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                 CDEBUG(D_INODE, "invalidate 0x%p: %*s -> %lu/%lu\n",
 116                        dentry, dentry->d_name.len, dentry->d_name.name,
 117                        (unsigned long) dentry->d_inode->i_ino,
 118                        (unsigned long) dentry->d_inode->i_generation);
 119                 if (!atomic_read(&dentry->d_count)) {
 120                         dget_locked(dentry);
 121                         __d_drop(dentry);
 122                         spin_unlock(&dcache_lock);
 123                         dput(dentry);
 124                         goto restart;
 125                 } else {
 126                         hlist_del_init(&dentry->d_hash);
 127                         dentry->d_flags |= DCACHE_LUSTRE_INVALID;
 128                         hlist_add_head(&dentry->d_hash,
 129                                        &sbi->ll_orphan_dentry_list);
 130                 }
 131         }
 132         spin_unlock(&dcache_lock);
 133         EXIT;
 134 }
 135
 136 extern struct dentry *ll_find_alias(struct inode *, struct dentry *);
 137
 138 static int revalidate_it_finish(struct ptlrpc_request *request, int offset,
 139                                 struct lookup_intent *it,
 140                                 struct dentry *de)
 141 {
 142         struct ll_sb_info *sbi;
 143         int rc = 0;
 144         ENTRY;
 145
 146         if (!request)
 147                 RETURN(0);
 148
 149         if (it_disposition(it, DISP_LOOKUP_NEG))
 150                 RETURN(-ENOENT);
 151
 152         sbi = ll_i2sbi(de->d_inode);
 153         rc = ll_prep_inode(sbi->ll_osc_exp, sbi->ll_mdc_exp,
 154                            &de->d_inode, request, offset,NULL);
 155
 156         RETURN(rc);
 157 }
 158
 159 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 160 {
 161         LASSERT(it != NULL);
 162         LASSERT(dentry != NULL);
 163
 164         if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
 165                 struct inode *inode = dentry->d_inode;
 166                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 167                        inode, inode->i_ino, inode->i_generation);
 168                 mdc_set_lock_data(&it->d.lustre.it_lock_handle, inode);
 169         }
 170
 171         /* drop lookup or getattr locks immediately */
 172         if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
 173 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 174                 /* on 2.6 there are situation when several lookups and
 175                  * revalidations may be requested during single operation.
 176                  * therefore, we don't release intent here -bzzz */
 177                 ll_intent_drop_lock(it);
 178 #else
 179                 ll_intent_release(it);
 180 #endif
 181         }
 182 }
 183
 184 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 185 {
 186         struct lookup_intent *it = *itp;
 187 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 188         if (it && it->it_magic != INTENT_MAGIC) {
 189                 CERROR("WARNING: uninitialized intent\n");
 190                 LBUG();
 191         }
 192         if (it && (it->it_op == IT_GETATTR || it->it_op == 0))
 193                 it->it_op = IT_LOOKUP;
 194 #endif
 195
 196         if (!it || it->it_op == IT_GETXATTR)
 197                 it = *itp = deft;
 198
 199         it->it_op_release = ll_intent_release;
 200 }
 201
 202 int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
 203 {
 204         int rc;
 205         struct ll_fid pfid, cfid;
 206         struct it_cb_data icbd;
 207         struct ll_uctxt ctxt;
 208         struct ptlrpc_request *req = NULL;
 209         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 210         struct obd_export *exp;
 211
 212         ENTRY;
 213         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
 214                LL_IT2STR(it));
 215
 216         /* Cached negative dentries are unsafe for now - look them up again */
 217         if (de->d_inode == NULL)
 218                 RETURN(0);
 219
 220         CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n",
 221                         de, de->d_name.len, de->d_name.name,
 222                         (unsigned long) de->d_inode->i_ino,
 223                         (unsigned long) de->d_inode->i_generation);
 224
 225         exp = ll_i2mdcexp(de->d_inode);
 226         ll_inode2fid(&pfid, de->d_parent->d_inode);
 227         ll_inode2fid(&cfid, de->d_inode);
 228         icbd.icbd_parent = de->d_parent->d_inode;
 229         icbd.icbd_childp = &de;
 230
 231         /* Never execute intents for mount points.
