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
  44         CDEBUG(D_DENTRY, "releasing dentry %p\n", de);
  45
  46         lld = ll_d2d(de);
  47         if (lld) { /* Root dentry does not have ll_dentry_data */
  48                 LASSERT(lld->lld_cwd_count == 0);
  49                 LASSERT(lld->lld_mnt_count == 0);
  50                 OBD_FREE(de->d_fsdata, sizeof(struct ll_dentry_data));
  51         }
  52
  53         EXIT;
  54 }
  55
  56 /* should NOT be called with the dcache lock, see fs/dcache.c */
  57 static int ll_ddelete(struct dentry *de)
  58 {
  59         ENTRY;
  60         LASSERT(de);
  61         CDEBUG(D_DENTRY, "%s dentry %*s (%p, parent %p, inode %p) %s%s\n",
  62                (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
  63                de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
  64                d_unhashed(de) ? "" : "hashed,",
  65                list_empty(&de->d_subdirs) ? "" : "subdirs");
  66         RETURN(0);
  67 }
  68
  69 void ll_set_dd(struct dentry *de)
  70 {
  71         ENTRY;
  72         LASSERT(de != NULL);
  73
  74         CDEBUG(D_DENTRY, "ldd on dentry %*s (%p) parent %p inode %p refc %d\n",
  75                de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
  76                atomic_read(&de->d_count));
  77         lock_kernel();
  78         if (de->d_fsdata == NULL) {
  79                 OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data));
  80         }
  81         unlock_kernel();
  82
  83         EXIT;
  84 }
  85
  86 void ll_intent_drop_lock(struct lookup_intent *it)
  87 {
  88         struct lustre_handle *handle;
  89         struct lustre_intent_data *itdata = LUSTRE_IT(it);
  90
  91         if (it->it_op && itdata && itdata->it_lock_mode) {
  92                 handle = (struct lustre_handle *)&itdata->it_lock_handle;
  93                 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
  94                        " from it %p\n", handle->cookie, it);
  95                 ldlm_lock_decref(handle, itdata->it_lock_mode);
  96
  97                 /* bug 494: intent_release may be called multiple times, from
  98                  * this thread and we don't want to double-decref this lock */
  99                 itdata->it_lock_mode = 0;
 100         }
 101 }
 102
 103 void ll_intent_release(struct lookup_intent *it)
 104 {
 105         ENTRY;
 106
 107         ll_intent_drop_lock(it);
 108         it->it_magic = 0;
 109         it->it_op_release = 0;
 110         ll_intent_free(it);
 111         EXIT;
 112 }
 113
 114 void ll_intent_free(struct lookup_intent *it)
 115 {
 116         if (it->d.fs_data) {
 117                 OBD_SLAB_FREE(it->d.fs_data, ll_intent_slab,
 118                                sizeof(struct lustre_intent_data));
 119                 it->d.fs_data = NULL;
 120         }
 121 }
 122
 123 void ll_unhash_aliases(struct inode *inode)
 124 {
 125         struct list_head *tmp, *head;
 126         struct ll_sb_info *sbi;
 127         ENTRY;
 128
 129         if (inode == NULL) {
 130                 CERROR("unexpected NULL inode, tell phil\n");
 131                 return;
 132         }
 133
 134         CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
 135                inode->i_ino, inode->i_generation, inode);
 136
 137         sbi = ll_i2sbi(inode);
 138         head = &inode->i_dentry;
 139 restart:
 140         spin_lock(&dcache_lock);
 141         tmp = head;
 142         while ((tmp = tmp->next) != head) {
 143                 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
 144                 if (atomic_read(&dentry->d_count) == 0) {
 145                         CDEBUG(D_DENTRY, "deleting dentry %*s (%p) parent %p "
 146                                "inode %p\n", dentry->d_name.len,
 147                                dentry->d_name.name, dentry, dentry->d_parent,
 148                                dentry->d_inode);
 149                         dget_locked(dentry);
 150                         __d_drop(dentry);
 151                         spin_unlock(&dcache_lock);
 152                         dput(dentry);
