On Mon, 28 Jul 2025 17:35:26 +0100, Al Viro wrote:
> On Mon, Jul 28, 2025 at 05:04:45PM +0100, Al Viro wrote:
> > On Mon, Jul 28, 2025 at 07:37:02PM +0800, Edward Adam Davis wrote:
> > > The writer and reader access FAT32 entry without any lock, so the data
> > > obtained by the reader is incomplete.
> >
> > Could you be more specific? "Incomplete" in which sense?
Because ent32_p and ent12_p are in the same union [1], their addresses are
the same, and they both have the "read/write race condition" problem, so I
used the same method as [2] to add a spinlock to solve it.
> >
> > > Add spin lock to solve the race condition that occurs when accessing
> > > FAT32 entry.
> >
> > Which race condition would that be?
data-race in fat32_ent_get / fat32_ent_put, detail: see [3]
> >
> > > FAT16 entry has the same issue and is handled together.
> >
> > FWIW, I strongly suspect that
> > * "issue" with FAT32 is a red herring coming from mindless parroting
> > of dumb tool output
> > * issue with FAT16 just might be real, if architecture-specific.
> > If 16bit stores are done as 32bit read-modify-write, we might need some
> > serialization. Assuming we still have such architectures, that is -
> > alpha used to be one, but support for pre-BWX models got dropped.
> > Sufficiently ancient ARM?
>
> Note that FAT12 situation is really different - we not just have an inevitable
> read-modify-write for stores (half-byte access), we are not even guaranteed that
> byte and half-byte will be within the same cacheline, so cmpxchg is not an
> option; we have to use a spinlock there.
I think for FAT12 they are always in the same cacheline, the offset of the
member ent12_p in struct fat_entry is 4 bytes, and no matter x86-64 or arm64,
the regular 64-byte cacheline is enough to ensure that they are in the same
cacheline.
[1]
345 struct fat_entry {
1 int entry;
2 union {
3 u8 *ent12_p[2];
4 __le16 *ent16_p;
5 __le32 *ent32_p;
6 } u;
7 int nr_bhs;
8 struct buffer_head *bhs[2];
9 struct inode *fat_inode;
10 };
[2] 98283bb49c6c ("fat: Fix the race of read/write the FAT12 entry")
[3] https://syzkaller.appspot.com/bug?extid=d3c29ed63db6ddf8406e
BR,
Edward
