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_sem 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_sem to avoid single
  33  *            threaded operation on the OST.
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_FILTER
  37
  38 #include <linux/config.h>
  39 #include <linux/module.h>
  40 #include <linux/fs.h>
  41 #include <linux/dcache.h>
  42 #include <linux/init.h>
  43 #include <linux/version.h>
  44 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
  45 # include <linux/mount.h>
  46 # include <linux/buffer_head.h>
  47 #endif
  48
  49 #include <linux/obd_class.h>
  50 #include <linux/obd_lov.h>
  51 #include <linux/lustre_dlm.h>
  52 #include <linux/lustre_fsfilt.h>
  53 #include <linux/lprocfs_status.h>
  54 #include <linux/lustre_log.h>
  55 #include <linux/lustre_commit_confd.h>
  56
  57 #include "filter_internal.h"
  58
  59 static struct lvfs_callback_ops filter_lvfs_ops;
  60
  61 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
  62                           struct lov_stripe_md *ea, struct obd_trans_info *);
  63
  64 static void filter_commit_cb(struct obd_device *obd, __u64 transno,
  65                              void *cb_data, int error)
  66 {
  67         obd_transno_commit_cb(obd, transno, error);
  68 }
  69
  70
  71 /* Assumes caller has already pushed us into the kernel context. */
  72 int filter_finish_transno(struct obd_export *exp, struct obd_trans_info *oti,
  73                           int rc)
  74 {
  75         struct filter_obd *filter = &exp->exp_obd->u.filter;
  76         struct filter_export_data *fed = &exp->exp_filter_data;
  77         struct filter_client_data *fcd = fed->fed_fcd;
  78         __u64 last_rcvd;
  79         loff_t off;
  80         int err, log_pri = D_HA;
  81
  82         /* Propagate error code. */
  83         if (rc)
  84                 RETURN(rc);
  85
  86         if (!exp->exp_obd->obd_replayable || oti == NULL)
  87                 RETURN(rc);
  88
  89         /* we don't allocate new transnos for replayed requests */
  90         if (oti->oti_transno == 0) {
  91                 spin_lock(&filter->fo_translock);
  92                 last_rcvd = le64_to_cpu(filter->fo_fsd->fsd_last_transno) + 1;
  93                 filter->fo_fsd->fsd_last_transno = cpu_to_le64(last_rcvd);
  94                 spin_unlock(&filter->fo_translock);
  95                 oti->oti_transno = last_rcvd;
  96         } else {
  97                 spin_lock(&filter->fo_translock);
  98                 last_rcvd = oti->oti_transno;
  99                 if (last_rcvd > le64_to_cpu(filter->fo_fsd->fsd_last_transno))
 100                         filter->fo_fsd->fsd_last_transno =
 101                                 cpu_to_le64(last_rcvd);
 102                 spin_unlock(&filter->fo_translock);
 103         }
 104         fcd->fcd_last_rcvd = cpu_to_le64(last_rcvd);
 105         fcd->fcd_mount_count = cpu_to_le64(filter->fo_mount_count);
 106
 107         /* could get xid from oti, if it's ever needed */
 108         fcd->fcd_last_xid = 0;
 109
 110         off = fed->fed_lr_off;
 111         fsfilt_add_journal_cb(exp->exp_obd, last_rcvd, oti->oti_handle,
 112                               filter_commit_cb, NULL);
 113         err = fsfilt_write_record(exp->exp_obd, filter->fo_rcvd_filp, fcd,
 114                                   sizeof(*fcd), &off, 0);
 115         if (err) {
 116                 log_pri = D_ERROR;
 117                 if (rc == 0)
 118                         rc = err;
 119         }
 120
 121         CDEBUG(log_pri, "wrote trans "LPU64" for client %s at #%d: err = %d\n",
 122                last_rcvd, fcd->fcd_uuid, fed->fed_lr_idx, err);
 123
 124         RETURN(rc);
 125 }
 126
 127 void f_dput(struct dentry *dentry)
 128 {
 129         /* Can't go inside filter_ddelete because it can block */
 130         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
 131                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
 132         LASSERT(atomic_read(&dentry->d_count) > 0);
 133
 134         dput(dentry);
 135 }
 136
 137 /* Add client data to the FILTER.  We use a bitmap to locate a free space
 138  * in the last_rcvd file if cl_idx is -1 (i.e. a new client).
 139  * Otherwise, we have just read the data from the last_rcvd file and
 140  * we know its offset. */
 141 static int filter_client_add(struct obd_device *obd, struct filter_obd *filter,
 142                              struct filter_export_data *fed, int cl_idx)
 143 {
 144         unsigned long *bitmap = filter->fo_last_rcvd_slots;
 145         int new_client = (cl_idx == -1);
 146         ENTRY;
 147
 148         LASSERT(bitmap != NULL);
 149
 150         /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 151         if (!strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid))
 152                 RETURN(0);
 153
 154         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 155          * there's no need for extra complication here
 156          */
 157         if (new_client) {
 158                 cl_idx = find_first_zero_bit(bitmap, FILTER_LR_MAX_CLIENTS);
 159         repeat:
 160                 if (cl_idx >= FILTER_LR_MAX_CLIENTS) {
 161                         CERROR("no client slots - fix FILTER_LR_MAX_CLIENTS\n");
 162                         RETURN(-ENOMEM);
 163                 }
 164                 if (test_and_set_bit(cl_idx, bitmap)) {
 165                         CERROR("FILTER client %d: found bit is set in bitmap\n",
 166                                cl_idx);
 167                         cl_idx = find_next_zero_bit(bitmap,
 168                                                     FILTER_LR_MAX_CLIENTS,
 169                                                     cl_idx);
 170                         goto repeat;
 171                 }
 172         } else {
 173                 if (test_and_set_bit(cl_idx, bitmap)) {
 174                         CERROR("FILTER client %d: bit already set in bitmap!\n",
 175                                cl_idx);
 176                         LBUG();
 177                 }
 178         }
 179
 180         fed->fed_lr_idx = cl_idx;
 181         fed->fed_lr_off = le32_to_cpu(filter->fo_fsd->fsd_client_start) +
 182                 cl_idx * le16_to_cpu(filter->fo_fsd->fsd_client_size);
 183
 184         CDEBUG(D_INFO, "client at index %d (%llu) with UUID '%s' added\n",
 185                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
 186
 187         if (new_client) {
 188                 struct obd_run_ctxt saved;
 189                 loff_t off = fed->fed_lr_off;
 190                 int err;
 191                 void *handle;
 192
 193                 CDEBUG(D_INFO, "writing client fcd at idx %u (%llu) (len %u)\n",
 194                        fed->fed_lr_idx,off,(unsigned int)sizeof(*fed->fed_fcd));
 195
 196                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
 197                 /* Transaction needed to fix bug 1403 */
 198                 handle = fsfilt_start(obd,
 199                                       filter->fo_rcvd_filp->f_dentry->d_inode,
 200                                       FSFILT_OP_SETATTR, NULL);
 201                 if (IS_ERR(handle)) {
 202                         err = PTR_ERR(handle);
 203                         CERROR("unable to start transaction: rc %d\n", err);
 204                 } else {
 205                         err = fsfilt_write_record(obd, filter->fo_rcvd_filp,
 206                                                   fed->fed_fcd,
 207                                                   sizeof(*fed->fed_fcd),
 208                                                   &off, 1);
 209                         fsfilt_commit(obd,
 210                                       filter->fo_rcvd_filp->f_dentry->d_inode,
 211                                       handle, 1);
 212                 }
 213                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 214
 215                 if (err) {
 216                         CERROR("error writing %s client idx %u: rc %d\n",
 217                                LAST_RCVD, fed->fed_lr_idx, err);
 218                         RETURN(err);
 219                 }
 220         }
 221         RETURN(0);
 222 }
 223
 224 static int filter_client_free(struct obd_export *exp, int flags)
 225 {
 226         struct filter_export_data *fed = &exp->exp_filter_data;
 227         struct filter_obd *filter = &exp->exp_obd->u.filter;
 228         struct obd_device *obd = exp->exp_obd;
 229         struct filter_client_data zero_fcd;
 230         struct obd_run_ctxt saved;
 231         int rc;
 232         loff_t off;
 233         ENTRY;
 234
 235         if (fed->fed_fcd == NULL)
 236                 RETURN(0);
 237
 238         if (flags & OBD_OPT_FAILOVER)
 239                 GOTO(free, 0);
 240
 241         /* XXX if fcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 242         if (strcmp(fed->fed_fcd->fcd_uuid, obd->obd_uuid.uuid ) == 0)
 243                 GOTO(free, 0);
 244
 245         LASSERT(filter->fo_last_rcvd_slots != NULL);
 246
 247         off = fed->fed_lr_off;
 248
 249         CDEBUG(D_INFO, "freeing client at idx %u (%lld) with UUID '%s'\n",
 250                fed->fed_lr_idx, fed->fed_lr_off, fed->fed_fcd->fcd_uuid);
 251
 252         if (!test_and_clear_bit(fed->fed_lr_idx, filter->fo_last_rcvd_slots)) {
 253                 CERROR("FILTER client %u: bit already clear in bitmap!!\n",
 254                        fed->fed_lr_idx);
 255                 LBUG();
 256         }
 257
 258         memset(&zero_fcd, 0, sizeof zero_fcd);
 259         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 260         rc = fsfilt_write_record(obd, filter->fo_rcvd_filp, &zero_fcd,
 261                                  sizeof(zero_fcd), &off, 1);
 262         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 263
 264         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 265                "zeroing disconnecting client %s at idx %u (%llu) in %s rc %d\n",
 266                fed->fed_fcd->fcd_uuid, fed->fed_lr_idx, fed->fed_lr_off,
 267                LAST_RCVD, rc);
 268
 269 free:
 270         OBD_FREE(fed->fed_fcd, sizeof(*fed->fed_fcd));
 271
 272         RETURN(0);
 273 }
 274
 275 static int filter_free_server_data(struct filter_obd *filter)
 276 {
 277         OBD_FREE(filter->fo_fsd, sizeof(*filter->fo_fsd));
 278         filter->fo_fsd = NULL;
 279         OBD_FREE(filter->fo_last_rcvd_slots,
 280                  FILTER_LR_MAX_CLIENT_WORDS * sizeof(unsigned long));
 281         filter->fo_last_rcvd_slots = NULL;
 282         return 0;
 283 }
 284
 285 /* assumes caller is already in kernel ctxt */
 286 int filter_update_server_data(struct obd_device *obd, struct file *filp,
 287                               struct filter_server_data *fsd, int force_sync)
 288 {
 289         loff_t off = 0;
 290         int rc;
 291         ENTRY;
 292
