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  */
  31
  32 #include <linux/fs.h>
  33 #include <linux/sched.h>
  34 #include <linux/mm.h>
  35 #include <linux/file.h>
  36 #include <linux/quotaops.h>
  37 #include <linux/highmem.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/user_namespace.h>
  40 #include <linux/uidgid.h>
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 #include <obd_support.h>
  45 #include <lustre_fid.h>
  46 #include <lustre_dlm.h>
  47 #include "llite_internal.h"
  48
  49 static int ll_create_it(struct inode *dir, struct dentry *dentry,
  50                         struct lookup_intent *it,
  51                         void *secctx, __u32 secctxlen, bool encrypt,
  52                         void *encctx, __u32 encctxlen);
  53
  54 /* called from iget5_locked->find_inode() under inode_lock spinlock */
  55 static int ll_test_inode(struct inode *inode, void *opaque)
  56 {
  57         struct ll_inode_info    *lli = ll_i2info(inode);
  58         struct lustre_md        *md = opaque;
  59
  60         if (unlikely(!(md->body->mbo_valid & OBD_MD_FLID))) {
  61                 CERROR("MDS body missing FID\n");
  62                 return 0;
  63         }
  64
  65         if (!lu_fid_eq(&lli->lli_fid, &md->body->mbo_fid1))
  66                 return 0;
  67
  68         return 1;
  69 }
  70
  71 static int ll_set_inode(struct inode *inode, void *opaque)
  72 {
  73         struct ll_inode_info *lli = ll_i2info(inode);
  74         struct mdt_body *body = ((struct lustre_md *)opaque)->body;
  75
  76         if (unlikely(!(body->mbo_valid & OBD_MD_FLID))) {
  77                 CERROR("MDS body missing FID\n");
  78                 return -EINVAL;
  79         }
  80
  81         lli->lli_fid = body->mbo_fid1;
  82         if (unlikely(!(body->mbo_valid & OBD_MD_FLTYPE))) {
  83                 CERROR("Can not initialize inode "DFID" without object type: "
  84                        "valid = %#llx\n",
  85                        PFID(&lli->lli_fid), body->mbo_valid);
  86                 return -EINVAL;
  87         }
  88
  89         inode->i_mode = (inode->i_mode & ~S_IFMT) | (body->mbo_mode & S_IFMT);
  90         if (unlikely(inode->i_mode == 0)) {
  91                 CERROR("Invalid inode "DFID" type\n", PFID(&lli->lli_fid));
  92                 return -EINVAL;
  93         }
  94
  95         ll_lli_init(lli);
  96
  97         return 0;
  98 }
  99
 100
 101 /**
 102  * Get an inode by inode number(@hash), which is already instantiated by
 103  * the intent lookup).
 104  */
 105 struct inode *ll_iget(struct super_block *sb, ino_t hash,
 106                       struct lustre_md *md)
 107 {
 108         struct inode    *inode;
 109         int             rc = 0;
 110
 111         ENTRY;
 112
 113         LASSERT(hash != 0);
 114         inode = iget5_locked(sb, hash, ll_test_inode, ll_set_inode, md);
 115         if (inode == NULL)
 116                 RETURN(ERR_PTR(-ENOMEM));
 117
 118         if (inode->i_state & I_NEW) {
 119                 rc = ll_read_inode2(inode, md);
 120                 if (rc == 0 && S_ISREG(inode->i_mode) &&
 121                     ll_i2info(inode)->lli_clob == NULL)
 122                         rc = cl_file_inode_init(inode, md);
 123
 124                 if (rc != 0) {
 125                         /* Let's clear directory lsm here, otherwise
 126                          * make_bad_inode() will reset the inode mode
 127                          * to regular, then ll_clear_inode will not
 128                          * be able to clear lsm_md */
 129                         if (S_ISDIR(inode->i_mode))
 130                                 ll_dir_clear_lsm_md(inode);
 131                         make_bad_inode(inode);
 132                         unlock_new_inode(inode);
 133                         iput(inode);
 134                         inode = ERR_PTR(rc);
 135                 } else {
 136                         inode_has_no_xattr(inode);
 137                         unlock_new_inode(inode);
 138                 }
 139         } else if (is_bad_inode(inode)) {
 140                 iput(inode);
 141                 inode = ERR_PTR(-ESTALE);
 142         } else if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
 143                 rc = ll_update_inode(inode, md);
 144                 CDEBUG(D_VFSTRACE, "got inode: "DFID"(%p): rc = %d\n",
 145                        PFID(&md->body->mbo_fid1), inode, rc);
 146                 if (rc != 0) {
 147                         if (S_ISDIR(inode->i_mode))
 148                                 ll_dir_clear_lsm_md(inode);
 149                         iput(inode);
 150                         inode = ERR_PTR(rc);
 151                 }
 152         }
 153
 154         RETURN(inode);
 155 }
 156
 157 /* mark negative sub file dentries invalid and prune unused dentries */
 158 static void ll_prune_negative_children(struct inode *dir)
 159 {
 160         struct dentry *dentry;
 161         struct dentry *child;
 162
 163         ENTRY;
 164
 165 restart:
 166         spin_lock(&dir->i_lock);
 167         hlist_for_each_entry(dentry, &dir->i_dentry, d_alias) {
 168                 spin_lock(&dentry->d_lock);
 169                 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
 170                         if (child->d_inode)
 171                                 continue;
 172
 173                         spin_lock_nested(&child->d_lock, DENTRY_D_LOCK_NESTED);
 174                         if (lld_is_init(child))
 175                                 ll_d2d(child)->lld_invalid = 1;
 176                         if (!ll_d_count(child)) {
 177                                 dget_dlock(child);
 178                                 __d_drop(child);
 179                                 spin_unlock(&child->d_lock);
 180                                 spin_unlock(&dentry->d_lock);
 181                                 spin_unlock(&dir->i_lock);
 182
 183                                 CDEBUG(D_DENTRY, "prune negative dentry %pd\n",
 184                                        child);
 185
 186                                 dput(child);
 187                                 goto restart;
 188                         }
 189                         spin_unlock(&child->d_lock);
 190                 }
 191                 spin_unlock(&dentry->d_lock);
 192         }
 193         spin_unlock(&dir->i_lock);
 194
 195         EXIT;
 196 }
 197
 198 int ll_test_inode_by_fid(struct inode *inode, void *opaque)
 199 {
 200         return lu_fid_eq(&ll_i2info(inode)->lli_fid, opaque);
 201 }
 202
 203 static int ll_dom_lock_cancel(struct inode *inode, struct ldlm_lock *lock)
 204 {
 205         struct lu_env *env;
 206         struct ll_inode_info *lli = ll_i2info(inode);
 207         __u16 refcheck;
 208         int rc;
 209         ENTRY;
 210
 211         env = cl_env_get(&refcheck);
 212         if (IS_ERR(env))
 213                 RETURN(PTR_ERR(env));
 214
 215         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_REPLAY_PAUSE, cfs_fail_val);
 216
 217         /* reach MDC layer to flush data under  the DoM ldlm lock */
 218         rc = cl_object_flush(env, lli->lli_clob, lock);
 219         if (rc == -ENODATA) {
 220                 CDEBUG(D_INODE, "inode "DFID" layout has no DoM stripe\n",
 221                        PFID(ll_inode2fid(inode)));
 222                 /* most likely result of layout change, do nothing */
 223                 rc = 0;
 224         }
 225
 226         cl_env_put(env, &refcheck);
 227         RETURN(rc);
 228 }
 229
 230 static void ll_lock_cancel_bits(struct ldlm_lock *lock, __u64 to_cancel)
 231 {
 232         struct inode *inode = ll_inode_from_resource_lock(lock);
 233         struct ll_inode_info *lli;
 234         __u64 bits = to_cancel;
 235         int rc;
 236
 237         ENTRY;
 238
 239         if (!inode) {
 240                 /* That means the inode is evicted most likely and may cause
 241                  * the skipping of lock cleanups below, so print the message
 242                  * about that in log.
 243                  */
 244                 if (lock->l_resource->lr_lvb_inode)
 245                         LDLM_DEBUG(lock,
 246                                    "can't take inode for the lock (%sevicted)\n",
 247                                    lock->l_resource->lr_lvb_inode->i_state &
 248                                    I_FREEING ? "" : "not ");
 249                 RETURN_EXIT;
 250         }
 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                 ll_xattr_cache_destroy(inode);
 261                 bits &= ~MDS_INODELOCK_XATTR;
 262         }
 263
 264         /* For OPEN locks we differentiate between lock modes
 265          * LCK_CR, LCK_CW, LCK_PR - bug 22891 */
 266         if (bits & MDS_INODELOCK_OPEN)
 267                 ll_have_md_lock(inode, &bits, lock->l_req_mode);
 268
 269         if (bits & MDS_INODELOCK_OPEN) {
 270                 fmode_t fmode;
 271
 272                 switch (lock->l_req_mode) {
 273                 case LCK_CW:
 274                         fmode = FMODE_WRITE;
 275                         break;
 276                 case LCK_PR:
 277                         fmode = FMODE_EXEC;
 278                         break;
 279                 case LCK_CR:
 280                         fmode = FMODE_READ;
 281                         break;
 282                 default:
 283                         LDLM_ERROR(lock, "bad lock mode for OPEN lock");
 284                         LBUG();
 285                 }
 286
 287                 ll_md_real_close(inode, fmode);
 288
 289                 bits &= ~MDS_INODELOCK_OPEN;
 290         }
 291
 292         if (bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
 293                     MDS_INODELOCK_LAYOUT | MDS_INODELOCK_PERM |
 294                     MDS_INODELOCK_DOM))
 295                 ll_have_md_lock(inode, &bits, LCK_MINMODE);
 296
 297         if (bits & MDS_INODELOCK_DOM) {
 298                 rc =  ll_dom_lock_cancel(inode, lock);
 299                 if (rc < 0)
 300                         CDEBUG(D_INODE, "cannot flush DoM data "
 301                                DFID": rc = %d\n",
 302                                PFID(ll_inode2fid(inode)), rc);
 303         }
 304
 305         if (bits & MDS_INODELOCK_LAYOUT) {
 306                 struct cl_object_conf conf = {
 307                         .coc_opc = OBJECT_CONF_INVALIDATE,
 308                         .coc_inode = inode,
 309                 };
 310
 311                 rc = ll_layout_conf(inode, &conf);
 312                 if (rc < 0)
 313                         CDEBUG(D_INODE, "cannot invalidate layout of "
 314                                DFID": rc = %d\n",
 315                                PFID(ll_inode2fid(inode)), rc);
 316         }
 317
 318         lli = ll_i2info(inode);
 319
 320         if (bits & MDS_INODELOCK_UPDATE)
 321                 set_bit(LLIF_UPDATE_ATIME, &lli->lli_flags);
 322
 323         if ((bits & MDS_INODELOCK_UPDATE) && S_ISDIR(inode->i_mode)) {
 324                 CDEBUG(D_INODE, "invalidating inode "DFID" lli = %p, "
 325                        "pfid  = "DFID"\n", PFID(ll_inode2fid(inode)),
 326                        lli, PFID(&lli->lli_pfid));
 327                 truncate_inode_pages(inode->i_mapping, 0);
 328
 329                 if (unlikely(!fid_is_zero(&lli->lli_pfid))) {
 330                         struct inode *master_inode = NULL;
 331                         unsigned long hash;
 332
 333                         /* This is slave inode, since all of the child dentry
 334                          * is connected on the master inode, so we have to
 335                          * invalidate the negative children on master inode */
 336                         CDEBUG(D_INODE, "Invalidate s"DFID" m"DFID"\n",
 337                                PFID(ll_inode2fid(inode)), PFID(&lli->lli_pfid));
 338
 339                         hash = cl_fid_build_ino(&lli->lli_pfid,
 340                                         ll_need_32bit_api(ll_i2sbi(inode)));
 341
 342                         /* Do not lookup the inode with ilookup5, otherwise
 343                          * it will cause dead lock,
 344                          * 1. Client1 send chmod req to the MDT0, then on MDT0,
 345                          * it enqueues master and all of its slaves lock,
 346                          * (mdt_attr_set() -> mdt_lock_slaves()), after gets
 347                          * master and stripe0 lock, it will send the enqueue
 348                          * req (for stripe1) to MDT1, then MDT1 finds the lock
 349                          * has been granted to client2. Then MDT1 sends blocking
 350                          * ast to client2.
