lustre/llite/dcache.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #include <linux/fs.h>
  23 #include <linux/sched.h>
  24 #include <linux/smp_lock.h>
  25 #include <linux/quotaops.h>
  26
  27 #define DEBUG_SUBSYSTEM S_LLITE
  28
  29 #include <linux/obd_support.h>
  30 #include <linux/lustre_lite.h>
  31 #include <linux/lustre_idl.h>
  32 #include <linux/lustre_dlm.h>
  33 #include <linux/lustre_version.h>
  34
  35 #include "llite_internal.h"
  36
  37 /* should NOT be called with the dcache lock, see fs/dcache.c */
  38 static void ll_release(struct dentry *de)
  39 {
  40         struct ll_dentry_data *lld;
  41         ENTRY;
  42         LASSERT(de != NULL);
  43         lld = ll_d2d(de);
  44         LASSERT(lld != NULL);
  45         LASSERT(lld->lld_cwd_count == 0);
  46         LASSERT(lld->lld_mnt_count == 0);
  47         OBD_FREE(de->d_fsdata, sizeof(struct ll_dentry_data));
  48
  49         EXIT;
  50 }
  51
  52 void ll_set_dd(struct dentry *de)
  53 {
  54         ENTRY;
  55         LASSERT(de != NULL);
  56
  57         lock_kernel();
  58         if (de->d_fsdata == NULL) {
  59                 OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data));
  60         }
  61         unlock_kernel();
  62
  63         EXIT;
  64 }
  65
  66 void ll_intent_drop_lock(struct lookup_intent *it)
  67 {
  68         struct lustre_handle *handle;
  69
  70         if (it->it_op && it->d.lustre.it_lock_mode) {
  71                 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
  72                 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
  73                        " from it %p\n", handle->cookie, it);
  74                 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
  75
  76                 /* bug 494: intent_release may be called multiple times, from
  77                  * this thread and we don't want to double-decref this lock */
  78                 it->d.lustre.it_lock_mode = 0;
  79         }
  80 }
  81
  82 void ll_intent_release(struct lookup_intent *it)
  83 {
  84         ENTRY;
  85
  86         ll_intent_drop_lock(it);
  87         it->it_magic = 0;
  88         it->it_op_release = 0;
  89         it->d.lustre.it_disposition = 0;
  90         it->d.lustre.it_data = NULL;
  91         EXIT;
  92 }
  93
  94 void ll_unhash_aliases(struct inode *inode)
  95 {
  96         struct list_head *tmp, *head;
  97         struct ll_sb_info *sbi;
  98         ENTRY;
  99
 100         sbi = ll_i2sbi(inode);
 101
 102         CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
 103                inode->i_ino, inode->i_generation, inode);
 104
 105         if (inode == NULL) {
 106                 CERROR("unexpected NULL inode, tell phil\n");
 107                 return;
 108         }
 109         head = &inode->i_dentry;
 110 restart:
 111         spin_lock(&dcache_lock);
 112         tmp = head;
 113         while ((tmp = tmp->next) != head) {
 114                 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
 115                 if (!atomic_read(&dentry->d_count)) {
 116                         dget_locked(dentry);
 117                         __d_drop(dentry);
 118                         spin_unlock(&dcache_lock);
 119                         dput(dentry);
 120                         goto restart;
 121                 } else {
 122                         hlist_del_init(&dentry->d_hash);
 123                         dentry->d_flags |= DCACHE_LUSTRE_INVALID;
 124                         hlist_add_head(&dentry->d_hash,
 125                                        &sbi->ll_orphan_dentry_list);
 126                 }
 127         }
 128         spin_unlock(&dcache_lock);
 129         EXIT;
 130 }
 131
 132 extern struct dentry *ll_find_alias(struct inode *, struct dentry *);
 133
 134 static int revalidate_it_finish(struct ptlrpc_request *request, int offset,
 135                                 struct lookup_intent *it,
 136                                 struct dentry *de)
 137 {
 138         struct ll_sb_info *sbi;
 139         int rc = 0;
 140         ENTRY;
 141
 142         if (!request)
 143                 RETURN(0);
 144
 145         if (it_disposition(it, DISP_LOOKUP_NEG))
 146                 RETURN(-ENOENT);
 147
 148         sbi = ll_i2sbi(de->d_inode);
 149         rc = ll_prep_inode(sbi->ll_osc_exp, &de->d_inode, request, offset,NULL);
 150
 151         RETURN(rc);
 152 }
 153
 154 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 155 {
 156         LASSERT(it != NULL);
 157         LASSERT(dentry != NULL);
 158
 159         if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
 160                 struct inode *inode = dentry->d_inode;
 161                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 162                        inode, inode->i_ino, inode->i_generation);
 163                 mdc_set_lock_data(&it->d.lustre.it_lock_handle, inode);
 164         }
 165
 166         /* drop lookup or getattr locks immediately */
 167         if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
 168 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 169                 /* on 2.6 there are situation when several lookups and
 170                  * revalidations may be requested during single operation.
