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 <obd_support.h>
  30 #include <lustre_lite.h>
  31 #include <lustre/lustre_idl.h>
  32 #include <lustre_dlm.h>
  33 #include <lustre_mdc.h>
  34 #include <lustre_ver.h>
  35
  36 #include "llite_internal.h"
  37
  38 /* should NOT be called with the dcache lock, see fs/dcache.c */
  39 static void ll_release(struct dentry *de)
  40 {
  41         struct ll_dentry_data *lld;
  42         ENTRY;
  43         LASSERT(de != NULL);
  44         lld = ll_d2d(de);
  45         if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */
  46                 EXIT;
  47                 return;
  48         }
  49         LASSERT(lld->lld_cwd_count == 0);
  50         LASSERT(lld->lld_mnt_count == 0);
  51         OBD_FREE(de->d_fsdata, sizeof(*lld));
  52
  53         EXIT;
  54 }
  55
  56 /* Compare if two dentries are the same.  Don't match if the existing dentry
  57  * is marked DCACHE_LUSTRE_INVALID.  Returns 1 if different, 0 if the same.
  58  *
  59  * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
  60  * an AST before calling d_revalidate_it().  The dentry still exists (marked
  61  * INVALID) so d_lookup() matches it, but we have no lock on it (so
  62  * lock_match() fails) and we spin around real_lookup(). */
  63 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
  64 {
  65         struct dentry *dchild;
  66         ENTRY;
  67
  68         if (d_name->len != name->len)
  69                 RETURN(1);
  70
  71         if (memcmp(d_name->name, name->name, name->len))
  72                 RETURN(1);
  73
  74         /* XXX: d_name must be in-dentry structure */
  75         dchild = container_of(d_name, struct dentry, d_name); /* ugh */
  76         if (dchild->d_flags & DCACHE_LUSTRE_INVALID) {
  77                 CDEBUG(D_DENTRY,"INVALID dentry %p not matched, was bug 3784\n",
  78                        dchild);
  79                 RETURN(1);
  80         }
  81
  82         RETURN(0);
  83 }
  84
  85 /* should NOT be called with the dcache lock, see fs/dcache.c */
  86 static int ll_ddelete(struct dentry *de)
  87 {
  88         ENTRY;
  89         LASSERT(de);
  90         CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
  91                (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
  92                de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
  93                d_unhashed(de) ? "" : "hashed,",
  94                list_empty(&de->d_subdirs) ? "" : "subdirs");
  95         RETURN(0);
  96 }
  97
  98 void ll_set_dd(struct dentry *de)
  99 {
 100         ENTRY;
 101         LASSERT(de != NULL);
 102
 103         CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
 104                de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
 105                atomic_read(&de->d_count));
 106         lock_kernel();
 107         if (de->d_fsdata == NULL) {
 108                 OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data));
 109         }
 110         unlock_kernel();
 111
 112         EXIT;
 113 }
 114
 115 void ll_intent_drop_lock(struct lookup_intent *it)
 116 {
 117         struct lustre_handle *handle;
 118
 119         if (it->it_op && it->d.lustre.it_lock_mode) {
 120                 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
 121                 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
 122                        " from it %p\n", handle->cookie, it);
 123                 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
 124
 125                 /* bug 494: intent_release may be called multiple times, from
 126                  * this thread and we don't want to double-decref this lock */
 127                 it->d.lustre.it_lock_mode = 0;
 128         }
 129 }
 130
 131 void ll_intent_release(struct lookup_intent *it)
 132 {
 133         ENTRY;
 134
 135         CDEBUG(D_INFO, "intent %p released\n", it);
 136         ll_intent_drop_lock(it);
 137         it->it_magic = 0;
 138         it->it_op_release = 0;
 139         it->d.lustre.it_disposition = 0;
 140         it->d.lustre.it_data = NULL;
 141         EXIT;
 142 }
 143
 144 /* Drop dentry if it is not used already, unhash otherwise.
