lustre/mds/mds_unlink_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mds/mds_unlink_open.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Andreas Dilger <adilger@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  */
  42
  43 /* code for handling open unlinked files */
  44
  45 #define DEBUG_SUBSYSTEM S_MDS
  46
  47 #ifndef AUTOCONF_INCLUDED
  48 #include <linux/config.h>
  49 #endif
  50 #include <linux/module.h>
  51 #include <linux/version.h>
  52
  53 #include <libcfs/list.h>
  54 #include <obd_class.h>
  55 #include <lustre_fsfilt.h>
  56 #include <lustre_mds.h>
  57 #include <lvfs.h>
  58
  59 #include "mds_internal.h"
  60
  61 int mds_osc_destroy_orphan(struct obd_device *obd,
  62                            umode_t mode,
  63                            struct lov_mds_md *lmm,
  64                            int lmm_size,
  65                            struct llog_cookie *logcookies,
  66                            int log_unlink)
  67 {
  68         struct mds_obd *mds = &obd->u.mds;
  69         struct lov_stripe_md *lsm = NULL;
  70         struct obd_trans_info oti = { 0 };
  71         struct obdo *oa;
  72         int rc;
  73         ENTRY;
  74
  75         if (lmm_size == 0)
  76                 RETURN(0);
  77
  78         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
  79         if (rc < 0) {
  80                 CERROR("Error unpack md %p\n", lmm);
  81                 RETURN(rc);
  82         } else {
  83                 LASSERT(rc >= sizeof(*lsm));
  84                 rc = 0;
  85         }
  86
  87         rc = obd_checkmd(mds->mds_osc_exp, obd->obd_self_export, lsm);
  88         if (rc)
  89                 GOTO(out_free_memmd, rc);
  90
  91         OBDO_ALLOC(oa);
  92         if (oa == NULL)
  93                 GOTO(out_free_memmd, rc = -ENOMEM);
  94         oa->o_id = lsm->lsm_object_id;
  95         oa->o_mode = mode & S_IFMT;
  96         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE;
  97
  98         if (log_unlink && logcookies) {
  99                 oa->o_valid |= OBD_MD_FLCOOKIE;
 100                 oti.oti_logcookies = logcookies;
 101         }
 102         rc = obd_destroy(mds->mds_osc_exp, oa, lsm, &oti, obd->obd_self_export);
 103         OBDO_FREE(oa);
 104         if (rc)
 105                 CDEBUG(D_INODE, "destroy orphan objid 0x"LPX64" on ost error "
 106                        "%d\n", lsm->lsm_object_id, rc);
 107 out_free_memmd:
 108         obd_free_memmd(mds->mds_osc_exp, &lsm);
 109         RETURN(rc);
 110 }
 111
 112 static int mds_unlink_orphan(struct obd_device *obd, struct dentry *dchild,
 113                              struct inode *inode, struct inode *pending_dir)
 114 {
 115         struct mds_obd *mds = &obd->u.mds;
 116         struct lov_mds_md *lmm = NULL;
 117         struct llog_cookie *logcookies = NULL;
 118         int lmm_size, log_unlink = 0, cookie_size = 0;
 119         void *handle = NULL;
 120         umode_t mode;
 121         int rc, err;
 122         ENTRY;
 123
 124         LASSERT(mds->mds_osc_obd != NULL);
 125
 126         /* We don't need to do any of these other things for orhpan dirs,
 127          * especially not mds_get_md (may get a default LOV EA, bug 4554) */
 128         mode = inode->i_mode;
 129         if (S_ISDIR(mode)) {
 130                 rc = ll_vfs_rmdir(pending_dir, dchild, mds->mds_vfsmnt);
 131                 if (rc)
 132                         CERROR("error %d unlinking dir %*s from PENDING\n",
 133                                rc, dchild->d_name.len, dchild->d_name.name);
 134                 RETURN(rc);
 135         }
 136
 137         lmm_size = mds->mds_max_mdsize;
 138         OBD_ALLOC(lmm, lmm_size);
 139         if (lmm == NULL)
 140                 RETURN(-ENOMEM);
 141
 142         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0, 0);
 143         if (rc < 0)
 144                 GOTO(out_free_lmm, rc);
 145
 146         handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK, NULL,
 147                                   le32_to_cpu(lmm->lmm_stripe_count));
