }
3950
3951         case OBD_IOC_CATLOGLIST: {
3952                 rc = llog_catalog_list(obd, 1, data);
3953                 RETURN(rc);
3954         }
3955
3956         case OBD_IOC_LLOG_CANCEL:
3957         case OBD_IOC_LLOG_REMOVE:
3958         case OBD_IOC_LLOG_INFO:
3959         case OBD_IOC_LLOG_PRINT: {
3960                 /* FIXME to be finished */
3961                 RETURN(-EOPNOTSUPP);
3962 /*
3963                 struct llog_ctxt *ctxt = NULL;
3964
3965                 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3966                 rc = llog_ioctl(ctxt, cmd, data);
3967                 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3968
3969                 RETURN(rc);
3970 */
3971         }
3972
3973
3974         default:
3975                 RETURN(-EINVAL);
3976         }
3977         RETURN(0);
3978 }
3979
3980 static int filter_health_check(struct obd_device *obd)
3981 {
3982 #ifdef USE_HEALTH_CHECK_WRITE
3983         struct filter_obd *filter = &obd->u.filter;
3984 #endif
3985         int rc = 0;
3986
3987         /*
3988          * health_check to return 0 on healthy
3989          * and 1 on unhealthy.
3990          */
3991         if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3992                 rc = 1;
3993
3994 #ifdef USE_HEALTH_CHECK_WRITE
3995         LASSERT(filter->fo_health_check_filp != NULL);
3996         rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3997 #endif
3998         return rc;
3999 }
4000
4001 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
4002                                              void *data)
4003 {
4004         return filter_fid2dentry(data, NULL, gr, id);
4005 }
4006
4007 static int filter_process_config(struct obd_device *obd, obd_count len,
4008                                  void *buf)
4009 {
4010         struct lustre_cfg *lcfg = buf;
4011         struct lprocfs_static_vars lvars;
4012         int rc = 0;
4013
4014         lprocfs_filter_init_vars(&lvars);
4015
4016         rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
4017         return rc;
4018 }
4019
4020 static struct lvfs_callback_ops filter_lvfs_ops = {
4021         l_fid2dentry:     filter_lvfs_fid2dentry,
4022 };
4023
4024 static struct obd_ops filter_obd_ops = {
4025         .o_owner          = THIS_MODULE,
4026         .o_get_info       = filter_get_info,
4027         .o_set_info_async = filter_set_info_async,
4028         .o_setup          = filter_setup,
4029         .o_precleanup     = filter_precleanup,
4030         .o_cleanup        = filter_cleanup,
4031         .o_connect        = filter_connect,
4032         .o_reconnect      = filter_reconnect,
4033         .o_disconnect     = filter_disconnect,
4034         .o_ping           = filter_ping,
4035         .o_init_export    = filter_init_export,
4036         .o_destroy_export = filter_destroy_export,
4037         .o_statfs         = filter_statfs,
4038         .o_getattr        = filter_getattr,
4039         .o_unpackmd       = filter_unpackmd,
4040         .o_create         = filter_create,
4041         .o_setattr        = filter_setattr,
4042         .o_destroy        = filter_destroy,
4043         .o_brw            = filter_brw,
4044         .o_punch          = filter_truncate,
4045         .o_sync           = filter_sync,
4046         .o_preprw         = filter_preprw,
4047         .o_commitrw       = filter_commitrw,
4048         .o_llog_init      = filter_llog_init,
4049         .o_llog_connect   = filter_llog_connect,
4050         .o_llog_finish    = filter_llog_finish,
4051         .o_iocontrol      = filter_iocontrol,
4052         .o_health_check   = filter_health_check,
4053         .o_process_config = filter_process_config,
4054 };
4055
4056 quota_interface_t *filter_quota_interface_ref;
4057 extern quota_interface_t filter_quota_interface;
4058
4059 static int __init obdfilter_init(void)
4060 {
4061         struct lprocfs_static_vars lvars;
4062         int rc;
4063
4064         lprocfs_filter_init_vars(&lvars);
4065
4066         request_module("lquota");
4067         OBD_ALLOC(obdfilter_created_scratchpad,
4068                   OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4069                   sizeof(*obdfilter_created_scratchpad));
4070         if (obdfilter_created_scratchpad == NULL)
4071                 return -ENOMEM;
4072
4073         ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4074                                              sizeof(struct filter_mod_data),
4075                                              0, 0);
4076         if (!ll_fmd_cachep)
4077                 GOTO(out, rc = -ENOMEM);
4078
4079         filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4080         init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4081
4082         rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4083                                  LUSTRE_OST_NAME, NULL);
4084         if (rc) {
4085                 int err;
4086
4087                 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4088                 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4089                 ll_fmd_cachep = NULL;
4090 out:
4091                 if (filter_quota_interface_ref)
4092                         PORTAL_SYMBOL_PUT(filter_quota_interface);
4093
4094                 OBD_FREE(obdfilter_created_scratchpad,
4095                          OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4096                          sizeof(*obdfilter_created_scratchpad));
4097         }
4098
4099         return rc;
4100 }
4101
4102 static void __exit obdfilter_exit(void)
4103 {
4104         if (filter_quota_interface_ref)
4105                 PORTAL_SYMBOL_PUT(filter_quota_interface);
4106
4107         if (ll_fmd_cachep) {
4108                 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4109                 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4110                 ll_fmd_cachep = NULL;
4111         }
4112
4113         class_unregister_type(LUSTRE_OST_NAME);
4114         OBD_FREE(obdfilter_created_scratchpad,
4115                  OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4116                  sizeof(*obdfilter_created_scratchpad));
4117 }
4118
4119 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4120 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4121 MODULE_LICENSE("GPL");
4122
4123 module_init(obdfilter_init);
4124 module_exit(obdfilter_exit);