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
1392         rc = filter_lock_dentry(obd, dparent);
1393         fsfilt_check_slow(obd, now, obd_timeout, "parent lock");
1394         return rc ? ERR_PTR(rc) : dparent;
1395 }
1396
1397 /* We never dget the object parent, so DON'T dput it either */
1398 static void filter_parent_unlock(struct dentry *dparent)
1399 {
1400         UNLOCK_INODE_MUTEX(dparent->d_inode);
1401 }
1402
1403 /* How to get files, dentries, inodes from object id's.
1404  *
1405  * If dir_dentry is passed, the caller has already locked the parent
1406  * appropriately for this operation (normally a write lock).  If
1407  * dir_dentry is NULL, we do a read lock while we do the lookup to
1408  * avoid races with create/destroy and such changing the directory
1409  * internal to the filesystem code. */
1410 struct dentry *filter_fid2dentry(struct obd_device *obd,
1411                                  struct dentry *dir_dentry,
1412                                  obd_gr group, obd_id id)
1413 {
1414         struct dentry *dparent = dir_dentry;
1415         struct dentry *dchild;
1416         char name[32];
1417         int len;
1418         ENTRY;
1419
1420         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) &&
1421             obd->u.filter.fo_destroys_in_progress == 0) {
1422                 /* don't fail lookups for orphan recovery, it causes
1423                  * later LBUGs when objects still exist during precreate */
1424                 CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT);
1425                 RETURN(ERR_PTR(-ENOENT));
1426         }
1427         if (id == 0) {
1428                 CERROR("fatal: invalid object id 0\n");
1429                 RETURN(ERR_PTR(-ESTALE));
1430         }
1431
1432         len = sprintf(name, LPU64, id);
1433         if (dir_dentry == NULL) {
1434                 dparent = filter_parent_lock(obd, group, id);
1435                 if (IS_ERR(dparent)) {
1436                         CERROR("%s: error getting object "LPU64":"LPU64
1437                                " parent: rc %ld\n", obd->obd_name,
1438                                id, group, PTR_ERR(dparent));
1439                         RETURN(dparent);
1440                 }
1441         }
1442         CDEBUG(D_INODE, "looking up object O/%.*s/%s\n",
1443                dparent->d_name.len, dparent->d_name.name, name);
1444         dchild = /*ll_*/lookup_one_len(name, dparent, len);
1445         if (dir_dentry == NULL)
1446                 filter_parent_unlock(dparent);
1447         if (IS_ERR(dchild)) {
1448                 CERROR("%s: child lookup error %ld\n", obd->obd_name,
1449                        PTR_ERR(dchild));
1450                 RETURN(dchild);
1451         }
1452
1453         if (dchild->d_inode != NULL && is_bad_inode(dchild->d_inode)) {
1454                 CERROR("%s: got bad object "LPU64" inode %lu\n",
1455                        obd->obd_name, id, dchild->d_inode->i_ino);
1456                 f_dput(dchild);
1457                 RETURN(ERR_PTR(-ENOENT));
1458         }
1459
1460         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1461                name, dchild, atomic_read(&dchild->d_count));
1462
1463         LASSERT(atomic_read(&dchild->d_count) > 0);
1464
1465         RETURN(dchild);
1466 }
1467
1468 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid,
1469                                   obd_id group)
1470 {
1471         struct lustre_handle lockh;
1472         int flags = LDLM_AST_DISCARD_DATA, rc;
1473         struct ldlm_res_id res_id = { .name = { objid, 0, group, 0} };
1474         ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1475
1476         ENTRY;
1477         /* Tell the clients that the object is gone now and that they should
1478          * throw away any cached pages. */
1479         rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_EXTENT,
1480                                     &policy, LCK_PW, &flags, ldlm_blocking_ast,
1481                                     ldlm_completion_ast, NULL, NULL, 0, NULL,
1482                                     &lockh);
1483
1484         /* We only care about the side-effects, just drop the lock. */
1485         if (rc == ELDLM_OK)
1486                 ldlm_lock_decref(&lockh, LCK_PW);
1487
1488         RETURN(rc);
1489 }
1490
1491 /* This is vfs_unlink() without down(i_sem).  If we call regular vfs_unlink()
1492  * we have 2.6 lock ordering issues with filter_commitrw_write() as it takes
1493  * i_sem before starting a handle, while filter_destroy() + vfs_unlink do the
1494  * reverse.  Caller must take i_sem before starting the transaction and we
1495  * drop it here before the inode is removed from the dentry.  bug 4180/6984 */
1496 int filter_vfs_unlink(struct inode *dir, struct dentry *dentry)
1497 {
1498         int rc;
1499         ENTRY;
1500
1501         /* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
1502          * itself we already hold dir->i_mutex for child create/unlink ops */
1503         LASSERT(dentry->d_inode != NULL);
1504         LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
1505         LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
1506
1507
1508         /* may_delete() */
1509         if (/*!dentry->d_inode ||*/dentry->d_parent->d_inode != dir)
1510                 GOTO(out, rc = -ENOENT);
1511
1512         rc = ll_permission(dir, MAY_WRITE | MAY_EXEC, NULL);
1513         if (rc)
1514                 GOTO(out, rc);
1515
1516         if (IS_APPEND(dir))
1517                 GOTO(out, rc = -EPERM);
1518
1519         /* check_sticky() */
1520         if ((dentry->d_inode->i_uid != current->fsuid && !capable(CAP_FOWNER))||
1521             IS_APPEND(dentry->d_inode) || IS_IMMUTABLE(dentry->d_inode))
1522                 GOTO(out, rc = -EPERM);
1523
1524         /* NOTE: This might need to go outside i_mutex, though it isn't clear if
1525          *       that was done because of journal_start (which is already done
1526          *       here) or some other ordering issue. */
1527         DQUOT_INIT(dir);
1528
1529         rc = security_inode_unlink(dir, dentry);
1530         if (rc)
1531                 GOTO(out, rc);
1532
1533         rc = dir->i_op->unlink(dir, dentry);
1534 out:
1535         /* need to drop i_mutex before we lose inode reference */
1536         UNLOCK_INODE_MUTEX(dentry->d_inode);
1537         if (rc == 0)
1538                 d_delete(dentry);
1539
1540         RETURN(rc);
1541 }
1542
1543 /* Caller must hold LCK_PW on parent and push us into kernel context.
1544  * Caller must hold child i_mutex, we drop it always.
1545  * Caller is also required to ensure that dchild->d_inode exists. */
1546 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1547                                    obd_gr group, struct dentry *dparent,
1548                                    struct dentry *dchild)
1549 {
1550         struct inode *inode = dchild->d_inode;
1551         int rc;
1552
1553         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1554                 CERROR("destroying objid %.*s ino %lu nlink %lu count %d\n",
1555                        dchild->d_name.len, dchild->d_name.name, inode->i_ino,
1556                        (unsigned long)inode->i_nlink,
1557                        atomic_read(&inode->i_count));
1558         }
1559
1560         rc = filter_vfs_unlink(dparent->d_inode, dchild);
1561         if (rc)
1562                 CERROR("error unlinking objid %.*s: rc %d\n",
1563                        dchild->d_name.len, dchild->d_name.name, rc);
1564         return(rc);
1565 }
1566
1567 static int filter_intent_policy(struct ldlm_namespace *ns,
1568                                 struct ldlm_lock **lockp, void *req_cookie,
1569                                 ldlm_mode_t mode, int flags, void *data)
1570 {
1571         struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
1572         struct ptlrpc_request *req = req_cookie;
1573         struct ldlm_lock *lock = *lockp, *l = NULL;
1574         struct ldlm_resource *res = lock->l_resource;
1575         ldlm_processing_policy policy;
1576         struct ost_lvb *res_lvb, *reply_lvb;
1577         struct ldlm_reply *rep;
1578         struct list_head *tmp;
1579         ldlm_error_t err;
1580         int rc, tmpflags = 0, only_liblustre = 0;
1581         int repsize[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1582                            [DLM_LOCKREPLY_OFF]   = sizeof(*rep),
1583                            [DLM_REPLY_REC_OFF]   = sizeof(*reply_lvb) };
1584         ENTRY;
1585
1586         policy = ldlm_get_processing_policy(res);
1587         LASSERT(policy != NULL);
1588         LASSERT(req != NULL);
1589
1590         rc = lustre_pack_reply(req, 3, repsize, NULL);
1591         if (rc)
1592                 RETURN(req->rq_status = rc);
1593
1594         rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
1595         LASSERT(rep != NULL);
1596
1597         reply_lvb = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1598                                    sizeof(*reply_lvb));
1599         LASSERT(reply_lvb != NULL);
1600
1601         //fixup_handle_for_resent_req(req, lock, &lockh);
1602
1603         /* If we grant any lock at all, it will be a whole-file read lock.
1604          * Call the extent policy function to see if our request can be
1605          * granted, or is blocked. */
1606         lock->l_policy_data.l_extent.start = 0;
1607         lock->l_policy_data.l_extent.end = OBD_OBJECT_EOF;
1608         lock->l_req_mode = LCK_PR;
1609
1610         LASSERT(ns == res->lr_namespace);
1611         lock_res(res);
1612         rc = policy(lock, &tmpflags, 0, &err, &rpc_list);
1613         check_res_locked(res);
1614
1615         /* FIXME: we should change the policy function slightly, to not make
1616          * this list at all, since we just turn around and free it */
1617         while (!list_empty(&rpc_list)) {
1618                 struct ldlm_lock *wlock =
1619                         list_entry(rpc_list.next, struct ldlm_lock, l_cp_ast);
1620                 LASSERT((lock->l_flags & LDLM_FL_AST_SENT) == 0);
1621                 LASSERT(lock->l_flags & LDLM_FL_CP_REQD);
1622                 lock->l_flags &= ~LDLM_FL_CP_REQD;
1623                 list_del_init(&wlock->l_cp_ast);
1624                 LDLM_LOCK_PUT(wlock);
1625         }
1626
1627         /* The lock met with no resistance; we're finished. */
1628         if (rc == LDLM_ITER_CONTINUE) {
1629                 /* do not grant locks to the liblustre clients: they cannot
1630                  * handle ASTs robustly.  We need to do this while still
1631                  * holding ns_lock to avoid the lock remaining on the res_link
1632                  * list (and potentially being added to l_pending_list by an
1633                  * AST) when we are going to drop this lock ASAP. */
1634                 if (lock->l_export->exp_libclient ||
1635                     OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2)) {
1636                         ldlm_resource_unlink_lock(lock);
1637                         err = ELDLM_LOCK_ABORTED;
1638                 } else {
1639                         err = ELDLM_LOCK_REPLACED;
1640                 }
1641                 unlock_res(res);
1642                 RETURN(err);
1643         }
1644
1645         /* Do not grant any lock, but instead send GL callbacks.  The extent
1646          * policy nicely created a list of all PW locks for us.  We will choose
1647          * the highest of those which are larger than the size in the LVB, if
1648          * any, and perform a glimpse callback. */
1649         res_lvb = res->lr_lvb_data;
1650         LASSERT(res_lvb != NULL);
1651         *reply_lvb = *res_lvb;
1652
1653         list_for_each(tmp, &res->lr_granted) {
1654                 struct ldlm_lock *tmplock =
1655                         list_entry(tmp, struct ldlm_lock, l_res_link);
1656
1657                 if (tmplock->l_granted_mode == LCK_PR)
1658                         continue;
1659                 /*
1660                  * ->ns_lock guarantees that no new locks are granted, and,
1661                  * therefore, that res->lr_lvb_data cannot increase beyond the
1662                  * end of already granted lock. As a result, it is safe to
1663                  * check against "stale" reply_lvb->lvb_size value without
1664                  * res->lr_lvb_sem.
1665                  */
1666                 if (tmplock->l_policy_data.l_extent.end <= reply_lvb->lvb_size)
1667                         continue;
1668
1669                 /* Don't send glimpse ASTs to liblustre clients.  They aren't
1670                  * listening for them, and they do entirely synchronous I/O
1671                  * anyways. */
1672                 if (tmplock->l_export == NULL ||
1673                     tmplock->l_export->exp_libclient == 1) {
1674                         only_liblustre = 1;
1675                         continue;
1676                 }
1677
1678                 if (l == NULL) {
1679                         l = LDLM_LOCK_GET(tmplock);
1680                         continue;
1681                 }
1682
1683                 if (l->l_policy_data.l_extent.start >
1684                     tmplock->l_policy_data.l_extent.start)
1685                         continue;
1686
1687                 LDLM_LOCK_PUT(l);
1688                 l = LDLM_LOCK_GET(tmplock);
1689         }
1690         unlock_res(res);
1691
1692         /* There were no PW locks beyond the size in the LVB; finished. */
1693         if (l == NULL) {
1694                 if (only_liblustre) {
1695                         /* If we discovered a liblustre client with a PW lock,
1696                          * however, the LVB may be out of date!  The LVB is
1697                          * updated only on glimpse (which we don't do for
1698                          * liblustre clients) and cancel (which the client
1699                          * obviously has not yet done).  So if it has written
1700                          * data but kept the lock, the LVB is stale and needs
1701                          * to be updated from disk.
1702                          *
1703                          * Of course, this will all disappear when we switch to
1704                          * taking liblustre locks on the OST. */
1705                         ldlm_res_lvbo_update(res, NULL, 0, 1);
1706                 }
1707                 RETURN(ELDLM_LOCK_ABORTED);
1708         }
1709
1710         /*
1711          * This check is for lock taken in filter_prepare_destroy() that does
1712          * not have l_glimpse_ast set. So the logic is: if there is a lock
1713          * with no l_glimpse_ast set, this object is being destroyed already.