 351                          * 2. At the same time, client2 tries to unlink
 352                          * the striped dir (rm -rf striped_dir), and during
 353                          * lookup, it will hold the master inode of the striped
 354                          * directory, whose inode state is NEW, then tries to
 355                          * revalidate all of its slaves, (ll_prep_inode()->
 356                          * ll_iget()->ll_read_inode2()-> ll_update_inode().).
 357                          * And it will be blocked on the server side because
 358                          * of 1.
 359                          * 3. Then the client get the blocking_ast req, cancel
 360                          * the lock, but being blocked if using ->ilookup5()),
 361                          * because master inode state is NEW. */
 362                         master_inode = ilookup5_nowait(inode->i_sb, hash,
 363                                                         ll_test_inode_by_fid,
 364                                                         (void *)&lli->lli_pfid);
 365                         if (master_inode) {
 366                                 ll_prune_negative_children(master_inode);
 367                                 iput(master_inode);
 368                         }
 369                 } else {
 370                         ll_prune_negative_children(inode);
 371                 }
 372         }
 373
 374         if ((bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM)) &&
 375             !is_root_inode(inode))
 376                 ll_prune_aliases(inode);
 377
 378         if (bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM))
 379                 forget_all_cached_acls(inode);
 380
 381         iput(inode);
 382         RETURN_EXIT;
 383 }
 384
 385 /* Check if the given lock may be downgraded instead of canceling and
 386  * that convert is really needed. */
 387 int ll_md_need_convert(struct ldlm_lock *lock)
 388 {
 389         struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
 390         struct inode *inode;
 391         __u64 wanted = lock->l_policy_data.l_inodebits.cancel_bits;
 392         __u64 bits = lock->l_policy_data.l_inodebits.bits & ~wanted;
 393         enum ldlm_mode mode = LCK_MINMODE;
 394
 395         if (!lock->l_conn_export ||
 396             !exp_connect_lock_convert(lock->l_conn_export))
 397                 return 0;
 398
 399         if (!wanted || !bits || ldlm_is_cancel(lock))
 400                 return 0;
 401
 402         /* do not convert locks other than DOM for now */
 403         if (!((bits | wanted) & MDS_INODELOCK_DOM))
 404                 return 0;
 405
 406         /* We may have already remaining bits in some other lock so
 407          * lock convert will leave us just extra lock for the same bit.
 408          * Check if client has other lock with the same bits and the same
 409          * or lower mode and don't convert if any.
 410          */
 411         switch (lock->l_req_mode) {
 412         case LCK_PR:
 413                 mode = LCK_PR;
 414                 /* fallthrough */
 415         case LCK_PW:
 416                 mode |= LCK_CR;
 417                 break;
 418         case LCK_CW:
 419                 mode = LCK_CW;
 420                 /* fallthrough */
 421         case LCK_CR:
 422                 mode |= LCK_CR;
 423                 break;
 424         default:
 425                 /* do not convert other modes */
 426                 return 0;
 427         }
 428
 429         /* is lock is too old to be converted? */
 430         lock_res_and_lock(lock);
 431         if (ktime_after(ktime_get(),
 432                         ktime_add(lock->l_last_used, ns->ns_dirty_age_limit))) {
 433                 unlock_res_and_lock(lock);
 434                 return 0;
 435         }
 436         unlock_res_and_lock(lock);
 437
 438         inode = ll_inode_from_resource_lock(lock);
 439         ll_have_md_lock(inode, &bits, mode);
 440         iput(inode);
 441         return !!(bits);
 442 }
 443
 444 int ll_md_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *ld,
 445                        void *data, int flag)
 446 {
 447         struct lustre_handle lockh;
 448         int rc;
 449
 450         ENTRY;
 451
 452         switch (flag) {
 453         case LDLM_CB_BLOCKING:
 454         {
 455                 __u64 cancel_flags = LCF_ASYNC;
 456
 457                 /* if lock convert is not needed then still have to
 458                  * pass lock via ldlm_cli_convert() to keep all states
 459                  * correct, set cancel_bits to full lock bits to cause
 460                  * full cancel to happen.
 461                  */
 462                 if (!ll_md_need_convert(lock)) {
 463                         lock_res_and_lock(lock);
 464                         lock->l_policy_data.l_inodebits.cancel_bits =
 465                                         lock->l_policy_data.l_inodebits.bits;
 466                         unlock_res_and_lock(lock);
 467                 }
 468                 rc = ldlm_cli_convert(lock, cancel_flags);
 469                 if (!rc)
 470                         RETURN(0);
 471                 /* continue with cancel otherwise */
 472                 ldlm_lock2handle(lock, &lockh);
 473                 rc = ldlm_cli_cancel(&lockh, cancel_flags);
 474                 if (rc < 0) {
 475                         CDEBUG(D_INODE, "ldlm_cli_cancel: rc = %d\n", rc);
 476                         RETURN(rc);
 477                 }
 478                 break;
 479         }
 480         case LDLM_CB_CANCELING:
 481         {
 482                 __u64 to_cancel = lock->l_policy_data.l_inodebits.bits;
 483
 484                 /* Nothing to do for non-granted locks */
 485                 if (!ldlm_is_granted(lock))
 486                         break;
 487
 488                 /* If 'ld' is supplied then bits to be cancelled are passed
 489                  * implicitly by lock converting and cancel_bits from 'ld'
 490                  * should be used. Otherwise full cancel is being performed
 491                  * and lock inodebits are used.
 492                  *
 493                  * Note: we cannot rely on cancel_bits in lock itself at this
 494                  * moment because they can be changed by concurrent thread,
 495                  * so ldlm_cli_inodebits_convert() pass cancel bits implicitly
 496                  * in 'ld' parameter.
 497                  */
 498                 if (ld) {
 499                         /* partial bits cancel allowed only during convert */
 500                         LASSERT(ldlm_is_converting(lock));
 501                         /* mask cancel bits by lock bits so only no any unused
 502                          * bits are passed to ll_lock_cancel_bits()
 503                          */
 504                         to_cancel &= ld->l_policy_data.l_inodebits.cancel_bits;
 505                 }
 506                 ll_lock_cancel_bits(lock, to_cancel);
 507                 break;
 508         }
 509         default:
 510                 LBUG();
 511         }
 512
 513         RETURN(0);
 514 }
 515
 516 __u32 ll_i2suppgid(struct inode *i)
 517 {
 518         if (in_group_p(i->i_gid))
 519                 return (__u32)from_kgid(&init_user_ns, i->i_gid);
 520         else
 521                 return (__u32) __kgid_val(INVALID_GID);
 522 }
 523
 524 /* Pack the required supplementary groups into the supplied groups array.
 525  * If we don't need to use the groups from the target inode(s) then we
 526  * instead pack one or more groups from the user's supplementary group
 527  * array in case it might be useful.  Not needed if doing an MDS-side upcall. */
 528 void ll_i2gids(__u32 *suppgids, struct inode *i1, struct inode *i2)
 529 {
 530         LASSERT(i1 != NULL);
 531         LASSERT(suppgids != NULL);
 532
 533         suppgids[0] = ll_i2suppgid(i1);
 534
 535         if (i2)
 536                 suppgids[1] = ll_i2suppgid(i2);
 537         else
 538                 suppgids[1] = -1;
 539 }
 540
 541 /*
 542  * try to reuse three types of dentry:
 543  * 1. unhashed alias, this one is unhashed by d_invalidate (but it may be valid
 544  *    by concurrent .revalidate).
 545  * 2. INVALID alias (common case for no valid ldlm lock held, but this flag may
 546  *    be cleared by others calling d_lustre_revalidate).
 547  * 3. DISCONNECTED alias.
 548  */
 549 static struct dentry *ll_find_alias(struct inode *inode, struct dentry *dentry)
 550 {
 551         struct dentry *alias, *discon_alias, *invalid_alias;
 552
 553         if (hlist_empty(&inode->i_dentry))
 554                 return NULL;
 555
 556         discon_alias = invalid_alias = NULL;
 557
 558         spin_lock(&inode->i_lock);
 559         hlist_for_each_entry(alias, &inode->i_dentry, d_alias) {
 560                 LASSERT(alias != dentry);
 561
 562                 spin_lock(&alias->d_lock);
 563                 if ((alias->d_flags & DCACHE_DISCONNECTED) &&
 564                     S_ISDIR(inode->i_mode))
 565                         /* LASSERT(last_discon == NULL); LU-405, bz 20055 */
 566                         discon_alias = alias;
 567                 else if (alias->d_parent == dentry->d_parent             &&
 568                          alias->d_name.hash == dentry->d_name.hash       &&
 569                          alias->d_name.len == dentry->d_name.len         &&
 570                          memcmp(alias->d_name.name, dentry->d_name.name,
 571                                 dentry->d_name.len) == 0)
 572                         invalid_alias = alias;
 573                 spin_unlock(&alias->d_lock);
 574
 575                 if (invalid_alias)
 576                         break;
 577         }
 578         alias = invalid_alias ?: discon_alias ?: NULL;
 579         if (alias) {
 580                 spin_lock(&alias->d_lock);
 581                 dget_dlock(alias);
 582                 spin_unlock(&alias->d_lock);
 583         }
 584         spin_unlock(&inode->i_lock);
 585
 586         return alias;
 587 }
 588
 589 /*
 590  * Similar to d_splice_alias(), but lustre treats invalid alias
 591  * similar to DCACHE_DISCONNECTED, and tries to use it anyway.
