Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_acl.h"
15#include "xfs_quota.h"
16#include "xfs_da_format.h"
17#include "xfs_da_btree.h"
18#include "xfs_attr.h"
19#include "xfs_trans.h"
20#include "xfs_trace.h"
21#include "xfs_icache.h"
22#include "xfs_symlink.h"
23#include "xfs_dir2.h"
24#include "xfs_iomap.h"
25#include "xfs_error.h"
26#include "xfs_ioctl.h"
27#include "xfs_xattr.h"
28
29#include <linux/posix_acl.h>
30#include <linux/security.h>
31#include <linux/iversion.h>
32#include <linux/fiemap.h>
33
34/*
35 * Directories have different lock order w.r.t. mmap_lock compared to regular
36 * files. This is due to readdir potentially triggering page faults on a user
37 * buffer inside filldir(), and this happens with the ilock on the directory
38 * held. For regular files, the lock order is the other way around - the
39 * mmap_lock is taken during the page fault, and then we lock the ilock to do
40 * block mapping. Hence we need a different class for the directory ilock so
41 * that lockdep can tell them apart.
42 */
43static struct lock_class_key xfs_nondir_ilock_class;
44static struct lock_class_key xfs_dir_ilock_class;
45
46static int
47xfs_initxattrs(
48 struct inode *inode,
49 const struct xattr *xattr_array,
50 void *fs_info)
51{
52 const struct xattr *xattr;
53 struct xfs_inode *ip = XFS_I(inode);
54 int error = 0;
55
56 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
57 struct xfs_da_args args = {
58 .dp = ip,
59 .attr_filter = XFS_ATTR_SECURE,
60 .name = xattr->name,
61 .namelen = strlen(xattr->name),
62 .value = xattr->value,
63 .valuelen = xattr->value_len,
64 };
65 error = xfs_attr_change(&args);
66 if (error < 0)
67 break;
68 }
69 return error;
70}
71
72/*
73 * Hook in SELinux. This is not quite correct yet, what we really need
74 * here (as we do for default ACLs) is a mechanism by which creation of
75 * these attrs can be journalled at inode creation time (along with the
76 * inode, of course, such that log replay can't cause these to be lost).
77 */
78int
79xfs_inode_init_security(
80 struct inode *inode,
81 struct inode *dir,
82 const struct qstr *qstr)
83{
84 return security_inode_init_security(inode, dir, qstr,
85 &xfs_initxattrs, NULL);
86}
87
88static void
89xfs_dentry_to_name(
90 struct xfs_name *namep,
91 struct dentry *dentry)
92{
93 namep->name = dentry->d_name.name;
94 namep->len = dentry->d_name.len;
95 namep->type = XFS_DIR3_FT_UNKNOWN;
96}
97
98static int
99xfs_dentry_mode_to_name(
100 struct xfs_name *namep,
101 struct dentry *dentry,
102 int mode)
103{
104 namep->name = dentry->d_name.name;
105 namep->len = dentry->d_name.len;
106 namep->type = xfs_mode_to_ftype(mode);
107
108 if (unlikely(namep->type == XFS_DIR3_FT_UNKNOWN))
109 return -EFSCORRUPTED;
110
111 return 0;
112}
113
114STATIC void
115xfs_cleanup_inode(
116 struct inode *dir,
117 struct inode *inode,
118 struct dentry *dentry)
119{
120 struct xfs_name teardown;
121
122 /* Oh, the horror.
123 * If we can't add the ACL or we fail in
124 * xfs_inode_init_security we must back out.
125 * ENOSPC can hit here, among other things.
126 */
127 xfs_dentry_to_name(&teardown, dentry);
128
129 xfs_remove(XFS_I(dir), &teardown, XFS_I(inode));
130}
131
132/*
133 * Check to see if we are likely to need an extended attribute to be added to
134 * the inode we are about to allocate. This allows the attribute fork to be
135 * created during the inode allocation, reducing the number of transactions we
136 * need to do in this fast path.
137 *
138 * The security checks are optimistic, but not guaranteed. The two LSMs that
139 * require xattrs to be added here (selinux and smack) are also the only two
140 * LSMs that add a sb->s_security structure to the superblock. Hence if security
141 * is enabled and sb->s_security is set, we have a pretty good idea that we are
142 * going to be asked to add a security xattr immediately after allocating the
143 * xfs inode and instantiating the VFS inode.
144 */
145static inline bool
146xfs_create_need_xattr(
147 struct inode *dir,
148 struct posix_acl *default_acl,
149 struct posix_acl *acl)
150{
151 if (acl)
152 return true;
153 if (default_acl)
154 return true;
155#if IS_ENABLED(CONFIG_SECURITY)
156 if (dir->i_sb->s_security)
157 return true;
158#endif
159 return false;
160}
161
162
163STATIC int
164xfs_generic_create(
165 struct user_namespace *mnt_userns,
166 struct inode *dir,
167 struct dentry *dentry,
168 umode_t mode,
169 dev_t rdev,
170 struct file *tmpfile) /* unnamed file */
171{
172 struct inode *inode;
173 struct xfs_inode *ip = NULL;
174 struct posix_acl *default_acl, *acl;
175 struct xfs_name name;
176 int error;
177
178 /*
179 * Irix uses Missed'em'V split, but doesn't want to see
180 * the upper 5 bits of (14bit) major.
181 */
182 if (S_ISCHR(mode) || S_ISBLK(mode)) {
183 if (unlikely(!sysv_valid_dev(rdev) || MAJOR(rdev) & ~0x1ff))
184 return -EINVAL;
185 } else {
186 rdev = 0;
187 }
188
189 error = posix_acl_create(dir, &mode, &default_acl, &acl);
190 if (error)
191 return error;
192
193 /* Verify mode is valid also for tmpfile case */
194 error = xfs_dentry_mode_to_name(&name, dentry, mode);
195 if (unlikely(error))
196 goto out_free_acl;
197
198 if (!tmpfile) {
199 error = xfs_create(mnt_userns, XFS_I(dir), &name, mode, rdev,
200 xfs_create_need_xattr(dir, default_acl, acl),
201 &ip);
202 } else {
203 error = xfs_create_tmpfile(mnt_userns, XFS_I(dir), mode, &ip);
204 }
205 if (unlikely(error))
206 goto out_free_acl;
207
208 inode = VFS_I(ip);
209
210 error = xfs_inode_init_security(inode, dir, &dentry->d_name);
211 if (unlikely(error))
212 goto out_cleanup_inode;
213
214 if (default_acl) {
215 error = __xfs_set_acl(inode, default_acl, ACL_TYPE_DEFAULT);
216 if (error)
217 goto out_cleanup_inode;
218 }
219 if (acl) {
220 error = __xfs_set_acl(inode, acl, ACL_TYPE_ACCESS);
221 if (error)
222 goto out_cleanup_inode;
223 }
224
225 xfs_setup_iops(ip);
226
227 if (tmpfile) {
228 /*
229 * The VFS requires that any inode fed to d_tmpfile must have
230 * nlink == 1 so that it can decrement the nlink in d_tmpfile.
231 * However, we created the temp file with nlink == 0 because
232 * we're not allowed to put an inode with nlink > 0 on the
233 * unlinked list. Therefore we have to set nlink to 1 so that
234 * d_tmpfile can immediately set it back to zero.
235 */
236 set_nlink(inode, 1);
237 d_tmpfile(tmpfile, inode);
238 } else
239 d_instantiate(dentry, inode);
240
241 xfs_finish_inode_setup(ip);
242
243 out_free_acl:
244 posix_acl_release(default_acl);
245 posix_acl_release(acl);
246 return error;
247
248 out_cleanup_inode:
249 xfs_finish_inode_setup(ip);
250 if (!tmpfile)
251 xfs_cleanup_inode(dir, inode, dentry);
252 xfs_irele(ip);
253 goto out_free_acl;
254}
255
256STATIC int
257xfs_vn_mknod(
258 struct user_namespace *mnt_userns,
259 struct inode *dir,
260 struct dentry *dentry,
261 umode_t mode,
262 dev_t rdev)
263{
264 return xfs_generic_create(mnt_userns, dir, dentry, mode, rdev, NULL);
265}
266
267STATIC int
268xfs_vn_create(
269 struct user_namespace *mnt_userns,
270 struct inode *dir,
271 struct dentry *dentry,
272 umode_t mode,
273 bool flags)
274{
275 return xfs_generic_create(mnt_userns, dir, dentry, mode, 0, NULL);
276}
277
278STATIC int
279xfs_vn_mkdir(
280 struct user_namespace *mnt_userns,
281 struct inode *dir,
282 struct dentry *dentry,
283 umode_t mode)
284{
285 return xfs_generic_create(mnt_userns, dir, dentry, mode | S_IFDIR, 0,
286 NULL);
287}
288
289STATIC struct dentry *
290xfs_vn_lookup(
291 struct inode *dir,
292 struct dentry *dentry,
293 unsigned int flags)
294{
295 struct inode *inode;
296 struct xfs_inode *cip;
297 struct xfs_name name;
298 int error;
299
300 if (dentry->d_name.len >= MAXNAMELEN)
301 return ERR_PTR(-ENAMETOOLONG);
302
303 xfs_dentry_to_name(&name, dentry);
304 error = xfs_lookup(XFS_I(dir), &name, &cip, NULL);
305 if (likely(!error))
306 inode = VFS_I(cip);
307 else if (likely(error == -ENOENT))
308 inode = NULL;
309 else
310 inode = ERR_PTR(error);
311 return d_splice_alias(inode, dentry);
312}
313
314STATIC struct dentry *
315xfs_vn_ci_lookup(
316 struct inode *dir,
317 struct dentry *dentry,
318 unsigned int flags)
319{
320 struct xfs_inode *ip;
321 struct xfs_name xname;
322 struct xfs_name ci_name;
323 struct qstr dname;
324 int error;
325
326 if (dentry->d_name.len >= MAXNAMELEN)
327 return ERR_PTR(-ENAMETOOLONG);
328
329 xfs_dentry_to_name(&xname, dentry);
330 error = xfs_lookup(XFS_I(dir), &xname, &ip, &ci_name);
331 if (unlikely(error)) {
332 if (unlikely(error != -ENOENT))
333 return ERR_PTR(error);
334 /*
335 * call d_add(dentry, NULL) here when d_drop_negative_children
336 * is called in xfs_vn_mknod (ie. allow negative dentries
337 * with CI filesystems).
338 */
339 return NULL;
340 }
341
342 /* if exact match, just splice and exit */
343 if (!ci_name.name)
344 return d_splice_alias(VFS_I(ip), dentry);
345
346 /* else case-insensitive match... */
347 dname.name = ci_name.name;
348 dname.len = ci_name.len;
349 dentry = d_add_ci(dentry, VFS_I(ip), &dname);
350 kmem_free(ci_name.name);
351 return dentry;
352}
353
354STATIC int
355xfs_vn_link(
356 struct dentry *old_dentry,
357 struct inode *dir,
358 struct dentry *dentry)
359{
360 struct inode *inode = d_inode(old_dentry);
361 struct xfs_name name;
362 int error;
363
364 error = xfs_dentry_mode_to_name(&name, dentry, inode->i_mode);
365 if (unlikely(error))
366 return error;
367
368 error = xfs_link(XFS_I(dir), XFS_I(inode), &name);
369 if (unlikely(error))
370 return error;
371
372 ihold(inode);
373 d_instantiate(dentry, inode);
374 return 0;
375}
376
377STATIC int
378xfs_vn_unlink(
379 struct inode *dir,
380 struct dentry *dentry)
381{
382 struct xfs_name name;
383 int error;
384
385 xfs_dentry_to_name(&name, dentry);
386
387 error = xfs_remove(XFS_I(dir), &name, XFS_I(d_inode(dentry)));
388 if (error)
389 return error;
390
391 /*
392 * With unlink, the VFS makes the dentry "negative": no inode,
393 * but still hashed. This is incompatible with case-insensitive
394 * mode, so invalidate (unhash) the dentry in CI-mode.
