Re: [Qemu-devel] [PATCH] linux-user: Add binfmt wrapper

[Alexander Graf <agraf@suse.de>]

On 14.07.14 17:38, Joakim Tjernlund wrote:
> Alexander Graf <agraf@suse.de> wrote on 2014/07/14 17:21:33:
>
>> From: Alexander Graf <agraf@suse.de>
>> To: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>,
>> Cc: qemu-devel@nongnu.org
>> Date: 2014/07/14 17:21
>> Subject: Re: [PATCH] linux-user: Add binfmt wrapper
>>
>>
>> On 14.07.14 16:38, Joakim Tjernlund wrote:
>>> The popular binfmt-wrapper patch adds an additional
>>> executable which mangle argv suitable for binfmt flag P.
>>> In a chroot you need the both (statically linked) qemu-$arch
>>> and qemu-$arch-binfmt-wrapper. This is sub optimal and a
>>> better approach is to recognize the -binfmt-wrapper extension
>>> within linux-user(qemu-$arch) and mangle argv there.
>>> This just produces on executable which can be either copied to
>>> the chroot or bind mounted with the appropriate -binfmt-wrapper
>>> suffix.
>>>
>>> Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
>> Please make sure to CC Riku on patches like this - he is the linux-user
>> maintainer.
> Doesn't he read the devel list? Anyhow CC:ed

He may or may not. Qemu-devel can be pretty high volume :).

>
>>> ---
>>>    linux-user/main.c | 13 +++++++++++++
>>>    1 file changed, 13 insertions(+)
>>>
>>> diff --git a/linux-user/main.c b/linux-user/main.c
>>> index 71a33c7..212067a 100644
>>> --- a/linux-user/main.c
>>> +++ b/linux-user/main.c
>>> @@ -3828,6 +3828,19 @@ int main(int argc, char **argv, char **envp)
>>>        int i;
>>>        int ret;
>>>        int execfd;
>>> +    char *binfmt;
>>> +
>>> +    i = strlen( argv[0] ) - strlen ( "-binfmt-wrapper" );
>> The spaces are odd. Did this patch pass checkpatch.pl? Same comment goes
>> for almost all function invocations.
> ehh, didn't run it through checkpatch.pl. Easy to fix next time.
>
>>> +    binfmt = argv[0] + i;
>>> +    if (i > 0 && strcmp ( binfmt, "-binfmt-wrapper" ) == 0) {
>> This magic needs to be documented somewhere. In fact, I find it pretty
>> hard to use in real world scenarios. Imagine a distribution - should it
>> package every target binary twice? Should it create hardlinks all over?
> How does dists. handle your original binfmt-wrapper? This is not much
> different I think. Here you got a choice to create a hardlink or a copy.
> Any chroot will only have to bind mount binfmt-wrapper into the chroot or
> lxc container.

Yeah, and there are reasons my original approach isn't upstream :).

>
>> I think we should try and find better magic :). Looking at the
>> binfmt_misc loading code, I think we can cheat a bit. If we pass the 'O'
>> flag (open target binary for handler), binfmt_misc will tell us the
>> binary fd in AT_EXECFD:
>>
>>                   NEW_AUX_ENT(AT_EXECFD, bprm->interp_data);
>>
>> We could then use this as a hint that we were spawned by binfmt_misc
>> rather than directly and interpret the first argv as target_argv[0].
>>
>> Then we can also add the P and O flags to scripts/qemu-binfmt-conf.sh
>> and have a solution that works well for everyone.
> What to do with P only then? Seems like most dists uses only P

If a distro uses the P flag it's not using upstream code, so they have 
to deal with their own breakage :). Fortunately the binfmt install 
scripts are usually part of a package too, so they can be updated easily.

If a distro cares a lot about backwards compatibility with their old 
name space, they can still compile the old -binfmt wrapper code and ship it.


Alex