[PATCH] crypto/sha1-mb: make sha1_x8_avx2() conform to C function ABI

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Mon, May 16, 2016 at 11:31:12AM -0700, Megha Dey wrote:
> On Mon, 2016-05-16 at 09:44 -0500, Josh Poimboeuf wrote:
> > On Fri, May 13, 2016 at 10:32:26AM -0700, Megha Dey wrote:
> > > On Fri, 2016-05-13 at 07:51 +0200, Ingo Molnar wrote:
> > > > * Herbert Xu <herbert@gondor.apana.org.au> wrote:
> > > > 
> > > > > On Thu, May 12, 2016 at 04:31:06PM -0700, Megha Dey wrote:
> > > > > > Hi,
> > > > > >  
> > > > > > When booting latest kernel with the CONFIG_CRYPTO_SHA1_MB enabled, I
> > > > > > observe a panic.
> > > > > >  
> > > > > > After having a quick look, on reverting the following patches, I am able
> > > > > > to complete the booting process.
> > > > > > aec4d0e301f17bb143341c82cc44685b8af0b945
> > > > > > 8691ccd764f9ecc69a6812dfe76214c86ac9ba06
> > > > > > 68874ac3304ade7ed5ebb12af00d6b9bbbca0a16
> > > > > >  
> > > > > > Of the 3 patches, aec4d0e301f17bb143341c82cc44685b8af0b945 seems wrong.
> > > > > > The r10 to r15 registers are used in sha1_x8_avx2.S, which is called
> > > > > > from sha1_mb_mgr_flush_avx2.S.
> > > > > >
> > > > > > I do not think the functionality of the SHA1-MB crypto algorithm has
> > > > > > been tested after applying these changes. (I am not sure if any of the
> > > > > > other crypto algorithms have been affected by these changes).
> > > > > 
> > > > > Josh, Ingo:
> > > > > 
> > > > > Any ideas on this? Should we revert?
> > > > 
> > > > Yeah, I think so - although another option would be to standardize sha1_x8_avx2() 
> > > > - the problem is that it is a function that clobbers registers without 
> > > > saving/restoring them, breaking the C function ABI. I realize it's written in 
> > > > assembly, but unless there are strong performance reasons to deviate from the 
> > > > regular calling convention it might make sense to fix that.
> > > > 
> > > > Do any warnings get generated after the revert, if you enable 
> > > > CONFIG_STACK_VALIDATION=y?
> > > 
> > > After the revert and enabling CONFIG_STACK_VALIDATION:
> > > arch/x86/crypto/sha1-mb/sha1_mb_mgr_flush_avx2.o: warning: objtool:
> > > sha1_mb_mgr_flush_avx2()+0x20d: call without frame pointer save/setup
> > > 
> > > arch/x86/crypto/sha1-mb/sha1_mb_mgr_submit_avx2.o: warning: objtool:
> > > sha1_mb_mgr_submit_avx2()+0x115: call without frame pointer save/setup
> > 
> > Megha,
> > 
> > Sorry for breaking it.  I completely missed the fact that the function
> > calls sha1_x8_avx2() which clobbers registers.
> > 
> > If the performance penalty isn't too bad, I'll submit a patch to
> > standardize sha1_x8_avx2() to follow the C ABI.
> > 
> > Do you have any tips for testing this code?  I've tried using the tcrypt
> > module, but no luck.
> > 
> Josh,
> Build the kernel with the following configs:
> CONFIG_CRYPTO_SHA1_MB=y
> CONFIG_CRYPTO_TEST=m
> CONFIG_CRYPTO_MANAGER_DISABLE_TESTS=n
> There was a kernel panic while booting. 
> So if after applying your new patch, we are able to get complete the
> boot, then we are good.
> 
> Could you please send a copy of the patch, I could test it on my end
> too. 

Thanks.  I was able to run the tests, though I didn't see a panic.  Can
you test with this patch?

----

From: Josh Poimboeuf <jpoimboe@redhat.com>
Subject: [PATCH] crypto/sha1-mb: make sha1_x8_avx2() conform to C function ABI

Megha Day reported a kernel panic in crypto code.  The problem is that
sha1_x8_avx2() clobbers registers r12-r15 without saving and restoring
them.

Before commit aec4d0e301f1 ("x86/asm/crypto: Simplify stack usage in
sha-mb functions"), those registers were saved and restored by the
callers of the function.  I removed them with that commit because I
didn't realize sha1_x8_avx2() clobbered them.

Fix the potential undefined behavior associated with clobbering the
registers and make the behavior less surprising by changing the
registers to be callee saved/restored to conform with the C function
call ABI.

Also, rdx (aka RSP_SAVE) doesn't need to be saved: I verified that none
of the callers rely on it being saved, and it's not a callee-saved
register in the C ABI.

Fixes: aec4d0e301f1 ("x86/asm/crypto: Simplify stack usage in sha-mb functions")
Cc: stable@vger.kernel.org # 4.6
Reported-by: Megha Dey <megha.dey@linux.intel.com>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
---
 arch/x86/crypto/sha-mb/sha1_x8_avx2.S | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/arch/x86/crypto/sha-mb/sha1_x8_avx2.S b/arch/x86/crypto/sha-mb/sha1_x8_avx2.S
index 8e1b477..c9dae1c 100644
--- a/arch/x86/crypto/sha-mb/sha1_x8_avx2.S
+++ b/arch/x86/crypto/sha-mb/sha1_x8_avx2.S
@@ -296,7 +296,11 @@ W14  = TMP_
 #
 ENTRY(sha1_x8_avx2)
 
-	push	RSP_SAVE
+	# save callee-saved clobbered registers to comply with C function ABI
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
 
 	#save rsp
 	mov	%rsp, RSP_SAVE
@@ -446,7 +450,12 @@ lloop:
 	## Postamble
 
 	mov     RSP_SAVE, %rsp
-	pop	RSP_SAVE
+
+	# restore callee-saved clobbered registers
+	pop	%r15
+	pop	%r14
+	pop	%r13
+	pop	%r12
 
 	ret
 ENDPROC(sha1_x8_avx2)
-- 
2.4.11