Re: [PATCH RFC 5/5] tools/nolibc: tests: add test for -fstack-protector

["Thomas Weißschuh" <linux@weissschuh.net>]

On Sun, Mar 12, 2023 at 02:07:16PM +0100, Willy Tarreau wrote:
> On Tue, Mar 07, 2023 at 10:22:34PM +0000, Thomas Weißschuh wrote:
> > Test the previously introduce stack protector functionality in nolibc.
> 
> s/introduce/introduced/
> 
> (I can adjust it myself when merging to avoid a respin if you want).

I respin is necessary anways.
I'll change it.

FYI there is also another patch to make nolibc-test buildable with
compilers that enable -fstack-protector by default.
Maybe this can be picked up until the proper stack-protector support is
hashed out.
Maybe even for 6.3:

https://lore.kernel.org/lkml/20230221-nolibc-no-stack-protector-v1-1-4e6a42f969e2@weissschuh.net/

> > Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
> > ---
> >  tools/testing/selftests/nolibc/nolibc-test.c | 74 +++++++++++++++++++++++++++-
> >  1 file changed, 72 insertions(+), 2 deletions(-)
> > 
> > diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c
> > index fb2d4872fac9..4990b2750279 100644
> > --- a/tools/testing/selftests/nolibc/nolibc-test.c
> > +++ b/tools/testing/selftests/nolibc/nolibc-test.c
> > @@ -45,6 +45,7 @@ char **environ;
> >  struct test {
> >  	const char *name;              // test name
> >  	int (*func)(int min, int max); // handler
> > +	char skip_by_default;         // don't run by default
> 
> Just a tiny detail but that comment is misaligned by one char on the left.

Ack.

> >  };
> >  
> >  #ifndef _NOLIBC_STDLIB_H
> > @@ -667,6 +668,70 @@ int run_stdlib(int min, int max)
> >  	return ret;
> >  }
> >  
> > +#if defined(__clang__)
> > +__attribute__((optnone))
> > +#elif defined(__GNUC__)
> > +__attribute__((optimize("O0")))
> > +#endif
> > +static int run_smash_stack(int min, int max)
> > +{
> > +	char buf[100];
> > +
> > +	for (size_t i = 0; i < 200; i++)
> > +		buf[i] = 15;
> 
> If the goal is to make it easy to spot in a crash dump, I suggest
> that you use a readable ASCII letter that's easy to recognize. 0xF
> will usually not be printed in hex dumps, making it less evident
> when scrolling quickly. For example I often use 'P' when poisoning
> memory but you get the idea.

Ack.

> > +int run_stackprotector(int min, int max)
> > +{
> > +	int llen = 0;
> > +
> > +	llen += printf("0 ");
> > +
> > +#if !defined(NOLIBC_STACKPROTECTOR)
> > +	llen += printf("stack smashing detection not supported");
> > +	pad_spc(llen, 64, "[SKIPPED]\n");
> > +	return 0;
> > +#endif
> 
> Shouldn't the whole function be enclosed instead ? I know it's more of
> a matter of taste, but avoiding to build and link it for archs that
> will not use it may be better.

The goal was to print a [SKIPPED] message if it's not supported.
The overhead of doing this should be neglectable.

> 
> > +
> > +	pid_t pid = fork();
> 
> Please avoid variable declarations after statements, for me these
> are really horrible to deal with when editing the code later, because
> instead of having to look up only the beginning of each containing
> block (i.e. in O(log(N))) you have to visually parse every single line
> (i.e. O(N)).

Ack.

> > +	switch (pid) {
> > +	case -1:
> > +		llen += printf("fork()");
> > +		pad_spc(llen, 64, "[FAIL]\n");
> > +		return 1;
> > +
> > +	case 0:
> > +		close(STDOUT_FILENO);
> > +		close(STDERR_FILENO);
> > +
> > +		char *const argv[] = {
> > +			"/proc/self/exe",
> > +			"_smash_stack",
> > +			NULL,
> > +		};
> 
> Same here.

Ack.

> > +		execve("/proc/self/exe", argv, NULL);
> > +		return 1;
> > +
> > +	default: {
> > +		int status;
> 
> And here by moving "status" upper in the function you can even
> get rid of the braces.

Ack.

> > +		pid = waitpid(pid, &status, 0);
> > +
> > +		if (pid == -1 || !WIFSIGNALED(status) || WTERMSIG(status) != SIGABRT) {
> > +			llen += printf("waitpid()");
> > +			pad_spc(llen, 64, "[FAIL]\n");
> > +			return 1;
> > +		}
> > +		llen += printf("stack smashing detected");
> > +		pad_spc(llen, 64, " [OK]\n");
> > +		return 0;
> > +	}
> > +	}
> > +}
> > +
> >  /* prepare what needs to be prepared for pid 1 (stdio, /dev, /proc, etc) */
> >  int prepare(void)
> >  {
> > @@ -719,8 +784,11 @@ int prepare(void)
> >  /* This is the definition of known test names, with their functions */
> >  static const struct test test_names[] = {
> >  	/* add new tests here */
> > -	{ .name = "syscall",   .func = run_syscall  },
> > -	{ .name = "stdlib",    .func = run_stdlib   },
> > +	{ .name = "syscall",        .func = run_syscall         },
> > +	{ .name = "stdlib",         .func = run_stdlib          },
> > +	{ .name = "stackprotector", .func = run_stackprotector, },
> > +	{ .name = "_smash_stack",   .func = run_smash_stack,
> 
> I think it would be better to keep the number of categories low
> and probably you should add just one called "protection" or so,
> and implement your various tests in it as is done for other
> categories. The goal is to help developers quickly spot and select
> the few activities they're interested in at a given moment. 

I'm not sure how this would be done. The goal here is that
"stackprotector" is the user-visible category. It can be changed to
"protection".
"_smash_stack" however is just an entrypoint that is used by the forked
process to call the crashing code.
We need the fork+exec+special entrypoint to avoid crashing the test
process itself.