 293         CDEBUG(D_INODE, "server uuid      : %s\n", fsd->fsd_uuid);
 294         CDEBUG(D_INODE, "server last_rcvd : "LPU64"\n",
 295                le64_to_cpu(fsd->fsd_last_transno));
 296         CDEBUG(D_INODE, "server last_mount: "LPU64"\n",
 297                le64_to_cpu(fsd->fsd_mount_count));
 298
 299         rc = fsfilt_write_record(obd, filp, fsd, sizeof(*fsd), &off,force_sync);
 300         if (rc)
 301                 CERROR("error writing filter_server_data: rc = %d\n", rc);
 302
 303         RETURN(rc);
 304 }
 305
 306 int filter_update_last_objid(struct obd_device *obd, obd_gr group,
 307                              int force_sync)
 308 {
 309         struct filter_obd *filter = &obd->u.filter;
 310         __u64 tmp;
 311         loff_t off = 0;
 312         int rc;
 313         ENTRY;
 314
 315         CDEBUG(D_INODE, "server last_objid for group "LPU64": "LPU64"\n",
 316                group, filter->fo_last_objids[group]);
 317
 318         tmp = cpu_to_le64(filter->fo_last_objids[group]);
 319         rc = fsfilt_write_record(obd, filter->fo_last_objid_files[group],
 320                                  &tmp, sizeof(tmp), &off, force_sync);
 321         if (rc)
 322                 CERROR("error writing group "LPU64" last objid: rc = %d\n",
 323                        group, rc);
 324         RETURN(rc);
 325 }
 326
 327 /* assumes caller has already in kernel ctxt */
 328 static int filter_init_server_data(struct obd_device *obd, struct file * filp)
 329 {
 330         struct filter_obd *filter = &obd->u.filter;
 331         struct filter_server_data *fsd;
 332         struct filter_client_data *fcd = NULL;
 333         struct inode *inode = filp->f_dentry->d_inode;
 334         unsigned long last_rcvd_size = inode->i_size;
 335         __u64 mount_count;
 336         int cl_idx;
 337         loff_t off = 0;
 338         int rc;
 339
 340         /* ensure padding in the struct is the correct size */
 341         LASSERT (offsetof(struct filter_server_data, fsd_padding) +
 342                  sizeof(fsd->fsd_padding) == FILTER_LR_SERVER_SIZE);
 343         LASSERT (offsetof(struct filter_client_data, fcd_padding) +
 344                  sizeof(fcd->fcd_padding) == FILTER_LR_CLIENT_SIZE);
 345
 346         OBD_ALLOC(fsd, sizeof(*fsd));
 347         if (!fsd)
 348                 RETURN(-ENOMEM);
 349         filter->fo_fsd = fsd;
 350
 351         OBD_ALLOC(filter->fo_last_rcvd_slots,
 352                   FILTER_LR_MAX_CLIENT_WORDS * sizeof(unsigned long));
 353         if (filter->fo_last_rcvd_slots == NULL) {
 354                 OBD_FREE(fsd, sizeof(*fsd));
 355                 RETURN(-ENOMEM);
 356         }
 357
 358         if (last_rcvd_size == 0) {
 359                 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
 360
 361                 memcpy(fsd->fsd_uuid, obd->obd_uuid.uuid,sizeof(fsd->fsd_uuid));
 362                 fsd->fsd_last_transno = 0;
 363                 mount_count = fsd->fsd_mount_count = 0;
 364                 fsd->fsd_server_size = cpu_to_le32(FILTER_LR_SERVER_SIZE);
 365                 fsd->fsd_client_start = cpu_to_le32(FILTER_LR_CLIENT_START);
 366                 fsd->fsd_client_size = cpu_to_le16(FILTER_LR_CLIENT_SIZE);
 367                 fsd->fsd_subdir_count = cpu_to_le16(FILTER_SUBDIR_COUNT);
 368                 filter->fo_subdir_count = FILTER_SUBDIR_COUNT;
 369         } else {
 370                 rc = fsfilt_read_record(obd, filp, fsd, sizeof(*fsd), &off);
 371                 if (rc) {
 372                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
 373                                LAST_RCVD, rc);
 374                         GOTO(err_fsd, rc);
 375                 }
 376                 if (strcmp(fsd->fsd_uuid, obd->obd_uuid.uuid) != 0) {
 377                         CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
 378                                obd->obd_uuid.uuid, fsd->fsd_uuid);
 379                         GOTO(err_fsd, rc = -EINVAL);
 380                 }
 381                 mount_count = le64_to_cpu(fsd->fsd_mount_count);
 382                 filter->fo_subdir_count = le16_to_cpu(fsd->fsd_subdir_count);
 383         }
 384
 385         if (fsd->fsd_feature_incompat & ~cpu_to_le32(FILTER_INCOMPAT_SUPP)) {
 386                 CERROR("unsupported feature %x\n",
 387                        le32_to_cpu(fsd->fsd_feature_incompat) &
 388                        ~FILTER_INCOMPAT_SUPP);
 389                 GOTO(err_fsd, rc = -EINVAL);
 390         }
 391         if (fsd->fsd_feature_rocompat & ~cpu_to_le32(FILTER_ROCOMPAT_SUPP)) {
 392                 CERROR("read-only feature %x\n",
 393                        le32_to_cpu(fsd->fsd_feature_rocompat) &
 394                        ~FILTER_ROCOMPAT_SUPP);
 395                 /* Do something like remount filesystem read-only */
 396                 GOTO(err_fsd, rc = -EINVAL);
 397         }
 398
 399         CDEBUG(D_INODE, "%s: server last_rcvd : "LPU64"\n",
 400                obd->obd_name, le64_to_cpu(fsd->fsd_last_transno));
 401         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 402                obd->obd_name, mount_count + 1);
 403         CDEBUG(D_INODE, "%s: server data size: %u\n",
 404                obd->obd_name, le32_to_cpu(fsd->fsd_server_size));
 405         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 406                obd->obd_name, le32_to_cpu(fsd->fsd_client_start));
 407         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 408                obd->obd_name, le32_to_cpu(fsd->fsd_client_size));
 409         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
 410                obd->obd_name, le16_to_cpu(fsd->fsd_subdir_count));
 411         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 412                last_rcvd_size <= FILTER_LR_CLIENT_START ? 0 :
 413                (last_rcvd_size-FILTER_LR_CLIENT_START) /FILTER_LR_CLIENT_SIZE);
 414
 415         if (!obd->obd_replayable) {
 416                 CWARN("%s: recovery support OFF\n", obd->obd_name);
 417                 GOTO(out, rc = 0);
 418         }
 419
 420         for (cl_idx = 0, off = le32_to_cpu(fsd->fsd_client_start);
 421              off < last_rcvd_size; cl_idx++) {
 422                 __u64 last_rcvd;
 423                 int mount_age;
 424
 425                 if (!fcd) {
 426                         OBD_ALLOC(fcd, sizeof(*fcd));
 427                         if (!fcd)
 428                                 GOTO(err_client, rc = -ENOMEM);
 429                 }
 430
 431                 /* Don't assume off is incremented properly by
 432                  * fsfilt_read_record(), in case sizeof(*fcd)
 433                  * isn't the same as fsd->fsd_client_size.  */
 434                 off = le32_to_cpu(fsd->fsd_client_start) +
 435                         cl_idx * le16_to_cpu(fsd->fsd_client_size);
 436                 rc = fsfilt_read_record(obd, filp, fcd, sizeof(*fcd), &off);
 437                 if (rc) {
 438                         CERROR("error reading FILT %s idx %d off %llu: rc %d\n",
 439                                LAST_RCVD, cl_idx, off, rc);
 440                         break; /* read error shouldn't cause startup to fail */
 441                 }
 442
 443                 if (fcd->fcd_uuid[0] == '\0') {
 444                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 445                                cl_idx);
 446                         continue;
 447                 }
 448
 449                 last_rcvd = le64_to_cpu(fcd->fcd_last_rcvd);
 450
 451                 /* These exports are cleaned up by filter_disconnect(), so they
 452                  * need to be set up like real exports as filter_connect() does.
 453                  */
 454                 mount_age = mount_count - le64_to_cpu(fcd->fcd_mount_count);
 455                 if (mount_age < FILTER_MOUNT_RECOV) {
 456                         struct obd_export *exp = class_new_export(obd);
 457                         struct filter_export_data *fed;
 458                         CERROR("RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 459                                " srv lr: "LPU64" mnt: "LPU64" last mount: "
 460                                LPU64"\n", fcd->fcd_uuid, cl_idx,
 461                                last_rcvd, le64_to_cpu(fsd->fsd_last_transno),
 462                                le64_to_cpu(fcd->fcd_mount_count), mount_count);
 463                         if (exp == NULL)
 464                                 GOTO(err_client, rc = -ENOMEM);
 465
 466                         memcpy(&exp->exp_client_uuid.uuid, fcd->fcd_uuid,
 467                                sizeof exp->exp_client_uuid.uuid);
 468                         fed = &exp->exp_filter_data;
 469                         fed->fed_fcd = fcd;
 470                         filter_client_add(obd, filter, fed, cl_idx);
 471                         /* create helper if export init gets more complex */
 472                         spin_lock_init(&fed->fed_lock);
 473
 474                         fcd = NULL;
 475                         obd->obd_recoverable_clients++;
 476                         class_export_put(exp);
 477                 } else {
 478                         CDEBUG(D_INFO, "discarded client %d UUID '%s' count "
 479                                LPU64"\n", cl_idx, fcd->fcd_uuid,
 480                                le64_to_cpu(fcd->fcd_mount_count));
 481                 }
 482
 483                 CDEBUG(D_OTHER, "client at idx %d has last_rcvd = "LPU64"\n",
 484                        cl_idx, last_rcvd);
 485
 486                 if (last_rcvd > le64_to_cpu(fsd->fsd_last_transno))
 487                         fsd->fsd_last_transno = cpu_to_le64(last_rcvd);
 488
 489         }
 490
 491         obd->obd_last_committed = le64_to_cpu(fsd->fsd_last_transno);
 492
 493         if (obd->obd_recoverable_clients) {
 494                 CERROR("RECOVERY: %d recoverable clients, last_rcvd "
 495                        LPU64"\n", obd->obd_recoverable_clients,
 496                        le64_to_cpu(fsd->fsd_last_transno));
 497                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 498                 obd->obd_recovering = 1;
 499         }
 500
 501         if (fcd)
 502                 OBD_FREE(fcd, sizeof(*fcd));
 503
 504 out:
 505         filter->fo_mount_count = mount_count + 1;
 506         fsd->fsd_mount_count = cpu_to_le64(filter->fo_mount_count);
 507
 508         /* save it, so mount count and last_transno is current */
 509         rc = filter_update_server_data(obd, filp, filter->fo_fsd, 1);
 510
 511         RETURN(rc);
 512
 513 err_client:
 514         class_disconnect_exports(obd, 0);
 515 err_fsd:
 516         filter_free_server_data(filter);
 517         RETURN(rc);
 518 }
 519
 520 static int filter_cleanup_groups(struct obd_device *obd)
 521 {
 522         struct filter_obd *filter = &obd->u.filter;
 523         int i;
 524         ENTRY;
 525