395 */
396 if (xfs_has_asciici(XFS_M(dir->i_sb)))
397 d_invalidate(dentry);
398 return 0;
399}
400
401STATIC int
402xfs_vn_symlink(
403 struct user_namespace *mnt_userns,
404 struct inode *dir,
405 struct dentry *dentry,
406 const char *symname)
407{
408 struct inode *inode;
409 struct xfs_inode *cip = NULL;
410 struct xfs_name name;
411 int error;
412 umode_t mode;
413
414 mode = S_IFLNK |
415 (irix_symlink_mode ? 0777 & ~current_umask() : S_IRWXUGO);
416 error = xfs_dentry_mode_to_name(&name, dentry, mode);
417 if (unlikely(error))
418 goto out;
419
420 error = xfs_symlink(mnt_userns, XFS_I(dir), &name, symname, mode, &cip);
421 if (unlikely(error))
422 goto out;
423
424 inode = VFS_I(cip);
425
426 error = xfs_inode_init_security(inode, dir, &dentry->d_name);
427 if (unlikely(error))
428 goto out_cleanup_inode;
429
430 xfs_setup_iops(cip);
431
432 d_instantiate(dentry, inode);
433 xfs_finish_inode_setup(cip);
434 return 0;
435
436 out_cleanup_inode:
437 xfs_finish_inode_setup(cip);
438 xfs_cleanup_inode(dir, inode, dentry);
439 xfs_irele(cip);
440 out:
441 return error;
442}
443
444STATIC int
445xfs_vn_rename(
446 struct user_namespace *mnt_userns,
447 struct inode *odir,
448 struct dentry *odentry,
449 struct inode *ndir,
450 struct dentry *ndentry,
451 unsigned int flags)
452{
453 struct inode *new_inode = d_inode(ndentry);
454 int omode = 0;
455 int error;
456 struct xfs_name oname;
457 struct xfs_name nname;
458
459 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
460 return -EINVAL;
461
462 /* if we are exchanging files, we need to set i_mode of both files */
463 if (flags & RENAME_EXCHANGE)
464 omode = d_inode(ndentry)->i_mode;
465
466 error = xfs_dentry_mode_to_name(&oname, odentry, omode);
467 if (omode && unlikely(error))
468 return error;
469
470 error = xfs_dentry_mode_to_name(&nname, ndentry,
471 d_inode(odentry)->i_mode);
472 if (unlikely(error))
473 return error;
474
475 return xfs_rename(mnt_userns, XFS_I(odir), &oname,
476 XFS_I(d_inode(odentry)), XFS_I(ndir), &nname,
477 new_inode ? XFS_I(new_inode) : NULL, flags);
478}
479
480/*
481 * careful here - this function can get called recursively, so
482 * we need to be very careful about how much stack we use.
483 * uio is kmalloced for this reason...
484 */
485STATIC const char *
486xfs_vn_get_link(
487 struct dentry *dentry,
488 struct inode *inode,
489 struct delayed_call *done)
490{
491 char *link;
492 int error = -ENOMEM;
493
494 if (!dentry)
495 return ERR_PTR(-ECHILD);
496
497 link = kmalloc(XFS_SYMLINK_MAXLEN+1, GFP_KERNEL);
498 if (!link)
499 goto out_err;
500
501 error = xfs_readlink(XFS_I(d_inode(dentry)), link);
502 if (unlikely(error))
503 goto out_kfree;
504
505 set_delayed_call(done, kfree_link, link);
506 return link;
507
508 out_kfree:
509 kfree(link);
510 out_err:
511 return ERR_PTR(error);
512}
513
514static uint32_t
515xfs_stat_blksize(
516 struct xfs_inode *ip)
517{
518 struct xfs_mount *mp = ip->i_mount;
519
520 /*
521 * If the file blocks are being allocated from a realtime volume, then
522 * always return the realtime extent size.
523 */
524 if (XFS_IS_REALTIME_INODE(ip))
525 return XFS_FSB_TO_B(mp, xfs_get_extsz_hint(ip));
526
527 /*
528 * Allow large block sizes to be reported to userspace programs if the
529 * "largeio" mount option is used.
530 *
531 * If compatibility mode is specified, simply return the basic unit of
532 * caching so that we don't get inefficient read/modify/write I/O from
533 * user apps. Otherwise....
534 *
535 * If the underlying volume is a stripe, then return the stripe width in
536 * bytes as the recommended I/O size. It is not a stripe and we've set a
537 * default buffered I/O size, return that, otherwise return the compat
538 * default.
539 */
540 if (xfs_has_large_iosize(mp)) {
541 if (mp->m_swidth)
542 return XFS_FSB_TO_B(mp, mp->m_swidth);
543 if (xfs_has_allocsize(mp))
544 return 1U << mp->m_allocsize_log;
545 }
546
547 return PAGE_SIZE;
548}
549
550STATIC int
551xfs_vn_getattr(
552 struct user_namespace *mnt_userns,
553 const struct path *path,
554 struct kstat *stat,
555 u32 request_mask,
556 unsigned int query_flags)
557{
558 struct inode *inode = d_inode(path->dentry);
559 struct xfs_inode *ip = XFS_I(inode);
560 struct xfs_mount *mp = ip->i_mount;
561 vfsuid_t vfsuid = i_uid_into_vfsuid(mnt_userns, inode);
562 vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
563
564 trace_xfs_getattr(ip);
565
566 if (xfs_is_shutdown(mp))
567 return -EIO;
568
569 stat->size = XFS_ISIZE(ip);
570 stat->dev = inode->i_sb->s_dev;
571 stat->mode = inode->i_mode;
572 stat->nlink = inode->i_nlink;
573 stat->uid = vfsuid_into_kuid(vfsuid);
574 stat->gid = vfsgid_into_kgid(vfsgid);
575 stat->ino = ip->i_ino;
576 stat->atime = inode->i_atime;
577 stat->mtime = inode->i_mtime;
578 stat->ctime = inode->i_ctime;
579 stat->blocks = XFS_FSB_TO_BB(mp, ip->i_nblocks + ip->i_delayed_blks);
580
581 if (xfs_has_v3inodes(mp)) {
582 if (request_mask & STATX_BTIME) {
583 stat->result_mask |= STATX_BTIME;
584 stat->btime = ip->i_crtime;
585 }
586 }
587
588 /*
589 * Note: If you add another clause to set an attribute flag, please
590 * update attributes_mask below.
591 */
592 if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
593 stat->attributes |= STATX_ATTR_IMMUTABLE;
594 if (ip->i_diflags & XFS_DIFLAG_APPEND)
595 stat->attributes |= STATX_ATTR_APPEND;
596 if (ip->i_diflags & XFS_DIFLAG_NODUMP)
597 stat->attributes |= STATX_ATTR_NODUMP;
598
599 stat->attributes_mask |= (STATX_ATTR_IMMUTABLE |
600 STATX_ATTR_APPEND |
601 STATX_ATTR_NODUMP);
602
603 switch (inode->i_mode & S_IFMT) {
604 case S_IFBLK:
605 case S_IFCHR:
606 stat->blksize = BLKDEV_IOSIZE;
607 stat->rdev = inode->i_rdev;
608 break;
609 case S_IFREG:
610 if (request_mask & STATX_DIOALIGN) {
611 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
612 struct block_device *bdev = target->bt_bdev;
613
614 stat->result_mask |= STATX_DIOALIGN;
615 stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
616 stat->dio_offset_align = bdev_logical_block_size(bdev);
617 }
618 fallthrough;
619 default:
620 stat->blksize = xfs_stat_blksize(ip);
621 stat->rdev = 0;
622 break;
623 }
624
625 return 0;
626}
627
628static int
629xfs_vn_change_ok(
630 struct user_namespace *mnt_userns,
631 struct dentry *dentry,
632 struct iattr *iattr)
633{
634 struct xfs_mount *mp = XFS_I(d_inode(dentry))->i_mount;
635
636 if (xfs_is_readonly(mp))
637 return -EROFS;
638
639 if (xfs_is_shutdown(mp))
640 return -EIO;
641
642 return setattr_prepare(mnt_userns, dentry, iattr);
643}
644
645/*
646 * Set non-size attributes of an inode.
647 *
648 * Caution: The caller of this function is responsible for calling
649 * setattr_prepare() or otherwise verifying the change is fine.
650 */
651static int
652xfs_setattr_nonsize(
653 struct user_namespace *mnt_userns,
654 struct dentry *dentry,
655 struct xfs_inode *ip,
656 struct iattr *iattr)
657{
658 xfs_mount_t *mp = ip->i_mount;
659 struct inode *inode = VFS_I(ip);
660 int mask = iattr->ia_valid;
661 xfs_trans_t *tp;
662 int error;
663 kuid_t uid = GLOBAL_ROOT_UID;
664 kgid_t gid = GLOBAL_ROOT_GID;
665 struct xfs_dquot *udqp = NULL, *gdqp = NULL;
666 struct xfs_dquot *old_udqp = NULL, *old_gdqp = NULL;
667
668 ASSERT((mask & ATTR_SIZE) == 0);
669
670 /*
671 * If disk quotas is on, we make sure that the dquots do exist on disk,
672 * before we start any other transactions. Trying to do this later
673 * is messy. We don't care to take a readlock to look at the ids
674 * in inode here, because we can't hold it across the trans_reserve.
675 * If the IDs do change before we take the ilock, we're covered
676 * because the i_*dquot fields will get updated anyway.
677 */
678 if (XFS_IS_QUOTA_ON(mp) && (mask & (ATTR_UID|ATTR_GID))) {
679 uint qflags = 0;
680
681 if ((mask & ATTR_UID) && XFS_IS_UQUOTA_ON(mp)) {
682 uid = from_vfsuid(mnt_userns, i_user_ns(inode),
683 iattr->ia_vfsuid);
684 qflags |= XFS_QMOPT_UQUOTA;
685 } else {
686 uid = inode->i_uid;
687 }
688 if ((mask & ATTR_GID) && XFS_IS_GQUOTA_ON(mp)) {
689 gid = from_vfsgid(mnt_userns, i_user_ns(inode),
690 iattr->ia_vfsgid);
691 qflags |= XFS_QMOPT_GQUOTA;
692 } else {
693 gid = inode->i_gid;
694 }
695
696 /*
697 * We take a reference when we initialize udqp and gdqp,
698 * so it is important that we never blindly double trip on
699 * the same variable. See xfs_create() for an example.
700 */
701 ASSERT(udqp == NULL);
702 ASSERT(gdqp == NULL);
703 error = xfs_qm_vop_dqalloc(ip, uid, gid, ip->i_projid,
704 qflags, &udqp, &gdqp, NULL);
705 if (error)
706 return error;
707 }
708
709 error = xfs_trans_alloc_ichange(ip, udqp, gdqp, NULL,
710 has_capability_noaudit(current, CAP_FOWNER), &tp);
711 if (error)
712 goto out_dqrele;
713
714 /*
715 * Register quota modifications in the transaction. Must be the owner
716 * or privileged. These IDs could have changed since we last looked at
717 * them. But, we're assured that if the ownership did change while we
718 * didn't have the inode locked, inode's dquot(s) would have changed
719 * also.
