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_lov_exp, sbi->ll_lmv_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 || it->it_op == IT_CHDIR) {
 192 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 193                 /* on 2.6 there are situation when several lookups and
 194                  * revalidations may be requested during single operation.
 195                  * therefore, we don't release intent here -bzzz */
 196                 ll_intent_drop_lock(it);
 197 #else
 198                 ll_intent_release(it);
 199 #endif
 200         }
 201 }
 202
 203 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 204 {
 205         struct lookup_intent *it = *itp;
 206
 207 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 208         if (it) {
 209                 LASSERTF(it->it_magic == INTENT_MAGIC, "bad intent magic: %x\n",
 210                          it->it_magic);
 211         }
 212 #endif
 213
 214         if (!it || it->it_op == IT_GETXATTR)
 215                 it = *itp = deft;
 216
 217         it->it_op_release = ll_intent_release;
 218 }
 219
 220 int ll_revalidate_it(struct dentry *de, int flags, struct nameidata *nd,
 221                      struct lookup_intent *it)
 222 {
 223         int rc;
 224         struct lustre_id pid;
 225         struct lustre_id cid;
 226         struct it_cb_data icbd;
 227         struct ptlrpc_request *req = NULL;
 228         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 229         struct obd_export *exp;
 230
 231         ENTRY;
 232         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
 233                LL_IT2STR(it));
 234
 235         /* Cached negative dentries are unsafe for now - look them up again */
 236         if (de->d_inode == NULL)
 237                 RETURN(0);
 238
 239         CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n",
 240                de, de->d_name.len, de->d_name.name,
 241                (unsigned long) de->d_inode->i_ino,
 242                (unsigned long) de->d_inode->i_generation);
 243
 244         exp = ll_i2lmvexp(de->d_inode);
 245         ll_inode2id(&pid, de->d_parent->d_inode);
 246         ll_inode2id(&cid, de->d_inode);
 247         icbd.icbd_parent = de->d_parent->d_inode;
 248         icbd.icbd_childp = &de;
 249
 250         /*
 251          * never execute intents for mount points. Attributes will be fixed up
 252          * in ll_inode_revalidate_it().
 253          */
 254         if (d_mountpoint(de))
 255                 RETURN(1);
 256
 257         if (nd != NULL)
 258                 nd->mnt->mnt_last_used = jiffies;
 259
 260         ll_frob_intent(&it, &lookup_it);
 261         LASSERT(it);
 262
 263         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock
 264                                           as well */
 265                 rc = md_intent_lock(exp, &pid, de->d_name.name,
 266                                     de->d_name.len, NULL, 0, &cid, &lookup_it,
 267                                     flags, &req, ll_mdc_blocking_ast);
 268                 /* If there was no lookup lock, no point in even checking for
 269                    UPDATE lock */
 270                 if (!rc) {
 271                         it = &lookup_it;
 272                         GOTO(out, rc);
 273                 }
 274                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 275                         ll_intent_release(&lookup_it);
 276                         it = &lookup_it;
 277                         GOTO(out, rc = 0);
 278                 }
 279
 280                 if (req)
 281                         ptlrpc_req_finished(req);
 282                 req = NULL;
 283                 ll_lookup_finish_locks(&lookup_it, de);
 284         }
 285
 286         /* at this point fid should be valid. */
 287         LASSERT(id_fid(&cid) != 0);
 288
 289         rc = md_intent_lock(exp, &pid, de->d_name.name, de->d_name.len,
 290                             NULL, 0, &cid, it, flags, &req,
 291                             ll_mdc_blocking_ast);
 292
 293         /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
 294          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 295         if (req == NULL && rc >= 0)
 296                 GOTO(out, rc);
 297
 298         if (rc < 0) {
 299                 if (rc != -ESTALE) {
 300                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
 301                                "%d\n", rc, it->d.lustre.it_status);
 302                 }
 303                 GOTO(out, rc = 0);
 304         }
 305
 306         rc = revalidate_it_finish(req, 1, it, de);
 307         if (rc != 0) {
 308                 ll_intent_release(it);
 309                 GOTO(out, rc = 0);
 310         }
 311         rc = 1;
 312
 313         /* unfortunately ll_intent_lock may cause a callback and revoke our
 314            dentry */
 315         spin_lock(&dcache_lock);
 316         hlist_del_init(&de->d_hash);
 317         __d_rehash(de, 0);
 318         spin_unlock(&dcache_lock);
 319
 320         GOTO(out, rc);
 321  out:
 322         if (req != NULL && rc == 1)
 323                 ptlrpc_req_finished(req);
 324         if (rc == 0) {
 325                 ll_unhash_aliases(de->d_inode);
 326                 return rc;
 327         }
 328
 329         CDEBUG(D_DENTRY, "revalidated dentry %*s (%p) parent %p "
 330                "inode %p refc %d\n", de->d_name.len,
 331                de->d_name.name, de, de->d_parent, de->d_inode,
 332                atomic_read(&de->d_count));
 333         ll_lookup_finish_locks(it, de);
 334         de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 335         if (!((de->d_inode->i_mode & S_ISUID) &&S_ISDIR(de->d_inode->i_mode)) ||
 336             !(flags & LOOKUP_CONTINUE || (it->it_op & (IT_CHDIR | IT_OPEN))))
 337                 return rc;
 338
 339         if (nd)
 340                 (void)ll_dir_process_mount_object(de, nd->mnt);
 341         return rc;
 342 }
 343
 344 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 345 {
 346         struct inode *inode= de->d_inode;
 347         struct ll_sb_info *sbi = ll_i2sbi(inode);
 348         struct ll_dentry_data *ldd = ll_d2d(de);
 349         struct obd_client_handle *handle;
 350         int rc = 0;
 351         ENTRY;
 352         LASSERT(ldd);
 353
 354         lock_kernel();
 355         /* Strictly speaking this introduces an additional race: the
 356          * increments should wait until the rpc has returned.