 153                         goto restart;
 154                 } else if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
 155                         CDEBUG(D_DENTRY, "unhashing dentry %*s (%p) parent %p "
 156                                "inode %p refc %d\n", dentry->d_name.len,
 157                                dentry->d_name.name, dentry, dentry->d_parent,
 158                                dentry->d_inode, atomic_read(&dentry->d_count));
 159                         hlist_del_init(&dentry->d_hash);
 160                         dentry->d_flags |= DCACHE_LUSTRE_INVALID;
 161                         hlist_add_head(&dentry->d_hash,
 162                                        &sbi->ll_orphan_dentry_list);
 163                 }
 164         }
 165         spin_unlock(&dcache_lock);
 166         EXIT;
 167 }
 168
 169 extern struct dentry *ll_find_alias(struct inode *, struct dentry *);
 170
 171 int revalidate_it_finish(struct ptlrpc_request *request, int offset,
 172                          struct lookup_intent *it, struct dentry *de)
 173 {
 174         struct ll_sb_info *sbi;
 175         int rc = 0;
 176         ENTRY;
 177
 178         if (!request)
 179                 RETURN(0);
 180
 181         if (it_disposition(it, DISP_LOOKUP_NEG))
 182                 RETURN(-ENOENT);
 183
 184         sbi = ll_i2sbi(de->d_inode);
 185         rc = ll_prep_inode(sbi->ll_dt_exp, sbi->ll_md_exp,
 186                            &de->d_inode, request, offset, NULL);
 187
 188         RETURN(rc);
 189 }
 190
 191 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 192 {
 193         LASSERT(it != NULL);
 194         LASSERT(dentry != NULL);
 195
 196         if (LUSTRE_IT(it)->it_lock_mode && dentry->d_inode != NULL) {
 197                 struct inode *inode = dentry->d_inode;
 198                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 199                        inode, inode->i_ino, inode->i_generation);
 200                 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle, inode);
 201         }
 202
 203         /* drop lookup or getattr locks immediately */
 204         if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
 205             it->it_op == IT_CHDIR) {
 206 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 207                 /*
 208                  * on 2.6 there are situations when several lookups and
 209                  * revalidations may be requested during single operation.
 210                  * Therefore, we don't release intent here -bzzz
 211                  */
 212                 ll_intent_drop_lock(it);
 213 #else
 214                 ll_intent_release(it);
 215 #endif
 216         }
 217 }
 218
 219 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 220 {
 221         struct lookup_intent *it = *itp;
 222
 223 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 224         if (it) {
 225                 LASSERTF(it->it_magic == INTENT_MAGIC, "bad intent magic: %x\n",
 226                          it->it_magic);
 227         }
 228 #endif
 229
 230         if (!it || it->it_op == IT_GETXATTR)
 231                 it = *itp = deft;
 232
 233         if (it->d.fs_data)
 234                 return;
 235
 236         if (ll_intent_alloc(it)) {
 237                 CERROR("Failed to allocate memory for lustre specific intent "
 238                        "data\n");
 239                 /* XXX: we cannot return status just yet */
 240                 LBUG();
 241         }
 242 }
 243
 244 int ll_intent_alloc(struct lookup_intent *it)
 245 {
 246         if (it->d.fs_data) {
 247                 CERROR("Intent alloc on already allocated intent\n");
 248                 return 0;
 249         }
 250         OBD_SLAB_ALLOC(it->d.fs_data, ll_intent_slab, SLAB_KERNEL,
 251                         sizeof(struct lustre_intent_data));
 252         if (!it->d.fs_data) {
 253                 CERROR("Failed to allocate memory for lustre specific intent "
 254                        "data\n");
 255                 return -ENOMEM;
 256         }
 257
 258         it->it_op_release = ll_intent_release;
 259
 260         return 0;
