Re: [Xen-devel] [PATCH RFC] xen-block: correctly define structures in public headers

["Roger Pau Monné" <roger.pau@citrix.com>]

On 04/12/13 10:28, Ian Campbell wrote:
> On Tue, 2013-12-03 at 15:11 -0500, Konrad Rzeszutek Wilk wrote:
>>>> If Konrad and Boris agree that breaking the kernel's ABI in this way is
>>>> acceptable in this specific case, I'll defer to them.
>>>
>>> My opinion as Xen on ARM hypervisor maintainer is that this is the right
>>> thing to do in this case.
>>
>> Heh. If somebody can guarantee me that (by testing the right variants and
>> mentioning this in the git commit) that this does not break x86, then
>> I am fine.
>>
>> And by 'break x86' I mean that this combination works:
>>  32-bit domU on 64-bit dom0
>>  64-bit domU on 32-bit dom0
>>
>> And perhaps also the obvious:
>>  64-bit domU on 64-bit dom0
>>  32-bit domU on 32-bit dom0
> 
> One way to test this is with gdb on a vmlinux for each arch with
> CONFIG_DEBUG_INFO=y. For each MEMBER of each interesting STRUCT:
>         (gdb) print &((struct STRUCT *)0)->MEMBER
> (this is effectively an open coded offsetof)
> 
> This could probably even be semi automated by producing a script to feed
> to gdb which run through all of the options and diffing the result.
> 
> If I could have the moon on a stick I would have a tool such as this
> running against the canonical Xen headers, to catch breakage as it is
> introduced upstream and a tool which could run against an arbitrary ELF
> binary to validate it against the upstream results.
> tools/include/xen-foreign/mkchecker.py goes some way towards that but
> isn't really extensible to the extent we would need/want.
> 
> While I'm asking for unicorns a gcc __attribute__((warn_on_holes)) which
> could be applied to a struct to enforce the need for explicit padding
> would probably be incredibly useful for this of thing.

Right now I would be happy to add something like:

#if !defined(CONFIG_X86_32) && !defined(CONFIG_X86_64) &&
!defined(CONFIG_ARM...
#error This architecture is not supported by the Xen PV block ABI
#endif

To the Linux copy of blkif.h

>> Since the xen-blkback has its own version of the structs there is no
>> need to change change newer and older version of it.
> 
> Someone should check that these are producing the right interface on ARM
> though!

AFAICT blkback on ARM is not using the structures defined in
blkback/common.h, since the protocol is "native", it's using the same
structures defined in the public header (just as blkfront). There's no
translation needed and blkback just does a memcpy from the ring to the
native struct.