4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 #include <linux/sched.h>
36 #include <linux/quotaops.h>
37 #include <linux/highmem.h>
38 #include <linux/pagemap.h>
39 #include <linux/security.h>
40 #include <linux/user_namespace.h>
41 #ifdef HAVE_UIDGID_HEADER
42 # include <linux/uidgid.h>
45 #define DEBUG_SUBSYSTEM S_LLITE
47 #include <obd_support.h>
48 #include <lustre_fid.h>
49 #include <lustre_dlm.h>
50 #include <lustre_ver.h>
51 #include "llite_internal.h"
53 static int ll_create_it(struct inode *dir, struct dentry *dentry,
54 struct lookup_intent *it,
55 void *secctx, __u32 secctxlen);
57 /* called from iget5_locked->find_inode() under inode_lock spinlock */
58 static int ll_test_inode(struct inode *inode, void *opaque)
60 struct ll_inode_info *lli = ll_i2info(inode);
61 struct lustre_md *md = opaque;
63 if (unlikely(!(md->body->mbo_valid & OBD_MD_FLID))) {
64 CERROR("MDS body missing FID\n");
68 if (!lu_fid_eq(&lli->lli_fid, &md->body->mbo_fid1))
74 static int ll_set_inode(struct inode *inode, void *opaque)
76 struct ll_inode_info *lli = ll_i2info(inode);
77 struct mdt_body *body = ((struct lustre_md *)opaque)->body;
79 if (unlikely(!(body->mbo_valid & OBD_MD_FLID))) {
80 CERROR("MDS body missing FID\n");
84 lli->lli_fid = body->mbo_fid1;
85 if (unlikely(!(body->mbo_valid & OBD_MD_FLTYPE))) {
86 CERROR("Can not initialize inode "DFID" without object type: "
88 PFID(&lli->lli_fid), body->mbo_valid);
92 inode->i_mode = (inode->i_mode & ~S_IFMT) | (body->mbo_mode & S_IFMT);
93 if (unlikely(inode->i_mode == 0)) {
94 CERROR("Invalid inode "DFID" type\n", PFID(&lli->lli_fid));
105 * Get an inode by inode number(@hash), which is already instantiated by
106 * the intent lookup).
108 struct inode *ll_iget(struct super_block *sb, ino_t hash,
109 struct lustre_md *md)
117 inode = iget5_locked(sb, hash, ll_test_inode, ll_set_inode, md);
119 RETURN(ERR_PTR(-ENOMEM));
121 if (inode->i_state & I_NEW) {
122 rc = ll_read_inode2(inode, md);
123 if (rc == 0 && S_ISREG(inode->i_mode) &&
124 ll_i2info(inode)->lli_clob == NULL)
125 rc = cl_file_inode_init(inode, md);
128 /* Let's clear directory lsm here, otherwise
129 * make_bad_inode() will reset the inode mode
130 * to regular, then ll_clear_inode will not
131 * be able to clear lsm_md */
132 if (S_ISDIR(inode->i_mode))
133 ll_dir_clear_lsm_md(inode);
134 make_bad_inode(inode);
135 unlock_new_inode(inode);
139 inode_has_no_xattr(inode);
140 unlock_new_inode(inode);
142 } else if (is_bad_inode(inode)) {
144 inode = ERR_PTR(-ESTALE);
145 } else if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
146 rc = ll_update_inode(inode, md);
147 CDEBUG(D_VFSTRACE, "got inode: "DFID"(%p): rc = %d\n",
148 PFID(&md->body->mbo_fid1), inode, rc);
150 if (S_ISDIR(inode->i_mode))
151 ll_dir_clear_lsm_md(inode);
160 static void ll_invalidate_negative_children(struct inode *dir)
162 struct dentry *dentry, *tmp_subdir;
163 DECLARE_LL_D_HLIST_NODE_PTR(p);
166 ll_d_hlist_for_each_entry(dentry, p, &dir->i_dentry) {
167 spin_lock(&dentry->d_lock);
168 if (!list_empty(&dentry->d_subdirs)) {
169 struct dentry *child;
171 list_for_each_entry_safe(child, tmp_subdir,
174 if (child->d_inode == NULL)
175 d_lustre_invalidate(child, 1);
178 spin_unlock(&dentry->d_lock);
180 ll_unlock_dcache(dir);
183 int ll_test_inode_by_fid(struct inode *inode, void *opaque)
185 return lu_fid_eq(&ll_i2info(inode)->lli_fid, opaque);
188 int ll_md_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
189 void *data, int flag)
191 struct lustre_handle lockh;
196 case LDLM_CB_BLOCKING:
197 ldlm_lock2handle(lock, &lockh);
198 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
200 CDEBUG(D_INODE, "ldlm_cli_cancel: rc = %d\n", rc);
204 case LDLM_CB_CANCELING: {
205 struct inode *inode = ll_inode_from_resource_lock(lock);
206 __u64 bits = lock->l_policy_data.l_inodebits.bits;
208 /* Inode is set to lock->l_resource->lr_lvb_inode
209 * for mdc - bug 24555 */
210 LASSERT(lock->l_ast_data == NULL);
215 /* Invalidate all dentries associated with this inode */
216 LASSERT(ldlm_is_canceling(lock));
218 if (!fid_res_name_eq(ll_inode2fid(inode),
219 &lock->l_resource->lr_name)) {
220 LDLM_ERROR(lock, "data mismatch with object "DFID"(%p)",
221 PFID(ll_inode2fid(inode)), inode);
225 if (bits & MDS_INODELOCK_XATTR) {
226 if (S_ISDIR(inode->i_mode))
227 ll_i2info(inode)->lli_def_stripe_offset = -1;
228 ll_xattr_cache_destroy(inode);
229 bits &= ~MDS_INODELOCK_XATTR;
232 /* For OPEN locks we differentiate between lock modes
233 * LCK_CR, LCK_CW, LCK_PR - bug 22891 */
234 if (bits & MDS_INODELOCK_OPEN)
235 ll_have_md_lock(inode, &bits, lock->l_req_mode);
237 if (bits & MDS_INODELOCK_OPEN) {
240 switch (lock->l_req_mode) {
251 LDLM_ERROR(lock, "bad lock mode for OPEN lock");
255 ll_md_real_close(inode, fmode);
