Biggest part, ~7%, of improvement resulted from omitting constants'
table index increment in each round. And minor part from rescheduling
instructions. Apparently POWER9 (and POWER8) manage to dispatch
instructions more efficiently if they are laid down as if they have
no latency...
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6406)
This comes at cost of minor 2.5% regression on G4, which is reasonable
trade-off. [Further improve compliance with ABI requirements.]
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6406)
As it turns out originally published results were skewed by "turbo"
mode. VM apparently remains oblivious to dynamic frequency scaling,
and reports that processor operates at "base" frequency at all times.
While actual frequency gets increased under load.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6406)
OPENSSL_memcmp is a must in GCM decrypt and general-purpose loop takes
quite a portion of execution time for short inputs, more than GHASH for
few-byte inputs according to profiler. Special 16-byte case takes it off
top five list in profiler output.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6312)
On contemporary platforms assembly GHASH processes multiple blocks
faster than one by one. For TLS payloads shorter than 16 bytes, e.g.
alerts, it's possible to reduce hashing operation to single call.
And for block lengths not divisible by 16 - fold two final calls to
one. Improvement is most noticeable with "reptoline", because call to
assembly GHASH is indirect.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6312)
Even though calls can be viewed as styling improvement, they do come
with cost. It's not big cost and shows only on short inputs, but it is
measurable, 2-3% on some platforms.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6312)
Only Linux and FreeBSD provide getrandom(), but they both also provide
getentropy() since the same version and we already tried to call that.
Reviewed-by: Andy Polyakov <appro@openssl.org>
GH: #6405
This will actually support most OSs, and at least adds support for
Solaris and OSX
Fixes: #6403
Reviewed-by: Andy Polyakov <appro@openssl.org>
GH: #6405
ECDSA_SIG_new() returns NULL on error.
Reviewed-by: Matthias St. Pierre <Matthias.St.Pierre@ncp-e.com>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6398)
Just because an engine implements algorithm methods, that doesn't mean
it also implements the ASN1 method. Therefore, be careful when looking
for an ASN1 method among all engines, don't try to use one that doesn't
exist.
Fixes#6381
Reviewed-by: Tim Hudson <tjh@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6383)
(cherry picked from commit 1ac3cd6277)
(cherry picked from commit 13b578ada3)
This is probably a "should not happen" scenario, but better check anyway.
Found by Coverity.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6373)
Found by coverity. This is an artifact left over from the original
decaf import which generated the source code for different curves. For
curve 448 this is dead.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6373)
XN_FLAG_COMPAT has a unique property, its zero for value. This means
it needs special treatment; if it has been set (which can only be
determined indirectly) and set alone (*), no other flags should be
set.
(*) if any other nameopt flag has been set by the user, compatibility
mode is blown away.
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6382)
In `aes_wrap_cipher()`, the minimal out buff length is `(inlen - 8)`.
Since it calls `CRYPTO_128_unwrap_pad()` underneath, it makes sense to
reduce the minimal out length in `CRYPTO_128_unwrap_pad()` to align to
its caller.
Signed-off-by: Yihong Wang <yh.wang@ibm.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6266)
The One&Done attack, which is described in a paper to appear in the
USENIX Security'18 conference, uses EM emanations to recover the values
of the bits that are obtained using BN_is_bit_set while constructing
the value of the window in BN_mod_exp_consttime. The EM signal changes
slightly depending on the value of the bit, and since the lookup of a
bit is surrounded by highly regular execution (constant-time Montgomery
multiplications) the attack is able to isolate the (very brief) part of
the signal that changes depending on the bit. Although the change is
slight, the attack recovers it successfully >90% of the time on several
phones and IoT devices (all with ARM processors with clock rates around
1GHz), so after only one RSA decryption more than 90% of the bits in
d_p and d_q are recovered correctly, which enables rapid recovery of
the full RSA key using an algorithm (also described in the paper) that
modifies the branch-and-prune approach for a situation in which the
exponents' bits are recovered with errors, i.e. where we do not know
a priori which bits are correctly recovered.
The mitigation for the attack is relatively simple - all the bits of
the window are obtained at once, along with other bits so that an
entire integer's worth of bits are obtained together using masking and
shifts, without unnecessarily considering each bit in isolation. This
improves performance somewhat (one call to bn_get_bits is faster than
several calls to BN_is_bit_set), so the attacker now gets one signal
snippet per window (rather than one per bit) in which the signal is
affected by all bits in the integer (rather than just the one bit).
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6276)
32-bit vector rotate instruction was defined from beginning, it
not being used from the start must be a brain-slip...
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6363)
Goal is to exercise AEAD ciphers in TLS-like sequence, i.e. 13-byte
AAD followed by payload. Update doc/man1/speed.pod accordingly.
[While we are at it, address even some styling and readability issues.]
Reviewed-by: Tim Hudson <tjh@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6311)
Under a mingw shell, the command line path conversion either mangles
file: URIs to something useless (file;C:\...) or not at all (which
can't be opened by the Windows C RTL unless we're really lucky), so we
simply skip testing them in that environment.
Fixes#6369
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6376)
OpenSSL 1.1.0 made the X509_LOOKUP_METHOD structure opaque, so
applications that were previously able to define a custom lookup method
are not able to be ported.
This commit adds getters and setters for each of the current fields of
X509_LOOKUP_METHOD, along with getters and setters on several associated
opaque types (such as X509_LOOKUP and X509_OBJECT).
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6152)
It's freed with OPENSSL_free()
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6331)
Return immediately upon discovery of bad message digest.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6298)