720 */
721 if (XFS_IS_UQUOTA_ON(mp) &&
722 i_uid_needs_update(mnt_userns, iattr, inode)) {
723 ASSERT(udqp);
724 old_udqp = xfs_qm_vop_chown(tp, ip, &ip->i_udquot, udqp);
725 }
726 if (XFS_IS_GQUOTA_ON(mp) &&
727 i_gid_needs_update(mnt_userns, iattr, inode)) {
728 ASSERT(xfs_has_pquotino(mp) || !XFS_IS_PQUOTA_ON(mp));
729 ASSERT(gdqp);
730 old_gdqp = xfs_qm_vop_chown(tp, ip, &ip->i_gdquot, gdqp);
731 }
732
733 setattr_copy(mnt_userns, inode, iattr);
734 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
735
736 XFS_STATS_INC(mp, xs_ig_attrchg);
737
738 if (xfs_has_wsync(mp))
739 xfs_trans_set_sync(tp);
740 error = xfs_trans_commit(tp);
741
742 /*
743 * Release any dquot(s) the inode had kept before chown.
744 */
745 xfs_qm_dqrele(old_udqp);
746 xfs_qm_dqrele(old_gdqp);
747 xfs_qm_dqrele(udqp);
748 xfs_qm_dqrele(gdqp);
749
750 if (error)
751 return error;
752
753 /*
754 * XXX(hch): Updating the ACL entries is not atomic vs the i_mode
755 * update. We could avoid this with linked transactions
756 * and passing down the transaction pointer all the way
757 * to attr_set. No previous user of the generic
758 * Posix ACL code seems to care about this issue either.
759 */
760 if (mask & ATTR_MODE) {
761 error = posix_acl_chmod(mnt_userns, dentry, inode->i_mode);
762 if (error)
763 return error;
764 }
765
766 return 0;
767
768out_dqrele:
769 xfs_qm_dqrele(udqp);
770 xfs_qm_dqrele(gdqp);
771 return error;
772}
773
774/*
775 * Truncate file. Must have write permission and not be a directory.
776 *
777 * Caution: The caller of this function is responsible for calling
778 * setattr_prepare() or otherwise verifying the change is fine.
779 */
780STATIC int
781xfs_setattr_size(
782 struct user_namespace *mnt_userns,
783 struct dentry *dentry,
784 struct xfs_inode *ip,
785 struct iattr *iattr)
786{
787 struct xfs_mount *mp = ip->i_mount;
788 struct inode *inode = VFS_I(ip);
789 xfs_off_t oldsize, newsize;
790 struct xfs_trans *tp;
791 int error;
792 uint lock_flags = 0;
793 bool did_zeroing = false;
794
795 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
796 ASSERT(xfs_isilocked(ip, XFS_MMAPLOCK_EXCL));
797 ASSERT(S_ISREG(inode->i_mode));
798 ASSERT((iattr->ia_valid & (ATTR_UID|ATTR_GID|ATTR_ATIME|ATTR_ATIME_SET|
799 ATTR_MTIME_SET|ATTR_TIMES_SET)) == 0);
800
801 oldsize = inode->i_size;
802 newsize = iattr->ia_size;
803
804 /*
805 * Short circuit the truncate case for zero length files.
806 */
807 if (newsize == 0 && oldsize == 0 && ip->i_df.if_nextents == 0) {
808 if (!(iattr->ia_valid & (ATTR_CTIME|ATTR_MTIME)))
809 return 0;
810
811 /*
812 * Use the regular setattr path to update the timestamps.
813 */
814 iattr->ia_valid &= ~ATTR_SIZE;
815 return xfs_setattr_nonsize(mnt_userns, dentry, ip, iattr);
816 }
817
818 /*
819 * Make sure that the dquots are attached to the inode.
820 */
821 error = xfs_qm_dqattach(ip);
822 if (error)
823 return error;
824
825 /*
826 * Wait for all direct I/O to complete.
827 */
828 inode_dio_wait(inode);
829
830 /*
831 * File data changes must be complete before we start the transaction to
832 * modify the inode. This needs to be done before joining the inode to
833 * the transaction because the inode cannot be unlocked once it is a
834 * part of the transaction.
835 *
836 * Start with zeroing any data beyond EOF that we may expose on file
837 * extension, or zeroing out the rest of the block on a downward
838 * truncate.
839 */
840 if (newsize > oldsize) {
841 trace_xfs_zero_eof(ip, oldsize, newsize - oldsize);
842 error = xfs_zero_range(ip, oldsize, newsize - oldsize,
843 &did_zeroing);
844 } else {
845 /*
846 * iomap won't detect a dirty page over an unwritten block (or a
847 * cow block over a hole) and subsequently skips zeroing the
848 * newly post-EOF portion of the page. Flush the new EOF to
849 * convert the block before the pagecache truncate.
850 */
851 error = filemap_write_and_wait_range(inode->i_mapping, newsize,
852 newsize);
853 if (error)
854 return error;
855 error = xfs_truncate_page(ip, newsize, &did_zeroing);
856 }
857
858 if (error)
859 return error;
860
861 /*
862 * We've already locked out new page faults, so now we can safely remove
863 * pages from the page cache knowing they won't get refaulted until we
864 * drop the XFS_MMAP_EXCL lock after the extent manipulations are
865 * complete. The truncate_setsize() call also cleans partial EOF page
866 * PTEs on extending truncates and hence ensures sub-page block size
867 * filesystems are correctly handled, too.
868 *
869 * We have to do all the page cache truncate work outside the
870 * transaction context as the "lock" order is page lock->log space
871 * reservation as defined by extent allocation in the writeback path.
872 * Hence a truncate can fail with ENOMEM from xfs_trans_alloc(), but
873 * having already truncated the in-memory version of the file (i.e. made
874 * user visible changes). There's not much we can do about this, except
875 * to hope that the caller sees ENOMEM and retries the truncate
876 * operation.
877 *
878 * And we update in-core i_size and truncate page cache beyond newsize
879 * before writeback the [i_disk_size, newsize] range, so we're
880 * guaranteed not to write stale data past the new EOF on truncate down.
881 */
882 truncate_setsize(inode, newsize);
883
884 /*
885 * We are going to log the inode size change in this transaction so
886 * any previous writes that are beyond the on disk EOF and the new
887 * EOF that have not been written out need to be written here. If we
888 * do not write the data out, we expose ourselves to the null files
889 * problem. Note that this includes any block zeroing we did above;
890 * otherwise those blocks may not be zeroed after a crash.
891 */
892 if (did_zeroing ||
893 (newsize > ip->i_disk_size && oldsize != ip->i_disk_size)) {
894 error = filemap_write_and_wait_range(VFS_I(ip)->i_mapping,
895 ip->i_disk_size, newsize - 1);
896 if (error)
897 return error;
898 }
899
900 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
901 if (error)
902 return error;
903
904 lock_flags |= XFS_ILOCK_EXCL;
905 xfs_ilock(ip, XFS_ILOCK_EXCL);
906 xfs_trans_ijoin(tp, ip, 0);
907
908 /*
909 * Only change the c/mtime if we are changing the size or we are
910 * explicitly asked to change it. This handles the semantic difference
911 * between truncate() and ftruncate() as implemented in the VFS.
912 *
913 * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
914 * special case where we need to update the times despite not having
915 * these flags set. For all other operations the VFS set these flags
916 * explicitly if it wants a timestamp update.
917 */
918 if (newsize != oldsize &&
919 !(iattr->ia_valid & (ATTR_CTIME | ATTR_MTIME))) {
920 iattr->ia_ctime = iattr->ia_mtime =
921 current_time(inode);
922 iattr->ia_valid |= ATTR_CTIME | ATTR_MTIME;
923 }
924
925 /*
926 * The first thing we do is set the size to new_size permanently on
927 * disk. This way we don't have to worry about anyone ever being able
928 * to look at the data being freed even in the face of a crash.
929 * What we're getting around here is the case where we free a block, it
930 * is allocated to another file, it is written to, and then we crash.
931 * If the new data gets written to the file but the log buffers
932 * containing the free and reallocation don't, then we'd end up with
933 * garbage in the blocks being freed. As long as we make the new size
934 * permanent before actually freeing any blocks it doesn't matter if
935 * they get written to.
936 */
937 ip->i_disk_size = newsize;
938 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
939
940 if (newsize <= oldsize) {
941 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, newsize);
942 if (error)
943 goto out_trans_cancel;
944
945 /*
946 * Truncated "down", so we're removing references to old data
947 * here - if we delay flushing for a long time, we expose
948 * ourselves unduly to the notorious NULL files problem. So,
949 * we mark this inode and flush it when the file is closed,
950 * and do not wait the usual (long) time for writeout.