 592  */
 593 struct dentry *ll_splice_alias(struct inode *inode, struct dentry *de)
 594 {
 595         struct dentry *new;
 596
 597         if (inode) {
 598                 new = ll_find_alias(inode, de);
 599                 if (new) {
 600                         if (!ll_d_setup(new, true))
 601                                 return ERR_PTR(-ENOMEM);
 602                         d_move(new, de);
 603                         iput(inode);
 604                         CDEBUG(D_DENTRY,
 605                                "Reuse dentry %p inode %p refc %d flags %#x\n",
 606                               new, new->d_inode, ll_d_count(new), new->d_flags);
 607                         return new;
 608                 }
 609         }
 610         if (!ll_d_setup(de, false))
 611                 return ERR_PTR(-ENOMEM);
 612         d_add(de, inode);
 613
 614         /* this needs only to be done for foreign symlink dirs as
 615          * DCACHE_SYMLINK_TYPE is already set by d_flags_for_inode()
 616          * kernel routine for files with symlink ops (ie, real symlink)
 617          */
 618         if (inode && S_ISDIR(inode->i_mode) &&
 619             ll_sbi_has_foreign_symlink(ll_i2sbi(inode)) &&
 620 #ifdef HAVE_IOP_GET_LINK
 621             inode->i_op->get_link) {
 622 #else
 623             inode->i_op->follow_link) {
 624 #endif
 625                 CDEBUG(D_INFO, "%s: inode "DFID": faking foreign dir as a symlink\n",
 626                        ll_i2sbi(inode)->ll_fsname, PFID(ll_inode2fid(inode)));
 627                 spin_lock(&de->d_lock);
 628                 /* like d_flags_for_inode() already does for files */
 629                 de->d_flags = (de->d_flags & ~DCACHE_ENTRY_TYPE) |
 630                               DCACHE_SYMLINK_TYPE;
 631                 spin_unlock(&de->d_lock);
 632         }
 633
 634         CDEBUG(D_DENTRY, "Add dentry %p inode %p refc %d flags %#x\n",
 635                de, de->d_inode, ll_d_count(de), de->d_flags);
 636         return de;
 637 }
 638
 639 static int ll_lookup_it_finish(struct ptlrpc_request *request,
 640                                struct lookup_intent *it,
 641                                struct inode *parent, struct dentry **de,
 642                                void *secctx, __u32 secctxlen,
 643                                void *encctx, __u32 encctxlen,
 644                                ktime_t kstart, bool encrypt)
 645 {
 646         struct inode             *inode = NULL;
 647         __u64                     bits = 0;
 648         int                       rc;
 649         struct dentry *alias;
 650         ENTRY;
 651
 652         /* NB 1 request reference will be taken away by ll_intent_lock()
 653          * when I return */
 654         CDEBUG(D_DENTRY, "it %p it_disposition %x\n", it,
 655                it->it_disposition);
 656         if (!it_disposition(it, DISP_LOOKUP_NEG)) {
 657                 struct req_capsule *pill = &request->rq_pill;
 658                 struct mdt_body *body = req_capsule_server_get(pill,
 659                                                                &RMF_MDT_BODY);
 660
 661                 rc = ll_prep_inode(&inode, &request->rq_pill, (*de)->d_sb, it);
 662                 if (rc)
 663                         RETURN(rc);
 664
 665                 /* If encryption context was returned by MDT, put it in
 666                  * inode now to save an extra getxattr and avoid deadlock.
 667                  */
 668                 if (body->mbo_valid & OBD_MD_ENCCTX) {
 669                         encctx = req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
 670                         encctxlen = req_capsule_get_size(pill,
 671                                                          &RMF_FILE_ENCCTX,
 672                                                          RCL_SERVER);
 673
 674                         if (encctxlen) {
 675                                 CDEBUG(D_SEC,
 676                                        "server returned encryption ctx for "DFID"\n",
 677                                        PFID(ll_inode2fid(inode)));
 678                                 rc = ll_xattr_cache_insert(inode,
 679                                                LL_XATTR_NAME_ENCRYPTION_CONTEXT,
 680                                                            encctx, encctxlen);
 681                                 if (rc)
 682                                         CWARN("%s: cannot set enc ctx for "DFID": rc = %d\n",
 683                                               ll_i2sbi(inode)->ll_fsname,
 684                                               PFID(ll_inode2fid(inode)), rc);
 685                                 else if (encrypt) {
 686                                         rc = llcrypt_get_encryption_info(inode);
 687                                         if (rc)
 688                                                 CDEBUG(D_SEC,
 689                                                  "cannot get enc info for "DFID": rc = %d\n",
 690                                                  PFID(ll_inode2fid(inode)), rc);
 691                                 }
 692                         }
 693                 }
 694
 695                 ll_set_lock_data(ll_i2sbi(parent)->ll_md_exp, inode, it, &bits);
 696                 /* OPEN can return data if lock has DoM+LAYOUT bits set */
 697                 if (it->it_op & IT_OPEN &&
 698                     bits & MDS_INODELOCK_DOM && bits & MDS_INODELOCK_LAYOUT)
 699                         ll_dom_finish_open(inode, request);
 700
 701                 /* We used to query real size from OSTs here, but actually
 702                  * this is not needed. For stat() calls size would be updated
 703                  * from subsequent do_revalidate()->ll_inode_revalidate_it() in
 704                  * 2.4 and
 705                  * vfs_getattr_it->ll_getattr()->ll_inode_revalidate_it() in 2.6
 706                  * Everybody else who needs correct file size would call
 707                  * ll_glimpse_size or some equivalent themselves anyway.
 708                  * Also see bug 7198.
 709                  */
 710
 711                 /* If security context was returned by MDT, put it in
 712                  * inode now to save an extra getxattr from security hooks,
 713                  * and avoid deadlock.
 714                  */
 715                 if (body->mbo_valid & OBD_MD_SECCTX) {
 716                         secctx = req_capsule_server_get(pill, &RMF_FILE_SECCTX);
 717                         secctxlen = req_capsule_get_size(pill,
 718                                                            &RMF_FILE_SECCTX,
 719                                                            RCL_SERVER);
 720
 721                         if (secctxlen)
 722                                 CDEBUG(D_SEC, "server returned security context"
 723                                        " for "DFID"\n",
 724                                        PFID(ll_inode2fid(inode)));
 725                 }
 726
 727                 if (secctx != NULL && secctxlen != 0) {
 728                         /* no need to protect selinux_inode_setsecurity() by
 729                          * inode_lock. Taking it would lead to a client deadlock
 730                          * LU-13617
 731                          */
 732                         rc = security_inode_notifysecctx(inode, secctx,
 733                                                          secctxlen);
 734                         if (rc)
 735                                 CWARN("%s: cannot set security context for "DFID": rc = %d\n",
 736                                       ll_i2sbi(inode)->ll_fsname,
 737                                       PFID(ll_inode2fid(inode)),
 738                                       rc);
 739                 }
 740         }
 741
 742         /* Only hash *de if it is unhashed (new dentry).
 743          * Atoimc_open may passin hashed dentries for open.
 744          */
 745         alias = ll_splice_alias(inode, *de);
 746         if (IS_ERR(alias))
 747                 GOTO(out, rc = PTR_ERR(alias));
 748
 749         *de = alias;
 750
 751         if (!it_disposition(it, DISP_LOOKUP_NEG)) {
 752                 /* we have lookup look - unhide dentry */
 753                 if (bits & MDS_INODELOCK_LOOKUP) {
 754                         d_lustre_revalidate(*de);
 755                         ll_update_dir_depth(parent, (*de)->d_inode);
 756                 }
 757
 758                 if (encrypt) {
 759                         rc = llcrypt_get_encryption_info(inode);
 760                         if (rc)
 761                                 GOTO(out, rc);
 762                         if (!llcrypt_has_encryption_key(inode))
 763                                 GOTO(out, rc = -ENOKEY);
 764                 }
 765         } else if (!it_disposition(it, DISP_OPEN_CREATE)) {
 766                 /*
 767                  * If file was created on the server, the dentry is revalidated
 768                  * in ll_create_it if the lock allows for it.
 769                  */
 770                 /* Check that parent has UPDATE lock. */
 771                 struct lookup_intent parent_it = {
 772                                         .it_op = IT_GETATTR,
 773                                         .it_lock_handle = 0 };
 774                 struct lu_fid   fid = ll_i2info(parent)->lli_fid;
 775
 776                 /* If it is striped directory, get the real stripe parent */
 777                 if (unlikely(ll_dir_striped(parent))) {
 778                         rc = md_get_fid_from_lsm(ll_i2mdexp(parent),
 779                                                  ll_i2info(parent)->lli_lsm_md,
 780                                                  (*de)->d_name.name,
 781                                                  (*de)->d_name.len, &fid);
 782                         if (rc != 0)
 783                                 GOTO(out, rc);
 784                 }
 785
 786                 if (md_revalidate_lock(ll_i2mdexp(parent), &parent_it, &fid,
 787                                        NULL)) {
 788                         d_lustre_revalidate(*de);
 789                         ll_intent_release(&parent_it);
 790                 }
 791         }
 792
 793         if (it_disposition(it, DISP_OPEN_CREATE)) {
 794                 ll_stats_ops_tally(ll_i2sbi(parent), LPROC_LL_MKNOD,
 795                                    ktime_us_delta(ktime_get(), kstart));
 796         }
 797
 798         GOTO(out, rc = 0);
 799
 800 out:
 801         if (rc != 0 && it->it_op & IT_OPEN) {
 802                 ll_intent_drop_lock(it);
 803                 ll_open_cleanup((*de)->d_sb, &request->rq_pill);
 804         }
 805
 806         return rc;
 807 }
 808
 809 static struct dentry *ll_lookup_it(struct inode *parent, struct dentry *dentry,
 810                                    struct lookup_intent *it,
 811                                    void **secctx, __u32 *secctxlen,
 812                                    struct pcc_create_attach *pca,
 813                                    bool encrypt,
 814                                    void **encctx, __u32 *encctxlen)
 815 {
 816         ktime_t kstart = ktime_get();
 817         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
 818         struct dentry *save = dentry, *retval;
 819         struct ptlrpc_request *req = NULL;
 820         struct md_op_data *op_data = NULL;
 821         struct lov_user_md *lum = NULL;
 822         __u32 opc;
 823         int rc;
 824         char secctx_name[XATTR_NAME_MAX + 1];
 825         struct llcrypt_name fname;
 826         struct lu_fid fid;
 827         ENTRY;
 828
 829         if (dentry->d_name.len > ll_i2sbi(parent)->ll_namelen)
 830                 RETURN(ERR_PTR(-ENAMETOOLONG));
 831
 832         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), intent=%s\n",
 833                dentry, PFID(ll_inode2fid(parent)), parent, LL_IT2STR(it));
 834
 835         if (d_mountpoint(dentry))
 836                 CERROR("Tell Peter, lookup on mtpt, it %s\n", LL_IT2STR(it));
 837
 838         if (it == NULL || it->it_op == IT_GETXATTR)
 839                 it = &lookup_it;
 840
 841         if (it->it_op == IT_GETATTR && dentry_may_statahead(parent, dentry)) {
 842                 rc = ll_revalidate_statahead(parent, &dentry, 0);
 843                 if (rc == 1)
 844                         RETURN(dentry == save ? NULL : dentry);
 845         }
 846
 847         if (it->it_op & IT_OPEN && it->it_flags & FMODE_WRITE &&
 848             dentry->d_sb->s_flags & SB_RDONLY)
 849                 RETURN(ERR_PTR(-EROFS));
 850
 851         if (it->it_op & IT_CREAT)
 852                 opc = LUSTRE_OPC_CREATE;
 853         else
 854                 opc = LUSTRE_OPC_LOOKUP;
 855
 856         /* Here we should be calling llcrypt_prepare_lookup(). But it installs a
 857          * custom ->d_revalidate() method, so we lose ll_d_ops.
