kernel virtual memory access (from app) does not generate segfault
Ben Dooks
ben-linux at fluff.org
Tue Apr 20 13:09:44 EDT 2010
On Tue, Apr 20, 2010 at 10:20:47PM +0800, anfei wrote:
> On Tue, Apr 20, 2010 at 11:27:40AM +0100, Dave P. Martin wrote:
> >
> >
> > > -----Original Message-----
> > > From: linux-arm-kernel-bounces at lists.infradead.org
> > > [mailto:linux-arm-kernel-bounces at lists.infradead.org] On
> > > Behalf Of Ben Dooks
> > > Sent: 20 April 2010 10:35
> > > To: Sasha Sirotkin
> > > Cc: linux-arm-kernel at lists.infradead.org
> > > Subject: Re: kernel virtual memory access (from app) does not
> >
> > [..]
> >
> > > > For instance, this code generates a segfault allright
> > > >
> > > > int * aa;
> > > > aa = 0xc0000000;
> > > > *aa=42;
> > > >
> > > > However this code does not, instead the process simply
> > > hangs (and can
> > > > be
> > > > killed)
> > > >
> > > > void (*func)(void);
> > > > func = 0xc0000000;
> > > > func();
> > >
> > > Your first example writes to an area, your second is
> > > execution. IIRC, this version of the ARM architecture equates
> > > read and execute permission and so you may actually have
> > > permission to read this area and thus execute code in it.
> > >
> > > > I stumbled across this by accident. Just curious to
> > > understand why it
> > > > happens. Isn't it a bug ?
> > >
> > > Don't think so, other than you might not want that area to be
> > > readable by user space?
> >
> > I tried reading that address (albeit on an old 2.6.28 kernel), and I get a
> > segfault.
> >
> > Trying to execute in kernel space is the only thing that appears to hang.
> > Attaching to the process in gdb, I observed that pc is always 0xc0000000
> > when the process is stopped.
> >
> > top accounts most of the CPU time as being consumed in the kernel.
> >
> > I think what is going on here is that the kernel is catching the expected
> > prefetch abort, but the handler fails to send SIGSEGV to the user process
> > --- the process is resumed with the same pc and we end up in an endless
> > spin.
> >
> > This only appears to apply to certain address ranges: substituting some
> > other random unmapped address for 0xc0000000 (0x48000000 worked for me), we
> > get the expected segfault.
> >
> > Does the prefetch abort handler assume that lr >= 0xc0000000 implies the
> > fault came from inside the kernel? Should it?
> >
> > arch/arm/mm/fault.c has:
> >
> > /*
> > ...
> > * If the address is in kernel space (>= TASK_SIZE), then we are
> > * probably faulting in the vmalloc() area.
> > ...
> > */
> > static int __kprobes
> > do_translation_fault(unsigned long addr, unsigned int fsr,
> > struct pt_regs *regs)
> > {
> > ...
> > if (addr < TASK_SIZE)
> > return do_page_fault(addr, fsr, regs);
> >
> > So the common case for userspace prefetch aborts is do_page_fault()
> >
> > This suggests that the weirdness is caused by something in the remainder of
> > do_translation_fault(), or something it calls.
> >
> >
> > The comment preceding do_translation_fault() suggests a possible unsafe
> > assumption which could lead to a security hole... but it really depends on
> > what the handler code is trying to do. Unfortunately, my understanding has
> > broken down by this point.
> >
> > Is someone else able to comment on how this code responds to a user fault >=
> > TASK_SIZE?
> >
> I think something like this is needed:
>
> diff --git a/arch/arm/mm/fault.c b/arch/arm/mm/fault.c
> index 9d40c34..cd4d15c 100644
> --- a/arch/arm/mm/fault.c
> +++ b/arch/arm/mm/fault.c
> @@ -393,6 +393,9 @@ do_translation_fault(unsigned long addr, unsigned int fsr,
> if (addr < TASK_SIZE)
> return do_page_fault(addr, fsr, regs);
>
> + if (user_mode(regs) && addr >= TASK_SIZE)
> + goto bad_area;
> +
technically, addr >= TASK_SIZE was guaranteed by the previous test
on addr. The user_mode(regs) may well be a good idea, although I'm
not sure if we get entered here if the kernel is attempting to access
user-mode memory by forcing unpriveldged accesses.
probably best to get Russell's opinion.
--
Ben
Q: What's a light-year?
A: One-third less calories than a regular year.
More information about the linux-arm-kernel
mailing list