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, 2017, 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 #include <linux/user_namespace.h>
  41 #ifdef HAVE_UIDGID_HEADER
  42 # include <linux/uidgid.h>
  43 #endif
  44
  45 #define DEBUG_SUBSYSTEM S_LLITE
  46
  47 #include <obd_support.h>
  48 #include <lustre_fid.h>
  49 #include <lustre_dlm.h>
  50 #include "llite_internal.h"
  51
  52 static int ll_create_it(struct inode *dir, struct dentry *dentry,
  53                         struct lookup_intent *it,
  54                         void *secctx, __u32 secctxlen);
  55
  56 /* called from iget5_locked->find_inode() under inode_lock spinlock */
  57 static int ll_test_inode(struct inode *inode, void *opaque)
  58 {
  59         struct ll_inode_info    *lli = ll_i2info(inode);
  60         struct lustre_md        *md = opaque;
  61
  62         if (unlikely(!(md->body->mbo_valid & OBD_MD_FLID))) {
  63                 CERROR("MDS body missing FID\n");
  64                 return 0;
  65         }
  66
  67         if (!lu_fid_eq(&lli->lli_fid, &md->body->mbo_fid1))
  68                 return 0;
  69
  70         return 1;
  71 }
  72
  73 static int ll_set_inode(struct inode *inode, void *opaque)
  74 {
  75         struct ll_inode_info *lli = ll_i2info(inode);
  76         struct mdt_body *body = ((struct lustre_md *)opaque)->body;
  77
  78         if (unlikely(!(body->mbo_valid & OBD_MD_FLID))) {
  79                 CERROR("MDS body missing FID\n");
  80                 return -EINVAL;
  81         }
  82
  83         lli->lli_fid = body->mbo_fid1;
  84         if (unlikely(!(body->mbo_valid & OBD_MD_FLTYPE))) {
  85                 CERROR("Can not initialize inode "DFID" without object type: "
  86                        "valid = %#llx\n",
  87                        PFID(&lli->lli_fid), body->mbo_valid);
  88                 return -EINVAL;
  89         }
  90
  91         inode->i_mode = (inode->i_mode & ~S_IFMT) | (body->mbo_mode & S_IFMT);
  92         if (unlikely(inode->i_mode == 0)) {
  93                 CERROR("Invalid inode "DFID" type\n", PFID(&lli->lli_fid));
  94                 return -EINVAL;
  95         }
  96
  97         ll_lli_init(lli);
  98
  99         return 0;
 100 }
 101
 102
 103 /**
 104  * Get an inode by inode number(@hash), which is already instantiated by
 105  * the intent lookup).
 106  */
 107 struct inode *ll_iget(struct super_block *sb, ino_t hash,
 108                       struct lustre_md *md)
 109 {
 110         struct inode    *inode;
 111         int             rc = 0;
 112
 113         ENTRY;
 114
 115         LASSERT(hash != 0);
 116         inode = iget5_locked(sb, hash, ll_test_inode, ll_set_inode, md);
 117         if (inode == NULL)
 118                 RETURN(ERR_PTR(-ENOMEM));
 119
 120         if (inode->i_state & I_NEW) {
 121                 rc = ll_read_inode2(inode, md);
 122                 if (rc == 0 && S_ISREG(inode->i_mode) &&
 123                     ll_i2info(inode)->lli_clob == NULL)
 124                         rc = cl_file_inode_init(inode, md);
 125
 126                 if (rc != 0) {
 127                         /* Let's clear directory lsm here, otherwise
 128                          * make_bad_inode() will reset the inode mode
 129                          * to regular, then ll_clear_inode will not
 130                          * be able to clear lsm_md */
 131                         if (S_ISDIR(inode->i_mode))
 132                                 ll_dir_clear_lsm_md(inode);
 133                         make_bad_inode(inode);
 134                         unlock_new_inode(inode);
 135                         iput(inode);
 136                         inode = ERR_PTR(rc);
 137                 } else {
 138                         inode_has_no_xattr(inode);
 139                         unlock_new_inode(inode);
 140                 }
 141         } else if (is_bad_inode(inode)) {
 142                 iput(inode);
 143                 inode = ERR_PTR(-ESTALE);
 144         } else if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
 145                 rc = ll_update_inode(inode, md);
 146                 CDEBUG(D_VFSTRACE, "got inode: "DFID"(%p): rc = %d\n",
 147                        PFID(&md->body->mbo_fid1), inode, rc);
 148                 if (rc != 0) {
 149                         if (S_ISDIR(inode->i_mode))
 150                                 ll_dir_clear_lsm_md(inode);
 151                         iput(inode);
 152                         inode = ERR_PTR(rc);
 153                 }
 154         }
 155
 156         RETURN(inode);
 157 }
 158
 159 static void ll_invalidate_negative_children(struct inode *dir)
 160 {
 161         struct dentry *dentry, *tmp_subdir;
 162         DECLARE_LL_D_HLIST_NODE_PTR(p);
 163
 164         ll_lock_dcache(dir);
 165         ll_d_hlist_for_each_entry(dentry, p, &dir->i_dentry) {
 166                 spin_lock(&dentry->d_lock);
 167                 if (!list_empty(&dentry->d_subdirs)) {
 168                         struct dentry *child;
 169
 170                         list_for_each_entry_safe(child, tmp_subdir,
 171                                                  &dentry->d_subdirs,
 172                                                  d_child) {
 173                                 if (child->d_inode == NULL)
 174                                         d_lustre_invalidate(child, 1);
 175                         }
 176                 }
 177                 spin_unlock(&dentry->d_lock);
 178         }
 179         ll_unlock_dcache(dir);
 180 }
 181
 182 int ll_test_inode_by_fid(struct inode *inode, void *opaque)
 183 {
 184         return lu_fid_eq(&ll_i2info(inode)->lli_fid, opaque);
 185 }
 186
 187 int ll_dom_lock_cancel(struct inode *inode, struct ldlm_lock *lock)
 188 {
 189         struct lu_env *env;
 190         struct ll_inode_info *lli = ll_i2info(inode);
 191         struct cl_layout clt = { .cl_layout_gen = 0, };
 192         int rc;
 193         __u16 refcheck;
 194
 195
 196         ENTRY;
 197
 198         if (!lli->lli_clob)
 199                 RETURN(0);
 200
 201         env = cl_env_get(&refcheck);
 202         if (IS_ERR(env))
 203                 RETURN(PTR_ERR(env));
 204
 205         rc = cl_object_layout_get(env, lli->lli_clob, &clt);
 206         if (rc) {
 207                 CDEBUG(D_INODE, "Cannot get layout for "DFID"\n",
 208                        PFID(ll_inode2fid(inode)));
 209                 rc = -ENODATA;
 210         } else if (clt.cl_size == 0 || clt.cl_dom_comp_size == 0) {
 211                 CDEBUG(D_INODE, "DOM lock without DOM layout for "DFID"\n",
 212                        PFID(ll_inode2fid(inode)));
 213         } else {
 214                 enum cl_fsync_mode mode;
 215                 loff_t end = clt.cl_dom_comp_size - 1;
 216
 217                 mode = ldlm_is_discard_data(lock) ?
