Re: [PATCH] x86/crypto: Add missing RETs

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Tue, Jun 26, 2018 at 08:49:30AM +0200, Ingo Molnar wrote:
> 
> * Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> 
> > On Mon, Jun 25, 2018 at 09:24:38AM +0200, Ingo Molnar wrote:
> > > 
> > > * Alexey Dobriyan <adobriyan@gmail.com> wrote:
> > > 
> > > > On Sun, Jun 24, 2018 at 09:11:05AM +0200, Ingo Molnar wrote:
> > > > > > Add explicit RETs to the tail calls of AEGIS and MORUS crypto algorithms
> > > > > > otherwise they run into INT3 padding due to
> > > > > > 
> > > > > >   51bad67ffbce ("x86/asm: Pad assembly functions with INT3 instructions")
> > > > > > 
> > > > > > leading to spurious debug exceptions.
> > > > > > 
> > > > > > Mike Galbraith <efault@gmx.de> took care of all the remaining callsites.
> > > > > 
> > > > > Note that 51bad67ffbce has been zapped because it caused too many problems like 
> > > > > this, but the explicit RETs make sense nevertheless.
> > > > 
> > > > So commit which found real bug(s) was zapped.
> > > > 
> > > > OK
> > > 
> > > No, what happened is that the commit was first moved into WIP.x86/debug showing 
> > > its work-in-progress status, because it was incomplete and caused bugs:
> > > 
> > >    https://lore.kernel.org/lkml/20180518073644.GA8593@gmail.com/T/#u
> > > 
> > > ... and finally, after weeks of inaction I zapped it because I didn't see progress 
> > > and you didn't answer my question.
> > > 
> > > If a fixed patch with updated tooling to detect these crashes before they occur on 
> > > live systems is submitted we'll reconsider - it didn't get NAK-ed, it's just 
> > > incomplete in the current form.
> > 
> > Hm, what happened to the objtool patch to detect these at build time?
> > Did it not work?
> > 
> >   https://lore.kernel.org/lkml/20180517134934.eog2fgoby5azq5a7@treble
> 
> So that's still incomplete in that doesn't analyze the 32-bit build yet, right?

We could do INT3s on 64-bit and NOPs on 32-bit.

Or, possibly even better, we could just keep NOPs everywhere and instead
make objtool smart enough to detect function fallthroughs.  That should
be pretty easy, actually.  It already does it for C files.

Something like the below should work, though it's still got a few
issues:

  a) objtool is currently disabled for crypto code because it doesn't
     yet understand crypto stack re-alignments (which really needs
     fixing anyway); and

  b) it complains about the blank xen hypercalls falling through.  Those
     aren't actual functions anyway, so we should probably annotate
     those somehow so that objtool ignores them anyway.

I'm a bit swamped at the moment but I can fix those once I get a little
more bandwidth.  I at least verified that this patch caught the crypto
missing RETs.


diff --git a/arch/x86/crypto/Makefile b/arch/x86/crypto/Makefile
index a450ad573dcb..a2c52eec2863 100644
--- a/arch/x86/crypto/Makefile
+++ b/arch/x86/crypto/Makefile
@@ -3,8 +3,6 @@
 # Arch-specific CryptoAPI modules.
 #
 
-OBJECT_FILES_NON_STANDARD := y
-
 avx_supported := $(call as-instr,vpxor %xmm0$(comma)%xmm0$(comma)%xmm0,yes,no)
 avx2_supported := $(call as-instr,vpgatherdd %ymm0$(comma)(%eax$(comma)%ymm1\
 				$(comma)4)$(comma)%ymm2,yes,no)
diff --git a/tools/objtool/check.c b/tools/objtool/check.c
index 2928939b98ec..f740fd828cba 100644
--- a/tools/objtool/check.c
+++ b/tools/objtool/check.c
@@ -1798,13 +1798,14 @@ static int validate_branch(struct objtool_file *file, struct instruction *first,
 	while (1) {
 		next_insn = next_insn_same_sec(file, insn);
 
-		if (file->c_file && func && insn->func && func != insn->func->pfunc) {
+		if (func && insn->func && func != insn->func->pfunc) {
 			WARN("%s() falls through to next function %s()",
 			     func->name, insn->func->name);
 			return 1;
 		}
 
-		func = insn->func ? insn->func->pfunc : NULL;
+		if (insn->type != INSN_NOP)
+			func = insn->func ? insn->func->pfunc : NULL;
 
 		if (func && insn->ignore) {
 			WARN_FUNC("BUG: why am I validating an ignored function?",