 232          * Attributes will be fixed up in ll_inode_revalidate_it */
 233         if (d_mountpoint(de))
 234                 RETURN(1);
 235
 236         ll_frob_intent(&it, &lookup_it);
 237         LASSERT(it);
 238
 239         ll_i2uctxt(&ctxt, de->d_parent->d_inode, de->d_inode);
 240
 241         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock
 242                                           as well */
 243                 rc = md_intent_lock(exp, &ctxt, &pfid, de->d_name.name,
 244                                      de->d_name.len, NULL, 0, &cfid, &lookup_it,
 245                                      flags, &req, ll_mdc_blocking_ast);
 246                 /* If there was no lookup lock, no point in even checking for
 247                    UPDATE lock */
 248                 if (!rc) {
 249                         it = &lookup_it;
 250                         GOTO(out, rc);
 251                 }
 252                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 253                         ll_intent_release(&lookup_it);
 254                         it = &lookup_it;
 255                         GOTO(out, rc = 0);
 256                 }
 257
 258                 if (req)
 259                         ptlrpc_req_finished(req);
 260                 req = NULL;
 261                 ll_lookup_finish_locks(&lookup_it, de);
 262         }
 263
 264         rc = md_intent_lock(exp, &ctxt, &pfid, de->d_name.name, de->d_name.len,
 265                             NULL, 0, &cfid, it, flags, &req, ll_mdc_blocking_ast);
 266         /* If req is NULL, then md_intent_lock only tried to do a lock match;
 267          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 268         if (req == NULL && rc >= 0)
 269                 GOTO(out, rc);
 270
 271         if (rc < 0) {
 272                 if (rc != -ESTALE) {
 273                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
 274                                "%d\n", rc, it->d.lustre.it_status);
 275                 }
 276                 GOTO(out, rc = 0);
 277         }
 278
 279         rc = revalidate_it_finish(req, 1, it, de);
 280         if (rc != 0) {
 281                 ll_intent_release(it);
 282                 GOTO(out, rc = 0);
 283         }
 284         rc = 1;
 285
 286         /* unfortunately ll_intent_lock may cause a callback and revoke our
 287            dentry */
 288         spin_lock(&dcache_lock);
 289         hlist_del_init(&de->d_hash);
 290         __d_rehash(de, 0);
 291         spin_unlock(&dcache_lock);
 292
 293  out:
 294         if (req != NULL && rc == 1)
 295                 ptlrpc_req_finished(req);
 296         if (rc == 0) {
 297                 ll_unhash_aliases(de->d_inode);
 298                 de->d_flags |= DCACHE_LUSTRE_INVALID;
 299         } else {
 300                 ll_lookup_finish_locks(it, de);
 301                 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 302         }
 303         RETURN(rc);
 304 }
 305
 306 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 307 {
 308         struct inode *inode= de->d_inode;
 309         struct ll_sb_info *sbi = ll_i2sbi(inode);
 310         struct ll_dentry_data *ldd = ll_d2d(de);
 311         struct obd_client_handle *handle;
 312         int rc = 0;
 313         ENTRY;
 314         LASSERT(ldd);
 315
 316         lock_kernel();
 317         /* Strictly speaking this introduces an additional race: the
 318          * increments should wait until the rpc has returned.
 319          * However, given that at present the function is void, this
 320          * issue is moot. */
 321         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 322                 unlock_kernel();
 323                 EXIT;
 324                 return;
 325         }
 326
 327         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 328                 unlock_kernel();
 329                 EXIT;
 330                 return;
 331         }
 332         unlock_kernel();
 333
 334         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 335         rc = obd_pin(sbi->ll_mdc_exp, inode->i_ino, inode->i_generation,
 336                      inode->i_mode & S_IFMT, handle, flag);
 337
 338         if (rc) {
 339                 lock_kernel();
 340                 memset(handle, 0, sizeof(*handle));
 341                 if (flag == 0)
 342                         ldd->lld_cwd_count--;
 343                 else
 344                         ldd->lld_mnt_count--;
 345                 unlock_kernel();
 346         }
 347
 348         EXIT;
 349         return;
 350 }
 351
 352 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 353 {
 354         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 355         struct ll_dentry_data *ldd = ll_d2d(de);
 356         struct obd_client_handle handle;
 357         int count, rc = 0;
 358         ENTRY;
 359         LASSERT(ldd);
 360
 361         lock_kernel();
 362         /* Strictly speaking this introduces an additional race: the
 363          * increments should wait until the rpc has returned.
 364          * However, given that at present the function is void, this
 365          * issue is moot. */
 366         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 367         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 368                 /* the "pin" failed */
 369                 unlock_kernel();
 370                 EXIT;
 371                 return;
 372         }
 373
 374         if (flag)
 375                 count = --ldd->lld_mnt_count;
 376         else
 377                 count = --ldd->lld_cwd_count;
 378         unlock_kernel();
 379
 380         if (count != 0) {
 381                 EXIT;
 382                 return;
 383         }
 384
 385         rc = obd_unpin(sbi->ll_mdc_exp, &handle, flag);
 386         EXIT;
 387         return;
 388 }
 389
 390 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 391 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 392 {
 393         int rc;
 394         ENTRY;
 395
 396         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 397                 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
 398         else
 399                 rc = ll_revalidate_it(dentry, 0, NULL);
 400
 401         RETURN(rc);
 402 }
 403 #endif
 404
 405 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 406 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 407 {
 408         struct ll_sb_info *sbi = ll_i2sbi(inode);
 409         struct ll_fid parent, child;
 410
 411         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 412         ll_inode2fid(&parent, dentry->d_parent->d_inode);
 413         ll_inode2fid(&child, inode);
 414         md_change_cbdata_name(sbi->ll_mdc_exp, &parent,
 415                               (char *)dentry->d_name.name, dentry->d_name.len,
 416                               &child, null_if_equal, inode);
 417         iput(inode);
 418 }
 419 #else
 420 #error "implement ->d_iput() for 2.6"
 421 #endif
 422
 423 struct dentry_operations ll_d_ops = {
 424 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 425         .d_revalidate = ll_revalidate_nd,
 426 #else
 427         .d_revalidate_it = ll_revalidate_it,
 428 #endif
 429         .d_release = ll_release,
 430         .d_iput = ll_dentry_iput,
 431 #if 0
 432         .d_pin = ll_pin,
 433         .d_unpin = ll_unpin,
 434 #endif
 435 };