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 /* should NOT be called with the dcache lock, see fs/dcache.c */
  53 static int ll_ddelete(struct dentry *de)
  54 {
  55         ENTRY;
  56         LASSERT(de);
  57         CDEBUG(D_DENTRY, "%s dentry %*s (%p, parent %p, inode %p) %s%s\n",
  58                (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
  59                de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
  60                d_unhashed(de) ? "" : "hashed,",
  61                list_empty(&de->d_subdirs) ? "" : "subdirs");
  62         RETURN(0);
  63 }
  64
  65 void ll_set_dd(struct dentry *de)
  66 {
  67         ENTRY;
  68         LASSERT(de != NULL);
  69
  70         CDEBUG(D_DENTRY, "ldd on dentry %*s (%p) parent %p inode %p refc %d\n",
  71                de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
  72                atomic_read(&de->d_count));
  73         lock_kernel();
  74         if (de->d_fsdata == NULL) {
  75                 OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data));
  76         }
  77         unlock_kernel();
  78
  79         EXIT;
  80 }
  81
  82 void ll_intent_drop_lock(struct lookup_intent *it)
  83 {
  84         struct lustre_handle *handle;
  85
  86         if (it->it_op && it->d.lustre.it_lock_mode) {
  87                 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
  88                 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
  89                        " from it %p\n", handle->cookie, it);
  90                 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
  91
  92                 /* bug 494: intent_release may be called multiple times, from
  93                  * this thread and we don't want to double-decref this lock */
  94                 it->d.lustre.it_lock_mode = 0;
  95         }
  96 }
  97
  98 void ll_intent_release(struct lookup_intent *it)
  99 {
 100         ENTRY;
 101
 102         ll_intent_drop_lock(it);
 103         it->it_magic = 0;
 104         it->it_op_release = 0;
 105         it->d.lustre.it_disposition = 0;
 106         it->d.lustre.it_data = NULL;
 107         EXIT;
 108 }
 109
 110 void ll_unhash_aliases(struct inode *inode)
 111 {
 112         struct list_head *tmp, *head;
 113         struct ll_sb_info *sbi;
 114         ENTRY;
 115
 116         if (inode == NULL) {
 117                 CERROR("unexpected NULL inode, tell phil\n");
 118                 return;
 119         }
 120
 121         CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
 122                inode->i_ino, inode->i_generation, inode);
 123
 124         sbi = ll_i2sbi(inode);
 125         head = &inode->i_dentry;
 126 restart:
 127         spin_lock(&dcache_lock);
 128         tmp = head;
 129         while ((tmp = tmp->next) != head) {
 130                 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
 131                 if (atomic_read(&dentry->d_count) == 0) {
 132                         CDEBUG(D_DENTRY, "deleting dentry %*s (%p) parent %p "
 133                                "inode %p\n", dentry->d_name.len,
 134                                dentry->d_name.name, dentry, dentry->d_parent,
 135                                dentry->d_inode);
 136                         dget_locked(dentry);
 137                         __d_drop(dentry);
 138                         spin_unlock(&dcache_lock);
 139                         dput(dentry);
 140                         goto restart;
 141                 } else if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
 142                         CDEBUG(D_DENTRY, "unhashing dentry %*s (%p) parent %p "
 143                                "inode %p refc %d\n", dentry->d_name.len,
 144                                dentry->d_name.name, dentry, dentry->d_parent,
 145                                dentry->d_inode, atomic_read(&dentry->d_count));
 146                         hlist_del_init(&dentry->d_hash);
 147                         dentry->d_flags |= DCACHE_LUSTRE_INVALID;
 148                         hlist_add_head(&dentry->d_hash,
 149                                        &sbi->ll_orphan_dentry_list);
 150                 }
 151         }
 152         spin_unlock(&dcache_lock);
 153         EXIT;
 154 }
 155
 156 extern struct dentry *ll_find_alias(struct inode *, struct dentry *);
 157
 158 int revalidate_it_finish(struct ptlrpc_request *request, int offset,
 159                          struct lookup_intent *it, struct dentry *de)
 160 {
 161         struct ll_sb_info *sbi;
 162         int rc = 0;
 163         ENTRY;
 164
 165         if (!request)
 166                 RETURN(0);
 167
 168         if (it_disposition(it, DISP_LOOKUP_NEG))
 169                 RETURN(-ENOENT);
 170
 171         sbi = ll_i2sbi(de->d_inode);
 172         rc = ll_prep_inode(sbi->ll_dt_exp, sbi->ll_md_exp,
 173                            &de->d_inode, request, offset, NULL);
 174
 175         RETURN(rc);
 176 }
 177
 178 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 179 {
 180         LASSERT(it != NULL);
 181         LASSERT(dentry != NULL);
 182
 183         if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
 184                 struct inode *inode = dentry->d_inode;
 185                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 186                        inode, inode->i_ino, inode->i_generation);
 187                 mdc_set_lock_data(NULL, &it->d.lustre.it_lock_handle, inode);
 188         }
 189
 190         /* drop lookup or getattr locks immediately */
 191         if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 192             it->it_op == IT_CHDIR) {
 193 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 194                 /*
 195                  * on 2.6 there are situations when several lookups and
 196                  * revalidations may be requested during single operation.