 526         if (filter->fo_dentry_O_groups != NULL &&
 527             filter->fo_last_objids != NULL &&
 528             filter->fo_last_objid_files != NULL) {
 529                 for (i = 0; i < FILTER_GROUPS; i++) {
 530                         struct dentry *dentry = filter->fo_dentry_O_groups[i];
 531                         struct file *filp = filter->fo_last_objid_files[i];
 532                         if (dentry != NULL) {
 533                                 f_dput(dentry);
 534                                 filter->fo_dentry_O_groups[i] = NULL;
 535                         }
 536                         if (filp != NULL) {
 537                                 filp_close(filp, 0);
 538                                 filter->fo_last_objid_files[i] = NULL;
 539                         }
 540                 }
 541         }
 542         if (filter->fo_dentry_O_sub != NULL && filter->fo_subdir_count) {
 543                 for (i = 0; i < filter->fo_subdir_count; i++) {
 544                         struct dentry *dentry = filter->fo_dentry_O_sub[i];
 545                         if (dentry != NULL) {
 546                                 f_dput(dentry);
 547                                 filter->fo_dentry_O_sub[i] = NULL;
 548                         }
 549                 }
 550                 OBD_FREE(filter->fo_dentry_O_sub,
 551                          filter->fo_subdir_count *
 552                          sizeof(*filter->fo_dentry_O_sub));
 553         }
 554         if (filter->fo_dentry_O_groups != NULL)
 555                 OBD_FREE(filter->fo_dentry_O_groups,
 556                          FILTER_GROUPS * sizeof(struct dentry *));
 557         if (filter->fo_last_objids != NULL)
 558                 OBD_FREE(filter->fo_last_objids,
 559                          FILTER_GROUPS * sizeof(__u64));
 560         if (filter->fo_last_objid_files != NULL)
 561                 OBD_FREE(filter->fo_last_objid_files,
 562                          FILTER_GROUPS * sizeof(struct file *));
 563         RETURN(0);
 564 }
 565
 566 /* FIXME: object groups */
 567 static int filter_prep_groups(struct obd_device *obd)
 568 {
 569         struct filter_obd *filter = &obd->u.filter;
 570         struct dentry *dentry, *O_dentry;
 571         struct file *filp;
 572         int i, rc = 0, cleanup_phase = 0;
 573         ENTRY;
 574
 575         O_dentry = simple_mkdir(current->fs->pwd, "O", 0700);
 576         CDEBUG(D_INODE, "got/created O: %p\n", O_dentry);
 577         if (IS_ERR(O_dentry)) {
 578                 rc = PTR_ERR(O_dentry);
 579                 CERROR("cannot open/create O: rc = %d\n", rc);
 580                 GOTO(cleanup, rc);
 581         }
 582         filter->fo_dentry_O = O_dentry;
 583         cleanup_phase = 1; /* O_dentry */
 584
 585         /* Lookup "R" to tell if we're on an old OST FS and need to convert
 586          * from O/R/<dir>/<objid> to O/0/<dir>/<objid>.  This can be removed
 587          * some time post 1.0 when all old-style OSTs have converted along
 588          * with the init_objid hack. */
 589         dentry = ll_lookup_one_len("R", O_dentry, 1);
 590         if (IS_ERR(dentry))
 591                 GOTO(cleanup, rc = PTR_ERR(dentry));
 592         if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
 593                 struct dentry *O0_dentry = lookup_one_len("0", O_dentry, 1);
 594                 ENTRY;
 595
 596                 CWARN("converting OST to new object layout\n");
 597                 if (IS_ERR(O0_dentry)) {
 598                         rc = PTR_ERR(O0_dentry);
 599                         CERROR("error looking up O/0: rc %d\n", rc);
 600                         GOTO(cleanup_R, rc);
 601                 }
 602
 603                 if (O0_dentry->d_inode) {
 604                         CERROR("Both O/R and O/0 exist. Fix manually.\n");
 605                         GOTO(cleanup_O0, rc = -EEXIST);
 606                 }
 607
 608                 down(&O_dentry->d_inode->i_sem);
 609                 rc = vfs_rename(O_dentry->d_inode, dentry,
 610                                 O_dentry->d_inode, O0_dentry);
 611                 up(&O_dentry->d_inode->i_sem);
 612
 613                 if (rc) {
 614                         CERROR("error renaming O/R to O/0: rc %d\n", rc);
 615                         GOTO(cleanup_O0, rc);
 616                 }
 617                 filter->fo_fsd->fsd_feature_incompat |=
 618                         cpu_to_le32(FILTER_INCOMPAT_GROUPS);
 619                 rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
 620                                                filter->fo_fsd, 1);
 621                 GOTO(cleanup_O0, rc);
 622
 623         cleanup_O0:
 624                 dput(O0_dentry);
 625         cleanup_R:
 626                 dput(dentry);
 627                 if (rc)
 628                         GOTO(cleanup, rc);
 629         } else {
 630                 dput(dentry);
 631         }
 632
 633         OBD_ALLOC(filter->fo_last_objids, FILTER_GROUPS * sizeof(__u64));
 634         if (filter->fo_last_objids == NULL)
 635                 GOTO(cleanup, rc = -ENOMEM);
 636         cleanup_phase = 2; /* groups */
 637
 638         OBD_ALLOC(filter->fo_dentry_O_groups, FILTER_GROUPS * sizeof(dentry));
 639         if (filter->fo_dentry_O_groups == NULL)
 640                 GOTO(cleanup, rc = -ENOMEM);
 641         OBD_ALLOC(filter->fo_last_objid_files, FILTER_GROUPS * sizeof(filp));
 642         if (filter->fo_last_objid_files == NULL)
 643                 GOTO(cleanup, rc = -ENOMEM);
 644
 645         for (i = 0; i < FILTER_GROUPS; i++) {
 646                 char name[25];
 647                 loff_t off = 0;
 648
 649                 sprintf(name, "%d", i);
 650                 dentry = simple_mkdir(O_dentry, name, 0700);
 651                 CDEBUG(D_INODE, "got/created O/%s: %p\n", name, dentry);
 652                 if (IS_ERR(dentry)) {
 653                         rc = PTR_ERR(dentry);
 654                         CERROR("cannot lookup/create O/%s: rc = %d\n",
 655                                name, rc);
 656                         GOTO(cleanup, rc);
 657                 }
 658                 filter->fo_dentry_O_groups[i] = dentry;
 659
 660                 sprintf(name, "O/%d/LAST_ID", i);
 661                 filp = filp_open(name, O_CREAT | O_RDWR, 0700);
 662                 if (IS_ERR(dentry)) {
 663                         rc = PTR_ERR(dentry);
 664                         CERROR("cannot create %s: rc = %d\n", name, rc);
 665                         GOTO(cleanup, rc);
 666                 }
 667                 filter->fo_last_objid_files[i] = filp;
 668
 669                 if (filp->f_dentry->d_inode->i_size == 0) {
 670                         if (i == 0 && filter->fo_fsd->fsd_unused != 0) {
 671                                 /* OST conversion, remove sometime post 1.0 */
 672                                 filter->fo_last_objids[i] =
 673                                         le64_to_cpu(filter->fo_fsd->fsd_unused);
 674                                 CWARN("saving old objid "LPU64" to LAST_ID\n",
 675                                       filter->fo_last_objids[i]);
 676                                 rc = filter_update_last_objid(obd, 0, 1);
 677                                 if (rc)
 678                                         GOTO(cleanup, rc);
 679                         } else {
 680                                 filter->fo_last_objids[i] = FILTER_INIT_OBJID;
 681                         }
 682                         continue;
 683                 }
 684
 685                 rc = fsfilt_read_record(obd, filp, &filter->fo_last_objids[i],
 686                                         sizeof(__u64), &off);
 687                 if (rc) {
 688                         CDEBUG(D_INODE,"OBD filter: error reading %s: rc %d\n",
 689                                name, rc);
 690                         GOTO(cleanup, rc);
 691                 }
 692                 filter->fo_last_objids[i] =
 693                         le64_to_cpu(filter->fo_last_objids[i]);
 694                 CDEBUG(D_INODE, "%s: server last_objid group %d: "LPU64"\n",
 695                        obd->obd_name, i, filter->fo_last_objids[i]);
 696         }
 697
 698         if (filter->fo_subdir_count) {
 699                 O_dentry = filter->fo_dentry_O_groups[0];
 700                 OBD_ALLOC(filter->fo_dentry_O_sub,
 701                           filter->fo_subdir_count * sizeof(dentry));
 702                 if (filter->fo_dentry_O_sub == NULL)
 703                         GOTO(cleanup, rc = -ENOMEM);
 704
 705                 for (i = 0; i < filter->fo_subdir_count; i++) {
 706                         char dir[20];
 707                         snprintf(dir, sizeof(dir), "d%u", i);
 708
 709                         dentry = simple_mkdir(O_dentry, dir, 0700);
 710                         CDEBUG(D_INODE, "got/created O/0/%s: %p\n", dir,dentry);
 711                         if (IS_ERR(dentry)) {
 712                                 rc = PTR_ERR(dentry);
 713                                 CERROR("can't lookup/create O/0/%s: rc = %d\n",
 714                                        dir, rc);
 715                                 GOTO(cleanup, rc);
 716                         }
 717                         filter->fo_dentry_O_sub[i] = dentry;
 718                 }
 719         }
 720         RETURN(0);
 721
 722  cleanup:
 723         switch (cleanup_phase) {
 724         case 2:
 725                 filter_cleanup_groups(obd);
 726         case 1:
 727                 f_dput(filter->fo_dentry_O);
 728                 filter->fo_dentry_O = NULL;
 729         default:
 730                 break;
 731         }
 732         return rc;
 733 }
 734
 735 /* setup the object store with correct subdirectories */
 736 static int filter_prep(struct obd_device *obd)
 737 {
 738         struct obd_run_ctxt saved;
 739         struct filter_obd *filter = &obd->u.filter;
 740         struct file *file;
 741         struct inode *inode;
 742         int rc = 0;
 743         ENTRY;
 744
 745         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 746         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT | O_LARGEFILE, 0700);
 747         if (!file || IS_ERR(file)) {
 748                 rc = PTR_ERR(file);
 749                 CERROR("OBD filter: cannot open/create %s: rc = %d\n",
 750                        LAST_RCVD, rc);
 751                 GOTO(out, rc);
 752         }
 753
 754         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 755                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 756                        file->f_dentry->d_inode->i_mode);
 757                 GOTO(err_filp, rc = -ENOENT);
 758         }
 759
 760         /* steal operations */
 761         inode = file->f_dentry->d_inode;
 762         filter->fo_fop = file->f_op;
 763         filter->fo_iop = inode->i_op;
 764         filter->fo_aops = inode->i_mapping->a_ops;
