drivers/staging/lustre/lustre/llite/namei.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #include <linux/fs.h>
  34 #include <linux/sched.h>
  35 #include <linux/mm.h>
  36 #include <linux/quotaops.h>
  37 #include <linux/highmem.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/security.h>
  40
  41 #define DEBUG_SUBSYSTEM S_LLITE
  42
  43 #include <obd_support.h>
  44 #include <lustre_fid.h>
  45 #include <lustre_dlm.h>
  46 #include "llite_internal.h"
  47
  48 static int ll_create_it(struct inode *dir, struct dentry *dentry,
  49                         struct lookup_intent *it);
  50
  51 /* called from iget5_locked->find_inode() under inode_hash_lock spinlock */
  52 static int ll_test_inode(struct inode *inode, void *opaque)
  53 {
  54         struct ll_inode_info *lli = ll_i2info(inode);
  55         struct lustre_md     *md = opaque;
  56
  57         if (unlikely(!(md->body->mbo_valid & OBD_MD_FLID))) {
  58                 CERROR("MDS body missing FID\n");
  59                 return 0;
  60         }
  61
  62         if (!lu_fid_eq(&lli->lli_fid, &md->body->mbo_fid1))
  63                 return 0;
  64
  65         return 1;
  66 }
  67
  68 static int ll_set_inode(struct inode *inode, void *opaque)
  69 {
  70         struct ll_inode_info *lli = ll_i2info(inode);
  71         struct mdt_body *body = ((struct lustre_md *)opaque)->body;
  72
  73         if (unlikely(!(body->mbo_valid & OBD_MD_FLID))) {
  74                 CERROR("MDS body missing FID\n");
  75                 return -EINVAL;
  76         }
  77
  78         lli->lli_fid = body->mbo_fid1;
  79         if (unlikely(!(body->mbo_valid & OBD_MD_FLTYPE))) {
  80                 CERROR("Can not initialize inode " DFID
  81                        " without object type: valid = %#llx\n",
  82                        PFID(&lli->lli_fid), body->mbo_valid);
  83                 return -EINVAL;
  84         }
  85
  86         inode->i_mode = (inode->i_mode & ~S_IFMT) | (body->mbo_mode & S_IFMT);
  87         if (unlikely(inode->i_mode == 0)) {
  88                 CERROR("Invalid inode " DFID " type\n", PFID(&lli->lli_fid));
  89                 return -EINVAL;
  90         }
  91
  92         ll_lli_init(lli);
  93
  94         return 0;
  95 }
  96
  97 /**
  98  * Get an inode by inode number(@hash), which is already instantiated by
  99  * the intent lookup).
 100  */
 101 struct inode *ll_iget(struct super_block *sb, ino_t hash,
 102                       struct lustre_md *md)
 103 {
 104         struct inode     *inode;
 105         int rc = 0;
 106
 107         LASSERT(hash != 0);
 108         inode = iget5_locked(sb, hash, ll_test_inode, ll_set_inode, md);
 109         if (!inode)
 110                 return ERR_PTR(-ENOMEM);
 111
 112         if (inode->i_state & I_NEW) {
 113                 rc = ll_read_inode2(inode, md);
 114                 if (!rc && S_ISREG(inode->i_mode) &&
 115                     !ll_i2info(inode)->lli_clob)
 116                         rc = cl_file_inode_init(inode, md);
 117
 118                 if (rc) {
 119                         /*
 120                          * Let's clear directory lsm here, otherwise
 121                          * make_bad_inode() will reset the inode mode
 122                          * to regular, then ll_clear_inode will not
 123                          * be able to clear lsm_md
 124                          */
 125                         if (S_ISDIR(inode->i_mode))
 126                                 ll_dir_clear_lsm_md(inode);
 127                         make_bad_inode(inode);
 128                         unlock_new_inode(inode);
 129                         iput(inode);
 130                         inode = ERR_PTR(rc);
 131                 } else {
 132                         unlock_new_inode(inode);
 133                 }
 134         } else if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
 135                 rc = ll_update_inode(inode, md);
 136                 CDEBUG(D_VFSTRACE, "got inode: " DFID "(%p): rc = %d\n",
 137                        PFID(&md->body->mbo_fid1), inode, rc);
 138                 if (rc) {
 139                         if (S_ISDIR(inode->i_mode))
 140                                 ll_dir_clear_lsm_md(inode);
 141                         iput(inode);
 142                         inode = ERR_PTR(rc);
 143                 }
 144         }
 145         return inode;
 146 }
 147
 148 static void ll_invalidate_negative_children(struct inode *dir)
 149 {
 150         struct dentry *dentry, *tmp_subdir;
 151
 152         spin_lock(&dir->i_lock);
 153         hlist_for_each_entry(dentry, &dir->i_dentry, d_u.d_alias) {
 154                 spin_lock(&dentry->d_lock);
 155                 if (!list_empty(&dentry->d_subdirs)) {
 156                         struct dentry *child;
 157
 158                         list_for_each_entry_safe(child, tmp_subdir,
 159                                                  &dentry->d_subdirs,
 160                                                  d_child) {
 161                                 if (d_really_is_negative(child))
 162                                         d_lustre_invalidate(child, 1);
 163                         }
 164                 }
 165                 spin_unlock(&dentry->d_lock);
 166         }
 167         spin_unlock(&dir->i_lock);
 168 }
 169
 170 int ll_test_inode_by_fid(struct inode *inode, void *opaque)
 171 {
 172         return lu_fid_eq(&ll_i2info(inode)->lli_fid, opaque);
 173 }
 174
 175 int ll_md_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 176                        void *data, int flag)
 177 {
 178         struct lustre_handle lockh;
 179         int rc;
 180
 181         switch (flag) {
 182         case LDLM_CB_BLOCKING:
 183                 ldlm_lock2handle(lock, &lockh);
 184                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 185                 if (rc < 0) {
 186                         CDEBUG(D_INODE, "ldlm_cli_cancel: rc = %d\n", rc);
 187                         return rc;
 188                 }
 189                 break;
 190         case LDLM_CB_CANCELING: {
