]> git.ipfire.org Git - thirdparty/kernel/stable.git/commit
crypto: x86/twofish-3way - Fix %rbp usage
authorEric Biggers <ebiggers@google.com>
Tue, 19 Dec 2017 00:40:26 +0000 (16:40 -0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 25 Feb 2018 10:03:36 +0000 (11:03 +0100)
commit6bfbf2aacb212002e774bb586dd2d0edfd470077
treee1a623d6ab139f8f89d708ebcd9b0296bc47e842
parent002924ab7b023d16d3291d5ae600c49cebed1c59
crypto: x86/twofish-3way - Fix %rbp usage

commit d8c7fe9f2a486a6e5f0d5229ca43807af5ab22c6 upstream.

Using %rbp as a temporary register breaks frame pointer convention and
breaks stack traces when unwinding from an interrupt in the crypto code.

In twofish-3way, we can't simply replace %rbp with another register
because there are none available.  Instead, we use the stack to hold the
values that %rbp, %r11, and %r12 were holding previously.  Each of these
values represents the half of the output from the previous Feistel round
that is being passed on unchanged to the following round.  They are only
used once per round, when they are exchanged with %rax, %rbx, and %rcx.

As a result, we free up 3 registers (one per block) and can reassign
them so that %rbp is not used, and additionally %r14 and %r15 are not
used so they do not need to be saved/restored.

There may be a small overhead caused by replacing 'xchg REG, REG' with
the needed sequence 'mov MEM, REG; mov REG, MEM; mov REG, REG' once per
round.  But, counterintuitively, when I tested "ctr-twofish-3way" on a
Haswell processor, the new version was actually about 2% faster.
(Perhaps 'xchg' is not as well optimized as plain moves.)

Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/x86/crypto/twofish-x86_64-asm_64-3way.S