951 */
952 xfs_iflags_set(ip, XFS_ITRUNCATED);
953
954 /* A truncate down always removes post-EOF blocks. */
955 xfs_inode_clear_eofblocks_tag(ip);
956 }
957
958 ASSERT(!(iattr->ia_valid & (ATTR_UID | ATTR_GID)));
959 setattr_copy(mnt_userns, inode, iattr);
960 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
961
962 XFS_STATS_INC(mp, xs_ig_attrchg);
963
964 if (xfs_has_wsync(mp))
965 xfs_trans_set_sync(tp);
966
967 error = xfs_trans_commit(tp);
968out_unlock:
969 if (lock_flags)
970 xfs_iunlock(ip, lock_flags);
971 return error;
972
973out_trans_cancel:
974 xfs_trans_cancel(tp);
975 goto out_unlock;
976}
977
978int
979xfs_vn_setattr_size(
980 struct user_namespace *mnt_userns,
981 struct dentry *dentry,
982 struct iattr *iattr)
983{
984 struct xfs_inode *ip = XFS_I(d_inode(dentry));
985 int error;
986
987 trace_xfs_setattr(ip);
988
989 error = xfs_vn_change_ok(mnt_userns, dentry, iattr);
990 if (error)
991 return error;
992 return xfs_setattr_size(mnt_userns, dentry, ip, iattr);
993}
994
995STATIC int
996xfs_vn_setattr(
997 struct user_namespace *mnt_userns,
998 struct dentry *dentry,
999 struct iattr *iattr)
1000{
1001 struct inode *inode = d_inode(dentry);
1002 struct xfs_inode *ip = XFS_I(inode);
1003 int error;
1004
1005 if (iattr->ia_valid & ATTR_SIZE) {
1006 uint iolock;
1007
1008 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1009 iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
1010
1011 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
1012 if (error) {
1013 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1014 return error;
1015 }
1016
1017 error = xfs_vn_setattr_size(mnt_userns, dentry, iattr);
1018 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1019 } else {
1020 trace_xfs_setattr(ip);
1021
1022 error = xfs_vn_change_ok(mnt_userns, dentry, iattr);
1023 if (!error)
1024 error = xfs_setattr_nonsize(mnt_userns, dentry, ip, iattr);
1025 }
1026
1027 return error;
1028}
1029
1030STATIC int
1031xfs_vn_update_time(
1032 struct inode *inode,
1033 struct timespec64 *now,
1034 int flags)
1035{
1036 struct xfs_inode *ip = XFS_I(inode);
1037 struct xfs_mount *mp = ip->i_mount;
1038 int log_flags = XFS_ILOG_TIMESTAMP;
1039 struct xfs_trans *tp;
1040 int error;
1041
1042 trace_xfs_update_time(ip);
1043
1044 if (inode->i_sb->s_flags & SB_LAZYTIME) {
1045 if (!((flags & S_VERSION) &&
1046 inode_maybe_inc_iversion(inode, false)))
1047 return generic_update_time(inode, now, flags);
1048
1049 /* Capture the iversion update that just occurred */
1050 log_flags |= XFS_ILOG_CORE;
1051 }
1052
1053 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_fsyncts, 0, 0, 0, &tp);
1054 if (error)
1055 return error;
1056
1057 xfs_ilock(ip, XFS_ILOCK_EXCL);
1058 if (flags & S_CTIME)
1059 inode->i_ctime = *now;
1060 if (flags & S_MTIME)
1061 inode->i_mtime = *now;
1062 if (flags & S_ATIME)
1063 inode->i_atime = *now;
1064
1065 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1066 xfs_trans_log_inode(tp, ip, log_flags);
1067 return xfs_trans_commit(tp);
1068}
1069
1070STATIC int
1071xfs_vn_fiemap(
1072 struct inode *inode,
1073 struct fiemap_extent_info *fieinfo,
1074 u64 start,
1075 u64 length)
1076{
1077 int error;
1078
1079 xfs_ilock(XFS_I(inode), XFS_IOLOCK_SHARED);
1080 if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
1081 fieinfo->fi_flags &= ~FIEMAP_FLAG_XATTR;
1082 error = iomap_fiemap(inode, fieinfo, start, length,
1083 &xfs_xattr_iomap_ops);
1084 } else {
1085 error = iomap_fiemap(inode, fieinfo, start, length,
1086 &xfs_read_iomap_ops);
1087 }
1088 xfs_iunlock(XFS_I(inode), XFS_IOLOCK_SHARED);
1089
1090 return error;
1091}
1092
1093STATIC int
1094xfs_vn_tmpfile(
1095 struct user_namespace *mnt_userns,
1096 struct inode *dir,
1097 struct file *file,
1098 umode_t mode)
1099{
1100 int err = xfs_generic_create(mnt_userns, dir, file->f_path.dentry, mode, 0, file);
1101
1102 return finish_open_simple(file, err);
1103}
1104
1105static const struct inode_operations xfs_inode_operations = {
1106 .get_inode_acl = xfs_get_acl,
1107 .set_acl = xfs_set_acl,
1108 .getattr = xfs_vn_getattr,
1109 .setattr = xfs_vn_setattr,
1110 .listxattr = xfs_vn_listxattr,
1111 .fiemap = xfs_vn_fiemap,
1112 .update_time = xfs_vn_update_time,
1113 .fileattr_get = xfs_fileattr_get,
1114 .fileattr_set = xfs_fileattr_set,
1115};
1116
1117static const struct inode_operations xfs_dir_inode_operations = {
1118 .create = xfs_vn_create,
1119 .lookup = xfs_vn_lookup,
1120 .link = xfs_vn_link,
1121 .unlink = xfs_vn_unlink,
1122 .symlink = xfs_vn_symlink,
1123 .mkdir = xfs_vn_mkdir,
1124 /*
1125 * Yes, XFS uses the same method for rmdir and unlink.
1126 *
1127 * There are some subtile differences deeper in the code,
1128 * but we use S_ISDIR to check for those.
1129 */
1130 .rmdir = xfs_vn_unlink,
1131 .mknod = xfs_vn_mknod,
1132 .rename = xfs_vn_rename,
1133 .get_inode_acl = xfs_get_acl,
1134 .set_acl = xfs_set_acl,
1135 .getattr = xfs_vn_getattr,
1136 .setattr = xfs_vn_setattr,
1137 .listxattr = xfs_vn_listxattr,
1138 .update_time = xfs_vn_update_time,
1139 .tmpfile = xfs_vn_tmpfile,
1140 .fileattr_get = xfs_fileattr_get,
1141 .fileattr_set = xfs_fileattr_set,
1142};
1143
1144static const struct inode_operations xfs_dir_ci_inode_operations = {
1145 .create = xfs_vn_create,
1146 .lookup = xfs_vn_ci_lookup,
1147 .link = xfs_vn_link,
1148 .unlink = xfs_vn_unlink,
1149 .symlink = xfs_vn_symlink,
1150 .mkdir = xfs_vn_mkdir,
1151 /*
1152 * Yes, XFS uses the same method for rmdir and unlink.
1153 *
1154 * There are some subtile differences deeper in the code,
1155 * but we use S_ISDIR to check for those.
1156 */
1157 .rmdir = xfs_vn_unlink,
1158 .mknod = xfs_vn_mknod,
1159 .rename = xfs_vn_rename,
1160 .get_inode_acl = xfs_get_acl,
1161 .set_acl = xfs_set_acl,
1162 .getattr = xfs_vn_getattr,
1163 .setattr = xfs_vn_setattr,
1164 .listxattr = xfs_vn_listxattr,
1165 .update_time = xfs_vn_update_time,
1166 .tmpfile = xfs_vn_tmpfile,
1167 .fileattr_get = xfs_fileattr_get,
1168 .fileattr_set = xfs_fileattr_set,
1169};
1170
1171static const struct inode_operations xfs_symlink_inode_operations = {
1172 .get_link = xfs_vn_get_link,
1173 .getattr = xfs_vn_getattr,
1174 .setattr = xfs_vn_setattr,
1175 .listxattr = xfs_vn_listxattr,
1176 .update_time = xfs_vn_update_time,
1177};
1178
1179/* Figure out if this file actually supports DAX. */
1180static bool
1181xfs_inode_supports_dax(
1182 struct xfs_inode *ip)
1183{
1184 struct xfs_mount *mp = ip->i_mount;
1185
1186 /* Only supported on regular files. */
1187 if (!S_ISREG(VFS_I(ip)->i_mode))
1188 return false;
1189
1190 /* Block size must match page size */
1191 if (mp->m_sb.sb_blocksize != PAGE_SIZE)
1192 return false;
1193
1194 /* Device has to support DAX too. */
1195 return xfs_inode_buftarg(ip)->bt_daxdev != NULL;
1196}
1197
1198static bool
1199xfs_inode_should_enable_dax(
1200 struct xfs_inode *ip)
1201{
1202 if (!IS_ENABLED(CONFIG_FS_DAX))
1203 return false;
1204 if (xfs_has_dax_never(ip->i_mount))
1205 return false;
1206 if (!xfs_inode_supports_dax(ip))
1207 return false;
1208 if (xfs_has_dax_always(ip->i_mount))
1209 return true;
1210 if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
1211 return true;
1212 return false;
1213}
1214
1215void
1216xfs_diflags_to_iflags(
1217 struct xfs_inode *ip,
1218 bool init)
1219{
1220 struct inode *inode = VFS_I(ip);
1221 unsigned int xflags = xfs_ip2xflags(ip);
1222 unsigned int flags = 0;
1223
1224 ASSERT(!(IS_DAX(inode) && init));
1225
1226 if (xflags & FS_XFLAG_IMMUTABLE)
1227 flags |= S_IMMUTABLE;
1228 if (xflags & FS_XFLAG_APPEND)
1229 flags |= S_APPEND;
1230 if (xflags & FS_XFLAG_SYNC)
1231 flags |= S_SYNC;
1232 if (xflags & FS_XFLAG_NOATIME)
1233 flags |= S_NOATIME;
1234 if (init && xfs_inode_should_enable_dax(ip))
1235 flags |= S_DAX;
1236
1237 /*
1238 * S_DAX can only be set during inode initialization and is never set by
1239 * the VFS, so we cannot mask off S_DAX in i_flags.
1240 */
1241 inode->i_flags &= ~(S_IMMUTABLE | S_APPEND | S_SYNC | S_NOATIME);
1242 inode->i_flags |= flags;
1243}
1244
1245/*
1246 * Initialize the Linux inode.
1247 *
1248 * When reading existing inodes from disk this is called directly from xfs_iget,
1249 * when creating a new inode it is called from xfs_init_new_inode after setting
1250 * up the inode. These callers have different criteria for clearing XFS_INEW, so
1251 * leave it up to the caller to deal with unlocking the inode appropriately.
1252 */
1253void
1254xfs_setup_inode(
1255 struct xfs_inode *ip)
1256{
1257 struct inode *inode = &ip->i_vnode;
1258 gfp_t gfp_mask;
1259
1260 inode->i_ino = ip->i_ino;
1261 inode->i_state |= I_NEW;
1262
1263 inode_sb_list_add(inode);
1264 /* make the inode look hashed for the writeback code */
1265 inode_fake_hash(inode);
1266
1267 i_size_write(inode, ip->i_disk_size);
1268 xfs_diflags_to_iflags(ip, true);
1269
1270 if (S_ISDIR(inode->i_mode)) {
1271 /*
1272 * We set the i_rwsem class here to avoid potential races with
1273 * lockdep_annotate_inode_mutex_key() reinitialising the lock
1274 * after a filehandle lookup has already found the inode in
1275 * cache before it has been unlocked via unlock_new_inode().
1276 */
1277 lockdep_set_class(&inode->i_rwsem,
1278 &inode->i_sb->s_type->i_mutex_dir_key);
1279 lockdep_set_class(&ip->i_lock.mr_lock, &xfs_dir_ilock_class);
1280 } else {
1281 lockdep_set_class(&ip->i_lock.mr_lock, &xfs_nondir_ilock_class);
1282 }
1283
1284 /*
1285 * Ensure all page cache allocations are done from GFP_NOFS context to
1286 * prevent direct reclaim recursion back into the filesystem and blowing
1287 * stacks or deadlocking.
1288 */
1289 gfp_mask = mapping_gfp_mask(inode->i_mapping);
1290 mapping_set_gfp_mask(inode->i_mapping, (gfp_mask & ~(__GFP_FS)));
1291
1292 /*
1293 * If there is no attribute fork no ACL can exist on this inode,
1294 * and it can't have any file capabilities attached to it either.
1295 */
1296 if (!xfs_inode_has_attr_fork(ip)) {
1297 inode_has_no_xattr(inode);
1298 cache_no_acl(inode);
1299 }
1300}
1301
1302void
1303xfs_setup_iops(
1304 struct xfs_inode *ip)
1305{
1306 struct inode *inode = &ip->i_vnode;
1307
1308 switch (inode->i_mode & S_IFMT) {
1309 case S_IFREG:
1310 inode->i_op = &xfs_inode_operations;
1311 inode->i_fop = &xfs_file_operations;
1312 if (IS_DAX(inode))
1313 inode->i_mapping->a_ops = &xfs_dax_aops;
1314 else
1315 inode->i_mapping->a_ops = &xfs_address_space_operations;
1316 break;
1317 case S_IFDIR:
1318 if (xfs_has_asciici(XFS_M(inode->i_sb)))
1319 inode->i_op = &xfs_dir_ci_inode_operations;
1320 else
1321 inode->i_op = &xfs_dir_inode_operations;
1322 inode->i_fop = &xfs_dir_file_operations;
1323 break;
1324 case S_IFLNK:
1325 inode->i_op = &xfs_symlink_inode_operations;
1326 break;
1327 default:
1328 inode->i_op = &xfs_inode_operations;
1329 init_special_inode(inode, inode->i_mode, inode->i_rdev);
1330 break;
1331 }
1332}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_acl.h"
15#include "xfs_quota.h"
16#include "xfs_attr.h"
17#include "xfs_trans.h"
18#include "xfs_trace.h"
19#include "xfs_icache.h"
20#include "xfs_symlink.h"
21#include "xfs_dir2.h"
22#include "xfs_iomap.h"
23
24#include <linux/xattr.h>
25#include <linux/posix_acl.h>
26#include <linux/security.h>
27#include <linux/iversion.h>
28
29/*
30 * Directories have different lock order w.r.t. mmap_sem compared to regular
31 * files. This is due to readdir potentially triggering page faults on a user
32 * buffer inside filldir(), and this happens with the ilock on the directory
33 * held. For regular files, the lock order is the other way around - the
34 * mmap_sem is taken during the page fault, and then we lock the ilock to do
35 * block mapping. Hence we need a different class for the directory ilock so
36 * that lockdep can tell them apart.