 765
 766         rc = filter_init_server_data(obd, file);
 767         if (rc) {
 768                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 769                 GOTO(err_filp, rc);
 770         }
 771         filter->fo_rcvd_filp = file;
 772
 773         rc = filter_prep_groups(obd);
 774         if (rc)
 775                 GOTO(err_server_data, rc);
 776
 777  out:
 778         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 779
 780         return(rc);
 781
 782  err_server_data:
 783         //class_disconnect_exports(obd, 0);
 784         filter_free_server_data(filter);
 785  err_filp:
 786         if (filp_close(file, 0))
 787                 CERROR("can't close %s after error\n", LAST_RCVD);
 788         filter->fo_rcvd_filp = NULL;
 789         goto out;
 790 }
 791
 792 /* cleanup the filter: write last used object id to status file */
 793 static void filter_post(struct obd_device *obd)
 794 {
 795         struct obd_run_ctxt saved;
 796         struct filter_obd *filter = &obd->u.filter;
 797         int rc, i;
 798
 799         /* XXX: filter_update_lastobjid used to call fsync_dev.  It might be
 800          * best to start a transaction with h_sync, because we removed this
 801          * from lastobjid */
 802
 803         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 804         rc = filter_update_server_data(obd, filter->fo_rcvd_filp,
 805                                        filter->fo_fsd, 0);
 806         if (rc)
 807                 CERROR("error writing server data: rc = %d\n", rc);
 808
 809         for (i = 0; i < FILTER_GROUPS; i++) {
 810                 rc = filter_update_last_objid(obd, i, (i == FILTER_GROUPS - 1));
 811                 if (rc)
 812                         CERROR("error writing group %d lastobjid: rc = %d\n",
 813                                i, rc);
 814         }
 815
 816         filp_close(filter->fo_rcvd_filp, 0);
 817         filter->fo_rcvd_filp = NULL;
 818         if (rc)
 819                 CERROR("error closing %s: rc = %d\n", LAST_RCVD, rc);
 820
 821         filter_cleanup_groups(obd);
 822         f_dput(filter->fo_dentry_O);
 823         filter_free_server_data(filter);
 824         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 825 }
 826
 827 static void filter_set_last_id(struct filter_obd *filter, struct obdo *oa,
 828                                obd_id id)
 829 {
 830         obd_gr group = 0;
 831         LASSERT(filter->fo_fsd != NULL);
 832
 833         if (oa != NULL) {
 834                 LASSERT(oa->o_gr <= FILTER_GROUPS);
 835                 group = oa->o_gr;
 836         }
 837
 838         spin_lock(&filter->fo_objidlock);
 839         filter->fo_last_objids[group] = id;
 840         spin_unlock(&filter->fo_objidlock);
 841 }
 842
 843 __u64 filter_last_id(struct filter_obd *filter, struct obdo *oa)
 844 {
 845         obd_id id;
 846         obd_gr group = 0;
 847         LASSERT(filter->fo_fsd != NULL);
 848
 849         if (oa != NULL) {
 850                 LASSERT(oa->o_gr <= FILTER_GROUPS);
 851                 group = oa->o_gr;
 852         }
 853
 854         /* FIXME: object groups */
 855         spin_lock(&filter->fo_objidlock);
 856         id = filter->fo_last_objids[group];
 857         spin_unlock(&filter->fo_objidlock);
 858
 859         return id;
 860 }
 861
 862 /* direct cut-n-paste of mds_blocking_ast() */
 863 static int filter_blocking_ast(struct ldlm_lock *lock,
 864                                struct ldlm_lock_desc *desc,
 865                                void *data, int flag)
 866 {
 867         int do_ast;
 868         ENTRY;
 869
 870         if (flag == LDLM_CB_CANCELING) {
 871                 /* Don't need to do anything here. */
 872                 RETURN(0);
 873         }
 874
 875         /* XXX layering violation!  -phil */
 876         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 877         /* Get this: if filter_blocking_ast is racing with ldlm_intent_policy,
 878          * such that filter_blocking_ast is called just before l_i_p takes the
 879          * ns_lock, then by the time we get the lock, we might not be the
 880          * correct blocking function anymore.  So check, and return early, if
 881          * so. */
 882         if (lock->l_blocking_ast != filter_blocking_ast) {
 883                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 884                 RETURN(0);
 885         }
 886
 887         lock->l_flags |= LDLM_FL_CBPENDING;
 888         do_ast = (!lock->l_readers && !lock->l_writers);
 889         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 890
 891         if (do_ast) {
 892                 struct lustre_handle lockh;
 893                 int rc;
 894
 895                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 896                 ldlm_lock2handle(lock, &lockh);
 897                 rc = ldlm_cli_cancel(&lockh);
 898                 if (rc < 0)
 899                         CERROR("ldlm_cli_cancel: %d\n", rc);
 900         } else {
 901                 LDLM_DEBUG(lock, "Lock still has references, will be "
 902                            "cancelled later");
 903         }
 904         RETURN(0);
 905 }
 906
 907 static int filter_lock_dentry(struct obd_device *obd, struct dentry *de,
 908                               ldlm_mode_t lock_mode,struct lustre_handle *lockh)
 909 {
 910         struct ldlm_res_id res_id = { .name = {0} };
 911         int flags = 0, rc;
 912         ENTRY;
 913
 914         res_id.name[0] = de->d_inode->i_ino;
 915         res_id.name[1] = de->d_inode->i_generation;
 916         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 917                               res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 918                               &flags, ldlm_completion_ast,
 919                               filter_blocking_ast, NULL, lockh);
 920
 921         RETURN(rc == ELDLM_OK ? 0 : -EIO);  /* XXX translate ldlm code */
 922 }
 923
 924 /* We never dget the object parent, so DON'T dput it either */
 925 static void filter_parent_unlock(struct dentry *dparent,
 926                                  struct lustre_handle *lockh,
 927                                  ldlm_mode_t lock_mode)
 928 {
 929         ldlm_lock_decref(lockh, lock_mode);
 930 }
 931
 932 /* We never dget the object parent, so DON'T dput it either */
 933 struct dentry *filter_parent(struct obd_device *obd, obd_gr group, obd_id objid)
 934 {
 935         struct filter_obd *filter = &obd->u.filter;
 936         LASSERT(group < FILTER_GROUPS); /* FIXME: object groups */
 937
 938         if (group > 0 || filter->fo_subdir_count == 0)
 939                 return filter->fo_dentry_O_groups[group];
 940
 941         return filter->fo_dentry_O_sub[objid & (filter->fo_subdir_count - 1)];
 942 }
 943
 944 /* We never dget the object parent, so DON'T dput it either */
 945 struct dentry *filter_parent_lock(struct obd_device *obd, obd_gr group,
 946                                   obd_id objid, ldlm_mode_t lock_mode,
 947                                   struct lustre_handle *lockh)
 948 {
 949         unsigned long now = jiffies;
 950         struct dentry *de = filter_parent(obd, group, objid);
 951         int rc;
 952
 953         if (IS_ERR(de))
 954                 return de;
 955
 956         rc = filter_lock_dentry(obd, de, lock_mode, lockh);
 957         if (time_after(jiffies, now + 15 * HZ))
 958                 CERROR("slow parent lock %lus\n", (jiffies - now) / HZ);
 959         return rc ? ERR_PTR(rc) : de;
 960 }
 961
 962 /* How to get files, dentries, inodes from object id's.
 963  *
 964  * If dir_dentry is passed, the caller has already locked the parent
 965  * appropriately for this operation (normally a write lock).  If
 966  * dir_dentry is NULL, we do a read lock while we do the lookup to
 967  * avoid races with create/destroy and such changing the directory
 968  * internal to the filesystem code. */
 969 struct dentry *filter_fid2dentry(struct obd_device *obd,
 970                                  struct dentry *dir_dentry,
 971                                  obd_gr group, obd_id id)
 972 {
 973         struct lustre_handle lockh;
 974         struct dentry *dparent = dir_dentry;
 975         struct dentry *dchild;
 976         char name[32];
 977         int len;
 978         ENTRY;
 979
 980         if (id == 0) {
 981                 CERROR("fatal: invalid object id 0\n");
 982                 RETURN(ERR_PTR(-ESTALE));
 983         }
 984
 985         len = sprintf(name, LPU64, id);
 986         if (dir_dentry == NULL) {
 987                 dparent = filter_parent_lock(obd, group, id, LCK_PR, &lockh);
 988                 if (IS_ERR(dparent))
 989                         RETURN(dparent);
 990         }
 991         CDEBUG(D_INODE, "looking up object O/%*s/%s\n",
 992                dparent->d_name.len, dparent->d_name.name, name);
 993         dchild = ll_lookup_one_len(name, dparent, len);
 994         if (dir_dentry == NULL)
 995                 filter_parent_unlock(dparent, &lockh, LCK_PR);
 996         if (IS_ERR(dchild)) {
 997                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 998                 RETURN(dchild);
 999         }
1000
1001         CDEBUG(D_INODE, "got child objid %s: %p, count = %d\n",
1002                name, dchild, atomic_read(&dchild->d_count));
1003
1004         LASSERT(atomic_read(&dchild->d_count) > 0);
1005
1006         RETURN(dchild);
1007 }
1008
1009 static int filter_prepare_destroy(struct obd_device *obd, obd_id objid)
1010 {
1011         struct lustre_handle lockh;
1012         int flags = LDLM_AST_DISCARD_DATA, rc;
1013         struct ldlm_res_id res_id = { .name = { objid } };
1014         struct ldlm_extent extent = { 0, OBD_OBJECT_EOF };
1015         ENTRY;
1016
1017         /* Tell the clients that the object is gone now and that they should
1018          * throw away any cached pages.  If we're the OST at stripe 0 in the
1019          * file then this enqueue will communicate the DISCARD to all the
1020          * clients.  This assumes that we always destroy all the objects for
1021          * a file at a time, as is currently the case.  If we're not the
1022          * OST at stripe 0 then we'll harmlessly get a very lonely lock in
1023          * the local DLM and immediately drop it. */
1024         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
1025                               res_id, LDLM_EXTENT, &extent,
1026                               sizeof(extent), LCK_PW, &flags,
1027                               ldlm_completion_ast, filter_blocking_ast,
1028                               NULL, &lockh);
1029
1030         /* We only care about the side-effects, just drop the lock. */
1031         if (rc == ELDLM_OK)
1032                 ldlm_lock_decref(&lockh, LCK_PW);
1033
1034         RETURN(rc);
1035 }
1036
1037 /* Caller must hold LCK_PW on parent and push us into kernel context.