257 bits &= ~MDS_INODELOCK_OPEN;
260 if (bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
261 MDS_INODELOCK_LAYOUT | MDS_INODELOCK_PERM))
262 ll_have_md_lock(inode, &bits, LCK_MINMODE);
264 if (bits & MDS_INODELOCK_LAYOUT) {
265 struct cl_object_conf conf = {
266 .coc_opc = OBJECT_CONF_INVALIDATE,
270 rc = ll_layout_conf(inode, &conf);
272 CDEBUG(D_INODE, "cannot invalidate layout of "
274 PFID(ll_inode2fid(inode)), rc);
277 if (bits & MDS_INODELOCK_UPDATE) {
278 struct ll_inode_info *lli = ll_i2info(inode);
279 lli->lli_update_atime = 1;
282 if ((bits & MDS_INODELOCK_UPDATE) && S_ISDIR(inode->i_mode)) {
283 struct ll_inode_info *lli = ll_i2info(inode);
285 CDEBUG(D_INODE, "invalidating inode "DFID" lli = %p, "
286 "pfid = "DFID"\n", PFID(ll_inode2fid(inode)),
287 lli, PFID(&lli->lli_pfid));
288 truncate_inode_pages(inode->i_mapping, 0);
290 if (unlikely(!fid_is_zero(&lli->lli_pfid))) {
291 struct inode *master_inode = NULL;
294 /* This is slave inode, since all of the child
295 * dentry is connected on the master inode, so
296 * we have to invalidate the negative children
298 CDEBUG(D_INODE, "Invalidate s"DFID" m"DFID"\n",
299 PFID(ll_inode2fid(inode)),
300 PFID(&lli->lli_pfid));
302 hash = cl_fid_build_ino(&lli->lli_pfid,
303 ll_need_32bit_api(ll_i2sbi(inode)));
305 /* Do not lookup the inode with ilookup5,
306 * otherwise it will cause dead lock,
308 * 1. Client1 send chmod req to the MDT0, then
309 * on MDT0, it enqueues master and all of its
310 * slaves lock, (mdt_attr_set() ->
311 * mdt_lock_slaves()), after gets master and
312 * stripe0 lock, it will send the enqueue req
313 * (for stripe1) to MDT1, then MDT1 finds the
314 * lock has been granted to client2. Then MDT1
315 * sends blocking ast to client2.
317 * 2. At the same time, client2 tries to unlink
318 * the striped dir (rm -rf striped_dir), and
319 * during lookup, it will hold the master inode
320 * of the striped directory, whose inode state
321 * is NEW, then tries to revalidate all of its
322 * slaves, (ll_prep_inode()->ll_iget()->
323 * ll_read_inode2()-> ll_update_inode().). And
324 * it will be blocked on the server side because
327 * 3. Then the client get the blocking_ast req,
328 * cancel the lock, but being blocked if using
329 * ->ilookup5()), because master inode state is
331 master_inode = ilookup5_nowait(inode->i_sb,
332 hash, ll_test_inode_by_fid,
333 (void *)&lli->lli_pfid);
335 ll_invalidate_negative_children(
340 ll_invalidate_negative_children(inode);
344 if ((bits & (MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM)) &&
345 inode->i_sb->s_root != NULL &&
346 inode != inode->i_sb->s_root->d_inode)
347 ll_invalidate_aliases(inode);
359 __u32 ll_i2suppgid(struct inode *i)
361 if (in_group_p(i->i_gid))
362 return (__u32)from_kgid(&init_user_ns, i->i_gid);
364 return (__u32) __kgid_val(INVALID_GID);
367 /* Pack the required supplementary groups into the supplied groups array.
368 * If we don't need to use the groups from the target inode(s) then we
369 * instead pack one or more groups from the user's supplementary group
370 * array in case it might be useful. Not needed if doing an MDS-side upcall. */
371 void ll_i2gids(__u32 *suppgids, struct inode *i1, struct inode *i2)
374 LASSERT(suppgids != NULL);
376 suppgids[0] = ll_i2suppgid(i1);
379 suppgids[1] = ll_i2suppgid(i2);
385 * try to reuse three types of dentry:
386 * 1. unhashed alias, this one is unhashed by d_invalidate (but it may be valid
387 * by concurrent .revalidate).
388 * 2. INVALID alias (common case for no valid ldlm lock held, but this flag may
389 * be cleared by others calling d_lustre_revalidate).
390 * 3. DISCONNECTED alias.
392 static struct dentry *ll_find_alias(struct inode *inode, struct dentry *dentry)
394 struct dentry *alias, *discon_alias, *invalid_alias;
395 DECLARE_LL_D_HLIST_NODE_PTR(p);
397 if (ll_d_hlist_empty(&inode->i_dentry))
400 discon_alias = invalid_alias = NULL;
402 ll_lock_dcache(inode);
403 ll_d_hlist_for_each_entry(alias, p, &inode->i_dentry) {
404 LASSERT(alias != dentry);
406 spin_lock(&alias->d_lock);
407 if ((alias->d_flags & DCACHE_DISCONNECTED) &&
408 S_ISDIR(inode->i_mode))
409 /* LASSERT(last_discon == NULL); LU-405, bz 20055 */
410 discon_alias = alias;
411 else if (alias->d_parent == dentry->d_parent &&
412 alias->d_name.hash == dentry->d_name.hash &&
413 alias->d_name.len == dentry->d_name.len &&
414 memcmp(alias->d_name.name, dentry->d_name.name,
415 dentry->d_name.len) == 0)
416 invalid_alias = alias;
417 spin_unlock(&alias->d_lock);
422 alias = invalid_alias ?: discon_alias ?: NULL;
424 spin_lock(&alias->d_lock);
426 spin_unlock(&alias->d_lock);
428 ll_unlock_dcache(inode);
434 * Similar to d_splice_alias(), but lustre treats invalid alias
435 * similar to DCACHE_DISCONNECTED, and tries to use it anyway.
