[PATCH 07/11] signal/arm64: Document conflicts with SI_USER and SIGFPE, SIGTRAP, SIGBUS

[Dave Martin]

On Tue, Jan 16, 2018 at 04:28:50PM -0600, Eric W. Biederman wrote:
> Dave Martin <Dave.Martin at arm.com> writes:
> 
> > On Mon, Jan 15, 2018 at 11:23:03AM -0600, Eric W. Biederman wrote:
> >> Dave Martin <Dave.Martin at arm.com> writes:
> >> 
> >> > On Thu, Jan 11, 2018 at 06:59:36PM -0600, Eric W. Biederman wrote:
> >> >> Setting si_code to 0 results in a userspace seeing an si_code of 0.
> >> >> This is the same si_code as SI_USER.  Posix and common sense requires
> >> >> that SI_USER not be a signal specific si_code.  As such this use of 0
> >> >> for the si_code is a pretty horribly broken ABI.
> >> >
> >> > I think this situation may have come about because 0 is used as a
> >> > padding value for "impossible" cases -- i.e., things that can't happen
> >> > unless the kernel is broken, or things that are too unrecoverable for
> >> > clean error reporting to be helpful.
> >> >
> >> > In general, I think these values are not expected to reach userspace in
> >> > practice.
> >> >
> >> > This is not an excuse though -- and not 100% true -- so it's certainly
> >> > worthy of cleanup.
> >> >
> >> >
> >> > It would be good to approach this similarly for arm and arm64, since
> >> > the arm64 fault code is derived from arm.
> >> 
> >> In this case the fault_info is something I have only seen on arm64.
> >> I have been approaching all architectures the same way.
> >
> > Bad guess on my part; this table-driven approach seems to be new for
> > arm64.
> >
> >> If there is insufficient information without architecture expertise
> >> to fix this class of error I have been ading FPE_FIXME to them.
> >
> > Fair enough.
> >
> >> >> Further use of si_code == 0 guaranteed that copy_siginfo_to_user saw a
> >> >> value of __SI_KILL and now sees a value of SIL_KILL with the result
> >> >> that uid and pid fields are copied and which might copying the si_addr
> >> >> field by accident but certainly not by design.  Making this a very
> >> >> flakey implementation.
> >> >> 
> >> >> Utilizing FPE_FIXME, BUS_FIXME, TRAP_FIXME siginfo_layout will now return
> >> >> SIL_FAULT and the appropriate fields will be reliably copied.
> >> >> 
> >> >> But folks this is a new and unique kind of bad.  This is massively
> >> >> untested code bad.  This is inventing new and unique was to get
> >> >> siginfo wrong bad.  This is don't even think about Posix or what
> >> >> siginfo means bad.  This is lots of eyeballs all missing the fact
> >> >> that the code does the wrong thing bad.  This is getting stuck
> >> >> and keep making the same mistake bad.
> >> >> 
> >> >> I really hope we can find a non userspace breaking fix for this on a
> >> >> port as new as arm64.
> >> >
> >> >> Possible ABI fixes include:
> >> >> - Send the signal without siginfo
> >> >> - Don't generate a signal
> >> >
> >> > The above two sould like ABI breaks?
> >> 
> >> They are ways I have seen code on other platforms deal with
> >> not information to generate siginfo.  Sending the signal without siginfo
> >> is roughly equivalent to your send SIGKILL suggestion below.
> >> 
> >> A good example of that is code that calls force_sigsegv.
> >> 
> >> Calling "force_sig(SIGBUS, current);" is perfectly valid.
> >> And then the parent when it reaped the process would have
> >> a little more information to go on when guessing what happened
> >> to the process.
> >> 
> >> >> - Possibly assign and use an appropriate si_code
> >> >> - Don't handle cases which can't happen
> >> >
> >> > I think a mixture of these two is the best approach.
> >> >
> >> > In any case, si_code == 0 here doesn't seem to have any explicit meaning.
> >> > I think we can translate all of the arm64 faults to proper si_codes --
> >> > see my sketch below.  Probably means a bit more thought though.
> >> 
> >> Yes I would be very happy to see that.
> >> 
> >> > The only counterargument would be if there is software relying on
> >> > these bogus signal cases getting si_code == 0 for a useful purpose.
> >> >
> >> > The main reason I see to check for SI_USER is to allow a process to
> >> > filter out spurious signals (say, an asynchronous I/O signal for
> >> > which si_value would be garbage), and to print out diagnostics
> >> > before (in the case of a well-behaved program) resetting the signal
> >> > to SIG_DFL and killing itself to report the signal to the waiter.
> >> >
> >> > Daemons may be more discerning about who is allowed to signal them,
> >> > but overloading SIGBUS (say) as an IPC channel sounds like a very odd
> >> > thing to do.  The same probably applies to any signal that has
> >> > nontrivial metadata.
> >> 
> >> Agreed.  Although I have seen ltp test cases that do crazy things like
> >> that.
> >> 
> >> > Have you found software that is impacted by this in practice?
> >> 
> >> No.
> >
> > Searching for si_code on codesearch.debian.net, I looked at a few
> > random hits.  Although this is far from exhaustive, I saw no instance
> > of code that assumes some arch-specific meaning for SI_USER (or 0).
> >
> > Most code seems to check for signal-specific si_code values before
> > assuming that signal-specific signifo fields are valid; or for
> > SI_USER (or si_code <= 0) before assuming that si_uid and si_pid
> > are valid.
> >
> > Returning proper values for si_code values in place of "0" would fix
> > rather than break such cases.
> >
> >
> >> I don't expect many userspace applications look at siginfo and
> >> everything I have found is some rare hard to trigger non-x86 case which
