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 ? "keeping" : "deleting"),
  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 static int revalidate_it_finish(struct ptlrpc_request *request, int offset,
 159                                 struct lookup_intent *it,
 160                                 struct dentry *de)
 161 {
 162         struct ll_sb_info *sbi;
 163         int rc = 0;
 164         ENTRY;
 165
 166         if (!request)
 167                 RETURN(0);
 168
 169         if (it_disposition(it, DISP_LOOKUP_NEG))
 170                 RETURN(-ENOENT);
 171
 172         sbi = ll_i2sbi(de->d_inode);
 173         rc = ll_prep_inode(sbi->ll_osc_exp, &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(&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 #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 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 207         if (it && it->it_magic != INTENT_MAGIC) {
 208                 CERROR("WARNING: uninitialized intent\n");
 209                 LBUG();
 210         }
 211         if (it && (it->it_op == IT_GETATTR || it->it_op == 0))
 212                 it->it_op = IT_LOOKUP;
 213 #endif
 214
 215         if (!it || it->it_op == IT_GETXATTR)
 216                 it = *itp = deft;
 217
 218         it->it_op_release = ll_intent_release;
 219 }
 220
 221 int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
 222 {
 223         int rc;
 224         struct ll_fid pfid, cfid;
 225         struct it_cb_data icbd;
 226         struct ll_uctxt ctxt;
 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         exp = ll_i2mdcexp(de->d_inode);
 240         ll_inode2fid(&pfid, de->d_parent->d_inode);
 241         ll_inode2fid(&cfid, de->d_inode);
 242         icbd.icbd_parent = de->d_parent->d_inode;
 243         icbd.icbd_childp = &de;
 244
 245         /* Never execute intents for mount points.
 246          * Attributes will be fixed up in ll_inode_revalidate_it */
 247         if (d_mountpoint(de))
 248                 RETURN(1);
 249
 250         ll_frob_intent(&it, &lookup_it);
 251         LASSERT(it);
 252
 253         ll_i2uctxt(&ctxt, de->d_parent->d_inode, de->d_inode);
 254
 255         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock
 256                                           as well */
 257                 rc = mdc_intent_lock(exp, &ctxt, &pfid, de->d_name.name,
 258                                      de->d_name.len, NULL, 0, &cfid, &lookup_it,
 259                                      flags, &req, ll_mdc_blocking_ast);
 260                 /* If there was no lookup lock, no point in even checking for
 261                    UPDATE lock */
 262                 if (!rc) {
 263                         it = &lookup_it;
 264                         GOTO(out, rc);
 265                 }
 266                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 267                         ll_intent_release(&lookup_it);
 268                         it = &lookup_it;
 269                         GOTO(out, rc = 0);
 270                 }
 271
 272                 if (req)
 273                         ptlrpc_req_finished(req);
 274                 req = NULL;
 275                 ll_lookup_finish_locks(&lookup_it, de);
 276         }
 277
 278         rc = mdc_intent_lock(exp, &ctxt, &pfid, de->d_name.name, de->d_name.len,
 279                              NULL, 0,
 280                              &cfid, it, flags, &req, ll_mdc_blocking_ast);
 281         /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
 282          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 283         if (req == NULL && rc >= 0)
 284                 GOTO(out, rc);
 285
 286         if (rc < 0) {
 287                 if (rc != -ESTALE) {
 288                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
 289                                "%d\n", rc, it->d.lustre.it_status);
 290                 }
 291                 GOTO(out, rc = 0);
 292         }
 293
 294         rc = revalidate_it_finish(req, 1, it, de);
 295         if (rc != 0) {
 296                 ll_intent_release(it);
 297                 GOTO(out, rc = 0);
 298         }
 299         rc = 1;
 300
 301         /* unfortunately ll_intent_lock may cause a callback and revoke our
 302            dentry */
 303         spin_lock(&dcache_lock);
 304         hlist_del_init(&de->d_hash);
 305         __d_rehash(de, 0);
 306         spin_unlock(&dcache_lock);
 307
 308  out:
 309         if (req != NULL && rc == 1)
 310                 ptlrpc_req_finished(req);
 311         if (rc == 0) {
 312                 ll_unhash_aliases(de->d_inode);
 313                 /* done in ll_unhash_aliases()
 314                 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
 315         } else {
 316                 CDEBUG(D_DENTRY, "revalidated dentry %*s (%p) parent %p "
 317                                "inode %p refc %d\n", de->d_name.len,
 318                                de->d_name.name, de, de->d_parent, de->d_inode,
 319                                atomic_read(&de->d_count));
 320                 ll_lookup_finish_locks(it, de);
 321                 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 322         }
 323         RETURN(rc);
 324 }
 325
 326 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 327 {
 328         struct inode *inode= de->d_inode;
 329         struct ll_sb_info *sbi = ll_i2sbi(inode);
 330         struct ll_dentry_data *ldd = ll_d2d(de);
 331         struct obd_client_handle *handle;
 332         int rc = 0;
 333         ENTRY;
 334         LASSERT(ldd);
 335
 336         lock_kernel();
 337         /* Strictly speaking this introduces an additional race: the
 338          * increments should wait until the rpc has returned.
 339          * However, given that at present the function is void, this
 340          * issue is moot. */
 341         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 342                 unlock_kernel();
 343                 EXIT;
 344                 return;
 345         }
 346
 347         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 348                 unlock_kernel();
 349                 EXIT;
 350                 return;
 351         }
 352         unlock_kernel();
 353
 354         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 355         rc = obd_pin(sbi->ll_mdc_exp, inode->i_ino, inode->i_generation,
 356                      inode->i_mode & S_IFMT, handle, flag);
 357
 358         if (rc) {
 359                 lock_kernel();
 360                 memset(handle, 0, sizeof(*handle));
 361                 if (flag == 0)
 362                         ldd->lld_cwd_count--;
 363                 else
 364                         ldd->lld_mnt_count--;
 365                 unlock_kernel();
 366         }
 367
 368         EXIT;
 369         return;
 370 }
 371
 372 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 373 {
 374         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 375         struct ll_dentry_data *ldd = ll_d2d(de);
 376         struct obd_client_handle handle;
 377         int count, rc = 0;
 378         ENTRY;
 379         LASSERT(ldd);
 380
 381         lock_kernel();
 382         /* Strictly speaking this introduces an additional race: the
 383          * increments should wait until the rpc has returned.
 384          * However, given that at present the function is void, this
 385          * issue is moot. */
 386         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 387         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 388                 /* the "pin" failed */
 389                 unlock_kernel();
 390                 EXIT;
 391                 return;
 392         }
 393
 394         if (flag)
 395                 count = --ldd->lld_mnt_count;
 396         else
 397                 count = --ldd->lld_cwd_count;
 398         unlock_kernel();
 399
 400         if (count != 0) {
 401                 EXIT;
 402                 return;
 403         }
 404
 405         rc = obd_unpin(sbi->ll_mdc_exp, &handle, flag);
 406         EXIT;
 407         return;
 408 }
 409
 410 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 411 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 412 {
 413         int rc;
 414         ENTRY;
 415
 416         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 417                 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
 418         else
 419                 rc = ll_revalidate_it(dentry, 0, NULL);
 420
 421         RETURN(rc);
 422 }
 423 #endif
 424
 425 struct dentry_operations ll_d_ops = {
 426 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 427         .d_revalidate = ll_revalidate_nd,
 428 #else
 429         .d_revalidate_it = ll_revalidate_it,
 430 #endif
 431         .d_release = ll_release,
 432         .d_delete = ll_ddelete,
 433 #if 0
 434         .d_pin = ll_pin,
 435         .d_unpin = ll_unpin,
 436 #endif
 437 };