1714          *
1715          * Hence, if you are grabbing DLM locks on the server, always set
1716          * non-NULL glimpse_ast (e.g., ldlm_request.c:ldlm_glimpse_ast()).
1717          */
1718         if (l->l_glimpse_ast == NULL) {
1719                 /* We are racing with unlink(); just return -ENOENT */
1720                 rep->lock_policy_res1 = -ENOENT;
1721                 goto out;
1722         }
1723
1724         LASSERTF(l->l_glimpse_ast != NULL, "l == %p", l);
1725         rc = l->l_glimpse_ast(l, NULL); /* this will update the LVB */
1726         /* Update the LVB from disk if the AST failed (this is a legal race) */
1727         /*
1728          * XXX nikita: situation when ldlm_server_glimpse_ast() failed before
1729          * sending ast is not handled. This can result in lost client writes.
1730          */
1731         if (rc != 0)
1732                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1733
1734         lock_res(res);
1735         *reply_lvb = *res_lvb;
1736         unlock_res(res);
1737
1738  out:
1739         LDLM_LOCK_PUT(l);
1740
1741         RETURN(ELDLM_LOCK_ABORTED);
1742 }
1743
1744 /*
1745  * per-obd_device iobuf pool.
1746  *
1747  * To avoid memory deadlocks in low-memory setups, amount of dynamic
1748  * allocations in write-path has to be minimized (see bug 5137).
1749  *
1750  * Pages, niobuf_local's and niobuf_remote's are pre-allocated and attached to
1751  * OST threads (see ost_thread_{init,done}()).
1752  *
1753  * "iobuf's" used by filter cannot be attached to OST thread, however, because
1754  * at the OST layer there are only (potentially) multiple obd_device of type
1755  * unknown at the time of OST thread creation.
1756  *
1757  * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
1758  * field). This array has size OST_MAX_THREADS, so that each OST thread uses
1759  * it's very own iobuf.
1760  *
1761  * Functions below
1762  *
1763  *     filter_kiobuf_pool_init()
1764  *
1765  *     filter_kiobuf_pool_done()
1766  *
1767  *     filter_iobuf_get()
1768  *
1769  * operate on this array. They are "generic" in a sense that they don't depend
1770  * on actual type of iobuf's (the latter depending on Linux kernel version).
1771  */
1772
1773 /*
1774  * destroy pool created by filter_iobuf_pool_init
1775  */
1776 static void filter_iobuf_pool_done(struct filter_obd *filter)
1777 {
1778         struct filter_iobuf **pool;
1779         int i;
1780
1781         ENTRY;
1782
1783         pool = filter->fo_iobuf_pool;
1784         if (pool != NULL) {
1785                 for (i = 0; i < filter->fo_iobuf_count; ++ i) {
1786                         if (pool[i] != NULL)
1787                                 filter_free_iobuf(pool[i]);
1788                 }
1789                 OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
1790                 filter->fo_iobuf_pool = NULL;
1791         }
1792         EXIT;
1793 }
1794
1795 /*
1796  * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
1797  */
1798 static int filter_iobuf_pool_init(struct filter_obd *filter)
1799 {
1800         void **pool;
1801
1802         ENTRY;
1803
1804
1805         OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
1806                       GFP_KERNEL);
1807         if (filter->fo_iobuf_pool == NULL)
1808                 RETURN(-ENOMEM);
1809
1810         filter->fo_iobuf_count = OSS_THREADS_MAX;
1811
1812         RETURN(0);
1813 }
1814
1815 /* Return iobuf allocated for @thread_id.  We don't know in advance how
1816  * many threads there will be so we allocate a large empty array and only
1817  * fill in those slots that are actually in use.
1818  * If we haven't allocated a pool entry for this thread before, do so now. */
1819 void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
1820 {
1821         int thread_id                    = oti ? oti->oti_thread_id : -1;
1822         struct filter_iobuf  *pool       = NULL;
1823         struct filter_iobuf **pool_place = NULL;
1824
1825         if (thread_id >= 0) {
1826                 LASSERT(thread_id < filter->fo_iobuf_count);
1827                 pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
1828         }
1829
1830         if (unlikely(pool == NULL)) {
1831                 pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
1832                                           PTLRPC_MAX_BRW_PAGES);
1833                 if (pool_place != NULL)
1834                         *pool_place = pool;
1835         }
1836
1837         return pool;
1838 }
1839
1840 /* mount the file system (secretly).  lustre_cfg parameters are:
1841  * 1 = device
1842  * 2 = fstype
1843  * 3 = flags: failover=f, failout=n
1844  * 4 = mount options
1845  */
1846 int filter_common_setup(struct obd_device *obd, struct lustre_cfg* lcfg,
1847                         void *option)
1848 {
1849         struct filter_obd *filter = &obd->u.filter;
1850         struct vfsmount *mnt;
1851         struct lustre_mount_info *lmi;
1852         struct obd_uuid uuid;
1853         __u8 *uuid_ptr;
1854         char *str, *label;
1855         char ns_name[48];
1856         int rc, i;
1857         ENTRY;
1858
1859         if (lcfg->lcfg_bufcount < 3 ||
1860             LUSTRE_CFG_BUFLEN(lcfg, 1) < 1 ||
1861             LUSTRE_CFG_BUFLEN(lcfg, 2) < 1)
1862                 RETURN(-EINVAL);
1863
1864         lmi = server_get_mount(obd->obd_name);
1865         if (lmi) {
1866                 /* We already mounted in lustre_fill_super.
1867                    lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1868                 struct lustre_sb_info *lsi = s2lsi(lmi->lmi_sb);
1869                 mnt = lmi->lmi_mnt;
1870                 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1871         } else {
1872                 /* old path - used by lctl */
1873                 CERROR("Using old MDS mount method\n");
1874                 mnt = ll_kern_mount(lustre_cfg_string(lcfg, 2),
1875                                     MS_NOATIME|MS_NODIRATIME,
1876                                     lustre_cfg_string(lcfg, 1), option);
1877                 if (IS_ERR(mnt)) {
1878                         rc = PTR_ERR(mnt);
1879                         LCONSOLE_ERROR_MSG(0x135, "Can't mount disk %s (%d)\n",
1880                                            lustre_cfg_string(lcfg, 1), rc);
1881                         RETURN(rc);
1882                 }
1883
1884                 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
1885         }
1886         if (IS_ERR(obd->obd_fsops))
1887                 GOTO(err_mntput, rc = PTR_ERR(obd->obd_fsops));
1888
1889         rc = filter_iobuf_pool_init(filter);
1890         if (rc != 0)
1891                 GOTO(err_ops, rc);
1892
1893         LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1894
1895         /* failover is the default */
1896         obd->obd_replayable = 1;
1897
1898         if (lcfg->lcfg_bufcount > 3 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1899                 str = lustre_cfg_string(lcfg, 3);
1900                 if (strchr(str, 'n')) {
1901                         CWARN("%s: recovery disabled\n", obd->obd_name);
1902                         obd->obd_replayable = 0;
1903                 }
1904         }
1905
1906         filter->fo_vfsmnt = mnt;
1907         obd->u.obt.obt_sb = mnt->mnt_sb;
1908         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1909         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1910
1911         fsfilt_setup(obd, obd->u.obt.obt_sb);
1912
1913         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1914         obd->obd_lvfs_ctxt.pwdmnt = mnt;
1915         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
1916         obd->obd_lvfs_ctxt.fs = get_ds();
1917         obd->obd_lvfs_ctxt.cb_ops = filter_lvfs_ops;
1918
1919         sema_init(&filter->fo_init_lock, 1);
1920         filter->fo_committed_group = 0;
1921
1922         rc = filter_prep(obd);
1923         if (rc)
1924                 GOTO(err_ops, rc);
1925
1926         filter->fo_destroys_in_progress = 0;
1927         for (i = 0; i < 32; i++)
1928                 sema_init(&filter->fo_create_locks[i], 1);
1929
1930         spin_lock_init(&filter->fo_translock);
1931         spin_lock_init(&filter->fo_objidlock);
1932         INIT_LIST_HEAD(&filter->fo_export_list);
1933         sema_init(&filter->fo_alloc_lock, 1);
1934         init_brw_stats(&filter->fo_filter_stats);
1935         filter->fo_readcache_max_filesize = FILTER_MAX_CACHE_SIZE;
1936         filter->fo_fmd_max_num = FILTER_FMD_MAX_NUM_DEFAULT;
1937         filter->fo_fmd_max_age = FILTER_FMD_MAX_AGE_DEFAULT;
1938
1939         INIT_LIST_HEAD(&filter->fo_llog_list);
1940         spin_lock_init(&filter->fo_llog_list_lock);
1941
1942         filter->fo_fl_oss_capa = 0;
1943         INIT_LIST_HEAD(&filter->fo_capa_keys);
1944         filter->fo_capa_hash = init_capa_hash();
1945         if (filter->fo_capa_hash == NULL)
1946                 GOTO(err_ops, rc = -ENOMEM);
1947
1948         sprintf(ns_name, "filter-%s", obd->obd_uuid.uuid);
1949         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1950                                                 LDLM_NAMESPACE_GREEDY);
1951         if (obd->obd_namespace == NULL)
1952                 GOTO(err_post, rc = -ENOMEM);
1953         obd->obd_namespace->ns_lvbp = obd;
1954         obd->obd_namespace->ns_lvbo = &filter_lvbo;
1955         ldlm_register_intent(obd->obd_namespace, filter_intent_policy);
1956
1957         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1958                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1959
1960         rc = llog_cat_initialize(obd, NULL, 1, NULL);
1961         if (rc) {
1962                 CERROR("failed to setup llogging subsystems\n");
1963                 GOTO(err_post, rc);
1964         }
1965
1966         rc = lquota_setup(filter_quota_interface_ref, obd);
1967         if (rc)
1968                 GOTO(err_post, rc);
1969
1970         uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
1971         if (uuid_ptr != NULL) {
1972                 class_uuid_unparse(uuid_ptr, &uuid);
1973                 str = uuid.uuid;
1974         } else {
1975                 str = "no UUID";
1976         }
1977
1978         label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
1979
1980         if (obd->obd_recovering) {
1981                 LCONSOLE_WARN("OST %s now serving %s (%s%s%s), but will be in "
1982                               "recovery until %d %s reconnect, or if no clients"
1983                               " reconnect for %d:%.02d; during that time new "
1984                               "clients will not be allowed to connect. "
1985                               "Recovery progress can be monitored by watching "
1986                               "/proc/fs/lustre/obdfilter/%s/recovery_status.\n",
1987                               obd->obd_name, lustre_cfg_string(lcfg, 1),
1988                               label ?: "", label ? "/" : "", str,
1989                               obd->obd_max_recoverable_clients,
1990                               (obd->obd_max_recoverable_clients == 1)
1991                               ? "client" : "clients",
1992                               (int)(OBD_RECOVERY_TIMEOUT) / 60,
1993                               (int)(OBD_RECOVERY_TIMEOUT) % 60,
1994                               obd->obd_name);
1995         } else {
1996                 LCONSOLE_INFO("OST %s now serving %s (%s%s%s) with recovery "
1997                               "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
1998                               label ?: "", label ? "/" : "", str,
1999                               obd->obd_replayable ? "enabled" : "disabled");
2000         }
2001
2002         RETURN(0);
2003
2004 err_post:
2005         filter_post(obd);
2006 err_ops:
2007         fsfilt_put_ops(obd->obd_fsops);
2008         filter_iobuf_pool_done(filter);
2009 err_mntput:
2010         server_put_mount(obd->obd_name, mnt);
2011         obd->u.obt.obt_sb = 0;
2012         return rc;
2013 }
2014
2015 static int filter_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2016 {
2017         struct lprocfs_static_vars lvars;
2018         unsigned long page;
2019         int rc;
2020
2021         CLASSERT(offsetof(struct obd_device, u.obt) ==
2022                  offsetof(struct obd_device, u.filter.fo_obt));
2023
2024         if (!LUSTRE_CFG_BUFLEN(lcfg, 1) || !LUSTRE_CFG_BUFLEN(lcfg, 2))
2025                 RETURN(-EINVAL);
2026
2027         /* 2.6.9 selinux wants a full option page for do_kern_mount (bug6471) */
2028         page = get_zeroed_page(GFP_KERNEL);
2029         if (!page)
2030                 RETURN(-ENOMEM);
2031
2032         /* lprocfs must be setup before the filter so state can be safely added
2033          * to /proc incrementally as the filter is setup */
2034         lprocfs_init_vars(filter, &lvars);
2035         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2036             lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST) == 0) {
2037                 /* Init obdfilter private stats here */
2038                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
2039                                      LPROCFS_CNTR_AVGMINMAX,
2040                                      "read_bytes", "bytes");
2041                 lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
2042                                      LPROCFS_CNTR_AVGMINMAX,
2043                                      "write_bytes", "bytes");
2044
2045                 lproc_filter_attach_seqstat(obd);
2046                 obd->obd_proc_exports = proc_mkdir("exports",
2047                                                    obd->obd_proc_entry);
2048         }
2049
2050         memcpy((void *)page, lustre_cfg_buf(lcfg, 4),
2051                LUSTRE_CFG_BUFLEN(lcfg, 4));
2052         rc = filter_common_setup(obd, lcfg, (void *)page);
2053         free_page(page);
2054
2055         if (rc) {
2056                 lprocfs_obd_cleanup(obd);
2057                 lprocfs_free_obd_stats(obd);