 357          * However, given that at present the function is void, this
 358          * issue is moot. */
 359         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 360                 unlock_kernel();
 361                 EXIT;
 362                 return;
 363         }
 364
 365         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 366                 unlock_kernel();
 367                 EXIT;
 368                 return;
 369         }
 370         unlock_kernel();
 371
 372         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 373         rc = obd_pin(sbi->ll_lmv_exp, inode->i_ino, inode->i_generation,
 374                      inode->i_mode & S_IFMT, handle, flag);
 375
 376         if (rc) {
 377                 lock_kernel();
 378                 memset(handle, 0, sizeof(*handle));
 379                 if (flag == 0)
 380                         ldd->lld_cwd_count--;
 381                 else
 382                         ldd->lld_mnt_count--;
 383                 unlock_kernel();
 384         }
 385
 386         EXIT;
 387         return;
 388 }
 389
 390 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 391 {
 392         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 393         struct ll_dentry_data *ldd = ll_d2d(de);
 394         struct obd_client_handle handle;
 395         int count, rc = 0;
 396         ENTRY;
 397         LASSERT(ldd);
 398
 399         lock_kernel();
 400         /* Strictly speaking this introduces an additional race: the
 401          * increments should wait until the rpc has returned.
 402          * However, given that at present the function is void, this
 403          * issue is moot. */
 404         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 405         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 406                 /* the "pin" failed */
 407                 unlock_kernel();
 408                 EXIT;
 409                 return;
 410         }
 411
 412         if (flag)
 413                 count = --ldd->lld_mnt_count;
 414         else
 415                 count = --ldd->lld_cwd_count;
 416         unlock_kernel();
 417
 418         if (count != 0) {
 419                 EXIT;
 420                 return;
 421         }
 422
 423         rc = obd_unpin(sbi->ll_lmv_exp, &handle, flag);
 424         EXIT;
 425         return;
 426 }
 427
 428 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 429 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 430 {
 431         int rc;
 432         ENTRY;
 433
 434         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 435                 rc = ll_revalidate_it(dentry, nd->flags, nd, &nd->intent);
 436         else
 437                 rc = ll_revalidate_it(dentry, 0, nd, NULL);
 438
 439         RETURN(rc);
 440 }
 441 #endif
 442
 443 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 444 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 445 {
 446         struct ll_sb_info *sbi = ll_i2sbi(inode);
 447         struct lustre_id parent, child;
 448
 449         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 450         ll_inode2id(&parent, dentry->d_parent->d_inode);
 451         ll_inode2id(&child, inode);
 452         md_change_cbdata_name(sbi->ll_lmv_exp, &parent,
 453                               (char *)dentry->d_name.name,
 454                               dentry->d_name.len, &child,
 455                               null_if_equal, inode);
 456         iput(inode);
 457 }
 458 #else
 459 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 460 {
 461         struct ll_sb_info *sbi = ll_i2sbi(inode);
 462         struct ll_fid parent, child;
 463
 464         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 465         ll_inode2fid(&parent, dentry->d_parent->d_inode);
 466         ll_inode2fid(&child, inode);
 467         md_change_cbdata_name(sbi->ll_mdc_exp, &parent,
 468                               (char *)dentry->d_name.name,
 469                               dentry->d_name.len, &child,
 470                               null_if_equal, inode);
 471         iput(inode);
 472 }
 473 #endif
 474
 475 struct dentry_operations ll_d_ops = {
 476 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 477         .d_revalidate = ll_revalidate_nd,
 478 #else
 479         .d_revalidate_it = ll_revalidate_it,
 480 #endif
 481         .d_release = ll_release,
 482         .d_iput = ll_dentry_iput,
 483         .d_delete = ll_ddelete,
 484 #if 0
 485         .d_pin = ll_pin,
 486         .d_unpin = ll_unpin,
 487 #endif
 488 };