 191                 struct inode *inode = ll_inode_from_resource_lock(lock);
 192                 __u64 bits = lock->l_policy_data.l_inodebits.bits;
 193
 194                 /* Inode is set to lock->l_resource->lr_lvb_inode
 195                  * for mdc - bug 24555
 196                  */
 197                 LASSERT(!lock->l_ast_data);
 198
 199                 if (!inode)
 200                         break;
 201
 202                 /* Invalidate all dentries associated with this inode */
 203                 LASSERT(ldlm_is_canceling(lock));
 204
 205                 if (!fid_res_name_eq(ll_inode2fid(inode),
 206                                      &lock->l_resource->lr_name)) {
 207                         LDLM_ERROR(lock, "data mismatch with object " DFID "(%p)",
 208                                    PFID(ll_inode2fid(inode)), inode);
 209                         LBUG();
 210                 }
 211
 212                 if (bits & MDS_INODELOCK_XATTR) {
 213                         if (S_ISDIR(inode->i_mode))
 214                                 ll_i2info(inode)->lli_def_stripe_offset = -1;
 215                         ll_xattr_cache_destroy(inode);
 216                         bits &= ~MDS_INODELOCK_XATTR;
 217                 }
 218
 219                 /* For OPEN locks we differentiate between lock modes
 220                  * LCK_CR, LCK_CW, LCK_PR - bug 22891
 221                  */
 222                 if (bits & MDS_INODELOCK_OPEN)
 223                         ll_have_md_lock(inode, &bits, lock->l_req_mode);
 224
 225                 if (bits & MDS_INODELOCK_OPEN) {
 226                         fmode_t fmode;
 227
 228                         switch (lock->l_req_mode) {
 229                         case LCK_CW:
 230                                 fmode = FMODE_WRITE;
 231                                 break;
 232                         case LCK_PR:
 233                                 fmode = FMODE_EXEC;
 234                                 break;
 235                         case LCK_CR:
 236                                 fmode = FMODE_READ;
 237                                 break;
 238                         default:
 239                                 LDLM_ERROR(lock, "bad lock mode for OPEN lock");
 240                                 LBUG();
 241                         }
 242
 243                         ll_md_real_close(inode, fmode);
 244                 }
 245
 246                 if (bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
 247                             MDS_INODELOCK_LAYOUT | MDS_INODELOCK_PERM))
 248                         ll_have_md_lock(inode, &bits, LCK_MINMODE);
 249
 250                 if (bits & MDS_INODELOCK_LAYOUT) {
 251                         struct cl_object_conf conf = {
 252                                 .coc_opc = OBJECT_CONF_INVALIDATE,
 253                                 .coc_inode = inode,
 254                         };
 255
 256                         rc = ll_layout_conf(inode, &conf);
 257                         if (rc < 0)
 258                                 CDEBUG(D_INODE, "cannot invalidate layout of "
 259                                        DFID ": rc = %d\n",
 260                                        PFID(ll_inode2fid(inode)), rc);
 261                 }
 262
 263                 if (bits & MDS_INODELOCK_UPDATE) {
 264                         struct ll_inode_info *lli = ll_i2info(inode);
 265
 266                         spin_lock(&lli->lli_lock);
 267                         LTIME_S(inode->i_mtime) = 0;
 268                         LTIME_S(inode->i_atime) = 0;
 269                         LTIME_S(inode->i_ctime) = 0;
 270                         spin_unlock(&lli->lli_lock);
 271                 }
 272
 273                 if ((bits & MDS_INODELOCK_UPDATE) && S_ISDIR(inode->i_mode)) {
 274                         struct ll_inode_info *lli = ll_i2info(inode);
 275
 276                         CDEBUG(D_INODE, "invalidating inode " DFID " lli = %p, pfid  = " DFID "\n",
 277                                PFID(ll_inode2fid(inode)), lli,
 278                                PFID(&lli->lli_pfid));
 279
 280                         truncate_inode_pages(inode->i_mapping, 0);
 281
 282                         if (unlikely(!fid_is_zero(&lli->lli_pfid))) {
 283                                 struct inode *master_inode = NULL;
 284                                 unsigned long hash;
 285
 286                                 /*
 287                                  * This is slave inode, since all of the child
 288                                  * dentry is connected on the master inode, so
 289                                  * we have to invalidate the negative children
 290                                  * on master inode
 291                                  */
 292                                 CDEBUG(D_INODE, "Invalidate s" DFID " m" DFID "\n",
 293                                        PFID(ll_inode2fid(inode)),
 294                                        PFID(&lli->lli_pfid));
 295
 296                                 hash = cl_fid_build_ino(&lli->lli_pfid,
 297                                                         ll_need_32bit_api(ll_i2sbi(inode)));
 298                                 /*
 299                                  * Do not lookup the inode with ilookup5,
 300                                  * otherwise it will cause dead lock,
 301                                  *
 302                                  * 1. Client1 send chmod req to the MDT0, then
 303                                  * on MDT0, it enqueues master and all of its
 304                                  * slaves lock, (mdt_attr_set() ->
 305                                  * mdt_lock_slaves()), after gets master and
 306                                  * stripe0 lock, it will send the enqueue req
 307                                  * (for stripe1) to MDT1, then MDT1 finds the
 308                                  * lock has been granted to client2. Then MDT1
 309                                  * sends blocking ast to client2.
