Whamcloud - gitweb
LU-15216 lmv: improve MDT QOS space balance
[fs/lustre-release.git] / 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, unsigned int open_flags);
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_empty(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             test_bit(LL_SBI_FILE_SECCTX, ll_i2sbi(parent)->ll_flags)) {
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                 if (unlikely(filename_is_volatile(dentry->d_name.name,
912                                                   dentry->d_name.len, NULL))) {
913                         /* get encryption context from reference file */
914                         int ctx_size = LLCRYPT_ENC_CTX_SIZE;
915                         struct lustre_sb_info *lsi;
916                         struct file *ref_file;
917                         struct inode *ref_inode;
918                         void *ctx;
919
920                         rc = volatile_ref_file(dentry->d_name.name,
921                                                dentry->d_name.len,
922                                                &ref_file);
923                         if (rc)
924                                 GOTO(out, retval = ERR_PTR(rc));
925
926                         ref_inode = file_inode(ref_file);
927                         if (!ref_inode) {
928                                 fput(ref_file);
929                                 GOTO(inherit, rc = -EINVAL);
930                         }
931
932                         lsi = s2lsi(ref_inode->i_sb);
933
934 getctx:
935                         OBD_ALLOC(ctx, ctx_size);
936                         if (!ctx)
937                                 GOTO(out, retval = ERR_PTR(-ENOMEM));
938
939 #ifdef CONFIG_LL_ENCRYPTION
940                         rc = lsi->lsi_cop->get_context(ref_inode,
941                                                        ctx, ctx_size);
942 #else
943                         rc = -ENODATA;
944 #endif
945                         if (rc == -ERANGE) {
946                                 OBD_FREE(ctx, ctx_size);
947                                 ctx_size *= 2;
948                                 goto getctx;
949                         }
950                         fput(ref_file);
951                         if (rc < 0) {
952                                 OBD_FREE(ctx, ctx_size);
953                                 GOTO(inherit, rc);
954                         }
955
956                         op_data->op_file_encctx_size = rc;
957                         if (rc == ctx_size) {
958                                 op_data->op_file_encctx = ctx;
959                         } else {
960                                 OBD_ALLOC(op_data->op_file_encctx,
961                                           op_data->op_file_encctx_size);
962                                 if (!op_data->op_file_encctx) {
963                                         OBD_FREE(ctx, ctx_size);
964                                         GOTO(out, retval = ERR_PTR(-ENOMEM));
965                                 }
966                                 memcpy(op_data->op_file_encctx, ctx,
967                                        op_data->op_file_encctx_size);
968                                 OBD_FREE(ctx, ctx_size);
969                         }
970                 } else {
971 inherit:
972                         rc = llcrypt_inherit_context(parent, NULL, op_data,
973                                                      false);
974                         if (rc)
975                                 GOTO(out, retval = ERR_PTR(rc));
976                 }
977                 if (encctx != NULL)
978                         *encctx = op_data->op_file_encctx;
979                 if (encctxlen != NULL)
980                         *encctxlen = op_data->op_file_encctx_size;
981         } else {
982                 if (encctx != NULL)
983                         *encctx = NULL;
984                 if (encctxlen != NULL)
985                         *encctxlen = 0;
986         }
987
988         /* ask for security context upon intent */
989         if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_OPEN)) {
990                 /* get name of security xattr to request to server */
991                 rc = ll_listsecurity(parent, secctx_name,
992                                      sizeof(secctx_name));
993                 if (rc < 0) {
994                         CDEBUG(D_SEC, "cannot get security xattr name for "
995                                DFID": rc = %d\n",
996                                PFID(ll_inode2fid(parent)), rc);
997                 } else if (rc > 0) {
998                         op_data->op_file_secctx_name = secctx_name;
999                         op_data->op_file_secctx_name_size = rc;
1000                         CDEBUG(D_SEC, "'%.*s' is security xattr for "DFID"\n",
1001                                rc, secctx_name, PFID(ll_inode2fid(parent)));
1002                 }
1003         }
1004
1005         if (pca && pca->pca_dataset) {
1006                 OBD_ALLOC_PTR(lum);
1007                 if (lum == NULL)
1008                         GOTO(out, retval = ERR_PTR(-ENOMEM));
1009
1010                 lum->lmm_magic = LOV_USER_MAGIC_V1;
1011                 lum->lmm_pattern = LOV_PATTERN_F_RELEASED | LOV_PATTERN_RAID0;
1012                 op_data->op_data = lum;
1013                 op_data->op_data_size = sizeof(*lum);
1014                 op_data->op_archive_id = pca->pca_dataset->pccd_rwid;
1015                 it->it_flags |= MDS_OPEN_PCC;
1016         }
1017
1018         rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
1019                             &ll_md_blocking_ast, 0);
1020         /* If the MDS allows the client to chgrp (CFS_SETGRP_PERM), but the
1021          * client does not know which suppgid should be sent to the MDS, or
1022          * some other(s) changed the target file's GID after this RPC sent
1023          * to the MDS with the suppgid as the original GID, then we should
1024          * try again with right suppgid. */
1025         if (rc == -EACCES && it->it_op & IT_OPEN &&
1026             it_disposition(it, DISP_OPEN_DENY)) {
1027                 struct mdt_body *body;
1028
1029                 LASSERT(req != NULL);
1030
1031                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1032                 if (op_data->op_suppgids[0] == body->mbo_gid ||
1033                     op_data->op_suppgids[1] == body->mbo_gid ||
1034                     !in_group_p(make_kgid(&init_user_ns, body->mbo_gid)))
1035                         GOTO(out, retval = ERR_PTR(-EACCES));
1036
1037                 fid_zero(&op_data->op_fid2);
1038                 op_data->op_suppgids[1] = body->mbo_gid;
1039                 ptlrpc_req_finished(req);
1040                 req = NULL;
1041                 ll_intent_release(it);
1042                 rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
1043                                     &ll_md_blocking_ast, 0);
1044         }
1045
1046         if (rc < 0)
1047                 GOTO(out, retval = ERR_PTR(rc));
1048
1049         if (pca && pca->pca_dataset) {
1050                 rc = pcc_inode_create(parent->i_sb, pca->pca_dataset,
1051                                       &op_data->op_fid2,
1052                                       &pca->pca_dentry);
1053                 if (rc)
1054                         GOTO(out, retval = ERR_PTR(rc));
1055         }
1056
1057         /* dir layout may change */
1058         ll_unlock_md_op_lsm(op_data);
1059         rc = ll_lookup_it_finish(req, it, parent, &dentry,
1060                                  secctx != NULL ? *secctx : NULL,
1061                                  secctxlen != NULL ? *secctxlen : 0,
1062                                  encctx != NULL ? *encctx : NULL,