 145    Should be called with dcache lock held!
 146    Returns: 1 if dentry was dropped, 0 if unhashed. */
 147 int ll_drop_dentry(struct dentry *dentry)
 148 {
 149         lock_dentry(dentry);
 150                 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
 151                        "inode %p flags %d\n", dentry->d_name.len,
 152                        dentry->d_name.name, dentry, dentry->d_parent,
 153                        dentry->d_inode, dentry->d_flags);
 154
 155         if (atomic_read(&dentry->d_count) == 0) {
 156                 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
 157                        "inode %p\n", dentry->d_name.len,
 158                        dentry->d_name.name, dentry, dentry->d_parent,
 159                        dentry->d_inode);
 160                 dget_locked(dentry);
 161                 __d_drop(dentry);
 162                 unlock_dentry(dentry);
 163                 spin_unlock(&dcache_lock);
 164                 dput(dentry);
 165                 spin_lock(&dcache_lock);
 166                 return 1;
 167         }
 168
 169 #ifdef LUSTRE_KERNEL_VERSION
 170         if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
 171 #else
 172         if (!d_unhashed(dentry)) {
 173                 struct inode *inode = dentry->d_inode;
 174 #endif
 175                 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
 176                        "inode %p refc %d\n", dentry->d_name.len,
 177                        dentry->d_name.name, dentry, dentry->d_parent,
 178                        dentry->d_inode, atomic_read(&dentry->d_count));
 179                 /* actually we don't unhash the dentry, rather just
 180                  * mark it inaccessible for to __d_lookup(). otherwise
 181                  * sys_getcwd() could return -ENOENT -bzzz */
 182 #ifdef LUSTRE_KERNEL_VERSION
 183                 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
 184 #else
 185                 if (!inode || !S_ISDIR(inode->i_mode))
 186                         __d_drop(dentry);
 187 #endif
 188
 189 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 190                 __d_drop(dentry);
 191                 if (inode) {
 192                         /* Put positive dentries to orphan list */
 193                         hlist_add_head(&dentry->d_hash,
 194                                        &ll_i2sbi(inode)->ll_orphan_dentry_list);
 195                 }
 196 #endif
 197         }
 198         unlock_dentry(dentry);
 199         return 0;
 200 }
 201
 202 void ll_unhash_aliases(struct inode *inode)
 203 {
 204         struct list_head *tmp, *head;
 205         ENTRY;
 206
 207         if (inode == NULL) {
 208                 CERROR("unexpected NULL inode, tell phil\n");
 209                 return;
 210         }
 211
 212         CDEBUG(D_INODE, "marking dentries for 111 ino %lu/%u(%p) invalid\n",
 213                inode->i_ino, inode->i_generation, inode);
 214
 215         head = &inode->i_dentry;
 216         spin_lock(&dcache_lock);
 217 restart:
 218         tmp = head;
 219         while ((tmp = tmp->next) != head) {
 220                 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
 221
 222                 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
 223                        "inode %p flags %d\n", dentry->d_name.len,
 224                        dentry->d_name.name, dentry, dentry->d_parent,
 225                        dentry->d_inode, dentry->d_flags);
 226
 227                 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
 228                         CERROR("called on root (?) dentry=%p, inode=%p "
 229                                "ino=%lu\n", dentry, inode, inode->i_ino);
 230                         lustre_dump_dentry(dentry, 1);
 231                         libcfs_debug_dumpstack(NULL);
 232                 } else if (d_mountpoint(dentry)) {
 233                         /* For mountpoints we skip removal of the dentry
 234                            which happens solely because we have a lock on it
 235                            obtained when this dentry was not a mountpoint yet */
 236                         CDEBUG(D_DENTRY, "Skippind mountpoint dentry removal "
 237                                          "%.*s (%p) parent %p\n",
 238                                           dentry->d_name.len,
 239                                           dentry->d_name.name,
 240                                           dentry, dentry->d_parent);
 241
 242                         continue;
 243                 }
 244
 245                 if (ll_drop_dentry(dentry))
 246                           goto restart;
 247         }
 248         spin_unlock(&dcache_lock);
 249         EXIT;
 250 }
 251
 252 int ll_revalidate_it_finish(struct ptlrpc_request *request,
 253                             int offset, struct lookup_intent *it,
 254                             struct dentry *de)
 255 {
 256         int rc = 0;
 257         ENTRY;
 258
 259         if (!request)
 260                 RETURN(0);
 261
 262         if (it_disposition(it, DISP_LOOKUP_NEG)) {
 263                 /*Sometimes, revalidate_it might also create node in MDS,
 264                  *So we need check whether it is really created, if not,
 265                  *just return the -ENOENT, if it is, return -ESTALE anyway.
 266                  *which is original done in mdc_intent_lock,
 267                  *but no chance for new fid allocation in client.