 310                                  *
 311                                  * 2. At the same time, client2 tries to unlink
 312                                  * the striped dir (rm -rf striped_dir), and
 313                                  * during lookup, it will hold the master inode
 314                                  * of the striped directory, whose inode state
 315                                  * is NEW, then tries to revalidate all of its
 316                                  * slaves, (ll_prep_inode()->ll_iget()->
 317                                  * ll_read_inode2()-> ll_update_inode().). And
 318                                  * it will be blocked on the server side because
 319                                  * of 1.
 320                                  *
 321                                  * 3. Then the client get the blocking_ast req,
 322                                  * cancel the lock, but being blocked if using
 323                                  * ->ilookup5()), because master inode state is
 324                                  *  NEW.
 325                                  */
 326                                 master_inode = ilookup5_nowait(inode->i_sb,
 327                                                                hash,
 328                                                                ll_test_inode_by_fid,
 329                                                                (void *)&lli->lli_pfid);
 330                                 if (master_inode) {
 331                                         ll_invalidate_negative_children(master_inode);
 332                                         iput(master_inode);
 333                                 }
 334                         } else {
 335                                 ll_invalidate_negative_children(inode);
 336                         }
 337                 }
 338
 339                 if ((bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM)) &&
 340                     inode->i_sb->s_root &&
 341                     !is_root_inode(inode))
 342                         ll_invalidate_aliases(inode);
 343
 344                 iput(inode);
 345                 break;
 346         }
 347         default:
 348                 LBUG();
 349         }
 350
 351         return 0;
 352 }
 353
 354 __u32 ll_i2suppgid(struct inode *i)
 355 {
 356         if (in_group_p(i->i_gid))
 357                 return (__u32)from_kgid(&init_user_ns, i->i_gid);
 358         else
 359                 return (__u32)(-1);
 360 }
 361
 362 /* Pack the required supplementary groups into the supplied groups array.
 363  * If we don't need to use the groups from the target inode(s) then we
 364  * instead pack one or more groups from the user's supplementary group
 365  * array in case it might be useful.  Not needed if doing an MDS-side upcall.
 366  */
 367 void ll_i2gids(__u32 *suppgids, struct inode *i1, struct inode *i2)
 368 {
 369         LASSERT(i1);
 370
 371         suppgids[0] = ll_i2suppgid(i1);
 372
 373         if (i2)
 374                 suppgids[1] = ll_i2suppgid(i2);
 375                 else
 376                         suppgids[1] = -1;
 377 }
 378
 379 /*
 380  * try to reuse three types of dentry:
 381  * 1. unhashed alias, this one is unhashed by d_invalidate (but it may be valid
 382  *    by concurrent .revalidate).
 383  * 2. INVALID alias (common case for no valid ldlm lock held, but this flag may
 384  *    be cleared by others calling d_lustre_revalidate).
 385  * 3. DISCONNECTED alias.
 386  */
 387 static struct dentry *ll_find_alias(struct inode *inode, struct dentry *dentry)
 388 {
 389         struct dentry *alias, *discon_alias, *invalid_alias;
 390
 391         if (hlist_empty(&inode->i_dentry))
 392                 return NULL;
 393
 394         discon_alias = NULL;
 395         invalid_alias = NULL;
 396
 397         spin_lock(&inode->i_lock);
 398         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
 399                 LASSERT(alias != dentry);
 400
 401                 spin_lock(&alias->d_lock);
 402                 if ((alias->d_flags & DCACHE_DISCONNECTED) &&
 403                     S_ISDIR(inode->i_mode))
 404                         /* LASSERT(last_discon == NULL); LU-405, bz 20055 */
 405                         discon_alias = alias;
 406                 else if (alias->d_parent == dentry->d_parent         &&
 407                          alias->d_name.hash == dentry->d_name.hash       &&
 408                          alias->d_name.len == dentry->d_name.len         &&
 409                          memcmp(alias->d_name.name, dentry->d_name.name,
 410                                 dentry->d_name.len) == 0)
 411                         invalid_alias = alias;
 412                 spin_unlock(&alias->d_lock);
 413
 414                 if (invalid_alias)
 415                         break;
 416         }
 417         alias = invalid_alias ?: discon_alias ?: NULL;
 418         if (alias) {
 419                 spin_lock(&alias->d_lock);
 420                 dget_dlock(alias);
 421                 spin_unlock(&alias->d_lock);
 422         }
 423         spin_unlock(&inode->i_lock);
 424
 425         return alias;
 426 }
 427
 428 /*
 429  * Similar to d_splice_alias(), but lustre treats invalid alias
 430  * similar to DCACHE_DISCONNECTED, and tries to use it anyway.
