* [Bug report] ntfs3: attr_insert_range() shifts nres.evcn of a resident attribute past the end of the MFT record buffer
@ 2026-11-15 12:00 Guikun Yang
0 siblings, 0 replies; only message in thread
From: Guikun Yang @ 2026-11-15 12:00 UTC (permalink / raw)
To: Konstantin Komarov; +Cc: ntfs3, linux-fsdevel, linux-kernel
Hi,
fallocate(FALLOC_FL_INSERT_RANGE) on a crafted NTFS image reaches a
segment-enumeration loop in attr_insert_range() that shifts the
svcn/evcn of every following DATA attribute without checking whether
the enumerated attribute is non-resident. A resident attribute placed
at the end of an extension MFT record makes le64_add_cpu(&nres.evcn,
len) do an 8-byte read-modify-write past the end of the 1024-byte
record buffer.
BUG: KASAN: slab-out-of-bounds in attr_insert_range+0x14a1/0x1670 [ntfs3]
Read of size 8 at addr ffff8880062d9400 by task init/1
...
allocated 1024-byte region [ffff8880062d9000, ffff8880062d9400)
The buggy address is located 0 bytes to the right
Reproduced 3/3. The negative control changes only that attribute's
type from ATTR_DATA to ATTR_END; fallocate() then returns 0 with no
KASAN report.
## This is the other branch of the function
bf9a6578e0 ("fs/ntfs3: fix slab-out-of-bounds write in
attr_insert_range()", in 7.4) fixed the **resident** branch: the
inline-data memmove used the insert length as both source offset and
copy size. The current tree has that fix:
memmove(data + vbo + bytes, data + vbo, data_size - vbo);
The **non-resident** branch - the loop below it - is a separate path
that still has no bound (fs/ntfs3/attrib.c):
next_svcn = le64_to_cpu(attr->nres.evcn) + 1;
while ((attr = ni_enum_attr_ex(ni, attr, &le, &mi)) &&
attr->type == ATTR_DATA && !attr->name_len) {
le64_add_cpu(&attr->nres.svcn, len);
le64_add_cpu(&attr->nres.evcn, len);
...
}
The loop condition checks only the attribute type and name. The
attributes come from ni_enum_attr_ex(), which walks the attribute
list into extension MFT records, and does not imply any residency
check. nres.evcn has no meaning for a resident attribute: in that
layout the first four bytes at the same offset are res.data_size,
and the 8-byte RMW reaches past the record buffer when the attribute
is placed at its tail.
Both branches are reachable from the same fallocate() syscall; the
resident one is fixed, this one is not. Current mainline
(fs/ntfs3/attrib.c) still has the loop verbatim.
## Suggested fix
Reject resident attributes in the enumeration loop:
while ((attr = ni_enum_attr_ex(ni, attr, &le, &mi)) &&
attr->type == ATTR_DATA && !attr->name_len) {
+ if (!attr->non_res) {
+ err = -EFSCORRUPTED;
+ goto out;
+ }
le64_add_cpu(&attr->nres.svcn, len);
le64_add_cpu(&attr->nres.evcn, len);
Checking at the ni_enum_attr_ex() level would cover other callers
with the same assumption, but the local check is the minimal fix
for this path.
## Reproducer
The crafted image places a resident DATA attribute at the end of an
extension MFT record so that the enumerated attribute is resident and
tail-aligned. The image builder (mk_patch_record.py), positive and
negative images, a trigger program and a syzkaller-format
`syz_mount_image$ntfs3` reproducer (image embedded as zlib+base64,
followed by openat + fallocate) are in the test bundle.
Available on request.
Tested on 7.3.0 (KASAN, ntfs3 built-in). mainline 2026-11-15 checked.
Thanks,
Guikun Yang
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2026-11-17 1:48 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-15 12:00 [Bug report] ntfs3: attr_insert_range() shifts nres.evcn of a resident attribute past the end of the MFT record buffer Guikun Yang
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®