 218                                         CL_FSYNC_DISCARD : CL_FSYNC_LOCAL;
 219                 rc = cl_sync_file_range(inode, 0, end, mode, 1);
 220                 truncate_inode_pages_range(inode->i_mapping, 0, end);
 221         }
 222         cl_env_put(env, &refcheck);
 223         RETURN(rc);
 224 }
 225
 226 int ll_md_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 227                        void *data, int flag)
 228 {
 229         struct lustre_handle lockh;
 230         int rc;
 231         ENTRY;
 232
 233         switch (flag) {
 234         case LDLM_CB_BLOCKING:
 235                 ldlm_lock2handle(lock, &lockh);
 236                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
 237                 if (rc < 0) {
 238                         CDEBUG(D_INODE, "ldlm_cli_cancel: rc = %d\n", rc);
 239                         RETURN(rc);
 240                 }
 241                 break;
 242         case LDLM_CB_CANCELING: {
 243                 struct inode *inode = ll_inode_from_resource_lock(lock);
 244                 __u64 bits = lock->l_policy_data.l_inodebits.bits;
 245
 246                 if (inode == NULL)
 247                         break;
 248
 249                 /* Invalidate all dentries associated with this inode */
 250                 LASSERT(ldlm_is_canceling(lock));
 251
 252                 if (!fid_res_name_eq(ll_inode2fid(inode),
 253                                      &lock->l_resource->lr_name)) {
 254                         LDLM_ERROR(lock, "data mismatch with object "DFID"(%p)",
 255                                    PFID(ll_inode2fid(inode)), inode);
 256                         LBUG();
 257                 }
 258
 259                 if (bits & MDS_INODELOCK_XATTR) {
 260                         if (S_ISDIR(inode->i_mode))
 261                                 ll_i2info(inode)->lli_def_stripe_offset = -1;
 262                         ll_xattr_cache_destroy(inode);
 263                         bits &= ~MDS_INODELOCK_XATTR;
 264                 }
 265
 266                 /* For OPEN locks we differentiate between lock modes
 267                  * LCK_CR, LCK_CW, LCK_PR - bug 22891 */
 268                 if (bits & MDS_INODELOCK_OPEN)
 269                         ll_have_md_lock(inode, &bits, lock->l_req_mode);
 270
 271                 if (bits & MDS_INODELOCK_OPEN) {
 272                         fmode_t fmode;
 273
 274                         switch (lock->l_req_mode) {
 275                         case LCK_CW:
 276                                 fmode = FMODE_WRITE;
 277                                 break;
 278                         case LCK_PR:
 279                                 fmode = FMODE_EXEC;
 280                                 break;
 281                         case LCK_CR:
 282                                 fmode = FMODE_READ;
 283                                 break;
 284                         default:
 285                                 LDLM_ERROR(lock, "bad lock mode for OPEN lock");
 286                                 LBUG();
 287                         }
 288
 289                         ll_md_real_close(inode, fmode);
 290
 291                         bits &= ~MDS_INODELOCK_OPEN;
 292                 }
 293
 294                 if (bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
 295                             MDS_INODELOCK_LAYOUT | MDS_INODELOCK_PERM |
 296                             MDS_INODELOCK_DOM))
 297                         ll_have_md_lock(inode, &bits, LCK_MINMODE);
 298
 299                 if (bits & MDS_INODELOCK_DOM) {
 300                         rc =  ll_dom_lock_cancel(inode, lock);
 301                         if (rc < 0)
 302                                 CDEBUG(D_INODE, "cannot flush DoM data "
 303                                        DFID": rc = %d\n",
 304                                        PFID(ll_inode2fid(inode)), rc);
 305                         lock_res_and_lock(lock);
 306                         ldlm_set_kms_ignore(lock);
 307                         unlock_res_and_lock(lock);
 308                         bits &= ~MDS_INODELOCK_DOM;
 309                 }
 310
 311                 if (bits & MDS_INODELOCK_LAYOUT) {
 312                         struct cl_object_conf conf = {
 313                                 .coc_opc = OBJECT_CONF_INVALIDATE,
 314                                 .coc_inode = inode,
 315                         };
 316
 317                         rc = ll_layout_conf(inode, &conf);
 318                         if (rc < 0)
 319                                 CDEBUG(D_INODE, "cannot invalidate layout of "
 320                                        DFID": rc = %d\n",
 321                                        PFID(ll_inode2fid(inode)), rc);
 322                 }
 323
 324                 if (bits & MDS_INODELOCK_UPDATE) {
 325                         struct ll_inode_info *lli = ll_i2info(inode);
 326                         lli->lli_update_atime = 1;
 327                 }
 328
 329                 if ((bits & MDS_INODELOCK_UPDATE) && S_ISDIR(inode->i_mode)) {
 330                         struct ll_inode_info *lli = ll_i2info(inode);
 331
 332                         CDEBUG(D_INODE, "invalidating inode "DFID" lli = %p, "
 333                                "pfid  = "DFID"\n", PFID(ll_inode2fid(inode)),
 334                                lli, PFID(&lli->lli_pfid));
 335                         truncate_inode_pages(inode->i_mapping, 0);
 336
 337                         if (unlikely(!fid_is_zero(&lli->lli_pfid))) {
 338                                 struct inode *master_inode = NULL;
 339                                 unsigned long hash;
 340
 341                                 /* This is slave inode, since all of the child
 342                                  * dentry is connected on the master inode, so
 343                                  * we have to invalidate the negative children
 344                                  * on master inode */
 345                                 CDEBUG(D_INODE, "Invalidate s"DFID" m"DFID"\n",
 346                                        PFID(ll_inode2fid(inode)),
 347                                        PFID(&lli->lli_pfid));
 348
 349                                 hash = cl_fid_build_ino(&lli->lli_pfid,
 350                                         ll_need_32bit_api(ll_i2sbi(inode)));
 351
 352                                 /* Do not lookup the inode with ilookup5,
 353                                  * otherwise it will cause dead lock,
 354                                  *
 355                                  * 1. Client1 send chmod req to the MDT0, then
 356                                  * on MDT0, it enqueues master and all of its
 357                                  * slaves lock, (mdt_attr_set() ->
 358                                  * mdt_lock_slaves()), after gets master and
 359                                  * stripe0 lock, it will send the enqueue req
 360                                  * (for stripe1) to MDT1, then MDT1 finds the
 361                                  * lock has been granted to client2. Then MDT1
 362                                  * sends blocking ast to client2.
 363                                  *
 364                                  * 2. At the same time, client2 tries to unlink
 365                                  * the striped dir (rm -rf striped_dir), and
 366                                  * during lookup, it will hold the master inode
 367                                  * of the striped directory, whose inode state
 368                                  * is NEW, then tries to revalidate all of its
 369                                  * slaves, (ll_prep_inode()->ll_iget()->
 370                                  * ll_read_inode2()-> ll_update_inode().). And
 371                                  * it will be blocked on the server side because
 372                                  * of 1.
 373                                  *
 374                                  * 3. Then the client get the blocking_ast req,
 375                                  * cancel the lock, but being blocked if using
 376                                  * ->ilookup5()), because master inode state is
 377                                  *  NEW. */
 378                                 master_inode = ilookup5_nowait(inode->i_sb,
 379                                                     hash, ll_test_inode_by_fid,
 380                                                         (void *)&lli->lli_pfid);
 381                                 if (master_inode) {
 382                                         ll_invalidate_negative_children(
 383                                                                 master_inode);
 384                                         iput(master_inode);
 385                                 }
 386                         } else {
 387                                 ll_invalidate_negative_children(inode);
 388                         }
 389                 }
 390
 391                 if ((bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM)) &&
 392                     inode->i_sb->s_root != NULL &&
 393                     inode != inode->i_sb->s_root->d_inode)
 394                         ll_invalidate_aliases(inode);
 395
 396                 iput(inode);
 397                 break;
 398         }
 399         default:
 400                 LBUG();
 401         }
 402
 403         RETURN(0);
 404 }
 405
 406 __u32 ll_i2suppgid(struct inode *i)
 407 {
 408         if (in_group_p(i->i_gid))
 409                 return (__u32)from_kgid(&init_user_ns, i->i_gid);
 410         else
 411                 return (__u32) __kgid_val(INVALID_GID);
 412 }
 413
 414 /* Pack the required supplementary groups into the supplied groups array.