 431  */
 432 struct dentry *ll_splice_alias(struct inode *inode, struct dentry *de)
 433 {
 434         if (inode) {
 435                 struct dentry *new = ll_find_alias(inode, de);
 436
 437                 if (new) {
 438                         d_move(new, de);
 439                         iput(inode);
 440                         CDEBUG(D_DENTRY,
 441                                "Reuse dentry %p inode %p refc %d flags %#x\n",
 442                               new, d_inode(new), d_count(new), new->d_flags);
 443                         return new;
 444                 }
 445         }
 446         d_add(de, inode);
 447         CDEBUG(D_DENTRY, "Add dentry %p inode %p refc %d flags %#x\n",
 448                de, d_inode(de), d_count(de), de->d_flags);
 449         return de;
 450 }
 451
 452 static int ll_lookup_it_finish(struct ptlrpc_request *request,
 453                                struct lookup_intent *it,
 454                                struct inode *parent, struct dentry **de)
 455 {
 456         struct inode *inode = NULL;
 457         __u64 bits = 0;
 458         int rc = 0;
 459         struct dentry *alias;
 460
 461         /* NB 1 request reference will be taken away by ll_intent_lock()
 462          * when I return
 463          */
 464         CDEBUG(D_DENTRY, "it %p it_disposition %x\n", it,
 465                it->it_disposition);
 466         if (!it_disposition(it, DISP_LOOKUP_NEG)) {
 467                 rc = ll_prep_inode(&inode, request, (*de)->d_sb, it);
 468                 if (rc)
 469                         return rc;
 470
 471                 ll_set_lock_data(ll_i2sbi(parent)->ll_md_exp, inode, it, &bits);
 472
 473                 /* We used to query real size from OSTs here, but actually
 474                  * this is not needed. For stat() calls size would be updated
 475                  * from subsequent do_revalidate()->ll_inode_revalidate_it() in
 476                  * 2.4 and
 477                  * vfs_getattr_it->ll_getattr()->ll_inode_revalidate_it() in 2.6
 478                  * Everybody else who needs correct file size would call
 479                  * ll_glimpse_size or some equivalent themselves anyway.
 480                  * Also see bug 7198.
 481                  */
 482         }
 483
 484         alias = ll_splice_alias(inode, *de);
 485         if (IS_ERR(alias)) {
 486                 rc = PTR_ERR(alias);
 487                 goto out;
 488         }
 489         *de = alias;
 490
 491         if (!it_disposition(it, DISP_LOOKUP_NEG)) {
 492                 /* We have the "lookup" lock, so unhide dentry */
 493                 if (bits & MDS_INODELOCK_LOOKUP)
 494                         d_lustre_revalidate(*de);
 495         } else if (!it_disposition(it, DISP_OPEN_CREATE)) {
 496                 /* If file created on server, don't depend on parent UPDATE
 497                  * lock to unhide it. It is left hidden and next lookup can
 498                  * find it in ll_splice_alias.
 499                  */
 500                 /* Check that parent has UPDATE lock. */
 501                 struct lookup_intent parent_it = {
 502                                         .it_op = IT_GETATTR,
 503                                         .it_lock_handle = 0 };
 504                 struct lu_fid fid = ll_i2info(parent)->lli_fid;
 505
 506                 /* If it is striped directory, get the real stripe parent */
 507                 if (unlikely(ll_i2info(parent)->lli_lsm_md)) {
 508                         rc = md_get_fid_from_lsm(ll_i2mdexp(parent),
 509                                                  ll_i2info(parent)->lli_lsm_md,
 510                                                  (*de)->d_name.name,
 511                                                  (*de)->d_name.len, &fid);
 512                         if (rc)
 513                                 return rc;
 514                 }
 515
 516                 if (md_revalidate_lock(ll_i2mdexp(parent), &parent_it, &fid,
 517                                        NULL)) {
 518                         d_lustre_revalidate(*de);
 519                         ll_intent_release(&parent_it);
 520                 }
 521         }
 522
 523 out:
 524         if (rc != 0 && it->it_op & IT_OPEN)
 525                 ll_open_cleanup((*de)->d_sb, request);
 526
 527         return rc;
 528 }
 529
 530 static struct dentry *ll_lookup_it(struct inode *parent, struct dentry *dentry,
 531                                    struct lookup_intent *it)
 532 {
 533         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 534         struct dentry *save = dentry, *retval;
 535         struct ptlrpc_request *req = NULL;
 536         struct md_op_data *op_data = NULL;
 537         struct inode *inode;
 538         __u32 opc;
 539         int rc;
 540
 541         if (dentry->d_name.len > ll_i2sbi(parent)->ll_namelen)
 542                 return ERR_PTR(-ENAMETOOLONG);
 543
 544         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p),intent=%s\n",
 545                dentry, PFID(ll_inode2fid(parent)), parent, LL_IT2STR(it));
 546
 547         if (d_mountpoint(dentry))
 548                 CERROR("Tell Peter, lookup on mtpt, it %s\n", LL_IT2STR(it));
 549
 550         if (!it || it->it_op == IT_GETXATTR)
 551                 it = &lookup_it;
 552
 553         if (it->it_op == IT_GETATTR && dentry_may_statahead(parent, dentry)) {
 554                 rc = ll_statahead(parent, &dentry, 0);
 555                 if (rc == 1) {
 556                         if (dentry == save)
 557                                 retval = NULL;
 558                         else
 559                                 retval = dentry;
 560                         goto out;
 561                 }
 562         }
 563
 564         if (it->it_op & IT_OPEN && it->it_flags & FMODE_WRITE && sb_rdonly(dentry->d_sb))
 565                 return ERR_PTR(-EROFS);
 566
 567         if (it->it_op & IT_CREAT)
 568                 opc = LUSTRE_OPC_CREATE;
 569         else
 570                 opc = LUSTRE_OPC_ANY;
 571
 572         op_data = ll_prep_md_op_data(NULL, parent, NULL, dentry->d_name.name,
 573                                      dentry->d_name.len, 0, opc, NULL);
 574         if (IS_ERR(op_data))
 575                 return (void *)op_data;
 576
 577         /* enforce umask if acl disabled or MDS doesn't support umask */
 578         if (!IS_POSIXACL(parent) || !exp_connect_umask(ll_i2mdexp(parent)))
 579                 it->it_create_mode &= ~current_umask();
 580
 581         rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
 582                             &ll_md_blocking_ast, 0);
 583         /*
 584          * If the MDS allows the client to chgrp (CFS_SETGRP_PERM), but the
 585          * client does not know which suppgid should be sent to the MDS, or
 586          * some other(s) changed the target file's GID after this RPC sent
 587          * to the MDS with the suppgid as the original GID, then we should
 588          * try again with right suppgid.
