Re: [PATCH v3] ntfs3: fix out-of-bounds write in do_action() bitmap operations

From: Greg KH

Date: Sat Nov 21 2026 - 07:57:39 EST


On Sat, Nov 21, 2026 at 01:57:24PM +0100, Greg KH wrote:
> On Sat, Nov 21, 2026 at 05:54:17PM +0530, Tanuj Sharma wrote:
> > In do_action(), when replaying log records for ATTR_ALLOC attributes,
> > the buffer is allocated based on bytes_per_index (typically 4096 bytes,
> > kmalloc-4k) but bitmap operation bounds checking uses lco -- a larger
> > value derived from lcns_follow << cluster_bits (typically 8192).
> >
> > This allows SetBitsInNonresidentBitMap and ClearBitsInNonresidentBitMap
> > opcodes to write past the allocated buffer into the adjacent slab object
> > when processing crafted NTFS images with malicious $LogFile records.
> >
> > The bounds check compares against lco (the full LCN extent size) but
> > should compare against the actual buffer capacity. For ATTR_ALLOC, the
> > buffer is only bytes_per_index bytes, allocated here:
> >
> > if (attr->type == ATTR_ALLOC) {
> > t32 = le32_to_cpu(oe->bytes_per_index);
> > if (bytes < t32)
> > bytes = t32;
> > }
> > buffer_le = kmalloc(bytes, GFP_NOFS);
> >
> > But the bounds check uses the larger lco:
> >
> > if (cbo + (off + 7) / 8 > lco || ...)
> >
> > With bytes_per_index=4096 and lco=8192, any bitmap_off >= 32768
> > (byte offset 4096) passes the lco check but writes past the buffer.
> >
> > Fix this by computing the bounds limit as the minimum of the buffer
> > extent (cbo + bytes_per_index) and the LCN extent (lco). This ensures
> > bitmap operations cannot exceed either the allocated buffer or the
> > on-disk extent, and also prevents a crafted bytes_per_index larger
> > than lco from weakening the existing check.
> >
> > Found by code audit and confirmed with a crafted NTFS image that
> > triggers the OOB write on mount.
> >
> > Cc: stable@xxxxxxxxxxxxxxx
> > Fixes: b46acd6a6a62 ("fs/ntfs3: Add NTFS journal")
> > Signed-off-by: Tanuj Sharma <ertanuj96@xxxxxxxxx>
> > ---
> > Changes in v3:
> > - Fix coding style: move bmp_limit declaration to function scope,
> > remove bare blocks
> >
> > Changes in v2:
> > - Resend with proper formatting (git send-email)
> >
> > fs/ntfs3/fslog.c | 17 +++++++++++++----
> > 1 file changed, 13 insertions(+), 4 deletions(-)
> >
> > diff --git a/fs/ntfs3/fslog.c b/fs/ntfs3/fslog.c
> > index 1879d04e3cd05..373d84545a8fe 100644
> > --- a/fs/ntfs3/fslog.c
> > +++ b/fs/ntfs3/fslog.c
> > @@ -3125,6 +3125,7 @@ static int do_action(struct ntfs_log *log, struct OPEN_ATTR_ENRTY *oe,
> > u16 roff = le16_to_cpu(lrh->record_off);
> > u16 aoff = le16_to_cpu(lrh->attr_off);
> > u64 lco = 0;
> > + u64 bmp_limit;
> > u64 cbo = (u64)le16_to_cpu(lrh->cluster_off) << SECTOR_SHIFT;
> > u64 tvo = le64_to_cpu(lrh->target_vcn) << sbi->cluster_bits;
> > u64 vbo = cbo + tvo;
> > @@ -3751,8 +3752,12 @@ static int do_action(struct ntfs_log *log, struct OPEN_ATTR_ENRTY *oe,
> > off = le32_to_cpu(((struct BITMAP_RANGE *)data)->bitmap_off);
> > bits = le32_to_cpu(((struct BITMAP_RANGE *)data)->bits);
> >
> > - if (cbo + (off + 7) / 8 > lco ||
> > - cbo + ((off + bits + 7) / 8) > lco) {
> > + bmp_limit = attr->type == ATTR_ALLOC ?
> > + min(cbo + (u64)le32_to_cpu(oe->bytes_per_index), lco) :
> > + lco;
> > +
> > + if (cbo + (off + 7) / 8 > bmp_limit ||
> > + cbo + ((off + bits + 7) / 8) > bmp_limit) {
> > goto dirty_vol;
> > }
>
> Did you run checkpatch on this? Why are these { } needed here?
>
> >
> > @@ -3764,8 +3769,12 @@ static int do_action(struct ntfs_log *log, struct OPEN_ATTR_ENRTY *oe,
> > off = le32_to_cpu(((struct BITMAP_RANGE *)data)->bitmap_off);
> > bits = le32_to_cpu(((struct BITMAP_RANGE *)data)->bits);
> >
> > - if (cbo + (off + 7) / 8 > lco ||
> > - cbo + ((off + bits + 7) / 8) > lco) {
> > + bmp_limit = attr->type == ATTR_ALLOC ?
> > + min(cbo + (u64)le32_to_cpu(oe->bytes_per_index), lco) :
> > + lco;
> > +
> > + if (cbo + (off + 7) / 8 > bmp_limit ||
> > + cbo + ((off + bits + 7) / 8) > bmp_limit) {
> > goto dirty_vol;
> > }
>
> Or here?

Also, slow down, there is no rush here. Always wait at LEAST 24 hours
before resending a new version, ideally longer.

thanks,

greg k-h