2058         }
2059
2060         return rc;
2061 }
2062
2063 static struct llog_operations filter_mds_ost_repl_logops /* initialized below*/;
2064 static struct llog_operations filter_size_orig_logops = {
2065         lop_setup: llog_obd_origin_setup,
2066         lop_cleanup: llog_obd_origin_cleanup,
2067         lop_add: llog_obd_origin_add
2068 };
2069
2070 static int filter_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
2071                             struct obd_device *tgt, int count,
2072                             struct llog_catid *catid,
2073                             struct obd_uuid *uuid)
2074 {
2075         struct llog_ctxt *ctxt;
2076         int rc;
2077         ENTRY;
2078
2079         filter_mds_ost_repl_logops = llog_client_ops;
2080         filter_mds_ost_repl_logops.lop_cancel = llog_obd_repl_cancel;
2081         filter_mds_ost_repl_logops.lop_connect = llog_repl_connect;
2082         filter_mds_ost_repl_logops.lop_sync = llog_obd_repl_sync;
2083
2084         rc = llog_setup(obd, llogs, LLOG_MDS_OST_REPL_CTXT, tgt, 0, NULL,
2085                         &filter_mds_ost_repl_logops);
2086         if (rc)
2087                 RETURN(rc);
2088
2089         /* FIXME - assign unlink_cb for filter's recovery */
2090         if (!llogs)
2091                 ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2092         else
2093                 ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2094
2095         LASSERT(ctxt != NULL);
2096         ctxt->llog_proc_cb = filter_recov_log_mds_ost_cb;
2097
2098         rc = llog_setup(obd, llogs, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2099                         &filter_size_orig_logops);
2100         RETURN(rc);
2101 }
2102
2103 static int filter_group_llog_cleanup(struct llog_ctxt *ctxt)
2104 {
2105         int rc = 0;
2106         ENTRY;
2107
2108         if (CTXTP(ctxt, cleanup))
2109                 rc = CTXTP(ctxt, cleanup)(ctxt);
2110
2111         if (ctxt->loc_exp)
2112                 class_export_put(ctxt->loc_exp);
2113         OBD_FREE(ctxt, sizeof(*ctxt));
2114
2115         RETURN(rc);
2116 }
2117
2118 static int filter_group_llog_finish(struct obd_llogs *llogs)
2119 {
2120         struct llog_ctxt *ctxt;
2121         int rc = 0, rc2 = 0;
2122         ENTRY;
2123
2124         ctxt = llog_get_context_from_llogs(llogs, LLOG_MDS_OST_REPL_CTXT);
2125         if (ctxt)
2126                 rc = filter_group_llog_cleanup(ctxt);
2127
2128         ctxt = llog_get_context_from_llogs(llogs, LLOG_SIZE_ORIG_CTXT);
2129         if (ctxt)
2130                 rc2 = filter_group_llog_cleanup(ctxt);
2131         if (!rc)
2132                 rc = rc2;
2133
2134         RETURN(rc);
2135 }
2136
2137 static int filter_llog_finish(struct obd_device *obd, int count)
2138 {
2139         struct llog_ctxt *ctxt;
2140         int rc = 0, rc2 = 0;
2141         ENTRY;
2142
2143         ctxt = llog_get_context(obd, LLOG_MDS_OST_REPL_CTXT);
2144         if (ctxt)
2145                 rc = llog_cleanup(ctxt);
2146
2147         ctxt = llog_get_context(obd, LLOG_SIZE_ORIG_CTXT);
2148         if (ctxt)
2149                 rc2 = llog_cleanup(ctxt);
2150         if (!rc)
2151                 rc = rc2;
2152
2153         RETURN(rc);
2154 }
2155
2156 struct obd_llogs *filter_grab_llog_for_group(struct obd_device *obd, int group,
2157                                              struct obd_export *export)
2158 {
2159         struct filter_group_llog *fglog, *nlog;
2160         struct filter_obd *filter;
2161         struct llog_ctxt *ctxt;
2162         struct list_head *cur;
2163         int rc;
2164
2165         filter = &obd->u.filter;
2166
2167         spin_lock(&filter->fo_llog_list_lock);
2168         list_for_each(cur, &filter->fo_llog_list) {
2169                 fglog = list_entry(cur, struct filter_group_llog, list);
2170                 if (fglog->group == group) {
2171                         if (!(fglog->exp == NULL || fglog->exp == export || export == NULL))
2172                                 CWARN("%s: export for group %d changes: 0x%p -> 0x%p\n",
2173                                       obd->obd_name, group, fglog->exp, export);
2174                         spin_unlock(&filter->fo_llog_list_lock);
2175                         goto init;
2176                 }
2177         }
2178         spin_unlock(&filter->fo_llog_list_lock);
2179
2180         if (export == NULL)
2181                 RETURN(NULL);
2182
2183         OBD_ALLOC_PTR(fglog);
2184         if (fglog == NULL)
2185                 RETURN(NULL);
2186         fglog->group = group;
2187
2188         OBD_ALLOC_PTR(fglog->llogs);
2189         if (fglog->llogs == NULL) {
2190                 OBD_FREE_PTR(fglog);
2191                 RETURN(NULL);
2192         }
2193
2194         spin_lock(&filter->fo_llog_list_lock);
2195         list_for_each(cur, &filter->fo_llog_list) {
2196                 nlog = list_entry(cur, struct filter_group_llog, list);
2197                 LASSERT(nlog->group != group);
2198         }
2199         list_add(&fglog->list, &filter->fo_llog_list);
2200         spin_unlock(&filter->fo_llog_list_lock);
2201
2202         rc = llog_cat_initialize(obd, fglog->llogs, 1, NULL);
2203         if (rc) {
2204                 OBD_FREE_PTR(fglog->llogs);
2205                 OBD_FREE_PTR(fglog);
2206                 RETURN(NULL);
2207         }
2208
2209 init:
2210         if (export) {
2211                 fglog->exp = export;
2212                 ctxt = llog_get_context_from_llogs(fglog->llogs,
2213                                                LLOG_MDS_OST_REPL_CTXT);
2214                 LASSERT(ctxt != NULL);
2215
2216                 llog_receptor_accept(ctxt, export->exp_imp_reverse);
2217         }
2218         CDEBUG(D_OTHER, "%s: new llog 0x%p for group %u\n",
2219                obd->obd_name, fglog->llogs, group);
2220
2221         RETURN(fglog->llogs);
2222 }
2223
2224 static int filter_llog_connect(struct obd_export *exp,
2225                                struct llogd_conn_body *body)
2226 {
2227         struct obd_device *obd = exp->exp_obd;
2228         struct llog_ctxt *ctxt;
2229         struct obd_llogs *llog;
2230         int rc;
2231         ENTRY;
2232
2233         CDEBUG(D_OTHER, "handle connect for %s: %u/%u/%u\n", obd->obd_name,
2234                 (unsigned) body->lgdc_logid.lgl_ogr,
2235                 (unsigned) body->lgdc_logid.lgl_oid,
2236                 (unsigned) body->lgdc_logid.lgl_ogen);
2237
2238         llog = filter_grab_llog_for_group(obd, body->lgdc_logid.lgl_ogr, exp);
2239         LASSERT(llog != NULL);
2240         ctxt = llog_get_context_from_llogs(llog, body->lgdc_ctxt_idx);
2241         LASSERTF(ctxt != NULL, "ctxt is not null, ctxt idx %d \n",
2242                  body->lgdc_ctxt_idx);
2243         rc = llog_connect(ctxt, 1, &body->lgdc_logid,
2244                           &body->lgdc_gen, NULL);
2245         if (rc != 0)
2246                 CERROR("failed to connect rc %d idx %d\n", rc,
2247                                 body->lgdc_ctxt_idx);
2248
2249         RETURN(rc);
2250 }
2251
2252 static int filter_llog_preclean (struct obd_device *obd)
2253 {
2254         struct filter_group_llog *log;
2255         struct filter_obd *filter;
2256         int rc = 0;
2257         ENTRY;
2258
2259         filter = &obd->u.filter;
2260         spin_lock(&filter->fo_llog_list_lock);
2261         while (!list_empty(&filter->fo_llog_list)) {
2262                 log = list_entry(filter->fo_llog_list.next,
2263                                  struct filter_group_llog, list);
2264                 list_del(&log->list);
2265                 spin_unlock(&filter->fo_llog_list_lock);
2266
2267                 rc = filter_group_llog_finish(log->llogs);
2268                 if (rc)
2269                         CERROR("failed to cleanup llogging subsystem for %u\n",
2270                                 log->group);
2271                 OBD_FREE_PTR(log->llogs);
2272                 OBD_FREE_PTR(log);
2273                 spin_lock(&filter->fo_llog_list_lock);
2274         }
2275         spin_unlock(&filter->fo_llog_list_lock);
2276
2277         rc = obd_llog_finish(obd, 0);
2278         if (rc)
2279                 CERROR("failed to cleanup llogging subsystem\n");
2280
2281         RETURN(rc);
2282 }
2283
2284 static int filter_precleanup(struct obd_device *obd,
2285                              enum obd_cleanup_stage stage)
2286 {
2287         int rc = 0;
2288         ENTRY;
2289
2290         switch(stage) {
2291         case OBD_CLEANUP_EARLY:
2292                 break;
2293         case OBD_CLEANUP_EXPORTS:
2294                 target_cleanup_recovery(obd);
2295                 break;
2296         case OBD_CLEANUP_SELF_EXP:
2297                 rc = filter_llog_preclean(obd);
2298                 break;
2299         case OBD_CLEANUP_OBD:
2300                 break;
2301         }
2302         RETURN(rc);
2303 }
2304
2305 static int filter_cleanup(struct obd_device *obd)
2306 {
2307         struct filter_obd *filter = &obd->u.filter;
2308         ENTRY;
2309
2310         if (obd->obd_fail)
2311                 LCONSOLE_WARN("%s: shutting down for failover; client state "
2312                               "will be preserved.\n", obd->obd_name);
2313
2314         if (!list_empty(&obd->obd_exports)) {
2315                 CERROR("%s: still has clients!\n", obd->obd_name);
2316                 class_disconnect_exports(obd);
2317                 if (!list_empty(&obd->obd_exports)) {
2318                         CERROR("still has exports after forced cleanup?\n");
2319                         RETURN(-EBUSY);
2320                 }
2321         }
2322
2323         lprocfs_obd_cleanup(obd);
2324         lprocfs_free_obd_stats(obd);
2325         lquota_cleanup(filter_quota_interface_ref, obd);
2326
2327         /* Stop recovery before namespace cleanup. */
2328         target_stop_recovery_thread(obd);
2329         target_cleanup_recovery(obd);
2330
2331         ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2332
2333         if (obd->u.obt.obt_sb == NULL)
2334                 RETURN(0);
2335
2336         filter_post(obd);
2337
2338         shrink_dcache_parent(obd->u.obt.obt_sb->s_root);
2339
2340         LL_DQUOT_OFF(obd->u.obt.obt_sb);
2341
2342         server_put_mount(obd->obd_name, filter->fo_vfsmnt);
2343         obd->u.obt.obt_sb = NULL;
2344
2345         fsfilt_put_ops(obd->obd_fsops);
2346
2347         filter_iobuf_pool_done(filter);
2348
2349         LCONSOLE_INFO("OST %s has stopped.\n", obd->obd_name);
2350
2351         RETURN(0);
2352 }
2353
2354 static int filter_connect_internal(struct obd_export *exp,
2355                                    struct obd_connect_data *data)
2356 {
2357         if (!data)
2358                 RETURN(0);
2359
2360         CDEBUG(D_RPCTRACE, "%s: cli %s/%p ocd_connect_flags: "LPX64
2361                " ocd_version: %x ocd_grant: %d ocd_index: %u\n",
2362                exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp,
2363                data->ocd_connect_flags, data->ocd_version,
2364                data->ocd_grant, data->ocd_index);
2365
2366         data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
2367         exp->exp_connect_flags = data->ocd_connect_flags;
2368         data->ocd_version = LUSTRE_VERSION_CODE;
2369
2370         if (exp->exp_connect_flags & OBD_CONNECT_GRANT) {
2371                 struct filter_export_data *fed = &exp->exp_filter_data;
2372                 obd_size left, want;
2373
2374                 spin_lock(&exp->exp_obd->obd_osfs_lock);
2375                 left = filter_grant_space_left(exp);
2376                 want = data->ocd_grant;
2377                 filter_grant(exp, fed->fed_grant, want, left);
2378                 data->ocd_grant = fed->fed_grant;
2379                 spin_unlock(&exp->exp_obd->obd_osfs_lock);
2380
2381                 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %d want: "
2382                        LPU64" left: "LPU64"\n", exp->exp_obd->obd_name,
2383                        exp->exp_client_uuid.uuid, exp,
2384                        data->ocd_grant, want, left);
2385         }
2386
2387         if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
2388                 struct filter_obd *filter = &exp->exp_obd->u.filter;
2389                 struct lr_server_data *lsd = filter->fo_fsd;
2390                 int index = le32_to_cpu(lsd->lsd_ost_index);
2391
2392                 if (!(lsd->lsd_feature_compat &
2393                       cpu_to_le32(OBD_COMPAT_OST))) {
2394                         /* this will only happen on the first connect */
2395                         lsd->lsd_ost_index = cpu_to_le32(data->ocd_index);
2396                         lsd->lsd_feature_compat |= cpu_to_le32(OBD_COMPAT_OST);
2397                         filter_update_server_data(exp->exp_obd,
2398                                                   filter->fo_rcvd_filp, lsd, 1);
2399                 } else if (index != data->ocd_index) {
2400                         LCONSOLE_ERROR_MSG(0x136, "Connection from %s to index"
2401                                            " %u doesn't match actual OST index"
2402                                            " %u in last_rcvd file, bad "
2403                                            "configuration?\n",
2404                                            obd_export_nid2str(exp), index,
2405                                            data->ocd_index);
2406                         RETURN(-EBADF);
2407                 }
2408         }
2409
2410         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_SIZE)) {
2411                 data->ocd_brw_size = 65536;
2412         } else if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
2413                 data->ocd_brw_size = min(data->ocd_brw_size,
2414                                (__u32)(PTLRPC_MAX_BRW_PAGES << CFS_PAGE_SHIFT));
2415                 LASSERT(data->ocd_brw_size);
2416         }
2417
2418         /* FIXME: Do the same with the MDS UUID and fsd_peeruuid.
