Received-SPF: pass (google.com: domain of linux-crypto-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Mon, 16 Jan 2023 14:59:46 +0000
To:     "linux-crypto@vger.kernel.org" <linux-crypto@vger.kernel.org>
From:   Peter Lafreniere <peter@n8pjl.ca>
Cc:     "ardb@kernel.org" <ardb@kernel.org>,
        "x86@kernel.org" <x86@kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Peter Lafreniere <peter@n8pjl.ca>
Subject: [PATCH] crypto: x86 - exit fpu context earlier in ECB/CBC macros
Message-ID: <IBooTlGWpNE7pOelt0gm21bxW7wBILNYJ1HaoPbbfdEEMwz0Pp92vpd_OUlhNFNAitFThTi27P6q6NvcYMKm-y7tjwiF9YbImWjhgC3UDMk=@n8pjl.ca>
Feedback-ID: 53133685:user:proton
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Currently the ecb/cbc macros hold fpu context unnecessarily when using=20
scalar cipher routines (e.g. when handling odd sizes of blocks per walk).

Change the macros to drop fpu context as soon as the fpu is out of use.

No performance impact found (on Intel Haswell).

Signed-off-by: Peter Lafreniere <peter@n8pjl.ca>
---
 arch/x86/crypto/ecb_cbc_helpers.h | 19 +++++++++++++++----
 1 file changed, 15 insertions(+), 4 deletions(-)

diff --git a/arch/x86/crypto/ecb_cbc_helpers.h b/arch/x86/crypto/ecb_cbc_he=
lpers.h
index eaa15c7b29d6..b83085e18ab0 100644
--- a/arch/x86/crypto/ecb_cbc_helpers.h
+++ b/arch/x86/crypto/ecb_cbc_helpers.h
@@ -14,12 +14,13 @@
 #define ECB_WALK_START(req, bsize, fpu_blocks) do {=09=09=09\
 =09void *ctx =3D crypto_skcipher_ctx(crypto_skcipher_reqtfm(req));=09\
+=09const int __fpu_blocks =3D (fpu_blocks);=09=09=09=09\
 =09const int __bsize =3D (bsize);=09=09=09=09=09\
 =09struct skcipher_walk walk;=09=09=09=09=09\
 =09int err =3D skcipher_walk_virt(&walk, (req), false);=09=09\
 =09while (walk.nbytes > 0) {=09=09=09=09=09\
 =09=09unsigned int nbytes =3D walk.nbytes;=09=09=09\
-=09=09bool do_fpu =3D (fpu_blocks) !=3D -1 &&=09=09=09\
-=09=09=09      nbytes >=3D (fpu_blocks) * __bsize;=09=09\
+=09=09bool do_fpu =3D __fpu_blocks !=3D -1 &&=09=09=09\
+=09=09=09      nbytes >=3D __fpu_blocks * __bsize;=09=09\
 =09=09const u8 *src =3D walk.src.virt.addr;=09=09=09\
 =09=09u8 *dst =3D walk.dst.virt.addr;=09=09=09=09\
 =09=09u8 __maybe_unused buf[(bsize)];=09=09=09=09\
@@ -35,7 +36,12 @@
 } while (0)
=20
 #define ECB_BLOCK(blocks, func) do {=09=09=09=09=09\
-=09while (nbytes >=3D (blocks) * __bsize) {=09=09=09=09\
+=09const int __blocks =3D (blocks);=09=09=09=09=09\
+=09if (do_fpu && __blocks < __fpu_blocks) {=09=09=09\
+=09=09kernel_fpu_end();=09=09=09=09=09\
+=09=09do_fpu =3D false;=09=09=09=09=09=09\
+=09}=09=09=09=09=09=09=09=09\
+=09while (nbytes >=3D __blocks * __bsize) {=09=09=09=09\
 =09=09(func)(ctx, dst, src);=09=09=09=09=09\
 =09=09ECB_WALK_ADVANCE(blocks);=09=09=09=09\
 =09}=09=09=09=09=09=09=09=09\
@@ -53,7 +59,12 @@
 } while (0)
=20
 #define CBC_DEC_BLOCK(blocks, func) do {=09=09=09=09\
-=09while (nbytes >=3D (blocks) * __bsize) {=09=09=09=09\
+=09const int __blocks =3D (blocks);=09=09=09=09=09\
+=09if (do_fpu && __blocks <  __fpu_blocks) {=09=09=09\
+=09=09kernel_fpu_end();=09=09=09=09=09\
+=09=09do_fpu =3D false;=09=09=09=09=09=09\
+=09}=09=09=09=09=09=09=09=09\
+=09while (nbytes >=3D __blocks * __bsize) {=09=09=09=09\
 =09=09const u8 *__iv =3D src + ((blocks) - 1) * __bsize;=09\
 =09=09if (dst =3D=3D src)=09=09=09=09=09=09\
 =09=09=09__iv =3D memcpy(buf, __iv, __bsize);=09=09\
--=20
2.39.0