 858          * To workaround this, call ll_setup_filename() and do the rest
 859          * manually. Also make a copy of llcrypt_d_revalidate() (unfortunately
 860          * not exported function) and call it from ll_revalidate_dentry(), to
 861          * ensure we do not cache stale dentries after a key has been added.
 862          */
 863         rc = ll_setup_filename(parent, &dentry->d_name, 1, &fname, &fid);
 864         if ((!rc || rc == -ENOENT) && fname.is_ciphertext_name) {
 865                 spin_lock(&dentry->d_lock);
 866                 dentry->d_flags |= DCACHE_ENCRYPTED_NAME;
 867                 spin_unlock(&dentry->d_lock);
 868         }
 869         if (rc == -ENOENT)
 870                 RETURN(NULL);
 871         if (rc)
 872                 RETURN(ERR_PTR(rc));
 873
 874         op_data = ll_prep_md_op_data(NULL, parent, NULL, fname.disk_name.name,
 875                                      fname.disk_name.len, 0, opc, NULL);
 876         if (IS_ERR(op_data)) {
 877                 llcrypt_free_filename(&fname);
 878                 RETURN(ERR_CAST(op_data));
 879         }
 880         if (!fid_is_zero(&fid)) {
 881                 op_data->op_fid2 = fid;
 882                 op_data->op_bias = MDS_FID_OP;
 883                 if (it->it_op & IT_OPEN)
 884                         it->it_flags |= MDS_OPEN_BY_FID;
 885         }
 886
 887         /* enforce umask if acl disabled or MDS doesn't support umask */
 888         if (!IS_POSIXACL(parent) || !exp_connect_umask(ll_i2mdexp(parent)))
 889                 it->it_create_mode &= ~current_umask();
 890
 891         if (it->it_op & IT_CREAT &&
 892             ll_i2sbi(parent)->ll_flags & LL_SBI_FILE_SECCTX) {
 893                 rc = ll_dentry_init_security(dentry, it->it_create_mode,
 894                                              &dentry->d_name,
 895                                              &op_data->op_file_secctx_name,
 896                                              &op_data->op_file_secctx,
 897                                              &op_data->op_file_secctx_size);
 898                 if (rc < 0)
 899                         GOTO(out, retval = ERR_PTR(rc));
 900                 if (secctx != NULL)
 901                         *secctx = op_data->op_file_secctx;
 902                 if (secctxlen != NULL)
 903                         *secctxlen = op_data->op_file_secctx_size;
 904         } else {
 905                 if (secctx != NULL)
 906                         *secctx = NULL;
 907                 if (secctxlen != NULL)
 908                         *secctxlen = 0;
 909         }
 910         if (it->it_op & IT_CREAT && encrypt) {
 911                 /* Volatile file name may look like:
 912                  * <parent>/LUSTRE_VOLATILE_HDR:<mdt_index>:<random>:fd=<fd>
 913                  * where fd is opened descriptor of reference file.
 914                  */
 915                 if (unlikely(filename_is_volatile(dentry->d_name.name,
 916                                                   dentry->d_name.len, NULL))) {
 917                         int ctx_size = LLCRYPT_ENC_CTX_SIZE;
 918                         struct lustre_sb_info *lsi;
 919                         struct file *ref_file;
 920                         struct inode *ref_inode;
 921                         char *p, *q, *fd_str;
 922                         void *ctx;
 923                         int fd;
 924
 925                         p = strnstr(dentry->d_name.name, ":fd=",
 926                                     dentry->d_name.len);
 927                         if (!p || strlen(p + 4) == 0)
 928                                 GOTO(out, retval = ERR_PTR(-EINVAL));
 929
 930                         q = strchrnul(p + 4, ':');
 931                         fd_str = kstrndup(p + 4, q - p - 4, GFP_NOFS);
 932                         if (!fd_str)
 933                                 GOTO(out, retval = ERR_PTR(-ENOMEM));
 934                         rc = kstrtouint(fd_str, 10, &fd);
 935                         kfree(fd_str);
 936                         if (rc)
 937                                 GOTO(inherit, rc = -EINVAL);
 938
 939                         ref_file = fget(fd);
 940                         if (!ref_file)
 941                                 GOTO(inherit, rc = -EINVAL);
 942
 943                         ref_inode = file_inode(ref_file);
 944                         if (!ref_inode) {
 945                                 fput(ref_file);
 946                                 GOTO(inherit, rc = -EINVAL);
 947                         }
 948
 949                         lsi = s2lsi(ref_inode->i_sb);
 950
 951 getctx:
 952                         OBD_ALLOC(ctx, ctx_size);
 953                         if (!ctx)
 954                                 GOTO(out, retval = ERR_PTR(-ENOMEM));
 955
 956 #ifdef CONFIG_LL_ENCRYPTION
 957                         rc = lsi->lsi_cop->get_context(ref_inode,
 958                                                        ctx, ctx_size);
 959 #else
 960                         rc = -ENODATA;
 961 #endif
 962                         if (rc == -ERANGE) {
 963                                 OBD_FREE(ctx, ctx_size);
 964                                 ctx_size *= 2;
 965                                 goto getctx;
 966                         }
 967                         fput(ref_file);
 968                         if (rc < 0) {
 969                                 OBD_FREE(ctx, ctx_size);
 970                                 GOTO(inherit, rc);
 971                         }
 972
 973                         op_data->op_file_encctx_size = rc;
 974                         if (rc == ctx_size) {
 975                                 op_data->op_file_encctx = ctx;
 976                         } else {
 977                                 OBD_ALLOC(op_data->op_file_encctx,
 978                                           op_data->op_file_encctx_size);
 979                                 if (!op_data->op_file_encctx) {
 980                                         OBD_FREE(ctx, ctx_size);
 981                                         GOTO(out, retval = ERR_PTR(-ENOMEM));
 982                                 }
 983                                 memcpy(op_data->op_file_encctx, ctx,
 984                                        op_data->op_file_encctx_size);
 985                                 OBD_FREE(ctx, ctx_size);
 986                         }
 987
 988                 } else {
 989 inherit:
 990                         rc = llcrypt_inherit_context(parent, NULL, op_data,
 991                                                      false);
 992                         if (rc)
 993                                 GOTO(out, retval = ERR_PTR(rc));
 994                 }
 995                 if (encctx != NULL)
 996                         *encctx = op_data->op_file_encctx;
 997                 if (encctxlen != NULL)
 998                         *encctxlen = op_data->op_file_encctx_size;
 999         } else {
1000                 if (encctx != NULL)
1001                         *encctx = NULL;
1002                 if (encctxlen != NULL)
1003                         *encctxlen = 0;
1004         }
1005
1006         /* ask for security context upon intent */
1007         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_OPEN)) {
1008                 /* get name of security xattr to request to server */
1009                 rc = ll_listsecurity(parent, secctx_name,
1010                                      sizeof(secctx_name));
1011                 if (rc < 0) {
1012                         CDEBUG(D_SEC, "cannot get security xattr name for "
1013                                DFID": rc = %d\n",
1014                                PFID(ll_inode2fid(parent)), rc);
1015                 } else if (rc > 0) {
1016                         op_data->op_file_secctx_name = secctx_name;
1017                         op_data->op_file_secctx_name_size = rc;
1018                         CDEBUG(D_SEC, "'%.*s' is security xattr for "DFID"\n",
1019                                rc, secctx_name, PFID(ll_inode2fid(parent)));
1020                 }
1021         }
1022
1023         if (pca && pca->pca_dataset) {
1024                 OBD_ALLOC_PTR(lum);
1025                 if (lum == NULL)
1026                         GOTO(out, retval = ERR_PTR(-ENOMEM));
1027
1028                 lum->lmm_magic = LOV_USER_MAGIC_V1;
1029                 lum->lmm_pattern = LOV_PATTERN_F_RELEASED | LOV_PATTERN_RAID0;
1030                 op_data->op_data = lum;
1031                 op_data->op_data_size = sizeof(*lum);
1032                 op_data->op_archive_id = pca->pca_dataset->pccd_rwid;
1033                 it->it_flags |= MDS_OPEN_PCC;
1034         }
1035
1036         rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
1037                             &ll_md_blocking_ast, 0);
1038         /* If the MDS allows the client to chgrp (CFS_SETGRP_PERM), but the
1039          * client does not know which suppgid should be sent to the MDS, or
1040          * some other(s) changed the target file's GID after this RPC sent
1041          * to the MDS with the suppgid as the original GID, then we should
1042          * try again with right suppgid. */
1043         if (rc == -EACCES && it->it_op & IT_OPEN &&
1044             it_disposition(it, DISP_OPEN_DENY)) {
1045                 struct mdt_body *body;
1046
1047                 LASSERT(req != NULL);
1048
1049                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1050                 if (op_data->op_suppgids[0] == body->mbo_gid ||
1051                     op_data->op_suppgids[1] == body->mbo_gid ||
1052                     !in_group_p(make_kgid(&init_user_ns, body->mbo_gid)))
1053                         GOTO(out, retval = ERR_PTR(-EACCES));
1054
1055                 fid_zero(&op_data->op_fid2);
1056                 op_data->op_suppgids[1] = body->mbo_gid;
1057                 ptlrpc_req_finished(req);
1058                 req = NULL;
1059                 ll_intent_release(it);
1060                 rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
1061                                     &ll_md_blocking_ast, 0);
1062         }
1063
1064         if (rc < 0)
1065                 GOTO(out, retval = ERR_PTR(rc));
1066
1067         if (pca && pca->pca_dataset) {
1068                 rc = pcc_inode_create(parent->i_sb, pca->pca_dataset,
1069                                       &op_data->op_fid2,
1070                                       &pca->pca_dentry);
1071                 if (rc)
1072                         GOTO(out, retval = ERR_PTR(rc));
1073         }
1074
1075         /* dir layout may change */
1076         ll_unlock_md_op_lsm(op_data);
1077         rc = ll_lookup_it_finish(req, it, parent, &dentry,
1078                                  secctx != NULL ? *secctx : NULL,
1079                                  secctxlen != NULL ? *secctxlen : 0,
1080                                  encctx != NULL ? *encctx : NULL,
1081                                  encctxlen != NULL ? *encctxlen : 0,
1082                                  kstart, encrypt);
1083         if (rc != 0) {
1084                 ll_intent_release(it);
1085                 GOTO(out, retval = ERR_PTR(rc));
1086         }
1087
1088         if ((it->it_op & IT_OPEN) && dentry->d_inode &&
1089             !S_ISREG(dentry->d_inode->i_mode) &&
1090             !S_ISDIR(dentry->d_inode->i_mode)) {
1091                 ll_release_openhandle(dentry, it);
1092         }
1093         ll_lookup_finish_locks(it, dentry);
1094
1095         GOTO(out, retval = (dentry == save) ? NULL : dentry);
1096
1097 out:
1098         if (op_data != NULL && !IS_ERR(op_data)) {
1099                 if (secctx != NULL && secctxlen != NULL) {
1100                         /* caller needs sec ctx info, so reset it in op_data to
1101                          * prevent it from being freed */
1102                         op_data->op_file_secctx = NULL;
1103                         op_data->op_file_secctx_size = 0;
1104                 }
1105                 if (encctx != NULL && encctxlen != NULL &&
1106                     it->it_op & IT_CREAT && encrypt) {
1107                         /* caller needs enc ctx info, so reset it in op_data to
1108                          * prevent it from being freed
1109                          */
1110                         op_data->op_file_encctx = NULL;
1111                         op_data->op_file_encctx_size = 0;
1112                 }
1113                 llcrypt_free_filename(&fname);
1114                 ll_finish_md_op_data(op_data);
1115         }
1116
1117         if (lum != NULL)
1118                 OBD_FREE_PTR(lum);
1119
1120         ptlrpc_req_finished(req);
1121         return retval;
1122 }
1123
1124 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
1125                                    unsigned int flags)
1126 {
1127         struct lookup_intent *itp, it = { .it_op = IT_GETATTR };
1128         struct dentry *de;
1129
1130         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), flags=%u\n",
1131                dentry, PFID(ll_inode2fid(parent)), parent, flags);
1132
1133         /*
1134          * Optimize away (CREATE && !OPEN). Let .create handle the race.