437 struct dentry *ll_splice_alias(struct inode *inode, struct dentry *de)
443 new = ll_find_alias(inode, de);
453 "Reuse dentry %p inode %p refc %d flags %#x\n",
454 new, new->d_inode, ll_d_count(new), new->d_flags);
462 CDEBUG(D_DENTRY, "Add dentry %p inode %p refc %d flags %#x\n",
463 de, de->d_inode, ll_d_count(de), de->d_flags);
467 static int ll_lookup_it_finish(struct ptlrpc_request *request,
468 struct lookup_intent *it,
469 struct inode *parent, struct dentry **de)
471 struct inode *inode = NULL;
474 struct dentry *alias;
477 /* NB 1 request reference will be taken away by ll_intent_lock()
479 CDEBUG(D_DENTRY, "it %p it_disposition %x\n", it,
481 if (!it_disposition(it, DISP_LOOKUP_NEG)) {
482 rc = ll_prep_inode(&inode, request, (*de)->d_sb, it);
486 ll_set_lock_data(ll_i2sbi(parent)->ll_md_exp, inode, it, &bits);
488 /* We used to query real size from OSTs here, but actually
489 this is not needed. For stat() calls size would be updated
490 from subsequent do_revalidate()->ll_inode_revalidate_it() in
492 vfs_getattr_it->ll_getattr()->ll_inode_revalidate_it() in 2.6
493 Everybody else who needs correct file size would call
494 ll_glimpse_size or some equivalent themselves anyway.
495 Also see bug 7198. */
498 /* Only hash *de if it is unhashed (new dentry).
499 * Atoimc_open may passin hashed dentries for open.
501 alias = ll_splice_alias(inode, *de);
503 GOTO(out, rc = PTR_ERR(alias));
507 if (!it_disposition(it, DISP_LOOKUP_NEG)) {
508 /* we have lookup look - unhide dentry */
509 if (bits & MDS_INODELOCK_LOOKUP)
510 d_lustre_revalidate(*de);
511 } else if (!it_disposition(it, DISP_OPEN_CREATE)) {
512 /* If file created on server, don't depend on parent UPDATE
513 * lock to unhide it. It is left hidden and next lookup can
514 * find it in ll_splice_alias.
516 /* Check that parent has UPDATE lock. */
517 struct lookup_intent parent_it = {
519 .it_lock_handle = 0 };
520 struct lu_fid fid = ll_i2info(parent)->lli_fid;
522 /* If it is striped directory, get the real stripe parent */
523 if (unlikely(ll_i2info(parent)->lli_lsm_md != NULL)) {
524 rc = md_get_fid_from_lsm(ll_i2mdexp(parent),
525 ll_i2info(parent)->lli_lsm_md,
527 (*de)->d_name.len, &fid);
532 if (md_revalidate_lock(ll_i2mdexp(parent), &parent_it, &fid,
534 d_lustre_revalidate(*de);
535 ll_intent_release(&parent_it);
542 if (rc != 0 && it->it_op & IT_OPEN)
543 ll_open_cleanup((*de)->d_sb, request);
548 static struct dentry *ll_lookup_it(struct inode *parent, struct dentry *dentry,
549 struct lookup_intent *it,
550 void **secctx, __u32 *secctxlen)
552 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
553 struct dentry *save = dentry, *retval;
554 struct ptlrpc_request *req = NULL;
555 struct md_op_data *op_data = NULL;
560 if (dentry->d_name.len > ll_i2sbi(parent)->ll_namelen)
561 RETURN(ERR_PTR(-ENAMETOOLONG));
563 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), intent=%s\n",
564 dentry->d_name.len, dentry->d_name.name,
565 PFID(ll_inode2fid(parent)), parent, LL_IT2STR(it));
567 if (d_mountpoint(dentry))
568 CERROR("Tell Peter, lookup on mtpt, it %s\n", LL_IT2STR(it));
570 if (it == NULL || it->it_op == IT_GETXATTR)
573 if (it->it_op == IT_GETATTR && dentry_may_statahead(parent, dentry)) {
574 rc = ll_statahead(parent, &dentry, 0);
576 RETURN(dentry == save ? NULL : dentry);
579 if (it->it_op & IT_OPEN && it->it_flags & FMODE_WRITE &&
580 dentry->d_sb->s_flags & MS_RDONLY)
581 RETURN(ERR_PTR(-EROFS));
583 if (it->it_op & IT_CREAT)
584 opc = LUSTRE_OPC_CREATE;
586 opc = LUSTRE_OPC_ANY;
588 op_data = ll_prep_md_op_data(NULL, parent, NULL, dentry->d_name.name,
589 dentry->d_name.len, 0, opc, NULL);
591 GOTO(out, retval = ERR_CAST(op_data));
593 /* enforce umask if acl disabled or MDS doesn't support umask */
594 if (!IS_POSIXACL(parent) || !exp_connect_umask(ll_i2mdexp(parent)))
595 it->it_create_mode &= ~current_umask();
597 if (it->it_op & IT_CREAT &&
598 ll_i2sbi(parent)->ll_flags & LL_SBI_FILE_SECCTX) {
599 rc = ll_dentry_init_security(dentry, it->it_create_mode,
601 &op_data->op_file_secctx_name,
602 &op_data->op_file_secctx,
603 &op_data->op_file_secctx_size);
605 GOTO(out, retval = ERR_PTR(rc));
607 *secctx = op_data->op_file_secctx;
608 if (secctxlen != NULL)
609 *secctxlen = op_data->op_file_secctx_size;
612 rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
613 &ll_md_blocking_ast, 0);
614 /* If the MDS allows the client to chgrp (CFS_SETGRP_PERM), but the
615 * client does not know which suppgid should be sent to the MDS, or
