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         unsigned mode;
  64         ENTRY;
  65
  66         LASSERT(!mds_inode_is_orphan(dchild->d_inode));
  67
  68         down(&pending_dir->i_sem);
  69         fidlen = ll_fid2str(fidname, dchild->d_inode->i_ino,
  70                             dchild->d_inode->i_generation);
  71
  72         CDEBUG(D_HA, "pending destroy of %dx open file %s = %s\n",
  73                mds_open_orphan_count(dchild->d_inode), rec->ur_name, fidname);
  74
  75         pending_child = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
  76         if (IS_ERR(pending_child))
  77                 GOTO(out_lock, rc = PTR_ERR(pending_child));
  78
  79         if (pending_child->d_inode != NULL) {
  80                 CERROR("re-destroying orphan file %s?\n", rec->ur_name);
  81                 LASSERT(pending_child->d_inode == dchild->d_inode);
  82                 GOTO(out_dput, rc = 0);
  83         }
  84
  85         /* link() is semanticaly-wrong for S_IFDIR, so we set S_IFREG
  86          * for linking and return real mode back then -bzzz */
  87         mode = dchild->d_inode->i_mode;
  88         dchild->d_inode->i_mode = S_IFREG;
  89         rc = vfs_link(dchild, pending_dir, pending_child);
  90         if (rc)
  91                 CERROR("error linking orphan %s to PENDING: rc = %d\n",
  92                        rec->ur_name, rc);
  93         else
  94                 mds_inode_set_orphan(dchild->d_inode);
  95
  96         /* return mode and correct i_nlink if inode is directory */
  97         LASSERT(dchild->d_inode->i_nlink == 1);
  98         dchild->d_inode->i_mode = mode;
  99         if ((mode & S_IFMT) == S_IFDIR) {
 100                 dchild->d_inode->i_nlink++;
 101                 pending_dir->i_nlink++;
 102         }
 103         mark_inode_dirty(dchild->d_inode);
 104
 105 out_dput:
 106         dput(pending_child);
 107 out_lock:
 108         up(&pending_dir->i_sem);
 109         RETURN(rc);
 110 }
 111
 112 static int mds_osc_destroy_orphan(struct mds_obd *mds,
 113                                   struct inode *inode,
 114                                   struct lov_mds_md *lmm,
 115                                   int lmm_size,
 116                                   struct llog_cookie *logcookies,
 117                                   int log_unlink)
 118 {
 119         struct lov_stripe_md *lsm = NULL;
 120         struct obd_trans_info oti = { 0 };
 121         struct obdo *oa;
 122         int rc;
 123         ENTRY;
 124
 125         if (lmm_size == 0)
 126                 RETURN(0);
 127
 128         rc = obd_unpackmd(mds->mds_osc_exp, &lsm, lmm, lmm_size);
 129         if (rc < 0) {
 130                 CERROR("Error unpack md %p\n", lmm);
 131                 RETURN(rc);
 132         } else {
 133                 LASSERT(rc >= sizeof(*lsm));
 134                 rc = 0;
 135         }
 136
 137         oa = obdo_alloc();
 138         if (oa == NULL)
 139                 GOTO(out_free_memmd, rc = -ENOMEM);
 140         oa->o_id = lsm->lsm_object_id;
 141         oa->o_mode = inode->i_mode & S_IFMT;
 142         oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE;
 143
 144         if (log_unlink && logcookies) {
 145                 oa->o_valid |= OBD_MD_FLCOOKIE;
 146                 oti.oti_logcookies = logcookies;
 147         }
 148
 149         rc = obd_destroy(mds->mds_osc_exp, oa, lsm, &oti);
 150         obdo_free(oa);
 151         if (rc)
 152                 CDEBUG(D_INODE, "destroy orphan objid 0x"LPX64" on ost error "
 153                        "%d\n", lsm->lsm_object_id, rc);
 154 out_free_memmd:
 155         obd_free_memmd(mds->mds_osc_exp, &lsm);
 156         RETURN(rc);