2419          * FIXME: We don't strictly need the COMPAT flag for that,
2420          * FIXME: as fsd_peeruuid[0] will tell us if that is set.
2421          * FIXME: We needed it for the index, as index 0 is valid. */
2422
2423         RETURN(0);
2424 }
2425
2426 static int filter_reconnect(struct obd_export *exp, struct obd_device *obd,
2427                             struct obd_uuid *cluuid,
2428                             struct obd_connect_data *data)
2429 {
2430         int rc;
2431         ENTRY;
2432
2433         if (exp == NULL || obd == NULL || cluuid == NULL)
2434                 RETURN(-EINVAL);
2435
2436         rc = filter_connect_internal(exp, data);
2437
2438         RETURN(rc);
2439 }
2440
2441 /* nearly identical to mds_connect */
2442 static int filter_connect(const struct lu_env *env,
2443                           struct lustre_handle *conn, struct obd_device *obd,
2444                           struct obd_uuid *cluuid,
2445                           struct obd_connect_data *data)
2446 {
2447         struct lvfs_run_ctxt saved;
2448         struct obd_export *exp;
2449         struct filter_export_data *fed;
2450         struct filter_client_data *fcd = NULL;
2451         __u32 group;
2452         int rc;
2453         ENTRY;
2454
2455         if (conn == NULL || obd == NULL || cluuid == NULL)
2456                 RETURN(-EINVAL);
2457
2458         rc = class_connect(conn, obd, cluuid);
2459         if (rc)
2460                 RETURN(rc);
2461         exp = class_conn2export(conn);
2462         LASSERT(exp != NULL);
2463
2464         fed = &exp->exp_filter_data;
2465
2466         rc = filter_connect_internal(exp, data);
2467         if (rc)
2468                 GOTO(cleanup, rc);
2469
2470         filter_export_stats_init(obd, exp);
2471         group = data->ocd_group;
2472         if (obd->obd_replayable) {
2473                 OBD_ALLOC(fcd, sizeof(*fcd));
2474                 if (!fcd) {
2475                         CERROR("filter: out of memory for client data\n");
2476                         GOTO(cleanup, rc = -ENOMEM);
2477                 }
2478
2479                 memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
2480                 fed->fed_fcd = fcd;
2481                 fed->fed_fcd->fcd_group = group;
2482                 rc = filter_client_add(obd, exp, -1);
2483                 if (rc)
2484                         GOTO(cleanup, rc);
2485         }
2486         CWARN("%s: Received MDS connection ("LPX64"); group %d\n",
2487                obd->obd_name, exp->exp_handle.h_cookie, group);
2488         if (group == 0)
2489                 GOTO(cleanup, rc);
2490
2491         if (fed->fed_group != 0 && fed->fed_group != group) {
2492                 CERROR("!!! This export (nid %s) used object group %d "
2493                        "earlier; now it's trying to use group %d!  This could "
2494                        "be a bug in the MDS.  Tell CFS.\n",
2495                        obd_export_nid2str(exp), fed->fed_group, group);
2496                 GOTO(cleanup, rc = -EPROTO);
2497         }
2498         fed->fed_group = group;
2499
2500         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2501         rc = filter_read_groups(obd, group, 1);
2502         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2503         if (rc != 0) {
2504                 CERROR("can't read group %u\n", group);
2505                 GOTO(cleanup, rc);
2506         }
2507
2508         GOTO(cleanup, rc);
2509
2510 cleanup:
2511         if (rc) {
2512                 if (fcd) {
2513                         OBD_FREE(fcd, sizeof(*fcd));
2514                         fed->fed_fcd = NULL;
2515                 }
2516                 class_disconnect(exp);
2517         } else {
2518                 class_export_put(exp);
2519         }
2520
2521         RETURN(rc);
2522 }
2523
2524 /* Do extra sanity checks for grant accounting.  We do this at connect,
2525  * disconnect, and statfs RPC time, so it shouldn't be too bad.  We can
2526  * always get rid of it or turn it off when we know accounting is good. */
2527 static void filter_grant_sanity_check(struct obd_device *obd, const char *func)
2528 {
2529         struct filter_export_data *fed;
2530         struct obd_export *exp;
2531         obd_size maxsize = obd->obd_osfs.os_blocks * obd->obd_osfs.os_bsize;
2532         obd_size tot_dirty = 0, tot_pending = 0, tot_granted = 0;
2533         obd_size fo_tot_dirty, fo_tot_pending, fo_tot_granted;
2534
2535         if (list_empty(&obd->obd_exports))
2536                 return;
2537
2538         /* We don't want to do this for large machines that do lots of
2539            mounts or unmounts.  It burns... */
2540         if (obd->obd_num_exports > 100)
2541                 return;
2542
2543         spin_lock(&obd->obd_osfs_lock);
2544         spin_lock(&obd->obd_dev_lock);
2545         list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
2546                 int error = 0;
2547                 fed = &exp->exp_filter_data;
2548                 if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
2549                     fed->fed_dirty < 0)
2550                         error = 1;
2551                 if (maxsize > 0) { /* we may not have done a statfs yet */
2552                         LASSERTF(fed->fed_grant + fed->fed_pending <= maxsize,
2553                                  "%s: cli %s/%p %ld+%ld > "LPU64"\n", func,
2554                                  exp->exp_client_uuid.uuid, exp,
2555                                  fed->fed_grant, fed->fed_pending, maxsize);
2556                         LASSERTF(fed->fed_dirty <= maxsize,
2557                                  "%s: cli %s/%p %ld > "LPU64"\n", func,
2558                                  exp->exp_client_uuid.uuid, exp,
2559                                  fed->fed_dirty, maxsize);
2560                 }
2561                 if (error)
2562                         CERROR("%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2563                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2564                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2565                 else
2566                         CDEBUG(D_CACHE, "%s: cli %s/%p dirty %ld pend %ld grant %ld\n",
2567                                obd->obd_name, exp->exp_client_uuid.uuid, exp,
2568                                fed->fed_dirty, fed->fed_pending,fed->fed_grant);
2569                 tot_granted += fed->fed_grant + fed->fed_pending;
2570                 tot_pending += fed->fed_pending;
2571                 tot_dirty += fed->fed_dirty;
2572         }
2573         fo_tot_granted = obd->u.filter.fo_tot_granted;
2574         fo_tot_pending = obd->u.filter.fo_tot_pending;
2575         fo_tot_dirty = obd->u.filter.fo_tot_dirty;
2576         spin_unlock(&obd->obd_dev_lock);
2577         spin_unlock(&obd->obd_osfs_lock);
2578
2579         /* Do these assertions outside the spinlocks so we don't kill system */
2580         if (tot_granted != fo_tot_granted)
2581                 CERROR("%s: tot_granted "LPU64" != fo_tot_granted "LPU64"\n",
2582                        func, tot_granted, fo_tot_granted);
2583         if (tot_pending != fo_tot_pending)
2584                 CERROR("%s: tot_pending "LPU64" != fo_tot_pending "LPU64"\n",
2585                        func, tot_pending, fo_tot_pending);
2586         if (tot_dirty != fo_tot_dirty)
2587                 CERROR("%s: tot_dirty "LPU64" != fo_tot_dirty "LPU64"\n",
2588                        func, tot_dirty, fo_tot_dirty);
2589         if (tot_pending > tot_granted)
2590                 CERROR("%s: tot_pending "LPU64" > tot_granted "LPU64"\n",
2591                        func, tot_pending, tot_granted);
2592         if (tot_granted > maxsize)
2593                 CERROR("%s: tot_granted "LPU64" > maxsize "LPU64"\n",
2594                        func, tot_granted, maxsize);
2595         if (tot_dirty > maxsize)
2596                 CERROR("%s: tot_dirty "LPU64" > maxsize "LPU64"\n",
2597                        func, tot_dirty, maxsize);
2598 }
2599
2600 /* Remove this client from the grant accounting totals.  We also remove
2601  * the export from the obd device under the osfs and dev locks to ensure
2602  * that the filter_grant_sanity_check() calculations are always valid.
2603  * The client should do something similar when it invalidates its import. */
2604 static void filter_grant_discard(struct obd_export *exp)
2605 {
2606         struct obd_device *obd = exp->exp_obd;
2607         struct filter_obd *filter = &obd->u.filter;
2608         struct filter_export_data *fed = &exp->exp_filter_data;
2609
2610         spin_lock(&obd->obd_osfs_lock);
2611         spin_lock(&obd->obd_dev_lock);
2612         list_del_init(&exp->exp_obd_chain);
2613         spin_unlock(&obd->obd_dev_lock);
2614
2615         LASSERTF(filter->fo_tot_granted >= fed->fed_grant,
2616                  "%s: tot_granted "LPU64" cli %s/%p fed_grant %ld\n",
2617                  obd->obd_name, filter->fo_tot_granted,
2618                  exp->exp_client_uuid.uuid, exp, fed->fed_grant);
2619         filter->fo_tot_granted -= fed->fed_grant;
2620         LASSERTF(filter->fo_tot_pending >= fed->fed_pending,
2621                  "%s: tot_pending "LPU64" cli %s/%p fed_pending %ld\n",
2622                  obd->obd_name, filter->fo_tot_pending,
2623                  exp->exp_client_uuid.uuid, exp, fed->fed_pending);
2624         /* fo_tot_pending is handled in filter_grant_commit as bulk finishes */
2625         LASSERTF(filter->fo_tot_dirty >= fed->fed_dirty,
2626                  "%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
2627                  obd->obd_name, filter->fo_tot_dirty,
2628                  exp->exp_client_uuid.uuid, exp, fed->fed_dirty);
2629         filter->fo_tot_dirty -= fed->fed_dirty;
2630         fed->fed_dirty = 0;
2631         fed->fed_grant = 0;
2632
2633         spin_unlock(&obd->obd_osfs_lock);
2634 }
2635
2636 static int filter_destroy_export(struct obd_export *exp)
2637 {
2638         ENTRY;
2639
2640         if (exp->exp_filter_data.fed_pending)
2641                 CERROR("%s: cli %s/%p has %lu pending on destroyed export\n",
2642                        exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
2643                        exp, exp->exp_filter_data.fed_pending);
2644
2645         target_destroy_export(exp);
2646
2647         if (obd_uuid_equals(&exp->exp_client_uuid, &exp->exp_obd->obd_uuid))
2648                 RETURN(0);
2649
2650         lprocfs_exp_cleanup(exp);
2651
2652         if (exp->exp_obd->obd_replayable)
2653                 filter_client_free(exp);
2654         else
2655                 fsfilt_sync(exp->exp_obd, exp->exp_obd->u.obt.obt_sb);
2656
2657         filter_grant_discard(exp);
2658         filter_fmd_cleanup(exp);
2659
2660         if (!(exp->exp_flags & OBD_OPT_FORCE))
2661                 filter_grant_sanity_check(exp->exp_obd, __FUNCTION__);
2662
2663         RETURN(0);
2664 }
2665
2666 static void filter_sync_llogs(struct obd_device *obd, struct obd_export *dexp)
2667 {
2668         struct filter_group_llog *fglog, *nlog;
2669         struct filter_obd *filter;
2670         int worked = 0, group;
2671         struct llog_ctxt *ctxt;
2672         ENTRY;
2673
2674         filter = &obd->u.filter;
2675
2676         /* we can't sync log holding spinlock. also, we do not want to get
2677          * into livelock. so we do following: loop over MDS's exports in
2678          * group order and skip already synced llogs -bzzz */
2679         do {
2680                 /* look for group with min. number, but > worked */
2681                 fglog = NULL;
2682                 group = 1 << 30;
2683                 spin_lock(&filter->fo_llog_list_lock);
2684                 list_for_each_entry(nlog, &filter->fo_llog_list, list) {
2685                         if (nlog->group <= worked) {
2686                                 /* this group is already synced */
2687                                 continue;
2688                         }
2689                         if (group < nlog->group) {
2690                                 /* we have group with smaller number to sync */
2691                                 continue;
2692                         }
2693                         /* store current minimal group */
2694                         fglog = nlog;
2695                         group = nlog->group;
2696                 }
2697                 spin_unlock(&filter->fo_llog_list_lock);
2698
2699                 if (fglog == NULL)
2700                         break;
2701
2702                 worked = fglog->group;
2703                 if (fglog->exp && (dexp == fglog->exp || dexp == NULL)) {
2704                         ctxt = llog_get_context_from_llogs(fglog->llogs,
2705                                                 LLOG_MDS_OST_REPL_CTXT);
2706                         LASSERT(ctxt != NULL);
2707                         llog_sync(ctxt, fglog->exp);
2708                 }
2709         } while (fglog != NULL);
2710 }
2711
2712 /* also incredibly similar to mds_disconnect */
2713 static int filter_disconnect(struct obd_export *exp)
2714 {
2715         struct obd_device *obd = exp->exp_obd;
2716         int rc;
2717         ENTRY;
2718
2719         LASSERT(exp);
2720         class_export_get(exp);
2721
2722         if (!(exp->exp_flags & OBD_OPT_FORCE))
2723                 filter_grant_sanity_check(obd, __FUNCTION__);
2724         filter_grant_discard(exp);
2725
2726         /* Disconnect early so that clients can't keep using export */
2727         rc = class_disconnect(exp);
2728         if (exp->exp_obd->obd_namespace != NULL)
2729                 ldlm_cancel_locks_for_export(exp);
2730
2731         fsfilt_sync(obd, obd->u.obt.obt_sb);
2732
2733         /* flush any remaining cancel messages out to the target */
2734         filter_sync_llogs(obd, exp);
2735         class_export_put(exp);
2736         RETURN(rc);
2737 }
2738
2739 /* reverse import is changed, sync all cancels */
2740 static void filter_revimp_update(struct obd_export *exp)
2741 {
2742         ENTRY;
2743
2744         LASSERT(exp);
2745         class_export_get(exp);
2746
2747         /* flush any remaining cancel messages out to the target */
2748         filter_sync_llogs(exp->exp_obd, exp);
2749         class_export_put(exp);
2750         EXIT;
2751 }
2752
2753 static int filter_ping(struct obd_export *exp)
2754 {
2755         filter_fmd_expire(exp);
2756
2757         return 0;
2758 }
2759
2760 struct dentry *__filter_oa2dentry(struct obd_device *obd, struct obdo *oa,
2761                                   const char *what, int quiet)
2762 {
2763         struct dentry *dchild = NULL;
2764         obd_gr group = 0;
2765
2766         if (oa->o_valid & OBD_MD_FLGROUP)
2767                 group = oa->o_gr;
2768
2769         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
2770
2771         if (IS_ERR(dchild)) {
2772                 CERROR("%s error looking up object: "LPU64"\n",
2773                        what, oa->o_id);
2774                 RETURN(dchild);
2775         }
2776
2777         if (dchild->d_inode == NULL) {
2778                 if (!quiet)
2779                         CERROR("%s: %s on non-existent object: "LPU64"\n",
2780                                obd->obd_name, what, oa->o_id);
2781                 f_dput(dchild);
2782                 RETURN(ERR_PTR(-ENOENT));
2783         }
2784
2785         return dchild;
2786 }
2787
2788 static int filter_getattr(struct obd_export *exp, struct obd_info *oinfo)
2789 {
2790         struct dentry *dentry = NULL;
2791         struct obd_device *obd;
2792         int rc = 0;
2793         ENTRY;
2794
2795         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
2796                               oinfo_capa(oinfo), CAPA_OPC_META_READ);
2797         if (rc)
2798                 RETURN(rc);
2799
2800         obd = class_exp2obd(exp);
2801         if (obd == NULL) {
2802                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
2803                 RETURN(-EINVAL);
2804         }
2805
2806         dentry = filter_oa2dentry(obd, oinfo->oi_oa);
2807         if (IS_ERR(dentry))
2808                 RETURN(PTR_ERR(dentry));
2809
2810         /* Limit the valid bits in the return data to what we actually use */
2811         oinfo->oi_oa->o_valid = OBD_MD_FLID;
2812         obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
2813
2814         f_dput(dentry);
2815         RETURN(rc);
2816 }
2817
2818 /* this should be enabled/disabled in condition to enabled/disabled large
2819  * inodes (fast EAs) in backing store FS. */
2820 int filter_update_fidea(struct obd_export *exp, struct inode *inode,
2821                         void *handle, struct obdo *oa)
2822 {
2823         struct obd_device *obd = exp->exp_obd;
2824         int rc = 0;
2825         ENTRY;
2826
2827         if (oa->o_valid & OBD_MD_FLFID) {
2828                 struct filter_fid ff;
2829
2830                 if (!(oa->o_valid & OBD_MD_FLGROUP))
2831                         oa->o_gr = 0;
2832                 /* packing fid and converting it to LE for storing into EA.