> >> limits the exposure to userspace applications tremendously.
> >> 
> >> The angle I am coming at all of this from is that the linux kernel code
> >> that filled out out struct siginfo was not comprehensible or correct.
> >
> > I think "not comprehensible or correct" is a pretty good summary of
> > all signal-related code...
> >
> >> Internal to the kernel it was using a magic value (not exportable to
> >> userspace) in the upper bits of si_code.  That was causing problems for
> >> signal injection and converting signals from 32bit to 64bit, and from
> >> 64bit to 32bit.
> >> 
> >> So I wrote kernel/signal.c:siginfo_layout() to figure out which fields
> >> of struct siginfo should be sent to userspace.  In doing so I discovered
> >> that using 0 in si_code (aka SI_USER) is ambiguous, and problematic.
> >> 
> >> Unfortuantely in most of the cases I have spotted using 0 in the si_code
> >> requires architectural knowledge that I don't currently have to sort
> >> out.  So the best I can do is change si_code from 0 to
> >> FPE_FIXME/BUS_FIXME/TRAP_FIXME and bring the architecture maintainers
> >> attention to this area.
> >> 
> >> One of the problems that results from all of this is that we copy
> >> unitialized data to userspace.   I am slowly unifying and cleaning the
> >> code up so that the code is simple enough we can be certain we are
> >> not copying unitialized data to userspace.
> >> 
> >> With si_coes of FPE_FIXME/BUS_FIXME/TRAP_FIXME I can at least attempt to
> >> keep the craziness from happening.
> >> 
> >> My next step is to unify struct siginfo and struct compat_siginfo
> >> and the functions that copy them to userspace because there are very
> >> siginficant problems there.
> >
> > All of which sounds valuable.
> >
> >> All of that said I like the way you are thinking about fixing these
> >> issues.
> >
> > Is it feasible to have a different internal constant for SI_USER?
> > Then the generic could warn if it sees si_code == 0.  If the
> > special nonzero KERNEL_SI_USER is seen instead, it is silently
> > translated to SI_USER (0) for userspace.  This might help us
> > track down cases where 0 is passed by accident.
> >
> > It may not be worth it though: if the affected cases are ones
> > that happen never or almost never, a runtime BUG_ON() may not be
> > helpful for tracking them down.
> >
> > Also, I'm making an assumption that si_code always flows through
> > some generic code before reaching userspace (maybe untrue).
> 
> The code does flow through a generic path, and I am in the middle of
> tightening that up right now.  As filling out siginfo is error prone,
> and I need a guarantee that all of struct siginfo is initialized.
> Adding a warning if a arch fault hander uses si_code == 0 aka SI_USER
> would not be hard.
> 
> Given what you have found.  Given that it seems to match my experience
> we can almost certainly change the code to just warn when the 0 is
> passed in for the si_code for fault handlers.
> 
> I want to ensure that all of the fields are filled in before I do that
> or else I risk passing unitialized values to userspace.  But I like that
> as the long term goal for this code.
> 
> >> >> +++ b/arch/arm64/kernel/fpsimd.c
> >> >> @@ -867,7 +867,7 @@ asmlinkage void do_fpsimd_acc(unsigned int esr, struct pt_regs *regs)
> >> >>  asmlinkage void do_fpsimd_exc(unsigned int esr, struct pt_regs *regs)
> >> >>  {
> >> >>  	siginfo_t info;
> >> >> -	unsigned int si_code = 0;
> >> >> +	unsigned int si_code = FPE_FIXME;
> >> >>  
> >> >>  	if (esr & FPEXC_IOF)
> >> >>  		si_code = FPE_FLTINV;
> >> >
> >> > This 0 can happen for vector operations where the implementation may
> >> > not be able to report exactly what happened, for example where
> >> > the implementer didn't want to pay the cost of tracking exactly
> >> > what went wrong in each lane.
> >> >
> >> > However, the FPEXC_* bits can be garbage in such a case rather
> >> > than being all zero: we should be checking the TFV bit in the ESR here.
> >> > This may be a bug.
> >> >
> >> > Perhaps FPE_FLTINV should be returned in si_code for such cases:  it's
> >> > not otherwise used on arm64 -- invalid instructions would be reported as
> >> > SIGILL/ILL_ILLOPC instead).
> >> >
> >> > Otherwise, we might want to define a new code or arbitrarily pick
> >> > one of the existing FLT_* since this is really a more benign condition
> >> > than executing an illegal instruction.  Alternatively, treat the
> >> > fault as spurious and suppress it, but that doesn't feel right either.
> >> 
> >> I would love to see this sorted out.  There is a very similar pattern
> >> on several different architectures.  I suspect if we have a clean
> >> solution on one architecture the other architectures will be able to use
> >> that solution as well.
> >
> > Since user code that relies on checking si_code for fp exceptions will
> > probably already break in these cases, we can probably fudge things a
> > bit here.
> >
> > I'll have a think.  IEEE may also define some rules that are relevant
> > here...
> >
> > For the proposed conversion of the si_code==0 cases for arm64, I'll
> > draft an RFC for discussion (hopefully sometime this week).
> 
> Sounds good.
> 
> I will keep FPE_FIXME as a place holder until this gets sorted out.
> 
> There is a second issue I am looking at in this location,
> and maybe I don't have to address it now.  But it looks like the code is
> calling send_sig_info instead of force_sig_info for a synchronous
> exception.  Am I reading that correctly?