1063                                  encctxlen != NULL ? *encctxlen : 0,
1064                                  kstart, encrypt);
1065         if (rc != 0) {
1066                 ll_intent_release(it);
1067                 GOTO(out, retval = ERR_PTR(rc));
1068         }
1069
1070         if ((it->it_op & IT_OPEN) && dentry->d_inode &&
1071             !S_ISREG(dentry->d_inode->i_mode) &&
1072             !S_ISDIR(dentry->d_inode->i_mode)) {
1073                 ll_release_openhandle(dentry, it);
1074         }
1075         ll_lookup_finish_locks(it, dentry);
1076
1077         GOTO(out, retval = (dentry == save) ? NULL : dentry);
1078
1079 out:
1080         if (op_data != NULL && !IS_ERR(op_data)) {
1081                 if (secctx != NULL && secctxlen != NULL) {
1082                         /* caller needs sec ctx info, so reset it in op_data to
1083                          * prevent it from being freed */
1084                         op_data->op_file_secctx = NULL;
1085                         op_data->op_file_secctx_size = 0;
1086                 }
1087                 if (encctx != NULL && encctxlen != NULL &&
1088                     it->it_op & IT_CREAT && encrypt) {
1089                         /* caller needs enc ctx info, so reset it in op_data to
1090                          * prevent it from being freed
1091                          */
1092                         op_data->op_file_encctx = NULL;
1093                         op_data->op_file_encctx_size = 0;
1094                 }
1095                 llcrypt_free_filename(&fname);
1096                 ll_finish_md_op_data(op_data);
1097         }
1098
1099         if (lum != NULL)
1100                 OBD_FREE_PTR(lum);
1101
1102         ptlrpc_req_finished(req);
1103         return retval;
1104 }
1105
1106 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
1107                                    unsigned int flags)
1108 {
1109         struct lookup_intent *itp, it = { .it_op = IT_GETATTR };
1110         struct dentry *de;
1111
1112         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), flags=%u\n",
1113                dentry, PFID(ll_inode2fid(parent)), parent, flags);
1114
1115         /*
1116          * Optimize away (CREATE && !OPEN). Let .create handle the race.
1117          * but only if we have write permissions there, otherwise we need
1118          * to proceed with lookup. LU-4185
1119          */
1120         if ((flags & LOOKUP_CREATE) && !(flags & LOOKUP_OPEN) &&
1121             (inode_permission(parent, MAY_WRITE | MAY_EXEC) == 0))
1122                 return NULL;
1123
1124         if (flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE))
1125                 itp = NULL;
1126         else
1127                 itp = &it;
1128         de = ll_lookup_it(parent, dentry, itp, NULL, NULL, NULL, false,
1129                           NULL, NULL);
1130
1131         if (itp != NULL)
1132                 ll_intent_release(itp);
1133
1134         return de;
1135 }
1136
1137 #ifdef FMODE_CREATED /* added in Linux v4.18-rc1-20-g73a09dd */
1138 # define ll_is_opened(o, f)             ((f)->f_mode & FMODE_OPENED)
1139 # define ll_finish_open(f, d, o)        finish_open((f), (d), NULL)
1140 # define ll_last_arg
1141 # define ll_set_created(o, f)                                           \
1142 do {                                                                    \
1143         (f)->f_mode |= FMODE_CREATED;                                   \
1144 } while (0)
1145
1146 #else
1147 # define ll_is_opened(o, f)             (*(o))
1148 # define ll_finish_open(f, d, o)        finish_open((f), (d), NULL, (o))
1149 # define ll_last_arg                    , int *opened
1150 # define ll_set_created(o, f)                                           \
1151 do {                                                                    \
1152         *(o) |= FILE_CREATED;                                           \
1153 } while (0)
1154
1155 #endif
1156
1157 /*
1158  * For cached negative dentry and new dentry, handle lookup/create/open
1159  * together.
1160  */
1161 static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
1162                           struct file *file, unsigned open_flags,
1163                           umode_t mode ll_last_arg)
1164 {
1165         struct lookup_intent *it;
1166         struct dentry *de;
1167         long long lookup_flags = LOOKUP_OPEN;
1168         void *secctx = NULL;
1169         __u32 secctxlen = 0;
1170         void *encctx = NULL;
1171         __u32 encctxlen = 0;
1172         struct ll_sb_info *sbi = NULL;
1173         struct pcc_create_attach pca = { NULL, NULL };
1174         bool encrypt = false;
1175         int rc = 0;
1176         ENTRY;
1177
1178         CDEBUG(D_VFSTRACE,
1179                "VFS Op:name=%pd, dir="DFID"(%p), file %p, open_flags %x, mode %x opened %d\n",
1180                dentry, PFID(ll_inode2fid(dir)), dir, file, open_flags, mode,
1181                ll_is_opened(opened, file));
1182
1183         /* Only negative dentries enter here */
1184         LASSERT(dentry->d_inode == NULL);
1185
1186         if (!d_unhashed(dentry)) {
1187                 /* A valid negative dentry that just passed revalidation,
1188                  * there's little point to try and open it server-side,
1189                  * even though there's a minuscule chance it might succeed.
1190                  * Either way it's a valid race to just return -ENOENT here.
1191                  */
1192                 if (!(open_flags & O_CREAT))
1193                         return -ENOENT;
1194
1195                 /* Otherwise we just unhash it to be rehashed afresh via
1196                  * lookup if necessary
1197                  */
1198                 d_drop(dentry);
1199         }
1200
1201         OBD_ALLOC(it, sizeof(*it));
1202         if (!it)
1203                 RETURN(-ENOMEM);
1204
1205         it->it_op = IT_OPEN;
1206         if (open_flags & O_CREAT) {
1207                 it->it_op |= IT_CREAT;
1208                 lookup_flags |= LOOKUP_CREATE;
1209                 sbi = ll_i2sbi(dir);
1210                 /* Volatile file is used for HSM restore, so do not use PCC */
1211                 if (!filename_is_volatile(dentry->d_name.name,
1212                                           dentry->d_name.len, NULL)) {
1213                         struct pcc_matcher item;
1214                         struct pcc_dataset *dataset;
1215
1216                         item.pm_uid = from_kuid(&init_user_ns, current_uid());
1217                         item.pm_gid = from_kgid(&init_user_ns, current_gid());
1218                         item.pm_projid = ll_i2info(dir)->lli_projid;
1219                         item.pm_name = &dentry->d_name;
1220                         dataset = pcc_dataset_match_get(&sbi->ll_pcc_super,
1221                                                         &item);
1222                         pca.pca_dataset = dataset;
1223                 }
1224         }
1225         it->it_create_mode = (mode & S_IALLUGO) | S_IFREG;
1226         it->it_flags = (open_flags & ~O_ACCMODE) | OPEN_FMODE(open_flags);
1227         it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
1228
1229         if (ll_sbi_has_encrypt(ll_i2sbi(dir)) && IS_ENCRYPTED(dir)) {
1230                 /* in case of create, this is going to be a regular file because
1231                  * we set S_IFREG bit on it->it_create_mode above
1232                  */
1233                 rc = llcrypt_get_encryption_info(dir);
1234                 if (rc)
1235                         GOTO(out_release, rc);
1236                 if (open_flags & O_CREAT) {
1237                         /* For migration or mirroring without enc key, we still
1238                          * need to be able to create a volatile file.
