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  *  lustre/mds/mds_orphan.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  *   Author: Phil Schwan <phil@clusterfs.com>
  10  *
  11  *   This file is part of the Lustre file system, http://www.lustre.org
  12  *   Lustre is a trademark of Cluster File Systems, Inc.
  13  *
  14  *   You may have signed or agreed to another license before downloading
  15  *   this software.  If so, you are bound by the terms and conditions
  16  *   of that agreement, and the following does not apply to you.  See the
  17  *   LICENSE file included with this distribution for more information.
  18  *
  19  *   If you did not agree to a different license, then this copy of Lustre
  20  *   is open source software; you can redistribute it and/or modify it
  21  *   under the terms of version 2 of the GNU General Public License as
  22  *   published by the Free Software Foundation.
  23  *
  24  *   In either case, Lustre is distributed in the hope that it will be
  25  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  26  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  27  *   license text for more details.
  28  */
  29
  30 /* code for handling open unlinked files */
  31
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/config.h>
  35 #include <linux/module.h>
  36 #include <linux/version.h>
  37
  38 #include <libcfs/list.h>
  39 #include <obd_class.h>
  40 #include <lustre_fsfilt.h>
  41 #include <lustre_mds.h>
  42 #include <lustre_commit_confd.h>
  43 #include <lvfs.h>
  44
  45 #include "mds_internal.h"
  46
  47 static int mds_osc_destroy_orphan(struct obd_device *obd,
  48                                   struct inode *inode,
  49                                   struct lov_mds_md *lmm,
  50                                   int lmm_size,
  51                                   struct llog_cookie *logcookies,
  52                                   int log_unlink)
  53 {
  54         struct mds_obd *mds = &obd->u.mds;
  55         struct lov_stripe_md *lsm = NULL;
  56         struct obd_trans_info oti = { 0 };
  57         struct obdo *oa;
  58         int rc;
  59         ENTRY;
  60
  61         if (lmm_size == 0)
  62                 RETURN(0);
  63
  64         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
  65         if (rc < 0) {
  66                 CERROR("Error unpack md %p\n", lmm);
  67                 RETURN(rc);
  68         } else {
  69                 LASSERT(rc >= sizeof(*lsm));
  70                 rc = 0;
  71         }
  72
  73         rc = obd_checkmd(mds->mds_osc_exp, obd->obd_self_export, lsm);
  74         if (rc)
  75                 GOTO(out_free_memmd, rc);
  76
  77         oa = obdo_alloc();
  78         if (oa == NULL)
  79                 GOTO(out_free_memmd, rc = -ENOMEM);
  80         oa->o_id = lsm->lsm_object_id;
  81         oa->o_mode = inode->i_mode & S_IFMT;
  82         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE;
  83
  84         if (log_unlink && logcookies) {
  85                 oa->o_valid |= OBD_MD_FLCOOKIE;
  86                 oti.oti_logcookies = logcookies;
  87         }
  88         rc = obd_destroy(mds->mds_osc_exp, oa, lsm, &oti, obd->obd_self_export);
  89         obdo_free(oa);
  90         if (rc)
  91                 CDEBUG(D_INODE, "destroy orphan objid 0x"LPX64" on ost error "
  92                        "%d\n", lsm->lsm_object_id, rc);
  93 out_free_memmd:
  94         obd_free_memmd(mds->mds_osc_exp, &lsm);
  95         RETURN(rc);
  96 }
  97
  98 static int mds_unlink_orphan(struct obd_device *obd, struct dentry *dchild,
  99                              struct inode *inode, struct inode *pending_dir)
 100 {
 101         struct mds_obd *mds = &obd->u.mds;
 102         struct lov_mds_md *lmm = NULL;
 103         struct llog_cookie *logcookies = NULL;
 104         int lmm_size, log_unlink = 0, cookie_size = 0;
