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,
 159                                 int offset, 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, sbi->ll_mdc_exp,
 174                            &de->d_inode, request, offset, NULL);
 175
 176         RETURN(rc);
 177 }
 178
 179 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 180 {
 181         LASSERT(it != NULL);
 182         LASSERT(dentry != NULL);
 183
 184         if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
 185                 struct inode *inode = dentry->d_inode;
 186                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 187                        inode, inode->i_ino, inode->i_generation);
 188                 mdc_set_lock_data(NULL, &it->d.lustre.it_lock_handle, inode);
 189         }
 190
 191         /* drop lookup or getattr locks immediately */
 192         if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
 193 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 194                 /* on 2.6 there are situation when several lookups and
 195                  * revalidations may be requested during single operation.
 196                  * therefore, we don't release intent here -bzzz */
 197                 ll_intent_drop_lock(it);
 198 #else
 199                 ll_intent_release(it);
 200 #endif
 201         }
 202 }
 203
 204 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 205 {
 206         struct lookup_intent *it = *itp;
 207 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 208         if (it && it->it_magic != INTENT_MAGIC) {
 209                 CERROR("WARNING: uninitialized intent\n");
 210                 LBUG();
 211         }
 212         if (it && (it->it_op == IT_GETATTR || it->it_op == 0))
 213                 it->it_op = IT_LOOKUP;
 214 #endif
 215
 216         if (!it || it->it_op == IT_GETXATTR)
 217                 it = *itp = deft;
 218
 219         it->it_op_release = ll_intent_release;
 220 }
 221
 222 int ll_revalidate_it(struct dentry *de, int flags, struct nameidata *nd,
 223                      struct lookup_intent *it)
 224 {
 225         int rc;
 226         struct ll_fid pfid, cfid;
 227         struct it_cb_data icbd;
 228         struct ll_uctxt ctxt;
 229         struct ptlrpc_request *req = NULL;
 230         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 231         struct obd_export *exp;
 232
 233         ENTRY;
 234         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
 235                LL_IT2STR(it));
 236
 237         /* Cached negative dentries are unsafe for now - look them up again */
 238         if (de->d_inode == NULL)
 239                 RETURN(0);
 240
 241         CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n",
 242                         de, de->d_name.len, de->d_name.name,
 243                         (unsigned long) de->d_inode->i_ino,
 244                         (unsigned long) de->d_inode->i_generation);
 245
 246         exp = ll_i2mdcexp(de->d_inode);
 247         ll_inode2fid(&pfid, de->d_parent->d_inode);
 248         ll_inode2fid(&cfid, de->d_inode);
 249         icbd.icbd_parent = de->d_parent->d_inode;
 250         icbd.icbd_childp = &de;
 251
 252         /* Never execute intents for mount points.
 253          * Attributes will be fixed up in ll_inode_revalidate_it */
 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         ll_i2uctxt(&ctxt, de->d_parent->d_inode, de->d_inode);
 264
 265         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock
 266                                           as well */
 267                 rc = md_intent_lock(exp, &ctxt, &pfid, de->d_name.name,
 268                                     de->d_name.len, NULL, 0, &cfid, &lookup_it,
 269                                     flags, &req, ll_mdc_blocking_ast);
 270                 /* If there was no lookup lock, no point in even checking for
 271                    UPDATE lock */
 272                 if (!rc) {
 273                         it = &lookup_it;
 274                         GOTO(out, rc);
 275                 }
 276                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 277                         ll_intent_release(&lookup_it);
 278                         it = &lookup_it;
 279                         GOTO(out, rc = 0);
 280                 }
 281
 282                 if (req)
 283                         ptlrpc_req_finished(req);
 284                 req = NULL;
 285                 ll_lookup_finish_locks(&lookup_it, de);
 286         }
 287
 288         rc = md_intent_lock(exp, &ctxt, &pfid, de->d_name.name, de->d_name.len,
 289                             NULL, 0, &cfid, it, flags, &req,
 290                             ll_mdc_blocking_ast);
 291         /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
 292          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 293         if (req == NULL && rc >= 0)
 294                 GOTO(out, rc);
 295
 296         if (rc < 0) {
 297                 if (rc != -ESTALE) {
 298                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
 299                                "%d\n", rc, it->d.lustre.it_status);
 300                 }
 301                 GOTO(out, rc = 0);
 302         }
 303
 304         rc = revalidate_it_finish(req, 1, it, de);
 305         if (rc != 0) {
 306                 ll_intent_release(it);
 307                 GOTO(out, rc = 0);
 308         }