1239                          */
1240                         if (!llcrypt_has_encryption_key(dir) &&
1241                             (!filename_is_volatile(dentry->d_name.name,
1242                                                    dentry->d_name.len, NULL) ||
1243                             (open_flags & O_FILE_ENC) != O_FILE_ENC ||
1244                             !(open_flags & O_DIRECT)))
1245                                 GOTO(out_release, rc = -ENOKEY);
1246                         encrypt = true;
1247                 }
1248         }
1249
1250         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE2, cfs_fail_val);
1251
1252         /* We can only arrive at this path when we have no inode, so
1253          * we only need to request open lock if it was requested
1254          * for every open
1255          */
1256         if (ll_i2sbi(dir)->ll_oc_thrsh_count == 1 &&
1257             exp_connect_flags2(ll_i2mdexp(dir)) &
1258             OBD_CONNECT2_ATOMIC_OPEN_LOCK)
1259                 it->it_flags |= MDS_OPEN_LOCK;
1260
1261         /* Dentry added to dcache tree in ll_lookup_it */
1262         de = ll_lookup_it(dir, dentry, it, &secctx, &secctxlen, &pca, encrypt,
1263                           &encctx, &encctxlen);
1264         if (IS_ERR(de))
1265                 rc = PTR_ERR(de);
1266         else if (de != NULL)
1267                 dentry = de;
1268
1269         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1270
1271         if (!rc) {
1272                 if (it_disposition(it, DISP_OPEN_CREATE)) {
1273                         /* Dentry instantiated in ll_create_it. */
1274                         rc = ll_create_it(dir, dentry, it, secctx, secctxlen,
1275                                           encrypt, encctx, encctxlen,
1276                                           open_flags);
1277                         ll_security_release_secctx(secctx, secctxlen);
1278                         llcrypt_free_ctx(encctx, encctxlen);
1279                         if (rc) {
1280                                 /* We dget in ll_splice_alias. */
1281                                 if (de != NULL)
1282                                         dput(de);
1283                                 goto out_release;
1284                         }
1285
1286                         rc = pcc_inode_create_fini(dentry->d_inode, &pca);
1287                         if (rc) {
1288                                 if (de != NULL)
1289                                         dput(de);
1290                                 GOTO(out_release, rc);
1291                         }
1292
1293                         ll_set_created(opened, file);
1294                 } else {
1295                         /* Open the file with O_CREAT, but the file already
1296                          * existed on MDT. This may happend in the case that
1297                          * the LOOKUP ibits lock is revoked and the
1298                          * corresponding dentry cache is deleted.
1299                          * i.e. In the current Lustre, the truncate operation
1300                          * will revoke the LOOKUP ibits lock, and the file
1301                          * dentry cache will be invalidated. The following open
1302                          * with O_CREAT flag will call into ->atomic_open, the
1303                          * file was wrongly though as newly created file and
1304                          * try to auto cache the file. So after client knows it
1305                          * is not a DISP_OPEN_CREATE, it should cleanup the
1306                          * already created PCC copy.
1307                          */
1308                         pcc_create_attach_cleanup(dir->i_sb, &pca);
1309
1310                         if (open_flags & O_CREAT && encrypt &&
1311                             dentry->d_inode) {
1312                                 rc = ll_set_encflags(dentry->d_inode, encctx,
1313                                                      encctxlen, true);
1314                                 llcrypt_free_ctx(encctx, encctxlen);
1315                                 if (rc)
1316                                         GOTO(out_release, rc);
1317                         }
1318                 }
1319
1320                 /* check also if a foreign file is openable */
1321                 if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN) &&
1322                     ll_foreign_is_openable(dentry, open_flags)) {
1323                         /* Open dentry. */
1324                         if (S_ISFIFO(dentry->d_inode->i_mode)) {
1325                                 /* We cannot call open here as it might
1326                                  * deadlock. This case is unreachable in
1327                                  * practice because of OBD_CONNECT_NODEVOH. */
1328                                 rc = finish_no_open(file, de);
1329                         } else {
1330                                 file->private_data = it;
1331                                 rc = ll_finish_open(file, dentry, opened);
1332                                 /* We dget in ll_splice_alias. finish_open takes
1333                                  * care of dget for fd open.
1334                                  */
1335                                 if (de != NULL)
1336                                         dput(de);
1337                         }
1338                 } else {
1339                         rc = finish_no_open(file, de);
1340                 }
1341         } else {
1342                 pcc_create_attach_cleanup(dir->i_sb, &pca);
1343         }
1344
1345 out_release:
1346         ll_intent_release(it);
1347         OBD_FREE(it, sizeof(*it));
1348
1349         RETURN(rc);
1350 }
1351
1352 /* We depend on "mode" being set with the proper file type/umask by now */
1353 static struct inode *ll_create_node(struct inode *dir, struct lookup_intent *it)
1354 {
1355         struct inode *inode = NULL;
1356         struct ptlrpc_request *request = NULL;
1357         struct ll_sb_info *sbi = ll_i2sbi(dir);
1358         int rc;
1359         ENTRY;
1360
1361         LASSERT(it && it->it_disposition);
1362
1363         LASSERT(it_disposition(it, DISP_ENQ_CREATE_REF));
1364         request = it->it_request;
1365         it_clear_disposition(it, DISP_ENQ_CREATE_REF);
1366         rc = ll_prep_inode(&inode, &request->rq_pill, dir->i_sb, it);
1367         if (rc)
1368                 GOTO(out, inode = ERR_PTR(rc));
1369
1370         /* Pause to allow for a race with concurrent access by fid */
1371         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_NODE_PAUSE, cfs_fail_val);
1372
1373         /* We asked for a lock on the directory, but were granted a
1374          * lock on the inode.  Since we finally have an inode pointer,
1375          * stuff it in the lock. */
1376         CDEBUG(D_DLMTRACE, "setting l_ast_data to inode "DFID"(%p)\n",
1377                PFID(ll_inode2fid(inode)), inode);
1378         ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
1379         EXIT;
1380  out:
1381         ptlrpc_req_finished(request);
1382         return inode;
1383 }
1384
1385 /*
1386  * By the time this is called, we already have created the directory cache
1387  * entry for the new file, but it is so far negative - it has no inode.