 589          */
 590         if (rc == -EACCES && it->it_op & IT_OPEN &&
 591             it_disposition(it, DISP_OPEN_DENY)) {
 592                 struct mdt_body *body;
 593
 594                 LASSERT(req);
 595
 596                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 597                 if (op_data->op_suppgids[0] == body->mbo_gid ||
 598                     op_data->op_suppgids[1] == body->mbo_gid ||
 599                     !in_group_p(make_kgid(&init_user_ns, body->mbo_gid))) {
 600                         retval = ERR_PTR(-EACCES);
 601                         goto out;
 602                 }
 603
 604                 fid_zero(&op_data->op_fid2);
 605                 op_data->op_suppgids[1] = body->mbo_gid;
 606                 ptlrpc_req_finished(req);
 607                 req = NULL;
 608                 ll_intent_release(it);
 609                 rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
 610                                     ll_md_blocking_ast, 0);
 611         }
 612
 613         if (rc < 0) {
 614                 retval = ERR_PTR(rc);
 615                 goto out;
 616         }
 617
 618         rc = ll_lookup_it_finish(req, it, parent, &dentry);
 619         if (rc != 0) {
 620                 ll_intent_release(it);
 621                 retval = ERR_PTR(rc);
 622                 goto out;
 623         }
 624
 625         inode = d_inode(dentry);
 626         if ((it->it_op & IT_OPEN) && inode &&
 627             !S_ISREG(inode->i_mode) &&
 628             !S_ISDIR(inode->i_mode)) {
 629                 ll_release_openhandle(inode, it);
 630         }
 631         ll_lookup_finish_locks(it, inode);
 632
 633         if (dentry == save)
 634                 retval = NULL;
 635         else
 636                 retval = dentry;
 637 out:
 638         if (op_data && !IS_ERR(op_data))
 639                 ll_finish_md_op_data(op_data);
 640
 641         ptlrpc_req_finished(req);
 642         return retval;
 643 }
 644
 645 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
 646                                    unsigned int flags)
 647 {
 648         struct lookup_intent *itp, it = { .it_op = IT_GETATTR };
 649         struct dentry *de;
 650
 651         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p),flags=%u\n",
 652                dentry, PFID(ll_inode2fid(parent)), parent, flags);
 653
 654         /* Optimize away (CREATE && !OPEN). Let .create handle the race.
 655          * but only if we have write permissions there, otherwise we need
 656          * to proceed with lookup. LU-4185
 657          */
 658         if ((flags & LOOKUP_CREATE) && !(flags & LOOKUP_OPEN) &&
 659             (inode_permission(parent, MAY_WRITE | MAY_EXEC) == 0))
 660                 return NULL;
 661
 662         if (flags & (LOOKUP_PARENT | LOOKUP_OPEN | LOOKUP_CREATE))
 663                 itp = NULL;
 664         else
 665                 itp = &it;
 666         de = ll_lookup_it(parent, dentry, itp);
 667
 668         if (itp)
 669                 ll_intent_release(itp);
 670
 671         return de;
 672 }
 673
 674 /*
 675  * For cached negative dentry and new dentry, handle lookup/create/open
 676  * together.
 677  */
 678 static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
 679                           struct file *file, unsigned int open_flags,
 680                           umode_t mode, int *opened)
 681 {
 682         struct lookup_intent *it;
 683         struct dentry *de;
 684         int rc = 0;
 685
 686         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p),file %p,open_flags %x,mode %x opened %d\n",
 687                dentry, PFID(ll_inode2fid(dir)), dir, file, open_flags, mode,
 688                *opened);
 689
 690         /* Only negative dentries enter here */
 691         LASSERT(!d_inode(dentry));
 692
 693         if (!d_in_lookup(dentry)) {
 694                 /* A valid negative dentry that just passed revalidation,
 695                  * there's little point to try and open it server-side,
 696                  * even though there's a minuscle chance it might succeed.
 697                  * Either way it's a valid race to just return -ENOENT here.
 698                  */
 699                 if (!(open_flags & O_CREAT))
 700                         return -ENOENT;
 701
 702                 /* Otherwise we just unhash it to be rehashed afresh via
 703                  * lookup if necessary
 704                  */
 705                 d_drop(dentry);
 706         }
 707
 708         it = kzalloc(sizeof(*it), GFP_NOFS);
 709         if (!it)
 710                 return -ENOMEM;
 711
 712         it->it_op = IT_OPEN;
 713         if (open_flags & O_CREAT)
 714                 it->it_op |= IT_CREAT;
 715         it->it_create_mode = (mode & S_IALLUGO) | S_IFREG;
 716         it->it_flags = (open_flags & ~O_ACCMODE) | OPEN_FMODE(open_flags);
 717         it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
 718
 719         /* Dentry added to dcache tree in ll_lookup_it */
 720         de = ll_lookup_it(dir, dentry, it);
 721         if (IS_ERR(de))
 722                 rc = PTR_ERR(de);
 723         else if (de)
 724                 dentry = de;
 725
 726         if (!rc) {
 727                 if (it_disposition(it, DISP_OPEN_CREATE)) {
 728                         /* Dentry instantiated in ll_create_it. */
 729                         rc = ll_create_it(dir, dentry, it);
 730                         if (rc) {
 731                                 /* We dget in ll_splice_alias. */
 732                                 if (de)
 733                                         dput(de);
 734                                 goto out_release;
 735                         }
 736
 737                         *opened |= FILE_CREATED;
 738                 }
 739                 if (d_really_is_positive(dentry) && it_disposition(it, DISP_OPEN_OPEN)) {
 740                         /* Open dentry. */
 741                         if (S_ISFIFO(d_inode(dentry)->i_mode)) {
 742                                 /* We cannot call open here as it might
 743                                  * deadlock. This case is unreachable in
 744                                  * practice because of OBD_CONNECT_NODEVOH.
