Re: [PATCH] selftests: proc: Make sure wchan works when it exists

[Kees Cook <keescook@chromium.org>]

On Sat, Oct 09, 2021 at 03:52:02PM +0300, Alexey Dobriyan wrote:
> On Fri, Oct 08, 2021 at 04:55:04PM -0700, Kees Cook wrote:
> > This makes sure that wchan contains a sensible symbol when a process is
> > blocked.
> 
> > Specifically this calls the sleep() syscall, and expects the
> > architecture to have called schedule() from a function that has "sleep"
> > somewhere in its name.
> 
> This exposes internal kernel symbol to userspace.

Correct; we're verifying the results of the wchan output, which produces
a kernel symbol for blocked processes.

> Why would want to test that?

This is part of a larger series refactoring/fixing wchan[1], and we've
now tripped over several different failure conditions, so I want to make
sure this doesn't regress in the future.

> Doing s/sleep/SLEEP/g doesn't change kernel but now the test is broken.

Yes; the test would be doing it's job, as that would mean there was a
userspace visible change to wchan, so we'd want to catch it and either
fix the kernel or update the test to reflect the new reality.

> 
> > For example, on the architectures I tested
> > (x86_64, arm64, arm, mips, and powerpc) this is "hrtimer_nanosleep":
> 
> > +/*
> > + * Make sure that wchan returns a reasonable symbol when blocked.
> > + */
> 
> Test should be "contains C identifier" then?

Nope, this was intentional. Expanding to a C identifier won't catch the
"we unwound the stack to the wrong depth and now all wchan shows is
'__switch_to'" bug[2]. We're specifically checking that wchan is doing
at least the right thing for the most common blocking state.

> 
> > +int main(void)
> > +{
> > +	char buf[64];
> > +	pid_t child;
> > +	int sync[2], fd;
> > +
> > +	if (pipe(sync) < 0)
> > +		perror_exit("pipe");
> > +
> > +	child = fork();
> > +	if (child < 0)
> > +		perror_exit("fork");
> > +	if (child == 0) {
> > +		/* Child */
> > +		if (close(sync[0]) < 0)
> > +			perror_exit("child close sync[0]");
> > +		if (close(sync[1]) < 0)
> > +			perror_exit("child close sync[1]");
> 
> Redundant close().

Hmm, did you maybe miss the differing array indexes? This closes the
reading end followed by the writing end of the child's pipe.

> 
> > +		sleep(10);
> > +		_exit(0);
> > +	}
> > +	/* Parent */
> > +	if (close(sync[1]) < 0)
> > +		perror_exit("parent close sync[1]");
> 
> Redundant close().

It's not, though. This closes the write side of the parent's pipe.

> 
> > +	if (read(sync[0], buf, 1) != 0)
> > +		perror_exit("parent read sync[0]");
> 
> Racy if child is scheduled out after first close in the child.

No, the first close will close the child's read-side of the pipe, which
isn't being used. For example, see[3].

The parent's read of /proc/$child/wchan could technically race if the
child is scheduled out after the second close() and before the sleep(),
but the parent is doing at least 2 syscalls before then. I'm open to
a more exact synchronization method, but this should be sufficient.
(e.g. Using ptrace to catch sleep syscall entry seemed like overkill.)

-Kees

[1] https://lore.kernel.org/lkml/20211008111527.438276127@infradead.org/
[2] https://lore.kernel.org/lkml/20211008124052.GA976@C02TD0UTHF1T.local/
[3] https://man7.org/tlpi/code/online/diff/pipes/pipe_sync.c.html

> 
> > +	snprintf(buf, sizeof(buf), "/proc/%d/wchan", child);
> > +	fd = open(buf, O_RDONLY);
> > +	if (fd < 0) {
> > +		if (errno == ENOENT)
> > +			return 4;
> > +		perror_exit(buf);
> > +	}
> > +
> > +	memset(buf, 0, sizeof(buf));
> > +	if (read(fd, buf, sizeof(buf) - 1) < 1)
> > +		perror_exit(buf);
> > +	if (strstr(buf, "sleep") == NULL) {
> > +		fprintf(stderr, "FAIL: did not find 'sleep' in wchan '%s'\n", buf);
> > +		return 1;
> > +	}
> > +	printf("ok: found 'sleep' in wchan '%s'\n", buf);
> > +
> > +	if (kill(child, SIGKILL) < 0)
> > +		perror_exit("kill");
> > +	if (waitpid(child, NULL, 0) != child) {
> > +		fprintf(stderr, "waitpid: got the wrong child!?\n");
> > +		return 1;
> > +	}
> > +
> > +	return 0;
> > +}

-- 
Kees Cook