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 #ifdef HAVE_KERNEL_CONFIG_H
  35 #include <linux/config.h>
  36 #endif
  37 #include <linux/module.h>
  38 #include <linux/version.h>
  39
  40 #include <libcfs/list.h>
  41 #include <obd_class.h>
  42 #include <lustre_fsfilt.h>
  43 #include <lustre_mds.h>
  44 #include <lustre_commit_confd.h>
  45 #include <lvfs.h>
  46
  47 #include "mds_internal.h"
  48
  49 int mds_osc_destroy_orphan(struct obd_device *obd,
  50                            umode_t mode,
  51                            struct lov_mds_md *lmm,
  52                            int lmm_size,
  53                            struct llog_cookie *logcookies,
  54                            int log_unlink)
  55 {
  56         struct mds_obd *mds = &obd->u.mds;
  57         struct lov_stripe_md *lsm = NULL;
  58         struct obd_trans_info oti = { 0 };
  59         struct obdo *oa;
  60         int rc;
  61         ENTRY;
  62
  63         if (lmm_size == 0)
  64                 RETURN(0);
  65
  66         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
  67         if (rc < 0) {
  68                 CERROR("Error unpack md %p\n", lmm);
  69                 RETURN(rc);
  70         } else {
  71                 LASSERT(rc >= sizeof(*lsm));
  72                 rc = 0;
  73         }
  74
  75         rc = obd_checkmd(mds->mds_osc_exp, obd->obd_self_export, lsm);
  76         if (rc)
  77                 GOTO(out_free_memmd, rc);
  78
  79         oa = obdo_alloc();
  80         if (oa == NULL)
  81                 GOTO(out_free_memmd, rc = -ENOMEM);
  82         oa->o_id = lsm->lsm_object_id;
  83         oa->o_mode = mode & S_IFMT;
  84         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE;
  85
  86         if (log_unlink && logcookies) {
  87                 oa->o_valid |= OBD_MD_FLCOOKIE;
  88                 oti.oti_logcookies = logcookies;
  89         }
  90         rc = obd_destroy(mds->mds_osc_exp, oa, lsm, &oti, obd->obd_self_export);
  91         obdo_free(oa);
  92         if (rc)
  93                 CDEBUG(D_INODE, "destroy orphan objid 0x"LPX64" on ost error "
  94                        "%d\n", lsm->lsm_object_id, rc);
  95 out_free_memmd:
  96         obd_free_memmd(mds->mds_osc_exp, &lsm);
  97         RETURN(rc);
  98 }
  99
 100 static int mds_unlink_orphan(struct obd_device *obd, struct dentry *dchild,
 101                              struct inode *inode, struct inode *pending_dir)
 102 {
 103         struct mds_obd *mds = &obd->u.mds;
 104         struct lov_mds_md *lmm = NULL;
 105         struct llog_cookie *logcookies = NULL;
 106         int lmm_size, log_unlink = 0, cookie_size = 0;
 107         void *handle = NULL;
 108         umode_t mode;
 109         int rc, err;
 110         ENTRY;
 111
 112         LASSERT(mds->mds_osc_obd != NULL);
 113
 114         /* We don't need to do any of these other things for orhpan dirs,
 115          * especially not mds_get_md (may get a default LOV EA, bug 4554) */
 116         mode = inode->i_mode;
 117         if (S_ISDIR(mode)) {
 118                 rc = vfs_rmdir(pending_dir, dchild);
 119                 if (rc)
 120                         CERROR("error %d unlinking dir %*s from PENDING\n",
 121                                rc, dchild->d_name.len, dchild->d_name.name);
 122                 RETURN(rc);
 123         }
 124
 125         lmm_size = mds->mds_max_mdsize;
 126         OBD_ALLOC(lmm, lmm_size);
 127         if (lmm == NULL)
 128                 RETURN(-ENOMEM);
 129
 130         rc = mds_get_md(obd, inode, lmm, &lmm_size, 1);
 131         if (rc < 0)
 132                 GOTO(out_free_lmm, rc);
 133
 134         handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK, NULL,
 135                                   le32_to_cpu(lmm->lmm_stripe_count));
 136         if (IS_ERR(handle)) {
 137                 rc = PTR_ERR(handle);
 138                 CERROR("error fsfilt_start: %d\n", rc);
 139                 handle = NULL;
 140                 GOTO(out_free_lmm, rc);
 141         }
 142
 143         rc = vfs_unlink(pending_dir, dchild);
 144         if (rc) {
 145                 CERROR("error %d unlinking orphan %.*s from PENDING\n",
 146                        rc, dchild->d_name.len, dchild->d_name.name);
 147         } else if (lmm_size) {
 148                 cookie_size = mds_get_cookie_size(obd, lmm);
 149                 OBD_ALLOC(logcookies, cookie_size);
 150                 if (logcookies == NULL)
 151                         rc = -ENOMEM;
 152                 else if (mds_log_op_unlink(obd, lmm,lmm_size,logcookies,
 153                                            cookie_size) > 0)
 154                         log_unlink = 1;
 155         }
 156
 157         err = fsfilt_commit(obd, pending_dir, handle, 0);
 158         if (err) {
 159                 CERROR("error committing orphan unlink: %d\n", err);
 160                 if (!rc)
 161                         rc = err;
 162         } else if (!rc) {
 163                 rc = mds_osc_destroy_orphan(obd, mode, lmm, lmm_size,
 164                                             logcookies, log_unlink);
 165         }
 166
 167         if (logcookies != NULL)
 168                 OBD_FREE(logcookies, cookie_size);