 415  * If we don't need to use the groups from the target inode(s) then we
 416  * instead pack one or more groups from the user's supplementary group
 417  * array in case it might be useful.  Not needed if doing an MDS-side upcall. */
 418 void ll_i2gids(__u32 *suppgids, struct inode *i1, struct inode *i2)
 419 {
 420         LASSERT(i1 != NULL);
 421         LASSERT(suppgids != NULL);
 422
 423         suppgids[0] = ll_i2suppgid(i1);
 424
 425         if (i2)
 426                 suppgids[1] = ll_i2suppgid(i2);
 427         else
 428                 suppgids[1] = -1;
 429 }
 430
 431 /*
 432  * try to reuse three types of dentry:
 433  * 1. unhashed alias, this one is unhashed by d_invalidate (but it may be valid
 434  *    by concurrent .revalidate).
 435  * 2. INVALID alias (common case for no valid ldlm lock held, but this flag may
 436  *    be cleared by others calling d_lustre_revalidate).
 437  * 3. DISCONNECTED alias.
 438  */
 439 static struct dentry *ll_find_alias(struct inode *inode, struct dentry *dentry)
 440 {
 441         struct dentry *alias, *discon_alias, *invalid_alias;
 442         DECLARE_LL_D_HLIST_NODE_PTR(p);
 443
 444         if (ll_d_hlist_empty(&inode->i_dentry))
 445                 return NULL;
 446
 447         discon_alias = invalid_alias = NULL;
 448
 449         ll_lock_dcache(inode);
 450         ll_d_hlist_for_each_entry(alias, p, &inode->i_dentry) {
 451                 LASSERT(alias != dentry);
 452
 453                 spin_lock(&alias->d_lock);
 454                 if ((alias->d_flags & DCACHE_DISCONNECTED) &&
 455                     S_ISDIR(inode->i_mode))
 456                         /* LASSERT(last_discon == NULL); LU-405, bz 20055 */
 457                         discon_alias = alias;
 458                 else if (alias->d_parent == dentry->d_parent             &&
 459                          alias->d_name.hash == dentry->d_name.hash       &&
 460                          alias->d_name.len == dentry->d_name.len         &&
 461                          memcmp(alias->d_name.name, dentry->d_name.name,
 462                                 dentry->d_name.len) == 0)
 463                         invalid_alias = alias;
 464                 spin_unlock(&alias->d_lock);
 465
 466                 if (invalid_alias)
 467                         break;
 468         }
 469         alias = invalid_alias ?: discon_alias ?: NULL;
 470         if (alias) {
 471                 spin_lock(&alias->d_lock);
 472                 dget_dlock(alias);
 473                 spin_unlock(&alias->d_lock);
 474         }
 475         ll_unlock_dcache(inode);
 476
 477         return alias;
 478 }
 479
 480 /*
 481  * Similar to d_splice_alias(), but lustre treats invalid alias
 482  * similar to DCACHE_DISCONNECTED, and tries to use it anyway.
 483  */
 484 struct dentry *ll_splice_alias(struct inode *inode, struct dentry *de)
 485 {
 486         struct dentry *new;
 487         int rc;
 488
 489         if (inode) {
 490                 new = ll_find_alias(inode, de);
 491                 if (new) {
 492                         rc = ll_d_init(new);
 493                         if (rc < 0) {
 494                                 dput(new);
 495                                 return ERR_PTR(rc);
 496                         }
 497                         d_move(new, de);
 498                         iput(inode);
 499                         CDEBUG(D_DENTRY,
 500                                "Reuse dentry %p inode %p refc %d flags %#x\n",
 501                               new, new->d_inode, ll_d_count(new), new->d_flags);
 502                         return new;
 503                 }
 504         }
 505         rc = ll_d_init(de);
 506         if (rc < 0)
 507                 return ERR_PTR(rc);
 508         d_add(de, inode);
 509         CDEBUG(D_DENTRY, "Add dentry %p inode %p refc %d flags %#x\n",
 510                de, de->d_inode, ll_d_count(de), de->d_flags);
 511         return de;
 512 }
 513
 514 static int ll_lookup_it_finish(struct ptlrpc_request *request,
 515                                struct lookup_intent *it,
 516                                struct inode *parent, struct dentry **de)
 517 {
 518         struct inode             *inode = NULL;
 519         __u64                     bits = 0;
 520         int                       rc;
 521         struct dentry *alias;
 522         ENTRY;
 523
 524         /* NB 1 request reference will be taken away by ll_intent_lock()
 525          * when I return */
 526         CDEBUG(D_DENTRY, "it %p it_disposition %x\n", it,
 527                it->it_disposition);
 528         if (!it_disposition(it, DISP_LOOKUP_NEG)) {
 529                 rc = ll_prep_inode(&inode, request, (*de)->d_sb, it);
 530                 if (rc)
 531                         RETURN(rc);
 532
 533                 ll_set_lock_data(ll_i2sbi(parent)->ll_md_exp, inode, it, &bits);
 534
 535                 /* We used to query real size from OSTs here, but actually
 536                    this is not needed. For stat() calls size would be updated
 537                    from subsequent do_revalidate()->ll_inode_revalidate_it() in
 538                    2.4 and
 539                    vfs_getattr_it->ll_getattr()->ll_inode_revalidate_it() in 2.6
 540                    Everybody else who needs correct file size would call
 541                    ll_glimpse_size or some equivalent themselves anyway.
 542                    Also see bug 7198. */
 543         }
 544
 545         /* Only hash *de if it is unhashed (new dentry).
 546          * Atoimc_open may passin hashed dentries for open.
 547          */
 548         alias = ll_splice_alias(inode, *de);
 549         if (IS_ERR(alias))
 550                 GOTO(out, rc = PTR_ERR(alias));
 551
 552         *de = alias;
 553
 554         if (!it_disposition(it, DISP_LOOKUP_NEG)) {
 555                 /* we have lookup look - unhide dentry */
 556                 if (bits & MDS_INODELOCK_LOOKUP)
 557                         d_lustre_revalidate(*de);
 558         } else if (!it_disposition(it, DISP_OPEN_CREATE)) {
 559                 /* If file created on server, don't depend on parent UPDATE
 560                  * lock to unhide it. It is left hidden and next lookup can
 561                  * find it in ll_splice_alias.