 148         if (IS_ERR(handle)) {
 149                 rc = PTR_ERR(handle);
 150                 CERROR("error fsfilt_start: %d\n", rc);
 151                 handle = NULL;
 152                 GOTO(out_free_lmm, rc);
 153         }
 154
 155         rc = ll_vfs_unlink(pending_dir, dchild, mds->mds_vfsmnt);
 156         if (rc) {
 157                 CERROR("error %d unlinking orphan %.*s from PENDING\n",
 158                        rc, dchild->d_name.len, dchild->d_name.name);
 159         } else if (lmm_size) {
 160                 cookie_size = mds_get_cookie_size(obd, lmm);
 161                 OBD_ALLOC(logcookies, cookie_size);
 162                 if (logcookies == NULL)
 163                         rc = -ENOMEM;
 164                 else if (mds_log_op_unlink(obd, lmm,lmm_size,logcookies,
 165                                            cookie_size) > 0)
 166                         log_unlink = 1;
 167         }
 168
 169         err = fsfilt_commit(obd, pending_dir, handle, 0);
 170         if (err) {
 171                 CERROR("error committing orphan unlink: %d\n", err);
 172                 if (!rc)
 173                         rc = err;
 174         } else if (!rc) {
 175                 rc = mds_osc_destroy_orphan(obd, mode, lmm, lmm_size,
 176                                             logcookies, log_unlink);
 177         }
 178
 179         if (logcookies != NULL)
 180                 OBD_FREE(logcookies, cookie_size);
 181 out_free_lmm:
 182         OBD_FREE(lmm, mds->mds_max_mdsize);
 183         RETURN(rc);
 184 }
 185
 186 static __u64 mds_orphans_max_version(struct obd_device *obd)
 187 {
 188         struct obd_export *exp;
 189         __u32 epoch = lr_epoch(obd->u.mds.mds_last_transno);
 190         spin_lock(&obd->obd_dev_lock);
 191         list_for_each_entry(exp, &obd->obd_delayed_exports, exp_obd_chain) {
 192                 struct lu_export_data *led = &exp->exp_target_data;
 193                 epoch = min(epoch, le32_to_cpu(led->led_lcd->lcd_first_epoch));
 194         }
 195         spin_unlock(&obd->obd_dev_lock);
 196         return (__u64)epoch << LR_EPOCH_BITS;
 197 }
 198
 199 /* Delete inodes which were previously open-unlinked but were not reopened
 200  * during MDS recovery for whatever reason (e.g. client also failed, recovery
 201  * aborted, etc). */
 202 int mds_cleanup_pending(struct obd_device *obd)
 203 {
 204         struct mds_obd *mds = &obd->u.mds;
 205         struct lvfs_run_ctxt saved;
 206         struct file *file;
 207         struct dentry *dchild, *dentry;
 208         struct vfsmount *mnt;
 209         struct inode *child_inode, *pending_dir = mds->mds_pending_dir->d_inode;
 210         struct l_linux_dirent *dirent, *n;
 211         struct list_head dentry_list;
 212         char d_name[LL_FID_NAMELEN];
 213         unsigned long inum;
 214         __u64 max_version;
 215         int i = 0, rc = 0, item = 0, namlen;
 216         ENTRY;
 217
 218         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 219         /* dentry and mnt ref dropped in dentry_open() on error, or
 220          * in filp_close() if dentry_open() succeeds */
 221         dentry = dget(mds->mds_pending_dir);
 222         if (IS_ERR(dentry))
 223                 GOTO(err_pop, rc = PTR_ERR(dentry));
 224         mnt = mntget(mds->mds_vfsmnt);
 225         if (IS_ERR(mnt))
 226                 GOTO(err_mntget, rc = PTR_ERR(mnt));
 227
 228         file = dentry_open(mds->mds_pending_dir, mds->mds_vfsmnt,
 229                            O_RDONLY | O_LARGEFILE);
 230         if (IS_ERR(file))
 231                 GOTO(err_pop, rc = PTR_ERR(file));
 232
 233         CFS_INIT_LIST_HEAD(&dentry_list);
 234         rc = l_readdir(file, &dentry_list);
 235         filp_close(file, 0);
 236         if (rc < 0)
 237                 GOTO(err_out, rc);
 238
 239         /** Get maximum version for orphans to delete. All other orphans may be
 240          *  needed for delayed clients */
 241         max_version = mds_orphans_max_version(obd);
 242
