Re: [msysGit] [PATCH] Prevent buffer overflows when path is too big

["Ondřej Bílka" <neleai@seznam.cz>]

On Sun, Oct 20, 2013 at 08:27:13AM +0200, Torsten Bögershausen wrote:
> Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on
> 	popelka.ms.mff.cuni.cz
> Status: O
> Content-Length: 2690
> Lines: 89
> 
> On 20.10.13 08:05, Ondřej Bílka wrote:
> > On Sun, Oct 20, 2013 at 07:47:06AM +0200, Torsten Bögershausen wrote:
> >> (may be s/path is too big/path is too long/ ?)
> >>
> >> On 19.10.13 12:52, Antoine Pelisse wrote:
> >>> Currently, most buffers created with PATH_MAX length, are not checked
> >>> when being written, and can overflow if PATH_MAX is not big enough to
> >>> hold the path.
> >>>
> >>> Fix that by using strlcpy() where strcpy() was used, and also run some
> >>> extra checks when copy is done with memcpy().
> >>>
> >>> Reported-by: Wataru Noguchi <wnoguchi.0727@gmail.com>
> >>> Signed-off-by: Antoine Pelisse <apelisse@gmail.com>
> >>> ---
> >>> diff --git a/abspath.c b/abspath.c
> >>> index 64adbe2..0e60ba4 100644
> >>> --- a/abspath.c
> >>> +++ b/abspath.c
> >>> @@ -216,11 +216,15 @@ const char *absolute_path(const char *path)
> >>>  const char *prefix_filename(const char *pfx, int pfx_len, const char *arg)
> >>>  {
> >>>  	static char path[PATH_MAX];
> > 
> > Why do you need static there?
> Good point.
> get_pathname() from path.c may be better.
> 
> >>> +
> >>> +	if (pfx_len > PATH_MAX)
> >> I think this should be 
> >> if (pfx_len > PATH_MAX-1) /* Keep 1 char for '\0'
> >>> +		die("Too long prefix path: %s", pfx);
> >>> +
> >>>  #ifndef GIT_WINDOWS_NATIVE
> >>>  	if (!pfx_len || is_absolute_path(arg))
> >>>  		return arg;
> >>>  	memcpy(path, pfx, pfx_len);
> >>> -	strcpy(path + pfx_len, arg);
> >>> +	strlcpy(path + pfx_len, arg, PATH_MAX - pfx_len);
> >>
> >> I'm not sure how to handle overlong path in general, there are several ways:
> >> a) Silently overwrite memory (with help of memcpy() and/or strcpy()
> >> b) Silently shorten the path using strlcpy() instead of strcpy()
> >> c) Avoid the overwriting and call die().
> >> d) Prepare a longer buffer using xmalloc()
> >>
> > There is also
> > e) modify allocation to place write protected page after buffer end.
> 
> Yes, I think this is what electric fence, DUMA or valgrind do:
> 
You need to be selective which buffers are important.

> http://sourceforge.jp/projects/freshmeat_efence/
> http://duma.sourceforge.net/

These are toys, this comes with fact that they need a 8kb of space for
each 8byte malloc. Just run a git diff and differences are huge.

$ /usr/bin/time git diff HEAD^ > x
0.06user 0.01system 0:00.07elapsed 97%CPU (0avgtext+0avgdata
19172maxresident)k
0inputs+8outputs (0major+5591minor)pagefaults 0swaps
$ LD_PRELOAD=libefence.so /usr/bin/time git diff HEAD^ > x

  Electric Fence 2.2 Copyright (C) 1987-1999 Bruce Perens
<bruce@perens.com>
3.49user 0.94system 0:04.45elapsed 99%CPU (0avgtext+0avgdata
91920maxresident)k
0inputs+8outputs (0major+118069minor)pagefaults 0swaps

> http://valgrind.sourceforge.net/
> 
> Theses are very good tools for developers, finding memory corruption
> (or other bugs like using uninitialized memory).
> 
> One of the motivation I asked for test cases is that a git developer can
> run these test cases under valgrind and can verify that we are never out of range.
> 
> For an end user a git "crash" caused by trying to write to a write protected page
> is better than silently corrupting memory.
> 
> And a range check, followed by die(), is even easier to debug.
> For an end user.
> /Torsten
>