1038  * Caller is also required to ensure that dchild->d_inode exists. */
1039 static int filter_destroy_internal(struct obd_device *obd, obd_id objid,
1040                                    struct dentry *dparent,
1041                                    struct dentry *dchild)
1042 {
1043         struct inode *inode = dchild->d_inode;
1044         int rc;
1045         ENTRY;
1046
1047         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
1048                 CERROR("destroying objid %*s nlink = %lu, count = %d\n",
1049                        dchild->d_name.len, dchild->d_name.name,
1050                        (unsigned long)inode->i_nlink,
1051                        atomic_read(&inode->i_count));
1052         }
1053
1054         rc = vfs_unlink(dparent->d_inode, dchild);
1055
1056         if (rc)
1057                 CERROR("error unlinking objid %*s: rc %d\n",
1058                        dchild->d_name.len, dchild->d_name.name, rc);
1059
1060         RETURN(rc);
1061 }
1062
1063 /* mount the file system (secretly) */
1064 int filter_common_setup(struct obd_device *obd, obd_count len, void *buf,
1065                         char *option)
1066 {
1067         struct lustre_cfg* lcfg = buf;
1068         struct filter_obd *filter = &obd->u.filter;
1069         struct vfsmount *mnt;
1070         int rc = 0;
1071         ENTRY;
1072
1073         dev_clear_rdonly(2);
1074
1075         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1076                 RETURN(-EINVAL);
1077
1078         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1079         if (IS_ERR(obd->obd_fsops))
1080                 RETURN(PTR_ERR(obd->obd_fsops));
1081
1082         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, MS_NOATIME | MS_NODIRATIME,
1083                             lcfg->lcfg_inlbuf1, option);
1084         rc = PTR_ERR(mnt);
1085         if (IS_ERR(mnt))
1086                 GOTO(err_ops, rc);
1087
1088         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1089                 if (*lcfg->lcfg_inlbuf3 == 'f') {
1090                         obd->obd_replayable = 1;
1091                         obd_sync_filter = 1;
1092                         CERROR("%s: recovery enabled\n", obd->obd_name);
1093                 } else {
1094                         if (*lcfg->lcfg_inlbuf3 != 'n') {
1095                                 CERROR("unrecognised flag '%c'\n",
1096                                        *lcfg->lcfg_inlbuf3);
1097                         }
1098                         // XXX Robert? Why do we get errors here
1099                         // GOTO(err_mntput, rc = -EINVAL);
1100                 }
1101         }
1102
1103         filter->fo_vfsmnt = mnt;
1104         filter->fo_sb = mnt->mnt_sb;
1105         filter->fo_fstype = mnt->mnt_sb->s_type->name;
1106         CDEBUG(D_SUPER, "%s: mnt = %p\n", filter->fo_fstype, mnt);
1107
1108         OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
1109         obd->obd_ctxt.pwdmnt = mnt;
1110         obd->obd_ctxt.pwd = mnt->mnt_root;
1111         obd->obd_ctxt.fs = get_ds();
1112         obd->obd_ctxt.cb_ops = filter_lvfs_ops;
1113
1114         rc = filter_prep(obd);
1115         if (rc)
1116                 GOTO(err_mntput, rc);
1117
1118         spin_lock_init(&filter->fo_translock);
1119         spin_lock_init(&filter->fo_objidlock);
1120         INIT_LIST_HEAD(&filter->fo_export_list);
1121         sema_init(&filter->fo_alloc_lock, 1);
1122         spin_lock_init(&filter->fo_r_pages.oh_lock);
1123         spin_lock_init(&filter->fo_w_pages.oh_lock);
1124         spin_lock_init(&filter->fo_r_discont_pages.oh_lock);
1125         spin_lock_init(&filter->fo_w_discont_pages.oh_lock);
1126         spin_lock_init(&filter->fo_r_discont_blocks.oh_lock);
1127         spin_lock_init(&filter->fo_w_discont_blocks.oh_lock);
1128
1129         obd->obd_namespace = ldlm_namespace_new("filter-tgt",
1130                                                 LDLM_NAMESPACE_SERVER);
1131         if (obd->obd_namespace == NULL)
1132                 GOTO(err_post, rc = -ENOMEM);
1133
1134         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1135                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
1136
1137         RETURN(0);
1138
1139 err_post:
1140         filter_post(obd);
1141 err_mntput:
1142         unlock_kernel();
1143         mntput(mnt);
1144         filter->fo_sb = 0;
1145         lock_kernel();
1146 err_ops:
1147         fsfilt_put_ops(obd->obd_fsops);
1148         return rc;
1149 }
1150
1151 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
1152 {
1153         struct lustre_cfg* lcfg = buf;
1154         const char *str = NULL;
1155         char *option = NULL;
1156         int n = 0;
1157         int rc;
1158
1159         if (!strcmp(lcfg->lcfg_inlbuf2, "ext3")) {
1160 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1161         /* bug 1577: implement async-delete for 2.5 */
1162                 str = "errors=remount-ro,asyncdel";
1163 #else
1164                 str = "errors=remount-ro";
1165 #endif
1166                 n = strlen(str) + 1;
1167                 OBD_ALLOC(option, n);
1168                 if (option == NULL)
1169                         RETURN(-ENOMEM);
1170                 strcpy(option, str);
1171         }
1172
1173         rc = filter_common_setup(obd, len, buf, option);
1174         if (option)
1175                 OBD_FREE(option, n);
1176         return rc;
1177 }
1178
1179 static int filter_postsetup(struct obd_device *obd)
1180 {
1181         int rc = 0;
1182         ENTRY;
1183
1184         // XXX add a storage location for the logid for size changes
1185 #ifdef ENABLE_ORPHANS
1186         rc = llog_cat_initialize(obd, 1);
1187         if (rc)
1188                 CERROR("failed to setup llogging subsystems\n");
1189 #endif
1190         RETURN(rc);
1191 }
1192
1193 static int filter_cleanup(struct obd_device *obd, int flags)
1194 {
1195         struct filter_obd *filter = &obd->u.filter;
1196         ENTRY;
1197
1198         if (flags & OBD_OPT_FAILOVER)
1199                 CERROR("%s: shutting down for failover; client state will"
1200                        " be preserved.\n", obd->obd_name);
1201
1202         if (!list_empty(&obd->obd_exports)) {
1203                 CERROR("%s: still has clients!\n", obd->obd_name);
1204                 class_disconnect_exports(obd, flags);
1205                 if (!list_empty(&obd->obd_exports)) {
1206                         CERROR("still has exports after forced cleanup?\n");
1207                         RETURN(-EBUSY);
1208                 }
1209         }
1210
1211         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1212
1213         if (filter->fo_sb == NULL)
1214                 RETURN(0);
1215
1216         filter_post(obd);
1217
1218         shrink_dcache_parent(filter->fo_sb->s_root);
1219         filter->fo_sb = 0;
1220
1221         if (atomic_read(&filter->fo_vfsmnt->mnt_count) > 1)
1222                 CERROR("%s: mount point %p busy, mnt_count: %d\n",
1223                        obd->obd_name, filter->fo_vfsmnt,
1224                        atomic_read(&filter->fo_vfsmnt->mnt_count));
1225
1226         unlock_kernel();
1227         mntput(filter->fo_vfsmnt);
1228         //destroy_buffers(filter->fo_sb->s_dev);
1229         filter->fo_sb = NULL;
1230         fsfilt_put_ops(obd->obd_fsops);
1231         lock_kernel();
1232
1233         dev_clear_rdonly(2);
1234
1235         RETURN(0);
1236 }
1237
1238 static int filter_attach(struct obd_device *obd, obd_count len, void *data)
1239 {
1240         struct lprocfs_static_vars lvars;
1241         int rc;
1242
1243         lprocfs_init_vars(filter, &lvars);
1244         rc = lprocfs_obd_attach(obd, lvars.obd_vars);
1245         if (rc != 0)
1246                 return rc;
1247
1248         rc = lprocfs_alloc_obd_stats(obd, LPROC_FILTER_LAST);
1249         if (rc != 0)
1250                 return rc;
1251
1252         /* Init obdfilter private stats here */
1253         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_READ_BYTES,
1254                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
1255         lprocfs_counter_init(obd->obd_stats, LPROC_FILTER_WRITE_BYTES,
1256                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
1257
1258         return lproc_filter_attach_seqstat(obd);
1259 }
1260
1261 static int filter_detach(struct obd_device *dev)
1262 {
1263         lprocfs_free_obd_stats(dev);
1264         return lprocfs_obd_detach(dev);
1265 }
1266
1267 /* nearly identical to mds_connect */
1268 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
1269                           struct obd_uuid *cluuid)
1270 {
1271         struct obd_export *exp;
1272         struct filter_export_data *fed;
1273         struct filter_client_data *fcd = NULL;
1274         struct filter_obd *filter = &obd->u.filter;
1275         int rc;
1276         ENTRY;
1277
1278         if (conn == NULL || obd == NULL || cluuid == NULL)
1279                 RETURN(-EINVAL);
1280
1281         rc = class_connect(conn, obd, cluuid);
1282         if (rc)
1283                 RETURN(rc);
1284         exp = class_conn2export(conn);
1285         LASSERT(exp != NULL);
1286
1287         fed = &exp->exp_filter_data;
1288
1289         spin_lock_init(&fed->fed_lock);
1290
1291         if (!obd->obd_replayable)
1292                 GOTO(cleanup, rc = 0);
1293
1294         OBD_ALLOC(fcd, sizeof(*fcd));
1295         if (!fcd) {
1296                 CERROR("filter: out of memory for client data\n");
1297                 GOTO(cleanup, rc = -ENOMEM);
1298         }
1299
1300         memcpy(fcd->fcd_uuid, cluuid, sizeof(fcd->fcd_uuid));
1301         fed->fed_fcd = fcd;
1302         fcd->fcd_mount_count = cpu_to_le64(filter->fo_mount_count);
1303
1304         rc = filter_client_add(obd, filter, fed, -1);
1305
1306 cleanup:
1307         if (rc) {
1308                 if (fcd)
1309                         OBD_FREE(fcd, sizeof(*fcd));
1310                 class_disconnect(exp, 0);
1311         } else {
1312                 class_export_put(exp);
1313         }
1314         return rc;
1315 }
1316
1317 static int filter_precleanup(struct obd_device *obd, int flags)
1318 {
1319         int rc = 0;
1320         ENTRY;
1321
1322 #ifdef ENABLE_ORPHANS
1323         rc = obd_llog_finish(obd, 0);
1324         if (rc)
1325                 CERROR("failed to cleanup llogging subsystem\n");
1326 #endif
1327
1328         RETURN(rc);