 745                                  */
 746                                 rc = finish_no_open(file, de);
 747                         } else {
 748                                 file->private_data = it;
 749                                 rc = finish_open(file, dentry, NULL, opened);
 750                                 /* We dget in ll_splice_alias. finish_open takes
 751                                  * care of dget for fd open.
 752                                  */
 753                                 if (de)
 754                                         dput(de);
 755                         }
 756                 } else {
 757                         rc = finish_no_open(file, de);
 758                 }
 759         }
 760
 761 out_release:
 762         ll_intent_release(it);
 763         kfree(it);
 764
 765         return rc;
 766 }
 767
 768 /* We depend on "mode" being set with the proper file type/umask by now */
 769 static struct inode *ll_create_node(struct inode *dir, struct lookup_intent *it)
 770 {
 771         struct inode *inode = NULL;
 772         struct ptlrpc_request *request = NULL;
 773         struct ll_sb_info *sbi = ll_i2sbi(dir);
 774         int rc;
 775
 776         LASSERT(it && it->it_disposition);
 777
 778         LASSERT(it_disposition(it, DISP_ENQ_CREATE_REF));
 779         request = it->it_request;
 780         it_clear_disposition(it, DISP_ENQ_CREATE_REF);
 781         rc = ll_prep_inode(&inode, request, dir->i_sb, it);
 782         if (rc) {
 783                 inode = ERR_PTR(rc);
 784                 goto out;
 785         }
 786
 787         LASSERT(hlist_empty(&inode->i_dentry));
 788
 789         /* We asked for a lock on the directory, but were granted a
 790          * lock on the inode.  Since we finally have an inode pointer,
 791          * stuff it in the lock.
 792          */
 793         CDEBUG(D_DLMTRACE, "setting l_ast_data to inode " DFID "(%p)\n",
 794                PFID(ll_inode2fid(dir)), inode);
 795         ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
 796  out:
 797         ptlrpc_req_finished(request);
 798         return inode;
 799 }
 800
 801 /*
 802  * By the time this is called, we already have created the directory cache
 803  * entry for the new file, but it is so far negative - it has no inode.
 804  *
 805  * We defer creating the OBD object(s) until open, to keep the intent and
 806  * non-intent code paths similar, and also because we do not have the MDS
 807  * inode number before calling ll_create_node() (which is needed for LOV),
 808  * so we would need to do yet another RPC to the MDS to store the LOV EA
 809  * data on the MDS.  If needed, we would pass the PACKED lmm as data and
 810  * lmm_size in datalen (the MDS still has code which will handle that).
 811  *
 812  * If the create succeeds, we fill in the inode information
 813  * with d_instantiate().
 814  */
 815 static int ll_create_it(struct inode *dir, struct dentry *dentry,
 816                         struct lookup_intent *it)
 817 {
 818         struct inode *inode;
 819         int rc = 0;
 820
 821         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p), intent=%s\n",
 822                dentry, PFID(ll_inode2fid(dir)), dir, LL_IT2STR(it));
 823
 824         rc = it_open_error(DISP_OPEN_CREATE, it);
 825         if (rc)
 826                 return rc;
 827
 828         inode = ll_create_node(dir, it);
 829         if (IS_ERR(inode))
 830                 return PTR_ERR(inode);
 831
 832         d_instantiate(dentry, inode);
 833
 834         return ll_init_security(dentry, inode, dir);
 835 }
 836
 837 void ll_update_times(struct ptlrpc_request *request, struct inode *inode)
 838 {
 839         struct mdt_body *body = req_capsule_server_get(&request->rq_pill,
 840                                                        &RMF_MDT_BODY);
 841
 842         LASSERT(body);
 843         if (body->mbo_valid & OBD_MD_FLMTIME &&
 844             body->mbo_mtime > LTIME_S(inode->i_mtime)) {
 845                 CDEBUG(D_INODE, "setting fid " DFID " mtime from %lu to %llu\n",
 846                        PFID(ll_inode2fid(inode)), LTIME_S(inode->i_mtime),
 847                        body->mbo_mtime);
 848                 LTIME_S(inode->i_mtime) = body->mbo_mtime;
 849         }
 850         if (body->mbo_valid & OBD_MD_FLCTIME &&
 851             body->mbo_ctime > LTIME_S(inode->i_ctime))
 852                 LTIME_S(inode->i_ctime) = body->mbo_ctime;
 853 }
 854
 855 static int ll_new_node(struct inode *dir, struct dentry *dentry,
 856                        const char *tgt, umode_t mode, int rdev,
 857                        __u32 opc)
 858 {
 859         struct ptlrpc_request *request = NULL;
 860         struct md_op_data *op_data;
 861         struct inode *inode = NULL;
 862         struct ll_sb_info *sbi = ll_i2sbi(dir);
 863         int tgt_len = 0;
 864         int err;
 865
 866         if (unlikely(tgt))
 867                 tgt_len = strlen(tgt) + 1;
 868 again:
 869         op_data = ll_prep_md_op_data(NULL, dir, NULL,
 870                                      dentry->d_name.name,
 871                                      dentry->d_name.len,
 872                                      0, opc, NULL);
 873         if (IS_ERR(op_data)) {
 874                 err = PTR_ERR(op_data);
 875                 goto err_exit;
 876         }
 877
 878         err = md_create(sbi->ll_md_exp, op_data, tgt, tgt_len, mode,
 879                         from_kuid(&init_user_ns, current_fsuid()),
 880                         from_kgid(&init_user_ns, current_fsgid()),
 881                         cfs_curproc_cap_pack(), rdev, &request);
 882         ll_finish_md_op_data(op_data);
 883         if (err < 0 && err != -EREMOTE)
 884                 goto err_exit;
 885
 886         /*
 887          * If the client doesn't know where to create a subdirectory (or
 888          * in case of a race that sends the RPC to the wrong MDS), the
 889          * MDS will return -EREMOTE and the client will fetch the layout
 890          * of the directory, then create the directory on the right MDT.