 157 }
 158
 159 static int mds_unlink_orphan(struct obd_device *obd, struct dentry *dchild,
 160                              struct inode *inode, struct inode *pending_dir)
 161 {
 162         struct mds_obd *mds = &obd->u.mds;
 163         struct lov_mds_md *lmm = NULL;
 164         struct llog_cookie *logcookies = NULL;
 165         int lmm_size = 0, log_unlink = 0;
 166         void *handle = NULL;
 167         int rc, err;
 168         ENTRY;
 169
 170         LASSERT(mds->mds_osc_obd != NULL);
 171
 172         OBD_ALLOC(lmm, mds->mds_max_mdsize);
 173         if (lmm == NULL)
 174                 RETURN(-ENOMEM);
 175
 176         down(&inode->i_sem);
 177         rc = fsfilt_get_md(obd, inode, lmm, mds->mds_max_mdsize);
 178         up(&inode->i_sem);
 179
 180         if (rc < 0) {
 181                 CERROR("Error %d reading eadata for ino %lu\n",
 182                        rc, inode->i_ino);
 183                 GOTO(out_free_lmm, rc);
 184         } else if (rc > 0) {
 185                 lmm_size = rc;
 186                 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
 187                 if (rc > 0)
 188                         lmm_size = rc;
 189                 rc = 0;
 190         }
 191
 192         handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK, NULL,
 193                                   le32_to_cpu(lmm->lmm_stripe_count));
 194         if (IS_ERR(handle)) {
 195                 rc = PTR_ERR(handle);
 196                 CERROR("error fsfilt_start: %d\n", rc);
 197                 handle = NULL;
 198                 GOTO(out_free_lmm, rc);
 199         }
 200
 201         down(&inode->i_sem);
 202         rc = fsfilt_get_md(obd, inode, lmm, mds->mds_max_mdsize);
 203         up(&inode->i_sem);
 204
 205         if (rc < 0) {
 206                 CERROR("Error %d reading eadata for ino %lu\n",
 207                        rc, inode->i_ino);
 208                 GOTO(out_free_lmm, rc);
 209         } else if (rc > 0) {
 210                 lmm_size = rc;
 211                 rc = 0;
 212         }
 213
 214         if (S_ISDIR(inode->i_mode))
 215                 rc = vfs_rmdir(pending_dir, dchild);
 216         else
 217                 rc = vfs_unlink(pending_dir, dchild);
 218
 219         if (rc)
 220                 CERROR("error %d unlinking orphan %*s from PENDING directory\n",
 221                        rc, dchild->d_name.len, dchild->d_name.name);
 222
 223         if (!rc && lmm_size) {
 224                 OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
 225                 if (logcookies == NULL)
 226                         rc = -ENOMEM;
 227                 else if (mds_log_op_unlink(obd, inode, lmm,lmm_size,logcookies,
 228                                            mds->mds_max_cookiesize) > 0)
 229                         log_unlink = 1;
 230         }
 231         err = fsfilt_commit(obd, pending_dir, handle, 0);
 232         if (err) {
 233                 CERROR("error committing orphan unlink: %d\n", err);
 234                 if (!rc)
 235                         rc = err;
 236         }
 237         if (!rc) {
 238                 rc = mds_osc_destroy_orphan(mds, inode, lmm, lmm_size,
 239                                             logcookies, log_unlink);
 240         }
 241
 242         if (logcookies != NULL)
 243                 OBD_FREE(logcookies, mds->mds_max_cookiesize);
 244 out_free_lmm:
 245         OBD_FREE(lmm, mds->mds_max_mdsize);
 246         RETURN(rc);
 247 }
 248
 249 int mds_cleanup_orphans(struct obd_device *obd)
 250 {
 251         struct mds_obd *mds = &obd->u.mds;
 252         struct lvfs_run_ctxt saved;
 253         struct file *file;
 254         struct dentry *dchild, *dentry;
