The C standard does not support dereferencing pointers that are not
aligned with respect to the pointed-to type, and doing so is technically
undefined behavior, even if the underlying hardware supports it.
This means that conditionally dereferencing such pointers based on
whether CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS=y is not the right thing
to do, and actually results in alignment faults on ARM, which are fixed
up on a slow path. Instead, we should use the unaligned accessors in
such cases: on architectures that don't care about alignment, they will
result in identical codegen whereas, e.g., codegen on ARM will avoid
doubleword loads and stores but use ordinary ones, which are able to
tolerate misalignment.
Link: https://lore.kernel.org/linux-crypto/CAHk-=wiKkdYLY0bv+nXrcJz3NH9mAqPAafX7PpW5EwVtxsEu7Q@mail.gmail.com/
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Commit fe8c8a1268 introduced a possible build error for archs
that do not have CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS set. :/
Fix this up by bringing else braces outside of the ifdef.
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Fixes: fe8c8a1268 ("crypto: more robust crypto_memneq")
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Acked-By: Cesar Eduardo Barros <cesarb@cesarb.eti.br>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Disabling compiler optimizations can be fragile, since a new
optimization could be added to -O0 or -Os that breaks the assumptions
the code is making.
Instead of disabling compiler optimizations, use a dummy inline assembly
(based on RELOC_HIDE) to block the problematic kinds of optimization,
while still allowing other optimizations to be applied to the code.
The dummy inline assembly is added after every OR, and has the
accumulator variable as its input and output. The compiler is forced to
assume that the dummy inline assembly could both depend on the
accumulator variable and change the accumulator variable, so it is
forced to compute the value correctly before the inline assembly, and
cannot assume anything about its value after the inline assembly.
This change should be enough to make crypto_memneq work correctly (with
data-independent timing) even if it is inlined at its call sites. That
can be done later in a followup patch.
Compile-tested on x86_64.
Signed-off-by: Cesar Eduardo Barros <cesarb@cesarb.eti.br>
Acked-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
When comparing MAC hashes, AEAD authentication tags, or other hash
values in the context of authentication or integrity checking, it
is important not to leak timing information to a potential attacker,
i.e. when communication happens over a network.
Bytewise memory comparisons (such as memcmp) are usually optimized so
that they return a nonzero value as soon as a mismatch is found. E.g,
on x86_64/i5 for 512 bytes this can be ~50 cyc for a full mismatch
and up to ~850 cyc for a full match (cold). This early-return behavior
can leak timing information as a side channel, allowing an attacker to
iteratively guess the correct result.
This patch adds a new method crypto_memneq ("memory not equal to each
other") to the crypto API that compares memory areas of the same length
in roughly "constant time" (cache misses could change the timing, but
since they don't reveal information about the content of the strings
being compared, they are effectively benign). Iow, best and worst case
behaviour take the same amount of time to complete (in contrast to
memcmp).
Note that crypto_memneq (unlike memcmp) can only be used to test for
equality or inequality, NOT for lexicographical order. This, however,
is not an issue for its use-cases within the crypto API.
We tried to locate all of the places in the crypto API where memcmp was
being used for authentication or integrity checking, and convert them
over to crypto_memneq.
crypto_memneq is declared noinline, placed in its own source file,
and compiled with optimizations that might increase code size disabled
("Os") because a smart compiler (or LTO) might notice that the return
value is always compared against zero/nonzero, and might then
reintroduce the same early-return optimization that we are trying to
avoid.
Using #pragma or __attribute__ optimization annotations of the code
for disabling optimization was avoided as it seems to be considered
broken or unmaintained for long time in GCC [1]. Therefore, we work
around that by specifying the compile flag for memneq.o directly in
the Makefile. We found that this seems to be most appropriate.
As we use ("Os"), this patch also provides a loop-free "fast-path" for
frequently used 16 byte digests. Similarly to kernel library string
functions, leave an option for future even further optimized architecture
specific assembler implementations.
This was a joint work of James Yonan and Daniel Borkmann. Also thanks
for feedback from Florian Weimer on this and earlier proposals [2].
[1] http://gcc.gnu.org/ml/gcc/2012-07/msg00211.html
[2] https://lkml.org/lkml/2013/2/10/131
Signed-off-by: James Yonan <james@openvpn.net>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Florian Weimer <fw@deneb.enyo.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Ard Biesheuvel