37 */
38static struct lock_class_key xfs_nondir_ilock_class;
39static struct lock_class_key xfs_dir_ilock_class;
40
41static int
42xfs_initxattrs(
43 struct inode *inode,
44 const struct xattr *xattr_array,
45 void *fs_info)
46{
47 const struct xattr *xattr;
48 struct xfs_inode *ip = XFS_I(inode);
49 int error = 0;
50
51 for (xattr = xattr_array; xattr->name != NULL; xattr++) {
52 error = xfs_attr_set(ip, xattr->name, xattr->value,
53 xattr->value_len, ATTR_SECURE);
54 if (error < 0)
55 break;
56 }
57 return error;
58}
59
60/*
61 * Hook in SELinux. This is not quite correct yet, what we really need
62 * here (as we do for default ACLs) is a mechanism by which creation of
63 * these attrs can be journalled at inode creation time (along with the
64 * inode, of course, such that log replay can't cause these to be lost).
65 */
66
67STATIC int
68xfs_init_security(
69 struct inode *inode,
70 struct inode *dir,
71 const struct qstr *qstr)
72{
73 return security_inode_init_security(inode, dir, qstr,
74 &xfs_initxattrs, NULL);
75}
76
77static void
78xfs_dentry_to_name(
79 struct xfs_name *namep,
80 struct dentry *dentry)
81{
82 namep->name = dentry->d_name.name;
83 namep->len = dentry->d_name.len;
84 namep->type = XFS_DIR3_FT_UNKNOWN;
85}
86
87static int
88xfs_dentry_mode_to_name(
89 struct xfs_name *namep,
90 struct dentry *dentry,
91 int mode)
92{
93 namep->name = dentry->d_name.name;
94 namep->len = dentry->d_name.len;
95 namep->type = xfs_mode_to_ftype(mode);
96
97 if (unlikely(namep->type == XFS_DIR3_FT_UNKNOWN))
98 return -EFSCORRUPTED;
99
100 return 0;
101}
102
103STATIC void
104xfs_cleanup_inode(
105 struct inode *dir,
106 struct inode *inode,
107 struct dentry *dentry)
108{
109 struct xfs_name teardown;
110
111 /* Oh, the horror.
112 * If we can't add the ACL or we fail in
113 * xfs_init_security we must back out.
114 * ENOSPC can hit here, among other things.
115 */
116 xfs_dentry_to_name(&teardown, dentry);
117
118 xfs_remove(XFS_I(dir), &teardown, XFS_I(inode));
119}
120
121STATIC int
122xfs_generic_create(
123 struct inode *dir,
124 struct dentry *dentry,
125 umode_t mode,
126 dev_t rdev,
127 bool tmpfile) /* unnamed file */
128{
129 struct inode *inode;
130 struct xfs_inode *ip = NULL;
131 struct posix_acl *default_acl, *acl;
132 struct xfs_name name;
133 int error;
134
135 /*
136 * Irix uses Missed'em'V split, but doesn't want to see
137 * the upper 5 bits of (14bit) major.
138 */
139 if (S_ISCHR(mode) || S_ISBLK(mode)) {
140 if (unlikely(!sysv_valid_dev(rdev) || MAJOR(rdev) & ~0x1ff))
141 return -EINVAL;
142 } else {
143 rdev = 0;
144 }
145
146 error = posix_acl_create(dir, &mode, &default_acl, &acl);
147 if (error)
148 return error;
149
150 /* Verify mode is valid also for tmpfile case */
151 error = xfs_dentry_mode_to_name(&name, dentry, mode);
152 if (unlikely(error))
153 goto out_free_acl;
154
155 if (!tmpfile) {
156 error = xfs_create(XFS_I(dir), &name, mode, rdev, &ip);
157 } else {
158 error = xfs_create_tmpfile(XFS_I(dir), mode, &ip);
159 }
160 if (unlikely(error))
161 goto out_free_acl;
162
163 inode = VFS_I(ip);
164
165 error = xfs_init_security(inode, dir, &dentry->d_name);
166 if (unlikely(error))
167 goto out_cleanup_inode;
168
169#ifdef CONFIG_XFS_POSIX_ACL
170 if (default_acl) {
171 error = __xfs_set_acl(inode, default_acl, ACL_TYPE_DEFAULT);
172 if (error)
173 goto out_cleanup_inode;
174 }
175 if (acl) {
176 error = __xfs_set_acl(inode, acl, ACL_TYPE_ACCESS);
177 if (error)
178 goto out_cleanup_inode;
179 }
180#endif
181
182 xfs_setup_iops(ip);
183
184 if (tmpfile) {
185 /*
186 * The VFS requires that any inode fed to d_tmpfile must have
187 * nlink == 1 so that it can decrement the nlink in d_tmpfile.
188 * However, we created the temp file with nlink == 0 because
189 * we're not allowed to put an inode with nlink > 0 on the
190 * unlinked list. Therefore we have to set nlink to 1 so that
191 * d_tmpfile can immediately set it back to zero.
192 */
193 set_nlink(inode, 1);
194 d_tmpfile(dentry, inode);
195 } else
196 d_instantiate(dentry, inode);
197
198 xfs_finish_inode_setup(ip);
199
200 out_free_acl:
201 if (default_acl)
202 posix_acl_release(default_acl);
203 if (acl)
204 posix_acl_release(acl);
205 return error;
206
207 out_cleanup_inode:
208 xfs_finish_inode_setup(ip);
209 if (!tmpfile)
210 xfs_cleanup_inode(dir, inode, dentry);
211 xfs_irele(ip);
212 goto out_free_acl;
213}
214
215STATIC int
216xfs_vn_mknod(
217 struct inode *dir,
218 struct dentry *dentry,
219 umode_t mode,
220 dev_t rdev)
221{
222 return xfs_generic_create(dir, dentry, mode, rdev, false);
223}
224
225STATIC int
226xfs_vn_create(
227 struct inode *dir,
228 struct dentry *dentry,
229 umode_t mode,
230 bool flags)
231{
232 return xfs_vn_mknod(dir, dentry, mode, 0);
233}
234
235STATIC int
236xfs_vn_mkdir(
237 struct inode *dir,
238 struct dentry *dentry,
239 umode_t mode)
240{
241 return xfs_vn_mknod(dir, dentry, mode|S_IFDIR, 0);
242}
243
244STATIC struct dentry *
245xfs_vn_lookup(
246 struct inode *dir,
247 struct dentry *dentry,
248 unsigned int flags)
249{
250 struct inode *inode;
251 struct xfs_inode *cip;
252 struct xfs_name name;
253 int error;
254
255 if (dentry->d_name.len >= MAXNAMELEN)
256 return ERR_PTR(-ENAMETOOLONG);
257
258 xfs_dentry_to_name(&name, dentry);
259 error = xfs_lookup(XFS_I(dir), &name, &cip, NULL);
260 if (likely(!error))
261 inode = VFS_I(cip);
262 else if (likely(error == -ENOENT))
263 inode = NULL;
264 else
265 inode = ERR_PTR(error);
266 return d_splice_alias(inode, dentry);
267}
268
269STATIC struct dentry *
270xfs_vn_ci_lookup(
271 struct inode *dir,
272 struct dentry *dentry,
273 unsigned int flags)
274{
275 struct xfs_inode *ip;
276 struct xfs_name xname;
277 struct xfs_name ci_name;
278 struct qstr dname;
279 int error;
280
281 if (dentry->d_name.len >= MAXNAMELEN)
282 return ERR_PTR(-ENAMETOOLONG);
283
284 xfs_dentry_to_name(&xname, dentry);
285 error = xfs_lookup(XFS_I(dir), &xname, &ip, &ci_name);
286 if (unlikely(error)) {
287 if (unlikely(error != -ENOENT))
288 return ERR_PTR(error);
289 /*
290 * call d_add(dentry, NULL) here when d_drop_negative_children
291 * is called in xfs_vn_mknod (ie. allow negative dentries
292 * with CI filesystems).
293 */
294 return NULL;
295 }
296
297 /* if exact match, just splice and exit */
298 if (!ci_name.name)
299 return d_splice_alias(VFS_I(ip), dentry);
300
301 /* else case-insensitive match... */
302 dname.name = ci_name.name;
303 dname.len = ci_name.len;
304 dentry = d_add_ci(dentry, VFS_I(ip), &dname);
305 kmem_free(ci_name.name);
306 return dentry;
307}
308
309STATIC int
310xfs_vn_link(
311 struct dentry *old_dentry,
312 struct inode *dir,
313 struct dentry *dentry)
314{
315 struct inode *inode = d_inode(old_dentry);
316 struct xfs_name name;
317 int error;
318
319 error = xfs_dentry_mode_to_name(&name, dentry, inode->i_mode);
320 if (unlikely(error))
321 return error;
322
323 error = xfs_link(XFS_I(dir), XFS_I(inode), &name);
324 if (unlikely(error))
325 return error;
326
327 ihold(inode);
328 d_instantiate(dentry, inode);
329 return 0;
330}
331
332STATIC int
333xfs_vn_unlink(
334 struct inode *dir,
335 struct dentry *dentry)
336{
337 struct xfs_name name;
338 int error;
339
340 xfs_dentry_to_name(&name, dentry);
341
342 error = xfs_remove(XFS_I(dir), &name, XFS_I(d_inode(dentry)));
343 if (error)
344 return error;
345
346 /*
347 * With unlink, the VFS makes the dentry "negative": no inode,
348 * but still hashed. This is incompatible with case-insensitive
349 * mode, so invalidate (unhash) the dentry in CI-mode.
350 */
351 if (xfs_sb_version_hasasciici(&XFS_M(dir->i_sb)->m_sb))
352 d_invalidate(dentry);
353 return 0;
354}
355
356STATIC int
357xfs_vn_symlink(
358 struct inode *dir,
359 struct dentry *dentry,
360 const char *symname)
361{
362 struct inode *inode;
363 struct xfs_inode *cip = NULL;
364 struct xfs_name name;
365 int error;
366 umode_t mode;
367
368 mode = S_IFLNK |
369 (irix_symlink_mode ? 0777 & ~current_umask() : S_IRWXUGO);
370 error = xfs_dentry_mode_to_name(&name, dentry, mode);
371 if (unlikely(error))
372 goto out;
373
374 error = xfs_symlink(XFS_I(dir), &name, symname, mode, &cip);
375 if (unlikely(error))
376 goto out;
377
378 inode = VFS_I(cip);
379
380 error = xfs_init_security(inode, dir, &dentry->d_name);
381 if (unlikely(error))
382 goto out_cleanup_inode;
383
384 xfs_setup_iops(cip);
385
386 d_instantiate(dentry, inode);
387 xfs_finish_inode_setup(cip);
388 return 0;
389
390 out_cleanup_inode:
391 xfs_finish_inode_setup(cip);
392 xfs_cleanup_inode(dir, inode, dentry);
393 xfs_irele(cip);
394 out:
395 return error;
396}
397
398STATIC int
399xfs_vn_rename(
400 struct inode *odir,
401 struct dentry *odentry,
402 struct inode *ndir,
403 struct dentry *ndentry,
404 unsigned int flags)
405{
406 struct inode *new_inode = d_inode(ndentry);
407 int omode = 0;
408 int error;
409 struct xfs_name oname;
410 struct xfs_name nname;
411
412 if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
413 return -EINVAL;
414
415 /* if we are exchanging files, we need to set i_mode of both files */
416 if (flags & RENAME_EXCHANGE)
417 omode = d_inode(ndentry)->i_mode;
418
419 error = xfs_dentry_mode_to_name(&oname, odentry, omode);
420 if (omode && unlikely(error))
421 return error;
422
423 error = xfs_dentry_mode_to_name(&nname, ndentry,
424 d_inode(odentry)->i_mode);
425 if (unlikely(error))
426 return error;
427
428 return xfs_rename(XFS_I(odir), &oname, XFS_I(d_inode(odentry)),
429 XFS_I(ndir), &nname,
430 new_inode ? XFS_I(new_inode) : NULL, flags);
431}
432
433/*
434 * careful here - this function can get called recursively, so
435 * we need to be very careful about how much stack we use.