1329 }
1330
1331 static int filter_destroy_export(struct obd_export *exp)
1332 {
1333         ENTRY;
1334
1335         target_destroy_export(exp);
1336
1337         if (exp->exp_obd->obd_replayable)
1338                 filter_client_free(exp, exp->exp_flags);
1339         RETURN(0);
1340 }
1341
1342 /* also incredibly similar to mds_disconnect */
1343 static int filter_disconnect(struct obd_export *exp, int flags)
1344 {
1345         struct filter_obd *filter = &exp->exp_obd->u.filter;
1346         struct filter_export_data *fed = &exp->exp_filter_data;
1347         unsigned long irqflags;
1348         struct llog_ctxt *ctxt;
1349         int rc;
1350         ENTRY;
1351
1352         LASSERT(exp);
1353
1354         /* XXX should this go into filter_destroy_export() instead? */
1355         /* forget what this client had cached, I bet this needs to be
1356          * matched with appropriate client behaviour in the face of
1357          * disconnects */
1358         spin_lock(&exp->exp_obd->obd_osfs_lock);
1359         filter->fo_tot_cached -= fed->fed_cached;
1360         filter->fo_tot_granted -= fed->fed_grant;
1361         fed->fed_cached = 0;
1362         fed->fed_grant = 0;
1363         fed->fed_grant_sent = 0;
1364         fed->fed_grant_waiting = 0;
1365         spin_unlock(&exp->exp_obd->obd_osfs_lock);
1366
1367         ldlm_cancel_locks_for_export(exp);
1368
1369         spin_lock_irqsave(&exp->exp_lock, irqflags);
1370         exp->exp_flags = flags;
1371         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
1372
1373         fsfilt_sync(exp->exp_obd, filter->fo_sb);
1374         /* XXX cleanup preallocated inodes */
1375
1376         /* flush any remaining cancel messages out to the target */
1377         ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_REPL_CTXT);
1378         llog_sync(ctxt, exp);
1379
1380         rc = class_disconnect(exp, flags);
1381         RETURN(rc);
1382 }
1383
1384 struct dentry *__filter_oa2dentry(struct obd_device *obd,
1385                                   struct obdo *oa, const char *what)
1386 {
1387         struct dentry *dchild = NULL;
1388         obd_gr group = 0;
1389
1390         if (oa->o_valid & OBD_MD_FLGROUP)
1391                 group = oa->o_gr;
1392
1393         dchild = filter_fid2dentry(obd, NULL, group, oa->o_id);
1394
1395         if (IS_ERR(dchild)) {
1396                 CERROR("%s error looking up object: "LPU64"\n", what, oa->o_id);
1397                 RETURN(dchild);
1398         }
1399
1400         if (dchild->d_inode == NULL) {
1401                 CERROR("%s on non-existent object: "LPU64"\n", what, oa->o_id);
1402                 f_dput(dchild);
1403                 RETURN(ERR_PTR(-ENOENT));
1404         }
1405
1406         return dchild;
1407 }
1408
1409 static int filter_getattr(struct obd_export *exp, struct obdo *oa,
1410                           struct lov_stripe_md *md)
1411 {
1412         struct dentry *dentry = NULL;
1413         struct obd_device *obd;
1414         int rc = 0;
1415         ENTRY;
1416
1417         obd = class_exp2obd(exp);
1418         if (obd == NULL) {
1419                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1420                        exp->exp_handle.h_cookie);
1421                 RETURN(-EINVAL);
1422         }
1423
1424         dentry = filter_oa2dentry(obd, oa);
1425         if (IS_ERR(dentry))
1426                 RETURN(PTR_ERR(dentry));
1427
1428         /* Limit the valid bits in the return data to what we actually use */
1429         oa->o_valid = OBD_MD_FLID;
1430         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1431
1432         f_dput(dentry);
1433         RETURN(rc);
1434 }
1435
1436 /* this is called from filter_truncate() until we have filter_punch() */
1437 static int filter_setattr(struct obd_export *exp, struct obdo *oa,
1438                           struct lov_stripe_md *md, struct obd_trans_info *oti)
1439 {
1440         struct obd_run_ctxt saved;
1441         struct filter_obd *filter;
1442         struct dentry *dentry;
1443         struct iattr iattr;
1444         void *handle;
1445         int rc, rc2;
1446         ENTRY;
1447
1448         LASSERT(oti != NULL);
1449
1450         dentry = filter_oa2dentry(exp->exp_obd, oa);
1451         if (IS_ERR(dentry))
1452                 RETURN(PTR_ERR(dentry));
1453
1454         filter = &exp->exp_obd->u.filter;
1455
1456         iattr_from_obdo(&iattr, oa, oa->o_valid);
1457
1458         push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1459         lock_kernel();
1460
1461         if (iattr.ia_valid & ATTR_SIZE)
1462                 down(&dentry->d_inode->i_sem);
1463         handle = fsfilt_start(exp->exp_obd, dentry->d_inode, FSFILT_OP_SETATTR,
1464                               oti);
1465         if (IS_ERR(handle))
1466                 GOTO(out_unlock, rc = PTR_ERR(handle));
1467
1468         /* XXX this could be a rwsem instead, if filter_preprw played along */
1469         if (iattr.ia_valid & ATTR_ATTR_FLAG)
1470                 rc = fsfilt_iocontrol(exp->exp_obd, dentry->d_inode, NULL,
1471                                       EXT3_IOC_SETFLAGS,
1472                                       (long)&iattr.ia_attr_flags);
1473         else
1474                 rc = fsfilt_setattr(exp->exp_obd, dentry, handle, &iattr, 1);
1475         rc = filter_finish_transno(exp, oti, rc);
1476         rc2 = fsfilt_commit(exp->exp_obd, dentry->d_inode, handle, 0);
1477         if (rc2) {
1478                 CERROR("error on commit, err = %d\n", rc2);
1479                 if (!rc)
1480                         rc = rc2;
1481         }
1482
1483         oa->o_valid = OBD_MD_FLID;
1484         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1485
1486 out_unlock:
1487         if (iattr.ia_valid & ATTR_SIZE)
1488                 up(&dentry->d_inode->i_sem);
1489         unlock_kernel();
1490         pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1491
1492         f_dput(dentry);
1493         RETURN(rc);
1494 }
1495
1496 /* XXX identical to osc_unpackmd */
1497 static int filter_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
1498                            struct lov_mds_md *lmm, int lmm_bytes)
1499 {
1500         int lsm_size;
1501         ENTRY;
1502
1503         if (lmm != NULL) {
1504                 if (lmm_bytes < sizeof (*lmm)) {
1505                         CERROR("lov_mds_md too small: %d, need %d\n",
1506                                lmm_bytes, (int)sizeof(*lmm));
1507                         RETURN(-EINVAL);
1508                 }
1509                 /* XXX LOV_MAGIC etc check? */
1510
1511                 if (lmm->lmm_object_id == cpu_to_le64(0)) {
1512                         CERROR("lov_mds_md: zero lmm_object_id\n");
1513                         RETURN(-EINVAL);
1514                 }
1515         }
1516
1517         lsm_size = lov_stripe_md_size(1);
1518         if (lsmp == NULL)
1519                 RETURN(lsm_size);
1520
1521         if (*lsmp != NULL && lmm == NULL) {
1522                 OBD_FREE(*lsmp, lsm_size);
1523                 *lsmp = NULL;
1524                 RETURN(0);
1525         }
1526
1527         if (*lsmp == NULL) {
1528                 OBD_ALLOC(*lsmp, lsm_size);
1529                 if (*lsmp == NULL)
1530                         RETURN(-ENOMEM);
1531
1532                 loi_init((*lsmp)->lsm_oinfo);
1533         }
1534
1535         if (lmm != NULL) {
1536                 /* XXX zero *lsmp? */
1537                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
1538                 LASSERT((*lsmp)->lsm_object_id);
1539         }
1540
1541         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
1542
1543         RETURN(lsm_size);
1544 }
1545
1546 static void filter_destroy_precreated(struct obd_export *exp, struct obdo *oa,
1547                                       struct filter_obd *filter)
1548 {
1549         struct obdo doa; /* XXX obdo on stack */
1550         __u64 last, id;
1551         ENTRY;
1552         LASSERT(oa);
1553
1554         memset(&doa, 0, sizeof(doa));
1555         if (oa->o_valid & OBD_MD_FLGROUP)
1556                 doa.o_gr = oa->o_gr;
1557         else
1558                 doa.o_gr = 0;
1559         doa.o_mode = S_IFREG;
1560         last = filter_last_id(filter, &doa); /* FIXME: object groups */
1561         CWARN("deleting orphan objects from "LPU64" to "LPU64"\n",
1562                oa->o_id + 1, last);
1563         for (id = oa->o_id + 1; id <= last; id++) {
1564                 doa.o_id = id;
1565                 filter_destroy(exp, &doa, NULL, NULL);
1566         }
1567         spin_lock(&filter->fo_objidlock);
1568         filter->fo_last_objids[0] = oa->o_id; /* FIXME: object groups */
1569         spin_unlock(&filter->fo_objidlock);
1570         EXIT;
1571 }
1572
1573 /* returns a negative error or a nonnegative number of files to create */
1574 static int filter_should_precreate(struct obd_export *exp, struct obdo *oa,
1575                                    int group)
1576 {
1577         struct obd_device *obd = exp->exp_obd;
1578         struct filter_obd *filter = &obd->u.filter;
1579         int diff, rc;
1580         ENTRY;
1581
1582         diff = oa->o_id - filter_last_id(filter, oa);
1583         CDEBUG(D_INFO, "filter_last_id() = "LPU64" -> diff = %d\n",
1584                filter_last_id(filter, oa), diff);
1585
1586         /* delete orphans request */
1587         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
1588             (oa->o_flags & OBD_FL_DELORPHAN)) {
1589                 if (diff >= 0)
1590                         RETURN(diff);
1591                 filter_destroy_precreated(exp, oa, filter);
1592                 rc = filter_update_last_objid(obd, group, 0);
1593                 if (rc)
1594                         CERROR("unable to write lastobjid, but orphans"
1595                                "were deleted\n");
1596                 RETURN(0);
1597         } else {
1598                 /* only precreate if group == 0 and o_id is specfied */
1599                 if (!(oa->o_valid & OBD_FL_DELORPHAN) &&
1600                     (group != 0 || oa->o_id == 0))
1601                         RETURN(1);
1602
1603                 LASSERT(diff >= 0);
1604                 RETURN(diff);
1605         }
1606
1607 }
1608
1609 /* We rely on the fact that only one thread will be creating files in a given
1610  * group at a time, which is why we don't need an atomic filter_get_new_id.