 891          */
 892         if (unlikely(err == -EREMOTE)) {
 893                 struct ll_inode_info *lli = ll_i2info(dir);
 894                 struct lmv_user_md *lum;
 895                 int lumsize, err2;
 896
 897                 ptlrpc_req_finished(request);
 898                 request = NULL;
 899
 900                 err2 = ll_dir_getstripe(dir, (void **)&lum, &lumsize, &request,
 901                                         OBD_MD_DEFAULT_MEA);
 902                 if (!err2) {
 903                         /* Update stripe_offset and retry */
 904                         lli->lli_def_stripe_offset = lum->lum_stripe_offset;
 905                 } else if (err2 == -ENODATA &&
 906                            lli->lli_def_stripe_offset != -1) {
 907                         /*
 908                          * If there are no default stripe EA on the MDT, but the
 909                          * client has default stripe, then it probably means
 910                          * default stripe EA has just been deleted.
 911                          */
 912                         lli->lli_def_stripe_offset = -1;
 913                 } else {
 914                         goto err_exit;
 915                 }
 916
 917                 ptlrpc_req_finished(request);
 918                 request = NULL;
 919                 goto again;
 920         }
 921
 922         ll_update_times(request, dir);
 923
 924         err = ll_prep_inode(&inode, request, dir->i_sb, NULL);
 925         if (err)
 926                 goto err_exit;
 927
 928         d_instantiate(dentry, inode);
 929
 930         err = ll_init_security(dentry, inode, dir);
 931 err_exit:
 932         if (request)
 933                 ptlrpc_req_finished(request);
 934
 935         return err;
 936 }
 937
 938 static int ll_mknod(struct inode *dir, struct dentry *dchild,
 939                     umode_t mode, dev_t rdev)
 940 {
 941         int err;
 942
 943         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p) mode %o dev %x\n",
 944                dchild, PFID(ll_inode2fid(dir)), dir, mode,
 945                old_encode_dev(rdev));
 946
 947         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
 948                 mode &= ~current_umask();
 949
 950         switch (mode & S_IFMT) {
 951         case 0:
 952                 mode |= S_IFREG; /* for mode = 0 case, fallthrough */
 953         case S_IFREG:
 954         case S_IFCHR:
 955         case S_IFBLK:
 956         case S_IFIFO:
 957         case S_IFSOCK:
 958                 err = ll_new_node(dir, dchild, NULL, mode,
 959                                   old_encode_dev(rdev),
 960                                   LUSTRE_OPC_MKNOD);
 961                 break;
 962         case S_IFDIR:
 963                 err = -EPERM;
 964                 break;
 965         default:
 966                 err = -EINVAL;
 967         }
 968
 969         if (!err)
 970                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKNOD, 1);
 971
 972         return err;
 973 }
 974
 975 /*
 976  * Plain create. Intent create is handled in atomic_open.
 977  */
 978 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
 979                         umode_t mode, bool want_excl)
 980 {
 981         int rc;
 982
 983         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p), flags=%u, excl=%d\n",
 984                dentry, PFID(ll_inode2fid(dir)), dir, mode, want_excl);
 985
 986         rc = ll_mknod(dir, dentry, mode, 0);
 987
 988         ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE, 1);
 989
 990         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, unhashed %d\n",
 991                dentry, d_unhashed(dentry));
 992
 993         return rc;
 994 }
 995
 996 static int ll_unlink(struct inode *dir, struct dentry *dchild)
 997 {
 998         struct ptlrpc_request *request = NULL;
 999         struct md_op_data *op_data;
1000         int rc;
1001
1002         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd,dir=%lu/%u(%p)\n",
1003                dchild, dir->i_ino, dir->i_generation, dir);
1004
1005         op_data = ll_prep_md_op_data(NULL, dir, NULL,
1006                                      dchild->d_name.name,
1007                                      dchild->d_name.len,
1008                                      0, LUSTRE_OPC_ANY, NULL);
1009         if (IS_ERR(op_data))
1010                 return PTR_ERR(op_data);
1011
1012         if (dchild->d_inode)
1013                 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1014
1015         op_data->op_fid2 = op_data->op_fid3;
1016         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1017         ll_finish_md_op_data(op_data);
1018         if (rc)
1019                 goto out;
1020
1021         ll_update_times(request, dir);
1022         ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_UNLINK, 1);
1023
1024  out:
1025         ptlrpc_req_finished(request);
1026         return rc;
1027 }
1028
1029 static int ll_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
1030 {
1031         int err;
1032
1033         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir" DFID "(%p)\n",
1034                dentry, PFID(ll_inode2fid(dir)), dir);
1035
1036         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1037                 mode &= ~current_umask();
1038         mode = (mode & (0777 | S_ISVTX)) | S_IFDIR;
1039
1040         err = ll_new_node(dir, dentry, NULL, mode, 0, LUSTRE_OPC_MKDIR);
1041         if (!err)
1042                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKDIR, 1);
1043
1044         return err;
1045 }
1046
1047 static int ll_rmdir(struct inode *dir, struct dentry *dchild)
1048 {
1049         struct ptlrpc_request *request = NULL;
1050         struct md_op_data *op_data;