2833                  * Here ->o_stripe_idx should be filled by LOV and rest of
2834                  * fields - by client. */
2835                 ff.ff_fid.id = cpu_to_le64(oa->o_fid);
2836                 ff.ff_fid.f_type = cpu_to_le32(oa->o_stripe_idx);
2837                 ff.ff_fid.generation = cpu_to_le32(oa->o_generation);
2838                 ff.ff_objid = cpu_to_le64(oa->o_id);
2839                 ff.ff_group = cpu_to_le64(oa->o_gr);
2840
2841                 CDEBUG(D_INODE, "storing filter fid EA ("LPU64"/%u/%u"
2842                        LPU64"/"LPU64")\n", oa->o_fid, oa->o_stripe_idx,
2843                        oa->o_generation, oa->o_id, oa->o_gr);
2844
2845                 rc = fsfilt_set_md(obd, inode, handle, &ff, sizeof(ff), "fid");
2846                 if (rc)
2847                         CERROR("store fid in object failed! rc: %d\n", rc);
2848         } else {
2849                 CDEBUG(D_HA, "OSS object without fid info!\n");
2850         }
2851
2852         RETURN(rc);
2853 }
2854
2855 /* this is called from filter_truncate() until we have filter_punch() */
2856 int filter_setattr_internal(struct obd_export *exp, struct dentry *dentry,
2857                             struct obdo *oa, struct obd_trans_info *oti)
2858 {
2859         unsigned int orig_ids[MAXQUOTAS] = {0, 0};
2860         struct llog_cookie *fcc = NULL;
2861         struct filter_obd *filter;
2862         int rc, err, locked = 0, sync = 0;
2863         unsigned int ia_valid;
2864         struct inode *inode;
2865         struct iattr iattr;
2866         void *handle;
2867         ENTRY;
2868
2869         LASSERT(dentry != NULL);
2870         LASSERT(!IS_ERR(dentry));
2871
2872         inode = dentry->d_inode;
2873         LASSERT(inode != NULL);
2874
2875         filter = &exp->exp_obd->u.filter;
2876         iattr_from_obdo(&iattr, oa, oa->o_valid);
2877         ia_valid = iattr.ia_valid;
2878
2879         if (oa->o_valid & OBD_MD_FLCOOKIE) {
2880                 OBD_ALLOC(fcc, sizeof(*fcc));
2881                 if (fcc != NULL)
2882                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
2883         }
2884
2885         if (ia_valid & ATTR_SIZE || ia_valid & (ATTR_UID | ATTR_GID)) {
2886                 DQUOT_INIT(inode);
2887                 LOCK_INODE_MUTEX(inode);
2888                 locked = 1;
2889         }
2890
2891         /* If the inode still has SUID+SGID bits set (see filter_precreate())
2892          * then we will accept the UID+GID sent by the client during write for
2893          * initializing the ownership of this inode.  We only allow this to
2894          * happen once so clear these bits in setattr. In 2.6 kernels it is
2895          * possible to get ATTR_UID and ATTR_GID separately, so we only clear
2896          * the flags that are actually being set. */
2897         if (ia_valid & (ATTR_UID | ATTR_GID)) {
2898                 CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n",
2899                        (unsigned long)oa->o_uid, (unsigned long)oa->o_gid);
2900
2901                 if ((inode->i_mode & S_ISUID) && (ia_valid & ATTR_UID)) {
2902                         if (!(ia_valid & ATTR_MODE)) {
2903                                 iattr.ia_mode = inode->i_mode;
2904                                 iattr.ia_valid |= ATTR_MODE;
2905                         }
2906                         iattr.ia_mode &= ~S_ISUID;
2907                 }
2908                 if ((inode->i_mode & S_ISGID) && (ia_valid & ATTR_GID)) {
2909                         if (!(iattr.ia_valid & ATTR_MODE)) {
2910                                 iattr.ia_mode = inode->i_mode;
2911                                 iattr.ia_valid |= ATTR_MODE;
2912                         }
2913                         iattr.ia_mode &= ~S_ISGID;
2914                 }
2915
2916                 orig_ids[USRQUOTA] = inode->i_uid;
2917                 orig_ids[GRPQUOTA] = inode->i_gid;
2918                 handle = fsfilt_start_log(exp->exp_obd, inode,
2919                                           FSFILT_OP_SETATTR, oti, 1);
2920                 if (IS_ERR(handle))
2921                         GOTO(out_unlock, rc = PTR_ERR(handle));
2922
2923                 /* update inode EA only once when inode is suid bit marked. As
2924                  * on 2.6.x UID and GID may be set separately, we check here
2925                  * only one of them to avoid double setting. */
2926                 if (inode->i_mode & S_ISUID)
2927                         filter_update_fidea(exp, inode, handle, oa);
2928         } else {
2929                 handle = fsfilt_start(exp->exp_obd, inode,
2930                                       FSFILT_OP_SETATTR, oti);
2931                 if (IS_ERR(handle))
2932                         GOTO(out_unlock, rc = PTR_ERR(handle));
2933         }
2934         if (oa->o_valid & OBD_MD_FLFLAGS) {
2935                 rc = fsfilt_iocontrol(exp->exp_obd, inode, NULL,
2936                                       EXT3_IOC_SETFLAGS, (long)&oa->o_flags);
2937         } else {
2938                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
2939                 if (fcc != NULL)
2940                         /* set cancel cookie callback function */
2941                         sync = fsfilt_add_journal_cb(exp->exp_obd, 0, handle,
2942                                                      filter_cancel_cookies_cb,
2943                                                      fcc);
2944         }
2945
2946         if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_CREDITS))
2947                 fsfilt_extend(exp->exp_obd, inode, 0, handle);
2948
2949         /* The truncate might have used up our transaction credits.  Make
2950          * sure we have one left for the last_rcvd update. */
2951         err = fsfilt_extend(exp->exp_obd, inode, 1, handle);
2952
2953         rc = filter_finish_transno(exp, oti, rc, sync);
2954         if (sync) {
2955                 filter_cancel_cookies_cb(exp->exp_obd, 0, fcc, rc);
2956                 fcc = NULL;
2957         }
2958
2959         err = fsfilt_commit(exp->exp_obd, inode, handle, 0);
2960         if (err) {
2961                 CERROR("error on commit, err = %d\n", err);
2962                 if (!rc)
2963                         rc = err;
2964         } else {
2965                 fcc = NULL;
2966         }
2967
2968         if (locked) {
2969                 /* Let's flush truncated page on disk immediately, then we can
2970                  * avoid need to search for page aliases before directio writes
2971                  * and this sort of stuff at expense of somewhat slower
2972                  * truncates not on a page boundary. I believe this is the only
2973                  * place in filter code that can lead to pages getting to
2974                  * pagecache so far. */
2975                 filter_clear_truncated_page(inode);
2976                 UNLOCK_INODE_MUTEX(inode);
2977                 locked = 0;
2978         }
2979
2980         EXIT;
2981 out_unlock:
2982         if (locked)
2983                 UNLOCK_INODE_MUTEX(inode);
2984
2985         if (fcc)
2986                 OBD_FREE(fcc, sizeof(*fcc));
2987
2988         /* trigger quota release */
2989         if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
2990                 unsigned int cur_ids[MAXQUOTAS] = {oa->o_uid, oa->o_gid};
2991                 int rc2 = lquota_adjust(filter_quota_interface_ref,
2992                                         exp->exp_obd, cur_ids,
2993                                         orig_ids, rc, FSFILT_OP_SETATTR);
2994                 CDEBUG(rc2 ? D_ERROR : D_QUOTA,
2995                        "filter adjust qunit. (rc:%d)\n", rc2);
2996         }
2997         return rc;
2998 }
2999
3000 /* this is called from filter_truncate() until we have filter_punch() */
3001 int filter_setattr(struct obd_export *exp, struct obd_info *oinfo,
3002                    struct obd_trans_info *oti)
3003 {
3004         struct ldlm_res_id res_id = { .name = { oinfo->oi_oa->o_id, 0,
3005                                                 oinfo->oi_oa->o_gr, 0 } };
3006         struct filter_mod_data *fmd;
3007         struct lvfs_run_ctxt saved;
3008         struct filter_obd *filter;
3009         struct ldlm_resource *res;
3010         struct dentry *dentry;
3011         int rc;
3012         ENTRY;
3013
3014         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3015                               oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
3016         if (rc)
3017                 RETURN(rc);
3018
3019         dentry = __filter_oa2dentry(exp->exp_obd, oinfo->oi_oa,
3020                                     __FUNCTION__, 1);
3021         if (IS_ERR(dentry))
3022                 RETURN(PTR_ERR(dentry));
3023
3024         filter = &exp->exp_obd->u.filter;
3025         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3026         lock_kernel();
3027
3028         if (oinfo->oi_oa->o_valid &
3029             (OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME)) {
3030                 fmd = filter_fmd_get(exp,oinfo->oi_oa->o_id,oinfo->oi_oa->o_gr);
3031                 if (fmd && fmd->fmd_mactime_xid < oti->oti_xid)
3032                         fmd->fmd_mactime_xid = oti->oti_xid;
3033                 filter_fmd_put(exp, fmd);
3034         }
3035
3036         /* setting objects attributes (including owner/group) */
3037         rc = filter_setattr_internal(exp, dentry, oinfo->oi_oa, oti);
3038         if (rc)
3039                 GOTO(out_unlock, rc);
3040
3041         res = ldlm_resource_get(exp->exp_obd->obd_namespace, NULL,
3042                                 &res_id, LDLM_EXTENT, 0);
3043
3044         if (res != NULL) {
3045                 rc = ldlm_res_lvbo_update(res, NULL, 0, 0);
3046                 ldlm_resource_putref(res);
3047         }
3048
3049         oinfo->oi_oa->o_valid = OBD_MD_FLID;
3050
3051         /* Quota release need uid/gid info */
3052         obdo_from_inode(oinfo->oi_oa, dentry->d_inode,
3053                         FILTER_VALID_FLAGS | OBD_MD_FLUID | OBD_MD_FLGID);
3054
3055         EXIT;
3056 out_unlock:
3057         unlock_kernel();
3058         f_dput(dentry);
3059         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3060         return rc;
3061 }
3062
3063 /* XXX identical to osc_unpackmd */
3064 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
3065                            struct lov_mds_md *lmm, int lmm_bytes)
3066 {
3067         int lsm_size;
3068         ENTRY;
3069
3070         if (lmm != NULL) {
3071                 if (lmm_bytes < sizeof (*lmm)) {
3072                         CERROR("lov_mds_md too small: %d, need %d\n",
3073                                lmm_bytes, (int)sizeof(*lmm));