 562                  */
 563                 /* Check that parent has UPDATE lock. */
 564                 struct lookup_intent parent_it = {
 565                                         .it_op = IT_GETATTR,
 566                                         .it_lock_handle = 0 };
 567                 struct lu_fid   fid = ll_i2info(parent)->lli_fid;
 568
 569                 /* If it is striped directory, get the real stripe parent */
 570                 if (unlikely(ll_i2info(parent)->lli_lsm_md != NULL)) {
 571                         rc = md_get_fid_from_lsm(ll_i2mdexp(parent),
 572                                                  ll_i2info(parent)->lli_lsm_md,
 573                                                  (*de)->d_name.name,
 574                                                  (*de)->d_name.len, &fid);
 575                         if (rc != 0)
 576                                 GOTO(out, rc);
 577                 }
 578
 579                 if (md_revalidate_lock(ll_i2mdexp(parent), &parent_it, &fid,
 580                                        NULL)) {
 581                         d_lustre_revalidate(*de);
 582                         ll_intent_release(&parent_it);
 583                 }
 584         }
 585
 586         GOTO(out, rc = 0);
 587
 588 out:
 589         if (rc != 0 && it->it_op & IT_OPEN)
 590                 ll_open_cleanup((*de)->d_sb, request);
 591
 592         return rc;
 593 }
 594
 595 static struct dentry *ll_lookup_it(struct inode *parent, struct dentry *dentry,
 596                                    struct lookup_intent *it,
 597                                    void **secctx, __u32 *secctxlen)
 598 {
 599         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 600         struct dentry *save = dentry, *retval;
 601         struct ptlrpc_request *req = NULL;
 602         struct md_op_data *op_data = NULL;
 603         __u32 opc;
 604         int rc;
 605         ENTRY;
 606
 607         if (dentry->d_name.len > ll_i2sbi(parent)->ll_namelen)
 608                 RETURN(ERR_PTR(-ENAMETOOLONG));
 609
 610         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), intent=%s\n",
 611                dentry->d_name.len, dentry->d_name.name,
 612                PFID(ll_inode2fid(parent)), parent, LL_IT2STR(it));
 613
 614         if (d_mountpoint(dentry))
 615                 CERROR("Tell Peter, lookup on mtpt, it %s\n", LL_IT2STR(it));
 616
 617         if (it == NULL || it->it_op == IT_GETXATTR)
 618                 it = &lookup_it;
 619
 620         if (it->it_op == IT_GETATTR && dentry_may_statahead(parent, dentry)) {
 621                 rc = ll_statahead(parent, &dentry, 0);
 622                 if (rc == 1)
 623                         RETURN(dentry == save ? NULL : dentry);
 624         }
 625
 626         if (it->it_op & IT_OPEN && it->it_flags & FMODE_WRITE &&
 627             dentry->d_sb->s_flags & MS_RDONLY)
 628                 RETURN(ERR_PTR(-EROFS));
 629
 630         if (it->it_op & IT_CREAT)
 631                 opc = LUSTRE_OPC_CREATE;
 632         else
 633                 opc = LUSTRE_OPC_ANY;
 634
 635         op_data = ll_prep_md_op_data(NULL, parent, NULL, dentry->d_name.name,
 636                                      dentry->d_name.len, 0, opc, NULL);
 637         if (IS_ERR(op_data))
 638                 GOTO(out, retval = ERR_CAST(op_data));
 639
 640         /* enforce umask if acl disabled or MDS doesn't support umask */
 641         if (!IS_POSIXACL(parent) || !exp_connect_umask(ll_i2mdexp(parent)))
 642                 it->it_create_mode &= ~current_umask();
 643
 644         if (it->it_op & IT_CREAT &&
 645             ll_i2sbi(parent)->ll_flags & LL_SBI_FILE_SECCTX) {
 646                 rc = ll_dentry_init_security(dentry, it->it_create_mode,
 647                                              &dentry->d_name,
 648                                              &op_data->op_file_secctx_name,
 649                                              &op_data->op_file_secctx,
 650                                              &op_data->op_file_secctx_size);
 651                 if (rc < 0)
 652                         GOTO(out, retval = ERR_PTR(rc));
 653                 if (secctx != NULL)
 654                         *secctx = op_data->op_file_secctx;
 655                 if (secctxlen != NULL)
 656                         *secctxlen = op_data->op_file_secctx_size;
 657         }
 658
 659         rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
 660                             &ll_md_blocking_ast, 0);
 661         /* If the MDS allows the client to chgrp (CFS_SETGRP_PERM), but the
 662          * client does not know which suppgid should be sent to the MDS, or
 663          * some other(s) changed the target file's GID after this RPC sent
 664          * to the MDS with the suppgid as the original GID, then we should
 665          * try again with right suppgid. */
 666         if (rc == -EACCES && it->it_op & IT_OPEN &&
 667             it_disposition(it, DISP_OPEN_DENY)) {
 668                 struct mdt_body *body;
 669
 670                 LASSERT(req != NULL);
 671
 672                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 673                 if (op_data->op_suppgids[0] == body->mbo_gid ||
 674                     op_data->op_suppgids[1] == body->mbo_gid ||
 675                     !in_group_p(make_kgid(&init_user_ns, body->mbo_gid)))
 676                         GOTO(out, retval = ERR_PTR(-EACCES));
 677
 678                 fid_zero(&op_data->op_fid2);
 679                 op_data->op_suppgids[1] = body->mbo_gid;
 680                 ptlrpc_req_finished(req);
 681                 req = NULL;
 682                 ll_intent_release(it);
 683                 rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
 684                                     &ll_md_blocking_ast, 0);
 685         }
 686
 687         if (rc < 0)
 688                 GOTO(out, retval = ERR_PTR(rc));
 689
 690         rc = ll_lookup_it_finish(req, it, parent, &dentry);
 691         if (rc != 0) {
 692                 ll_intent_release(it);
 693                 GOTO(out, retval = ERR_PTR(rc));
 694         }
 695
 696         if ((it->it_op & IT_OPEN) && dentry->d_inode &&
 697             !S_ISREG(dentry->d_inode->i_mode) &&
 698             !S_ISDIR(dentry->d_inode->i_mode)) {
 699                 ll_release_openhandle(dentry, it);
 700         }
 701         ll_lookup_finish_locks(it, dentry);
 702
 703         GOTO(out, retval = (dentry == save) ? NULL : dentry);
 704
 705 out:
 706         if (op_data != NULL && !IS_ERR(op_data)) {
 707                 if (secctx != NULL && secctxlen != NULL) {
 708                         /* caller needs sec ctx info, so reset it in op_data to
 709                          * prevent it from being freed */
 710                         op_data->op_file_secctx = NULL;
 711                         op_data->op_file_secctx_size = 0;
 712                 }
 713                 ll_finish_md_op_data(op_data);
 714         }
 715
 716         ptlrpc_req_finished(req);
 717         return retval;
 718 }
 719
 720 #ifdef HAVE_IOP_ATOMIC_OPEN
 721 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
 722                                    unsigned int flags)
 723 {
 724         struct lookup_intent *itp, it = { .it_op = IT_GETATTR };
 725         struct dentry *de;
 726
 727         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), flags=%u\n",
 728                dentry->d_name.len, dentry->d_name.name,
 729                PFID(ll_inode2fid(parent)), parent, flags);
 730
 731         /*
 732          * Optimize away (CREATE && !OPEN). Let .create handle the race.
 733          * but only if we have write permissions there, otherwise we need
 734          * to proceed with lookup. LU-4185
 735          */
 736         if ((flags & LOOKUP_CREATE) && !(flags & LOOKUP_OPEN) &&
 737             (inode_permission(parent, MAY_WRITE | MAY_EXEC) == 0))
 738                 return NULL;
 739
 740         if (flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE))
 741                 itp = NULL;
 742         else
 743                 itp = &it;
 744         de = ll_lookup_it(parent, dentry, itp, NULL, NULL);
 745
 746         if (itp != NULL)
 747                 ll_intent_release(itp);
 748
 749         return de;
 750 }
 751
 752 /*
 753  * For cached negative dentry and new dentry, handle lookup/create/open
 754  * together.
 755  */
 756 static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
 757                           struct file *file, unsigned open_flags,
 758                           umode_t mode, int *opened)
 759 {
 760         struct lookup_intent *it;
 761         struct dentry *de;
 762         long long lookup_flags = LOOKUP_OPEN;
 763         void *secctx = NULL;
 764         __u32 secctxlen = 0;
 765         int rc = 0;
 766         ENTRY;
 767
 768         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), file %p,"
 769                            "open_flags %x, mode %x opened %d\n",
 770                dentry->d_name.len, dentry->d_name.name,
 771                PFID(ll_inode2fid(dir)), dir, file, open_flags, mode, *opened);
 772
 773         /* Only negative dentries enter here */
 774         LASSERT(dentry->d_inode == NULL);
 775
 776         if (!d_unhashed(dentry)) {
 777                 /* A valid negative dentry that just passed revalidation,
 778                  * there's little point to try and open it server-side,
 779                  * even though there's a minuscule chance it might succeed.
 780                  * Either way it's a valid race to just return -ENOENT here.