1051         int rc;
1052
1053         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p)\n",
1054                dchild, PFID(ll_inode2fid(dir)), dir);
1055
1056         op_data = ll_prep_md_op_data(NULL, dir, NULL,
1057                                      dchild->d_name.name,
1058                                      dchild->d_name.len,
1059                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1060         if (IS_ERR(op_data))
1061                 return PTR_ERR(op_data);
1062
1063         if (dchild->d_inode)
1064                 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1065
1066         op_data->op_fid2 = op_data->op_fid3;
1067         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1068         ll_finish_md_op_data(op_data);
1069         if (rc == 0) {
1070                 ll_update_times(request, dir);
1071                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR, 1);
1072         }
1073
1074         ptlrpc_req_finished(request);
1075         return rc;
1076 }
1077
1078 static int ll_symlink(struct inode *dir, struct dentry *dentry,
1079                       const char *oldname)
1080 {
1081         int err;
1082
1083         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir=" DFID "(%p),target=%.*s\n",
1084                dentry, PFID(ll_inode2fid(dir)), dir, 3000, oldname);
1085
1086         err = ll_new_node(dir, dentry, oldname, S_IFLNK | 0777,
1087                           0, LUSTRE_OPC_SYMLINK);
1088
1089         if (!err)
1090                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_SYMLINK, 1);
1091
1092         return err;
1093 }
1094
1095 static int ll_link(struct dentry *old_dentry, struct inode *dir,
1096                    struct dentry *new_dentry)
1097 {
1098         struct inode *src = d_inode(old_dentry);
1099         struct ll_sb_info *sbi = ll_i2sbi(dir);
1100         struct ptlrpc_request *request = NULL;
1101         struct md_op_data *op_data;
1102         int err;
1103
1104         CDEBUG(D_VFSTRACE, "VFS Op: inode=" DFID "(%p), dir=" DFID "(%p), target=%pd\n",
1105                PFID(ll_inode2fid(src)), src, PFID(ll_inode2fid(dir)), dir,
1106                new_dentry);
1107
1108         op_data = ll_prep_md_op_data(NULL, src, dir, new_dentry->d_name.name,
1109                                      new_dentry->d_name.len,
1110                                      0, LUSTRE_OPC_ANY, NULL);
1111         if (IS_ERR(op_data))
1112                 return PTR_ERR(op_data);
1113
1114         err = md_link(sbi->ll_md_exp, op_data, &request);
1115         ll_finish_md_op_data(op_data);
1116         if (err)
1117                 goto out;
1118
1119         ll_update_times(request, dir);
1120         ll_stats_ops_tally(sbi, LPROC_LL_LINK, 1);
1121 out:
1122         ptlrpc_req_finished(request);
1123         return err;
1124 }
1125
1126 static int ll_rename(struct inode *src, struct dentry *src_dchild,
1127                      struct inode *tgt, struct dentry *tgt_dchild,
1128                      unsigned int flags)
1129 {
1130         struct ptlrpc_request *request = NULL;
1131         struct ll_sb_info *sbi = ll_i2sbi(src);
1132         struct md_op_data *op_data;
1133         int err;
1134
1135         if (flags)
1136                 return -EINVAL;
1137
1138         CDEBUG(D_VFSTRACE,
1139                "VFS Op:oldname=%pd, src_dir=" DFID "(%p), newname=%pd, tgt_dir=" DFID "(%p)\n",
1140                src_dchild, PFID(ll_inode2fid(src)), src,
1141                tgt_dchild, PFID(ll_inode2fid(tgt)), tgt);
1142
1143         op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, 0,
1144                                      LUSTRE_OPC_ANY, NULL);
1145         if (IS_ERR(op_data))
1146                 return PTR_ERR(op_data);
1147
1148         if (src_dchild->d_inode)
1149                 op_data->op_fid3 = *ll_inode2fid(src_dchild->d_inode);
1150         if (tgt_dchild->d_inode)
1151                 op_data->op_fid4 = *ll_inode2fid(tgt_dchild->d_inode);
1152
1153         err = md_rename(sbi->ll_md_exp, op_data,
1154                         src_dchild->d_name.name,
1155                         src_dchild->d_name.len,
1156                         tgt_dchild->d_name.name,
1157                         tgt_dchild->d_name.len, &request);
1158         ll_finish_md_op_data(op_data);
1159         if (!err) {
1160                 ll_update_times(request, src);
1161                 ll_update_times(request, tgt);
1162                 ll_stats_ops_tally(sbi, LPROC_LL_RENAME, 1);
1163         }
1164
1165         ptlrpc_req_finished(request);
1166         if (!err)
1167                 d_move(src_dchild, tgt_dchild);
1168         return err;
1169 }
1170
1171 const struct inode_operations ll_dir_inode_operations = {
1172         .mknod        = ll_mknod,
1173         .atomic_open        = ll_atomic_open,
1174         .lookup      = ll_lookup_nd,
1175         .create      = ll_create_nd,
1176         /* We need all these non-raw things for NFSD, to not patch it. */
1177         .unlink      = ll_unlink,
1178         .mkdir        = ll_mkdir,
1179         .rmdir        = ll_rmdir,
1180         .symlink            = ll_symlink,
1181         .link          = ll_link,
1182         .rename         = ll_rename,
1183         .setattr            = ll_setattr,
1184         .getattr            = ll_getattr,
1185         .permission      = ll_inode_permission,
1186         .listxattr        = ll_listxattr,
1187         .get_acl            = ll_get_acl,
1188 };
1189
1190 const struct inode_operations ll_special_inode_operations = {
1191         .setattr        = ll_setattr,
1192         .getattr        = ll_getattr,
1193         .permission     = ll_inode_permission,
1194         .listxattr      = ll_listxattr,
1195         .get_acl            = ll_get_acl,
1196 };