[Linux-kernel-mentees] deprecated.rst: deprecated strcpy ? (was: [PATCH] checkpatch: add a new check for strcpy/strlcpy uses)

[Kees Cook]

On Tue, Jan 05, 2021 at 01:28:18AM -0800, Joe Perches wrote:
> On Tue, 2021-01-05 at 14:29 +0530, Dwaipayan Ray wrote:
> > On Tue, Jan 5, 2021 at 2:14 PM Joe Perches <joe at perches.com> wrote:
> > > 
> > > On Tue, 2021-01-05 at 13:53 +0530, Dwaipayan Ray wrote:
> > > > strcpy() performs no bounds checking on the destination buffer.
> > > > This could result in linear overflows beyond the end of the buffer.
> > > > 
> > > > strlcpy() reads the entire source buffer first. This read
> > > > may exceed the destination size limit. This can be both inefficient
> > > > and lead to linear read overflows.
> > > > 
> > > > The safe replacement to both of these is to use strscpy() instead.
> > > > Add a new checkpatch warning which alerts the user on finding usage of
> > > > strcpy() or strlcpy().
> > > 
> > > I do not believe that strscpy is preferred over strcpy.
> > > 
> > > When the size of the output buffer is known to be larger
> > > than the input, strcpy is faster.
> > > 
> > > There are about 2k uses of strcpy.
> > > Is there a use where strcpy use actually matters?
> > > I don't know offhand...
> > > 
> > > But I believe compilers do not optimize away the uses of strscpy
> > > to a simple memcpy like they do for strcpy with a const from
> > > 
> > >         strcpy(foo, "bar");
> > > 
> > 
> > Yes the optimization here definitely helps. So in case the programmer
> > knows that the destination buffer is always larger, then strcpy() should be
> > preferred? I think the documentation might have been too strict about
> > strcpy() uses here:
> > 
> > Documentation/process/deprecated.rst:
> > "strcpy() performs no bounds checking on the destination buffer. This
> > could result in linear overflows beyond the end of the buffer, leading to
> > all kinds of misbehaviors. While `CONFIG_FORTIFY_SOURCE=y` and various
> > compiler flags help reduce the risk of using this function, there is
> > no good reason to add new uses of this function. The safe replacement
> > is strscpy(),..."
> 
> Kees/Jonathan:
> 
> Perhaps this text is overly restrictive.
> 
> There are ~2k uses of strcpy in the kernel.
> 
> About half of these are where the buffer length of foo is known and the
> use is 'strcpy(foo, "bar")' so the compiler converts/optimizes away the
> strcpy to memcpy and may not even put "bar" into the string table.
> 
> I believe strscpy uses do not have this optimization.
> 
> Is there a case where the runtime costs actually matters?
> I expect so.

The original goal was to use another helper that worked on static
strings like this. Linus rejected that idea, so we're in a weird place.
I think we could perhaps build a strcpy() replacement that requires
compile-time validated arguments, and to break the build if not.

i.e.

given:
	char array[8];
	char *ptr;

allow:


	strcpy(array, "1234567");

disallow:

	strcpy(array, "12345678");	/* too long */
	strcpy(array, src);		/* not optimized, so use strscpy? */
	strcpy(ptr, "1234567");		/* unknown destination size */
	strcpy(ptr, src);		/* unknown destination size */

What do you think?

-- 
Kees Cook