 268                  */
 269                 if (it_disposition(it, DISP_OPEN_CREATE) &&
 270                     !it_open_error(DISP_OPEN_CREATE, it)) {
 271                          /*These 2 req finished is for balancing 2 add ref
 272                           *in mdc_intent_lock, since there are no create_node
 273                           *and file_open following revalidate_it*/
 274                          it_set_disposition(it, DISP_ENQ_COMPLETE);
 275                          if (it_disposition(it, DISP_OPEN_CREATE) &&
 276                             !it_open_error(DISP_OPEN_CREATE, it))
 277                                 ptlrpc_req_finished(request);
 278                          if (it_disposition(it, DISP_OPEN_OPEN) &&
 279                             !it_open_error(DISP_OPEN_OPEN, it))
 280                                 ptlrpc_req_finished(request);
 281
 282                          RETURN(-ESTALE);
 283                 }
 284                 RETURN(-ENOENT);
 285         }
 286
 287         rc = ll_prep_inode(&de->d_inode,
 288                            request, offset, NULL);
 289
 290         RETURN(rc);
 291 }
 292
 293 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 294 {
 295         LASSERT(it != NULL);
 296         LASSERT(dentry != NULL);
 297
 298         if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
 299                 struct inode *inode = dentry->d_inode;
 300                 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
 301
 302                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
 303                        inode, inode->i_ino, inode->i_generation);
 304                 md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle,
 305                                  inode);
 306         }
 307
 308         /* drop lookup or getattr locks immediately */
 309         if (it->it_op & (IT_LOOKUP | IT_GETATTR)) {
 310 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 311                 /* on 2.6 there are situation when several lookups and
 312                  * revalidations may be requested during single operation.
 313                  * therefore, we don't release intent here -bzzz */
 314                 ll_intent_drop_lock(it);
 315 #else
 316                 ll_intent_release(it);
 317 #endif
 318         }
 319 }
 320
 321 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 322 {
 323         struct lookup_intent *it = *itp;
 324 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 325         if (it) {
 326                 LASSERTF(it->it_magic == INTENT_MAGIC,
 327                          "%p has bad intent magic: %x\n",
 328                          it, it->it_magic);
 329         }
 330 #endif
 331
 332         if (!it || it->it_op == IT_GETXATTR)
 333                 it = *itp = deft;
 334
 335         it->it_op_release = ll_intent_release;
 336 }
 337
 338 int ll_revalidate_it(struct dentry *de, int lookup_flags,
 339                      struct lookup_intent *it)
 340 {
 341         int rc;
 342         struct it_cb_data icbd;
 343         struct md_op_data *op_data;
 344         struct ptlrpc_request *req = NULL;
 345         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 346         struct obd_export *exp;
 347         struct inode *parent;
 348
 349         ENTRY;
 350         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
 351                LL_IT2STR(it));
 352
 353         /* Cached negative dentries are unsafe for now - look them up again */
 354         if (de->d_inode == NULL)
 355                 RETURN(0);
 356
 357         exp = ll_i2mdexp(de->d_inode);
 358         icbd.icbd_parent = de->d_parent->d_inode;
 359         icbd.icbd_childp = &de;
 360
 361         /* Never execute intents for mount points.
 362          * Attributes will be fixed up in ll_inode_revalidate_it */
 363         if (d_mountpoint(de))
 364                 RETURN(1);
 365
 366         OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
 367         ll_frob_intent(&it, &lookup_it);
 368         LASSERT(it);
 369
 370         parent = de->d_parent->d_inode;
 371
 372         OBD_ALLOC_PTR(op_data);
 373         if (op_data == NULL)
 374                 RETURN(-ENOMEM);
 375
 376         if (it->it_op & IT_CREAT) {
 377                 struct lu_placement_hint hint = { .ph_pname = NULL,
 378                                           .ph_pfid = ll_inode2fid(parent),
 379                                           .ph_cname = &de->d_name,
 380                                           .ph_opc = LUSTRE_OPC_CREATE };
 381
 382                 ll_prepare_md_op_data(op_data, parent, NULL,
 383                                       de->d_name.name, de->d_name.len, 0);
 384                 rc = ll_fid_md_alloc(ll_i2sbi(parent), &op_data->fid2,
 385                                      &hint);
 386                 if (rc) {
 387                         CERROR("can't allocate new fid, rc %d\n", rc);
 388                         LBUG();
 389                 }
 390         } else {
 391                 ll_prepare_md_op_data(op_data, parent, de->d_inode,
 392                                       de->d_name.name, de->d_name.len, 0);
 393         }
 394
 395         rc = md_intent_lock(exp, op_data, NULL, 0, it, lookup_flags,
 396                             &req, ll_md_blocking_ast, 0);
 397
 398         OBD_FREE_PTR(op_data);
 399         /* If req is NULL, then md_intent_lock only tried to do a lock match;
 400          * if all was well, it will return 1 if it found locks, 0 otherwise. */
 401         if (req == NULL && rc >= 0) {
 402                 GOTO(out, rc);
 403         }
 404
 405         if (rc < 0) {
 406                 if (rc != -ESTALE) {
 407                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
 408                                "%d\n", rc, it->d.lustre.it_status);