616 * some other(s) changed the target file's GID after this RPC sent
617 * to the MDS with the suppgid as the original GID, then we should
618 * try again with right suppgid. */
619 if (rc == -EACCES && it->it_op & IT_OPEN &&
620 it_disposition(it, DISP_OPEN_DENY)) {
621 struct mdt_body *body;
623 LASSERT(req != NULL);
625 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
626 if (op_data->op_suppgids[0] == body->mbo_gid ||
627 op_data->op_suppgids[1] == body->mbo_gid ||
628 !in_group_p(make_kgid(&init_user_ns, body->mbo_gid)))
629 GOTO(out, retval = ERR_PTR(-EACCES));
631 fid_zero(&op_data->op_fid2);
632 op_data->op_suppgids[1] = body->mbo_gid;
633 ptlrpc_req_finished(req);
635 ll_intent_release(it);
636 rc = md_intent_lock(ll_i2mdexp(parent), op_data, it, &req,
637 &ll_md_blocking_ast, 0);
641 GOTO(out, retval = ERR_PTR(rc));
643 rc = ll_lookup_it_finish(req, it, parent, &dentry);
645 ll_intent_release(it);
646 GOTO(out, retval = ERR_PTR(rc));
649 if ((it->it_op & IT_OPEN) && dentry->d_inode &&
650 !S_ISREG(dentry->d_inode->i_mode) &&
651 !S_ISDIR(dentry->d_inode->i_mode)) {
652 ll_release_openhandle(dentry, it);
654 ll_lookup_finish_locks(it, dentry);
656 GOTO(out, retval = (dentry == save) ? NULL : dentry);
659 if (op_data != NULL && !IS_ERR(op_data)) {
660 if (secctx != NULL && secctxlen != NULL) {
661 /* caller needs sec ctx info, so reset it in op_data to
662 * prevent it from being freed */
663 op_data->op_file_secctx = NULL;
664 op_data->op_file_secctx_size = 0;
666 ll_finish_md_op_data(op_data);
669 ptlrpc_req_finished(req);
673 #ifdef HAVE_IOP_ATOMIC_OPEN
674 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
677 struct lookup_intent *itp, it = { .it_op = IT_GETATTR };
680 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), flags=%u\n",
681 dentry->d_name.len, dentry->d_name.name,
682 PFID(ll_inode2fid(parent)), parent, flags);
685 * Optimize away (CREATE && !OPEN). Let .create handle the race.
686 * but only if we have write permissions there, otherwise we need
687 * to proceed with lookup. LU-4185
689 if ((flags & LOOKUP_CREATE) && !(flags & LOOKUP_OPEN) &&
690 (inode_permission(parent, MAY_WRITE | MAY_EXEC) == 0))
693 if (flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE))
697 de = ll_lookup_it(parent, dentry, itp, NULL, NULL);
700 ll_intent_release(itp);
706 * For cached negative dentry and new dentry, handle lookup/create/open
709 static int ll_atomic_open(struct inode *dir, struct dentry *dentry,
710 struct file *file, unsigned open_flags,
711 umode_t mode, int *opened)
713 struct lookup_intent *it;
715 long long lookup_flags = LOOKUP_OPEN;
721 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), file %p,"
722 "open_flags %x, mode %x opened %d\n",
723 dentry->d_name.len, dentry->d_name.name,
724 PFID(ll_inode2fid(dir)), dir, file, open_flags, mode, *opened);
726 /* Only negative dentries enter here */
727 LASSERT(dentry->d_inode == NULL);
729 if (!d_unhashed(dentry)) {
730 /* A valid negative dentry that just passed revalidation,
731 * there's little point to try and open it server-side,
732 * even though there's a minuscule chance it might succeed.
733 * Either way it's a valid race to just return -ENOENT here.
735 if (!(open_flags & O_CREAT))
738 /* Otherwise we just unhash it to be rehashed afresh via
739 * lookup if necessary
744 OBD_ALLOC(it, sizeof(*it));
749 if (open_flags & O_CREAT) {
750 it->it_op |= IT_CREAT;
751 lookup_flags |= LOOKUP_CREATE;
753 it->it_create_mode = (mode & S_IALLUGO) | S_IFREG;
754 it->it_flags = (open_flags & ~O_ACCMODE) | OPEN_FMODE(open_flags);
755 it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
757 /* Dentry added to dcache tree in ll_lookup_it */
758 de = ll_lookup_it(dir, dentry, it, &secctx, &secctxlen);
764 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
767 if (it_disposition(it, DISP_OPEN_CREATE)) {
768 /* Dentry instantiated in ll_create_it. */
769 rc = ll_create_it(dir, dentry, it, secctx, secctxlen);
770 security_release_secctx(secctx, secctxlen);
772 /* We dget in ll_splice_alias. */
778 *opened |= FILE_CREATED;
780 if (dentry->d_inode && it_disposition(it, DISP_OPEN_OPEN)) {
782 if (S_ISFIFO(dentry->d_inode->i_mode)) {
783 /* We cannot call open here as it might
784 * deadlock. This case is unreachable in
785 * practice because of OBD_CONNECT_NODEVOH. */
786 rc = finish_no_open(file, de);
788 file->private_data = it;
789 rc = finish_open(file, dentry, NULL, opened);
790 /* We dget in ll_splice_alias. finish_open takes
791 * care of dget for fd open.