1388  *
1389  * We defer creating the OBD object(s) until open, to keep the intent and
1390  * non-intent code paths similar, and also because we do not have the MDS
1391  * inode number before calling ll_create_node() (which is needed for LOV),
1392  * so we would need to do yet another RPC to the MDS to store the LOV EA
1393  * data on the MDS.  If needed, we would pass the PACKED lmm as data and
1394  * lmm_size in datalen (the MDS still has code which will handle that).
1395  *
1396  * If the create succeeds, we fill in the inode information
1397  * with d_instantiate().
1398  */
1399 static int ll_create_it(struct inode *dir, struct dentry *dentry,
1400                         struct lookup_intent *it,
1401                         void *secctx, __u32 secctxlen, bool encrypt,
1402                         void *encctx, __u32 encctxlen, unsigned int open_flags)
1403 {
1404         struct inode *inode;
1405         __u64 bits = 0;
1406         int rc = 0;
1407         ENTRY;
1408
1409         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), intent=%s\n",
1410                dentry, PFID(ll_inode2fid(dir)), dir, LL_IT2STR(it));
1411
1412         rc = it_open_error(DISP_OPEN_CREATE, it);
1413         if (rc)
1414                 RETURN(rc);
1415
1416         inode = ll_create_node(dir, it);
1417         if (IS_ERR(inode))
1418                 RETURN(PTR_ERR(inode));
1419
1420         if (test_bit(LL_SBI_FILE_SECCTX, ll_i2sbi(inode)->ll_flags) &&
1421             secctx) {
1422                 /* must be done before d_instantiate, because it calls
1423                  * security_d_instantiate, which means a getxattr if security
1424                  * context is not set yet */
1425                 /* no need to protect selinux_inode_setsecurity() by
1426                  * inode_lock. Taking it would lead to a client deadlock
1427                  * LU-13617
1428                  */
1429                 rc = security_inode_notifysecctx(inode, secctx, secctxlen);
1430                 if (rc)
1431                         RETURN(rc);
1432         }
1433
1434         d_instantiate(dentry, inode);
1435
1436         if (encrypt) {
1437                 bool preload = true;
1438
1439                 /* For migration or mirroring without enc key, we
1440                  * create a volatile file without enc context.
1441                  */
1442                 if (!llcrypt_has_encryption_key(dir) &&
1443                     filename_is_volatile(dentry->d_name.name,
1444                                          dentry->d_name.len, NULL) &&
1445                     (open_flags & O_FILE_ENC) == O_FILE_ENC &&
1446                     open_flags & O_DIRECT)
1447                         preload = false;
1448                 rc = ll_set_encflags(inode, encctx, encctxlen, preload);
1449                 if (rc)
1450                         RETURN(rc);
1451         }
1452
1453         if (!test_bit(LL_SBI_FILE_SECCTX, ll_i2sbi(inode)->ll_flags)) {
1454                 rc = ll_inode_init_security(dentry, inode, dir);
1455                 if (rc)
1456                         RETURN(rc);
1457         }
1458
1459         ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, inode, it, &bits);
1460         if (bits & MDS_INODELOCK_LOOKUP) {
1461                 d_lustre_revalidate(dentry);
1462                 ll_update_dir_depth(dir, inode);
1463         }
1464
1465         RETURN(0);
1466 }
1467
1468 void ll_update_times(struct ptlrpc_request *request, struct inode *inode)
1469 {
1470         struct mdt_body *body = req_capsule_server_get(&request->rq_pill,
1471                                                        &RMF_MDT_BODY);
1472
1473         LASSERT(body);
1474         if (body->mbo_valid & OBD_MD_FLMTIME &&
1475             body->mbo_mtime > inode->i_mtime.tv_sec) {
1476                 CDEBUG(D_INODE,
1477                        "setting fid " DFID " mtime from %lld to %llu\n",
1478                        PFID(ll_inode2fid(inode)),
1479                        (s64)inode->i_mtime.tv_sec, body->mbo_mtime);
1480                 inode->i_mtime.tv_sec = body->mbo_mtime;
1481         }
1482
1483         if (body->mbo_valid & OBD_MD_FLCTIME &&
1484             body->mbo_ctime > inode->i_ctime.tv_sec)
1485                 inode->i_ctime.tv_sec = body->mbo_ctime;
1486 }
1487
1488 /* once default LMV (space balanced) is set on ROOT, it should take effect if
1489  * default LMV is not set on parent directory.