 261 }
 262
 263 int ll_revalidate_it(struct dentry *de, int flags, struct nameidata *nd,
 264                      struct lookup_intent *it)
 265 {
 266         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 267         struct ptlrpc_request *req = NULL;
 268         struct it_cb_data icbd;
 269         struct obd_export *exp;
 270         struct lustre_id pid;
 271         struct lustre_id cid;
 272         int orig_it, rc = 0;
 273         ENTRY;
 274
 275         spin_lock(&de->d_lock);
 276
 277         if ((de->d_flags & DCACHE_GNS_PENDING) &&
 278             !(de->d_flags & DCACHE_GNS_MOUNTING))
 279         {
 280                 spin_unlock(&de->d_lock);
 281
 282                 if (nd) {
 283                         int err = ll_gns_mount_object(de, nd->mnt);
 284                         if (err)
 285                                 CERROR("can't mount %s, err = %d\n",
 286                                        de->d_name.name, err);
 287                 }
 288                 RETURN(1);
 289         }
 290         spin_unlock(&de->d_lock);
 291
 292         CDEBUG(D_VFSTRACE, "VFS Op:name=%s (%p), intent=%s\n", de->d_name.name,
 293                de, LL_IT2STR(it));
 294
 295         /* Cached negative dentries are unsafe for now - look them up again */
 296         if (de->d_inode == NULL)
 297                 RETURN(0);
 298
 299         /* Root of the tree is always valid, attributes would be fixed in
 300           ll_inode_revalidate_it */
 301         if (de->d_sb->s_root == de)
 302                 RETURN(1);
 303
 304         CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n",
 305                de, de->d_name.len, de->d_name.name,
 306                (unsigned long) de->d_inode->i_ino,
 307                (unsigned long) de->d_inode->i_generation);
 308
 309         exp = ll_i2mdexp(de->d_inode);
 310         ll_inode2id(&pid, de->d_parent->d_inode);
 311         ll_inode2id(&cid, de->d_inode);
 312         LASSERT(id_fid(&cid) != 0);
 313
 314         icbd.icbd_parent = de->d_parent->d_inode;
 315         icbd.icbd_childp = &de;
 316
 317         /*
 318          * never execute intents for mount points. Attributes will be fixed up
 319          * in ll_inode_revalidate_it().
 320          */
 321         if (d_mountpoint(de))
 322                 RETURN(1);
 323
 324         if (nd != NULL)
 325                 nd->mnt->mnt_last_used = jiffies;
 326
 327         orig_it = it ? it->it_op : IT_OPEN;
 328         ll_frob_intent(&it, &lookup_it);
 329         LASSERT(it != NULL);
 330
 331         if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock as
 332                                           well */
 333                 rc = ll_intent_alloc(&lookup_it);
 334                 if (rc)
 335                         LBUG(); /* Can't think of better idea just yet */
 336
 337
 338                 rc = md_intent_lock(exp, &pid, de->d_name.name,
 339                                     de->d_name.len, NULL, 0, &cid, &lookup_it,
 340                                     flags, &req, ll_mdc_blocking_ast);
 341                 /* If there was no lookup lock, no point in even checking for
 342                    UPDATE lock */
 343                 if (!rc) {
 344                         it = &lookup_it;
 345                         if (!req) {
 346                                 ll_intent_free(it);
 347                                 goto do_lookup;
 348                         }
 349                         GOTO(out, rc);
 350                 }
 351                 if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
 352                         it = &lookup_it;
 353                         ll_intent_free(it);
 354                         GOTO(out, rc = 0);
 355                 }
 356
 357                 if (req)
 358                         ptlrpc_req_finished(req);
 359                 req = NULL;
 360                 ll_lookup_finish_locks(&lookup_it, de);
 361                 /* XXX: on 2.6 ll_lookup_finish_locks does not call ll_intent_release */
 362                 ll_intent_release(&lookup_it);
 363         }
 364