 105         void *handle = NULL;
 106         int rc, err;
 107         ENTRY;
 108
 109         LASSERT(mds->mds_osc_obd != NULL);
 110
 111         /* We don't need to do any of these other things for orhpan dirs,
 112          * especially not mds_get_md (may get a default LOV EA, bug 4554) */
 113         if (S_ISDIR(inode->i_mode)) {
 114                 rc = vfs_rmdir(pending_dir, dchild);
 115                 if (rc)
 116                         CERROR("error %d unlinking dir %*s from PENDING\n",
 117                                rc, dchild->d_name.len, dchild->d_name.name);
 118                 RETURN(rc);
 119         }
 120
 121         lmm_size = mds->mds_max_mdsize;
 122         OBD_ALLOC(lmm, lmm_size);
 123         if (lmm == NULL)
 124                 RETURN(-ENOMEM);
 125
 126         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1);
 127         if (rc < 0)
 128                 GOTO(out_free_lmm, rc);
 129
 130         handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK, NULL,
 131                                   le32_to_cpu(lmm->lmm_stripe_count));
 132         if (IS_ERR(handle)) {
 133                 rc = PTR_ERR(handle);
 134                 CERROR("error fsfilt_start: %d\n", rc);
 135                 handle = NULL;
 136                 GOTO(out_free_lmm, rc);
 137         }
 138
 139         rc = vfs_unlink(pending_dir, dchild);
 140         if (rc) {
 141                 CERROR("error %d unlinking orphan %.*s from PENDING\n",
 142                        rc, dchild->d_name.len, dchild->d_name.name);
 143         } else if (lmm_size) {
 144                 cookie_size = mds_get_cookie_size(obd, lmm);
 145                 OBD_ALLOC(logcookies, cookie_size);
 146                 if (logcookies == NULL)
 147                         rc = -ENOMEM;
 148                 else if (mds_log_op_unlink(obd, inode, lmm,lmm_size,logcookies,
 149                                            cookie_size) > 0)
 150                         log_unlink = 1;
 151         }
 152
 153         err = fsfilt_commit(obd, pending_dir, handle, 0);
 154         if (err) {
 155                 CERROR("error committing orphan unlink: %d\n", err);
 156                 if (!rc)
 157                         rc = err;
 158         } else if (!rc) {
 159                 rc = mds_osc_destroy_orphan(obd, inode, lmm, lmm_size,
 160                                             logcookies, log_unlink);
 161         }
 162
 163         if (logcookies != NULL)
 164                 OBD_FREE(logcookies, cookie_size);
 165 out_free_lmm:
 166         OBD_FREE(lmm, mds->mds_max_mdsize);
 167         RETURN(rc);
 168 }
 169
 170 /* Delete inodes which were previously open-unlinked but were not reopened
 171  * during MDS recovery for whatever reason (e.g. client also failed, recovery
 172  * aborted, etc). */
 173 int mds_cleanup_pending(struct obd_device *obd)
 174 {
 175         struct mds_obd *mds = &obd->u.mds;
 176         struct lvfs_run_ctxt saved;
 177         struct file *file;
 178         struct dentry *dchild, *dentry;
 179         struct vfsmount *mnt;
 180         struct inode *child_inode, *pending_dir = mds->mds_pending_dir->d_inode;
 181         struct l_linux_dirent *dirent, *n;
 182         struct list_head dentry_list;
 183         char d_name[LL_FID_NAMELEN];
 184         unsigned long inum;
 185         int i = 0, rc = 0, item = 0, namlen;
 186         ENTRY;
 187
 188         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 189         /* dentry and mnt ref dropped in dentry_open() on error, or
 190          * in filp_close() if dentry_open() succeeds */
 191         dentry = dget(mds->mds_pending_dir);
 192         if (IS_ERR(dentry))
 193                 GOTO(err_pop, rc = PTR_ERR(dentry));
 194         mnt = mntget(mds->mds_vfsmnt);
 195         if (IS_ERR(mnt))