436 * uio is kmalloced for this reason...
437 */
438STATIC const char *
439xfs_vn_get_link(
440 struct dentry *dentry,
441 struct inode *inode,
442 struct delayed_call *done)
443{
444 char *link;
445 int error = -ENOMEM;
446
447 if (!dentry)
448 return ERR_PTR(-ECHILD);
449
450 link = kmalloc(XFS_SYMLINK_MAXLEN+1, GFP_KERNEL);
451 if (!link)
452 goto out_err;
453
454 error = xfs_readlink(XFS_I(d_inode(dentry)), link);
455 if (unlikely(error))
456 goto out_kfree;
457
458 set_delayed_call(done, kfree_link, link);
459 return link;
460
461 out_kfree:
462 kfree(link);
463 out_err:
464 return ERR_PTR(error);
465}
466
467STATIC const char *
468xfs_vn_get_link_inline(
469 struct dentry *dentry,
470 struct inode *inode,
471 struct delayed_call *done)
472{
473 char *link;
474
475 ASSERT(XFS_I(inode)->i_df.if_flags & XFS_IFINLINE);
476
477 /*
478 * The VFS crashes on a NULL pointer, so return -EFSCORRUPTED if
479 * if_data is junk.
480 */
481 link = XFS_I(inode)->i_df.if_u1.if_data;
482 if (!link)
483 return ERR_PTR(-EFSCORRUPTED);
484 return link;
485}
486
487STATIC int
488xfs_vn_getattr(
489 const struct path *path,
490 struct kstat *stat,
491 u32 request_mask,
492 unsigned int query_flags)
493{
494 struct inode *inode = d_inode(path->dentry);
495 struct xfs_inode *ip = XFS_I(inode);
496 struct xfs_mount *mp = ip->i_mount;
497
498 trace_xfs_getattr(ip);
499
500 if (XFS_FORCED_SHUTDOWN(mp))
501 return -EIO;
502
503 stat->size = XFS_ISIZE(ip);
504 stat->dev = inode->i_sb->s_dev;
505 stat->mode = inode->i_mode;
506 stat->nlink = inode->i_nlink;
507 stat->uid = inode->i_uid;
508 stat->gid = inode->i_gid;
509 stat->ino = ip->i_ino;
510 stat->atime = inode->i_atime;
511 stat->mtime = inode->i_mtime;
512 stat->ctime = inode->i_ctime;
513 stat->blocks =
514 XFS_FSB_TO_BB(mp, ip->i_d.di_nblocks + ip->i_delayed_blks);
515
516 if (ip->i_d.di_version == 3) {
517 if (request_mask & STATX_BTIME) {
518 stat->result_mask |= STATX_BTIME;
519 stat->btime.tv_sec = ip->i_d.di_crtime.t_sec;
520 stat->btime.tv_nsec = ip->i_d.di_crtime.t_nsec;
521 }
522 }
523
524 /*
525 * Note: If you add another clause to set an attribute flag, please
526 * update attributes_mask below.
527 */
528 if (ip->i_d.di_flags & XFS_DIFLAG_IMMUTABLE)
529 stat->attributes |= STATX_ATTR_IMMUTABLE;
530 if (ip->i_d.di_flags & XFS_DIFLAG_APPEND)
531 stat->attributes |= STATX_ATTR_APPEND;
532 if (ip->i_d.di_flags & XFS_DIFLAG_NODUMP)
533 stat->attributes |= STATX_ATTR_NODUMP;
534
535 stat->attributes_mask |= (STATX_ATTR_IMMUTABLE |
536 STATX_ATTR_APPEND |
537 STATX_ATTR_NODUMP);
538
539 switch (inode->i_mode & S_IFMT) {
540 case S_IFBLK:
541 case S_IFCHR:
542 stat->blksize = BLKDEV_IOSIZE;
543 stat->rdev = inode->i_rdev;
544 break;
545 default:
546 if (XFS_IS_REALTIME_INODE(ip)) {
547 /*
548 * If the file blocks are being allocated from a
549 * realtime volume, then return the inode's realtime
550 * extent size or the realtime volume's extent size.
551 */
552 stat->blksize =
553 xfs_get_extsz_hint(ip) << mp->m_sb.sb_blocklog;
554 } else
555 stat->blksize = xfs_preferred_iosize(mp);
556 stat->rdev = 0;
557 break;
558 }
559
560 return 0;
561}
562
563static void
564xfs_setattr_mode(
565 struct xfs_inode *ip,
566 struct iattr *iattr)
567{
568 struct inode *inode = VFS_I(ip);
569 umode_t mode = iattr->ia_mode;
570
571 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
572
573 inode->i_mode &= S_IFMT;
574 inode->i_mode |= mode & ~S_IFMT;
575}
576
577void
578xfs_setattr_time(
579 struct xfs_inode *ip,
580 struct iattr *iattr)
581{
582 struct inode *inode = VFS_I(ip);
583
584 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
585
586 if (iattr->ia_valid & ATTR_ATIME)
587 inode->i_atime = iattr->ia_atime;
588 if (iattr->ia_valid & ATTR_CTIME)
589 inode->i_ctime = iattr->ia_ctime;
590 if (iattr->ia_valid & ATTR_MTIME)
591 inode->i_mtime = iattr->ia_mtime;
592}
593
594static int
595xfs_vn_change_ok(
596 struct dentry *dentry,
597 struct iattr *iattr)
598{
599 struct xfs_mount *mp = XFS_I(d_inode(dentry))->i_mount;
600
601 if (mp->m_flags & XFS_MOUNT_RDONLY)
602 return -EROFS;
603
604 if (XFS_FORCED_SHUTDOWN(mp))
605 return -EIO;
606
607 return setattr_prepare(dentry, iattr);
608}
609
610/*
611 * Set non-size attributes of an inode.
612 *
613 * Caution: The caller of this function is responsible for calling
614 * setattr_prepare() or otherwise verifying the change is fine.
615 */
616int
617xfs_setattr_nonsize(
618 struct xfs_inode *ip,
619 struct iattr *iattr,
620 int flags)
621{
622 xfs_mount_t *mp = ip->i_mount;
623 struct inode *inode = VFS_I(ip);
624 int mask = iattr->ia_valid;
625 xfs_trans_t *tp;
626 int error;
627 kuid_t uid = GLOBAL_ROOT_UID, iuid = GLOBAL_ROOT_UID;
628 kgid_t gid = GLOBAL_ROOT_GID, igid = GLOBAL_ROOT_GID;
629 struct xfs_dquot *udqp = NULL, *gdqp = NULL;
630 struct xfs_dquot *olddquot1 = NULL, *olddquot2 = NULL;
631
632 ASSERT((mask & ATTR_SIZE) == 0);
633
634 /*
635 * If disk quotas is on, we make sure that the dquots do exist on disk,
636 * before we start any other transactions. Trying to do this later
637 * is messy. We don't care to take a readlock to look at the ids
638 * in inode here, because we can't hold it across the trans_reserve.
639 * If the IDs do change before we take the ilock, we're covered
640 * because the i_*dquot fields will get updated anyway.
641 */
642 if (XFS_IS_QUOTA_ON(mp) && (mask & (ATTR_UID|ATTR_GID))) {
643 uint qflags = 0;
644
645 if ((mask & ATTR_UID) && XFS_IS_UQUOTA_ON(mp)) {
646 uid = iattr->ia_uid;
647 qflags |= XFS_QMOPT_UQUOTA;
648 } else {
649 uid = inode->i_uid;
650 }
651 if ((mask & ATTR_GID) && XFS_IS_GQUOTA_ON(mp)) {
652 gid = iattr->ia_gid;
653 qflags |= XFS_QMOPT_GQUOTA;
654 } else {
655 gid = inode->i_gid;
656 }
657
658 /*
659 * We take a reference when we initialize udqp and gdqp,
660 * so it is important that we never blindly double trip on
661 * the same variable. See xfs_create() for an example.
662 */
663 ASSERT(udqp == NULL);
664 ASSERT(gdqp == NULL);
665 error = xfs_qm_vop_dqalloc(ip, xfs_kuid_to_uid(uid),
666 xfs_kgid_to_gid(gid),
667 xfs_get_projid(ip),
668 qflags, &udqp, &gdqp, NULL);
669 if (error)
670 return error;
671 }
672
673 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
674 if (error)
675 goto out_dqrele;
676
677 xfs_ilock(ip, XFS_ILOCK_EXCL);
678 xfs_trans_ijoin(tp, ip, 0);
679
680 /*
681 * Change file ownership. Must be the owner or privileged.
682 */
683 if (mask & (ATTR_UID|ATTR_GID)) {
684 /*
685 * These IDs could have changed since we last looked at them.
686 * But, we're assured that if the ownership did change
687 * while we didn't have the inode locked, inode's dquot(s)
688 * would have changed also.
689 */
690 iuid = inode->i_uid;
691 igid = inode->i_gid;
692 gid = (mask & ATTR_GID) ? iattr->ia_gid : igid;
693 uid = (mask & ATTR_UID) ? iattr->ia_uid : iuid;
694
695 /*
696 * Do a quota reservation only if uid/gid is actually
697 * going to change.
698 */
699 if (XFS_IS_QUOTA_RUNNING(mp) &&
700 ((XFS_IS_UQUOTA_ON(mp) && !uid_eq(iuid, uid)) ||
701 (XFS_IS_GQUOTA_ON(mp) && !gid_eq(igid, gid)))) {
702 ASSERT(tp);
703 error = xfs_qm_vop_chown_reserve(tp, ip, udqp, gdqp,
704 NULL, capable(CAP_FOWNER) ?
705 XFS_QMOPT_FORCE_RES : 0);
706 if (error) /* out of quota */
707 goto out_cancel;
708 }
709 }
710
711 /*
712 * Change file ownership. Must be the owner or privileged.
713 */
714 if (mask & (ATTR_UID|ATTR_GID)) {
715 /*
716 * CAP_FSETID overrides the following restrictions:
717 *
718 * The set-user-ID and set-group-ID bits of a file will be
719 * cleared upon successful return from chown()
720 */
721 if ((inode->i_mode & (S_ISUID|S_ISGID)) &&
722 !capable(CAP_FSETID))
723 inode->i_mode &= ~(S_ISUID|S_ISGID);
724
725 /*
726 * Change the ownerships and register quota modifications
727 * in the transaction.
