Re: [PATCH] selftests: mm: Fix build errors on armhf

[Liam R. Howlett]

* Jeff Xu <jeffxu@xxxxxxxxxxxx> [240910 10:15]:
> Hi Muhammad
> 
> On Mon, Aug 19, 2024 at 3:05 AM Muhammad Usama Anjum
> <Usama.Anjum@xxxxxxxxxxxxx> wrote:
> >
> > On 8/14/24 3:29 AM, Jeff Xu wrote:
> > > Hi Muhammad
> > >
> > > On Fri, Aug 9, 2024 at 1:25 AM Muhammad Usama Anjum
> > > <usama.anjum@xxxxxxxxxxxxx> wrote:
> > >>
> > >> The __NR_mmap isn't found on armhf. The mmap() is commonly available
> > >> system call and its wrapper is presnet on all architectures. So it
> > >> should be used directly. It solves problem for armhf and doesn't create
> > >> problem for architectures as well. Remove sys_mmap() functions as they
> > >> aren't doing anything else other than calling mmap(). There is no need
> > >> to set errno = 0 manually as glibc always resets it.
> > >>
> > > The mseal_test should't have dependency on libc, and mmap() is
> > > implemented by glibc, right ?
> > >
> > > I just fixed a bug to switch mremap() to sys_mremap to address an
> > > issue that different glibc version's behavior is slightly different
> > > for mremap().
> > >
> > > What is the reason that __NR_mmap not available in armhf ? (maybe it
> > > is another name ?)  there must be a way to call syscall directly on
> > > armhf, can we use that instead ?
> >
> > It seems __NR_mmap syscall is deprecated for arm. Found this comment in
> > arch/arm/include/asm/unistd.h:
> > /*
> >  * The following syscalls are obsolete and no longer available for EABI:
> >  *  __NR_time
> >  *  __NR_umount
> >  *  __NR_stime
> >  *  __NR_alarm
> >  *  __NR_utime
> >  *  __NR_getrlimit
> >  *  __NR_select
> >  *  __NR_readdir
> >  *  __NR_mmap
> >  *  __NR_socketcall
> >  *  __NR_syscall
> >  *  __NR_ipc
> >  */
> >
> > The glibc mmap() calls mmap2() these days by adjusting the parameters
> > internally. From man mmap:
> > C library/kernel differences:
> > This  page  describes the interface provided by the glibc mmap() wrapper
> > function.  Originally, this function invoked a system call of the same
> > name.  Since Linux 2.4, that system call has been superseded  by
> > mmap2(2), and nowadays the glibc mmap() wrapper function invokes
> > mmap2(2) with a suitably adjusted value for offset.
> >
> > I'm not sure if behaviour of glibc mmap() and syscall mmap2() would be
> > same, but we should use glibc at most places which accounts for
> > different architectures correctly. Maybe the differences were only
> > present in case of mremap().
> >
> We shouldn't use glibc to test mseal, mseal is a security feature, and
> an attacker can access syscall directly, so the test needs to test
> with as little layer as possible.

This sounds like you are concerned about the use of a library
artificially reducing the attack surface that exists by bypassing the
library to access the syscall directly.

If you have an example of something that is restricted by the library
that can be used by the attacker, then we will need to roll our own
caller with wrappers so that this works on random archs.

It appears that the existing tests can use the library without reducing
the test coverage, so why would we maintain our own abstraction?

This patch is also upstream, so it is obviously not going to change at
this point.  I'm sure you are aware of that as you raised this concern
on the failed backport email [1].  Neither of these locations are the
right one for a discussion that you are trying to start.  It would be
better to produce an RFC patch and send it to the mm mailing list and cc
the people on this patch.

Thanks,
Liam

[1]. https://lore.kernel.org/all/CABi2SkV-FdDQy2bjDkpgpqz7hX7ybeTjCrUgUf6WcYqGkuxWMQ@xxxxxxxxxxxxxx/