 365         rc = md_intent_lock(exp, &pid, de->d_name.name, de->d_name.len,
 366                             NULL, 0, &cid, it, flags, &req,
 367                             ll_mdc_blocking_ast);
 368
 369         /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
 370          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 371         if (req == NULL && rc >= 0) {
 372                 if (!rc)
 373                         goto do_lookup;
 374                 GOTO(out, rc);
 375         }
 376
 377         if (rc < 0) {
 378                 if (rc != -ESTALE) {
 379                         CDEBUG(D_INFO, "ll_intent_lock(): rc %d : it->it_status "
 380                                "%d\n", rc, LUSTRE_IT(it)->it_status);
 381                 }
 382                 GOTO(out, rc = 0);
 383         }
 384 revalidate_finish:
 385         rc = revalidate_it_finish(req, 1, it, de);
 386         if (rc != 0) {
 387                 ll_intent_release(it);
 388                 GOTO(out, rc = 0);
 389         }
 390         rc = 1;
 391
 392         /* unfortunately ll_intent_lock may cause a callback and revoke our
 393            dentry */
 394         spin_lock(&dcache_lock);
 395         hlist_del_init(&de->d_hash);
 396         __d_rehash(de);
 397         spin_unlock(&dcache_lock);
 398
 399         GOTO(out, rc);
 400 out:
 401         if (req != NULL && rc == 1)
 402                 ptlrpc_req_finished(req);
 403
 404         if (rc == 0) {
 405                 if (it == &lookup_it) {
 406                         ll_intent_release(it);
 407                         if (req) /* Special case: We did lookup and it failed,
 408                                     need to free request */
 409                                 ptlrpc_req_finished(req);
 410                 }
 411                 ll_unhash_aliases(de->d_inode);
 412                 return rc;
 413         }
 414
 415         CDEBUG(D_DENTRY, "revalidated dentry %*s (%p) parent %p "
 416                "inode %p refc %d\n", de->d_name.len,
 417                de->d_name.name, de, de->d_parent, de->d_inode,
 418                atomic_read(&de->d_count));
 419         ll_lookup_finish_locks(it, de);
 420         de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 421         if (it == &lookup_it)
 422                 ll_intent_release(it);
 423
 424         if (!((de->d_inode->i_mode & S_ISUID) && S_ISDIR(de->d_inode->i_mode)) ||
 425             !(flags & LOOKUP_CONTINUE || (orig_it & (IT_CHDIR | IT_OPEN))))
 426                 return rc;
 427
 428         if (nd && !(de->d_flags & DCACHE_GNS_MOUNTING)) {
 429                 int err = ll_gns_mount_object(de, nd->mnt);
 430                 if (err)
 431                         CERROR("can't mount %s, err = %d\n",
 432                                de->d_name.name, err);
 433         }
 434         return rc;
 435 do_lookup:
 436         it = &lookup_it;
 437         if (ll_intent_alloc(it))
 438                 LBUG();
 439 // We did that already, right?  ll_inode2id(&pid, de->d_parent->d_inode);
 440         rc = md_intent_lock(exp, &pid, de->d_name.name,
 441                             de->d_name.len, NULL, 0, NULL,
 442                             it, 0, &req, ll_mdc_blocking_ast);
 443         if (rc >= 0) {
 444                 struct mds_body *mds_body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*mds_body));
 445
 446                 /* See if we got same inode, if not - return error */
 447                 if (id_equal_stc(&cid, &mds_body->id1))
 448                         goto revalidate_finish;
 449         }
 450
 451         GOTO(out, rc = 0);
 452 }
 453
 454 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 455 {
 456         struct inode *inode= de->d_inode;
 457         struct ll_sb_info *sbi = ll_i2sbi(inode);
 458         struct ll_dentry_data *ldd = ll_d2d(de);
 459         struct obd_client_handle *handle;
 460         int rc = 0;
 461         ENTRY;
 462         LASSERT(ldd);
 463
 464         lock_kernel();
 465         /* Strictly speaking this introduces an additional race: the
 466          * increments should wait until the rpc has returned.