 781                  */
 782                 if (!(open_flags & O_CREAT))
 783                         return -ENOENT;
 784
 785                 /* Otherwise we just unhash it to be rehashed afresh via
 786                  * lookup if necessary
 787                  */
 788                 d_drop(dentry);
 789         }
 790
 791         OBD_ALLOC(it, sizeof(*it));
 792         if (!it)
 793                 RETURN(-ENOMEM);
 794
 795         it->it_op = IT_OPEN;
 796         if (open_flags & O_CREAT) {
 797                 it->it_op |= IT_CREAT;
 798                 lookup_flags |= LOOKUP_CREATE;
 799         }
 800         it->it_create_mode = (mode & S_IALLUGO) | S_IFREG;
 801         it->it_flags = (open_flags & ~O_ACCMODE) | OPEN_FMODE(open_flags);
 802         it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
 803
 804         /* Dentry added to dcache tree in ll_lookup_it */
 805         de = ll_lookup_it(dir, dentry, it, &secctx, &secctxlen);
 806         if (IS_ERR(de))
 807                 rc = PTR_ERR(de);
 808         else if (de != NULL)
 809                 dentry = de;
 810
 811         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
 812
 813         if (!rc) {
 814                 if (it_disposition(it, DISP_OPEN_CREATE)) {
 815                         /* Dentry instantiated in ll_create_it. */
 816                         rc = ll_create_it(dir, dentry, it, secctx, secctxlen);
 817                         security_release_secctx(secctx, secctxlen);
 818                         if (rc) {
 819                                 /* We dget in ll_splice_alias. */
 820                                 if (de != NULL)
 821                                         dput(de);
 822                                 goto out_release;
 823                         }
 824
 825                         *opened |= FILE_CREATED;
 826                 }
 827                 if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN)) {
 828                         /* Open dentry. */
 829                         if (S_ISFIFO(dentry->d_inode->i_mode)) {
 830                                 /* We cannot call open here as it might
 831                                  * deadlock. This case is unreachable in
 832                                  * practice because of OBD_CONNECT_NODEVOH. */
 833                                 rc = finish_no_open(file, de);
 834                         } else {
 835                                 file->private_data = it;
 836                                 rc = finish_open(file, dentry, NULL, opened);
 837                                 /* We dget in ll_splice_alias. finish_open takes
 838                                  * care of dget for fd open.
 839                                  */
 840                                 if (de != NULL)
 841                                         dput(de);
 842                         }
 843                 } else {
 844                         rc = finish_no_open(file, de);
 845                 }
 846         }
 847
 848 out_release:
 849         ll_intent_release(it);
 850         OBD_FREE(it, sizeof(*it));
 851
 852         RETURN(rc);
 853 }
 854
 855 #else /* !HAVE_IOP_ATOMIC_OPEN */
 856 static struct lookup_intent *
 857 ll_convert_intent(struct open_intent *oit, int lookup_flags, bool is_readonly)
 858 {
 859         struct lookup_intent *it;
 860
 861         OBD_ALLOC_PTR(it);
 862         if (!it)
 863                 return ERR_PTR(-ENOMEM);
 864
 865         if (lookup_flags & LOOKUP_OPEN) {
 866                 it->it_op = IT_OPEN;
 867                 /* Avoid file creation for ro bind mount point(is_readonly) */
 868                 if ((lookup_flags & LOOKUP_CREATE) && !is_readonly)
 869                         it->it_op |= IT_CREAT;
 870                 it->it_create_mode = (oit->create_mode & S_IALLUGO) | S_IFREG;
 871                 it->it_flags = ll_namei_to_lookup_intent_flag(oit->flags &
 872                                                 ~(is_readonly ? O_CREAT : 0));
 873                 it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
 874         } else {
 875                 it->it_op = IT_GETATTR;
 876         }
 877
 878         return it;
 879 }
 880
 881 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
 882                                    struct nameidata *nd)
 883 {
 884         struct dentry *de;
 885         ENTRY;
 886
 887         if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
 888                 struct lookup_intent *it;
 889
 890                 if (ll_d2d(dentry) && ll_d2d(dentry)->lld_it) {
 891                         it = ll_d2d(dentry)->lld_it;
 892                         ll_d2d(dentry)->lld_it = NULL;
 893                 } else {
 894                         /*
 895                          * Optimize away (CREATE && !OPEN). Let .create handle
 896                          * the race. But only if we have write permissions
 897                          * there, otherwise we need to proceed with lookup.
 898                          * LU-4185
 899                          */
 900                         if ((nd->flags & LOOKUP_CREATE) &&
 901                             !(nd->flags & LOOKUP_OPEN) &&
 902                             (inode_permission(parent,
 903                                               MAY_WRITE | MAY_EXEC) == 0))
 904                                 RETURN(NULL);
 905
 906                         it = ll_convert_intent(&nd->intent.open, nd->flags,
 907                                 (nd->path.mnt->mnt_flags & MNT_READONLY) ||
 908                                 (nd->path.mnt->mnt_sb->s_flags & MS_RDONLY));
 909                         if (IS_ERR(it))
 910                                 RETURN((struct dentry *)it);
 911                 }
 912
 913                 de = ll_lookup_it(parent, dentry, it, NULL, NULL);
 914                 if (de)
 915                         dentry = de;
 916                 if ((nd->flags & LOOKUP_OPEN) && !IS_ERR(dentry)) { /* Open */
 917                         if (dentry->d_inode &&
 918                             it_disposition(it, DISP_OPEN_OPEN)) { /* nocreate */
 919                                 if (S_ISFIFO(dentry->d_inode->i_mode)) {
 920                                         /* We cannot call open here as it might
 921                                          * deadlock. This case is unreachable in
 922                                          * practice because of
 923                                          * OBD_CONNECT_NODEVOH. */
 924                                 } else {
 925                                         struct file *filp;
 926
 927                                         nd->intent.open.file->private_data = it;
 928                                         filp = lookup_instantiate_filp(nd,
 929                                                                        dentry,
 930                                                                        NULL);
 931                                         if (IS_ERR(filp)) {
 932                                                 if (de)
 933                                                         dput(de);
 934                                                 de = (struct dentry *)filp;
 935                                         }
 936                                 }
 937                         } else if (it_disposition(it, DISP_OPEN_CREATE)) {
 938                                 /* XXX This can only reliably work on assumption
 939                                  * that there are NO hashed negative dentries.*/
 940                                 ll_d2d(dentry)->lld_it = it;
 941                                 it = NULL; /* Will be freed in ll_create_nd */
 942                                 /* We absolutely depend on ll_create_nd to be
 943                                  * called to not leak this intent and possible
 944                                  * data attached to it */
 945                         }
 946                 }
 947
 948                 if (it) {
 949                         ll_intent_release(it);
 950                         OBD_FREE(it, sizeof(*it));
 951                 }
 952         } else {
 953                 de = ll_lookup_it(parent, dentry, NULL, NULL, NULL);
 954         }
 955
 956         RETURN(de);
 957 }
 958 #endif /* HAVE_IOP_ATOMIC_OPEN */
 959
 960 /* We depend on "mode" being set with the proper file type/umask by now */
 961 static struct inode *ll_create_node(struct inode *dir, struct lookup_intent *it)
 962 {
 963         struct inode *inode = NULL;
 964         struct ptlrpc_request *request = NULL;
 965         struct ll_sb_info *sbi = ll_i2sbi(dir);
 966         int rc;
 967         ENTRY;
 968
 969         LASSERT(it && it->it_disposition);
 970
 971         LASSERT(it_disposition(it, DISP_ENQ_CREATE_REF));
 972         request = it->it_request;
 973         it_clear_disposition(it, DISP_ENQ_CREATE_REF);
 974         rc = ll_prep_inode(&inode, request, dir->i_sb, it);
 975         if (rc)
 976                 GOTO(out, inode = ERR_PTR(rc));
 977
 978         /* Pause to allow for a race with concurrent access by fid */
 979         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_NODE_PAUSE, cfs_fail_val);
 980
 981         /* We asked for a lock on the directory, but were granted a
 982          * lock on the inode.  Since we finally have an inode pointer,
 983          * stuff it in the lock. */
 984         CDEBUG(D_DLMTRACE, "setting l_ast_data to inode "DFID"(%p)\n",
 985                PFID(ll_inode2fid(inode)), inode);
 986         ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
 987         EXIT;
 988  out:
 989         ptlrpc_req_finished(request);
 990         return inode;
 991 }
 992
 993 /*
 994  * By the time this is called, we already have created the directory cache
 995  * entry for the new file, but it is so far negative - it has no inode.