1490  */
1491 static void ll_qos_mkdir_prep(struct md_op_data *op_data, struct inode *dir)
1492 {
1493         struct inode *root = dir->i_sb->s_root->d_inode;
1494         struct ll_inode_info *rlli = ll_i2info(root);
1495         struct ll_inode_info *lli = ll_i2info(dir);
1496         struct lmv_stripe_md *lsm;
1497
1498         op_data->op_dir_depth = lli->lli_dir_depth;
1499
1500         /* parent directory is striped */
1501         if (unlikely(lli->lli_lsm_md))
1502                 return;
1503
1504         /* default LMV set on parent directory */
1505         if (unlikely(lli->lli_default_lsm_md))
1506                 return;
1507
1508         /* parent is ROOT */
1509         if (unlikely(dir == root))
1510                 return;
1511
1512         /* default LMV not set on ROOT */
1513         if (!rlli->lli_default_lsm_md)
1514                 return;
1515
1516         down_read(&rlli->lli_lsm_sem);
1517         lsm = rlli->lli_default_lsm_md;
1518         if (!lsm)
1519                 goto unlock;
1520
1521         /* not space balanced */
1522         if (lsm->lsm_md_master_mdt_index != LMV_OFFSET_DEFAULT)
1523                 goto unlock;
1524
1525         if (lsm->lsm_md_max_inherit != LMV_INHERIT_NONE &&
1526             (lsm->lsm_md_max_inherit == LMV_INHERIT_UNLIMITED ||
1527              lsm->lsm_md_max_inherit >= lli->lli_dir_depth)) {
1528                 op_data->op_flags |= MF_QOS_MKDIR;
1529                 if (lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE &&
1530                     (lsm->lsm_md_max_inherit_rr == LMV_INHERIT_RR_UNLIMITED ||
1531                      lsm->lsm_md_max_inherit_rr >= lli->lli_dir_depth))
1532                         op_data->op_flags |= MF_RR_MKDIR;
1533                 CDEBUG(D_INODE, DFID" requests qos mkdir %#x\n",
1534                        PFID(&lli->lli_fid), op_data->op_flags);
1535         }
1536 unlock:
1537         up_read(&rlli->lli_lsm_sem);
1538 }
1539
1540 static int ll_new_node(struct inode *dir, struct dentry *dchild,
1541                        const char *tgt, umode_t mode, __u64 rdev, __u32 opc)
1542 {
1543         struct qstr *name = &dchild->d_name;
1544         struct ptlrpc_request *request = NULL;
1545         struct md_op_data *op_data = NULL;
1546         struct inode *inode = NULL;
1547         struct ll_sb_info *sbi = ll_i2sbi(dir);
1548         struct llcrypt_str *disk_link = NULL;
1549         bool encrypt = false;
1550         int err;
1551
1552         ENTRY;
1553         if (unlikely(tgt != NULL)) {
1554                 disk_link = (struct llcrypt_str *)rdev;
1555                 rdev = 0;
1556                 if (!disk_link)
1557                         RETURN(-EINVAL);
1558         }
1559
1560 again:
1561         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name,
1562                                      name->len, 0, opc, NULL);
1563         if (IS_ERR(op_data))
1564                 GOTO(err_exit, err = PTR_ERR(op_data));
1565
1566         if (S_ISDIR(mode))
1567                 ll_qos_mkdir_prep(op_data, dir);
1568
1569         if (test_bit(LL_SBI_FILE_SECCTX, sbi->ll_flags)) {
1570                 err = ll_dentry_init_security(dchild, mode, &dchild->d_name,
1571                                               &op_data->op_file_secctx_name,
1572                                               &op_data->op_file_secctx,
1573                                               &op_data->op_file_secctx_size);
1574                 if (err < 0)
1575                         GOTO(err_exit, err);
1576         }
1577
1578         if (ll_sbi_has_encrypt(sbi) &&
1579             ((IS_ENCRYPTED(dir) &&
1580             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) ||
1581             (unlikely(llcrypt_dummy_context_enabled(dir)) && S_ISDIR(mode)))) {
1582                 err = llcrypt_get_encryption_info(dir);
1583                 if (err)
1584                         GOTO(err_exit, err);
1585                 if (!llcrypt_has_encryption_key(dir))
1586                         GOTO(err_exit, err = -ENOKEY);
1587                 encrypt = true;
1588         }
1589
1590         if (encrypt) {
1591                 err = llcrypt_inherit_context(dir, NULL, op_data, false);
1592                 if (err)
1593                         GOTO(err_exit, err);
1594
1595                 if (S_ISLNK(mode)) {
1596                         /* llcrypt needs inode to encrypt target name, so create
1597                          * a fake inode and associate encryption context got
1598                          * from llcrypt_inherit_context.
1599                          */
1600                         struct inode *fakeinode =
1601                                 dchild->d_sb->s_op->alloc_inode(dchild->d_sb);
1602
1603                         if (!fakeinode)
1604                                 GOTO(err_exit, err = -ENOMEM);
1605                         fakeinode->i_sb = dchild->d_sb;
1606                         fakeinode->i_mode |= S_IFLNK;
1607 #ifdef IOP_XATTR
1608                         fakeinode->i_opflags |= IOP_XATTR;
1609 #endif
1610                         ll_lli_init(ll_i2info(fakeinode));
1611                         err = ll_set_encflags(fakeinode,
1612                                               op_data->op_file_encctx,
1613                                               op_data->op_file_encctx_size,
1614                                               true);
1615                         if (!err)
1616                                 err = __llcrypt_encrypt_symlink(fakeinode, tgt,
1617                                                                 strlen(tgt),
1618                                                                 disk_link);
1619
1620                         ll_xattr_cache_destroy(fakeinode);
1621                         llcrypt_put_encryption_info(fakeinode);
1622                         dchild->d_sb->s_op->destroy_inode(fakeinode);
1623                         if (err)
1624                                 GOTO(err_exit, err);
1625                 }
1626         }
1627
1628         err = md_create(sbi->ll_md_exp, op_data, tgt ? disk_link->name : NULL,
1629                         tgt ? disk_link->len : 0, mode,
1630                         from_kuid(&init_user_ns, current_fsuid()),
1631                         from_kgid(&init_user_ns, current_fsgid()),
1632                         current_cap(), rdev, &request);
1633 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 14, 58, 0)
1634         /*
1635          * server < 2.12.58 doesn't pack default LMV in intent_getattr reply,
1636          * fetch default LMV here.
1637          */
1638         if (unlikely(err == -EREMOTE)) {
1639                 struct ll_inode_info    *lli = ll_i2info(dir);
1640                 struct lmv_user_md      *lum;
1641                 int                     lumsize;
1642                 int                     err2;
1643
1644                 ptlrpc_req_finished(request);
1645                 request = NULL;
1646
1647                 err2 = ll_dir_getstripe(dir, (void **)&lum, &lumsize, &request,
1648                                         OBD_MD_DEFAULT_MEA);
1649                 ll_finish_md_op_data(op_data);
1650                 op_data = NULL;
1651                 if (err2 == 0) {
1652                         struct lustre_md md = { NULL };
1653
1654                         md.body = req_capsule_server_get(&request->rq_pill,
1655                                                          &RMF_MDT_BODY);
1656                         if (!md.body)
1657                                 GOTO(err_exit, err = -EPROTO);
1658
1659                         OBD_ALLOC_PTR(md.default_lmv);
1660                         if (!md.default_lmv)
1661                                 GOTO(err_exit, err = -ENOMEM);
1662
1663                         md.default_lmv->lsm_md_magic = lum->lum_magic;
1664                         md.default_lmv->lsm_md_stripe_count =
1665                                 lum->lum_stripe_count;
1666                         md.default_lmv->lsm_md_master_mdt_index =
1667                                 lum->lum_stripe_offset;
1668                         md.default_lmv->lsm_md_hash_type = lum->lum_hash_type;
1669                         md.default_lmv->lsm_md_max_inherit =
1670                                 lum->lum_max_inherit;
1671                         md.default_lmv->lsm_md_max_inherit_rr =
1672                                 lum->lum_max_inherit_rr;
1673
1674                         err = ll_update_inode(dir, &md);
1675                         md_free_lustre_md(sbi->ll_md_exp, &md);
1676                         if (err)
1677                                 GOTO(err_exit, err);
1678                 } else if (err2 == -ENODATA && lli->lli_default_lsm_md) {
1679                         /*
1680                          * If there are no default stripe EA on the MDT, but the
1681                          * client has default stripe, then it probably means
1682                          * default stripe EA has just been deleted.
