Re: [PATCH RFC] make error codes a formal part of the ABI

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 14/01/15 11:24, Ian Campbell wrote:
> On Wed, 2015-01-14 at 10:52 +0000, Julien Grall wrote:
>> Hi Ian,
>>
>> On 14/01/2015 10:28, Ian Campbell wrote:
>>>>> (I suppose someone needs to patch libxc et al to actually use this)
>>>> The primary consumer, as said in the description, is meant to be
>>>> hvmloader. But yes, other tools parts may also want to follow
>>>> that.
>>> /me wonders how libxc has been getting away with out this until now,
>>> especially on non-Linux platforms.
>> On FreeBSD, we translate the XEN errno to the guest one in the privcmd 
>> drivers.
> Ah, I looked in tools/libxc/*bsd* but didn't think to look in the kernel
> itself.
>
>> See 
>> http://xenbits.xen.org/gitweb/?p=people/royger/freebsd.git;a=patch;h=f09eeed01bc884b37e978f6ec6e3e7a86778ef4b
> OK, so essentially the requirement is that libxc should see the guest's
> own ERRNO numbers and changing libxc to use xen's numbers would just be
> wrong.
>
> Which is good, because making that change would be a nightmare!

It is not so simple.

Currently, the hypercall ioctl overloads errno from both the kernel and
Xen.  This makes it impossible for libxc to programmaticly distinguish
an kernel error from a Xen error (e.g. where did an EFAULT come from).

Ideally, the correct solution would be for the ioctl to return -1/errno
for a kernel failure alone, and then have the top level stub inspect
args[0] for the Xen error.  This would require libxc to know about both
the local errnos and Xen errnos.

~Andrew