Re: [f2fs-dev] [PATCH v3] f2fs: move ioctl interface definitions to separated file

[Chao Yu]

On 2020/11/3 11:22, Eric Biggers wrote:
> On Mon, Nov 02, 2020 at 02:21:31PM +0800, Chao Yu wrote:
> > +#define F2FS_IOC_MOVE_RANGE		_IOWR(F2FS_IOCTL_MAGIC, 9,	\
> > +						struct f2fs_move_range)
> [...]
> > +#define F2FS_IOC_GARBAGE_COLLECT_RANGE	_IOW(F2FS_IOCTL_MAGIC, 11,	\
> > +						struct f2fs_gc_range)
> [...]
> > +
> > +struct f2fs_gc_range {
> > +	__u32 sync;
> > +	__u64 start;
> > +	__u64 len;
> > +};
> [...]
> > +struct f2fs_move_range {
> > +	__u32 dst_fd;		/* destination fd */
> > +	__u64 pos_in;		/* start position in src_fd */
> > +	__u64 pos_out;		/* start position in dst_fd */
> > +	__u64 len;		/* size to move */
> > +};
>
> These two structs are weird because there is implicit padding between the __u32
> field and the following __u64 field on some 32-bit architectures (e.g. x86_32)
> but not others (e.g. arm32).
>
> But f2fs_compat_ioctl() doesn't handle these two ioctls specially, but rather
> just calls through to f2fs_ioctl().  That's wrong, and it means that
> F2FS_IOC_MOVE_RANGE and F2FS_IOC_GARBAGE_COLLECT_RANGE won't work when called
> from an x86_32 binary on an x86_64 kernel.

Nice catch!

> So something needs to be fixed.  I wonder if it's safe to just explicitly add
> the padding field after the fact.  If no one is actually using these two ioctls
> in a case where both userspace and the kernel lack the implicit padding (e.g.,
> x86_32 userspace with x86_32 kernel), it should be fine...

IIRC, Jaegeuk added those interfaces, I hope it's not the requirement from other
f2fs userspace developers...if it is, there may be users.

I found one patch in ext4 which fixes the similar issue, I guess we can fix this
with the same way, thoughts?

commit 4d92dc0f00a775dc2e1267b0e00befb783902fe7
Author: Ben Hutchings <ben@xxxxxxxxxxxxxxx>
Date:   Mon May 17 06:00:00 2010 -0400

    ext4: Fix compat EXT4_IOC_ADD_GROUP

    struct ext4_new_group_input needs to be converted because u64 has
    only 32-bit alignment on some 32-bit architectures, notably i386.

Thanks,

> - Eric
> .