1683                          */
1684                         down_write(&lli->lli_lsm_sem);
1685                         if (lli->lli_default_lsm_md)
1686                                 OBD_FREE_PTR(lli->lli_default_lsm_md);
1687                         lli->lli_default_lsm_md = NULL;
1688                         up_write(&lli->lli_lsm_sem);
1689                 } else {
1690                         GOTO(err_exit, err);
1691                 }
1692
1693                 ptlrpc_req_finished(request);
1694                 request = NULL;
1695                 goto again;
1696         }
1697 #endif
1698
1699         if (err < 0)
1700                 GOTO(err_exit, err);
1701
1702         ll_update_times(request, dir);
1703
1704         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_NEWNODE_PAUSE, cfs_fail_val);
1705
1706         err = ll_prep_inode(&inode, &request->rq_pill, dchild->d_sb, NULL);
1707         if (err)
1708                 GOTO(err_exit, err);
1709
1710         if (test_bit(LL_SBI_FILE_SECCTX, sbi->ll_flags)) {
1711                 /* must be done before d_instantiate, because it calls
1712                  * security_d_instantiate, which means a getxattr if security
1713                  * context is not set yet */
1714                 /* no need to protect selinux_inode_setsecurity() by
1715                  * inode_lock. Taking it would lead to a client deadlock
1716                  * LU-13617
1717                  */
1718                 err = security_inode_notifysecctx(inode,
1719                                                   op_data->op_file_secctx,
1720                                                   op_data->op_file_secctx_size);
1721                 if (err)
1722                         GOTO(err_exit, err);
1723         }
1724
1725         d_instantiate(dchild, inode);
1726
1727         if (encrypt) {
1728                 err = ll_set_encflags(inode, op_data->op_file_encctx,
1729                                       op_data->op_file_encctx_size, true);
1730                 if (err)
1731                         GOTO(err_exit, err);
1732
1733                 if (S_ISLNK(mode)) {
1734                         struct ll_inode_info *lli = ll_i2info(inode);
1735
1736                         /* Cache the plaintext symlink target
1737                          * for later use by get_link()
1738                          */
1739                         OBD_ALLOC(lli->lli_symlink_name, strlen(tgt) + 1);
1740                         /* do not return an error if we cannot
1741                          * cache the symlink locally
1742                          */
1743                         if (lli->lli_symlink_name)
1744                                 memcpy(lli->lli_symlink_name,
1745                                        tgt, strlen(tgt) + 1);
1746                 }
1747         }
1748
1749         if (!test_bit(LL_SBI_FILE_SECCTX, sbi->ll_flags)) {
1750                 err = ll_inode_init_security(dchild, inode, dir);
1751                 if (err)
1752                         GOTO(err_exit, err);
1753         }
1754
1755         EXIT;
1756 err_exit:
1757         if (request != NULL)
1758                 ptlrpc_req_finished(request);
1759
1760         if (!IS_ERR_OR_NULL(op_data))
1761                 ll_finish_md_op_data(op_data);
1762
1763         RETURN(err);
1764 }
1765
1766 static int ll_mknod(struct inode *dir, struct dentry *dchild, umode_t mode,
1767                     dev_t rdev)
1768 {
1769         ktime_t kstart = ktime_get();
1770         int err;
1771         ENTRY;
1772
1773         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p) mode %o dev %x\n",
1774                dchild, PFID(ll_inode2fid(dir)), dir, mode, rdev);
1775
1776         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1777                 mode &= ~current_umask();
1778
1779         switch (mode & S_IFMT) {
1780         case 0:
1781                 mode |= S_IFREG;
1782                 /* fallthrough */
1783         case S_IFREG:
1784         case S_IFCHR:
1785         case S_IFBLK:
1786         case S_IFIFO:
1787         case S_IFSOCK:
1788                 err = ll_new_node(dir, dchild, NULL, mode, old_encode_dev(rdev),
1789                                   LUSTRE_OPC_MKNOD);
1790                 break;
1791         case S_IFDIR:
1792                 err = -EPERM;
1793                 break;
1794         default:
1795                 err = -EINVAL;
1796         }
1797
1798         if (!err)
1799                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKNOD,
1800                                    ktime_us_delta(ktime_get(), kstart));
1801
1802         RETURN(err);
1803 }
1804
1805 /*
1806  * Plain create. Intent create is handled in atomic_open.
1807  */
1808 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
1809                         umode_t mode, bool want_excl)
1810 {
1811         ktime_t kstart = ktime_get();
1812         int rc;
1813
1814         CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1815
1816         CDEBUG(D_VFSTRACE,
1817                "VFS Op:name=%pd, dir="DFID"(%p), flags=%u, excl=%d\n",
1818                dentry, PFID(ll_inode2fid(dir)), dir, mode, want_excl);
1819
1820         /* Using mknod(2) to create a regular file is designed to not recognize
1821          * volatile file name, so we use ll_mknod() here. */
1822         rc = ll_mknod(dir, dentry, mode, 0);
1823
1824         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, unhashed %d\n",
1825                dentry, d_unhashed(dentry));
1826
1827         if (!rc)
1828                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE,
1829                                    ktime_us_delta(ktime_get(), kstart));
1830
1831         return rc;
1832 }
1833
1834 static int ll_symlink(struct inode *dir, struct dentry *dchild,
1835                       const char *oldpath)
1836 {
1837         ktime_t kstart = ktime_get();
1838         int len = strlen(oldpath);
1839         struct llcrypt_str disk_link;
1840         int err;
1841         ENTRY;
1842
1843         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p), target=%.*s\n",
1844                dchild, PFID(ll_inode2fid(dir)), dir, 3000, oldpath);
1845
1846         err = llcrypt_prepare_symlink(dir, oldpath, len, dir->i_sb->s_blocksize,
1847                                       &disk_link);
1848         if (err)
1849                 RETURN(err);
1850
1851         err = ll_new_node(dir, dchild, oldpath, S_IFLNK | S_IRWXUGO,
1852                           (__u64)&disk_link, LUSTRE_OPC_SYMLINK);
1853
1854         if (disk_link.name != (unsigned char *)oldpath)
1855                 kfree(disk_link.name);
1856
1857         if (!err)
1858                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_SYMLINK,
1859                                    ktime_us_delta(ktime_get(), kstart));
1860
1861         RETURN(err);
1862 }
1863
1864 static int ll_link(struct dentry *old_dentry, struct inode *dir,
1865                    struct dentry *new_dentry)
1866 {
1867         struct inode *src = old_dentry->d_inode;
1868         struct qstr *name = &new_dentry->d_name;
1869         struct ll_sb_info *sbi = ll_i2sbi(dir);
1870         struct ptlrpc_request *request = NULL;
1871         struct md_op_data *op_data;