797 rc = finish_no_open(file, de);
802 ll_intent_release(it);
803 OBD_FREE(it, sizeof(*it));
808 #else /* !HAVE_IOP_ATOMIC_OPEN */
809 static struct lookup_intent *
810 ll_convert_intent(struct open_intent *oit, int lookup_flags, bool is_readonly)
812 struct lookup_intent *it;
816 return ERR_PTR(-ENOMEM);
818 if (lookup_flags & LOOKUP_OPEN) {
820 /* Avoid file creation for ro bind mount point(is_readonly) */
821 if ((lookup_flags & LOOKUP_CREATE) && !is_readonly)
822 it->it_op |= IT_CREAT;
823 it->it_create_mode = (oit->create_mode & S_IALLUGO) | S_IFREG;
824 it->it_flags = ll_namei_to_lookup_intent_flag(oit->flags &
825 ~(is_readonly ? O_CREAT : 0));
826 it->it_flags &= ~MDS_OPEN_FL_INTERNAL;
828 it->it_op = IT_GETATTR;
834 static struct dentry *ll_lookup_nd(struct inode *parent, struct dentry *dentry,
835 struct nameidata *nd)
840 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
841 struct lookup_intent *it;
843 if (ll_d2d(dentry) && ll_d2d(dentry)->lld_it) {
844 it = ll_d2d(dentry)->lld_it;
845 ll_d2d(dentry)->lld_it = NULL;
848 * Optimize away (CREATE && !OPEN). Let .create handle
849 * the race. But only if we have write permissions
850 * there, otherwise we need to proceed with lookup.
853 if ((nd->flags & LOOKUP_CREATE) &&
854 !(nd->flags & LOOKUP_OPEN) &&
855 (inode_permission(parent,
856 MAY_WRITE | MAY_EXEC) == 0))
859 it = ll_convert_intent(&nd->intent.open, nd->flags,
860 (nd->path.mnt->mnt_flags & MNT_READONLY) ||
861 (nd->path.mnt->mnt_sb->s_flags & MS_RDONLY));
863 RETURN((struct dentry *)it);
866 de = ll_lookup_it(parent, dentry, it, NULL, NULL);
869 if ((nd->flags & LOOKUP_OPEN) && !IS_ERR(dentry)) { /* Open */
870 if (dentry->d_inode &&
871 it_disposition(it, DISP_OPEN_OPEN)) { /* nocreate */
872 if (S_ISFIFO(dentry->d_inode->i_mode)) {
873 /* We cannot call open here as it might
874 * deadlock. This case is unreachable in
875 * practice because of
876 * OBD_CONNECT_NODEVOH. */
880 nd->intent.open.file->private_data = it;
881 filp = lookup_instantiate_filp(nd,
887 de = (struct dentry *)filp;
890 } else if (it_disposition(it, DISP_OPEN_CREATE)) {
891 /* XXX This can only reliably work on assumption
892 * that there are NO hashed negative dentries.*/
893 ll_d2d(dentry)->lld_it = it;
894 it = NULL; /* Will be freed in ll_create_nd */
895 /* We absolutely depend on ll_create_nd to be
896 * called to not leak this intent and possible
897 * data attached to it */
902 ll_intent_release(it);
903 OBD_FREE(it, sizeof(*it));
906 de = ll_lookup_it(parent, dentry, NULL, NULL, NULL);
911 #endif /* HAVE_IOP_ATOMIC_OPEN */
913 /* We depend on "mode" being set with the proper file type/umask by now */
914 static struct inode *ll_create_node(struct inode *dir, struct lookup_intent *it)
916 struct inode *inode = NULL;
917 struct ptlrpc_request *request = NULL;
918 struct ll_sb_info *sbi = ll_i2sbi(dir);
922 LASSERT(it && it->it_disposition);
924 LASSERT(it_disposition(it, DISP_ENQ_CREATE_REF));
925 request = it->it_request;
926 it_clear_disposition(it, DISP_ENQ_CREATE_REF);
927 rc = ll_prep_inode(&inode, request, dir->i_sb, it);
929 GOTO(out, inode = ERR_PTR(rc));
931 /* Pause to allow for a race with concurrent access by fid */
932 OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_NODE_PAUSE, cfs_fail_val);
934 /* We asked for a lock on the directory, but were granted a
935 * lock on the inode. Since we finally have an inode pointer,
936 * stuff it in the lock. */
937 CDEBUG(D_DLMTRACE, "setting l_ast_data to inode "DFID"(%p)\n",
938 PFID(ll_inode2fid(inode)), inode);
939 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
942 ptlrpc_req_finished(request);
947 * By the time this is called, we already have created the directory cache
948 * entry for the new file, but it is so far negative - it has no inode.
950 * We defer creating the OBD object(s) until open, to keep the intent and
951 * non-intent code paths similar, and also because we do not have the MDS
952 * inode number before calling ll_create_node() (which is needed for LOV),
953 * so we would need to do yet another RPC to the MDS to store the LOV EA
954 * data on the MDS. If needed, we would pass the PACKED lmm as data and
955 * lmm_size in datalen (the MDS still has code which will handle that).
957 * If the create succeeds, we fill in the inode information
958 * with d_instantiate().