 996  *
 997  * We defer creating the OBD object(s) until open, to keep the intent and
 998  * non-intent code paths similar, and also because we do not have the MDS
 999  * inode number before calling ll_create_node() (which is needed for LOV),
1000  * so we would need to do yet another RPC to the MDS to store the LOV EA
1001  * data on the MDS.  If needed, we would pass the PACKED lmm as data and
1002  * lmm_size in datalen (the MDS still has code which will handle that).
1003  *
1004  * If the create succeeds, we fill in the inode information
1005  * with d_instantiate().
1006  */
1007 static int ll_create_it(struct inode *dir, struct dentry *dentry,
1008                         struct lookup_intent *it,
1009                         void *secctx, __u32 secctxlen)
1010 {
1011         struct inode *inode;
1012         int rc = 0;
1013         ENTRY;
1014
1015         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), intent=%s\n",
1016                dentry->d_name.len, dentry->d_name.name,
1017                PFID(ll_inode2fid(dir)), dir, LL_IT2STR(it));
1018
1019         rc = it_open_error(DISP_OPEN_CREATE, it);
1020         if (rc)
1021                 RETURN(rc);
1022
1023         inode = ll_create_node(dir, it);
1024         if (IS_ERR(inode))
1025                 RETURN(PTR_ERR(inode));
1026
1027         if ((ll_i2sbi(inode)->ll_flags & LL_SBI_FILE_SECCTX) &&
1028             secctx != NULL) {
1029                 inode_lock(inode);
1030                 /* must be done before d_instantiate, because it calls
1031                  * security_d_instantiate, which means a getxattr if security
1032                  * context is not set yet */
1033                 rc = security_inode_notifysecctx(inode, secctx, secctxlen);
1034                 inode_unlock(inode);
1035                 if (rc)
1036                         RETURN(rc);
1037         }
1038
1039         d_instantiate(dentry, inode);
1040
1041         if (!(ll_i2sbi(inode)->ll_flags & LL_SBI_FILE_SECCTX)) {
1042                 rc = ll_inode_init_security(dentry, inode, dir);
1043                 if (rc)
1044                         RETURN(rc);
1045         }
1046
1047         RETURN(0);
1048 }
1049
1050 void ll_update_times(struct ptlrpc_request *request, struct inode *inode)
1051 {
1052         struct mdt_body *body = req_capsule_server_get(&request->rq_pill,
1053                                                        &RMF_MDT_BODY);
1054
1055         LASSERT(body);
1056         if (body->mbo_valid & OBD_MD_FLMTIME &&
1057             body->mbo_mtime > LTIME_S(inode->i_mtime)) {
1058                 CDEBUG(D_INODE, "setting fid "DFID" mtime from %lu to %llu"
1059                        "\n", PFID(ll_inode2fid(inode)),
1060                        LTIME_S(inode->i_mtime), body->mbo_mtime);
1061                 LTIME_S(inode->i_mtime) = body->mbo_mtime;
1062         }
1063
1064         if (body->mbo_valid & OBD_MD_FLCTIME &&
1065             body->mbo_ctime > LTIME_S(inode->i_ctime))
1066                 LTIME_S(inode->i_ctime) = body->mbo_ctime;
1067 }
1068
1069 static int ll_new_node(struct inode *dir, struct dentry *dchild,
1070                        const char *tgt, umode_t mode, int rdev, __u32 opc)
1071 {
1072         struct qstr *name = &dchild->d_name;
1073         struct ptlrpc_request *request = NULL;
1074         struct md_op_data *op_data;
1075         struct inode *inode = NULL;
1076         struct ll_sb_info *sbi = ll_i2sbi(dir);
1077         int tgt_len = 0;
1078         int err;
1079
1080         ENTRY;
1081         if (unlikely(tgt != NULL))
1082                 tgt_len = strlen(tgt) + 1;
1083
1084 again:
1085         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name,
1086                                      name->len, 0, opc, NULL);
1087         if (IS_ERR(op_data))
1088                 GOTO(err_exit, err = PTR_ERR(op_data));
1089
1090         if (sbi->ll_flags & LL_SBI_FILE_SECCTX) {
1091                 err = ll_dentry_init_security(dchild, mode, &dchild->d_name,
1092                                               &op_data->op_file_secctx_name,
1093                                               &op_data->op_file_secctx,
1094                                               &op_data->op_file_secctx_size);
1095                 if (err < 0)
1096                         GOTO(err_exit, err);
1097         }
1098
1099         err = md_create(sbi->ll_md_exp, op_data, tgt, tgt_len, mode,
1100                         from_kuid(&init_user_ns, current_fsuid()),
1101                         from_kgid(&init_user_ns, current_fsgid()),
1102                         cfs_curproc_cap_pack(), rdev, &request);
1103         if (err < 0 && err != -EREMOTE)
1104                 GOTO(err_exit, err);
1105
1106         /* If the client doesn't know where to create a subdirectory (or
1107          * in case of a race that sends the RPC to the wrong MDS), the
1108          * MDS will return -EREMOTE and the client will fetch the layout
1109          * of the directory, then create the directory on the right MDT. */
1110         if (unlikely(err == -EREMOTE)) {
1111                 struct ll_inode_info    *lli = ll_i2info(dir);
1112                 struct lmv_user_md      *lum;
1113                 int                     lumsize;
1114                 int                     err2;
1115
1116                 ptlrpc_req_finished(request);
1117                 request = NULL;
1118
1119                 err2 = ll_dir_getstripe(dir, (void **)&lum, &lumsize, &request,
1120                                         OBD_MD_DEFAULT_MEA);
1121                 if (err2 == 0) {
1122                         /* Update stripe_offset and retry */
1123                         lli->lli_def_stripe_offset = lum->lum_stripe_offset;
1124                 } else if (err2 == -ENODATA &&
1125                            lli->lli_def_stripe_offset != -1) {
1126                         /* If there are no default stripe EA on the MDT, but the
1127                          * client has default stripe, then it probably means
1128                          * default stripe EA has just been deleted. */
1129                         lli->lli_def_stripe_offset = -1;
1130                 } else {
1131                         GOTO(err_exit, err);
1132                 }
1133
1134                 ptlrpc_req_finished(request);
1135                 request = NULL;
1136                 ll_finish_md_op_data(op_data);
1137                 goto again;
1138         }
1139
1140         ll_update_times(request, dir);
1141
1142         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_NEWNODE_PAUSE, cfs_fail_val);
1143
1144         err = ll_prep_inode(&inode, request, dchild->d_sb, NULL);
1145         if (err)
1146                 GOTO(err_exit, err);
1147
1148         if (sbi->ll_flags & LL_SBI_FILE_SECCTX) {
1149                 inode_lock(inode);
1150                 /* must be done before d_instantiate, because it calls
1151                  * security_d_instantiate, which means a getxattr if security
1152                  * context is not set yet */
1153                 err = security_inode_notifysecctx(inode,
1154                                                   op_data->op_file_secctx,
1155                                                   op_data->op_file_secctx_size);
1156                 inode_unlock(inode);
1157                 if (err)
1158                         GOTO(err_exit, err);
1159         }
1160
1161         d_instantiate(dchild, inode);
1162
1163         if (!(sbi->ll_flags & LL_SBI_FILE_SECCTX)) {
1164                 err = ll_inode_init_security(dchild, inode, dir);
1165                 if (err)
1166                         GOTO(err_exit, err);
1167         }
1168
1169         EXIT;
1170 err_exit:
1171         if (request != NULL)
1172                 ptlrpc_req_finished(request);
1173
1174         if (!IS_ERR_OR_NULL(op_data))
1175                 ll_finish_md_op_data(op_data);
1176
1177         return err;
1178 }
1179
1180 static int ll_mknod(struct inode *dir, struct dentry *dchild, ll_umode_t mode,
1181                     dev_t rdev)
1182 {
1183         struct qstr *name = &dchild->d_name;
1184         int err;
1185         ENTRY;
1186
1187         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p) mode %o dev %x\n",
1188                name->len, name->name, PFID(ll_inode2fid(dir)), dir,
1189                mode, rdev);
1190
1191         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1192                 mode &= ~current_umask();
1193
1194         switch (mode & S_IFMT) {
1195         case 0:
1196                 mode |= S_IFREG; /* for mode = 0 case, fallthrough */
1197         case S_IFREG:
1198         case S_IFCHR:
1199         case S_IFBLK:
1200         case S_IFIFO:
1201         case S_IFSOCK:
1202                 err = ll_new_node(dir, dchild, NULL, mode, old_encode_dev(rdev),
1203                                   LUSTRE_OPC_MKNOD);
1204                 break;
1205         case S_IFDIR:
1206                 err = -EPERM;
1207                 break;
1208         default:
1209                 err = -EINVAL;
1210         }
1211
1212         if (!err)
1213                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKNOD, 1);
1214
1215         RETURN(err);
1216 }
1217
1218 #ifdef HAVE_IOP_ATOMIC_OPEN
1219 /*
1220  * Plain create. Intent create is handled in atomic_open.