1872         ktime_t kstart = ktime_get();
1873         int err;
1874
1875         ENTRY;
1876         CDEBUG(D_VFSTRACE,
1877                "VFS Op: inode="DFID"(%p), dir="DFID"(%p), target=%pd\n",
1878                PFID(ll_inode2fid(src)), src,
1879                PFID(ll_inode2fid(dir)), dir, new_dentry);
1880
1881         err = llcrypt_prepare_link(old_dentry, dir, new_dentry);
1882         if (err)
1883                 RETURN(err);
1884
1885         op_data = ll_prep_md_op_data(NULL, src, dir, name->name, name->len,
1886                                      0, LUSTRE_OPC_ANY, NULL);
1887         if (IS_ERR(op_data))
1888                 RETURN(PTR_ERR(op_data));
1889
1890         err = md_link(sbi->ll_md_exp, op_data, &request);
1891         ll_finish_md_op_data(op_data);
1892         if (err)
1893                 GOTO(out, err);
1894
1895         ll_update_times(request, dir);
1896         ll_stats_ops_tally(sbi, LPROC_LL_LINK,
1897                            ktime_us_delta(ktime_get(), kstart));
1898         EXIT;
1899 out:
1900         ptlrpc_req_finished(request);
1901         RETURN(err);
1902 }
1903
1904 static int ll_mkdir(struct inode *dir, struct dentry *dchild, umode_t mode)
1905 {
1906         ktime_t kstart = ktime_get();
1907         int err;
1908         ENTRY;
1909
1910         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p)\n",
1911                dchild, PFID(ll_inode2fid(dir)), dir);
1912
1913         if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1914                 mode &= ~current_umask();
1915
1916         mode = (mode & (S_IRWXUGO|S_ISVTX)) | S_IFDIR;
1917
1918         err = ll_new_node(dir, dchild, NULL, mode, 0, LUSTRE_OPC_MKDIR);
1919         if (err == 0)
1920                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKDIR,
1921                                    ktime_us_delta(ktime_get(), kstart));
1922
1923         RETURN(err);
1924 }
1925
1926 static int ll_rmdir(struct inode *dir, struct dentry *dchild)
1927 {
1928         struct qstr *name = &dchild->d_name;
1929         struct ptlrpc_request *request = NULL;
1930         struct md_op_data *op_data;
1931         ktime_t kstart = ktime_get();
1932         int rc;
1933
1934         ENTRY;
1935
1936         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p)\n",
1937                dchild, PFID(ll_inode2fid(dir)), dir);
1938
1939         if (unlikely(d_mountpoint(dchild)))
1940                 RETURN(-EBUSY);
1941
1942         /* some foreign dir may not be allowed to be removed */
1943         if (!ll_foreign_is_removable(dchild, false))
1944                 RETURN(-EPERM);
1945
1946         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len,
1947                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1948         if (IS_ERR(op_data))
1949                 RETURN(PTR_ERR(op_data));
1950
1951         if (dchild->d_inode != NULL)
1952                 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1953
1954         if (fid_is_zero(&op_data->op_fid2))
1955                 op_data->op_fid2 = op_data->op_fid3;
1956         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1957         ll_finish_md_op_data(op_data);
1958         if (!rc) {
1959                 struct mdt_body *body;
1960
1961                 ll_update_times(request, dir);
1962                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR,
1963                                    ktime_us_delta(ktime_get(), kstart));
1964
1965                 /*
1966                  * The server puts attributes in on the last unlink, use them
1967                  * to update the link count so the inode can be freed
1968                  * immediately.
1969                  */
1970                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
1971                 if (body->mbo_valid & OBD_MD_FLNLINK) {
1972                         spin_lock(&dchild->d_inode->i_lock);
1973                         set_nlink(dchild->d_inode, body->mbo_nlink);
1974                         spin_unlock(&dchild->d_inode->i_lock);
1975                 }
1976         }
1977
1978         ptlrpc_req_finished(request);
1979
1980         RETURN(rc);
1981 }
1982
1983 /**
1984  * Remove dir entry
1985  **/
1986 int ll_rmdir_entry(struct inode *dir, char *name, int namelen)
1987 {
1988         struct ptlrpc_request *request = NULL;
1989         struct md_op_data *op_data;
1990         ktime_t kstart = ktime_get();
1991         int rc;
1992         ENTRY;
1993
1994         CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1995                namelen, name, PFID(ll_inode2fid(dir)), dir);
1996
1997         op_data = ll_prep_md_op_data(NULL, dir, NULL, name, strlen(name),
1998                                      S_IFDIR, LUSTRE_OPC_ANY, NULL);
1999         if (IS_ERR(op_data))
2000                 RETURN(PTR_ERR(op_data));
2001         op_data->op_cli_flags |= CLI_RM_ENTRY;
2002         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
2003         ll_finish_md_op_data(op_data);
2004         if (!rc)
2005                 ll_update_times(request, dir);
2006
2007         ptlrpc_req_finished(request);
2008         if (!rc)
2009                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR,
2010                                    ktime_us_delta(ktime_get(), kstart));
2011         RETURN(rc);
2012 }
2013
2014 static int ll_unlink(struct inode *dir, struct dentry *dchild)
2015 {
2016         struct qstr *name = &dchild->d_name;
2017         struct ptlrpc_request *request = NULL;
2018         struct md_op_data *op_data;
2019         struct mdt_body *body;
2020         ktime_t kstart = ktime_get();
2021         int rc;
2022
2023         ENTRY;
2024
2025         CDEBUG(D_VFSTRACE, "VFS Op:name=%pd, dir="DFID"(%p)\n",
2026                dchild, PFID(ll_inode2fid(dir)), dir);
2027
2028         /*
2029          * XXX: unlink bind mountpoint maybe call to here,
2030          * just check it as vfs_unlink does.
2031          */
2032         if (unlikely(d_mountpoint(dchild)))
2033                 RETURN(-EBUSY);
2034
2035         /* some foreign file/dir may not be allowed to be unlinked */
2036         if (!ll_foreign_is_removable(dchild, false))
2037                 RETURN(-EPERM);
2038
2039         op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len, 0,
2040                                      LUSTRE_OPC_ANY, NULL);
2041         if (IS_ERR(op_data))
2042                 RETURN(PTR_ERR(op_data));
2043
2044         op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
2045         /* notify lower layer if inode has dirty pages */
2046         if (S_ISREG(dchild->d_inode->i_mode) &&
2047             ll_i2info(dchild->d_inode)->lli_clob &&
2048             dirty_cnt(dchild->d_inode))
2049                 op_data->op_cli_flags |= CLI_DIRTY_DATA;
2050         if (fid_is_zero(&op_data->op_fid2))
2051                 op_data->op_fid2 = op_data->op_fid3;
2052         rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
2053         ll_finish_md_op_data(op_data);
2054         if (rc)
2055                 GOTO(out, rc);
2056
2057         /*
2058          * The server puts attributes in on the last unlink, use them to update
2059          * the link count so the inode can be freed immediately.