1135          * but only if we have write permissions there, otherwise we need
1136          * to proceed with lookup. LU-4185
1137          */
1138         if ((flags & LOOKUP_CREATE) && !(flags & LOOKUP_OPEN) &&
1139             (inode_permission(parent, MAY_WRITE | MAY_EXEC) == 0))
1140                 return NULL;
1141
1142         if (flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE))
1143                 itp = NULL;
1144         else
1145                 itp = &it;
1146         de = ll_lookup_it(parent, dentry, itp, NULL, NULL, NULL, false,
1147                           NULL, NULL);
1148
1149         if (itp != NULL)
1150                 ll_intent_release(itp);
1151
1152         return de;
1153 }
1154
1155 #ifdef FMODE_CREATED /* added in Linux v4.18-rc1-20-g73a09dd */
1156 # define ll_is_opened(o, f)             ((f)->f_mode & FMODE_OPENED)
1157 # define ll_finish_open(f, d, o)        finish_open((f), (d), NULL)
1158 # define ll_last_arg
1159 # define ll_set_created(o, f)                                           \
1160 do {                                                                    \
1161         (f)->f_mode |= FMODE_CREATED;                                   \
1162 } while (0)
1163
1164 #else
1165 # define ll_is_opened(o, f)             (*(o))
1166 # define ll_finish_open(f, d, o)        finish_open((f), (d), NULL, (o))
1167 # define ll_last_arg                    , int *opened
1168 # define ll_set_created(o, f)                                           \
1169 do {                                                                    \
1170         *(o) |= FILE_CREATED;                                           \
1171 } while (0)
1172
1173 #endif
1174
1175 /*
1176  * For cached negative dentry and new dentry, handle lookup/create/open
1177  * together.
1178  */
1179 static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
1180                           struct file *file, unsigned open_flags,
1181                           umode_t mode ll_last_arg)
1182 {
1183         struct lookup_intent *it;
1184         struct dentry *de;
1185         long long lookup_flags = LOOKUP_OPEN;
1186         void *secctx = NULL;
1187         __u32 secctxlen = 0;
1188         void *encctx = NULL;
1189         __u32 encctxlen = 0;
1190         struct ll_sb_info *sbi = NULL;
1191         struct pcc_create_attach pca = { NULL, NULL };
1192         bool encrypt = false;
1193         int rc = 0;
1194         ENTRY;
1195
1196         CDEBUG(D_VFSTRACE,
1197                "VFS Op:name=%pd, dir="DFID"(%p), file %p, open_flags %x, mode %x opened %d\n",
1198                dentry, PFID(ll_inode2fid(dir)), dir, file, open_flags, mode,
1199                ll_is_opened(opened, file));
1200
1201         /* Only negative dentries enter here */
1202         LASSERT(dentry->d_inode == NULL);
1203
1204         if (!d_unhashed(dentry)) {
1205                 /* A valid negative dentry that just passed revalidation,
1206                  * there's little point to try and open it server-side,
1207                  * even though there's a minuscule chance it might succeed.
1208                  * Either way it's a valid race to just return -ENOENT here.
1209                  */
1210                 if (!(open_flags & O_CREAT))
1211                         return -ENOENT;
1212
1213                 /* Otherwise we just unhash it to be rehashed afresh via
1214                  * lookup if necessary
1215                  */
1216                 d_drop(dentry);
1217         }
1218
1219         OBD_ALLOC(it, sizeof(*it));
1220         if (!it)
1221                 RETURN(-ENOMEM);
1222
1223         it->it_op = IT_OPEN;
1224         if (open_flags & O_CREAT) {
1225                 it->it_op |= IT_CREAT;
1226                 lookup_flags |= LOOKUP_CREATE;
1227                 sbi = ll_i2sbi(dir);
1228                 /* Volatile file is used for HSM restore, so do not use PCC */
1229                 if (!filename_is_volatile(dentry->d_name.name,
1230                                           dentry->d_name.len, NULL)) {
1231                         struct pcc_matcher item;
1232                         struct pcc_dataset *dataset;
1233
1234                         item.pm_uid = from_kuid(&init_user_ns, current_uid());
1235                         item.pm_gid = from_kgid(&init_user_ns, current_gid());
1236                         item.pm_projid = ll_i2info(dir)->lli_projid;
1237                         item.pm_name = &dentry->d_name;
1238                         dataset = pcc_dataset_match_get(&sbi->ll_pcc_super,
1239                                                         &item);
1240                         pca.pca_dataset = dataset;
1241                 }
1242         }
1243         it->it_create_mode = (mode & S_IALLUGO) | S_IFREG;
1244         it->it_flags = (open_flags & ~O_ACCMODE) | OPEN_FMODE(open_flags);
1245         it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
1246
1247         if (ll_sbi_has_encrypt(ll_i2sbi(dir)) && IS_ENCRYPTED(dir)) {
1248                 /* in case of create, this is going to be a regular file because
1249                  * we set S_IFREG bit on it->it_create_mode above
1250                  */
1251                 rc = llcrypt_get_encryption_info(dir);
1252                 if (rc)
1253                         GOTO(out_release, rc);
1254                 if (open_flags & O_CREAT) {
1255                         if (!llcrypt_has_encryption_key(dir))
1256                                 GOTO(out_release, rc = -ENOKEY);
1257                         encrypt = true;
1258                 }
1259         }
1260
1261         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE2, cfs_fail_val);
1262
1263         /* We can only arrive at this path when we have no inode, so
1264          * we only need to request open lock if it was requested
1265          * for every open
1266          */
1267         if (ll_i2sbi(dir)->ll_oc_thrsh_count == 1 &&
1268             exp_connect_flags2(ll_i2mdexp(dir)) &
1269             OBD_CONNECT2_ATOMIC_OPEN_LOCK)
1270                 it->it_flags |= MDS_OPEN_LOCK;
1271
1272         /* Dentry added to dcache tree in ll_lookup_it */
1273         de = ll_lookup_it(dir, dentry, it, &secctx, &secctxlen, &pca, encrypt,
1274                           &encctx, &encctxlen);
1275         if (IS_ERR(de))
1276                 rc = PTR_ERR(de);
1277         else if (de != NULL)
1278                 dentry = de;
1279
1280         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1281
1282         if (!rc) {
1283                 if (it_disposition(it, DISP_OPEN_CREATE)) {
1284                         /* Dentry instantiated in ll_create_it. */
1285                         rc = ll_create_it(dir, dentry, it, secctx, secctxlen,
1286                                           encrypt, encctx, encctxlen);
1287                         ll_security_release_secctx(secctx, secctxlen);
1288                         llcrypt_free_ctx(encctx, encctxlen);
1289                         if (rc) {
1290                                 /* We dget in ll_splice_alias. */
1291                                 if (de != NULL)
1292                                         dput(de);
1293                                 goto out_release;
1294                         }
1295
1296                         rc = pcc_inode_create_fini(dentry->d_inode, &pca);
1297                         if (rc) {
1298                                 if (de != NULL)
1299                                         dput(de);
1300                                 GOTO(out_release, rc);
1301                         }
1302
1303                         ll_set_created(opened, file);
1304                 } else {
1305                         /* Open the file with O_CREAT, but the file already
1306                          * existed on MDT. This may happend in the case that
1307                          * the LOOKUP ibits lock is revoked and the
1308                          * corresponding dentry cache is deleted.
1309                          * i.e. In the current Lustre, the truncate operation
1310                          * will revoke the LOOKUP ibits lock, and the file
1311                          * dentry cache will be invalidated. The following open
1312                          * with O_CREAT flag will call into ->atomic_open, the
1313                          * file was wrongly though as newly created file and
1314                          * try to auto cache the file. So after client knows it
1315                          * is not a DISP_OPEN_CREATE, it should cleanup the
1316                          * already created PCC copy.
1317                          */
1318                         pcc_create_attach_cleanup(dir->i_sb, &pca);
1319
1320                         if (open_flags & O_CREAT && encrypt &&
1321                             dentry->d_inode) {
1322                                 rc = ll_set_encflags(dentry->d_inode, encctx,
1323                                                      encctxlen, true);
1324                                 llcrypt_free_ctx(encctx, encctxlen);
1325                                 if (rc)
1326                                         GOTO(out_release, rc);
1327                         }
1328                 }
1329
1330                 /* check also if a foreign file is openable */
1331                 if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN) &&
1332                     ll_foreign_is_openable(dentry, open_flags)) {
1333                         /* Open dentry. */
1334                         if (S_ISFIFO(dentry->d_inode->i_mode)) {
1335                                 /* We cannot call open here as it might
1336                                  * deadlock. This case is unreachable in
1337                                  * practice because of OBD_CONNECT_NODEVOH. */
1338                                 rc = finish_no_open(file, de);
1339                         } else {
1340                                 file->private_data = it;
1341                                 rc = ll_finish_open(file, dentry, opened);
1342                                 /* We dget in ll_splice_alias. finish_open takes
1343                                  * care of dget for fd open.