 171                  * therefore, we don't release intent here -bzzz */
 172                 ll_intent_drop_lock(it);
 173 #else
 174                 ll_intent_release(it);
 175 #endif
 176         }
 177 }
 178
 179 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 180 {
 181         struct lookup_intent *it = *itp;
 182 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 183         if (it && it->it_magic != INTENT_MAGIC) {
 184                 CERROR("WARNING: uninitialized intent\n");
 185                 LBUG();
 186         }
 187         if (it && (it->it_op == IT_GETATTR || it->it_op == 0))
 188                 it->it_op = IT_LOOKUP;
 189 #endif
 190
 191         if (!it || it->it_op == IT_GETXATTR)
 192                 it = *itp = deft;
 193
 194         it->it_op_release = ll_intent_release;
 195 }
 196
 197 int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
 198 {
 199         int rc;
 200         struct ll_fid pfid, cfid;
 201         struct it_cb_data icbd;
 202         struct ll_uctxt ctxt;
 203         struct ptlrpc_request *req = NULL;
 204         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 205         struct obd_export *exp;
 206
 207         ENTRY;
 208         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
 209                LL_IT2STR(it));
 210
 211         /* Cached negative dentries are unsafe for now - look them up again */
 212         if (de->d_inode == NULL)
 213                 RETURN(0);
 214
 215         exp = ll_i2mdcexp(de->d_inode);
 216         ll_inode2fid(&pfid, de->d_parent->d_inode);
 217         ll_inode2fid(&cfid, de->d_inode);
 218         icbd.icbd_parent = de->d_parent->d_inode;
 219         icbd.icbd_childp = &de;
 220
 221         /* Never execute intents for mount points.
 222          * Attributes will be fixed up in ll_inode_revalidate_it */
 223         if (d_mountpoint(de))
 224                 RETURN(1);
 225
 226         ll_frob_intent(&it, &lookup_it);
 227         LASSERT(it);
 228
 229         ll_i2uctxt(&ctxt, de->d_parent->d_inode, de->d_inode);
 230
 231         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock
 232                                           as well */
 233                 rc = mdc_intent_lock(exp, &ctxt, &pfid, de->d_name.name,
 234                                      de->d_name.len, NULL, 0, &cfid, &lookup_it,
 235                                      flags, &req, ll_mdc_blocking_ast);
 236                 /* If there was no lookup lock, no point in even checking for
 237                    UPDATE lock */
 238                 if (!rc) {
 239                         it = &lookup_it;
 240                         GOTO(out, rc);
 241                 }
 242                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 243                         ll_intent_release(&lookup_it);
 244                         it = &lookup_it;
 245                         GOTO(out, rc = 0);
 246                 }
 247
 248                 if (req)
 249                         ptlrpc_req_finished(req);
 250                 req = NULL;
 251                 ll_lookup_finish_locks(&lookup_it, de);
 252         }
 253
 254         rc = mdc_intent_lock(exp, &ctxt, &pfid, de->d_name.name, de->d_name.len,
 255                              NULL, 0,
 256                              &cfid, it, flags, &req, ll_mdc_blocking_ast);
 257         /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
 258          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 259         if (req == NULL && rc >= 0)
 260                 GOTO(out, rc);
 261
 262         if (rc < 0) {
 263                 if (rc != -ESTALE) {
 264                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
 265                                "%d\n", rc, it->d.lustre.it_status);
 266                 }
 267                 GOTO(out, rc = 0);
 268         }
 269
 270         rc = revalidate_it_finish(req, 1, it, de);
 271         if (rc != 0) {
 272                 ll_intent_release(it);
 273                 GOTO(out, rc = 0);
 274         }
 275         rc = 1;
 276
 277         /* unfortunately ll_intent_lock may cause a callback and revoke our
 278            dentry */
 279         spin_lock(&dcache_lock);