2060          */
2061         body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
2062         if (body->mbo_valid & OBD_MD_FLNLINK) {
2063                 spin_lock(&dchild->d_inode->i_lock);
2064                 set_nlink(dchild->d_inode, body->mbo_nlink);
2065                 spin_unlock(&dchild->d_inode->i_lock);
2066         }
2067
2068         ll_update_times(request, dir);
2069
2070 out:
2071         ptlrpc_req_finished(request);
2072         if (!rc)
2073                 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_UNLINK,
2074                                    ktime_us_delta(ktime_get(), kstart));
2075         RETURN(rc);
2076 }
2077
2078 static int ll_rename(struct inode *src, struct dentry *src_dchild,
2079                      struct inode *tgt, struct dentry *tgt_dchild
2080 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
2081                      , unsigned int flags
2082 #endif
2083                      )
2084 {
2085         struct ptlrpc_request *request = NULL;
2086         struct ll_sb_info *sbi = ll_i2sbi(src);
2087         struct md_op_data *op_data;
2088         ktime_t kstart = ktime_get();
2089         umode_t mode = 0;
2090         struct llcrypt_name foldname, fnewname;
2091         int err;
2092         ENTRY;
2093
2094 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
2095         if (flags)
2096                 return -EINVAL;
2097 #endif
2098
2099         CDEBUG(D_VFSTRACE,
2100                "VFS Op:oldname=%pd, src_dir="DFID"(%p), newname=%pd, tgt_dir="DFID"(%p)\n",
2101                src_dchild, PFID(ll_inode2fid(src)), src,
2102                tgt_dchild, PFID(ll_inode2fid(tgt)), tgt);
2103
2104         if (unlikely(d_mountpoint(src_dchild) || d_mountpoint(tgt_dchild)))
2105                 RETURN(-EBUSY);
2106
2107 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
2108         err = llcrypt_prepare_rename(src, src_dchild, tgt, tgt_dchild, flags);
2109 #else
2110         err = llcrypt_prepare_rename(src, src_dchild, tgt, tgt_dchild, 0);
2111 #endif
2112         if (err)
2113                 RETURN(err);
2114         /* we prevent an encrypted file from being renamed
2115          * into an unencrypted dir
2116          */
2117         if (IS_ENCRYPTED(src) && !IS_ENCRYPTED(tgt))
2118                 RETURN(-EXDEV);
2119
2120         if (src_dchild->d_inode)
2121                 mode = src_dchild->d_inode->i_mode;
2122
2123         if (tgt_dchild->d_inode)
2124                 mode = tgt_dchild->d_inode->i_mode;
2125
2126         op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, mode,
2127                                      LUSTRE_OPC_ANY, NULL);
2128         if (IS_ERR(op_data))
2129                 RETURN(PTR_ERR(op_data));
2130
2131         if (src_dchild->d_inode)
2132                 op_data->op_fid3 = *ll_inode2fid(src_dchild->d_inode);
2133
2134         if (tgt_dchild->d_inode)
2135                 op_data->op_fid4 = *ll_inode2fid(tgt_dchild->d_inode);
2136
2137         err = ll_setup_filename(src, &src_dchild->d_name, 1, &foldname, NULL);
2138         if (err)
2139                 RETURN(err);
2140         err = ll_setup_filename(tgt, &tgt_dchild->d_name, 1, &fnewname, NULL);
2141         if (err) {
2142                 llcrypt_free_filename(&foldname);
2143                 RETURN(err);
2144         }
2145         err = md_rename(sbi->ll_md_exp, op_data,
2146                         foldname.disk_name.name, foldname.disk_name.len,
2147                         fnewname.disk_name.name, fnewname.disk_name.len,
2148                         &request);
2149         llcrypt_free_filename(&foldname);
2150         llcrypt_free_filename(&fnewname);
2151         ll_finish_md_op_data(op_data);
2152         if (!err) {
2153                 ll_update_times(request, src);
2154                 ll_update_times(request, tgt);
2155         }
2156
2157         ptlrpc_req_finished(request);
2158
2159         if (!err) {
2160                 d_move(src_dchild, tgt_dchild);
2161                 ll_stats_ops_tally(sbi, LPROC_LL_RENAME,
2162                                    ktime_us_delta(ktime_get(), kstart));
2163         }
2164
2165         RETURN(err);
2166 }
2167
2168 const struct inode_operations ll_dir_inode_operations = {
2169         .mknod          = ll_mknod,
2170         .atomic_open    = ll_atomic_open,
2171         .lookup         = ll_lookup_nd,
2172         .create         = ll_create_nd,
2173         /* We need all these non-raw things for NFSD, to not patch it. */
2174         .unlink         = ll_unlink,
2175         .mkdir          = ll_mkdir,
2176         .rmdir          = ll_rmdir,
2177         .symlink        = ll_symlink,
2178         .link           = ll_link,
2179         .rename         = ll_rename,
2180         .setattr        = ll_setattr,
2181         .getattr        = ll_getattr,
2182         .permission     = ll_inode_permission,
2183 #ifdef HAVE_IOP_XATTR
2184         .setxattr       = ll_setxattr,
2185         .getxattr       = ll_getxattr,
2186         .removexattr    = ll_removexattr,
2187 #endif
2188         .listxattr      = ll_listxattr,
2189         .get_acl        = ll_get_acl,
2190 #ifdef HAVE_IOP_SET_ACL
2191         .set_acl        = ll_set_acl,
2192 #endif
2193 };
2194
2195 const struct inode_operations ll_special_inode_operations = {
2196         .setattr        = ll_setattr,
2197         .getattr        = ll_getattr,
2198         .permission     = ll_inode_permission,
2199 #ifdef HAVE_IOP_XATTR
2200         .setxattr       = ll_setxattr,
2201         .getxattr       = ll_getxattr,
2202         .removexattr    = ll_removexattr,
2203 #endif
2204         .listxattr      = ll_listxattr,
2205         .get_acl        = ll_get_acl,
2206 #ifdef HAVE_IOP_SET_ACL
2207         .set_acl        = ll_set_acl,
2208 #endif
2209 };