Linux Ethernet Bridging - Jul 2020 - [Bridge] linux-next: umh: fix processed error when UMH_WAIT_PROC is used seems to break linux bridge on s390x (bisected)

On 2020/07/01 19:08, Christian Borntraeger wrote:
> 
> 
> On 30.06.20 19:57, Luis Chamberlain wrote:
>> On Fri, Jun 26, 2020 at 02:54:10AM +0000, Luis Chamberlain wrote:
>>> On Wed, Jun 24, 2020 at 08:37:55PM +0200, Christian Borntraeger
wrote:
>>>>
>>>>
>>>> On 24.06.20 20:32, Christian Borntraeger wrote:
>>>> [...]> 
>>>>> So the translations look correct. But your change is
actually a sematic change
>>>>> if(ret) will only trigger if there is an error
>>>>> if (KWIFEXITED(ret)) will always trigger when the process
ends. So we will always overwrite -ECHILD
>>>>> and we did not do it before. 
>>>>>
>>>>
>>>> So the right fix is
>>>>
>>>> diff --git a/kernel/umh.c b/kernel/umh.c
>>>> index f81e8698e36e..a3a3196e84d1 100644
>>>> --- a/kernel/umh.c
>>>> +++ b/kernel/umh.c
>>>> @@ -154,7 +154,7 @@ static void
call_usermodehelper_exec_sync(struct subprocess_info *sub_info)
>>>>                  * the real error code is already in
sub_info->retval or
>>>>                  * sub_info->retval is 0 anyway, so
don't mess with it then.
>>>>                  */
>>>> -               if (KWIFEXITED(ret))
>>>> +               if (KWEXITSTATUS(ret))
>>>>                         sub_info->retval =
KWEXITSTATUS(ret);
Well, it is not br_stp_call_user() but br_stp_start() which is expecting
to set sub_info->retval for both KWIFEXITED() case and KWIFSIGNALED() case.
That is, sub_info->retval needs to carry raw value (i.e. without "umh:
fix
processed error when UMH_WAIT_PROC is used" will be the correct behavior).

On Wed, Jul 01, 2020 at 10:24:29PM +0900, Tetsuo Handa
wrote:
> On 2020/07/01 19:08, Christian Borntraeger wrote:
> > 
> > 
> > On 30.06.20 19:57, Luis Chamberlain wrote:
> >> On Fri, Jun 26, 2020 at 02:54:10AM +0000, Luis Chamberlain wrote:
> >>> On Wed, Jun 24, 2020 at 08:37:55PM +0200, Christian
Borntraeger wrote:
> >>>>
> >>>>
> >>>> On 24.06.20 20:32, Christian Borntraeger wrote:
> >>>> [...]> 
> >>>>> So the translations look correct. But your change is
actually a sematic change
> >>>>> if(ret) will only trigger if there is an error
> >>>>> if (KWIFEXITED(ret)) will always trigger when the
process ends. So we will always overwrite -ECHILD
> >>>>> and we did not do it before. 
> >>>>>
> >>>>
> >>>> So the right fix is
> >>>>
> >>>> diff --git a/kernel/umh.c b/kernel/umh.c
> >>>> index f81e8698e36e..a3a3196e84d1 100644
> >>>> --- a/kernel/umh.c
> >>>> +++ b/kernel/umh.c
> >>>> @@ -154,7 +154,7 @@ static void
call_usermodehelper_exec_sync(struct subprocess_info *sub_info)
> >>>>                  * the real error code is already in
sub_info->retval or
> >>>>                  * sub_info->retval is 0 anyway, so
don't mess with it then.
> >>>>                  */
> >>>> -               if (KWIFEXITED(ret))
> >>>> +               if (KWEXITSTATUS(ret))
> >>>>                         sub_info->retval =
KWEXITSTATUS(ret);
> 
> Well, it is not br_stp_call_user() but br_stp_start() which is expecting
> to set sub_info->retval for both KWIFEXITED() case and KWIFSIGNALED()
case.
> That is, sub_info->retval needs to carry raw value (i.e. without
"umh: fix
> processed error when UMH_WAIT_PROC is used" will be the correct
behavior).
br_stp_start() doesn't check for the raw value, it just checks for err
or !err. So the patch, "umh: fix processed error when UMH_WAIT_PROC is
used" propagates the correct error now.

Christian, can you try removing the binary temporarily and seeing if
you get your bridge working?

  Luis

On 2020/07/01 22:53, Luis Chamberlain wrote:
>> Well, it is not br_stp_call_user() but br_stp_start() which is
expecting
>> to set sub_info->retval for both KWIFEXITED() case and
KWIFSIGNALED() case.
>> That is, sub_info->retval needs to carry raw value (i.e. without
"umh: fix
>> processed error when UMH_WAIT_PROC is used" will be the correct
behavior).
> 
> br_stp_start() doesn't check for the raw value, it just checks for err
> or !err. So the patch, "umh: fix processed error when UMH_WAIT_PROC is
> used" propagates the correct error now.
No. If "/sbin/bridge-stp virbr0 start" terminated due to e.g. SIGSEGV
(for example, by inserting "kill -SEGV $$" into right after
"#!/bin/sh" line),
br_stp_start() needs to select BR_KERNEL_STP path. We can't assume that
/sbin/bridge-stp is always terminated by exit() syscall (and hence we can't
ignore KWIFSIGNALED() case in call_usermodehelper_exec_sync()).