 197                  * Therefore, we don't release intent here -bzzz
 198                  */
 199                 ll_intent_drop_lock(it);
 200 #else
 201                 ll_intent_release(it);
 202 #endif
 203         }
 204 }
 205
 206 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 207 {
 208         struct lookup_intent *it = *itp;
 209
 210 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 211         if (it) {
 212                 LASSERTF(it->it_magic == INTENT_MAGIC, "bad intent magic: %x\n",
 213                          it->it_magic);
 214         }
 215 #endif
 216
 217         if (!it || it->it_op == IT_GETXATTR)
 218                 it = *itp = deft;
 219
 220         it->it_op_release = ll_intent_release;
 221 }
 222
 223 int ll_revalidate_it(struct dentry *de, int flags, struct nameidata *nd,
 224                      struct lookup_intent *it)
 225 {
 226         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 227         struct ptlrpc_request *req = NULL;
 228         struct it_cb_data icbd;
 229         struct obd_export *exp;
 230         struct lustre_id pid;
 231         struct lustre_id cid;
 232         int rc;
 233         ENTRY;
 234
 235         CDEBUG(D_VFSTRACE, "VFS Op:name=%s, intent=%s\n", de->d_name.name,
 236                LL_IT2STR(it));
 237
 238         /* Cached negative dentries are unsafe for now - look them up again */
 239         if (de->d_inode == NULL)
 240                 RETURN(0);
 241
 242         CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n",
 243                de, de->d_name.len, de->d_name.name,
 244                (unsigned long) de->d_inode->i_ino,
 245                (unsigned long) de->d_inode->i_generation);
 246
 247         exp = ll_i2mdexp(de->d_inode);
 248         ll_inode2id(&pid, de->d_parent->d_inode);
 249         ll_inode2id(&cid, de->d_inode);
 250         LASSERT(id_fid(&cid) != 0);
 251
 252         icbd.icbd_parent = de->d_parent->d_inode;
 253         icbd.icbd_childp = &de;
 254
 255         /*
 256          * never execute intents for mount points. Attributes will be fixed up
 257          * in ll_inode_revalidate_it().
 258          */
 259         if (d_mountpoint(de))
 260                 RETURN(1);
 261
 262         if (nd != NULL)
 263                 nd->mnt->mnt_last_used = jiffies;
 264
 265         ll_frob_intent(&it, &lookup_it);
 266         LASSERT(it != NULL);
 267
 268         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock as
 269                                           well */
 270                 rc = md_intent_lock(exp, &pid, de->d_name.name,
 271                                     de->d_name.len, NULL, 0, &cid, &lookup_it,
 272                                     flags, &req, ll_mdc_blocking_ast);
 273                 /* If there was no lookup lock, no point in even checking for
 274                    UPDATE lock */
 275                 if (!rc) {
 276                         it = &lookup_it;
 277                         GOTO(out, rc);
 278                 }
 279                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 280                         ll_intent_release(&lookup_it);
 281                         it = &lookup_it;
 282                         GOTO(out, rc = 0);
 283                 }
 284
 285                 if (req)
 286                         ptlrpc_req_finished(req);
 287                 req = NULL;
 288                 ll_lookup_finish_locks(&lookup_it, de);
 289         }
 290
 291         rc = md_intent_lock(exp, &pid, de->d_name.name, de->d_name.len,
 292                             NULL, 0, &cid, it, flags, &req,
 293                             ll_mdc_blocking_ast);
 294
 295         /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
 296          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 297         if (req == NULL && rc >= 0)
 298                 GOTO(out, rc);
 299
 300         if (rc < 0) {
 301                 if (rc != -ESTALE) {
 302                         CDEBUG(D_INFO, "ll_intent_lock(): rc %d : it->it_status "
 303                                "%d\n", rc, it->d.lustre.it_status);
 304                 }
 305                 GOTO(out, rc = 0);
 306         }
 307
 308         rc = revalidate_it_finish(req, 1, it, de);
 309         if (rc != 0) {
 310                 ll_intent_release(it);
 311                 GOTO(out, rc = 0);
 312         }
 313         rc = 1;
 314
 315         /* unfortunately ll_intent_lock may cause a callback and revoke our
 316            dentry */
 317         spin_lock(&dcache_lock);
 318         hlist_del_init(&de->d_hash);
 319         __d_rehash(de, 0);
 320         spin_unlock(&dcache_lock);
 321
 322         GOTO(out, rc);
 323 out:
 324         if (req != NULL && rc == 1)
 325                 ptlrpc_req_finished(req);
 326         if (rc == 0) {
 327                 ll_unhash_aliases(de->d_inode);
 328                 return rc;
 329         }
 330
 331         CDEBUG(D_DENTRY, "revalidated dentry %*s (%p) parent %p "
 332                "inode %p refc %d\n", de->d_name.len,
 333                de->d_name.name, de, de->d_parent, de->d_inode,
 334                atomic_read(&de->d_count));
 335         ll_lookup_finish_locks(it, de);
 336         de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 337         if (!((de->d_inode->i_mode & S_ISUID) &&S_ISDIR(de->d_inode->i_mode)) ||
 338             !(flags & LOOKUP_CONTINUE || (it->it_op & (IT_CHDIR | IT_OPEN))))
 339                 return rc;
 340
 341         if (nd)
 342                 (void)ll_dir_process_mount_object(de, nd->mnt);
 343         return rc;
 344 }
 345
 346 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 347 {
 348         struct inode *inode= de->d_inode;
 349         struct ll_sb_info *sbi = ll_i2sbi(inode);
 350         struct ll_dentry_data *ldd = ll_d2d(de);
 351         struct obd_client_handle *handle;
 352         int rc = 0;
 353         ENTRY;
 354         LASSERT(ldd);
 355
 356         lock_kernel();
 357         /* Strictly speaking this introduces an additional race: the
 358          * increments should wait until the rpc has returned.
