Re: strace-4.18 test suite oopses sparc64 4.12 and 4.13-rc kernels

From: Sam Ravnborg
Date: Wed Aug 02 2017 - 17:36:56 EST

On Tue, Aug 01, 2017 at 10:58:29PM +0200, Sam Ravnborg wrote:
> Hi Mikael.
> I think this translates to the following code
> from linux/uaccess.h
> first part is the inlined _copy_from_user()
> >
> > (gdb) x/10i do_sys_poll+0x80-16
> > 0x516ed0 <do_sys_poll+112>: brz %o0, 0x5170fc <do_sys_poll+668>
> if (unlikely(res))
> > 0x516ed4 <do_sys_poll+116>: mov %o0, %o2
> > 0x516ed8 <do_sys_poll+120>: sub %i4, %o0, %i4
> > 0x516edc <do_sys_poll+124>: clr %o1
> > 0x516ee0 <do_sys_poll+128>: call 0x7570b8 <memset>
> > 0x516ee4 <do_sys_poll+132>: add %l3, %i4, %o0
> memset(to + (n - res), 0, res);

And memset calls down to bzero, where %o0=buf, %o1=len

%o0 = 0xc
%o1 = 0xfff000123c897a80
%o2 = 0x0
%o3 = 0xc

So from this we know that:
res = 0xfff000123c897a80
to + (n - 0xfff000123c897a80)) = 0xc

The value "fff000123c897a80" really looks like a constructed address
from somewhere in the strace code, and where this constructed address
is used to provoke some unusual behaviour.
The "fff0" part may be a sparc thing.

So far the analysis seems to match the intial conclusion that
we in this special case try to zero out the remaining memory
based on the return value of raw_copy_from_user.
And therefore we use the return value (res) which triggers the oops.

So rather than manipulating with the assembler code as suggested
in the previous mail this simpler patch could be tested:

diff --git a/include/linux/uaccess.h b/include/linux/uaccess.h
index acdd6f915a8d..13d299ff1f21 100644
--- a/include/linux/uaccess.h
+++ b/include/linux/uaccess.h
@@ -115,7 +115,7 @@ _copy_from_user(void *to, const void __user *from, unsigned long n)
res = raw_copy_from_user(to, from, n);
if (unlikely(res))
- memset(to + (n - res), 0, res);
+ void: /*memset(to + (n - res), 0, res);*/
return res;

It would be good to know if this makes the opps go away.

And maybe you could try to print the parameters
supplied to _copy_from_user in case memset would be called,
so we have an idea what error path is taken.

I have tried to dechiper U3memcpy.S - but that is non-trivial.
So it would be good with a bit more data to verify the theory.