 196                 GOTO(err_mntget, rc = PTR_ERR(mnt));
 197
 198         file = dentry_open(mds->mds_pending_dir, mds->mds_vfsmnt,
 199                            O_RDONLY | O_LARGEFILE);
 200         if (IS_ERR(file))
 201                 GOTO(err_pop, rc = PTR_ERR(file));
 202
 203         INIT_LIST_HEAD(&dentry_list);
 204         rc = l_readdir(file, &dentry_list);
 205         filp_close(file, 0);
 206         if (rc < 0)
 207                 GOTO(err_out, rc);
 208
 209         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 210                 i++;
 211                 list_del(&dirent->lld_list);
 212
 213                 namlen = strlen(dirent->lld_name);
 214                 LASSERT(sizeof(d_name) >= namlen + 1);
 215                 strcpy(d_name, dirent->lld_name);
 216                 inum = dirent->lld_ino;
 217                 OBD_FREE(dirent, sizeof(*dirent));
 218
 219                 CDEBUG(D_INODE, "entry %d of PENDING DIR: %s\n", i, d_name);
 220
 221                 if (((namlen == 1) && !strcmp(d_name, ".")) ||
 222                     ((namlen == 2) && !strcmp(d_name, "..")) || inum == 0)
 223                         continue;
 224
 225                 LOCK_INODE_MUTEX(pending_dir);
 226                 dchild = lookup_one_len(d_name, mds->mds_pending_dir, namlen);
 227                 if (IS_ERR(dchild)) {
 228                         UNLOCK_INODE_MUTEX(pending_dir);
 229                         GOTO(err_out, rc = PTR_ERR(dchild));
 230                 }
 231                 if (!dchild->d_inode) {
 232                         CWARN("%s: orphan %s has already been removed\n",
 233                               obd->obd_name, d_name);
 234                         GOTO(next, rc = 0);
 235                 }
 236
 237                 if (is_bad_inode(dchild->d_inode)) {
 238                         CERROR("%s: bad orphan inode found %lu/%u\n",
 239                                obd->obd_name, dchild->d_inode->i_ino,
 240                                dchild->d_inode->i_generation);
 241                         GOTO(next, rc = -ENOENT);
 242                 }
 243
 244                 child_inode = dchild->d_inode;
 245                 MDS_DOWN_READ_ORPHAN_SEM(child_inode);
 246                 if (mds_inode_is_orphan(child_inode) &&
 247                     mds_orphan_open_count(child_inode)) {
 248                         MDS_UP_READ_ORPHAN_SEM(child_inode);
 249                         CWARN("%s: orphan %s re-opened during recovery\n",
 250                               obd->obd_name, d_name);
 251                         GOTO(next, rc = 0);
 252                 }
 253                 MDS_UP_READ_ORPHAN_SEM(child_inode);
 254
 255                 rc = mds_unlink_orphan(obd, dchild, child_inode, pending_dir);
 256                 if (rc == 0) {
 257                         item ++;
 258                         CDEBUG(D_HA, "%s: removed orphan %s\n",
 259                                obd->obd_name, d_name);
 260                 } else {
 261                         CDEBUG(D_INODE, "%s: removed orphan %s failed,"
 262                                " rc = %d\n", obd->obd_name, d_name, rc);
 263                         rc = 0;
 264                 }
 265 next:
 266                 l_dput(dchild);
 267                 UNLOCK_INODE_MUTEX(pending_dir);
 268         }
 269         rc = 0;
 270 err_out:
 271         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 272                 list_del(&dirent->lld_list);
 273                 OBD_FREE(dirent, sizeof(*dirent));
 274         }
 275 err_pop:
 276         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 277         if (item > 0)
 278                 CWARN("%s: removed %d pending open-unlinked files\n",
 279                       obd->obd_name, item);
 280         RETURN(rc);
 281
 282 err_mntget:
 283         l_dput(mds->mds_pending_dir);
 284         goto err_pop;
 285 }