960 static int ll_create_it(struct inode *dir, struct dentry *dentry,
961 struct lookup_intent *it,
962 void *secctx, __u32 secctxlen)
968 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), intent=%s\n",
969 dentry->d_name.len, dentry->d_name.name,
970 PFID(ll_inode2fid(dir)), dir, LL_IT2STR(it));
972 rc = it_open_error(DISP_OPEN_CREATE, it);
976 inode = ll_create_node(dir, it);
978 RETURN(PTR_ERR(inode));
980 if ((ll_i2sbi(inode)->ll_flags & LL_SBI_FILE_SECCTX) &&
983 /* must be done before d_instantiate, because it calls
984 * security_d_instantiate, which means a getxattr if security
985 * context is not set yet */
986 rc = security_inode_notifysecctx(inode, secctx, secctxlen);
992 d_instantiate(dentry, inode);
994 if (!(ll_i2sbi(inode)->ll_flags & LL_SBI_FILE_SECCTX)) {
995 rc = ll_inode_init_security(dentry, inode, dir);
1003 void ll_update_times(struct ptlrpc_request *request, struct inode *inode)
1005 struct mdt_body *body = req_capsule_server_get(&request->rq_pill,
1009 if (body->mbo_valid & OBD_MD_FLMTIME &&
1010 body->mbo_mtime > LTIME_S(inode->i_mtime)) {
1011 CDEBUG(D_INODE, "setting fid "DFID" mtime from %lu to %llu"
1012 "\n", PFID(ll_inode2fid(inode)),
1013 LTIME_S(inode->i_mtime), body->mbo_mtime);
1014 LTIME_S(inode->i_mtime) = body->mbo_mtime;
1017 if (body->mbo_valid & OBD_MD_FLCTIME &&
1018 body->mbo_ctime > LTIME_S(inode->i_ctime))
1019 LTIME_S(inode->i_ctime) = body->mbo_ctime;
1022 static int ll_new_node(struct inode *dir, struct dentry *dchild,
1023 const char *tgt, umode_t mode, int rdev, __u32 opc)
1025 struct qstr *name = &dchild->d_name;
1026 struct ptlrpc_request *request = NULL;
1027 struct md_op_data *op_data;
1028 struct inode *inode = NULL;
1029 struct ll_sb_info *sbi = ll_i2sbi(dir);
1034 if (unlikely(tgt != NULL))
1035 tgt_len = strlen(tgt) + 1;
1038 op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name,
1039 name->len, 0, opc, NULL);
1040 if (IS_ERR(op_data))
1041 GOTO(err_exit, err = PTR_ERR(op_data));
1043 if (sbi->ll_flags & LL_SBI_FILE_SECCTX) {
1044 err = ll_dentry_init_security(dchild, mode, &dchild->d_name,
1045 &op_data->op_file_secctx_name,
1046 &op_data->op_file_secctx,
1047 &op_data->op_file_secctx_size);
1049 GOTO(err_exit, err);
1052 err = md_create(sbi->ll_md_exp, op_data, tgt, tgt_len, mode,
1053 from_kuid(&init_user_ns, current_fsuid()),
1054 from_kgid(&init_user_ns, current_fsgid()),
1055 cfs_curproc_cap_pack(), rdev, &request);
1056 if (err < 0 && err != -EREMOTE)
1057 GOTO(err_exit, err);
1059 /* If the client doesn't know where to create a subdirectory (or
1060 * in case of a race that sends the RPC to the wrong MDS), the
1061 * MDS will return -EREMOTE and the client will fetch the layout
1062 * of the directory, then create the directory on the right MDT. */
1063 if (unlikely(err == -EREMOTE)) {
1064 struct ll_inode_info *lli = ll_i2info(dir);
1065 struct lmv_user_md *lum;
1069 ptlrpc_req_finished(request);
1072 err2 = ll_dir_getstripe(dir, (void **)&lum, &lumsize, &request,
1073 OBD_MD_DEFAULT_MEA);
1075 /* Update stripe_offset and retry */
1076 lli->lli_def_stripe_offset = lum->lum_stripe_offset;
1077 } else if (err2 == -ENODATA &&
1078 lli->lli_def_stripe_offset != -1) {
1079 /* If there are no default stripe EA on the MDT, but the
1080 * client has default stripe, then it probably means
1081 * default stripe EA has just been deleted. */
1082 lli->lli_def_stripe_offset = -1;
1084 GOTO(err_exit, err);
1087 ptlrpc_req_finished(request);
1089 ll_finish_md_op_data(op_data);
1093 ll_update_times(request, dir);
1095 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_NEWNODE_PAUSE, cfs_fail_val);
1097 err = ll_prep_inode(&inode, request, dchild->d_sb, NULL);
1099 GOTO(err_exit, err);
1101 if (sbi->ll_flags & LL_SBI_FILE_SECCTX) {
1103 /* must be done before d_instantiate, because it calls
1104 * security_d_instantiate, which means a getxattr if security
1105 * context is not set yet */
1106 err = security_inode_notifysecctx(inode,
1107 op_data->op_file_secctx,
1108 op_data->op_file_secctx_size);
1109 inode_unlock(inode);
1111 GOTO(err_exit, err);
1114 d_instantiate(dchild, inode);
1116 if (!(sbi->ll_flags & LL_SBI_FILE_SECCTX)) {
1117 err = ll_inode_init_security(dchild, inode, dir);
1119 GOTO(err_exit, err);
1124 if (request != NULL)
1125 ptlrpc_req_finished(request);
1127 if (!IS_ERR_OR_NULL(op_data))
1128 ll_finish_md_op_data(op_data);
1133 static int ll_mknod(struct inode *dir, struct dentry *dchild, ll_umode_t mode,
1136 struct qstr *name = &dchild->d_name;
1140 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p) mode %o dev %x\n",
1141 name->len, name->name, PFID(ll_inode2fid(dir)), dir,
1144 if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1145 mode &= ~current_umask();
1147 switch (mode & S_IFMT) {
1149 mode |= S_IFREG; /* for mode = 0 case, fallthrough */
1155 err = ll_new_node(dir, dchild, NULL, mode, old_encode_dev(rdev),
1166 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKNOD, 1);
1171 #ifdef HAVE_IOP_ATOMIC_OPEN
1173 * Plain create. Intent create is handled in atomic_open.
