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 Lustre, http://www.lustre.org.
  12  *
  13  *   Lustre is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Lustre is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Lustre; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 /* code for handling open unlinked files */
  28
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/config.h>
  32 #include <linux/module.h>
  33 #include <linux/version.h>
  34
  35 #include <portals/list.h>
  36 #include <linux/obd_class.h>
  37 #include <linux/lustre_fsfilt.h>
  38 #include <linux/lustre_commit_confd.h>
  39 #include <linux/lvfs.h>
  40
  41 #include "mds_internal.h"
  42
  43
  44 /* If we are unlinking an open file/dir (i.e. creating an orphan) then
  45  * we instead link the inode into the PENDING directory until it is
  46  * finally released.  We can't simply call mds_reint_rename() or some
  47  * part thereof, because we don't have the inode to check for link
  48  * count/open status until after it is locked.
  49  *
  50  * For lock ordering, we always get the PENDING, then pending_child lock
  51  * last to avoid deadlocks.
  52  */
  53
  54 int mds_open_unlink_rename(struct mds_update_record *rec,
  55                            struct obd_device *obd, struct dentry *dparent,
  56                            struct dentry *dchild, void **handle)
  57 {
  58         struct mds_obd *mds = &obd->u.mds;
  59         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
  60         struct dentry *pending_child;
  61         char fidname[LL_FID_NAMELEN];
  62         int fidlen = 0, rc;
  63         ENTRY;
  64
  65         LASSERT(!mds_inode_is_orphan(dchild->d_inode));
  66
  67         down(&pending_dir->i_sem);
  68         fidlen = ll_fid2str(fidname, dchild->d_inode->i_ino,
  69                             dchild->d_inode->i_generation);
  70
  71         CDEBUG(D_HA, "pending destroy of %dx open file %s = %s\n",
  72                mds_open_orphan_count(dchild->d_inode), rec->ur_name, fidname);
  73
  74         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
  75         if (IS_ERR(pending_child))
  76                 GOTO(out_lock, rc = PTR_ERR(pending_child));
  77
  78         if (pending_child->d_inode != NULL) {
  79                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
  80                 LASSERT(pending_child->d_inode == dchild->d_inode);
  81                 GOTO(out_dput, rc = 0);
  82         }
  83
  84         *handle = fsfilt_start(obd, pending_dir, FSFILT_OP_RENAME, NULL);
  85         if (IS_ERR(*handle))
  86                 GOTO(out_dput, rc = PTR_ERR(*handle));
  87
  88         lock_kernel();
  89         rc = vfs_rename(dparent->d_inode, dchild, pending_dir, pending_child);
  90         unlock_kernel();
  91         if (rc)
  92                 CERROR("error renaming orphan %lu/%s to PENDING: rc = %d\n",
  93                        dparent->d_inode->i_ino, rec->ur_name, rc);
  94         else
  95                 mds_inode_set_orphan(dchild->d_inode);
  96 out_dput:
  97         dput(pending_child);
  98 out_lock:
  99         up(&pending_dir->i_sem);
 100         RETURN(rc);
 101 }
 102
 103 static int mds_osc_destroy_orphan(struct mds_obd *mds,
 104                                   struct inode *inode,
 105                                   struct lov_mds_md *lmm,
 106                                   int lmm_size,
 107                                   struct llog_cookie *logcookies,
 108                                   int log_unlink)
 109 {
 110         struct lov_stripe_md *lsm = NULL;
 111         struct obd_trans_info oti = { 0 };
 112         struct obdo *oa;
 113         int rc;
 114         ENTRY;
 115
 116         if (lmm_size == 0)
 117                 RETURN(0);
 118
 119         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
 120         if (rc < 0) {
 121                 CERROR("Error unpack md %p\n", lmm);
 122                 RETURN(rc);
 123         } else {
 124                 LASSERT(rc >= sizeof(*lsm));
 125                 rc = 0;
 126         }
 127