 169 out_free_lmm:
 170         OBD_FREE(lmm, mds->mds_max_mdsize);
 171         RETURN(rc);
 172 }
 173
 174 /* Delete inodes which were previously open-unlinked but were not reopened
 175  * during MDS recovery for whatever reason (e.g. client also failed, recovery
 176  * aborted, etc). */
 177 int mds_cleanup_pending(struct obd_device *obd)
 178 {
 179         struct mds_obd *mds = &obd->u.mds;
 180         struct lvfs_run_ctxt saved;
 181         struct file *file;
 182         struct dentry *dchild, *dentry;
 183         struct vfsmount *mnt;
 184         struct inode *child_inode, *pending_dir = mds->mds_pending_dir->d_inode;
 185         struct l_linux_dirent *dirent, *n;
 186         struct list_head dentry_list;
 187         char d_name[LL_FID_NAMELEN];
 188         unsigned long inum;
 189         int i = 0, rc = 0, item = 0, namlen;
 190         ENTRY;
 191
 192         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 193         /* dentry and mnt ref dropped in dentry_open() on error, or
 194          * in filp_close() if dentry_open() succeeds */
 195         dentry = dget(mds->mds_pending_dir);
 196         if (IS_ERR(dentry))
 197                 GOTO(err_pop, rc = PTR_ERR(dentry));
 198         mnt = mntget(mds->mds_vfsmnt);
 199         if (IS_ERR(mnt))
 200                 GOTO(err_mntget, rc = PTR_ERR(mnt));
 201
 202         file = dentry_open(mds->mds_pending_dir, mds->mds_vfsmnt,
 203                            O_RDONLY | O_LARGEFILE);
 204         if (IS_ERR(file))
 205                 GOTO(err_pop, rc = PTR_ERR(file));
 206
 207         INIT_LIST_HEAD(&dentry_list);
 208         rc = l_readdir(file, &dentry_list);
 209         filp_close(file, 0);
 210         if (rc < 0)
 211                 GOTO(err_out, rc);
 212
 213         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 214                 i++;
 215                 list_del(&dirent->lld_list);
 216
 217                 namlen = strlen(dirent->lld_name);
 218                 LASSERT(sizeof(d_name) >= namlen + 1);
 219                 strcpy(d_name, dirent->lld_name);
 220                 inum = dirent->lld_ino;
 221                 OBD_FREE(dirent, sizeof(*dirent));
 222
 223                 CDEBUG(D_INODE, "entry %d of PENDING DIR: %s\n", i, d_name);
 224
 225                 if (((namlen == 1) && !strcmp(d_name, ".")) ||
 226                     ((namlen == 2) && !strcmp(d_name, "..")) || inum == 0)
 227                         continue;
 228
 229                 LOCK_INODE_MUTEX(pending_dir);
 230                 dchild = lookup_one_len(d_name, mds->mds_pending_dir, namlen);
 231                 if (IS_ERR(dchild)) {
 232                         UNLOCK_INODE_MUTEX(pending_dir);
 233                         GOTO(err_out, rc = PTR_ERR(dchild));
 234                 }
 235                 if (!dchild->d_inode) {
 236                         CWARN("%s: orphan %s has already been removed\n",
 237                               obd->obd_name, d_name);
 238                         GOTO(next, rc = 0);
 239                 }
 240
 241                 if (is_bad_inode(dchild->d_inode)) {
 242                         CERROR("%s: bad orphan inode found %lu/%u\n",
 243                                obd->obd_name, dchild->d_inode->i_ino,
 244                                dchild->d_inode->i_generation);
 245                         GOTO(next, rc = -ENOENT);
 246                 }
 247
 248                 child_inode = dchild->d_inode;
 249                 MDS_DOWN_READ_ORPHAN_SEM(child_inode);
 250                 if (mds_inode_is_orphan(child_inode) &&
 251                     mds_orphan_open_count(child_inode)) {
 252                         MDS_UP_READ_ORPHAN_SEM(child_inode);
 253                         CWARN("%s: orphan %s re-opened during recovery\n",
 254                               obd->obd_name, d_name);
 255                         GOTO(next, rc = 0);
 256                 }
 257                 MDS_UP_READ_ORPHAN_SEM(child_inode);
 258
 259                 rc = mds_unlink_orphan(obd, dchild, child_inode, pending_dir);
 260                 if (rc == 0) {
 261                         item ++;
 262                         CDEBUG(D_HA, "%s: removed orphan %s\n",
 263                                obd->obd_name, d_name);
 264                 } else {
 265                         CDEBUG(D_INODE, "%s: removed orphan %s failed,"
 266                                " rc = %d\n", obd->obd_name, d_name, rc);
 267                         rc = 0;
 268                 }
 269 next:
 270                 l_dput(dchild);
 271                 UNLOCK_INODE_MUTEX(pending_dir);
 272         }
 273         rc = 0;
 274 err_out:
 275         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 276                 list_del(&dirent->lld_list);
 277                 OBD_FREE(dirent, sizeof(*dirent));
 278         }
 279 err_pop:
 280         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 281         if (item > 0)
 282                 CWARN("%s: removed %d pending open-unlinked files\n",
 283                       obd->obd_name, item);
 284         RETURN(rc);
 285
 286 err_mntget:
 287         l_dput(mds->mds_pending_dir);
 288         goto err_pop;
 289 }