728 */
729 if (!uid_eq(iuid, uid)) {
730 if (XFS_IS_QUOTA_RUNNING(mp) && XFS_IS_UQUOTA_ON(mp)) {
731 ASSERT(mask & ATTR_UID);
732 ASSERT(udqp);
733 olddquot1 = xfs_qm_vop_chown(tp, ip,
734 &ip->i_udquot, udqp);
735 }
736 ip->i_d.di_uid = xfs_kuid_to_uid(uid);
737 inode->i_uid = uid;
738 }
739 if (!gid_eq(igid, gid)) {
740 if (XFS_IS_QUOTA_RUNNING(mp) && XFS_IS_GQUOTA_ON(mp)) {
741 ASSERT(xfs_sb_version_has_pquotino(&mp->m_sb) ||
742 !XFS_IS_PQUOTA_ON(mp));
743 ASSERT(mask & ATTR_GID);
744 ASSERT(gdqp);
745 olddquot2 = xfs_qm_vop_chown(tp, ip,
746 &ip->i_gdquot, gdqp);
747 }
748 ip->i_d.di_gid = xfs_kgid_to_gid(gid);
749 inode->i_gid = gid;
750 }
751 }
752
753 if (mask & ATTR_MODE)
754 xfs_setattr_mode(ip, iattr);
755 if (mask & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME))
756 xfs_setattr_time(ip, iattr);
757
758 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
759
760 XFS_STATS_INC(mp, xs_ig_attrchg);
761
762 if (mp->m_flags & XFS_MOUNT_WSYNC)
763 xfs_trans_set_sync(tp);
764 error = xfs_trans_commit(tp);
765
766 xfs_iunlock(ip, XFS_ILOCK_EXCL);
767
768 /*
769 * Release any dquot(s) the inode had kept before chown.
770 */
771 xfs_qm_dqrele(olddquot1);
772 xfs_qm_dqrele(olddquot2);
773 xfs_qm_dqrele(udqp);
774 xfs_qm_dqrele(gdqp);
775
776 if (error)
777 return error;
778
779 /*
780 * XXX(hch): Updating the ACL entries is not atomic vs the i_mode
781 * update. We could avoid this with linked transactions
782 * and passing down the transaction pointer all the way
783 * to attr_set. No previous user of the generic
784 * Posix ACL code seems to care about this issue either.
785 */
786 if ((mask & ATTR_MODE) && !(flags & XFS_ATTR_NOACL)) {
787 error = posix_acl_chmod(inode, inode->i_mode);
788 if (error)
789 return error;
790 }
791
792 return 0;
793
794out_cancel:
795 xfs_trans_cancel(tp);
796 xfs_iunlock(ip, XFS_ILOCK_EXCL);
797out_dqrele:
798 xfs_qm_dqrele(udqp);
799 xfs_qm_dqrele(gdqp);
800 return error;
801}
802
803int
804xfs_vn_setattr_nonsize(
805 struct dentry *dentry,
806 struct iattr *iattr)
807{
808 struct xfs_inode *ip = XFS_I(d_inode(dentry));
809 int error;
810
811 trace_xfs_setattr(ip);
812
813 error = xfs_vn_change_ok(dentry, iattr);
814 if (error)
815 return error;
816 return xfs_setattr_nonsize(ip, iattr, 0);
817}
818
819/*
820 * Truncate file. Must have write permission and not be a directory.
821 *
822 * Caution: The caller of this function is responsible for calling
823 * setattr_prepare() or otherwise verifying the change is fine.
824 */
825STATIC int
826xfs_setattr_size(
827 struct xfs_inode *ip,
828 struct iattr *iattr)
829{
830 struct xfs_mount *mp = ip->i_mount;
831 struct inode *inode = VFS_I(ip);
832 xfs_off_t oldsize, newsize;
833 struct xfs_trans *tp;
834 int error;
835 uint lock_flags = 0;
836 bool did_zeroing = false;
837
838 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
839 ASSERT(xfs_isilocked(ip, XFS_MMAPLOCK_EXCL));
840 ASSERT(S_ISREG(inode->i_mode));
841 ASSERT((iattr->ia_valid & (ATTR_UID|ATTR_GID|ATTR_ATIME|ATTR_ATIME_SET|
842 ATTR_MTIME_SET|ATTR_KILL_PRIV|ATTR_TIMES_SET)) == 0);
843
844 oldsize = inode->i_size;
845 newsize = iattr->ia_size;
846
847 /*
848 * Short circuit the truncate case for zero length files.
849 */
850 if (newsize == 0 && oldsize == 0 && ip->i_d.di_nextents == 0) {
851 if (!(iattr->ia_valid & (ATTR_CTIME|ATTR_MTIME)))
852 return 0;
853
854 /*
855 * Use the regular setattr path to update the timestamps.
856 */
857 iattr->ia_valid &= ~ATTR_SIZE;
858 return xfs_setattr_nonsize(ip, iattr, 0);
859 }
860
861 /*
862 * Make sure that the dquots are attached to the inode.
863 */
864 error = xfs_qm_dqattach(ip);
865 if (error)
866 return error;
867
868 /*
869 * Wait for all direct I/O to complete.
870 */
871 inode_dio_wait(inode);
872
873 /*
874 * File data changes must be complete before we start the transaction to
875 * modify the inode. This needs to be done before joining the inode to
876 * the transaction because the inode cannot be unlocked once it is a
877 * part of the transaction.
878 *
879 * Start with zeroing any data beyond EOF that we may expose on file
880 * extension, or zeroing out the rest of the block on a downward
881 * truncate.
882 */
883 if (newsize > oldsize) {
884 trace_xfs_zero_eof(ip, oldsize, newsize - oldsize);
885 error = iomap_zero_range(inode, oldsize, newsize - oldsize,
886 &did_zeroing, &xfs_iomap_ops);
887 } else {
888 error = iomap_truncate_page(inode, newsize, &did_zeroing,
889 &xfs_iomap_ops);
890 }
891
892 if (error)
893 return error;
894
895 /*
896 * We've already locked out new page faults, so now we can safely remove
897 * pages from the page cache knowing they won't get refaulted until we
898 * drop the XFS_MMAP_EXCL lock after the extent manipulations are
899 * complete. The truncate_setsize() call also cleans partial EOF page
900 * PTEs on extending truncates and hence ensures sub-page block size
901 * filesystems are correctly handled, too.
902 *
903 * We have to do all the page cache truncate work outside the
904 * transaction context as the "lock" order is page lock->log space
905 * reservation as defined by extent allocation in the writeback path.
906 * Hence a truncate can fail with ENOMEM from xfs_trans_alloc(), but
907 * having already truncated the in-memory version of the file (i.e. made
908 * user visible changes). There's not much we can do about this, except
909 * to hope that the caller sees ENOMEM and retries the truncate
910 * operation.
911 *
912 * And we update in-core i_size and truncate page cache beyond newsize
913 * before writeback the [di_size, newsize] range, so we're guaranteed
914 * not to write stale data past the new EOF on truncate down.
915 */
916 truncate_setsize(inode, newsize);
917
918 /*
919 * We are going to log the inode size change in this transaction so
920 * any previous writes that are beyond the on disk EOF and the new
921 * EOF that have not been written out need to be written here. If we
922 * do not write the data out, we expose ourselves to the null files
923 * problem. Note that this includes any block zeroing we did above;
924 * otherwise those blocks may not be zeroed after a crash.
925 */
926 if (did_zeroing ||
927 (newsize > ip->i_d.di_size && oldsize != ip->i_d.di_size)) {
928 error = filemap_write_and_wait_range(VFS_I(ip)->i_mapping,
929 ip->i_d.di_size, newsize - 1);
930 if (error)
931 return error;
932 }
933
934 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
935 if (error)
936 return error;
937
938 lock_flags |= XFS_ILOCK_EXCL;
939 xfs_ilock(ip, XFS_ILOCK_EXCL);
940 xfs_trans_ijoin(tp, ip, 0);
941
942 /*
943 * Only change the c/mtime if we are changing the size or we are
944 * explicitly asked to change it. This handles the semantic difference
945 * between truncate() and ftruncate() as implemented in the VFS.
946 *
947 * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
948 * special case where we need to update the times despite not having
949 * these flags set. For all other operations the VFS set these flags
950 * explicitly if it wants a timestamp update.
951 */
952 if (newsize != oldsize &&
953 !(iattr->ia_valid & (ATTR_CTIME | ATTR_MTIME))) {
954 iattr->ia_ctime = iattr->ia_mtime =
955 current_time(inode);
956 iattr->ia_valid |= ATTR_CTIME | ATTR_MTIME;
957 }
958
959 /*
960 * The first thing we do is set the size to new_size permanently on
961 * disk. This way we don't have to worry about anyone ever being able
962 * to look at the data being freed even in the face of a crash.
963 * What we're getting around here is the case where we free a block, it
964 * is allocated to another file, it is written to, and then we crash.
965 * If the new data gets written to the file but the log buffers
966 * containing the free and reallocation don't, then we'd end up with
967 * garbage in the blocks being freed. As long as we make the new size
968 * permanent before actually freeing any blocks it doesn't matter if
969 * they get written to.
970 */
971 ip->i_d.di_size = newsize;
972 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
973
974 if (newsize <= oldsize) {
975 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, newsize);
976 if (error)
977 goto out_trans_cancel;
978
979 /*
980 * Truncated "down", so we're removing references to old data
981 * here - if we delay flushing for a long time, we expose
982 * ourselves unduly to the notorious NULL files problem. So,
983 * we mark this inode and flush it when the file is closed,
984 * and do not wait the usual (long) time for writeout.