 128         oa = obdo_alloc();
 129         if (oa == NULL)
 130                 GOTO(out_free_memmd, rc = -ENOMEM);
 131         oa->o_id = lsm->lsm_object_id;
 132         oa->o_mode = inode->i_mode & S_IFMT;
 133         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE;
 134
 135         if (log_unlink && logcookies) {
 136                 oa->o_valid |= OBD_MD_FLCOOKIE;
 137                 oti.oti_logcookies = logcookies;
 138         }
 139
 140         rc = obd_destroy(mds->mds_osc_exp, oa, lsm, &oti);
 141         obdo_free(oa);
 142         if (rc)
 143                 CERROR("destroy orphan objid 0x"LPX64" on ost error "
 144                        "%d\n", lsm->lsm_object_id, rc);
 145 out_free_memmd:
 146         obd_free_memmd(mds->mds_osc_exp, &lsm);
 147         RETURN(rc);
 148 }
 149
 150 static int mds_unlink_orphan(struct obd_device *obd, struct dentry *dchild,
 151                              struct inode *inode, struct inode *pending_dir)
 152 {
 153         struct mds_obd *mds = &obd->u.mds;
 154         struct lov_mds_md *lmm = NULL;
 155         struct llog_cookie *logcookies = NULL;
 156         int lmm_size = 0, log_unlink = 0;
 157         void *handle = NULL;
 158         int rc, err;
 159         ENTRY;
 160
 161         LASSERT(mds->mds_osc_obd != NULL);
 162
 163         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 164         if (lmm == NULL)
 165                 RETURN(-ENOMEM);
 166
 167         down(&inode->i_sem);
 168         rc = fsfilt_get_md(obd, inode, lmm, mds->mds_max_mdsize);
 169         up(&inode->i_sem);
 170
 171         if (rc < 0) {
 172                 CERROR("Error %d reading eadata for ino %lu\n",
 173                        rc, inode->i_ino);
 174                 GOTO(out_free_lmm, rc);
 175         } else if (rc > 0) {
 176                 lmm_size = rc;
 177                 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
 178                 if (rc > 0)
 179                         lmm_size = rc;
 180                 rc = 0;
 181         }
 182
 183         handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK, NULL,
 184                                   le32_to_cpu(lmm->lmm_stripe_count));
 185         if (IS_ERR(handle)) {
 186                 rc = PTR_ERR(handle);
 187                 CERROR("error fsfilt_start: %d\n", rc);
 188                 handle = NULL;
 189                 GOTO(out_free_lmm, rc);
 190         }
 191
 192         down(&inode->i_sem);
 193         rc = fsfilt_get_md(obd, inode, lmm, mds->mds_max_mdsize);
 194         up(&inode->i_sem);
 195
 196         if (rc < 0) {
 197                 CERROR("Error %d reading eadata for ino %lu\n",
 198                        rc, inode->i_ino);
 199                 GOTO(out_free_lmm, rc);
 200         } else if (rc > 0) {
 201                 lmm_size = rc;
 202                 rc = 0;
 203         }
 204
 205         if (S_ISDIR(inode->i_mode))
 206                 rc = vfs_rmdir(pending_dir, dchild);
 207         else
 208                 rc = vfs_unlink(pending_dir, dchild);
 209
 210         if (rc)
 211                 CERROR("error %d unlinking orphan %*s from PENDING directory\n",
 212                        rc, dchild->d_name.len, dchild->d_name.name);
 213
 214         if (!rc && lmm_size) {
 215                 OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
 216                 if (logcookies == NULL)
 217                         rc = -ENOMEM;
 218                 else if (mds_log_op_unlink(obd, inode, lmm,lmm_size,logcookies,
 219                                            mds->mds_max_cookiesize) > 0)
 220                         log_unlink = 1;
 221         }
 222         err = fsfilt_commit(obd, pending_dir, handle, 0);
 223         if (err) {
 224                 CERROR("error committing orphan unlink: %d\n", err);
 225                 if (!rc)
 226                         rc = err;
 227         }
 228         if (!rc) {
 229                 rc = mds_osc_destroy_orphan(mds, inode, lmm, lmm_size,
 230                                             logcookies, log_unlink);
 231         }
 232