1175 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
1176 umode_t mode, bool want_excl)
1180 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1182 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), "
1183 "flags=%u, excl=%d\n", dentry->d_name.len,
1184 dentry->d_name.name, PFID(ll_inode2fid(dir)),
1185 dir, mode, want_excl);
1187 /* Using mknod(2) to create a regular file is designed to not recognize
1188 * volatile file name, so we use ll_mknod() here. */
1189 rc = ll_mknod(dir, dentry, mode, 0);
1191 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE, 1);
1193 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, unhashed %d\n",
1194 dentry->d_name.len, dentry->d_name.name, d_unhashed(dentry));
1198 #else /* !HAVE_IOP_ATOMIC_OPEN */
1199 static int ll_create_nd(struct inode *dir, struct dentry *dentry,
1200 ll_umode_t mode, struct nameidata *nd)
1202 struct ll_dentry_data *lld = ll_d2d(dentry);
1203 struct lookup_intent *it = NULL;
1206 CFS_FAIL_TIMEOUT(OBD_FAIL_LLITE_CREATE_FILE_PAUSE, cfs_fail_val);
1212 /* LU-8559: use LUSTRE_OPC_CREATE for non atomic open case
1213 * so that volatile file name is recoginized.
1214 * Mknod(2), however, is designed to not recognize volatile
1215 * file name to avoid inode leak under orphan directory until
1217 return ll_new_node(dir, dentry, NULL, mode, 0,
1223 /* Was there an error? Propagate it! */
1224 if (it->it_status) {
1229 rc = ll_create_it(dir, dentry, it, NULL, 0);
1230 if (nd && (nd->flags & LOOKUP_OPEN) && dentry->d_inode) { /* Open */
1233 nd->intent.open.file->private_data = it;
1234 filp = lookup_instantiate_filp(nd, dentry, NULL);
1240 ll_intent_release(it);
1241 OBD_FREE(it, sizeof(*it));
1244 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_CREATE, 1);
1248 #endif /* HAVE_IOP_ATOMIC_OPEN */
1250 static int ll_symlink(struct inode *dir, struct dentry *dchild,
1251 const char *oldpath)
1253 struct qstr *name = &dchild->d_name;
1257 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p), target=%.*s\n",
1258 name->len, name->name, PFID(ll_inode2fid(dir)),
1259 dir, 3000, oldpath);
1261 err = ll_new_node(dir, dchild, oldpath, S_IFLNK | S_IRWXUGO, 0,
1262 LUSTRE_OPC_SYMLINK);
1265 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_SYMLINK, 1);
1270 static int ll_link(struct dentry *old_dentry, struct inode *dir,
1271 struct dentry *new_dentry)
1273 struct inode *src = old_dentry->d_inode;
1274 struct qstr *name = &new_dentry->d_name;
1275 struct ll_sb_info *sbi = ll_i2sbi(dir);
1276 struct ptlrpc_request *request = NULL;
1277 struct md_op_data *op_data;
1281 CDEBUG(D_VFSTRACE, "VFS Op: inode="DFID"(%p), dir="DFID"(%p), "
1282 "target=%.*s\n", PFID(ll_inode2fid(src)), src,
1283 PFID(ll_inode2fid(dir)), dir, name->len, name->name);
1285 op_data = ll_prep_md_op_data(NULL, src, dir, name->name, name->len,
1286 0, LUSTRE_OPC_ANY, NULL);
1287 if (IS_ERR(op_data))
1288 RETURN(PTR_ERR(op_data));
1290 err = md_link(sbi->ll_md_exp, op_data, &request);
1291 ll_finish_md_op_data(op_data);
1295 ll_update_times(request, dir);
1296 ll_stats_ops_tally(sbi, LPROC_LL_LINK, 1);
1299 ptlrpc_req_finished(request);
1303 static int ll_mkdir(struct inode *dir, struct dentry *dchild, ll_umode_t mode)
1305 struct qstr *name = &dchild->d_name;
1309 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1310 name->len, name->name, PFID(ll_inode2fid(dir)), dir);
1312 if (!IS_POSIXACL(dir) || !exp_connect_umask(ll_i2mdexp(dir)))
1313 mode &= ~current_umask();
1315 mode = (mode & (S_IRWXUGO|S_ISVTX)) | S_IFDIR;
1317 err = ll_new_node(dir, dchild, NULL, mode, 0, LUSTRE_OPC_MKDIR);
1319 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_MKDIR, 1);
1324 static int ll_rmdir(struct inode *dir, struct dentry *dchild)
1326 struct qstr *name = &dchild->d_name;
1327 struct ptlrpc_request *request = NULL;
1328 struct md_op_data *op_data;
1332 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1333 name->len, name->name, PFID(ll_inode2fid(dir)), dir);
1335 if (unlikely(d_mountpoint(dchild)))
1338 op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len,
1339 S_IFDIR, LUSTRE_OPC_ANY, NULL);
1340 if (IS_ERR(op_data))
1341 RETURN(PTR_ERR(op_data));
1343 if (dchild->d_inode != NULL)
1344 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1346 op_data->op_fid2 = op_data->op_fid3;
1347 rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1348 ll_finish_md_op_data(op_data);
1350 ll_update_times(request, dir);
1351 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR, 1);
1354 ptlrpc_req_finished(request);
1361 int ll_rmdir_entry(struct inode *dir, char *name, int namelen)
1363 struct ptlrpc_request *request = NULL;
1364 struct md_op_data *op_data;
1368 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1369 namelen, name, PFID(ll_inode2fid(dir)), dir);
1371 op_data = ll_prep_md_op_data(NULL, dir, NULL, name, strlen(name),
1372 S_IFDIR, LUSTRE_OPC_ANY, NULL);
1373 if (IS_ERR(op_data))
1374 RETURN(PTR_ERR(op_data));
1375 op_data->op_cli_flags |= CLI_RM_ENTRY;
1376 rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1377 ll_finish_md_op_data(op_data);
1379 ll_update_times(request, dir);
1380 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_RMDIR, 1);
1383 ptlrpc_req_finished(request);
1387 static int ll_unlink(struct inode *dir, struct dentry *dchild)
1389 struct qstr *name = &dchild->d_name;
1390 struct ptlrpc_request *request = NULL;
1391 struct md_op_data *op_data;
1394 CDEBUG(D_VFSTRACE, "VFS Op:name=%.*s, dir="DFID"(%p)\n",
1395 name->len, name->name, PFID(ll_inode2fid(dir)), dir);
1398 * XXX: unlink bind mountpoint maybe call to here,
1399 * just check it as vfs_unlink does.