 409                 }
 410                 GOTO(out, rc = 0);
 411         }
 412
 413         rc = ll_revalidate_it_finish(req, 1, it, de);
 414         if (rc != 0) {
 415                 if (rc != -ESTALE && rc != -ENOENT)
 416                         ll_intent_release(it);
 417                 GOTO(out, rc = 0);
 418         }
 419         if ((it->it_op & IT_OPEN) && de->d_inode &&
 420             !S_ISREG(de->d_inode->i_mode) &&
 421             !S_ISDIR(de->d_inode->i_mode)) {
 422                 ll_release_openhandle(de, it);
 423         }
 424         rc = 1;
 425
 426         /* unfortunately ll_intent_lock may cause a callback and revoke our
 427          * dentry */
 428         spin_lock(&dcache_lock);
 429         lock_dentry(de);
 430         __d_drop(de);
 431         unlock_dentry(de);
 432         __d_rehash(de, 0);
 433         spin_unlock(&dcache_lock);
 434
 435  out:
 436         /* We do not free request as it may be reused during following lookup
 437           (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
 438            be freed in ll_lookup_it or in ll_intent_release. But if
 439            request was not completed, we need to free it. (bug 5154) */
 440         if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
 441                 ptlrpc_req_finished(req);
 442         if (rc == 0) {
 443                 ll_unhash_aliases(de->d_inode);
 444                 /* done in ll_unhash_aliases()
 445                 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
 446         } else {
 447                 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
 448                                "inode %p refc %d\n", de->d_name.len,
 449                                de->d_name.name, de, de->d_parent, de->d_inode,
 450                                atomic_read(&de->d_count));
 451                 ll_lookup_finish_locks(it, de);
 452                 lock_dentry(de);
 453                 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
 454                 unlock_dentry(de);
 455         }
 456         RETURN(rc);
 457 }
 458
 459 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
 460 {
 461         struct inode *inode= de->d_inode;
 462         struct ll_sb_info *sbi = ll_i2sbi(inode);
 463         struct ll_dentry_data *ldd = ll_d2d(de);
 464         struct obd_client_handle *handle;
 465         int rc = 0;
 466         ENTRY;
 467         LASSERT(ldd);
 468
 469         lock_kernel();
 470         /* Strictly speaking this introduces an additional race: the
 471          * increments should wait until the rpc has returned.
 472          * However, given that at present the function is void, this
 473          * issue is moot. */
 474         if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
 475                 unlock_kernel();
 476                 EXIT;
 477                 return;
 478         }
 479
 480         if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
 481                 unlock_kernel();
 482                 EXIT;
 483                 return;
 484         }
 485         unlock_kernel();
 486
 487         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
 488         rc = obd_pin(sbi->ll_md_exp, &ll_i2info(inode)->lli_fid,
 489                      handle, flag);
 490
 491         if (rc) {
 492                 lock_kernel();
 493                 memset(handle, 0, sizeof(*handle));
 494                 if (flag == 0)
 495                         ldd->lld_cwd_count--;
 496                 else
 497                         ldd->lld_mnt_count--;
 498                 unlock_kernel();
 499         }
 500
 501         EXIT;
 502         return;
 503 }
 504
 505 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
 506 {
 507         struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
 508         struct ll_dentry_data *ldd = ll_d2d(de);
 509         struct obd_client_handle handle;
 510         int count, rc = 0;
 511         ENTRY;
 512         LASSERT(ldd);
 513
 514         lock_kernel();
 515         /* Strictly speaking this introduces an additional race: the
 516          * increments should wait until the rpc has returned.
 517          * However, given that at present the function is void, this
 518          * issue is moot. */
 519         handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
 520         if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
 521                 /* the "pin" failed */
 522                 unlock_kernel();
 523                 EXIT;
 524                 return;
 525         }
 526
 527         if (flag)
 528                 count = --ldd->lld_mnt_count;
 529         else
 530                 count = --ldd->lld_cwd_count;
 531         unlock_kernel();
 532
 533         if (count != 0) {
 534                 EXIT;
 535                 return;
 536         }
 537
 538         rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
 539         EXIT;
 540         return;
 541 }
 542
 543 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 544 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 545 {
 546         int rc;
 547         ENTRY;
 548
 549         if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
 550                 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
 551         else
 552                 rc = ll_revalidate_it(dentry, 0, NULL);
 553
 554         RETURN(rc);
 555 }
 556 #endif
 557
 558 struct dentry_operations ll_d_ops = {
 559 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
 560         .d_revalidate = ll_revalidate_nd,
 561 #else
 562         .d_revalidate_it = ll_revalidate_it,
 563 #endif
 564         .d_release = ll_release,
 565         .d_delete = ll_ddelete,
 566         .d_compare = ll_dcompare,
 567 #if 0
 568         .d_pin = ll_pin,
 569         .d_unpin = ll_unpin,
 570 #endif
 571 };