On 32-bit ARM, AES in GCM mode takes full advantage of the ARMv8 Crypto
Extensions when available, resulting in a performance of 6-7 cycles per
byte for typical IPsec frames on cores such as Cortex-A53, using the
generic GCM template encapsulating the accelerated AES-CTR and GHASH
implementations.
At such high rates, any time spent copying data or doing other poorly
optimized work in the generic layer hurts disproportionately, and we can
get a significant performance improvement by combining the optimized
AES-CTR and GHASH implementations into a single one.
On Cortex-A53, this results in a performance improvement of around 70%,
or 4.2 cycles per byte for AES-256-GCM-128 with RFC4106 encapsulation.
The fastest mode on this core is bare AES-128-GCM using 8k blocks, which
manages 2.66 cycles per byte.
Signed-off-by: Ard Biesheuvel <[email protected]>
---
Note: this patch depends on the softirq context patches for kernel mode
NEON I sent last week. More specifically, this implements a sync AEAD
that does not implement a !simd fallback, as AEADs are not callable in
IRQ context anyway.
arch/arm/crypto/Kconfig | 2 +
arch/arm/crypto/ghash-ce-core.S | 381 +++++++++++++++++++-
arch/arm/crypto/ghash-ce-glue.c | 350 +++++++++++++++++-
3 files changed, 718 insertions(+), 15 deletions(-)
diff --git a/arch/arm/crypto/Kconfig b/arch/arm/crypto/Kconfig
index 3858c4d4cb98854d..c693a4fdf3771e63 100644
--- a/arch/arm/crypto/Kconfig
+++ b/arch/arm/crypto/Kconfig
@@ -16,9 +16,11 @@ config CRYPTO_CURVE25519_NEON
config CRYPTO_GHASH_ARM_CE
tristate "Hash functions: GHASH (PMULL/NEON/ARMv8 Crypto Extensions)"
depends on KERNEL_MODE_NEON
+ select CRYPTO_AEAD
select CRYPTO_HASH
select CRYPTO_CRYPTD
select CRYPTO_GF128MUL
+ select CRYPTO_LIB_AES
help
GCM GHASH function (NIST SP800-38D)
diff --git a/arch/arm/crypto/ghash-ce-core.S b/arch/arm/crypto/ghash-ce-core.S
index 9f51e3fa45268de9..e0a4dc20a3a96832 100644
--- a/arch/arm/crypto/ghash-ce-core.S
+++ b/arch/arm/crypto/ghash-ce-core.S
@@ -44,7 +44,7 @@
t2q .req q7
t3q .req q8
t4q .req q9
- T2 .req q9
+ XH2 .req q9
s1l .req d20
s1h .req d21
@@ -80,7 +80,7 @@
XL2 .req q5
XM2 .req q6
- XH2 .req q7
+ T2 .req q7
T3 .req q8
XL2_L .req d10
@@ -192,23 +192,42 @@
vshr.u64 XL, XL, #1
.endm
- .macro ghash_update, pn
+ .macro ghash_update, pn, enc, aggregate=1, head=1
vld1.64 {XL}, [r1]
+ .if \head
/* do the head block first, if supplied */
- ldr ip, [sp]
teq ip, #0
beq 0f
vld1.64 {T1}, [ip]
teq r0, #0
b 3f
+ .endif
0: .ifc \pn, p64
+ .if \aggregate
tst r0, #3 // skip until #blocks is a
bne 2f // round multiple of 4
vld1.8 {XL2-XM2}, [r2]!
-1: vld1.8 {T3-T2}, [r2]!
+1: vld1.8 {T2-T3}, [r2]!
+
+ .ifnb \enc
+ \enc\()_4x XL2, XM2, T2, T3
+
+ add ip, r3, #16
+ vld1.64 {HH}, [ip, :128]!
+ vld1.64 {HH3-HH4}, [ip, :128]
+
+ veor SHASH2_p64, SHASH_L, SHASH_H
+ veor SHASH2_H, HH_L, HH_H
+ veor HH34_L, HH3_L, HH3_H
+ veor HH34_H, HH4_L, HH4_H
+
+ vmov.i8 MASK, #0xe1
+ vshl.u64 MASK, MASK, #57
+ .endif
+
vrev64.8 XL2, XL2
vrev64.8 XM2, XM2
@@ -218,8 +237,8 @@
veor XL2_H, XL2_H, XL_L
veor XL, XL, T1
- vrev64.8 T3, T3
- vrev64.8 T1, T2
+ vrev64.8 T1, T3
+ vrev64.8 T3, T2
vmull.p64 XH, HH4_H, XL_H // a1 * b1
veor XL2_H, XL2_H, XL_H
@@ -267,14 +286,22 @@
b 1b
.endif
+ .endif
+
+2: vld1.8 {T1}, [r2]!
