[PATCH v2] All arch: remove system call sys_sysctl

[Eric W. Biederman]

Rich Felker <dalias at libc.org> writes:

> On Thu, Jun 11, 2020 at 12:01:11PM -0500, Eric W. Biederman wrote:
>> Rich Felker <dalias at libc.org> writes:
>> 
>> > On Thu, Jun 11, 2020 at 06:43:00AM -0500, Eric W. Biederman wrote:
>> >> Xiaoming Ni <nixiaoming at huawei.com> writes:
>> >> 
>> >> > Since the commit 61a47c1ad3a4dc ("sysctl: Remove the sysctl system call"),
>> >> > sys_sysctl is actually unavailable: any input can only return an error.
>> >> >
>> >> > We have been warning about people using the sysctl system call for years
>> >> > and believe there are no more users.  Even if there are users of this
>> >> > interface if they have not complained or fixed their code by now they
>> >> > probably are not going to, so there is no point in warning them any
>> >> > longer.
>> >> >
>> >> > So completely remove sys_sysctl on all architectures.
>> >> 
>> >> 
>> >> 
>> >> >
>> >> > Signed-off-by: Xiaoming Ni <nixiaoming at huawei.com>
>> >> >
>> >> > changes in v2:
>> >> >   According to Kees Cook's suggestion, completely remove sys_sysctl on all arch
>> >> >   According to Eric W. Biederman's suggestion, update the commit log
>> >> >
>> >> > V1: https://lore.kernel.org/lkml/1591683605-8585-1-git-send-email-nixiaoming@huawei.com/
>> >> >   Delete the code of sys_sysctl and return -ENOSYS directly at the function entry
>> >> > ---
>> >> >  include/uapi/linux/sysctl.h                        |  15 --
>> >> [snip]
>> >> 
>> >> > diff --git a/include/uapi/linux/sysctl.h b/include/uapi/linux/sysctl.h
>> >> > index 27c1ed2..84b44c3 100644
>> >> > --- a/include/uapi/linux/sysctl.h
>> >> > +++ b/include/uapi/linux/sysctl.h
>> >> > @@ -27,21 +27,6 @@
>> >> >  #include <linux/types.h>
>> >> >  #include <linux/compiler.h>
>> >> >  
>> >> > -#define CTL_MAXNAME 10		/* how many path components do we allow in a
>> >> > -				   call to sysctl?   In other words, what is
>> >> > -				   the largest acceptable value for the nlen
>> >> > -				   member of a struct __sysctl_args to have? */
>> >> > -
>> >> > -struct __sysctl_args {
>> >> > -	int __user *name;
>> >> > -	int nlen;
>> >> > -	void __user *oldval;
>> >> > -	size_t __user *oldlenp;
>> >> > -	void __user *newval;
>> >> > -	size_t newlen;
>> >> > -	unsigned long __unused[4];
>> >> > -};
>> >> > -
>> >> >  /* Define sysctl names first */
>> >> >  
>> >> >  /* Top-level names: */
>> >> [snip]
>> >> 
>> >> The uapi header change does not make sense.  The entire point of the
>> >> header is to allow userspace programs to be able to call sys_sysctl.
>> >> It either needs to all stay or all go.
>> >> 
>> >> As the concern with the uapi header is about userspace programs being
>> >> able to compile please leave the header for now.
>> >> 
>> >> We should leave auditing userspace and seeing if userspace code will
>> >> still compile if we remove this header for a separate patch.  The
>> >> concerns and justifications for the uapi header are completely different
>> >> then for the removing the sys_sysctl implementation.
>> >> 
>> >> Otherwise
>> >> Acked-by: "Eric W. Biederman" <ebiederm at xmission.com>
>> >
>> > The UAPI header should be kept because it's defining an API not just
>> > for the kernel the headers are supplied with, but for all past
>> > kernels. In particular programs needing a failsafe CSPRNG source that
>> > works on old kernels may (do) use this as a fallback only if modern
>> > syscalls are missing. Removing the syscall is no problem since it
>> > won't be used, but if you remove the types/macros from the UAPI
>> > headers, they'll have to copy that into their own sources.
>> 
>> May we assume you know of a least one piece of userspace that will fail
>> to compile if this header file is removed?
>
> I know at least one piece of software is using SYS_sysctl for a
> fallback CSPRNG source. I'm not 100% sure that they're using the
> kernel headers; they might have copied it already. I'm also not sure
> how many there are.
>
> Regardless, I think the principle stands. There's no need to remove
> definitions that are essentially maintenance-free now that the
> interface is no longer available in new kernels, and doing so
> contributes to the myth that you're supposed to use kernel headers
> matching runtime kernel rather than it always being safe to use latest
> headers.

If there is no one using the definitions removing them saves people
having to remember what they are there for.

The big rule is don't break userspace.  The goal is to allow people to
upgrade their kernel without needing to worry about userspace breaking,
and to be able to downgrade to the extent possible to help in tracking
bugs.

Not being able to compile userspace seems like a pretty clear cut case.
Although there are some fuzzy edges given the history of the kernel
headers.  Things like your libc requiring kernel headers to be processed
before they can be used.  I think there are still some kernel headers
that have that restriction when used with glibc as glibc uses different
sizes for types like dev_t.

The bottom line is we can't do it casually so that any work in the
direction of removing from or deleting uapi headers needs to be it's own
separate patch.

Given how much effort it can be to show that userspace is not using
something I don't expect us to be mucking with the uapi headers any time
soon.

Eric