Re: [PATCH 1/5] __fdelt_chk: Removed range check

["Carlos O'Donell" <carlos at redhat dot com>]

On 05/01/2013 06:13 PM, KOSAKI Motohiro wrote:
>>>> That means that we would have to recompile all of the
>>>> applications again in order to get checking again using
>>>> the new symbols proposed in PATCH #2?
>>>
>>> Right. Because, unfortunately, __fdelt_chk() doesn't have
>>> buffer size argument, so we can't implement buffer overflow
>>> checks on top of this interface.
>>>
>>> Then, I made new __fdelt_buffer_chk() function at patch #2.
>>>
>>> The rest problem is, how should we treat old interfaces? From
>>> point of Ubuntu and OpenSUSE view, it should be disable, at least,
>>> by default. Otherwise all applications need to recompile for disabling.
>>>
>>>
>>>> This is not sufficiently conservative. We want it the other
>>>> way around. A simple recompile of ruby should result in
>>>> a ruby that no longer needs to disable _FORTIFY_SOURCE
>>>> to work around FD_SETSIZE checks.
>>>
>>> If anyone have an alternative and better implementation idea, that's
>>> welcome. I definitely agree this is ideal result.
>>  
>> I don't think we want to disable the check.
>>
>> We added it for good reasons and it matches POSIX behaviour.
>>
>> At the end of the day we implement POSIX behaviour by default.
> 
> Could you clarify me my two questionsn? 

Adding distribution maintainers from:
http://sourceware.org/glibc/wiki/MAINTAINERS#Distribution_Maintainers
to the TO, including Andreas Jaegar (openSUSE) and 
Adam Conrad (Ubuntu(Canonical)).

> 1) If not disabling, Ubuntu/OpenSUSE need to recompile ALL of affected 
> packages. Do you suggest to recompile all of them? 

No. They need only recompile those applications that don't
conform to POSIX and expect a Linux-style select with support for
fds > 1024. Once recompiled we rely on __bos0 to make the
check dynamic, if it doesn't, then the package will need to disable
the check with a special fortify source flag.

At the end of the day the community has worked hard to enable
_FORTIFY_SOURCE. The distributions made a *conscious* choice to
turn on these checks. We are doing them a disservice if we disable
the checks in glibc because applications are having problems.

It was the distribution's choice to enable the check. The package
maintainers need to work with the distribution and glibc to understand
the issues and fix their code. Nothing is fixed by Ruby calling
glibc "wrong" e.g. "the implementation is wrong... it wrongly assume
fd is always less than FD_SETSIZE (i.e. 1024)...". It's not wrong,
it's POSIX ;-)
 
> Dear Ubuntu and OpenSUSE guys, please tell me your opinion. This patch
> doesn't affect Fedora/RHEL/Debian. So I want to know which is close to
> your desire.

Added Andreas and Adam to the TO.

> 2) I think my code correctly catch buffer overflow of posix codes too.
> Do you disagree it? In the other words, what kind of application wants 
> to static checks?

Your patch disables the check for existing binaries, therefore it
doesn't catch buffer overflow for existing binaries e.g. indexing
into the fds array with an out of bounds index.

The check enforces POSIX. It enforces an expectation for all
applications that expect POSIX behaviour, which is what glibc
implements.

The root of the problem is the question:
What does glibc implement? POSIX? or The Linux API?

The project implements POSIX and more, but where the two are in
conflict we need to work together to make an informed choice.

Here the consensus is coming around to:

* When compiling with _FORTIFY_SOURCE, accept a strict definition
  for the API and assume POSIX compliance.

* Allow an application to loosen the check if it expects Linux-style
  behaviour.

What kind of application wants static checks?

* Any application using `fd_set foo;' and conforming to POSIX.

There is a lot of example code that uses `fd_set xxx;' and then
manipulates xxx using the provided macros. That code benefits
from a static check. I'll bet, but have no proof, that you'd
find more of that code than code that malloc's the fd_set.
Which is the whole reason the default check in _FORTIFY_SOURCE
is to be conservative and check against <= FD_SETSIZE.

Does that answer your question?
 
>> Ruby has already worked around this by disabling _FORTIFY_SOURCE
>> in their code to avoid the assert.
> 
> Right. I don't mind ruby case.
> 
> 
>> What we want to do is prevent them from needing to disable
>> *ALL* of _FORTIFY_SOURCE and provide a macro that allows them
>> to have finer grained control over the checks that apply to their
>> code.
> 
> It can. but I want to understand which is better and why before to do.
> 
> 
>