François Legal <[email protected]> writes:

> Le Vendredi, Octobre 16, 2020 18:24 CEST, Philippe Gerum <[email protected]> a 
> écrit: 
>  
>> 
>> François Legal <[email protected]> writes:
>> 
>> > Le Vendredi, Octobre 16, 2020 14:03 CEST, Philippe Gerum 
>> > <[email protected]> a écrit: 
>> >  
>> >> 
>> >> François Legal <[email protected]> writes:
>> >> 
>> >> > Le Vendredi, Octobre 16, 2020 10:59 CEST, Philippe Gerum 
>> >> > <[email protected]> a écrit: 
>> >> >  
>> >> >> 
>> >> >> François Legal <[email protected]> writes:
>> >> >> 
>> >> >> > Le Mercredi, Octobre 14, 2020 16:16 CEST, Greg Gallagher 
>> >> >> > <[email protected]> a écrit: 
>> >> >> >  
>> >> >> >> On Wed, Oct 14, 2020 at 5:37 AM Jan Kiszka <[email protected]> 
>> >> >> >> wrote:
>> >> >> >> >
>> >> >> >> > On 14.10.20 10:43, François Legal via Xenomai wrote:
>> >> >> >> > > Anybody can help on this ?
>> >> >> >> > >
>> > >> >> >
>> >> >> >> > I'm unfortunately not familiar with the armv7 details of 
>> >> >> >> > copy-from-user,
>> >> >> >> > not too speak of how spectre contributed to it. Greg, Philippe, 
>> >> >> >> > did you
>> >> >> >> > come across this already?
>> >> >> >> >
>> >> >> >> > Jan
>> >> >> >> >
>> >> >> >> I'll take a look tonight but I haven't hit this in my testing.  This
>> >> >> >> was found on 4.4? Have you tried the 4.19 kernels?
>> >> >> >> 
>> >> >> >> -Greg
>> >> >> >  
>> >> >> > So I tried the same case on 4.19.121, with the same result, and I 
>> >> >> > guess for the same reason.
>> >> >> > Could anybody explain why, on ARMv7, we need to copy the message 
>> >> >> > header at the syscall entry, and not let the xxxmsg routine do it on 
>> >> >> > its own ?
>> >> >> > Also, I could not find how those COBALT_SYSCALL32emu logic work.
>> >> >> 
>> >> >> There is no way an armv7 system would run the sys32emu code in
>> >> >> Cobalt. This code path is specific to architectures which support mixed
>> >> >> ABI models. Only Cobalt/x86 supports this so far, issuing x86_32
>> >> >> syscalls to an x86_64 kernel. You may want to check
>> >> >> CONFIG_XENO_ARCH_SYS3264, it is set to "def_bool n" in the Kconfig
>> >> >> stuff.
>> >> >> 
>> >> >
>> >> > Maybe I don't use the right terms here, but what I can see from the 
>> >> > code is (in linux tree kernel/xenomai/posix/syscall32.c)
>> >> > COBALT_SYSCALL32emu(sendmsg, handover,
>> >> >                     (int fd, struct compat_msghdr __user *umsg, int 
>> >> > flags))
>> >> > {
>> >> >         struct user_msghdr m;
>> >> >         int ret;
>> >> >
>> >> >         ret = sys32_get_msghdr(&m, umsg);
>> >> >
>> >> >         return ret ?: rtdm_fd_sendmsg(fd, &m, flags);
>> >> > }
>> >> >
>> >> > And the problem regarding SPECTRE mitigation is with the "ret =
>> >> > sys32_get_msghdr(&m, umsg);" line, as af_packet (in my case, but I
>> >> > believe the other handlers should do the same) will also call
>> >> > copy_from_user on the "msghdr" argument, and the SPECTRE mitigation
>> >> > will check that this pointer is in the userland MM area.
>> >> 
>> >> There is indeed a problem with this code passing the kernel memory
>> >> address of a bounce buffer to RTDM handlers which would expect __user
>> >> tagged memory instead. This ends up confusing any low-level
>> >> copy_to/from_user routine which includes attack
>> >> mitigation. rtnet_get_arg() does call such routine under the hood. This
>> >> is something some Xenomai contributor may want to address.
>> >> 
>> >> But, again, this sys32emu code cannot run for armv7 under the current
>> >> stock implementation. So what we are discussing is purely hypothetical
>> >> at this stage for this architecture, and should definitely never happen
>> >> by construction if you are running armv7 (which does not make the
>> >> original issue go away, that is granted).
>> >> 
>> >
>> > I'm not sure I quite understand that point. The code reproduced above is 
>> > well built in the kernel. Are you saying this code is not called whenever 
>> > userland calls sendmsg on an rt socket ? Am I looking in the wrong 
>> > direction ? In that case, where should I be looking ? I mean, I tracked 
>> > the whole thing with a JTAG debugger, and it seemed to me that this was 
>> > what was really happening, with the SPECTRE code setting the pointer to 0 
>> > which was later being caught by arm_copy_from_user.
>> >
>> 
>> How could syscall32.c and compat.c be built into the kernel with
>> CONFIG_XENO_ARCH_SYS3264 forcibly unset in the Kconfig, which is always
>> the case when building for anything else than x86?
>> 
>> Checking kernel/cobalt/posix/Makefile may help in understanding why it
>> is simply not possible. arm_copy_from_user is built in, no question,
>> and your analysis regarding SVC context memory being spuriously fed into
>> arm_copy_from_user is likely right.
>> 
>> But the sys32 wrappers are neither for armv7, armv8 nor ppc32. So yes,
>> you are certainly following the wrong path when looking at
>> kernel/cobalt/posix/syscall32.c. This 32-to-64bit syscall support is NOT
>> built into a kernel targeting armv7, at least when it comes to the
>> vanilla Xenomai code.
>> 
>> You may want to double-check which call site actually invokes
>> arm_copy_from_user.
>> 
>> -- 
>> Philippe.
>  
> Thanks Philippe for pointing me in the right direction.
> So, if I'm correct this time, the problem is about the same, but in 
> posix/io.c.
> In 
>  
> COBALT_SYSCALL(sendmsg, handover,
>              (int fd, struct user_msghdr __user *umsg, int flags))
> {
>       struct user_msghdr m;
>       int ret;
>
>       ret = cobalt_copy_from_user(&m, umsg, sizeof(m));
>
>       return ret ?: rtdm_fd_sendmsg(fd, &m, flags);
> }
>
> Same thing, the user_msghdr is allocated on the SVC stack, then copied from 
> user, then handed over to the sendmsg handling function pertaining to that 
> fd, and whenever that handling function call copy_from_user for the same 
> user_msghdr pointer, it triggers the SPECTRE mitigation protection.
>
> What is still unclear to me is why this user_msghdr struct is copied here, 
> and not left to the handling function, as in the sendmmsg syscalls, the 
> struct mmsghdr is not copied from user in syscall entry.
>

Unlike sendmsg(), sendmmsg() has to deal with 32bit compat mode, so
__rtdm_fd_sendmmsg() receives a couple of helpers along with the user
pointer in order to do the right thing, depending on the 32/64 syscall
entry. sendmsg() was implemented much earlier too, when SPECTRE
mitigation was not there: passing SVC memory to copy_from/to_user
services was still wrong, but went unnoticed.

-- 
Philippe.

Reply via email to