3074                         RETURN(-EINVAL);
3075                 }
3076                 /* XXX LOV_MAGIC etc check? */
3077
3078                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
3079                         CERROR("lov_mds_md: zero lmm_object_id\n");
3080                         RETURN(-EINVAL);
3081                 }
3082         }
3083
3084         lsm_size = lov_stripe_md_size(1);
3085         if (lsmp == NULL)
3086                 RETURN(lsm_size);
3087
3088         if (*lsmp != NULL && lmm == NULL) {
3089                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3090                 OBD_FREE(*lsmp, lsm_size);
3091                 *lsmp = NULL;
3092                 RETURN(0);
3093         }
3094
3095         if (*lsmp == NULL) {
3096                 OBD_ALLOC(*lsmp, lsm_size);
3097                 if (*lsmp == NULL)
3098                         RETURN(-ENOMEM);
3099
3100                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
3101                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
3102                         OBD_FREE(*lsmp, lsm_size);
3103                         RETURN(-ENOMEM);
3104                 }
3105                 loi_init((*lsmp)->lsm_oinfo[0]);
3106         }
3107
3108         if (lmm != NULL) {
3109                 /* XXX zero *lsmp? */
3110                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
3111                 LASSERT((*lsmp)->lsm_object_id);
3112         }
3113
3114         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
3115
3116         RETURN(lsm_size);
3117 }
3118
3119 /* caller must hold fo_create_locks[oa->o_gr] */
3120 static int filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
3121                                       struct filter_obd *filter)
3122 {
3123         struct obdo doa; /* XXX obdo on stack */
3124         obd_id last, id;
3125         int rc;
3126         ENTRY;
3127
3128         LASSERT(oa);
3129         LASSERT(oa->o_gr != 0);
3130         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3131         LASSERT(down_trylock(&filter->fo_create_locks[oa->o_gr]) != 0);
3132
3133         memset(&doa, 0, sizeof(doa));
3134
3135         doa.o_valid |= OBD_MD_FLGROUP;
3136         doa.o_gr = oa->o_gr;
3137         doa.o_mode = S_IFREG;
3138
3139         if (!test_bit(doa.o_gr, &filter->fo_destroys_in_progress)) {
3140                 CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3141                        exp->exp_obd->obd_name, doa.o_gr);
3142                 RETURN(0);
3143         }
3144
3145         last = filter_last_id(filter, doa.o_gr);
3146
3147         CWARN("%s: deleting orphan objects from "LPU64" to "LPU64"\n",
3148                exp->exp_obd->obd_name, oa->o_id + 1, last);
3149
3150         for (id = last; id > oa->o_id; id--) {
3151                 doa.o_id = id;
3152                 rc = filter_destroy(exp, &doa, NULL, NULL, NULL);
3153                 if (rc && rc != -ENOENT) /* this is pretty fatal... */
3154                         CEMERG("error destroying precreate objid "LPU64": %d\n",
3155                                id, rc);
3156                 filter_set_last_id(filter, id - 1, doa.o_gr);
3157                 /* update last_id on disk periodically so that if we restart
3158                  * we don't need to re-scan all of the just-deleted objects. */
3159                 if ((id & 511) == 0)
3160                         filter_update_last_objid(exp->exp_obd, doa.o_gr, 0);
3161         }
3162
3163         CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
3164                exp->exp_obd->obd_name, doa.o_gr, oa->o_id);
3165
3166         rc = filter_update_last_objid(exp->exp_obd, doa.o_gr, 1);
3167         clear_bit(doa.o_gr, &filter->fo_destroys_in_progress);
3168
3169         RETURN(rc);
3170 }
3171
3172 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3173                             obd_gr group, int *num);
3174 /* returns a negative error or a nonnegative number of files to create */
3175 static int filter_handle_precreate(struct obd_export *exp, struct obdo *oa,
3176                                    obd_gr group, struct obd_trans_info *oti)
3177 {
3178         struct obd_device *obd = exp->exp_obd;
3179         struct filter_obd *filter = &obd->u.filter;
3180         int diff, rc;
3181         ENTRY;
3182
3183         /* delete orphans request */
3184         if ((oa->o_valid & OBD_MD_FLFLAGS) && (oa->o_flags & OBD_FL_DELORPHAN)){
3185                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3186                         CERROR("%s: dropping old orphan cleanup request\n",
3187                                obd->obd_name);
3188                         RETURN(0);
3189                 }
3190                 /* This causes inflight precreates to abort and drop lock */
3191                 set_bit(group, &filter->fo_destroys_in_progress);
3192                 down(&filter->fo_create_locks[group]);
3193                 if (!test_bit(group, &filter->fo_destroys_in_progress)) {
3194                         CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
3195                                exp->exp_obd->obd_name, group);
3196                         up(&filter->fo_create_locks[group]);
3197                         RETURN(0);
3198                 }
3199                 diff = oa->o_id - filter_last_id(filter, group);
3200                 CDEBUG(D_HA, "filter_last_id() = "LPU64" -> diff = %d\n",
3201                        filter_last_id(filter, group), diff);
3202
3203                 if (-diff > OST_MAX_PRECREATE) {
3204                         CERROR("%s: ignoring bogus orphan destroy request: "
3205                                "obdid "LPU64" last_id "LPU64"\n", obd->obd_name,
3206                                oa->o_id, filter_last_id(filter, group));
3207                         /* FIXME: should reset precreate_next_id on MDS */
3208                         GOTO(out, rc = -EINVAL);
3209                 }
3210                 if (diff < 0) {
3211                         rc = filter_destroy_precreated(exp, oa, filter);
3212                         if (rc)
3213                                 CERROR("%s: unable to write lastobjid, but "
3214                                        "orphans were deleted\n", obd->obd_name);
3215                         GOTO(out, rc);
3216                 } else {
3217                         /* XXX: Used by MDS for the first time! */
3218                         clear_bit(group, &filter->fo_destroys_in_progress);
3219                 }
3220         } else {
3221                 down(&filter->fo_create_locks[group]);
3222                 if (oti->oti_conn_cnt < exp->exp_conn_cnt) {
3223                         CERROR("%s: dropping old precreate request\n",
3224                                obd->obd_name);
3225                         GOTO(out, rc = 0);
3226                 }
3227                 /* only precreate if group == 0 and o_id is specfied */
3228                 if (group < FILTER_GROUP_MDS0 || oa->o_id == 0)
3229                         diff = 1;
3230                 else
3231                         diff = oa->o_id - filter_last_id(filter, group);
3232                 CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
3233                        filter_last_id(filter, group), diff);
3234
3235                 LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
3236                          oa->o_id, filter_last_id(filter, group), diff);
3237         }
3238
3239         if (diff > 0) {
3240                 oa->o_id = filter_last_id(&obd->u.filter, group);
3241                 rc = filter_precreate(obd, oa, group, &diff);
3242                 oa->o_id = filter_last_id(&obd->u.filter, group);
3243                 oa->o_gr = group;
3244                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3245                 GOTO(out, rc);
3246         }
3247         /* else diff == 0 */
3248         GOTO(out, rc = 0);
3249 out:
3250         up(&filter->fo_create_locks[group]);
3251         return rc;
3252 }
3253
3254 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3255                          __u64 max_age)
3256 {
3257         struct filter_obd *filter = &obd->u.filter;
3258         int blockbits = obd->u.obt.obt_sb->s_blocksize_bits;
3259         int rc;
3260         ENTRY;
3261
3262         /* at least try to account for cached pages.  its still racey and
3263          * might be under-reporting if clients haven't announced their
3264          * caches with brw recently */
3265         spin_lock(&obd->obd_osfs_lock);
3266         rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
3267         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
3268         spin_unlock(&obd->obd_osfs_lock);
3269
3270         CDEBUG(D_SUPER | D_CACHE, "blocks cached "LPU64" granted "LPU64
3271                " pending "LPU64" free "LPU64" avail "LPU64"\n",
3272                filter->fo_tot_dirty, filter->fo_tot_granted,
3273                filter->fo_tot_pending,
3274                osfs->os_bfree << blockbits, osfs->os_bavail << blockbits);
3275
3276         filter_grant_sanity_check(obd, __FUNCTION__);
3277
3278         osfs->os_bavail -= min(osfs->os_bavail, GRANT_FOR_LLOG(obd) +
3279                                ((filter->fo_tot_dirty + filter->fo_tot_pending +
3280                                  osfs->os_bsize - 1) >> blockbits));
3281
3282         /* set EROFS to state field if FS is mounted as RDONLY. The goal is to
3283          * stop creating files on MDS if OST is not good shape to create
3284          * objects.*/
3285         osfs->os_state = (filter->fo_obt.obt_sb->s_flags & MS_RDONLY) ?
3286                 EROFS : 0;
3287         RETURN(rc);
3288 }
3289
3290 static int filter_use_existing_obj(struct obd_device *obd,
3291                                    struct dentry *dchild, void **handle,
3292                                    int *cleanup_phase)
3293 {
3294         struct inode *inode = dchild->d_inode;
3295         struct iattr iattr;
3296         int rc;
3297
3298         if ((inode->i_mode & (S_ISUID | S_ISGID)) == (S_ISUID|S_ISGID))
3299                 return 0;
3300
3301         *handle = fsfilt_start_log(obd, inode, FSFILT_OP_SETATTR, NULL, 1);
3302         if (IS_ERR(*handle))
3303                 return PTR_ERR(*handle);
3304
3305         iattr.ia_valid = ATTR_MODE;
3306         iattr.ia_mode = S_ISUID | S_ISGID |0666;
3307         rc = fsfilt_setattr(obd, dchild, *handle, &iattr, 1);
3308         if (rc == 0)
3309                 *cleanup_phase = 3;
3310
3311         return rc;
3312 }
3313
3314
3315 /* We rely on the fact that only one thread will be creating files in a given
3316  * group at a time, which is why we don't need an atomic filter_get_new_id.