 467          * However, given that at present the function is void, this
 468          * issue is moot. */
 469         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 470                 unlock_kernel();
 471                 EXIT;
 472                 return;
 473         }
 474
 475         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 476                 unlock_kernel();
 477                 EXIT;
 478                 return;
 479         }
 480         unlock_kernel();
 481
 482         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 483         rc = obd_pin(sbi->ll_md_exp, inode->i_ino, inode->i_generation,
 484                      inode->i_mode & S_IFMT, handle, flag);
 485
 486         if (rc) {
 487                 lock_kernel();
 488                 memset(handle, 0, sizeof(*handle));
 489                 if (flag == 0)
 490                         ldd->lld_cwd_count--;
 491                 else
 492                         ldd->lld_mnt_count--;
 493                 unlock_kernel();
 494         }
 495
 496         EXIT;
 497         return;
 498 }
 499
 500 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 501 {
 502         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 503         struct ll_dentry_data *ldd = ll_d2d(de);
 504         struct obd_client_handle handle;
 505         int count, rc = 0;
 506         ENTRY;
 507         LASSERT(ldd);
 508
 509         lock_kernel();
 510         /* Strictly speaking this introduces an additional race: the
 511          * increments should wait until the rpc has returned.
 512          * However, given that at present the function is void, this
 513          * issue is moot. */
 514         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 515         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 516                 /* the "pin" failed */
 517                 unlock_kernel();
 518                 EXIT;
 519                 return;
 520         }
 521
 522         if (flag)
 523                 count = --ldd->lld_mnt_count;
 524         else
 525                 count = --ldd->lld_cwd_count;
 526         unlock_kernel();
 527
 528         if (count != 0) {
 529                 EXIT;
 530                 return;
 531         }
 532
 533         rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
 534         EXIT;
 535 }
 536
 537 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 538 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 539 {
 540         int rc;
 541         ENTRY;
 542
 543         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 544                 rc = ll_revalidate_it(dentry, nd->flags, nd, &nd->intent.open);
 545         else
 546                 rc = ll_revalidate_it(dentry, 0, nd, NULL);
 547
 548         RETURN(rc);
 549 }
 550 #endif
 551
 552 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 553 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 554 {
 555         struct ll_sb_info *sbi = ll_i2sbi(inode);
 556         struct lustre_id parent, child;
 557
 558         LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 559         ll_inode2id(&parent, dentry->d_parent->d_inode);
 560         ll_inode2id(&child, inode);
 561         md_change_cbdata_name(sbi->ll_md_exp, &parent,
 562                               (char *)dentry->d_name.name,
 563                               dentry->d_name.len, &child,
 564                               null_if_equal, inode);
 565         iput(inode);
 566 }
 567 #else
 568 static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
 569 {
 570         struct ll_sb_info *sbi = ll_i2sbi(inode);
 571         struct lustre_id parent, child;
 572
 573         if (dentry->d_parent != dentry) {
 574                 /* Do not do this for root of the tree */
 575                 LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
 576                 ll_inode2id(&parent, dentry->d_parent->d_inode);
 577                 ll_inode2id(&child, inode);
 578                 md_change_cbdata_name(sbi->ll_md_exp, &parent,
 579                                       (char *)dentry->d_name.name,
 580                                       dentry->d_name.len, &child,
 581                                       null_if_equal, inode);
 582         }
 583         iput(inode);
 584
 585 }
 586 #endif
 587
 588 struct dentry_operations ll_d_ops = {
 589 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 590         .d_revalidate = ll_revalidate_nd,
 591 #else
 592         .d_revalidate_it = ll_revalidate_it,
 593 #endif
 594         .d_release = ll_release,
 595         .d_iput = ll_dentry_iput,
 596         .d_delete = ll_ddelete,
 597 #if 0
 598         .d_pin = ll_pin,
 599         .d_unpin = ll_unpin,
 600 #endif
 601 };