Re: [PATCH 8/9 v2] Implement vFile:setfs in gdbserver

[Pedro Alves <palves@redhat.com>]

On 06/10/2015 10:40 AM, Gary Benson wrote:
> Gary Benson wrote:
>> Pedro Alves wrote:
>>> On 06/09/2015 03:11 PM, Gary Benson wrote:
>>>> Pedro Alves wrote:
>>>>> For the setns/ENOSYS case, how about somehow probing whether
>>>>> setns actually works here (with a new method, or probing one
>>>>> of the existing ones with getpid()) and return empty packet,
>>>>> instead of ending up with open failing with ENOSYS later on?
>>>>
>>>> Probing and then telling GDB we don't understand "vFile:setfs:"
>>>> isn't that good of an idea.  If we do that with an inferior in
>>>> another mount namespace then the end result is that gdbserver will
>>>> access the wrong filesystem.  You might get "file not found", or
>>>> you might get an open filehandle on another file (i.e. the file
>>>> with the same name but in gdbserver's filesystem).  If the
>>>> inferior's filesystem is different from gdbserver and gdbserver
>>>> cannot access that filesystem then the only correct response is
>>>> to fail.
>>>
>>> If the running system does not support "setfs", because it fails
>>> setns with ENOSYS, meaning, the system call isn't implemented at
>>> all, how can one end up in the situation that an inferior on the
>>> same kernel is running in a different filesystem namespace?
>>
>> If the running kernel has namespaces support but GDB and/or glibc
>> was built on a kernel without.  It's an edge case but it's possible.

I see.

> 
> Note that GDB/gdbserver will not attempt to call setns unless the
> inferior is actually in a different mount namespace.  If you're
> running without namespaces support it won't even start the helper
> let alone try to setns.  Same goes for if you have namespaces
> support but the inferior is in GDB/gdbserver's mount namespace.
> Nothing calls setns unless it is 100% necessary.

To probe for setns I was thinking would be done by on
gdb/gdbserver's pid, by gdb/gdbserver directly.

But I agree, when the kernel supports namespaces, and we can
tell the inferior is in a different namespace, but then setns
doesn't work, it's better to error out than disabling the setfs
packet.  So I'm happy with the errno translation alone.

Thanks,
Pedro Alves