1611  * Even if we had that atomic function, the following race would exist:
1612  *
1613  * thread 1: gets id x from filter_next_id
1614  * thread 2: gets id (x + 1) from filter_next_id
1615  * thread 2: creates object (x + 1)
1616  * thread 1: tries to create object x, gets -ENOSPC
1617  */
1618 static int filter_precreate(struct obd_device *obd, struct obdo *oa,
1619                             obd_gr group, int *num)
1620 {
1621         struct lustre_handle parent_lockh;
1622         struct dentry *dchild = NULL;
1623         struct filter_obd *filter;
1624         struct dentry *dparent;
1625         int err = 0, rc = 0, i;
1626         __u64 next_id;
1627         void *handle = NULL;
1628         ENTRY;
1629
1630         filter = &obd->u.filter;
1631
1632         for (i = 0; i < *num && err == 0; i++) {
1633                 int cleanup_phase = 0;
1634
1635                 next_id = filter_last_id(filter, oa) + 1;
1636                 CDEBUG(D_INFO, "precreate objid "LPU64"\n", next_id);
1637
1638                 dparent = filter_parent_lock(obd, group, next_id, LCK_PW,
1639                                              &parent_lockh);
1640                 if (IS_ERR(dparent))
1641                         GOTO(cleanup, rc = PTR_ERR(dparent));
1642                 cleanup_phase = 1;
1643
1644                 dchild = filter_fid2dentry(obd, dparent, group, next_id);
1645                 if (IS_ERR(dchild))
1646                         GOTO(cleanup, rc = PTR_ERR(dchild));
1647                 cleanup_phase = 2;
1648
1649                 if (dchild->d_inode != NULL) {
1650                         /* This would only happen if lastobjid was bad on disk*/
1651                         CERROR("Serious error: objid %*s already exists; is "
1652                                "this filesystem corrupt?\n",
1653                                dchild->d_name.len, dchild->d_name.name);
1654                         GOTO(cleanup, rc = -EEXIST);
1655                 }
1656
1657                 handle = fsfilt_start(obd, dparent->d_inode,
1658                                       FSFILT_OP_CREATE_LOG, NULL);
1659                 if (IS_ERR(handle))
1660                         GOTO(cleanup, rc = PTR_ERR(handle));
1661                 cleanup_phase = 3;
1662
1663                 rc = ll_vfs_create(dparent->d_inode, dchild, S_IFREG, NULL);
1664                 if (rc) {
1665                         CERROR("create failed rc = %d\n", rc);
1666                         GOTO(cleanup, rc);
1667                 }
1668
1669                 filter_set_last_id(filter, oa, next_id);
1670                 err = filter_update_last_objid(obd, group, 0);
1671                 if (err)
1672                         CERROR("unable to write lastobjid but file created\n");
1673
1674         cleanup:
1675                 switch(cleanup_phase) {
1676                 case 3:
1677                         err = fsfilt_commit(obd, dparent->d_inode, handle, 0);
1678                         if (err) {
1679                                 CERROR("error on commit, err = %d\n", err);
1680                                 if (!rc)
1681                                         rc = err;
1682                         }
1683                 case 2:
1684                         f_dput(dchild);
1685                 case 1:
1686                         filter_parent_unlock(dparent, &parent_lockh, LCK_PW);
1687                 case 0:
1688                         break;
1689                 }
1690
1691                 if (rc)
1692                         break;
1693         }
1694         *num = i;
1695
1696         CDEBUG(D_INFO, "filter_precreate() created %d objects\n", i);
1697         RETURN(rc);
1698 }
1699
1700 static int filter_create(struct obd_export *exp, struct obdo *oa,
1701                          struct lov_stripe_md **ea, struct obd_trans_info *oti)
1702 {
1703         struct obd_device *obd = NULL;
1704         struct obd_run_ctxt saved;
1705         struct lov_stripe_md *lsm = NULL;
1706         obd_gr group = 0;
1707         int rc = 0, diff;
1708         ENTRY;
1709
1710         if (oa->o_valid & OBD_MD_FLGROUP)
1711                 group = oa->o_gr;
1712
1713         CDEBUG(D_INFO, "filter_create(od->o_gr="LPU64",od->o_id="LPU64")\n",
1714                group, oa->o_id);
1715         if (ea != NULL) {
1716                 lsm = *ea;
1717                 if (lsm == NULL) {
1718                         rc = obd_alloc_memmd(exp, &lsm);
1719                         if (rc < 0)
1720                                 RETURN(rc);
1721                 }
1722         }
1723
1724         obd = exp->exp_obd;
1725         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1726
1727         diff = filter_should_precreate(exp, oa, group);
1728         if (diff > 0) {
1729                 oa->o_id = filter_last_id(&obd->u.filter, oa);
1730                 rc = filter_precreate(obd, oa, group, &diff);
1731                 oa->o_id += diff;
1732                 oa->o_valid = OBD_MD_FLID;
1733         }
1734
1735         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1736         if (rc && ea != NULL && *ea != lsm) {
1737                 obd_free_memmd(exp, &lsm);
1738         } else if (rc == 0 && ea != NULL) {
1739                 /* XXX LOV STACKING: the lsm that is passed to us from
1740                  * LOV does not have valid lsm_oinfo data structs, so
1741                  * don't go touching that.  This needs to be fixed in a
1742                  * big way. */
1743                 lsm->lsm_object_id = oa->o_id;
1744                 *ea = lsm;
1745         }
1746
1747         RETURN(rc);
1748 }
1749
1750 static int filter_destroy(struct obd_export *exp, struct obdo *oa,
1751                           struct lov_stripe_md *ea, struct obd_trans_info *oti)
1752 {
1753         struct obd_device *obd;
1754         struct filter_obd *filter;
1755         struct dentry *dchild = NULL, *dparent = NULL;
1756         struct obd_run_ctxt saved;
1757         void *handle = NULL;
1758         struct lustre_handle parent_lockh;
1759         struct llog_cookie *fcc = NULL;
1760         int rc, rc2, cleanup_phase = 0, have_prepared = 0;
1761         obd_gr group = 0;
1762         ENTRY;
1763
1764         if (oa->o_valid & OBD_MD_FLGROUP)
1765                 group = oa->o_gr;
1766
1767         obd = exp->exp_obd;
1768         filter = &obd->u.filter;
1769
1770         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1771
1772  acquire_locks:
1773         dparent = filter_parent_lock(obd, group, oa->o_id, LCK_PW,
1774                                      &parent_lockh);
1775         if (IS_ERR(dparent))
1776                 GOTO(cleanup, rc = PTR_ERR(dparent));
1777         cleanup_phase = 1;
1778
1779         dchild = filter_fid2dentry(obd, dparent, group, oa->o_id);
1780         if (IS_ERR(dchild))
1781                 GOTO(cleanup, rc = -ENOENT);
1782         cleanup_phase = 2;
1783
1784         if (dchild->d_inode == NULL) {
1785                 CERROR("destroying non-existent object "LPU64"\n", oa->o_id);
1786                 GOTO(cleanup, rc = -ENOENT);
1787         }
1788
1789         if (!have_prepared) {
1790                 /* If we're really going to destroy the object, get ready
1791                  * by getting the clients to discard their cached data.
1792                  *
1793                  * We have to drop the parent lock, because
1794                  * filter_prepare_destroy will acquire a PW on the object, and
1795                  * we don't want to deadlock with an incoming write to the
1796                  * object, which has the extent PW and then wants to get the
1797                  * parent dentry to do the lookup.
1798                  *
1799                  * We dput the child because it's not worth the extra
1800                  * complication of condition the above code to skip it on the
1801                  * second time through. */
1802                 f_dput(dchild);
1803                 filter_parent_unlock(dparent, &parent_lockh, LCK_PW);
1804
1805                 filter_prepare_destroy(obd, oa->o_id);
1806                 have_prepared = 1;
1807                 goto acquire_locks;
1808         }
1809
1810         handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_UNLINK_LOG, oti);
1811         if (IS_ERR(handle))
1812                 GOTO(cleanup, rc = PTR_ERR(handle));
1813         cleanup_phase = 3;
1814
1815         /* Our MDC connection is established by the MDS to us */
1816         if (oa->o_valid & OBD_MD_FLCOOKIE) {
1817                 OBD_ALLOC(fcc, sizeof(*fcc));
1818                 if (fcc != NULL)
1819                         memcpy(fcc, obdo_logcookie(oa), sizeof(*fcc));
1820         }
1821
1822         rc = filter_destroy_internal(obd, oa->o_id, dparent, dchild);
1823
1824 cleanup:
1825         switch(cleanup_phase) {
1826         case 3:
1827                 if (fcc != NULL)
1828                         fsfilt_add_journal_cb(obd, 0, oti->oti_handle,
1829                                               filter_cancel_cookies_cb, fcc);
1830                 rc = filter_finish_transno(exp, oti, rc);
1831                 rc2 = fsfilt_commit(obd, dparent->d_inode, handle, 0);
1832                 if (rc2) {
1833                         CERROR("error on commit, err = %d\n", rc2);
1834                         if (!rc)
1835                                 rc = rc2;
1836                 }
1837         case 2:
1838                 f_dput(dchild);
1839         case 1:
1840                 if (rc || oti == NULL) {
1841                         filter_parent_unlock(dparent, &parent_lockh, LCK_PW);
1842                 } else {
1843                         memcpy(&oti->oti_ack_locks[0].lock, &parent_lockh,
1844                                sizeof(parent_lockh));
1845                         oti->oti_ack_locks[0].mode = LCK_PW;
1846                 }
1847         case 0:
1848                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1849                 break;
1850         default:
1851                 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
1852                 LBUG();
1853         }
1854
1855         RETURN(rc);
1856 }
1857
1858 /* NB start and end are used for punch, but not truncate */
1859 static int filter_truncate(struct obd_export *exp, struct obdo *oa,
1860                            struct lov_stripe_md *lsm,
1861                            obd_off start, obd_off end,
1862                            struct obd_trans_info *oti)
1863 {
1864         int error;
1865         ENTRY;
1866
1867         if (end != OBD_OBJECT_EOF)
1868                 CERROR("PUNCH not supported, only truncate: end = "LPX64"\n",
1869                        end);
1870
1871         CDEBUG(D_INODE, "calling truncate for object "LPU64", valid = %x, "
1872                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
1873         oa->o_size = start;
1874         error = filter_setattr(exp, oa, NULL, oti);
1875         RETURN(error);
1876 }
1877
1878 static int filter_sync(struct obd_export *exp, struct obdo *oa,
1879                        struct lov_stripe_md *lsm, obd_off start, obd_off end)
1880 {
1881         struct obd_run_ctxt saved;
1882         struct filter_obd *filter;
1883         struct dentry *dentry;
1884         int rc, rc2;
1885         ENTRY;
1886
1887         filter = &exp->exp_obd->u.filter;
1888
1889         /* an objid of zero is taken to mean "sync whole filesystem" */
1890         if (!oa || !(oa->o_valid & OBD_MD_FLID)) {
1891                 rc = fsfilt_sync(exp->exp_obd, filter->fo_sb);
1892                 RETURN(rc);
1893         }
1894
1895         dentry = filter_oa2dentry(exp->exp_obd, oa);
1896         if (IS_ERR(dentry))
1897                 RETURN(PTR_ERR(dentry));
1898
1899         push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1900
1901         down(&dentry->d_inode->i_sem);
1902         rc = filemap_fdatasync(dentry->d_inode->i_mapping);
1903         if (rc == 0) {
1904                 /* just any file to grab fsync method - "file" arg unused */
1905                 struct file *file = filter->fo_rcvd_filp;
1906
1907                 if (file->f_op && file->f_op->fsync)
1908                         rc = file->f_op->fsync(NULL, dentry, 1);
1909
1910                 rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
1911                 if (!rc)
1912                         rc = rc2;
1913         }
1914         up(&dentry->d_inode->i_sem);
1915
1916         oa->o_valid = OBD_MD_FLID;
1917         obdo_from_inode(oa, dentry->d_inode, FILTER_VALID_FLAGS);
1918
1919         pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
1920
1921         f_dput(dentry);
1922         RETURN(rc);