3317  * Even if we had that atomic function, the following race would exist:
3318  *
3319  * thread 1: gets id x from filter_next_id
3320  * thread 2: gets id (x + 1) from filter_next_id
3321  * thread 2: creates object (x + 1)
3322  * thread 1: tries to create object x, gets -ENOSPC
3323  *
3324  * Caller must hold fo_create_locks[group]
3325  */
3326 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
3327                             obd_gr group, int *num)
3328 {
3329         struct dentry *dchild = NULL, *dparent = NULL;
3330         struct filter_obd *filter;
3331         struct obd_statfs *osfs;
3332         int err = 0, rc = 0, recreate_obj = 0, i;
3333         unsigned long enough_time = jiffies + min(obd_timeout * HZ / 4, 10U*HZ);
3334         obd_id next_id;
3335         void *handle = NULL;
3336         ENTRY;
3337
3338         filter = &obd->u.filter;
3339
3340         LASSERT(down_trylock(&filter->fo_create_locks[group]) != 0);
3341
3342         OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_PRECREATE, obd_timeout / 2);
3343
3344         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3345             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3346                 recreate_obj = 1;
3347         } else {
3348                 OBD_ALLOC(osfs, sizeof(*osfs));
3349                 if (osfs == NULL)
3350                         RETURN(-ENOMEM);
3351                 rc = filter_statfs(obd, osfs, cfs_time_current_64() - HZ);
3352                 if (rc == 0 && osfs->os_bavail < (osfs->os_blocks >> 10)) {
3353                         CDEBUG(D_RPCTRACE,"%s: not enough space for create "
3354                                LPU64"\n", obd->obd_name, osfs->os_bavail <<
3355                                filter->fo_vfsmnt->mnt_sb->s_blocksize_bits);
3356                         *num = 0;
3357                         rc = -ENOSPC;
3358                 }
3359                 OBD_FREE(osfs, sizeof(*osfs));
3360                 if (rc)
3361                         RETURN(rc);
3362         }
3363
3364         CDEBUG(D_RPCTRACE, "%s: precreating %d objects in group "LPU64
3365                " at "LPU64"\n", obd->obd_name, *num, group, oa->o_id);
3366
3367         for (i = 0; i < *num && err == 0; i++) {
3368                 int cleanup_phase = 0;
3369
3370                 if (test_bit(group, &filter->fo_destroys_in_progress)) {
3371                         CWARN("%s: create aborted by destroy\n",
3372                               obd->obd_name);
3373                         rc = -EAGAIN;
3374                         break;
3375                 }
3376
3377                 if (recreate_obj) {
3378                         __u64 last_id;
3379                         next_id = oa->o_id;
3380                         last_id = filter_last_id(filter, group);
3381                         if (next_id > last_id) {
3382                                 CERROR("Error: Trying to recreate obj greater"
3383                                        "than last id "LPD64" > "LPD64"\n",
3384                                        next_id, last_id);
3385                                 GOTO(cleanup, rc = -EINVAL);
3386                         }
3387                 } else
3388                         next_id = filter_last_id(filter, group) + 1;
3389
3390                 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
3391
3392                 dparent = filter_parent_lock(obd, group, next_id);
3393                 if (IS_ERR(dparent))
3394                         GOTO(cleanup, rc = PTR_ERR(dparent));
3395                 cleanup_phase = 1; /* filter_parent_unlock(dparent) */
3396
3397                 dchild = filter_fid2dentry(obd, dparent, group, next_id);
3398                 if (IS_ERR(dchild))
3399                         GOTO(cleanup, rc = PTR_ERR(dchild));
3400                 cleanup_phase = 2;  /* f_dput(dchild) */
3401
3402                 if (dchild->d_inode != NULL) {
3403                         /* This would only happen if lastobjid was bad on disk*/
3404                         /* Could also happen if recreating missing obj but it
3405                          * already exists. */
3406                         if (recreate_obj) {
3407                                 CERROR("%s: recreating existing object %.*s?\n",
3408                                        obd->obd_name, dchild->d_name.len,
3409                                        dchild->d_name.name);
3410                         } else {
3411                                 /* Use these existing objects if they are
3412                                  * zero length. */
3413                                 if (dchild->d_inode->i_size == 0) {
3414                                         rc = filter_use_existing_obj(obd,dchild,
3415                                                       &handle, &cleanup_phase);
3416                                         if (rc == 0)
3417                                                 goto set_last_id;
3418                                         else
3419                                                 GOTO(cleanup, rc);
3420                                 }
3421
3422                                 CERROR("%s: Serious error: objid %.*s already "
3423                                        "exists; is this filesystem corrupt?\n",
3424                                        obd->obd_name, dchild->d_name.len,
3425                                        dchild->d_name.name);
3426                                 LBUG();
3427                         }
3428                         GOTO(cleanup, rc = -EEXIST);
3429                 }
3430
3431                 handle = fsfilt_start_log(obd, dparent->d_inode,
3432                                           FSFILT_OP_CREATE, NULL, 1);
3433                 if (IS_ERR(handle))
3434                         GOTO(cleanup, rc = PTR_ERR(handle));
3435                 cleanup_phase = 3;
3436
3437                 /* We mark object SUID+SGID to flag it for accepting UID+GID
3438                  * from client on first write.  Currently the permission bits
3439                  * on the OST are never used, so this is OK. */
3440                 rc = ll_vfs_create(dparent->d_inode, dchild,
3441                                    S_IFREG |  S_ISUID | S_ISGID | 0666, NULL);
3442                 if (rc) {
3443                         CERROR("create failed rc = %d\n", rc);
3444                         GOTO(cleanup, rc);
3445                 }
3446
3447 set_last_id:
3448                 if (!recreate_obj) {
3449                         filter_set_last_id(filter, next_id, group);
3450                         err = filter_update_last_objid(obd, group, 0);
3451                         if (err)
3452                                 CERROR("unable to write lastobjid "
3453                                        "but file created\n");
3454                 }
3455
3456         cleanup:
3457                 switch(cleanup_phase) {
3458                 case 3:
3459                         err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3460                         if (err) {
3461                                 CERROR("error on commit, err = %d\n", err);
3462                                 if (!rc)
3463                                         rc = err;
3464                         }
3465                 case 2:
3466                         f_dput(dchild);
3467                 case 1:
3468                         filter_parent_unlock(dparent);
3469                 case 0:
3470                         break;
3471                 }
3472
3473                 if (rc)
3474                         break;
3475                 if (time_after(jiffies, enough_time)) {
3476                         CDEBUG(D_RPCTRACE,
3477                                "%s: precreate slow - want %d got %d \n",
3478                                obd->obd_name, *num, i);
3479                         break;
3480                 }
3481         }
3482         *num = i;
3483
3484         CDEBUG(D_RPCTRACE,
3485                "%s: created %d objects for group "LPU64": "LPU64" rc %d\n",
3486                obd->obd_name, i, group, filter->fo_last_objids[group], rc);
3487
3488         RETURN(rc);
3489 }
3490
3491 static int filter_create(struct obd_export *exp, struct obdo *oa,
3492                          struct lov_stripe_md **ea, struct obd_trans_info *oti)
3493 {
3494         struct filter_export_data *fed;
3495         struct obd_device *obd = NULL;
3496         struct filter_obd *filter;
3497         struct lvfs_run_ctxt saved;
3498         struct lov_stripe_md *lsm = NULL;
3499         int rc = 0, diff, group = oa->o_gr;
3500         ENTRY;
3501
3502         if (!(oa->o_valid & OBD_MD_FLGROUP) || group == 0) {
3503                 CERROR("!!! nid %s sent invalid object group %d\n",
3504                         obd_export_nid2str(exp), group);
3505                 RETURN(-EINVAL);
3506         }
3507
3508         obd = exp->exp_obd;
3509         fed = &exp->exp_filter_data;
3510         filter = &obd->u.filter;
3511
3512         if (fed->fed_group != group) {
3513                 CERROR("!!! this export (nid %s) used object group %d "
3514                         "earlier; now it's trying to use group %d!  This could "
3515                         "be a bug in the MDS.  Tell CFS.\n",
3516                         obd_export_nid2str(exp), fed->fed_group, group);
3517                 RETURN(-ENOTUNIQ);
3518         }
3519
3520         CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
3521                oa->o_gr, oa->o_id);
3522         if (ea != NULL) {
3523                 lsm = *ea;
3524                 if (lsm == NULL) {
3525                         rc = obd_alloc_memmd(exp, &lsm);
3526                         if (rc < 0)
3527                                 RETURN(rc);
3528                 }
3529         }
3530
3531         obd = exp->exp_obd;
3532         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3533
3534         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
3535             (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
3536                 if (oa->o_id > filter_last_id(filter, oa->o_gr)) {
3537                         CERROR("recreate objid "LPU64" > last id "LPU64"\n",
3538                                oa->o_id, filter_last_id(filter,
3539                                                         oa->o_gr));
3540                         rc = -EINVAL;
3541                 } else {
3542                         diff = 1;
3543                         down(&filter->fo_create_locks[oa->o_gr]);
3544                         rc = filter_precreate(obd, oa, oa->o_gr, &diff);
3545                         up(&filter->fo_create_locks[oa->o_gr]);
3546                 }
3547         } else {
3548                 rc = filter_handle_precreate(exp, oa, oa->o_gr, oti);
3549         }
3550
3551         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3552         if (rc && ea != NULL && *ea != lsm) {
3553                 obd_free_memmd(exp, &lsm);
3554         } else if (rc == 0 && ea != NULL) {
3555                 /* XXX LOV STACKING: the lsm that is passed to us from
3556                  * LOV does not have valid lsm_oinfo data structs, so
3557                  * don't go touching that.  This needs to be fixed in a
3558                  * big way. */
3559                 lsm->lsm_object_id = oa->o_id;
3560                 *ea = lsm;
3561         }
3562
3563         RETURN(rc);
3564 }
3565
3566 int filter_destroy(struct obd_export *exp, struct obdo *oa,
3567                    struct lov_stripe_md *md, struct obd_trans_info *oti,
3568                    struct obd_export *md_exp)
3569 {
3570         unsigned int qcids[MAXQUOTAS] = {0, 0};
3571         struct obd_device *obd;
3572         struct filter_obd *filter;
3573         struct dentry *dchild = NULL, *dparent = NULL;
3574         struct lvfs_run_ctxt saved;
3575         void *handle = NULL;
3576         struct llog_cookie *fcc = NULL;
3577         int rc, rc2, cleanup_phase = 0, sync = 0;
3578         struct iattr iattr;
3579         ENTRY;
3580
3581         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
3582
3583         obd = exp->exp_obd;
3584         filter = &obd->u.filter;
3585
3586         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3587         cleanup_phase = 1;
3588
3589         dchild = filter_fid2dentry(obd, NULL, oa->o_gr, oa->o_id);
3590         if (IS_ERR(dchild))
3591                 GOTO(cleanup, rc = PTR_ERR(dchild));
3592         cleanup_phase = 2;
3593
3594         if (dchild->d_inode == NULL) {
3595                 CDEBUG(D_INODE, "destroying non-existent object "LPU64"\n",
3596                        oa->o_id);
3597                 /* If object already gone, cancel cookie right now */
3598                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
3599                         fcc = obdo_logcookie(oa);
3600                         llog_cancel(llog_get_context(obd, fcc->lgc_subsys + 1),
3601                                     NULL, 1, fcc, 0);
3602                         fcc = NULL; /* we didn't allocate fcc, don't free it */
3603                 }
3604                 GOTO(cleanup, rc = -ENOENT);
3605         }
3606
3607         filter_prepare_destroy(obd, oa->o_id, oa->o_gr);
3608
3609         /* Our MDC connection is established by the MDS to us */
3610         if (oa->o_valid & OBD_MD_FLCOOKIE) {
3611                 OBD_ALLOC(fcc, sizeof(*fcc));
3612                 if (fcc != NULL)
3613                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
3614         }
3615         DQUOT_INIT(dchild->d_inode);
3616
3617         /* we're gonna truncate it first in order to avoid possible deadlock:
3618          *      P1                      P2
3619          * open trasaction      open transaction
3620          * down(i_zombie)       down(i_zombie)
3621          *                      restart transaction
3622          * (see BUG 4180) -bzzz
3623          */
3624         LOCK_INODE_MUTEX(dchild->d_inode);
3625         handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
3626                                   NULL, 1);
3627         if (IS_ERR(handle)) {
3628                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3629                 GOTO(cleanup, rc = PTR_ERR(handle));
3630         }
3631
3632         iattr.ia_valid = ATTR_SIZE;
3633         iattr.ia_size = 0;
3634         rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
3635         rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
3636         UNLOCK_INODE_MUTEX(dchild->d_inode);
3637         if (rc)
3638                 GOTO(cleanup, rc);
3639         if (rc2)
3640                 GOTO(cleanup, rc = rc2);
3641
3642         /* We don't actually need to lock the parent until we are unlinking
3643          * here, and not while truncating above.  That avoids holding the
3644          * parent lock for a long time during truncate, which can block other
3645          * threads from doing anything to objects in that directory. bug 7171 */
3646         dparent = filter_parent_lock(obd, oa->o_gr, oa->o_id);
3647         if (IS_ERR(dparent))
3648                 GOTO(cleanup, rc = PTR_ERR(dparent));
3649         cleanup_phase = 3; /* filter_parent_unlock */
3650
3651         LOCK_INODE_MUTEX(dchild->d_inode);
3652         handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
3653         if (IS_ERR(handle)) {
3654                 UNLOCK_INODE_MUTEX(dchild->d_inode);
3655                 GOTO(cleanup, rc = PTR_ERR(handle));
3656         }
3657         cleanup_phase = 4; /* fsfilt_commit */
3658
3659         /* Quota release need uid/gid of inode */
3660         obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID|OBD_MD_FLGID);
3661
3662         filter_fmd_drop(exp, oa->o_id, oa->o_gr);
3663
3664         /* this drops dchild->d_inode->i_mutex unconditionally */
3665         rc = filter_destroy_internal(obd, oa->o_id, oa->o_gr, dparent, dchild);
3666
3667         EXIT;
3668 cleanup:
3669         switch(cleanup_phase) {
3670         case 4:
3671                 if (fcc != NULL)
3672                         sync = fsfilt_add_journal_cb(obd, 0, oti ?
3673                                                      oti->oti_handle : handle,
3674                                                      filter_cancel_cookies_cb,
3675                                                      fcc);
3676                 /* If add_journal_cb failed, then filter_finish_transno
3677                  * will commit the handle and we will do a sync
3678                  * on commit. then we call callback directly to free
3679                  * the fcc.