 359          * However, given that at present the function is void, this
 360          * issue is moot. */
 361         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 362                 unlock_kernel();
 363                 EXIT;
 364                 return;
 365         }
 366
 367         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 368                 unlock_kernel();
 369                 EXIT;
 370                 return;
 371         }
 372         unlock_kernel();
 373
 374         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 375         rc = obd_pin(sbi->ll_md_exp, inode->i_ino, inode->i_generation,
 376                      inode->i_mode & S_IFMT, handle, flag);
 377
 378         if (rc) {
 379                 lock_kernel();
 380                 memset(handle, 0, sizeof(*handle));
 381                 if (flag == 0)
 382                         ldd->lld_cwd_count--;
 383                 else
 384                         ldd->lld_mnt_count--;
 385                 unlock_kernel();
 386         }
 387
 388         EXIT;
 389         return;
 390 }
 391
 392 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 393 {
 394         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 395         struct ll_dentry_data *ldd = ll_d2d(de);
 396         struct obd_client_handle handle;
 397         int count, rc = 0;
 398         ENTRY;
 399         LASSERT(ldd);
 400
 401         lock_kernel();
 402         /* Strictly speaking this introduces an additional race: the
 403          * increments should wait until the rpc has returned.
 404          * However, given that at present the function is void, this
 405          * issue is moot. */
 406         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 407         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 408                 /* the "pin" failed */
 409                 unlock_kernel();
 410                 EXIT;
 411                 return;
 412         }
 413
 414         if (flag)
 415                 count = --ldd->lld_mnt_count;
 416         else
 417                 count = --ldd->lld_cwd_count;
 418         unlock_kernel();
 419
 420         if (count != 0) {
 421                 EXIT;
 422                 return;
 423         }
 424
 425         rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
 426         EXIT;
 427 }
 428
 429 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 430 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 431 {
 432         int rc;
 433         ENTRY;
 434
 435         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 436                 rc = ll_revalidate_it(dentry, nd->flags, nd, &nd->intent);
 437         else
 438                 rc = ll_revalidate_it(dentry, 0, nd, NULL);
 439
 440         RETURN(rc);
 441 }
 442 #endif
 443
 444 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 445 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 446 {
 447         struct ll_sb_info *sbi = ll_i2sbi(inode);
 448         struct lustre_id parent, child;
 449
 450         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 451         ll_inode2id(&parent, dentry->d_parent->d_inode);
 452         ll_inode2id(&child, inode);
 453         md_change_cbdata_name(sbi->ll_md_exp, &parent,
 454                               (char *)dentry->d_name.name,
 455                               dentry->d_name.len, &child,
 456                               null_if_equal, inode);
 457         iput(inode);
 458 }
 459 #else
 460 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 461 {
 462         struct ll_sb_info *sbi = ll_i2sbi(inode);
 463         struct lustre_id parent, child;
 464
 465         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 466         ll_inode2id(&parent, dentry->d_parent->d_inode);
 467         ll_inode2id(&child, inode);
 468         md_change_cbdata_name(sbi->ll_md_exp, &parent,
 469                               (char *)dentry->d_name.name,
 470                               dentry->d_name.len, &child,
 471                               null_if_equal, inode);
 472         iput(inode);
 473 }
 474 #endif
 475
 476 struct dentry_operations ll_d_ops = {
 477 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 478         .d_revalidate = ll_revalidate_nd,
 479 #else
 480         .d_revalidate_it = ll_revalidate_it,
 481 #endif
 482         .d_release = ll_release,
 483         .d_iput = ll_dentry_iput,
 484         .d_delete = ll_ddelete,
 485 #if 0
 486         .d_pin = ll_pin,
 487         .d_unpin = ll_unpin,
 488 #endif
 489 };