 243         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 244                 __u64 version;
 245
 246                 i++;
 247                 list_del(&dirent->lld_list);
 248
 249                 namlen = strlen(dirent->lld_name);
 250                 LASSERT(sizeof(d_name) >= namlen + 1);
 251                 strcpy(d_name, dirent->lld_name);
 252                 inum = dirent->lld_ino;
 253                 OBD_FREE_PTR(dirent);
 254
 255                 CDEBUG(D_INODE, "entry %d of PENDING DIR: %s\n", i, d_name);
 256
 257                 if (((namlen == 1) && !strcmp(d_name, ".")) ||
 258                     ((namlen == 2) && !strcmp(d_name, "..")) || inum == 0)
 259                         continue;
 260
 261                 LOCK_INODE_MUTEX(pending_dir);
 262                 dchild = lookup_one_len(d_name, mds->mds_pending_dir, namlen);
 263                 if (IS_ERR(dchild)) {
 264                         UNLOCK_INODE_MUTEX(pending_dir);
 265                         GOTO(err_out, rc = PTR_ERR(dchild));
 266                 }
 267                 if (!dchild->d_inode) {
 268                         CWARN("%s: orphan %s has already been removed\n",
 269                               obd->obd_name, d_name);
 270                         GOTO(next, rc = 0);
 271                 }
 272
 273                 if (is_bad_inode(dchild->d_inode)) {
 274                         CERROR("%s: bad orphan inode found %lu/%u\n",
 275                                obd->obd_name, dchild->d_inode->i_ino,
 276                                dchild->d_inode->i_generation);
 277                         GOTO(next, rc = -ENOENT);
 278                 }
 279
 280                 child_inode = dchild->d_inode;
 281                 MDS_DOWN_READ_ORPHAN_SEM(child_inode);
 282                 if (mds_inode_is_orphan(child_inode) &&
 283                     mds_orphan_open_count(child_inode)) {
 284                         MDS_UP_READ_ORPHAN_SEM(child_inode);
 285                         CWARN("%s: orphan %s re-opened during recovery\n",
 286                               obd->obd_name, d_name);
 287                         GOTO(next, rc = 0);
 288                 }
 289                 /** Keep orphans for possible use by delayed exports. Remove
 290                  * orphans with version lower than minimal one of all exports */
 291                 version = fsfilt_get_version(obd, child_inode);
 292                 if ((__s64)version != -EOPNOTSUPP &&
 293                     version >= max_version) {
 294                         MDS_UP_READ_ORPHAN_SEM(child_inode);
 295                         CDEBUG(D_INFO,
 296                                "%s: orphan %s is needed for delayed exports\n",
 297                                obd->obd_name, d_name);
 298                         GOTO(next, rc = 0);
 299                 }
 300                 MDS_UP_READ_ORPHAN_SEM(child_inode);
 301
 302                 rc = mds_unlink_orphan(obd, dchild, child_inode, pending_dir);
 303                 CDEBUG(D_INODE, "%s: removed orphan %s: rc %d\n",
 304                        obd->obd_name, d_name, rc);
 305                 if (rc == 0)
 306                         item++;
 307                 else
 308                         rc = 0;
 309 next:
 310                 l_dput(dchild);
 311                 UNLOCK_INODE_MUTEX(pending_dir);
 312         }
 313         rc = 0;
 314 err_out:
 315         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 316                 list_del(&dirent->lld_list);
 317                 OBD_FREE(dirent, sizeof(*dirent));
 318         }
 319 err_pop:
 320         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 321         if (item > 0)
 322                 CWARN("%s: removed %d pending open-unlinked files\n",
 323                       obd->obd_name, item);
 324         RETURN(rc);
 325
 326 err_mntget:
 327         l_dput(mds->mds_pending_dir);
 328         goto err_pop;
 329 }
 330
 331 /**
 332  * Determine there is no orphan with the same inode number. That may happens
 333  * since unlink replay don't delete inode but keep orphan for delayed clients.