1401 if (unlikely(d_mountpoint(dchild)))
1404 op_data = ll_prep_md_op_data(NULL, dir, NULL, name->name, name->len, 0,
1405 LUSTRE_OPC_ANY, NULL);
1406 if (IS_ERR(op_data))
1407 RETURN(PTR_ERR(op_data));
1409 if (dchild->d_inode != NULL)
1410 op_data->op_fid3 = *ll_inode2fid(dchild->d_inode);
1412 op_data->op_fid2 = op_data->op_fid3;
1413 rc = md_unlink(ll_i2sbi(dir)->ll_md_exp, op_data, &request);
1414 ll_finish_md_op_data(op_data);
1418 ll_update_times(request, dir);
1419 ll_stats_ops_tally(ll_i2sbi(dir), LPROC_LL_UNLINK, 1);
1422 ptlrpc_req_finished(request);
1426 static int ll_rename(struct inode *src, struct dentry *src_dchild,
1427 struct inode *tgt, struct dentry *tgt_dchild
1428 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
1429 , unsigned int flags
1433 struct qstr *src_name = &src_dchild->d_name;
1434 struct qstr *tgt_name = &tgt_dchild->d_name;
1435 struct ptlrpc_request *request = NULL;
1436 struct ll_sb_info *sbi = ll_i2sbi(src);
1437 struct md_op_data *op_data;
1441 #ifdef HAVE_IOPS_RENAME_WITH_FLAGS
1446 CDEBUG(D_VFSTRACE, "VFS Op:oldname=%.*s, src_dir="DFID
1447 "(%p), newname=%.*s, tgt_dir="DFID"(%p)\n",
1448 src_name->len, src_name->name,
1449 PFID(ll_inode2fid(src)), src, tgt_name->len,
1450 tgt_name->name, PFID(ll_inode2fid(tgt)), tgt);
1452 if (unlikely(d_mountpoint(src_dchild) || d_mountpoint(tgt_dchild)))
1455 op_data = ll_prep_md_op_data(NULL, src, tgt, NULL, 0, 0,
1456 LUSTRE_OPC_ANY, NULL);
1457 if (IS_ERR(op_data))
1458 RETURN(PTR_ERR(op_data));
1460 if (src_dchild->d_inode != NULL)
1461 op_data->op_fid3 = *ll_inode2fid(src_dchild->d_inode);
1463 if (tgt_dchild->d_inode != NULL)
1464 op_data->op_fid4 = *ll_inode2fid(tgt_dchild->d_inode);
1466 err = md_rename(sbi->ll_md_exp, op_data,
1467 src_name->name, src_name->len,
1468 tgt_name->name, tgt_name->len, &request);
1469 ll_finish_md_op_data(op_data);
1471 ll_update_times(request, src);
1472 ll_update_times(request, tgt);
1473 ll_stats_ops_tally(sbi, LPROC_LL_RENAME, 1);
1476 ptlrpc_req_finished(request);
1479 d_move(src_dchild, tgt_dchild);
1484 const struct inode_operations ll_dir_inode_operations = {
1486 #ifdef HAVE_IOP_ATOMIC_OPEN
1487 .atomic_open = ll_atomic_open,
1489 .lookup = ll_lookup_nd,
1490 .create = ll_create_nd,
1491 /* We need all these non-raw things for NFSD, to not patch it. */
1492 .unlink = ll_unlink,
1495 .symlink = ll_symlink,
1497 .rename = ll_rename,
1498 .setattr = ll_setattr,
1499 .getattr = ll_getattr,
1500 .permission = ll_inode_permission,
1501 #ifdef HAVE_IOP_XATTR
1502 .setxattr = ll_setxattr,
1503 .getxattr = ll_getxattr,
1504 .removexattr = ll_removexattr,
1506 .listxattr = ll_listxattr,
1507 #ifdef HAVE_IOP_GET_ACL
1508 .get_acl = ll_get_acl,
1510 #ifdef HAVE_IOP_SET_ACL
1511 .set_acl = ll_set_acl,
1515 const struct inode_operations ll_special_inode_operations = {
1516 .setattr = ll_setattr,
1517 .getattr = ll_getattr,
1518 .permission = ll_inode_permission,
1519 #ifdef HAVE_IOP_XATTR
1520 .setxattr = ll_setxattr,
1521 .getxattr = ll_getxattr,
1522 .removexattr = ll_removexattr,
1524 .listxattr = ll_listxattr,
1525 #ifdef HAVE_IOP_GET_ACL
1526 .get_acl = ll_get_acl,
1528 #ifdef HAVE_IOP_SET_ACL
1529 .set_acl = ll_set_acl,