 280         hlist_del_init(&de->d_hash);
 281         __d_rehash(de, 0);
 282         spin_unlock(&dcache_lock);
 283
 284  out:
 285         if (req != NULL && rc == 1)
 286                 ptlrpc_req_finished(req);
 287         if (rc == 0) {
 288                 ll_unhash_aliases(de->d_inode);
 289                 de->d_flags |= DCACHE_LUSTRE_INVALID;
 290         } else {
 291                 ll_lookup_finish_locks(it, de);
 292                 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 293         }
 294         RETURN(rc);
 295 }
 296
 297 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 298 {
 299         struct inode *inode= de->d_inode;
 300         struct ll_sb_info *sbi = ll_i2sbi(inode);
 301         struct ll_dentry_data *ldd = ll_d2d(de);
 302         struct obd_client_handle *handle;
 303         int rc = 0;
 304         ENTRY;
 305         LASSERT(ldd);
 306
 307         lock_kernel();
 308         /* Strictly speaking this introduces an additional race: the
 309          * increments should wait until the rpc has returned.
 310          * However, given that at present the function is void, this
 311          * issue is moot. */
 312         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 313                 unlock_kernel();
 314                 EXIT;
 315                 return;
 316         }
 317
 318         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 319                 unlock_kernel();
 320                 EXIT;
 321                 return;
 322         }
 323         unlock_kernel();
 324
 325         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 326         rc = obd_pin(sbi->ll_mdc_exp, inode->i_ino, inode->i_generation,
 327                      inode->i_mode & S_IFMT, handle, flag);
 328
 329         if (rc) {
 330                 lock_kernel();
 331                 memset(handle, 0, sizeof(*handle));
 332                 if (flag == 0)
 333                         ldd->lld_cwd_count--;
 334                 else
 335                         ldd->lld_mnt_count--;
 336                 unlock_kernel();
 337         }
 338
 339         EXIT;
 340         return;
 341 }
 342
 343 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 344 {
 345         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 346         struct ll_dentry_data *ldd = ll_d2d(de);
 347         struct obd_client_handle handle;
 348         int count, rc = 0;
 349         ENTRY;
 350         LASSERT(ldd);
 351
 352         lock_kernel();
 353         /* Strictly speaking this introduces an additional race: the
 354          * increments should wait until the rpc has returned.
 355          * However, given that at present the function is void, this
 356          * issue is moot. */
 357         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 358         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 359                 /* the "pin" failed */
 360                 unlock_kernel();
 361                 EXIT;
 362                 return;
 363         }
 364
 365         if (flag)
 366                 count = --ldd->lld_mnt_count;
 367         else
 368                 count = --ldd->lld_cwd_count;
 369         unlock_kernel();
 370
 371         if (count != 0) {
 372                 EXIT;
 373                 return;
 374         }
 375
 376         rc = obd_unpin(sbi->ll_mdc_exp, &handle, flag);
 377         EXIT;
 378         return;
 379 }
 380
 381 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 382 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 383 {
 384         int rc;
 385         ENTRY;
 386
 387         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 388                 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
 389         else
 390                 rc = ll_revalidate_it(dentry, 0, NULL);
 391
 392         RETURN(rc);
 393 }
 394 #endif
 395
 396 struct dentry_operations ll_d_ops = {
 397 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 398         .d_revalidate = ll_revalidate_nd,
 399 #else
 400         .d_revalidate_it = ll_revalidate_it,
 401 #endif
 402         .d_release = ll_release,
 403 #if 0
 404         .d_pin = ll_pin,
 405         .d_unpin = ll_unpin,
 406 #endif
 407 };