1344                                  */
1345                                 if (de != NULL)
1346                                         dput(de);
1347                         }
1348                 } else {
1349                         rc = finish_no_open(file, de);
1350                 }
1351         } else {
1352                 pcc_create_attach_cleanup(dir->i_sb, &pca);
1353         }
1354
1355 out_release:
1356         ll_intent_release(it);
1357         OBD_FREE(it, sizeof(*it));
1358
1359         RETURN(rc);
1360 }
1361
1362 /* We depend on "mode" being set with the proper file type/umask by now */
1363 static struct inode *ll_create_node(struct inode *dir, struct lookup_intent *it)
1364 {
1365         struct inode *inode = NULL;
1366         struct ptlrpc_request *request = NULL;
1367         struct ll_sb_info *sbi = ll_i2sbi(dir);
1368         int rc;
1369         ENTRY;
1370
1371         LASSERT(it && it->it_disposition);
1372
1373         LASSERT(it_disposition(it, DISP_ENQ_CREATE_REF));
1374         request = it->it_request;
1375         it_clear_disposition(it, DISP_ENQ_CREATE_REF);
1376         rc = ll_prep_inode(&inode, &request->rq_pill, dir->i_sb, it);
1377         if (rc)
1378                 GOTO(out, inode = ERR_PTR(rc));
1379
1380         /* Pause to allow for a race with concurrent access by fid */
1381         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_NODE_PAUSE, cfs_fail_val);
1382
1383         /* We asked for a lock on the directory, but were granted a
1384          * lock on the inode.  Since we finally have an inode pointer,
1385          * stuff it in the lock. */
1386         CDEBUG(D_DLMTRACE, "setting l_ast_data to inode "DFID"(%p)\n",
1387                PFID(ll_inode2fid(inode)), inode);
1388         ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
1389         EXIT;
1390  out:
1391         ptlrpc_req_finished(request);
1392         return inode;
1393 }
1394
1395 /*
1396  * By the time this is called, we already have created the directory cache
1397  * entry for the new file, but it is so far negative - it has no inode.
1398  *
1399  * We defer creating the OBD object(s) until open, to keep the intent and
1400  * non-intent code paths similar, and also because we do not have the MDS
1401  * inode number before calling ll_create_node() (which is needed for LOV),
1402  * so we would need to do yet another RPC to the MDS to store the LOV EA
1403  * data on the MDS.  If needed, we would pass the PACKED lmm as data and
1404  * lmm_size in datalen (the MDS still has code which will handle that).
1405  *
1406  * If the create succeeds, we fill in the inode information
1407  * with d_instantiate().
1408  */
1409 static int ll_create_it(struct inode *dir, struct dentry *dentry,
1410                         struct lookup_intent *it,
1411                         void *secctx, __u32 secctxlen, bool encrypt,
1412                         void *encctx, __u32 encctxlen)
1413 {
1414         struct inode *inode;
1415         __u64 bits = 0;
1416         int rc = 0;
1417         ENTRY;
1418
1419         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), intent=%s\n",
1420                dentry, PFID(ll_inode2fid(dir)), dir, LL_IT2STR(it));
1421
1422         rc = it_open_error(DISP_OPEN_CREATE, it);
1423         if (rc)
1424                 RETURN(rc);
1425
1426         inode = ll_create_node(dir, it);
1427         if (IS_ERR(inode))
1428                 RETURN(PTR_ERR(inode));
1429
1430         if ((ll_i2sbi(inode)->ll_flags & LL_SBI_FILE_SECCTX) &&
1431             secctx != NULL) {
1432                 /* must be done before d_instantiate, because it calls
1433                  * security_d_instantiate, which means a getxattr if security
1434                  * context is not set yet */
1435                 /* no need to protect selinux_inode_setsecurity() by
1436                  * inode_lock. Taking it would lead to a client deadlock
1437                  * LU-13617
1438                  */
1439                 rc = security_inode_notifysecctx(inode, secctx, secctxlen);
1440                 if (rc)
1441                         RETURN(rc);
1442         }
1443
1444         d_instantiate(dentry, inode);
1445
1446         if (encrypt) {
1447                 rc = ll_set_encflags(inode, encctx, encctxlen, true);
1448                 if (rc)
1449                         RETURN(rc);
1450         }
1451
1452         if (!(ll_i2sbi(inode)->ll_flags & LL_SBI_FILE_SECCTX)) {
1453                 rc = ll_inode_init_security(dentry, inode, dir);
1454                 if (rc)
1455                         RETURN(rc);
1456         }
1457
1458         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, inode, it, &bits);
1459         if (bits & MDS_INODELOCK_LOOKUP) {
1460                 d_lustre_revalidate(dentry);
1461                 ll_update_dir_depth(dir, inode);
1462         }
1463
1464         RETURN(0);
1465 }
1466
1467 void ll_update_times(struct ptlrpc_request *request, struct inode *inode)
1468 {
1469         struct mdt_body *body = req_capsule_server_get(&request->rq_pill,
1470                                                        &RMF_MDT_BODY);
1471
1472         LASSERT(body);
1473         if (body->mbo_valid & OBD_MD_FLMTIME &&
1474             body->mbo_mtime > inode->i_mtime.tv_sec) {
1475                 CDEBUG(D_INODE,
1476                        "setting fid " DFID " mtime from %lld to %llu\n",
1477                        PFID(ll_inode2fid(inode)),
1478                        (s64)inode->i_mtime.tv_sec, body->mbo_mtime);
1479                 inode->i_mtime.tv_sec = body->mbo_mtime;
1480         }
1481
1482         if (body->mbo_valid & OBD_MD_FLCTIME &&
1483             body->mbo_ctime > inode->i_ctime.tv_sec)
1484                 inode->i_ctime.tv_sec = body->mbo_ctime;
1485 }
1486
1487 /* once default LMV (space balanced) is set on ROOT, it should take effect if
1488  * default LMV is not set on parent directory.
1489  */
1490 static void ll_qos_mkdir_prep(struct md_op_data *op_data, struct inode *dir)
1491 {
1492         struct inode *root = dir->i_sb->s_root->d_inode;
1493         struct ll_inode_info *rlli = ll_i2info(root);
1494         struct ll_inode_info *lli = ll_i2info(dir);
1495         struct lmv_stripe_md *lsm;
1496
1497         op_data->op_dir_depth = lli->lli_depth;
1498
1499         /* parent directory is striped */
1500         if (unlikely(lli->lli_lsm_md))
1501                 return;
1502
1503         /* default LMV set on parent directory */
1504         if (unlikely(lli->lli_default_lsm_md))
1505                 return;
1506
1507         /* parent is ROOT */
1508         if (unlikely(dir == root))
1509                 return;
1510
1511         /* default LMV not set on ROOT */
1512         if (!rlli->lli_default_lsm_md)
1513                 return;
1514
1515         down_read(&rlli->lli_lsm_sem);
1516         lsm = rlli->lli_default_lsm_md;
1517         if (!lsm)
1518                 goto unlock;
1519
1520         /* not space balanced */
1521         if (lsm->lsm_md_master_mdt_index != LMV_OFFSET_DEFAULT)
1522                 goto unlock;
1523
1524         if (lsm->lsm_md_max_inherit != LMV_INHERIT_NONE &&
1525             (lsm->lsm_md_max_inherit == LMV_INHERIT_UNLIMITED ||
1526              lsm->lsm_md_max_inherit >= lli->lli_depth)) {
1527                 op_data->op_flags |= MF_QOS_MKDIR;
1528                 if (lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE &&
1529                     (lsm->lsm_md_max_inherit_rr == LMV_INHERIT_RR_UNLIMITED ||
1530                      lsm->lsm_md_max_inherit_rr >= lli->lli_depth))
1531                         op_data->op_flags |= MF_RR_MKDIR;
1532                 CDEBUG(D_INODE, DFID" requests qos mkdir %#x\n",
1533                        PFID(&lli->lli_fid), op_data->op_flags);
1534         }
1535 unlock:
1536         up_read(&rlli->lli_lsm_sem);
1537 }
1538
1539 static int ll_new_node(struct inode *dir, struct dentry *dchild,
1540                        const char *tgt, umode_t mode, __u64 rdev, __u32 opc)
1541 {
1542         struct qstr *name = &dchild->d_name;
1543         struct ptlrpc_request *request = NULL;
1544         struct md_op_data *op_data = NULL;
1545         struct inode *inode = NULL;
1546         struct ll_sb_info *sbi = ll_i2sbi(dir);
1547         struct llcrypt_str *disk_link = NULL;
1548         bool encrypt = false;
1549         int err;
1550
1551         ENTRY;
1552         if (unlikely(tgt != NULL)) {
1553                 disk_link = (struct llcrypt_str *)rdev;
1554                 rdev = 0;
1555                 if (!disk_link)
1556                         RETURN(-EINVAL);
1557         }
1558
1559 again:
1560         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name,
1561                                      name->len, 0, opc, NULL);
1562         if (IS_ERR(op_data))
1563                 GOTO(err_exit, err = PTR_ERR(op_data));
1564
1565         if (S_ISDIR(mode))
1566                 ll_qos_mkdir_prep(op_data, dir);
1567
1568         if (sbi->ll_flags & LL_SBI_FILE_SECCTX) {
1569                 err = ll_dentry_init_security(dchild, mode, &dchild->d_name,
1570                                               &op_data->op_file_secctx_name,
1571                                               &op_data->op_file_secctx,
1572                                               &op_data->op_file_secctx_size);
1573                 if (err < 0)
1574                         GOTO(err_exit, err);
1575         }
1576
1577         if (ll_sbi_has_encrypt(sbi) &&
1578             ((IS_ENCRYPTED(dir) &&
1579             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) ||
1580             (unlikely(llcrypt_dummy_context_enabled(dir)) && S_ISDIR(mode)))) {
1581                 err = llcrypt_get_encryption_info(dir);
1582                 if (err)
1583                         GOTO(err_exit, err);
1584                 if (!llcrypt_has_encryption_key(dir))
1585                         GOTO(err_exit, err = -ENOKEY);
1586                 encrypt = true;
1587         }
1588
1589         if (encrypt) {
1590                 err = llcrypt_inherit_context(dir, NULL, op_data, false);
1591                 if (err)
1592                         GOTO(err_exit, err);
1593
1594                 if (S_ISLNK(mode)) {
1595                         /* llcrypt needs inode to encrypt target name, so create
1596                          * a fake inode and associate encryption context got
1597                          * from llcrypt_inherit_context.
