On Wed, Nov 4, 2020 at 10:23 AM Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
>
> On Tue, Nov 03, 2020 at 10:39:52AM -0800, Peter Collingbourne wrote:
> > On Tue, Nov 3, 2020 at 9:54 AM Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> > > On Mon, Nov 02, 2020 at 08:09:43PM -0800, Peter Collingbourne wrote:
> > > > This field will contain flags that may be used by signal handlers to
> > > > determine whether other fields in the _sigfault portion of siginfo are
> > > > valid. An example use case is the following patch, which introduces
> > > > the si_addr_tag_bits{,_mask} fields.
> > > >
> > > > A new sigcontext flag, SA_FAULTFLAGS, is introduced in order to allow
> > > > a signal handler to require the kernel to set the field (but note
> > > > that the field will be set anyway if the kernel supports the flag,
> > > > regardless of its value). In combination with the previous patches,
> > > > this allows a userspace program to determine whether the kernel will
> > > > set the field.
> > >
> > > As per patch 5, a user is supposed to call sigaction() twice to figure
> > > out whether _faultflags is meaningful. That's the part I'm not
> > > particularly fond of. Are the unused parts of siginfo always zeroed when
> > > the kernel delivers a signal? If yes, we could simply check the new
> > > field for non-zero bits.
> >
> > The unused parts of siginfo are zeroed in current kernels, but
> > unfortunately not in older kernels. The zeroing behavior was
> > introduced in commit c999b933faa5e281e3af2e110eccaf91698b0a81 which
> > first appeared in kernel version 4.18, and at least in Android land we
> > do need to support kernel versions older than that.
>
> I see. I was hoping for an easy way out.
>
> Now, with always populating the si_faultflags field, you are going back
> to writing non-zero stuff in siginfo for unaware apps. I don't think
> that's an issue (the alternative is to only write it of SA_FAULTFLAGS
> was set).
>
> Yet another option would be to pass a new AT_ZEROED_SI via AT_FLAGS (we
> don't use them for anything) so that the user can infer whether
> si_faultflags has meaningful information without two sigaction() calls.
That's one option, although one benefit of having this involve
sigaction is that in many cases where sigaction is wrapped or
interposed we end up with correct behavior. Imagine a wrapper that
stashes the provided struct sigaction somewhere and provides its own
struct sigaction with its own handler to the kernel, and that handler
copies siginfo field by field before calling the user's handler. In
this scenario the handler would observe an uninitialized faultflags.
With the SA_UNSUPPORTED/SA_FAULTFLAGS protocol we would detect this
scenario in the same way as an old kernel and avoid reading
faultflags.
Of course this isn't a perfect defense but it's probably the best we can do.
> > > > It is possible for an si_faultflags-unaware program to cause a signal
> > > > handler in an si_faultflags-aware program to be called with a provided
> > > > siginfo data structure by using one of the following syscalls:
> > > >
> > > > - ptrace(PTRACE_SETSIGINFO)
> > > > - pidfd_send_signal
> > > > - rt_sigqueueinfo
> > > > - rt_tgsigqueueinfo
> > > >
> > > > So we need to prevent the si_faultflags-unaware program from causing an
> > > > uninitialized read of si_faultflags in the si_faultflags-aware program when
> > > > it uses one of these syscalls.
> > > >
> > > > The last three cases can be handled by observing that each of these
> > > > syscalls fails if si_code >= 0. We also observe that kill(2) and
> > > > tgkill(2) may be used to send a signal where si_code == 0 (SI_USER),
> > > > so we define si_faultflags to only be valid if si_code > 0.
> > > >
> > > > There is no such check on si_code in ptrace(PTRACE_SETSIGINFO), so
> > > > we make ptrace(PTRACE_SETSIGINFO) clear the si_faultflags field if it
> > > > detects that the signal would use the _sigfault layout, and introduce
> > > > a new ptrace request type, PTRACE_SETSIGINFO2, that a si_faultflags-aware
> > > > program may use to opt out of this behavior.
> > >
> > > I find this pretty fragile but maybe I have to read it a few more times
> > > to fully understand the implications ;).
> > >
> > > Could we instead copy all the fields, potentially uninitialised, and
> > > instead filter them when delivering the signal based on the
> > > SA_FAULTFLAGS? That means that the kernel only writes si_faultflags if
> > > the user requested it.
> >
> > I don't see how that would help. The goal is to protect new signal
> > handlers from old signal "injectors" that will have potentially
> > uninitialized data where the si_faultflags field is. The new signal
> > handler will have SA_FAULTFLAGS set so that wouldn't prevent the
> > signal handler from seeing the uninitialized data.
>
> You are right, it doesn't help if the handler will have set
> SA_FAULTFLAGS.
>
> > > > v12:
> > > > - Change type of si_xflags to u32 to avoid increasing alignment
> > > [...]
> > > > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h
> > > > index 7aacf9389010..f43778355b77 100644
> > > > --- a/include/uapi/asm-generic/siginfo.h
> > > > +++ b/include/uapi/asm-generic/siginfo.h
> > > > @@ -91,7 +91,9 @@ union __sifields {
> > > > char _dummy_pkey[__ADDR_BND_PKEY_PAD];
> > > > __u32 _pkey;
> > > > } _addr_pkey;
> > > > + void *_pad[6];
> > > > };
> > > > + __u32 _faultflags;
> > > > } _sigfault;
> > >
> > > Sorry, I haven't checked the previous discussion on alignment here but
> > > don't we already require 64-bit alignment because of other members in
> > > the _sigfault union? We already have void * throughout this and with the
> > > next patch we just have a gap (unless I miscalculated the offsets).
> >
> > This is about avoiding increasing alignment on 32-bit platforms.
> > Currently the alignment is 4 but a u64 field would bump it to 8.
> >
> > Unfortunately we can't do much about the gap on 64-bit platforms. This
> > was previously a uintptr_t but that would mean that the upper 32 bits
> > cannot be used safely on all platforms so we would effectively end up
> > with a gap anyway.
>
> We could add a dummy pad on 64-bit.
And then later once we add a 32-bit field here we use it like so?
__u32 _faultflags;
#ifdef __LP64__
__u32 _newfield;
#endif
unsigned long _addr_tag_bits, _addr_tag_bits_mask;
#ifndef __LP64__
__u32 _newfield;
#endif
Okay, I'll go ahead with that for now.
> BTW, the tags only make sense on
> 64-bit hardware, 32-bit doesn't have enough room.
>From an architectural perspective it really depends on which kinds of
applications you are targeting. For example if you have something like
a 32-bit microcontroller you might not need all of the address space
for memory so it may be worthwhile to allow some bits to be used for
tags. According to the comments on [1] RISC-V is planning to have
their first implementation only support 64-bit but they haven't ruled
out 32-bit in the future.
[1] https://docs.google.com/document/d/1RZcEgljHY9ACeKKoLebBNPLqjl6nMMyG/edit#heading=h.1fob9te
Peter