1221  */
1222 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
1223                         umode_t mode, bool want_excl)
1224 {
1225         int rc;
1226
1227         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1228
1229         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), "
1230                            "flags=%u, excl=%d\n", dentry->d_name.len,
1231                dentry->d_name.name, PFID(ll_inode2fid(dir)),
1232                dir, mode, want_excl);
1233
1234         /* Using mknod(2) to create a regular file is designed to not recognize
1235          * volatile file name, so we use ll_mknod() here. */
1236         rc = ll_mknod(dir, dentry, mode, 0);
1237
1238         ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE, 1);
1239
1240         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, unhashed %d\n",
1241                dentry->d_name.len, dentry->d_name.name, d_unhashed(dentry));
1242
1243         return rc;
1244 }
1245 #else /* !HAVE_IOP_ATOMIC_OPEN */
1246 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
1247                         ll_umode_t mode, struct nameidata *nd)
1248 {
1249         struct ll_dentry_data *lld = ll_d2d(dentry);
1250         struct lookup_intent *it = NULL;
1251         int rc;
1252
1253         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1254
1255         if (lld != NULL)
1256                 it = lld->lld_it;
1257
1258         if (!it) {
1259                 /* LU-8559: use LUSTRE_OPC_CREATE for non atomic open case
1260                  * so that volatile file name is recoginized.
1261                  * Mknod(2), however, is designed to not recognize volatile
1262                  * file name to avoid inode leak under orphan directory until
1263                  * MDT reboot */
1264                 return ll_new_node(dir, dentry, NULL, mode, 0,
1265                                    LUSTRE_OPC_CREATE);
1266         }
1267
1268         lld->lld_it = NULL;
1269
1270         /* Was there an error? Propagate it! */
1271         if (it->it_status) {
1272                 rc = it->it_status;
1273                 goto out;
1274         }
1275
1276         rc = ll_create_it(dir, dentry, it, NULL, 0);
1277         if (nd && (nd->flags & LOOKUP_OPEN) && dentry->d_inode) { /* Open */
1278                 struct file *filp;
1279
1280                 nd->intent.open.file->private_data = it;
1281                 filp = lookup_instantiate_filp(nd, dentry, NULL);
1282                 if (IS_ERR(filp))
1283                         rc = PTR_ERR(filp);
1284         }
1285
1286 out:
1287         ll_intent_release(it);
1288         OBD_FREE(it, sizeof(*it));
1289
1290         if (!rc)
1291                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE, 1);
1292
1293         return rc;
1294 }
1295 #endif /* HAVE_IOP_ATOMIC_OPEN */
1296
1297 static int ll_symlink(struct inode *dir, struct dentry *dchild,
1298                       const char *oldpath)
1299 {
1300         struct qstr *name = &dchild->d_name;
1301         int err;
1302         ENTRY;
1303
1304         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), target=%.*s\n",
1305                name->len, name->name, PFID(ll_inode2fid(dir)),
1306                dir, 3000, oldpath);
1307
1308         err = ll_new_node(dir, dchild, oldpath, S_IFLNK | S_IRWXUGO, 0,
1309                           LUSTRE_OPC_SYMLINK);
1310
1311         if (!err)
1312                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_SYMLINK, 1);
1313
1314         RETURN(err);
1315 }
1316
1317 static int ll_link(struct dentry *old_dentry, struct inode *dir,
1318                    struct dentry *new_dentry)
1319 {
1320         struct inode *src = old_dentry->d_inode;
1321         struct qstr *name = &new_dentry->d_name;
1322         struct ll_sb_info *sbi = ll_i2sbi(dir);
1323         struct ptlrpc_request *request = NULL;
1324         struct md_op_data *op_data;
1325         int err;
1326
1327         ENTRY;
1328         CDEBUG(D_VFSTRACE, "VFS Op: inode="DFID"(%p), dir="DFID"(%p), "
1329                "target=%.*s\n", PFID(ll_inode2fid(src)), src,
1330                PFID(ll_inode2fid(dir)), dir, name->len, name->name);
1331
1332         op_data = ll_prep_md_op_data(NULL, src, dir, name->name, name->len,
1333                                      0, LUSTRE_OPC_ANY, NULL);
1334         if (IS_ERR(op_data))
1335                 RETURN(PTR_ERR(op_data));
1336
1337         err = md_link(sbi->ll_md_exp, op_data, &request);
1338         ll_finish_md_op_data(op_data);
1339         if (err)
1340                 GOTO(out, err);
1341
1342         ll_update_times(request, dir);
1343         ll_stats_ops_tally(sbi, LPROC_LL_LINK, 1);
1344         EXIT;
1345 out:
1346         ptlrpc_req_finished(request);
1347         RETURN(err);
1348 }
1349
1350 static int ll_mkdir(struct inode *dir, struct dentry *dchild, ll_umode_t mode)
1351 {
1352         struct qstr *name = &dchild->d_name;
1353         int err;
1354         ENTRY;
1355
1356         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1357                name->len, name->name, PFID(ll_inode2fid(dir)), dir);
1358
1359         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1360                 mode &= ~current_umask();
1361
1362         mode = (mode & (S_IRWXUGO|S_ISVTX)) | S_IFDIR;
1363
1364         err = ll_new_node(dir, dchild, NULL, mode, 0, LUSTRE_OPC_MKDIR);
1365         if (err == 0)
1366                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKDIR, 1);
1367
1368         RETURN(err);
1369 }
1370
1371 static int ll_rmdir(struct inode *dir, struct dentry *dchild)
1372 {
1373         struct qstr *name = &dchild->d_name;
1374         struct ptlrpc_request *request = NULL;
1375         struct md_op_data *op_data;
1376         int rc;
1377         ENTRY;
1378
1379         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1380                name->len, name->name, PFID(ll_inode2fid(dir)), dir);
1381
1382         if (unlikely(d_mountpoint(dchild)))
1383                 RETURN(-EBUSY);
1384
1385         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len,
1386                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1387         if (IS_ERR(op_data))
1388                 RETURN(PTR_ERR(op_data));
1389
1390         if (dchild->d_inode != NULL)
1391                 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1392
1393         op_data->op_fid2 = op_data->op_fid3;
1394         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1395         ll_finish_md_op_data(op_data);
1396         if (rc == 0) {
1397                 ll_update_times(request, dir);
1398                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR, 1);
1399         }
1400
1401         ptlrpc_req_finished(request);
1402         RETURN(rc);
1403 }
1404
1405 /**