1598                          */
1599                         struct inode *fakeinode =
1600                                 dchild->d_sb->s_op->alloc_inode(dchild->d_sb);
1601
1602                         if (!fakeinode)
1603                                 GOTO(err_exit, err = -ENOMEM);
1604                         fakeinode->i_sb = dchild->d_sb;
1605                         fakeinode->i_mode |= S_IFLNK;
1606 #ifdef IOP_XATTR
1607                         fakeinode->i_opflags |= IOP_XATTR;
1608 #endif
1609                         err = ll_set_encflags(fakeinode,
1610                                               op_data->op_file_encctx,
1611                                               op_data->op_file_encctx_size,
1612                                               true);
1613                         if (!err)
1614                                 err = __llcrypt_encrypt_symlink(fakeinode, tgt,
1615                                                                 strlen(tgt),
1616                                                                 disk_link);
1617
1618                         ll_xattr_cache_destroy(fakeinode);
1619                         llcrypt_put_encryption_info(fakeinode);
1620                         dchild->d_sb->s_op->destroy_inode(fakeinode);
1621                         if (err)
1622                                 GOTO(err_exit, err);
1623                 }
1624         }
1625
1626         err = md_create(sbi->ll_md_exp, op_data, tgt ? disk_link->name : NULL,
1627                         tgt ? disk_link->len : 0, mode,
1628                         from_kuid(&init_user_ns, current_fsuid()),
1629                         from_kgid(&init_user_ns, current_fsgid()),
1630                         current_cap(), rdev, &request);
1631 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 14, 58, 0)
1632         /*
1633          * server < 2.12.58 doesn't pack default LMV in intent_getattr reply,
1634          * fetch default LMV here.
1635          */
1636         if (unlikely(err == -EREMOTE)) {
1637                 struct ll_inode_info    *lli = ll_i2info(dir);
1638                 struct lmv_user_md      *lum;
1639                 int                     lumsize;
1640                 int                     err2;
1641
1642                 ptlrpc_req_finished(request);
1643                 request = NULL;
1644
1645                 err2 = ll_dir_getstripe(dir, (void **)&lum, &lumsize, &request,
1646                                         OBD_MD_DEFAULT_MEA);
1647                 ll_finish_md_op_data(op_data);
1648                 op_data = NULL;
1649                 if (err2 == 0) {
1650                         struct lustre_md md = { NULL };
1651
1652                         md.body = req_capsule_server_get(&request->rq_pill,
1653                                                          &RMF_MDT_BODY);
1654                         if (!md.body)
1655                                 GOTO(err_exit, err = -EPROTO);
1656
1657                         OBD_ALLOC_PTR(md.default_lmv);
1658                         if (!md.default_lmv)
1659                                 GOTO(err_exit, err = -ENOMEM);
1660
1661                         md.default_lmv->lsm_md_magic = lum->lum_magic;
1662                         md.default_lmv->lsm_md_stripe_count =
1663                                 lum->lum_stripe_count;
1664                         md.default_lmv->lsm_md_master_mdt_index =
1665                                 lum->lum_stripe_offset;
1666                         md.default_lmv->lsm_md_hash_type = lum->lum_hash_type;
1667                         md.default_lmv->lsm_md_max_inherit =
1668                                 lum->lum_max_inherit;
1669                         md.default_lmv->lsm_md_max_inherit_rr =
1670                                 lum->lum_max_inherit_rr;
1671
1672                         err = ll_update_inode(dir, &md);
1673                         md_free_lustre_md(sbi->ll_md_exp, &md);
1674                         if (err)
1675                                 GOTO(err_exit, err);
1676                 } else if (err2 == -ENODATA && lli->lli_default_lsm_md) {
1677                         /*
1678                          * If there are no default stripe EA on the MDT, but the
1679                          * client has default stripe, then it probably means
1680                          * default stripe EA has just been deleted.
1681                          */
1682                         down_write(&lli->lli_lsm_sem);
1683                         if (lli->lli_default_lsm_md)
1684                                 OBD_FREE_PTR(lli->lli_default_lsm_md);
1685                         lli->lli_default_lsm_md = NULL;
1686                         up_write(&lli->lli_lsm_sem);
1687                 } else {
1688                         GOTO(err_exit, err);
1689                 }
1690
1691                 ptlrpc_req_finished(request);
1692                 request = NULL;
1693                 goto again;
1694         }
1695 #endif
1696
1697         if (err < 0)
1698                 GOTO(err_exit, err);
1699
1700         ll_update_times(request, dir);
1701
1702         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_NEWNODE_PAUSE, cfs_fail_val);
1703
1704         err = ll_prep_inode(&inode, &request->rq_pill, dchild->d_sb, NULL);
1705         if (err)
1706                 GOTO(err_exit, err);
1707
1708         if (sbi->ll_flags & LL_SBI_FILE_SECCTX) {
1709                 /* must be done before d_instantiate, because it calls
1710                  * security_d_instantiate, which means a getxattr if security
1711                  * context is not set yet */
1712                 /* no need to protect selinux_inode_setsecurity() by
1713                  * inode_lock. Taking it would lead to a client deadlock
1714                  * LU-13617
1715                  */
1716                 err = security_inode_notifysecctx(inode,
1717                                                   op_data->op_file_secctx,
1718                                                   op_data->op_file_secctx_size);
1719                 if (err)
1720                         GOTO(err_exit, err);
1721         }
1722
1723         d_instantiate(dchild, inode);
1724
1725         if (encrypt) {
1726                 err = ll_set_encflags(inode, op_data->op_file_encctx,
1727                                       op_data->op_file_encctx_size, true);
1728                 if (err)
1729                         GOTO(err_exit, err);
1730
1731                 if (S_ISLNK(mode)) {
1732                         struct ll_inode_info *lli = ll_i2info(inode);
1733
1734                         /* Cache the plaintext symlink target
1735                          * for later use by get_link()
1736                          */
1737                         OBD_ALLOC(lli->lli_symlink_name, strlen(tgt) + 1);
1738                         /* do not return an error if we cannot
1739                          * cache the symlink locally
1740                          */
1741                         if (lli->lli_symlink_name)
1742                                 memcpy(lli->lli_symlink_name,
1743                                        tgt, strlen(tgt) + 1);
1744                 }
1745         }
1746
1747         if (!(sbi->ll_flags & LL_SBI_FILE_SECCTX)) {
1748                 err = ll_inode_init_security(dchild, inode, dir);
1749                 if (err)
1750                         GOTO(err_exit, err);
1751         }
1752
1753         EXIT;
1754 err_exit:
1755         if (request != NULL)
1756                 ptlrpc_req_finished(request);
1757
1758         if (!IS_ERR_OR_NULL(op_data))
1759                 ll_finish_md_op_data(op_data);
1760
1761         RETURN(err);
1762 }
1763
1764 static int ll_mknod(struct inode *dir, struct dentry *dchild, umode_t mode,
1765                     dev_t rdev)
1766 {
1767         ktime_t kstart = ktime_get();
1768         int err;
1769         ENTRY;
1770
1771         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p) mode %o dev %x\n",
1772                dchild, PFID(ll_inode2fid(dir)), dir, mode, rdev);
1773
1774         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1775                 mode &= ~current_umask();
1776
1777         switch (mode & S_IFMT) {
1778         case 0:
1779                 mode |= S_IFREG;
1780                 /* fallthrough */
1781         case S_IFREG:
1782         case S_IFCHR:
1783         case S_IFBLK:
1784         case S_IFIFO:
1785         case S_IFSOCK:
1786                 err = ll_new_node(dir, dchild, NULL, mode, old_encode_dev(rdev),
1787                                   LUSTRE_OPC_MKNOD);
1788                 break;
1789         case S_IFDIR:
1790                 err = -EPERM;
1791                 break;
1792         default:
1793                 err = -EINVAL;
1794         }
1795
1796         if (!err)
1797                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKNOD,
1798                                    ktime_us_delta(ktime_get(), kstart));
1799
1800         RETURN(err);
1801 }
1802
1803 /*
1804  * Plain create. Intent create is handled in atomic_open.
1805  */
1806 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
1807                         umode_t mode, bool want_excl)
1808 {
1809         ktime_t kstart = ktime_get();
1810         int rc;
1811
1812         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1813
1814         CDEBUG(D_VFSTRACE,
1815                "VFS Op:name=%pd, dir="DFID"(%p), flags=%u, excl=%d\n",
1816                dentry, PFID(ll_inode2fid(dir)), dir, mode, want_excl);
1817
1818         /* Using mknod(2) to create a regular file is designed to not recognize
1819          * volatile file name, so we use ll_mknod() here. */
1820         rc = ll_mknod(dir, dentry, mode, 0);
1821
1822         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, unhashed %d\n",
1823                dentry, d_unhashed(dentry));
1824
1825         if (!rc)
1826                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE,
1827                                    ktime_us_delta(ktime_get(), kstart));
1828
1829         return rc;
1830 }
1831
1832 static int ll_symlink(struct inode *dir, struct dentry *dchild,
1833                       const char *oldpath)
1834 {
1835         ktime_t kstart = ktime_get();
1836         int len = strlen(oldpath);
1837         struct llcrypt_str disk_link;
1838         int err;
1839         ENTRY;
1840
1841         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), target=%.*s\n",
1842                dchild, PFID(ll_inode2fid(dir)), dir, 3000, oldpath);
1843
1844         err = llcrypt_prepare_symlink(dir, oldpath, len, dir->i_sb->s_blocksize,
1845                                       &disk_link);
1846         if (err)
1847                 RETURN(err);
1848
1849         err = ll_new_node(dir, dchild, oldpath, S_IFLNK | S_IRWXUGO,
1850                           (__u64)&disk_link, LUSTRE_OPC_SYMLINK);
1851
1852         if (disk_link.name != (unsigned char *)oldpath)
1853                 kfree(disk_link.name);
1854
1855         if (!err)
1856                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_SYMLINK,
1857                                    ktime_us_delta(ktime_get(), kstart));
1858
1859         RETURN(err);
1860 }
1861
1862 static int ll_link(struct dentry *old_dentry, struct inode *dir,
1863                    struct dentry *new_dentry)
1864 {
1865         struct inode *src = old_dentry->d_inode;
1866         struct qstr *name = &new_dentry->d_name;
1867         struct ll_sb_info *sbi = ll_i2sbi(dir);
1868         struct ptlrpc_request *request = NULL;
1869         struct md_op_data *op_data;
1870         ktime_t kstart = ktime_get();
1871         int err;
1872
1873         ENTRY;
1874         CDEBUG(D_VFSTRACE,
1875                "VFS Op: inode="DFID"(%p), dir="DFID"(%p), target=%pd\n",
1876                PFID(ll_inode2fid(src)), src,
1877                PFID(ll_inode2fid(dir)), dir, new_dentry);
1878
1879         err = llcrypt_prepare_link(old_dentry, dir, new_dentry);
1880         if (err)
1881                 RETURN(err);
1882
1883         op_data = ll_prep_md_op_data(NULL, src, dir, name->name, name->len,
1884                                      0, LUSTRE_OPC_ANY, NULL);
1885         if (IS_ERR(op_data))
1886                 RETURN(PTR_ERR(op_data));
1887
1888         err = md_link(sbi->ll_md_exp, op_data, &request);
1889         ll_finish_md_op_data(op_data);
1890         if (err)
1891                 GOTO(out, err);
1892
1893         ll_update_times(request, dir);
1894         ll_stats_ops_tally(sbi, LPROC_LL_LINK,
1895                            ktime_us_delta(ktime_get(), kstart));
1896         EXIT;
1897 out:
1898         ptlrpc_req_finished(request);
1899         RETURN(err);
1900 }
1901
1902 static int ll_mkdir(struct inode *dir, struct dentry *dchild, umode_t mode)
1903 {
1904         ktime_t kstart = ktime_get();
1905         int err;
1906         ENTRY;
1907
1908         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p)\n",
1909                dchild, PFID(ll_inode2fid(dir)), dir);
1910
1911         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1912                 mode &= ~current_umask();
1913
1914         mode = (mode & (S_IRWXUGO|S_ISVTX)) | S_IFDIR;
1915
1916         err = ll_new_node(dir, dchild, NULL, mode, 0, LUSTRE_OPC_MKDIR);
1917         if (err == 0)
1918                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKDIR,
1919                                    ktime_us_delta(ktime_get(), kstart));
1920
1921         RETURN(err);
1922 }
1923
1924 static int ll_rmdir(struct inode *dir, struct dentry *dchild)
1925 {
1926         struct qstr *name = &dchild->d_name;
1927         struct ptlrpc_request *request = NULL;
1928         struct md_op_data *op_data;
1929         ktime_t kstart = ktime_get();
1930         int rc;
1931
1932         ENTRY;
1933
1934         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p)\n",
1935                dchild, PFID(ll_inode2fid(dir)), dir);
1936
1937         if (unlikely(d_mountpoint(dchild)))
1938                 RETURN(-EBUSY);
1939
1940         /* some foreign dir may not be allowed to be removed */
1941         if (!ll_foreign_is_removable(dchild, false))
1942                 RETURN(-EPERM);
1943
1944         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len,
1945                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1946         if (IS_ERR(op_data))
1947                 RETURN(PTR_ERR(op_data));
1948
1949         if (dchild->d_inode != NULL)
1950                 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1951
1952         if (fid_is_zero(&op_data->op_fid2))
1953                 op_data->op_fid2 = op_data->op_fid3;
1954         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1955         ll_finish_md_op_data(op_data);
1956         if (!rc) {
1957                 struct mdt_body *body;
1958
1959                 ll_update_times(request, dir);
1960                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR,
1961                                    ktime_us_delta(ktime_get(), kstart));
1962
1963                 /*
1964                  * The server puts attributes in on the last unlink, use them
1965                  * to update the link count so the inode can be freed
1966                  * immediately.