3680                  */
3681                 rc = filter_finish_transno(exp, oti, rc, sync);
3682                 if (sync) {
3683                         filter_cancel_cookies_cb(obd, 0, fcc, rc);
3684                         fcc = NULL;
3685                 }
3686                 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
3687                 if (rc2) {
3688                         CERROR("error on commit, err = %d\n", rc2);
3689                         if (!rc)
3690                                 rc = rc2;
3691                 } else {
3692                         fcc = NULL;
3693                 }
3694         case 3:
3695                 filter_parent_unlock(dparent);
3696         case 2:
3697                 f_dput(dchild);
3698                 if (fcc != NULL)
3699                         OBD_FREE(fcc, sizeof(*fcc));
3700         case 1:
3701                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3702                 break;
3703         default:
3704                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
3705                 LBUG();
3706         }
3707
3708         /* trigger quota release */
3709         qcids[USRQUOTA] = oa->o_uid;
3710         qcids[GRPQUOTA] = oa->o_gid;
3711         rc2 = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc,
3712                             FSFILT_OP_UNLINK);
3713
3714         if (rc2)
3715                 CDEBUG(D_QUOTA, "filter adjust qunit! (rc:%d)\n", rc2);
3716         return rc;
3717 }
3718
3719 /* NB start and end are used for punch, but not truncate */
3720 static int filter_truncate(struct obd_export *exp, struct obd_info *oinfo,
3721                            struct obd_trans_info *oti,
3722                            struct ptlrpc_request_set *rqset)
3723 {
3724         int rc;
3725         ENTRY;
3726
3727         if (oinfo->oi_policy.l_extent.end != OBD_OBJECT_EOF) {
3728                 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
3729                        oinfo->oi_policy.l_extent.end);
3730                 RETURN(-EFAULT);
3731         }
3732
3733         CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = "LPX64
3734                ", o_size = "LPD64"\n", oinfo->oi_oa->o_id,
3735                oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start);
3736
3737         rc = filter_auth_capa(exp, NULL, oinfo_mdsno(oinfo),
3738                               oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
3739         if (rc)
3740                 RETURN(rc);
3741
3742         oinfo->oi_oa->o_size = oinfo->oi_policy.l_extent.start;
3743         rc = filter_setattr(exp, oinfo, oti);
3744         RETURN(rc);
3745 }
3746
3747 static int filter_sync(struct obd_export *exp, struct obdo *oa,
3748                        struct lov_stripe_md *lsm, obd_off start, obd_off end,
3749                        void *capa)
3750 {
3751         struct lvfs_run_ctxt saved;
3752         struct filter_obd *filter;
3753         struct dentry *dentry;
3754         struct llog_ctxt *ctxt;
3755         int rc, rc2;
3756         ENTRY;
3757
3758         rc = filter_auth_capa(exp, NULL, obdo_mdsno(oa),
3759                               (struct lustre_capa *)capa, CAPA_OPC_OSS_WRITE);
3760         if (rc)
3761                 RETURN(rc);
3762
3763         filter = &exp->exp_obd->u.filter;
3764
3765         /* an objid of zero is taken to mean "sync whole filesystem" */
3766         if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
3767                 rc = fsfilt_sync(exp->exp_obd, filter->fo_obt.obt_sb);
3768                 /* flush any remaining cancel messages out to the target */
3769                 ctxt = llog_get_context(exp->exp_obd, LLOG_MDS_OST_REPL_CTXT);
3770                 llog_sync(ctxt, exp);
3771                 RETURN(rc);
3772         }
3773
3774         dentry = filter_oa2dentry(exp->exp_obd, oa);
3775         if (IS_ERR(dentry))
3776                 RETURN(PTR_ERR(dentry));
3777
3778         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3779
3780         LOCK_INODE_MUTEX(dentry->d_inode);
3781
3782         rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
3783         if (rc == 0) {
3784                 /* just any file to grab fsync method - "file" arg unused */
3785                 struct file *file = filter->fo_rcvd_filp;
3786
3787                 if (file->f_op && file->f_op->fsync)
3788                         rc = file->f_op->fsync(NULL, dentry, 1);
3789
3790                 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
3791                 if (!rc)
3792                         rc = rc2;
3793         }
3794         UNLOCK_INODE_MUTEX(dentry->d_inode);
3795
3796         oa->o_valid = OBD_MD_FLID;
3797         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
3798
3799         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
3800
3801         f_dput(dentry);
3802         RETURN(rc);
3803 }
3804
3805 static int filter_get_info(struct obd_export *exp, __u32 keylen,
3806                            void *key, __u32 *vallen, void *val)
3807 {
3808         struct obd_device *obd;
3809         ENTRY;
3810
3811         obd = class_exp2obd(exp);
3812         if (obd == NULL) {
3813                 CDEBUG(D_IOCTL, "invalid client export %p\n", exp);
3814                 RETURN(-EINVAL);
3815         }
3816
3817         if (keylen == strlen("blocksize") &&
3818             memcmp(key, "blocksize", keylen) == 0) {
3819                 __u32 *blocksize = val;
3820                 *vallen = sizeof(*blocksize);
3821                 *blocksize = obd->u.obt.obt_sb->s_blocksize;
3822                 RETURN(0);
3823         }
3824
3825         if (keylen == strlen("blocksize_bits") &&
3826             memcmp(key, "blocksize_bits", keylen) == 0) {
3827                 __u32 *blocksize_bits = val;
3828                 *vallen = sizeof(*blocksize_bits);
3829                 *blocksize_bits = obd->u.obt.obt_sb->s_blocksize_bits;
3830                 RETURN(0);
3831         }
3832
3833         if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
3834                 obd_id *last_id = val;
3835                 /* FIXME: object groups */
3836                 *last_id = filter_last_id(&obd->u.filter, 0);
3837                 RETURN(0);
3838         }
3839         CDEBUG(D_IOCTL, "invalid key\n");
3840         RETURN(-EINVAL);
3841 }
3842
3843 static int filter_set_info_async(struct obd_export *exp, __u32 keylen,
3844                                  void *key, __u32 vallen, void *val,
3845                                  struct ptlrpc_request_set *set)
3846 {
3847         struct obd_device *obd;
3848         struct obd_llogs *llog;
3849         struct llog_ctxt *ctxt;
3850         int rc = 0, group;
3851         ENTRY;
3852
3853         obd = exp->exp_obd;
3854         if (obd == NULL) {
3855                 CDEBUG(D_IOCTL, "invalid export %p\n", exp);
3856                 RETURN(-EINVAL);
3857         }
3858
3859         if (KEY_IS(KEY_CAPA_KEY)) {
3860                 rc = filter_update_capa_key(obd, (struct lustre_capa_key *)val);
3861                 if (rc)
3862                         CERROR("filter update capability key failed: %d\n", rc);
3863                 RETURN(rc);
3864         }
3865
3866         if (KEY_IS(KEY_REVIMP_UPD)) {
3867                 filter_revimp_update(exp);
3868                 RETURN(0);
3869         }
3870
3871         if (keylen < strlen(KEY_MDS_CONN) ||
3872             memcmp(key, KEY_MDS_CONN, keylen) != 0)
3873                 RETURN(-EINVAL);
3874
3875         LCONSOLE_WARN("%s: received MDS connection from %s\n", obd->obd_name,
3876                       obd_export_nid2str(exp));
3877         obd->u.filter.fo_mdc_conn.cookie = exp->exp_handle.h_cookie;
3878
3879         /* setup llog imports */
3880         LASSERT(val != NULL);
3881         group = (int)(*(__u32 *)val);
3882         LASSERT(group >= FILTER_GROUP_MDS0);
3883
3884         llog = filter_grab_llog_for_group(obd, group, exp);
3885         LASSERT(llog != NULL);
3886         ctxt = llog_get_context_from_llogs(llog, LLOG_MDS_OST_REPL_CTXT);
3887         LASSERTF(ctxt != NULL, "ctxt is not null\n"),
3888
3889         rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
3890
3891         lquota_setinfo(filter_quota_interface_ref, exp, obd);
3892
3893         RETURN(rc);
3894 }
3895
3896 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
3897                      int len, void *karg, void *uarg)
3898 {
3899         struct obd_device *obd = exp->exp_obd;
3900         struct obd_ioctl_data *data = karg;
3901         int rc = 0;
3902
3903         switch (cmd) {
3904         case OBD_IOC_ABORT_RECOVERY: {
3905                 CERROR("aborting recovery for device %s\n", obd->obd_name);
3906                 target_stop_recovery_thread(obd);
3907                 RETURN(0);
3908         }
3909
3910         case OBD_IOC_SYNC: {
3911                 CDEBUG(D_RPCTRACE, "syncing ost %s\n", obd->obd_name);
3912                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3913                 RETURN(rc);
3914         }
3915
3916         case OBD_IOC_SET_READONLY: {
3917                 void *handle;
3918                 struct super_block *sb = obd->u.obt.obt_sb;
3919                 struct inode *inode = sb->s_root->d_inode;
3920                 BDEVNAME_DECLARE_STORAGE(tmp);
3921                 CERROR("*** setting device %s read-only ***\n",
3922                        ll_bdevname(sb, tmp));
3923
3924                 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
3925                 if (!IS_ERR(handle))
3926                         rc = fsfilt_commit(obd, inode, handle, 1);
3927
3928                 CDEBUG(D_HA, "syncing ost %s\n", obd->obd_name);
3929                 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
3930
3931                 lvfs_set_rdonly(obd, obd->u.obt.obt_sb);
3932                 RETURN(0);
3933         }
3934
3935         case OBD_IOC_CATLOGLIST: {
3936                 rc = llog_catalog_list(obd, 1, data);
3937                 RETURN(rc);
3938         }
3939
3940         case OBD_IOC_LLOG_CANCEL:
3941         case OBD_IOC_LLOG_REMOVE:
3942         case OBD_IOC_LLOG_INFO:
3943         case OBD_IOC_LLOG_PRINT: {
3944                 /* FIXME to be finished */
3945                 RETURN(-EOPNOTSUPP);
3946 /*
3947                 struct llog_ctxt *ctxt = NULL;
3948
3949                 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3950                 rc = llog_ioctl(ctxt, cmd, data);
3951                 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_lvfs_ctxt, NULL);
3952
3953                 RETURN(rc);
3954 */
3955         }
3956
3957
3958         default:
3959                 RETURN(-EINVAL);
3960         }
3961         RETURN(0);
3962 }
3963
3964 static int filter_health_check(struct obd_device *obd)
3965 {
3966 #ifdef USE_HEALTH_CHECK_WRITE
3967         struct filter_obd *filter = &obd->u.filter;
3968 #endif
3969         int rc = 0;
3970
3971         /*
3972          * health_check to return 0 on healthy
3973          * and 1 on unhealthy.
3974          */
3975         if (obd->u.obt.obt_sb->s_flags & MS_RDONLY)
3976                 rc = 1;
3977
3978 #ifdef USE_HEALTH_CHECK_WRITE
3979         LASSERT(filter->fo_health_check_filp != NULL);
3980         rc |= !!lvfs_check_io_health(obd, filter->fo_health_check_filp);
3981 #endif
3982         return rc;
3983 }
3984
3985 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
3986                                              void *data)
3987 {
3988         return filter_fid2dentry(data, NULL, gr, id);
3989 }
3990
3991 static int filter_process_config(struct obd_device *obd, obd_count len,
3992                                  void *buf)
3993 {
3994         struct lustre_cfg *lcfg = buf;
3995         struct lprocfs_static_vars lvars;
3996         int rc = 0;
3997
3998         lprocfs_init_vars(filter, &lvars);
3999
4000         rc = class_process_proc_param(PARAM_OST, lvars.obd_vars, lcfg, obd);
4001         return rc;
4002 }
4003
4004 static struct lvfs_callback_ops filter_lvfs_ops = {
4005         l_fid2dentry:     filter_lvfs_fid2dentry,
4006 };
4007
4008 static struct obd_ops filter_obd_ops = {
4009         .o_owner          = THIS_MODULE,
4010         .o_get_info       = filter_get_info,
4011         .o_set_info_async = filter_set_info_async,
4012         .o_setup          = filter_setup,
4013         .o_precleanup     = filter_precleanup,
4014         .o_cleanup        = filter_cleanup,
4015         .o_connect        = filter_connect,
4016         .o_reconnect      = filter_reconnect,
4017         .o_disconnect     = filter_disconnect,
4018         .o_ping           = filter_ping,
4019         .o_init_export    = filter_init_export,
4020         .o_destroy_export = filter_destroy_export,
4021         .o_statfs         = filter_statfs,
4022         .o_getattr        = filter_getattr,
4023         .o_unpackmd       = filter_unpackmd,
4024         .o_create         = filter_create,
4025         .o_setattr        = filter_setattr,
4026         .o_destroy        = filter_destroy,
4027         .o_brw            = filter_brw,
4028         .o_punch          = filter_truncate,
4029         .o_sync           = filter_sync,
4030         .o_preprw         = filter_preprw,
4031         .o_commitrw       = filter_commitrw,
4032         .o_llog_init      = filter_llog_init,
4033         .o_llog_connect   = filter_llog_connect,
4034         .o_llog_finish    = filter_llog_finish,
4035         .o_iocontrol      = filter_iocontrol,
4036         .o_health_check   = filter_health_check,
4037         .o_process_config = filter_process_config,
4038 };
4039
4040 quota_interface_t *filter_quota_interface_ref;
4041 extern quota_interface_t filter_quota_interface;
4042
4043 static int __init obdfilter_init(void)
4044 {
4045         struct lprocfs_static_vars lvars;
4046         int rc;
4047
4048         lprocfs_init_vars(filter, &lvars);
4049
4050         request_module("lquota");
4051         OBD_ALLOC(obdfilter_created_scratchpad,
4052                   OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4053                   sizeof(*obdfilter_created_scratchpad));
4054         if (obdfilter_created_scratchpad == NULL)
4055                 return -ENOMEM;
4056
4057         ll_fmd_cachep = cfs_mem_cache_create("ll_fmd_cache",
4058                                              sizeof(struct filter_mod_data),
4059                                              0, 0);
4060         if (!ll_fmd_cachep)
4061                 GOTO(out, rc = -ENOMEM);
4062
4063         filter_quota_interface_ref = PORTAL_SYMBOL_GET(filter_quota_interface);
4064         init_obd_quota_ops(filter_quota_interface_ref, &filter_obd_ops);
4065
4066         rc = class_register_type(&filter_obd_ops, NULL, lvars.module_vars,
4067                                  LUSTRE_OST_NAME, NULL);
4068         if (rc) {
4069                 int err;
4070
4071                 err = cfs_mem_cache_destroy(ll_fmd_cachep);
4072                 LASSERTF(err == 0, "Cannot destroy ll_fmd_cachep: rc %d\n",err);
4073                 ll_fmd_cachep = NULL;
4074 out:
4075                 if (filter_quota_interface_ref)
4076                         PORTAL_SYMBOL_PUT(filter_quota_interface);
4077
4078                 OBD_FREE(obdfilter_created_scratchpad,
4079                          OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4080                          sizeof(*obdfilter_created_scratchpad));
4081         }
4082
4083         return rc;
4084 }
4085
4086 static void __exit obdfilter_exit(void)
4087 {
4088         if (filter_quota_interface_ref)
4089                 PORTAL_SYMBOL_PUT(filter_quota_interface);
4090
4091         if (ll_fmd_cachep) {
4092                 int rc = cfs_mem_cache_destroy(ll_fmd_cachep);
4093                 LASSERTF(rc == 0, "Cannot destroy ll_fmd_cachep: rc %d\n", rc);
4094                 ll_fmd_cachep = NULL;
4095         }
4096
4097         class_unregister_type(LUSTRE_OST_NAME);
4098         OBD_FREE(obdfilter_created_scratchpad,
4099                  OBDFILTER_CREATED_SCRATCHPAD_ENTRIES *
4100                  sizeof(*obdfilter_created_scratchpad));
4101 }
4102
4103 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4104 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
4105 MODULE_LICENSE("GPL");
4106
4107 module_init(obdfilter_init);
4108 module_exit(obdfilter_exit);