985 */
986 xfs_iflags_set(ip, XFS_ITRUNCATED);
987
988 /* A truncate down always removes post-EOF blocks. */
989 xfs_inode_clear_eofblocks_tag(ip);
990 }
991
992 if (iattr->ia_valid & ATTR_MODE)
993 xfs_setattr_mode(ip, iattr);
994 if (iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME))
995 xfs_setattr_time(ip, iattr);
996
997 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
998
999 XFS_STATS_INC(mp, xs_ig_attrchg);
1000
1001 if (mp->m_flags & XFS_MOUNT_WSYNC)
1002 xfs_trans_set_sync(tp);
1003
1004 error = xfs_trans_commit(tp);
1005out_unlock:
1006 if (lock_flags)
1007 xfs_iunlock(ip, lock_flags);
1008 return error;
1009
1010out_trans_cancel:
1011 xfs_trans_cancel(tp);
1012 goto out_unlock;
1013}
1014
1015int
1016xfs_vn_setattr_size(
1017 struct dentry *dentry,
1018 struct iattr *iattr)
1019{
1020 struct xfs_inode *ip = XFS_I(d_inode(dentry));
1021 int error;
1022
1023 trace_xfs_setattr(ip);
1024
1025 error = xfs_vn_change_ok(dentry, iattr);
1026 if (error)
1027 return error;
1028 return xfs_setattr_size(ip, iattr);
1029}
1030
1031STATIC int
1032xfs_vn_setattr(
1033 struct dentry *dentry,
1034 struct iattr *iattr)
1035{
1036 int error;
1037
1038 if (iattr->ia_valid & ATTR_SIZE) {
1039 struct inode *inode = d_inode(dentry);
1040 struct xfs_inode *ip = XFS_I(inode);
1041 uint iolock;
1042
1043 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1044 iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
1045
1046 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
1047 if (error) {
1048 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1049 return error;
1050 }
1051
1052 error = xfs_vn_setattr_size(dentry, iattr);
1053 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1054 } else {
1055 error = xfs_vn_setattr_nonsize(dentry, iattr);
1056 }
1057
1058 return error;
1059}
1060
1061STATIC int
1062xfs_vn_update_time(
1063 struct inode *inode,
1064 struct timespec64 *now,
1065 int flags)
1066{
1067 struct xfs_inode *ip = XFS_I(inode);
1068 struct xfs_mount *mp = ip->i_mount;
1069 int log_flags = XFS_ILOG_TIMESTAMP;
1070 struct xfs_trans *tp;
1071 int error;
1072
1073 trace_xfs_update_time(ip);
1074
1075 if (inode->i_sb->s_flags & SB_LAZYTIME) {
1076 if (!((flags & S_VERSION) &&
1077 inode_maybe_inc_iversion(inode, false)))
1078 return generic_update_time(inode, now, flags);
1079
1080 /* Capture the iversion update that just occurred */
1081 log_flags |= XFS_ILOG_CORE;
1082 }
1083
1084 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_fsyncts, 0, 0, 0, &tp);
1085 if (error)
1086 return error;
1087
1088 xfs_ilock(ip, XFS_ILOCK_EXCL);
1089 if (flags & S_CTIME)
1090 inode->i_ctime = *now;
1091 if (flags & S_MTIME)
1092 inode->i_mtime = *now;
1093 if (flags & S_ATIME)
1094 inode->i_atime = *now;
1095
1096 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1097 xfs_trans_log_inode(tp, ip, log_flags);
1098 return xfs_trans_commit(tp);
1099}
1100
1101STATIC int
1102xfs_vn_fiemap(
1103 struct inode *inode,
1104 struct fiemap_extent_info *fieinfo,
1105 u64 start,
1106 u64 length)
1107{
1108 int error;
1109
1110 xfs_ilock(XFS_I(inode), XFS_IOLOCK_SHARED);
1111 if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
1112 fieinfo->fi_flags &= ~FIEMAP_FLAG_XATTR;
1113 error = iomap_fiemap(inode, fieinfo, start, length,
1114 &xfs_xattr_iomap_ops);
1115 } else {
1116 error = iomap_fiemap(inode, fieinfo, start, length,
1117 &xfs_iomap_ops);
1118 }
1119 xfs_iunlock(XFS_I(inode), XFS_IOLOCK_SHARED);
1120
1121 return error;
1122}
1123
1124STATIC int
1125xfs_vn_tmpfile(
1126 struct inode *dir,
1127 struct dentry *dentry,
1128 umode_t mode)
1129{
1130 return xfs_generic_create(dir, dentry, mode, 0, true);
1131}
1132
1133static const struct inode_operations xfs_inode_operations = {
1134 .get_acl = xfs_get_acl,
1135 .set_acl = xfs_set_acl,
1136 .getattr = xfs_vn_getattr,
1137 .setattr = xfs_vn_setattr,
1138 .listxattr = xfs_vn_listxattr,
1139 .fiemap = xfs_vn_fiemap,
1140 .update_time = xfs_vn_update_time,
1141};
1142
1143static const struct inode_operations xfs_dir_inode_operations = {
1144 .create = xfs_vn_create,
1145 .lookup = xfs_vn_lookup,
1146 .link = xfs_vn_link,
1147 .unlink = xfs_vn_unlink,
1148 .symlink = xfs_vn_symlink,
1149 .mkdir = xfs_vn_mkdir,
1150 /*
1151 * Yes, XFS uses the same method for rmdir and unlink.
1152 *
1153 * There are some subtile differences deeper in the code,
1154 * but we use S_ISDIR to check for those.
1155 */
1156 .rmdir = xfs_vn_unlink,
1157 .mknod = xfs_vn_mknod,
1158 .rename = xfs_vn_rename,
1159 .get_acl = xfs_get_acl,
1160 .set_acl = xfs_set_acl,
1161 .getattr = xfs_vn_getattr,
1162 .setattr = xfs_vn_setattr,
1163 .listxattr = xfs_vn_listxattr,
1164 .update_time = xfs_vn_update_time,
1165 .tmpfile = xfs_vn_tmpfile,
1166};
1167
1168static const struct inode_operations xfs_dir_ci_inode_operations = {
1169 .create = xfs_vn_create,
1170 .lookup = xfs_vn_ci_lookup,
1171 .link = xfs_vn_link,
1172 .unlink = xfs_vn_unlink,
1173 .symlink = xfs_vn_symlink,
1174 .mkdir = xfs_vn_mkdir,
1175 /*
1176 * Yes, XFS uses the same method for rmdir and unlink.
1177 *
1178 * There are some subtile differences deeper in the code,
1179 * but we use S_ISDIR to check for those.
1180 */
1181 .rmdir = xfs_vn_unlink,
1182 .mknod = xfs_vn_mknod,
1183 .rename = xfs_vn_rename,
1184 .get_acl = xfs_get_acl,
1185 .set_acl = xfs_set_acl,
1186 .getattr = xfs_vn_getattr,
1187 .setattr = xfs_vn_setattr,
1188 .listxattr = xfs_vn_listxattr,
1189 .update_time = xfs_vn_update_time,
1190 .tmpfile = xfs_vn_tmpfile,
1191};
1192
1193static const struct inode_operations xfs_symlink_inode_operations = {
1194 .get_link = xfs_vn_get_link,
1195 .getattr = xfs_vn_getattr,
1196 .setattr = xfs_vn_setattr,
1197 .listxattr = xfs_vn_listxattr,
1198 .update_time = xfs_vn_update_time,
1199};
1200
1201static const struct inode_operations xfs_inline_symlink_inode_operations = {
1202 .get_link = xfs_vn_get_link_inline,
1203 .getattr = xfs_vn_getattr,
1204 .setattr = xfs_vn_setattr,
1205 .listxattr = xfs_vn_listxattr,
1206 .update_time = xfs_vn_update_time,
1207};
1208
1209/* Figure out if this file actually supports DAX. */
1210static bool
1211xfs_inode_supports_dax(
1212 struct xfs_inode *ip)
1213{
1214 struct xfs_mount *mp = ip->i_mount;
1215
1216 /* Only supported on non-reflinked files. */
1217 if (!S_ISREG(VFS_I(ip)->i_mode) || xfs_is_reflink_inode(ip))
1218 return false;
1219
1220 /* DAX mount option or DAX iflag must be set. */
1221 if (!(mp->m_flags & XFS_MOUNT_DAX) &&
1222 !(ip->i_d.di_flags2 & XFS_DIFLAG2_DAX))
1223 return false;
1224
1225 /* Block size must match page size */
1226 if (mp->m_sb.sb_blocksize != PAGE_SIZE)
1227 return false;
1228
1229 /* Device has to support DAX too. */
1230 return xfs_find_daxdev_for_inode(VFS_I(ip)) != NULL;
1231}
1232
1233STATIC void
1234xfs_diflags_to_iflags(
1235 struct inode *inode,
1236 struct xfs_inode *ip)
1237{
1238 uint16_t flags = ip->i_d.di_flags;
1239
1240 inode->i_flags &= ~(S_IMMUTABLE | S_APPEND | S_SYNC |
1241 S_NOATIME | S_DAX);
1242
1243 if (flags & XFS_DIFLAG_IMMUTABLE)
1244 inode->i_flags |= S_IMMUTABLE;
1245 if (flags & XFS_DIFLAG_APPEND)
1246 inode->i_flags |= S_APPEND;
1247 if (flags & XFS_DIFLAG_SYNC)
1248 inode->i_flags |= S_SYNC;
1249 if (flags & XFS_DIFLAG_NOATIME)
1250 inode->i_flags |= S_NOATIME;
1251 if (xfs_inode_supports_dax(ip))
1252 inode->i_flags |= S_DAX;
1253}
1254
1255/*
1256 * Initialize the Linux inode.
1257 *
1258 * When reading existing inodes from disk this is called directly from xfs_iget,
1259 * when creating a new inode it is called from xfs_ialloc after setting up the
1260 * inode. These callers have different criteria for clearing XFS_INEW, so leave
1261 * it up to the caller to deal with unlocking the inode appropriately.
1262 */
1263void
1264xfs_setup_inode(
1265 struct xfs_inode *ip)
1266{
1267 struct inode *inode = &ip->i_vnode;
1268 gfp_t gfp_mask;
1269
1270 inode->i_ino = ip->i_ino;
1271 inode->i_state = I_NEW;
1272
1273 inode_sb_list_add(inode);
1274 /* make the inode look hashed for the writeback code */
1275 inode_fake_hash(inode);
1276
1277 inode->i_uid = xfs_uid_to_kuid(ip->i_d.di_uid);
1278 inode->i_gid = xfs_gid_to_kgid(ip->i_d.di_gid);
1279
1280 i_size_write(inode, ip->i_d.di_size);
1281 xfs_diflags_to_iflags(inode, ip);
1282
1283 if (S_ISDIR(inode->i_mode)) {
1284 /*
1285 * We set the i_rwsem class here to avoid potential races with
1286 * lockdep_annotate_inode_mutex_key() reinitialising the lock
1287 * after a filehandle lookup has already found the inode in
1288 * cache before it has been unlocked via unlock_new_inode().
1289 */
1290 lockdep_set_class(&inode->i_rwsem,
1291 &inode->i_sb->s_type->i_mutex_dir_key);
1292 lockdep_set_class(&ip->i_lock.mr_lock, &xfs_dir_ilock_class);
1293 ip->d_ops = ip->i_mount->m_dir_inode_ops;
1294 } else {
1295 ip->d_ops = ip->i_mount->m_nondir_inode_ops;
1296 lockdep_set_class(&ip->i_lock.mr_lock, &xfs_nondir_ilock_class);
1297 }
1298
1299 /*
1300 * Ensure all page cache allocations are done from GFP_NOFS context to
1301 * prevent direct reclaim recursion back into the filesystem and blowing
1302 * stacks or deadlocking.
1303 */
1304 gfp_mask = mapping_gfp_mask(inode->i_mapping);
1305 mapping_set_gfp_mask(inode->i_mapping, (gfp_mask & ~(__GFP_FS)));
1306
1307 /*
1308 * If there is no attribute fork no ACL can exist on this inode,
1309 * and it can't have any file capabilities attached to it either.
1310 */
1311 if (!XFS_IFORK_Q(ip)) {
1312 inode_has_no_xattr(inode);
1313 cache_no_acl(inode);
1314 }
1315}
1316
1317void
1318xfs_setup_iops(
1319 struct xfs_inode *ip)
1320{
1321 struct inode *inode = &ip->i_vnode;
1322
1323 switch (inode->i_mode & S_IFMT) {
1324 case S_IFREG:
1325 inode->i_op = &xfs_inode_operations;
1326 inode->i_fop = &xfs_file_operations;
1327 if (IS_DAX(inode))
1328 inode->i_mapping->a_ops = &xfs_dax_aops;
1329 else
1330 inode->i_mapping->a_ops = &xfs_address_space_operations;
1331 break;
1332 case S_IFDIR:
1333 if (xfs_sb_version_hasasciici(&XFS_M(inode->i_sb)->m_sb))
1334 inode->i_op = &xfs_dir_ci_inode_operations;
1335 else
1336 inode->i_op = &xfs_dir_inode_operations;
1337 inode->i_fop = &xfs_dir_file_operations;
1338 break;
1339 case S_IFLNK:
1340 if (ip->i_df.if_flags & XFS_IFINLINE)
1341 inode->i_op = &xfs_inline_symlink_inode_operations;
1342 else
1343 inode->i_op = &xfs_symlink_inode_operations;
1344 break;
1345 default:
1346 inode->i_op = &xfs_inode_operations;
1347 init_special_inode(inode, inode->i_mode, inode->i_rdev);
1348 break;
1349 }
1350}