 309         rc = 1;
 310
 311         /* unfortunately ll_intent_lock may cause a callback and revoke our
 312            dentry */
 313         spin_lock(&dcache_lock);
 314         hlist_del_init(&de->d_hash);
 315         __d_rehash(de, 0);
 316         spin_unlock(&dcache_lock);
 317
 318         GOTO(out, rc);
 319  out:
 320         if (req != NULL && rc == 1)
 321                 ptlrpc_req_finished(req);
 322         if (rc == 0) {
 323                 ll_unhash_aliases(de->d_inode);
 324                 return rc;
 325         }
 326
 327         CDEBUG(D_DENTRY, "revalidated dentry %*s (%p) parent %p "
 328                "inode %p refc %d\n", de->d_name.len,
 329                de->d_name.name, de, de->d_parent, de->d_inode,
 330                atomic_read(&de->d_count));
 331         ll_lookup_finish_locks(it, de);
 332         de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 333         if (!((de->d_inode->i_mode & S_ISUID) &&S_ISDIR(de->d_inode->i_mode)) ||
 334             !(flags & LOOKUP_CONTINUE || (it->it_op & (IT_CHDIR | IT_OPEN))))
 335                 return rc;
 336
 337         if (nd)
 338                 (void)ll_dir_process_mount_object(de, nd->mnt);
 339         return rc;
 340 }
 341
 342 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 343 {
 344         struct inode *inode= de->d_inode;
 345         struct ll_sb_info *sbi = ll_i2sbi(inode);
 346         struct ll_dentry_data *ldd = ll_d2d(de);
 347         struct obd_client_handle *handle;
 348         int 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         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 358                 unlock_kernel();
 359                 EXIT;
 360                 return;
 361         }
 362
 363         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 364                 unlock_kernel();
 365                 EXIT;
 366                 return;
 367         }
 368         unlock_kernel();
 369
 370         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 371         rc = obd_pin(sbi->ll_mdc_exp, inode->i_ino, inode->i_generation,
 372                      inode->i_mode & S_IFMT, handle, flag);
 373
 374         if (rc) {
 375                 lock_kernel();
 376                 memset(handle, 0, sizeof(*handle));
 377                 if (flag == 0)
 378                         ldd->lld_cwd_count--;
 379                 else
 380                         ldd->lld_mnt_count--;
 381                 unlock_kernel();
 382         }
 383
 384         EXIT;
 385         return;
 386 }
 387
 388 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 389 {
 390         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 391         struct ll_dentry_data *ldd = ll_d2d(de);
 392         struct obd_client_handle handle;
 393         int count, rc = 0;
 394         ENTRY;
 395         LASSERT(ldd);
 396
 397         lock_kernel();
 398         /* Strictly speaking this introduces an additional race: the
 399          * increments should wait until the rpc has returned.
 400          * However, given that at present the function is void, this
 401          * issue is moot. */
 402         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 403         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 404                 /* the "pin" failed */
 405                 unlock_kernel();
 406                 EXIT;
 407                 return;
 408         }
 409
 410         if (flag)
 411                 count = --ldd->lld_mnt_count;
 412         else
 413                 count = --ldd->lld_cwd_count;
 414         unlock_kernel();
 415
 416         if (count != 0) {
 417                 EXIT;
 418                 return;
 419         }
 420
 421         rc = obd_unpin(sbi->ll_mdc_exp, &handle, flag);
 422         EXIT;
 423         return;
 424 }
 425
 426 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 427 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 428 {
 429         int rc;
 430         ENTRY;
 431
 432         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 433                 rc = ll_revalidate_it(dentry, nd->flags, nd, &nd->intent);
 434         else
 435                 rc = ll_revalidate_it(dentry, 0, nd, NULL);
 436
 437         RETURN(rc);
 438 }
 439 #endif
 440
 441 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 442 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 443 {
 444         struct ll_sb_info *sbi = ll_i2sbi(inode);
 445         struct ll_fid parent, child;
 446
 447         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 448         ll_inode2fid(&parent, dentry->d_parent->d_inode);
 449         ll_inode2fid(&child, inode);
 450         md_change_cbdata_name(sbi->ll_mdc_exp, &parent,
 451                               (char *)dentry->d_name.name,
 452                               dentry->d_name.len, &child,
 453                               null_if_equal, inode);
 454         iput(inode);
 455 }
 456 #else
 457 #error "implement ->d_iput() for 2.6"
 458 #endif
 459
 460 struct dentry_operations ll_d_ops = {
 461 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 462         .d_revalidate = ll_revalidate_nd,
 463 #else
 464         .d_revalidate_it = ll_revalidate_it,
 465 #endif
 466         .d_release = ll_release,
 467         .d_iput = ll_dentry_iput,
 468         .d_delete = ll_ddelete,
 469 #if 0
 470         .d_pin = ll_pin,
 471         .d_unpin = ll_unpin,
 472 #endif
 473 };