 334  * Therefore replays like 'create, unlink, create' will fail due to inode can't
 335  * be reused.
 336  */
 337 int mds_check_stale_orphan(struct obd_device *obd, struct ll_fid *fid)
 338 {
 339         struct mds_obd *mds = &obd->u.mds;
 340         char fidname[32];
 341         struct dentry *result;
 342         struct inode *inode, *pending_dir = mds->mds_pending_dir->d_inode;
 343         int fidlen = 0, rc = 0;
 344
 345         /* no need in checks*/
 346         if (fid->id == 0 || obd->obd_recovering == 0)
 347                 RETURN(0);
 348
 349         /** open by fid like mds_fid2dentry does */
 350         snprintf(fidname, sizeof(fidname), "0x%lx", (unsigned long)(fid->id));
 351         fidlen = strlen(fidname);
 352         result = mds_lookup(obd, fidname, mds->mds_fid_de, fidlen);
 353         if (IS_ERR(result))
 354                 RETURN(0);
 355         inode = result->d_inode;
 356         if (!inode)
 357                 GOTO(out, rc = 0);
 358
 359         LOCK_INODE_MUTEX(pending_dir);
 360         MDS_DOWN_READ_ORPHAN_SEM(inode);
 361         if (mds_inode_is_orphan(inode)) {
 362                 struct dentry *orphan;
 363
 364                 /* The exactly same inode can't be orphan */
 365                 LASSERT(inode->i_generation != fid->generation);
 366
 367                 if (mds_orphan_open_count(inode) > 0) {
 368                         CERROR("Orphan "LPU64"/%u is in use!\n",
 369                                fid->id, fid->generation);
 370                         GOTO(unlock_child, rc = -EFAULT);
 371                 }
 372
 373                 /** Found orphan in pending dir and delete it */
 374                 fidlen = ll_fid2str(fidname, fid->id, inode->i_generation);
 375                 orphan = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
 376                 if (IS_ERR(orphan)) {
 377                         rc = PTR_ERR(orphan);
 378                         CERROR("error looking up %s in PENDING: rc = %d\n",
 379                                 fidname, rc);
 380                         GOTO(unlock_child, rc);
 381                 }
 382                 if (orphan->d_inode != inode) {
 383                         l_dput(orphan);
 384                         CWARN("%s: Found wrong orphan %s %p/%p\n",
 385                               obd->obd_name, fidname, orphan->d_inode, inode);
 386                         GOTO(unlock_child, rc = -EFAULT);
 387                 }
 388                 MDS_UP_READ_ORPHAN_SEM(inode);
 389
 390                 rc = mds_unlink_orphan(obd, orphan, inode, pending_dir);
 391                 CDEBUG(D_INODE, "%s: removed orphan %s: rc %d\n",
 392                        obd->obd_name, fidname, rc);
 393                 l_dput(orphan);
 394                 GOTO(unlock, rc);
 395         }
 396 unlock_child:
 397         MDS_UP_READ_ORPHAN_SEM(inode);
 398 unlock:
 399         UNLOCK_INODE_MUTEX(pending_dir);
 400 out:
 401         l_dput(result);
 402         RETURN(0);
 403 }