 255         struct vfsmount *mnt;
 256         struct inode *child_inode, *pending_dir = mds->mds_pending_dir->d_inode;
 257         struct l_linux_dirent *dirent, *n;
 258         struct list_head dentry_list;
 259         char d_name[LL_FID_NAMELEN];
 260         __u64 i = 0;
 261         int rc = 0, item = 0, namlen;
 262         ENTRY;
 263
 264         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 265         dentry = dget(mds->mds_pending_dir);
 266         if (IS_ERR(dentry))
 267                 GOTO(err_pop, rc = PTR_ERR(dentry));
 268         mnt = mntget(mds->mds_vfsmnt);
 269         if (IS_ERR(mnt))
 270                 GOTO(err_mntget, rc = PTR_ERR(mnt));
 271
 272         file = dentry_open(mds->mds_pending_dir, mds->mds_vfsmnt,
 273                            O_RDONLY | O_LARGEFILE);
 274         if (IS_ERR(file))
 275                 GOTO(err_pop, rc = PTR_ERR(file));
 276
 277         INIT_LIST_HEAD(&dentry_list);
 278         rc = l_readdir(file, &dentry_list);
 279         filp_close(file, 0);
 280         if (rc < 0)
 281                 GOTO(err_out, rc);
 282
 283         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 284                 i ++;
 285                 list_del(&dirent->lld_list);
 286
 287                 namlen = strlen(dirent->lld_name);
 288                 LASSERT(sizeof(d_name) >= namlen + 1);
 289                 strcpy(d_name, dirent->lld_name);
 290                 OBD_FREE(dirent, sizeof(*dirent));
 291
 292                 CDEBUG(D_INODE, "entry "LPU64" of PENDING DIR: %s\n",
 293                        i, d_name);
 294
 295                 if (((namlen == 1) && !strcmp(d_name, ".")) ||
 296                     ((namlen == 2) && !strcmp(d_name, ".."))) {
 297                         continue;
 298                 }
 299
 300                 down(&pending_dir->i_sem);
 301                 dchild = lookup_one_len(d_name, mds->mds_pending_dir, namlen);
 302                 if (IS_ERR(dchild)) {
 303                         up(&pending_dir->i_sem);
 304                         GOTO(err_out, rc = PTR_ERR(dchild));
 305                 }
 306                 if (!dchild->d_inode) {
 307                         CERROR("orphan %s has been removed\n", d_name);
 308                         GOTO(next, rc = 0);
 309                 }
 310
 311                 child_inode = dchild->d_inode;
 312                 if (mds_inode_is_orphan(child_inode) &&
 313                     mds_open_orphan_count(child_inode)) {
 314                         CWARN("orphan %s was re-opened during recovery\n", d_name);
 315                         GOTO(next, rc = 0);
 316                 }
 317
 318                 rc = mds_unlink_orphan(obd, dchild, child_inode, pending_dir);
 319                 if (rc == 0) {
 320                         item ++;
 321                         CWARN("removed orphan %s from MDS and OST\n", d_name);
 322                 } else {
 323                         CDEBUG(D_INODE, "removed orphan %s from MDS/OST failed,"
 324                                " rc = %d\n", d_name, rc);
 325                         rc = 0;
 326                 }
 327 next:
 328                 l_dput(dchild);
 329                 up(&pending_dir->i_sem);
 330         }
 331 err_out:
 332         list_for_each_entry_safe(dirent, n, &dentry_list, lld_list) {
 333                 list_del(&dirent->lld_list);
 334                 OBD_FREE(dirent, sizeof(*dirent));
 335         }
 336 err_pop:
 337         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 338         if (rc == 0)
 339                 rc = item;
 340         RETURN(rc);
 341
 342 err_mntget:
 343         l_dput(mds->mds_pending_dir);
 344         goto err_pop;
 345 }