You mean floating-point exception? (almost anything that is not an
interrupt can be a sunchronous exception on arm64).

Interesting though.  arm and arm64 have send_sig_info() here, but
all other arches seem to have force_sig_info().

sigaction(2) says:
"According to POSIX, the behavior of a process is undefined after it
ignores a SIGFPE, SIGILL, or SIGSEGV signal that was not generated by
kill(2) or raise(3)."

I can't find the POSIX reference for this, but the above suggests we
don't _have_ to kill the process.  arm/arm64 do seem inconsistent on
this though, compared with other architectures.  force_sig_info() might
be needed to avoid a spin on architectures where writeback is
suppressed when a precise fp exception trap is taken and the exception
return address points back to the offending instruction.  I'm
struggling to deduce from the ARM ARM whether this applies here.

Maybe it doesn't matter what we do: the state of the standards seems
to mean that user code that attempts to recover from FP exceptions
will be nonportable and probably broken in any case...


Russell, do you remember the motivation for this?  I'm looking at
da41119af788 ("[PATCH] ARM: Don't force SIGFPE")

Cheers
---Dave

-8<-

From: Russell King <rmk at dyn-67.arm.linux.org.uk>
Date: Wed, 29 Jun 2005 23:02:02 +0100
Subject: [PATCH 1/1] [PATCH] ARM: Don't force SIGFPE

We were forcing SIGFPE on to a user program for no good reason.
Use send_sig_info() instead.

Signed-off-by: Russell King <rmk+kernel at arm.linux.org.uk>
---
 arch/arm/vfp/vfpmodule.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/arm/vfp/vfpmodule.c b/arch/arm/vfp/vfpmodule.c
index 3aeedd2..22f3da4 100644
--- a/arch/arm/vfp/vfpmodule.c
+++ b/arch/arm/vfp/vfpmodule.c
@@ -89,7 +89,7 @@ void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
 	current->thread.error_code = 0;
 	current->thread.trap_no = 6;
 
-	force_sig_info(SIGFPE, &info, current);
+	send_sig_info(SIGFPE, &info, current);
 }
 
 static void vfp_panic(char *reason)
-- 
2.1.4