1967                  */
1968                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
1969                 if (body->mbo_valid & OBD_MD_FLNLINK)
1970                         set_nlink(dchild->d_inode, body->mbo_nlink);
1971         }
1972
1973         ptlrpc_req_finished(request);
1974
1975         RETURN(rc);
1976 }
1977
1978 /**
1979  * Remove dir entry
1980  **/
1981 int ll_rmdir_entry(struct inode *dir, char *name, int namelen)
1982 {
1983         struct ptlrpc_request *request = NULL;
1984         struct md_op_data *op_data;
1985         ktime_t kstart = ktime_get();
1986         int rc;
1987         ENTRY;
1988
1989         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1990                namelen, name, PFID(ll_inode2fid(dir)), dir);
1991
1992         op_data = ll_prep_md_op_data(NULL, dir, NULL, name, strlen(name),
1993                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1994         if (IS_ERR(op_data))
1995                 RETURN(PTR_ERR(op_data));
1996         op_data->op_cli_flags |= CLI_RM_ENTRY;
1997         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1998         ll_finish_md_op_data(op_data);
1999         if (!rc)
2000                 ll_update_times(request, dir);
2001
2002         ptlrpc_req_finished(request);
2003         if (!rc)
2004                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR,
2005                                    ktime_us_delta(ktime_get(), kstart));
2006         RETURN(rc);
2007 }
2008
2009 static int ll_unlink(struct inode *dir, struct dentry *dchild)
2010 {
2011         struct qstr *name = &dchild->d_name;
2012         struct ptlrpc_request *request = NULL;
2013         struct md_op_data *op_data;
2014         struct mdt_body *body;
2015         ktime_t kstart = ktime_get();
2016         int rc;
2017
2018         ENTRY;
2019
2020         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p)\n",
2021                dchild, PFID(ll_inode2fid(dir)), dir);
2022
2023         /*
2024          * XXX: unlink bind mountpoint maybe call to here,
2025          * just check it as vfs_unlink does.
2026          */
2027         if (unlikely(d_mountpoint(dchild)))
2028                 RETURN(-EBUSY);
2029
2030         /* some foreign file/dir may not be allowed to be unlinked */
2031         if (!ll_foreign_is_removable(dchild, false))
2032                 RETURN(-EPERM);
2033
2034         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len, 0,
2035                                      LUSTRE_OPC_ANY, NULL);
2036         if (IS_ERR(op_data))
2037                 RETURN(PTR_ERR(op_data));
2038
2039         op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
2040         /* notify lower layer if inode has dirty pages */
2041         if (S_ISREG(dchild->d_inode->i_mode) &&
2042             ll_i2info(dchild->d_inode)->lli_clob &&
2043             dirty_cnt(dchild->d_inode))
2044                 op_data->op_cli_flags |= CLI_DIRTY_DATA;
2045         if (fid_is_zero(&op_data->op_fid2))
2046                 op_data->op_fid2 = op_data->op_fid3;
2047         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
2048         ll_finish_md_op_data(op_data);
2049         if (rc)
2050                 GOTO(out, rc);
2051
2052         /*
2053          * The server puts attributes in on the last unlink, use them to update
2054          * the link count so the inode can be freed immediately.
2055          */
2056         body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
2057         if (body->mbo_valid & OBD_MD_FLNLINK)
2058                 set_nlink(dchild->d_inode, body->mbo_nlink);
2059
2060         ll_update_times(request, dir);
2061
2062 out:
2063         ptlrpc_req_finished(request);
2064         if (!rc)
2065                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_UNLINK,
2066                                    ktime_us_delta(ktime_get(), kstart));
2067         RETURN(rc);
2068 }
2069
2070 static int ll_rename(struct inode *src, struct dentry *src_dchild,
2071                      struct inode *tgt, struct dentry *tgt_dchild
2072 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
2073                      , unsigned int flags
2074 #endif
2075                      )
2076 {
2077         struct ptlrpc_request *request = NULL;
2078         struct ll_sb_info *sbi = ll_i2sbi(src);
2079         struct md_op_data *op_data;
2080         ktime_t kstart = ktime_get();
2081         umode_t mode = 0;
2082         struct llcrypt_name foldname, fnewname;
2083         int err;
2084         ENTRY;
2085
2086 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
2087         if (flags)
2088                 return -EINVAL;
2089 #endif
2090
2091         CDEBUG(D_VFSTRACE,
2092                "VFS Op:oldname=%pd, src_dir="DFID"(%p), newname=%pd, tgt_dir="DFID"(%p)\n",
2093                src_dchild, PFID(ll_inode2fid(src)), src,
2094                tgt_dchild, PFID(ll_inode2fid(tgt)), tgt);
2095
2096         if (unlikely(d_mountpoint(src_dchild) || d_mountpoint(tgt_dchild)))
2097                 RETURN(-EBUSY);
2098
2099 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
2100         err = llcrypt_prepare_rename(src, src_dchild, tgt, tgt_dchild, flags);
2101 #else
2102         err = llcrypt_prepare_rename(src, src_dchild, tgt, tgt_dchild, 0);
2103 #endif
2104         if (err)
2105                 RETURN(err);
2106         /* we prevent an encrypted file from being renamed
2107          * into an unencrypted dir
2108          */
2109         if (IS_ENCRYPTED(src) && !IS_ENCRYPTED(tgt))
2110                 RETURN(-EXDEV);
2111
2112         if (src_dchild->d_inode)
2113                 mode = src_dchild->d_inode->i_mode;
2114
2115         if (tgt_dchild->d_inode)
2116                 mode = tgt_dchild->d_inode->i_mode;
2117
2118         op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, mode,
2119                                      LUSTRE_OPC_ANY, NULL);
2120         if (IS_ERR(op_data))
2121                 RETURN(PTR_ERR(op_data));
2122
2123         if (src_dchild->d_inode)
2124                 op_data->op_fid3 = *ll_inode2fid(src_dchild->d_inode);
2125
2126         if (tgt_dchild->d_inode)
2127                 op_data->op_fid4 = *ll_inode2fid(tgt_dchild->d_inode);
2128
2129         err = ll_setup_filename(src, &src_dchild->d_name, 1, &foldname, NULL);
2130         if (err)
2131                 RETURN(err);
2132         err = ll_setup_filename(tgt, &tgt_dchild->d_name, 1, &fnewname, NULL);
2133         if (err) {
2134                 llcrypt_free_filename(&foldname);
2135                 RETURN(err);
2136         }
2137         err = md_rename(sbi->ll_md_exp, op_data,
2138                         foldname.disk_name.name, foldname.disk_name.len,
2139                         fnewname.disk_name.name, fnewname.disk_name.len,
2140                         &request);
2141         llcrypt_free_filename(&foldname);
2142         llcrypt_free_filename(&fnewname);
2143         ll_finish_md_op_data(op_data);
2144         if (!err) {
2145                 ll_update_times(request, src);
2146                 ll_update_times(request, tgt);
2147         }
2148
2149         ptlrpc_req_finished(request);
2150
2151         if (!err) {
2152                 d_move(src_dchild, tgt_dchild);
2153                 ll_stats_ops_tally(sbi, LPROC_LL_RENAME,
2154                                    ktime_us_delta(ktime_get(), kstart));
2155         }
2156
2157         RETURN(err);
2158 }
2159
2160 const struct inode_operations ll_dir_inode_operations = {
2161         .mknod          = ll_mknod,
2162         .atomic_open    = ll_atomic_open,
2163         .lookup         = ll_lookup_nd,
2164         .create         = ll_create_nd,
2165         /* We need all these non-raw things for NFSD, to not patch it. */
2166         .unlink         = ll_unlink,
2167         .mkdir          = ll_mkdir,
2168         .rmdir          = ll_rmdir,
2169         .symlink        = ll_symlink,
2170         .link           = ll_link,
2171         .rename         = ll_rename,
2172         .setattr        = ll_setattr,
2173         .getattr        = ll_getattr,
2174         .permission     = ll_inode_permission,
2175 #ifdef HAVE_IOP_XATTR
2176         .setxattr       = ll_setxattr,
2177         .getxattr       = ll_getxattr,
2178         .removexattr    = ll_removexattr,
2179 #endif
2180         .listxattr      = ll_listxattr,
2181         .get_acl        = ll_get_acl,
2182 #ifdef HAVE_IOP_SET_ACL
2183         .set_acl        = ll_set_acl,
2184 #endif
2185 };
2186
2187 const struct inode_operations ll_special_inode_operations = {
2188         .setattr        = ll_setattr,
2189         .getattr        = ll_getattr,
2190         .permission     = ll_inode_permission,
2191 #ifdef HAVE_IOP_XATTR
2192         .setxattr       = ll_setxattr,
2193         .getxattr       = ll_getxattr,
2194         .removexattr    = ll_removexattr,
2195 #endif
2196         .listxattr      = ll_listxattr,
2197         .get_acl        = ll_get_acl,
2198 #ifdef HAVE_IOP_SET_ACL
2199         .set_acl        = ll_set_acl,
2200 #endif
2201 };