1406  * Remove dir entry
1407  **/
1408 int ll_rmdir_entry(struct inode *dir, char *name, int namelen)
1409 {
1410         struct ptlrpc_request *request = NULL;
1411         struct md_op_data *op_data;
1412         int rc;
1413         ENTRY;
1414
1415         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1416                namelen, name, PFID(ll_inode2fid(dir)), dir);
1417
1418         op_data = ll_prep_md_op_data(NULL, dir, NULL, name, strlen(name),
1419                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1420         if (IS_ERR(op_data))
1421                 RETURN(PTR_ERR(op_data));
1422         op_data->op_cli_flags |= CLI_RM_ENTRY;
1423         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1424         ll_finish_md_op_data(op_data);
1425         if (rc == 0) {
1426                 ll_update_times(request, dir);
1427                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR, 1);
1428         }
1429
1430         ptlrpc_req_finished(request);
1431         RETURN(rc);
1432 }
1433
1434 static int ll_unlink(struct inode *dir, struct dentry *dchild)
1435 {
1436         struct qstr *name = &dchild->d_name;
1437         struct ptlrpc_request *request = NULL;
1438         struct md_op_data *op_data;
1439         struct mdt_body *body;
1440         int rc;
1441         ENTRY;
1442         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1443                name->len, name->name, PFID(ll_inode2fid(dir)), dir);
1444
1445         /*
1446          * XXX: unlink bind mountpoint maybe call to here,
1447          * just check it as vfs_unlink does.
1448          */
1449         if (unlikely(d_mountpoint(dchild)))
1450                 RETURN(-EBUSY);
1451
1452         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len, 0,
1453                                      LUSTRE_OPC_ANY, NULL);
1454         if (IS_ERR(op_data))
1455                 RETURN(PTR_ERR(op_data));
1456
1457         op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1458
1459         op_data->op_fid2 = op_data->op_fid3;
1460         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1461         ll_finish_md_op_data(op_data);
1462         if (rc)
1463                 GOTO(out, rc);
1464
1465         /*
1466          * The server puts attributes in on the last unlink, use them to update
1467          * the link count so the inode can be freed immediately.
1468          */
1469         body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
1470         if (body->mbo_valid & OBD_MD_FLNLINK)
1471                 set_nlink(dchild->d_inode, body->mbo_nlink);
1472
1473         ll_update_times(request, dir);
1474         ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_UNLINK, 1);
1475
1476 out:
1477         ptlrpc_req_finished(request);
1478         RETURN(rc);
1479 }
1480
1481 static int ll_rename(struct inode *src, struct dentry *src_dchild,
1482                      struct inode *tgt, struct dentry *tgt_dchild
1483 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
1484                      , unsigned int flags
1485 #endif
1486                      )
1487 {
1488         struct qstr *src_name = &src_dchild->d_name;
1489         struct qstr *tgt_name = &tgt_dchild->d_name;
1490         struct ptlrpc_request *request = NULL;
1491         struct ll_sb_info *sbi = ll_i2sbi(src);
1492         struct md_op_data *op_data;
1493         int err;
1494         ENTRY;
1495
1496 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
1497         if (flags)
1498                 return -EINVAL;
1499 #endif
1500
1501         CDEBUG(D_VFSTRACE, "VFS Op:oldname=%.*s, src_dir="DFID
1502                "(%p), newname=%.*s, tgt_dir="DFID"(%p)\n",
1503                src_name->len, src_name->name,
1504                PFID(ll_inode2fid(src)), src, tgt_name->len,
1505                tgt_name->name, PFID(ll_inode2fid(tgt)), tgt);
1506
1507         if (unlikely(d_mountpoint(src_dchild) || d_mountpoint(tgt_dchild)))
1508                 RETURN(-EBUSY);
1509
1510         op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, 0,
1511                                      LUSTRE_OPC_ANY, NULL);
1512         if (IS_ERR(op_data))
1513                 RETURN(PTR_ERR(op_data));
1514
1515         if (src_dchild->d_inode != NULL)
1516                 op_data->op_fid3 = *ll_inode2fid(src_dchild->d_inode);
1517
1518         if (tgt_dchild->d_inode != NULL)
1519                 op_data->op_fid4 = *ll_inode2fid(tgt_dchild->d_inode);
1520
1521         err = md_rename(sbi->ll_md_exp, op_data,
1522                         src_name->name, src_name->len,
1523                         tgt_name->name, tgt_name->len, &request);
1524         ll_finish_md_op_data(op_data);
1525         if (!err) {
1526                 ll_update_times(request, src);
1527                 ll_update_times(request, tgt);
1528                 ll_stats_ops_tally(sbi, LPROC_LL_RENAME, 1);
1529         }
1530
1531         ptlrpc_req_finished(request);
1532
1533         if (err == 0)
1534                 d_move(src_dchild, tgt_dchild);
1535
1536         RETURN(err);
1537 }
1538
1539 const struct inode_operations ll_dir_inode_operations = {
1540         .mknod          = ll_mknod,
1541 #ifdef HAVE_IOP_ATOMIC_OPEN
1542         .atomic_open    = ll_atomic_open,
1543 #endif
1544         .lookup         = ll_lookup_nd,
1545         .create         = ll_create_nd,
1546         /* We need all these non-raw things for NFSD, to not patch it. */
1547         .unlink         = ll_unlink,
1548         .mkdir          = ll_mkdir,
1549         .rmdir          = ll_rmdir,
1550         .symlink        = ll_symlink,
1551         .link           = ll_link,
1552         .rename         = ll_rename,
1553         .setattr        = ll_setattr,
1554         .getattr        = ll_getattr,
1555         .permission     = ll_inode_permission,
1556 #ifdef HAVE_IOP_XATTR
1557         .setxattr       = ll_setxattr,
1558         .getxattr       = ll_getxattr,
1559         .removexattr    = ll_removexattr,
1560 #endif
1561         .listxattr      = ll_listxattr,
1562 #ifdef HAVE_IOP_GET_ACL
1563         .get_acl        = ll_get_acl,
1564 #endif
1565 #ifdef HAVE_IOP_SET_ACL
1566         .set_acl        = ll_set_acl,
1567 #endif
1568 };
1569
1570 const struct inode_operations ll_special_inode_operations = {
1571         .setattr        = ll_setattr,
1572         .getattr        = ll_getattr,
1573         .permission     = ll_inode_permission,
1574 #ifdef HAVE_IOP_XATTR
1575         .setxattr       = ll_setxattr,
1576         .getxattr       = ll_getxattr,
1577         .removexattr    = ll_removexattr,
1578 #endif
1579         .listxattr      = ll_listxattr,
1580 #ifdef HAVE_IOP_GET_ACL
1581         .get_acl        = ll_get_acl,
1582 #endif
1583 #ifdef HAVE_IOP_SET_ACL
1584         .set_acl        = ll_set_acl,
1585 #endif
1586 };