 233         if (logcookies != NULL)
 234                 OBD_FREE(logcookies, mds->mds_max_cookiesize);
 235 out_free_lmm:
 236         OBD_FREE(lmm, mds->mds_max_mdsize);
 237         RETURN(rc);
 238 }
 239
 240 int mds_cleanup_orphans(struct obd_device *obd)
 241 {
 242         struct mds_obd *mds = &obd->u.mds;
 243         struct obd_run_ctxt saved;
 244         struct file *file;
 245         struct dentry *dchild, *dentry;
 246         struct vfsmount *mnt;
 247         struct inode *child_inode, *pending_dir = mds->mds_pending_dir->d_inode;
 248         struct l_linux_dirent *dirent, *n;
 249         struct list_head dentry_list;
 250         char d_name[LL_FID_NAMELEN];
 251         __u64 i = 0;
 252         int rc = 0, item = 0, namlen;
 253         ENTRY;
 254
 255         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 256         dentry = dget(mds->mds_pending_dir);
 257         if (IS_ERR(dentry))
 258                 GOTO(err_pop, rc = PTR_ERR(dentry));
 259         mnt = mntget(mds->mds_vfsmnt);
 260         if (IS_ERR(mnt))
 261                 GOTO(err_mntget, rc = PTR_ERR(mnt));
 262
 263         file = dentry_open(mds->mds_pending_dir, mds->mds_vfsmnt,
 264                            O_RDONLY | O_LARGEFILE);
 265         if (IS_ERR(file))
 266                 GOTO(err_pop, rc = PTR_ERR(file));
 267
 268         INIT_LIST_HEAD(&dentry_list);
 269         rc = l_readdir(file, &dentry_list);
 270         filp_close(file, 0);
 271         if (rc < 0)
 272                 GOTO(err_out, rc);
 273
 274         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 275                 i ++;
 276                 list_del(&dirent->lld_list);
 277
 278                 namlen = strlen(dirent->lld_name);
 279                 LASSERT(sizeof(d_name) >= namlen + 1);
 280                 strcpy(d_name, dirent->lld_name);
 281                 OBD_FREE(dirent, sizeof(*dirent));
 282
 283                 CDEBUG(D_INODE, "entry "LPU64" of PENDING DIR: %s\n",
 284                        i, d_name);
 285
 286                 if (((namlen == 1) && !strcmp(d_name, ".")) ||
 287                     ((namlen == 2) && !strcmp(d_name, ".."))) {
 288                         continue;
 289                 }
 290
 291                 down(&pending_dir->i_sem);
 292                 dchild = lookup_one_len(d_name, mds->mds_pending_dir, namlen);
 293                 if (IS_ERR(dchild)) {
 294                         up(&pending_dir->i_sem);
 295                         GOTO(err_out, rc = PTR_ERR(dchild));
 296                 }
 297                 if (!dchild->d_inode) {
 298                         CERROR("orphan %s has been removed\n", d_name);
 299                         GOTO(next, rc = 0);
 300                 }
 301
 302                 child_inode = dchild->d_inode;
 303                 if (mds_inode_is_orphan(child_inode) &&
 304                     mds_open_orphan_count(child_inode)) {
 305                         CWARN("orphan %s was re-opened during recovery\n", d_name);
 306                         GOTO(next, rc = 0);
 307                 }
 308
 309                 rc = mds_unlink_orphan(obd, dchild, child_inode, pending_dir);
 310                 if (rc == 0) {
 311                         item ++;
 312                         CWARN("removed orphan %s from MDS and OST\n", d_name);
 313                 } else {
 314                         CERROR("removed orphan %s from MDS and OST failed,"
 315                                " rc = %d\n", d_name, rc);
 316                         rc = 0;
 317                 }
 318 next:
 319                 l_dput(dchild);
 320                 up(&pending_dir->i_sem);
 321         }
 322 err_out:
 323         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 324                 list_del(&dirent->lld_list);
 325                 OBD_FREE(dirent, sizeof(*dirent));
 326         }
 327 err_pop:
 328         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 329         if (rc == 0)
 330                 rc = item;
 331         RETURN(rc);
 332
 333 err_mntget:
 334         l_dput(mds->mds_pending_dir);
 335         goto err_pop;
 336 }