Re: [RFC][PATCH 0/4] Prepare for supporting more filesystems with fanotify

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri 28-04-23 08:15:50, Jeff Layton wrote:
> On Fri, 2023-04-28 at 13:40 +0200, Jan Kara wrote:
> > On Thu 27-04-23 22:11:46, Amir Goldstein wrote:
> > > On Thu, Apr 27, 2023 at 7:36 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > > > > There is also a way to extend the existing API with:
> > > > > 
> > > > > Perhstruct file_handle {
> > > > >         unsigned int handle_bytes:8;
> > > > >         unsigned int handle_flags:24;
> > > > >         int handle_type;
> > > > >         unsigned char f_handle[];
> > > > > };
> > > > > 
> > > > > AFAICT, this is guaranteed to be backward compat
> > > > > with old kernels and old applications.
> > > > > 
> > > > 
> > > > That could work. It would probably look cleaner as a union though.
> > > > Something like this maybe?
> > > > 
> > > > union {
> > > >         unsigned int legacy_handle_bytes;
> > > >         struct {
> > > >                 u8      handle_bytes;
> > > >                 u8      __reserved;
> > > >                 u16     handle_flags;
> > > >         };
> > > > }
> > > 
> > > I have no problem with the union, but does this struct
> > > guarantee that the lowest byte of legacy_handle_bytes
> > > is in handle_bytes for all architectures?
> > > 
> > > That's the reason I went with
> > > 
> > > struct {
> > >          unsigned int handle_bytes:8;
> > >          unsigned int handle_flags:24;
> > > }
> > > 
> > > Is there a problem with this approach?
> > 
> > As I'm thinking about it there are problems with both approaches in the
> > uAPI. The thing is: A lot of bitfield details (even whether they are packed
> > to a single int or not) are implementation defined (depends on the
> > architecture as well as the compiler) so they are not really usable in the
> > APIs.
> > 
> > With the union, things are well-defined but they would not work for
> > big-endian architectures. We could make the structure layout depend on the
> > endianity but that's quite ugly...
> > 
> 
> Good point. Bitfields just have a bad code-smell anyway.
> 
> Another idea would be to allow someone to set handle_bytes to a
> specified value that's larger than the current max of 128 (maybe ~0 or
> something), and use that as an indicator that this is a v2 struct.
> 
> So the v2 struct would look something like:
> 
> struct file_handle_v2 {
> 	unsigned int	legacy_handle_bytes;	// always set to ~0 or whatever
> 	unsigned int	flags;
> 	int		handle_type;
> 	unsigned int	handle_bytes;
> 	unsigned char	f_handle[];
> 	
> };

Well, there's also always the option of having:

struct file_handle {
	unsigned int handle_bytes_flags;
	int handle_type;
	unsigned char f_handle[];
};

And then helper functions like:

unsigned int file_handle_bytes(struct file_handle *handle)
{
	return handle->handle_bytes_flags & 0xffff;
}

unsigned int file_handle_flags(struct file_handle *handle)
{
	return handle->handle_bytes_flags >> 16;
}

(and similar for forming the handle_bytes_flags value).

That is well defined and compatible across architectures and compilers and
bearable (although a bit clumsy).

> ...but now I'm wondering...why do we return -EINVAL when
> f_handle.handle_bytes is > MAX_HANDLE_SZ? Is it really wrong for the
> caller to allocate more space for the resulting file_handle than will be
> needed? That seems wrong too. In fact, name_to_handle_at(2) says:
> 
> "The constant MAX_HANDLE_SZ, defined in <fcntl.h>, specifies the maximum
> expected size for a file handle.  It  is  not guaranteed upper limit as
> future filesystems may require more space."
> 
> So by returning -EINVAL when handle_bytes is too large, we're probably
> doing the wrong thing there.

Yeah, you're right. But at this point it can serve us well by making sure
there's no userspace passing absurdly high handle_bytes ;).

								Honza

-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR



[Index of Archives]     [Linux Filesystems Devel]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux