Re: [RFC] ip tunnel flag byte order issue

On Wed, 10 Oct 2012 21:26:36 +0100
Ben Hutchings [off-list ref] wrote:

On Wed, 2012-10-10 at 12:06 -0700, Stephen Hemminger wrote:

quoted

Sparse found a real problem with the ABI for tunnelling.

The SIT and VTI tunnel ioctl's both overload the i_flags field in the
ip_tunnel parameters structure. This field is defined as big endian
(be16) and the various GRE_XXX macros do the necessary byte swapping.

The problem is that both SIT and VTI are using an additional flag bit
that is defined in host byte order, and is then or'd in. It happens to
work because both possible locations hit holes in the current usage of
GRE.  For big endian cpu's it overlaps the GRE_VERSION which is always
zero, and for little endian it overlaps the GRE recursion field also
always zero.

Why do these fields exist if they're always going to be 0?

They exist in the RFC. GRE implementation mixes bits on the wire
with bits from ioctl().

quoted

Having the field in different places on different CPU architectures
was a mistake. The problem is fixing it will break the ABI on one or
the other architecture.  I choose to break big endian since it the
minority.

Or we can define the 'flag' to have both bits set (0x0101, with a
__cpu_to_be16 to keep sparse happy) while accepting either set on input.

quoted

Also both VTI and SIT are overloading the same bit which is an
accident waiting to happen.  Since VTI is newer, I propose giving a
different bit to VTI.

Indeed VTI is new in 3.6, so there is still a short window in which it's
fairly safe to tweak its ABI.

quoted

The other alternative is keeping the same ABI, but putting a big note
as to why it works in spite of our stupidity.

[...]

Does it even matter that different tunnel types have different meanings
for flags?

Ben.