Messages in this thread | | | From | Hawkins Jiawei <> | Subject | Re: [PATCH 2/3] ntfs: fix out-of-bounds read in ntfs_attr_find() | Date | Wed, 31 Aug 2022 20:03:25 +0800 |
| |
On Wed, 31 Aug 2022 at 19:08, Dan Carpenter <dan.carpenter@oracle.com> wrote: > > On Wed, Aug 31, 2022 at 10:43:36AM +0800, Hawkins Jiawei wrote: > > Kernel will iterates over ATTR_RECORDs in mft record in ntfs_attr_find(). > > To ensure access on these ATTR_RECORDs are within bounds, kernel will > > do some checking during iteration. > > > > The problem is that during checking whether ATTR_RECORD's name is within > > bounds, kernel will dereferences the ATTR_RECORD name_offset field, > > before checking this ATTR_RECORD strcture is within bounds. This problem > > may result out-of-bounds read in ntfs_attr_find(), reported by > > Syzkaller: > > > > ================================================================== > > BUG: KASAN: use-after-free in ntfs_attr_find+0xc02/0xce0 fs/ntfs/attrib.c:597 > > Read of size 2 at addr ffff88807e352009 by task syz-executor153/3607 > > > > [...] > > Call Trace: > > <TASK> > > __dump_stack lib/dump_stack.c:88 [inline] > > dump_stack_lvl+0xcd/0x134 lib/dump_stack.c:106 > > print_address_description mm/kasan/report.c:317 [inline] > > print_report.cold+0x2ba/0x719 mm/kasan/report.c:433 > > kasan_report+0xb1/0x1e0 mm/kasan/report.c:495 > > ntfs_attr_find+0xc02/0xce0 fs/ntfs/attrib.c:597 > > ntfs_attr_lookup+0x1056/0x2070 fs/ntfs/attrib.c:1193 > > ntfs_read_inode_mount+0x89a/0x2580 fs/ntfs/inode.c:1845 > > ntfs_fill_super+0x1799/0x9320 fs/ntfs/super.c:2854 > > mount_bdev+0x34d/0x410 fs/super.c:1400 > > legacy_get_tree+0x105/0x220 fs/fs_context.c:610 > > vfs_get_tree+0x89/0x2f0 fs/super.c:1530 > > do_new_mount fs/namespace.c:3040 [inline] > > path_mount+0x1326/0x1e20 fs/namespace.c:3370 > > do_mount fs/namespace.c:3383 [inline] > > __do_sys_mount fs/namespace.c:3591 [inline] > > __se_sys_mount fs/namespace.c:3568 [inline] > > __x64_sys_mount+0x27f/0x300 fs/namespace.c:3568 > > do_syscall_x64 arch/x86/entry/common.c:50 [inline] > > do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80 > > entry_SYSCALL_64_after_hwframe+0x63/0xcd > > [...] > > </TASK> > > > > The buggy address belongs to the physical page: > > page:ffffea0001f8d400 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x7e350 > > head:ffffea0001f8d400 order:3 compound_mapcount:0 compound_pincount:0 > > flags: 0xfff00000010200(slab|head|node=0|zone=1|lastcpupid=0x7ff) > > raw: 00fff00000010200 0000000000000000 dead000000000122 ffff888011842140 > > raw: 0000000000000000 0000000000040004 00000001ffffffff 0000000000000000 > > page dumped because: kasan: bad access detected > > Memory state around the buggy address: > > ffff88807e351f00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc > > ffff88807e351f80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc > > >ffff88807e352000: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb > > ^ > > ffff88807e352080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb > > ffff88807e352100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb > > ================================================================== > > > > This patch solves it by moving the ATTR_RECORD strcture's bounds > > checking earlier, then checking whether ATTR_RECORD's name > > is within bounds. What's more, this patch also add some comments > > to improve its maintainability. > > > > Reported-and-tested-by: syzbot+5f8dcabe4a3b2c51c607@syzkaller.appspotmail.com > > Signed-off-by: chenxiaosong (A) <chenxiaosong2@huawei.com> > > Link: https://lore.kernel.org/all/1636796c-c85e-7f47-e96f-e074fee3c7d3@huawei.com/ > > Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> > > Link: https://groups.google.com/g/syzkaller-bugs/c/t_XdeKPGTR4/m/LECAuIGcBgAJ > > Signed-off-by: Hawkins Jiawei <yin31149@gmail.com> > > --- > > fs/ntfs/attrib.c | 20 ++++++++++++++++---- > > 1 file changed, 16 insertions(+), 4 deletions(-) > > > > diff --git a/fs/ntfs/attrib.c b/fs/ntfs/attrib.c > > index 52615e6090e1..904734e34507 100644 > > --- a/fs/ntfs/attrib.c > > +++ b/fs/ntfs/attrib.c > > @@ -594,11 +594,23 @@ static int ntfs_attr_find(const ATTR_TYPE type, const ntfschar *name, > > for (;; a = (ATTR_RECORD*)((u8*)a + le32_to_cpu(a->length))) { > > u8 *mrec_end = (u8 *)ctx->mrec + > > le32_to_cpu(ctx->mrec->bytes_allocated); > > - u8 *name_end = (u8 *)a + le16_to_cpu(a->name_offset) + > > - a->name_length * sizeof(ntfschar); > > - if ((u8*)a < (u8*)ctx->mrec || (u8*)a > mrec_end || > > - name_end > mrec_end) > > + u8 *name_end, *arec_head_end; > > + > > + /* check for wrap around */ > > + if ((u8 *)a < (u8 *)ctx->mrec) > > + break; > > + > > + /* check whether Attribute Record Header is within bounds */ > > + arec_head_end = (u8 *)a + sizeof(ATTR_RECORD); > > + if (arec_head_end < (u8 *)a || arec_head_end > mrec_end) > > This works but I feel like it would be more natural to just check if > a was valid and if a + sizeof(ATTR_RECORD) was also valid. > > if (a > mrec_end || (u8 *)a + sizeof(ATTR_RECORD) > mrec_end) Hi Dan, Thanks for your suggestion. This looks more natural than original patch, yet I wonder if there may be an overflow?
To be more specific, if "a" and "mrec_end" is large enough, it seems that some fields of "a" may be out-of-bounds and also bypass this check because of the overflow.(Please correct me if I am wrong) > But what you have also works so if you want to go with that then that's > fine. > > regards, > dan carpenter > > > break; > > + > > + /* check whether ATTR_RECORD's name is within bounds */ > > + name_end = (u8 *)a + le16_to_cpu(a->name_offset) + > > + a->name_length * sizeof(ntfschar); > > + if (name_end > mrec_end) > > + break; > > + > > ctx->attr = a; > > if (unlikely(le32_to_cpu(a->type) > le32_to_cpu(type) || > > a->type == AT_END)) >
| |