1923 }
1924
1925 /* debugging to make sure that nothing bad happens, can be turned off soon */
1926 static void filter_grant_sanity_check(struct obd_device *obd,
1927                                       struct filter_obd *filter, __u64 maxsize)
1928 {
1929         struct filter_export_data *fed;
1930         struct obd_export *exp_pos;
1931         obd_size tot_cached = 0, tot_granted = 0;
1932
1933         list_for_each_entry(exp_pos, &obd->obd_exports, exp_obd_chain) {
1934                 fed = &exp_pos->exp_filter_data;
1935                 LASSERT(fed->fed_cached <= maxsize);
1936                 LASSERT(fed->fed_grant <= maxsize);
1937                 tot_cached += fed->fed_cached;
1938                 tot_granted += fed->fed_grant;
1939         }
1940         LASSERT(tot_cached == filter->fo_tot_cached);
1941         LASSERT(tot_granted == filter->fo_tot_granted);
1942         LASSERT(tot_cached <= maxsize);
1943         LASSERT(tot_granted <= maxsize);
1944 }
1945
1946 static int filter_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1947                          unsigned long max_age)
1948 {
1949         struct filter_obd *filter = &obd->u.filter;
1950         __u64 cached;
1951         int blockbits = filter->fo_sb->s_blocksize_bits;
1952         int rc;
1953         ENTRY;
1954
1955         /* at least try to account for cached pages.  its still racey and
1956          * might be under-reporting if clients haven't announced their
1957          * caches with brw recently */
1958         spin_lock(&obd->obd_osfs_lock);
1959         rc = fsfilt_statfs(obd, filter->fo_sb, max_age);
1960         memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1961         filter_grant_sanity_check(obd, filter, osfs->os_blocks << blockbits);
1962         cached = filter->fo_tot_cached + osfs->os_bsize - 1;
1963         spin_unlock(&obd->obd_osfs_lock);
1964
1965         cached >>= blockbits;
1966         if (cached > osfs->os_bavail) {
1967                 CERROR("cached "LPU64" > bavail "LPU64", clamping\n", cached,
1968                        osfs->os_bavail);
1969                 cached = osfs->os_bavail;
1970         }
1971         osfs->os_bavail -= cached;
1972         osfs->os_bfree -= cached;
1973
1974         RETURN(rc);
1975 }
1976
1977 static int filter_get_info(struct obd_export *exp, __u32 keylen,
1978                            void *key, __u32 *vallen, void *val)
1979 {
1980         struct obd_device *obd;
1981         ENTRY;
1982
1983         obd = class_exp2obd(exp);
1984         if (obd == NULL) {
1985                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1986                        exp->exp_handle.h_cookie);
1987                 RETURN(-EINVAL);
1988         }
1989
1990         if (keylen == strlen("blocksize") &&
1991             memcmp(key, "blocksize", keylen) == 0) {
1992                 __u32 *blocksize = val;
1993                 *vallen = sizeof(*blocksize);
1994                 *blocksize = obd->u.filter.fo_sb->s_blocksize;
1995                 RETURN(0);
1996         }
1997
1998         if (keylen == strlen("blocksize_bits") &&
1999             memcmp(key, "blocksize_bits", keylen) == 0) {
2000                 __u32 *blocksize_bits = val;
2001                 *vallen = sizeof(*blocksize_bits);
2002                 *blocksize_bits = obd->u.filter.fo_sb->s_blocksize_bits;
2003                 RETURN(0);
2004         }
2005
2006         if (keylen >= strlen("last_id") && memcmp(key, "last_id", 7) == 0) {
2007                 obd_id *last_id = val;
2008                 /* FIXME: object groups */
2009                 *last_id = filter_last_id(&obd->u.filter, 0);
2010                 RETURN(0);
2011         }
2012         CDEBUG(D_IOCTL, "invalid key\n");
2013         RETURN(-EINVAL);
2014 }
2015
2016 static int filter_set_info(struct obd_export *exp, __u32 keylen,
2017                            void *key, __u32 vallen, void *val)
2018 {
2019         struct obd_device *obd;
2020         struct lustre_handle conn;
2021 #ifdef ENABLE_ORPHANS
2022         struct llog_ctxt *ctxt;
2023 #endif
2024         int rc = 0;
2025         ENTRY;
2026
2027         conn.cookie = exp->exp_handle.h_cookie;
2028
2029         obd = exp->exp_obd;
2030         if (obd == NULL) {
2031                 CDEBUG(D_IOCTL, "invalid exp %p cookie "LPX64"\n",
2032                        exp, conn.cookie);
2033                 RETURN(-EINVAL);
2034         }
2035
2036         if (keylen < strlen("mds_conn") ||
2037             memcmp(key, "mds_conn", keylen) != 0)
2038                 RETURN(-EINVAL);
2039
2040         CWARN("Received MDS connection ("LPX64")\n", conn.cookie);
2041         memcpy(&obd->u.filter.fo_mdc_conn, &conn, sizeof(conn));
2042 #ifdef ENABLE_ORPHANS
2043         ctxt = llog_get_context(obd, LLOG_UNLINK_REPL_CTXT);
2044         rc = llog_receptor_accept(ctxt, exp->exp_imp_reverse);
2045 #endif
2046         RETURN(rc);
2047 }
2048
2049 int filter_iocontrol(unsigned int cmd, struct obd_export *exp,
2050                      int len, void *karg, void *uarg)
2051 {
2052         struct obd_device *obd = exp->exp_obd;
2053         struct obd_ioctl_data *data = karg;
2054         int rc = 0;
2055
2056         switch (cmd) {
2057         case OBD_IOC_ABORT_RECOVERY:
2058                 CERROR("aborting recovery for device %s\n", obd->obd_name);
2059                 target_abort_recovery(obd);
2060                 RETURN(0);
2061
2062         case OBD_IOC_SET_READONLY: {
2063                 void *handle;
2064                 struct super_block *sb = obd->u.filter.fo_sb;
2065                 struct inode *inode = sb->s_root->d_inode;
2066                 BDEVNAME_DECLARE_STORAGE(tmp);
2067                 CERROR("setting device %s read-only\n",
2068                        ll_bdevname(sb, tmp));
2069
2070                 handle = fsfilt_start(obd, inode, FSFILT_OP_MKNOD, NULL);
2071                 LASSERT(handle);
2072                 (void)fsfilt_commit(obd, inode, handle, 1);
2073
2074                 dev_set_rdonly(ll_sbdev(obd->u.filter.fo_sb), 2);
2075                 RETURN(0);
2076         }
2077
2078         case OBD_IOC_CATLOGLIST: {
2079                 rc = llog_catlog_list(obd, 1, data);
2080                 RETURN(rc);
2081         }
2082
2083         case OBD_IOC_LLOG_CANCEL:
2084         case OBD_IOC_LLOG_REMOVE:
2085         case OBD_IOC_LLOG_INFO:
2086         case OBD_IOC_LLOG_PRINT: {
2087                 /* FIXME to be finished */
2088                 RETURN(-EOPNOTSUPP);
2089 /*
2090                 struct llog_ctxt *ctxt = NULL;
2091
2092                 push_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2093                 rc = llog_ioctl(ctxt, cmd, data);
2094                 pop_ctxt(&saved, &ctxt->loc_exp->exp_obd->obd_ctxt, NULL);
2095
2096                 RETURN(rc);
2097 */
2098         }
2099
2100
2101         default:
2102                 RETURN(-EINVAL);
2103         }
2104         RETURN(0);
2105 }
2106
2107 static struct llog_operations filter_unlink_repl_logops;
2108 static struct llog_operations filter_size_orig_logops = {
2109         lop_setup: llog_obd_origin_setup,
2110         lop_cleanup: llog_obd_origin_cleanup,
2111         lop_add: llog_obd_origin_add
2112 };
2113
2114 static int filter_llog_init(struct obd_device *obd, struct obd_device *tgt,
2115                             int count, struct llog_logid *logid)
2116 {
2117         struct llog_ctxt *ctxt;
2118         int rc;
2119         ENTRY;
2120
2121         filter_unlink_repl_logops = llog_client_ops;
2122         filter_unlink_repl_logops.lop_cancel = llog_obd_repl_cancel;
2123         filter_unlink_repl_logops.lop_connect = llog_repl_connect;
2124         filter_unlink_repl_logops.lop_sync = llog_obd_repl_sync;
2125
2126         rc = llog_setup(obd, LLOG_UNLINK_REPL_CTXT, tgt, 0, NULL,
2127                         &filter_unlink_repl_logops);
2128         if (rc)
2129                 RETURN(rc);
2130         /* FIXME - assign unlink_cb for filter's recovery */
2131         ctxt = llog_get_context(obd, LLOG_UNLINK_REPL_CTXT);
2132         ctxt->llog_proc_cb = filter_recov_log_unlink_cb;
2133
2134         rc = llog_setup(obd, LLOG_SIZE_ORIG_CTXT, tgt, 0, NULL,
2135                         &filter_size_orig_logops);
2136         RETURN(rc);
2137 }
2138
2139 static int filter_llog_finish(struct obd_device *obd, int count)
2140 {
2141         int rc;
2142         ENTRY;
2143
2144         rc = llog_cleanup(llog_get_context(obd, LLOG_UNLINK_REPL_CTXT));
2145         if (rc)
2146                 RETURN(rc);
2147
2148         rc = llog_cleanup(llog_get_context(obd, LLOG_SIZE_ORIG_CTXT));
2149         RETURN(rc);
2150 }
2151
2152 static struct dentry *filter_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2153                                              void *data)
2154 {
2155         return filter_fid2dentry(data, NULL, gr, id);
2156 }
2157
2158 static struct lvfs_callback_ops filter_lvfs_ops = {
2159         l_fid2dentry:     filter_lvfs_fid2dentry,
2160 };
2161
2162 static struct obd_ops filter_obd_ops = {
2163         o_owner:          THIS_MODULE,
2164         o_attach:         filter_attach,
2165         o_detach:         filter_detach,
2166         o_get_info:       filter_get_info,
2167         o_set_info:       filter_set_info,
2168         o_setup:          filter_setup,
2169         o_postsetup:      filter_postsetup,
2170         o_precleanup:     filter_precleanup,
2171         o_cleanup:        filter_cleanup,
2172         o_connect:        filter_connect,
2173         o_disconnect:     filter_disconnect,
2174         o_statfs:         filter_statfs,
2175         o_getattr:        filter_getattr,
2176         o_unpackmd:       filter_unpackmd,
2177         o_create:         filter_create,
2178         o_setattr:        filter_setattr,
2179         o_destroy:        filter_destroy,
2180         o_brw:            filter_brw,
2181         o_punch:          filter_truncate,
2182         o_sync:           filter_sync,
2183         o_preprw:         filter_preprw,
2184         o_commitrw:       filter_commitrw,
2185         o_destroy_export: filter_destroy_export,
2186         o_llog_init:      filter_llog_init,
2187         o_llog_finish:    filter_llog_finish,
2188         o_iocontrol:      filter_iocontrol,
2189 };
2190
2191 static struct obd_ops filter_sanobd_ops = {
2192         o_owner:          THIS_MODULE,
2193         o_attach:         filter_attach,
2194         o_detach:         filter_detach,
2195         o_get_info:       filter_get_info,
2196         o_set_info:       filter_set_info,
2197         o_setup:          filter_san_setup,
2198         o_precleanup:     filter_precleanup,
2199         o_cleanup:        filter_cleanup,
2200         o_connect:        filter_connect,
2201         o_disconnect:     filter_disconnect,
2202         o_statfs:         filter_statfs,
2203         o_getattr:        filter_getattr,
2204         o_unpackmd:       filter_unpackmd,
2205         o_create:         filter_create,
2206         o_setattr:        filter_setattr,
2207         o_destroy:        filter_destroy,
2208         o_brw:            filter_brw,
2209         o_punch:          filter_truncate,
2210         o_sync:           filter_sync,
2211         o_preprw:         filter_preprw,
2212         o_commitrw:       filter_commitrw,
2213         o_san_preprw:     filter_san_preprw,
2214         o_destroy_export: filter_destroy_export,
2215         o_llog_init:      filter_llog_init,
2216         o_llog_finish:    filter_llog_finish,
2217         o_iocontrol:      filter_iocontrol,
2218 };
2219
2220 static int __init obdfilter_init(void)
2221 {
2222         struct lprocfs_static_vars lvars;
2223         int rc;
2224
2225         printk(KERN_INFO "Lustre: Filtering OBD driver; info@clusterfs.com\n");
2226
2227         lprocfs_init_vars(filter, &lvars);
2228
2229         rc = class_register_type(&filter_obd_ops, lvars.module_vars,
2230                                  OBD_FILTER_DEVICENAME);
2231         if (rc)
2232                 return rc;
2233
2234         rc = class_register_type(&filter_sanobd_ops, lvars.module_vars,
2235                                  OBD_FILTER_SAN_DEVICENAME);
2236         if (rc)
2237                 class_unregister_type(OBD_FILTER_DEVICENAME);
2238         return rc;
2239 }
2240
2241 static void __exit obdfilter_exit(void)
2242 {
2243         class_unregister_type(OBD_FILTER_SAN_DEVICENAME);
2244         class_unregister_type(OBD_FILTER_DEVICENAME);
2245 }
2246
2247 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2248 MODULE_DESCRIPTION("Lustre Filtering OBD driver");
2249 MODULE_LICENSE("GPL");
2250
2251 module_init(obdfilter_init);
2252 module_exit(obdfilter_exit);