+
+ .ifnb \enc
+ \enc\()_1x T1
+ veor SHASH2_p64, SHASH_L, SHASH_H
+ vmov.i8 MASK, #0xe1
+ vshl.u64 MASK, MASK, #57
+ .endif
-2: vld1.64 {T1}, [r2]!
subs r0, r0, #1
3: /* multiply XL by SHASH in GF(2^128) */
-#ifndef CONFIG_CPU_BIG_ENDIAN
vrev64.8 T1, T1
-#endif
+
vext.8 IN1, T1, T1, #8
veor T1_L, T1_L, XL_H
veor XL, XL, IN1
@@ -293,9 +320,6 @@
veor XL, XL, T1
bne 0b
-
- vst1.64 {XL}, [r1]
- bx lr
.endm
/*
@@ -315,7 +339,11 @@ ENTRY(pmull_ghash_update_p64)
vmov.i8 MASK, #0xe1
vshl.u64 MASK, MASK, #57
+ ldr ip, [sp]
ghash_update p64
+ vst1.64 {XL}, [r1]
+
+ bx lr
ENDPROC(pmull_ghash_update_p64)
ENTRY(pmull_ghash_update_p8)
@@ -335,5 +363,332 @@ ENTRY(pmull_ghash_update_p8)
vmov.i64 k32, #0xffffffff
vmov.i64 k48, #0xffffffffffff
+ ldr ip, [sp]
ghash_update p8
+ vst1.64 {XL}, [r1]
+
+ bx lr
ENDPROC(pmull_ghash_update_p8)
+
+ e0 .req q9
+ e1 .req q10
+ e2 .req q11
+ e3 .req q12
+ e0l .req d18
+ e0h .req d19
+ e2l .req d22
+ e2h .req d23
+ e3l .req d24
+ e3h .req d25
+ ctr .req q13
+ ctr0 .req d26
+ ctr1 .req d27
+
+ ek0 .req q14
+ ek1 .req q15
+
+ .macro round, rk:req, regs:vararg
+ .irp r, \regs
+ aese.8 \r, \rk
+ aesmc.8 \r, \r
+ .endr
+ .endm
+
+ .macro aes_encrypt, rkp, rounds, regs:vararg
+ vld1.8 {ek0-ek1}, [\rkp, :128]!
+ cmp \rounds, #12
+ blt .L\@_1 // AES-128
+ beq .L\@_0 // AES-192
+
+ round ek0, \regs
+ vld1.8 {ek0}, [\rkp, :128]!
+ round ek1, \regs
+ vld1.8 {ek1}, [\rkp, :128]!
+
+.L\@_0: round ek0, \regs
+ vld1.8 {ek0}, [\rkp, :128]!
+ round ek1, \regs
+ vld1.8 {ek1}, [\rkp, :128]!
+
+.L\@_1: .rept 4
+ round ek0, \regs
+ vld1.8 {ek0}, [\rkp, :128]!
+ round ek1, \regs
+ vld1.8 {ek1}, [\rkp, :128]!
+ .endr
+
+ round ek0, \regs
+ vld1.8 {ek0}, [\rkp, :128]
+
+ .irp r, \regs
+ aese.8 \r, ek1
+ .endr
+ .irp r, \regs
+ veor \r, \r, ek0
+ .endr
+ .endm
+
+pmull_aes_encrypt:
+ add ip, r5, #4
+ vld1.8 {ctr0}, [r5] // load 12 byte IV
+ vld1.8 {ctr1}, [ip]
+ rev r8, r7
+ vext.8 ctr1, ctr1, ctr1, #4
+ add r7, r7, #1
+ vmov.32 ctr1[1], r8
+ vmov e0, ctr
+
+ add ip, r3, #64
+ aes_encrypt ip, r6, e0
+ bx lr
+ENDPROC(pmull_aes_encrypt)
+
+pmull_aes_encrypt_4x:
+ add ip, r5, #4
+ vld1.8 {ctr0}, [r5]
+ vld1.8 {ctr1}, [ip]
+ rev r8, r7
+ vext.8 ctr1, ctr1, ctr1, #4
+ add r7, r7, #1
+ vmov.32 ctr1[1], r8
+ rev ip, r7
+ vmov e0, ctr
+ add r7, r7, #1
+ vmov.32 ctr1[1], ip
+ rev r8, r7
+ vmov e1, ctr
+ add r7, r7, #1
+ vmov.32 ctr1[1], r8
+ rev ip, r7
+ vmov e2, ctr
+ add r7, r7, #1
+ vmov.32 ctr1[1], ip
+ vmov e3, ctr
+
+ add ip, r3, #64
+ aes_encrypt ip, r6, e0, e1, e2, e3
+ bx lr
+ENDPROC(pmull_aes_encrypt_4x)
+
+pmull_aes_encrypt_final:
+ add ip, r5, #4
+ vld1.8 {ctr0}, [r5]
+ vld1.8 {ctr1}, [ip]
+ rev r8, r7
+ vext.8 ctr1, ctr1, ctr1, #4
+ mov r7, #1 << 24 // BE #1 for the tag
+ vmov.32 ctr1[1], r8
+ vmov e0, ctr
+ vmov.32 ctr1[1], r7
+ vmov e1, ctr
+
+ add ip, r3, #64
+ aes_encrypt ip, r6, e0, e1
+ bx lr
+ENDPROC(pmull_aes_encrypt_final)
+
+ .macro enc_1x, in0
+ bl pmull_aes_encrypt
+ veor \in0, \in0, e0
+ vst1.8 {\in0}, [r4]!
+ .endm
+
+ .macro dec_1x, in0
+ bl pmull_aes_encrypt
+ veor e0, e0, \in0
+ vst1.8 {e0}, [r4]!
+ .endm
+
+ .macro enc_4x, in0, in1, in2, in3
+ bl pmull_aes_encrypt_4x
+
+ veor \in0, \in0, e0
+ veor \in1, \in1, e1
+ veor \in2, \in2, e2
+ veor \in3, \in3, e3
+
+ vst1.8 {\in0-\in1}, [r4]!
+ vst1.8 {\in2-\in3}, [r4]!
+ .endm
+
+ .macro dec_4x, in0, in1, in2, in3
+ bl pmull_aes_encrypt_4x
+
+ veor e0, e0, \in0
+ veor e1, e1, \in1
+ veor e2, e2, \in2
+ veor e3, e3, \in3
+
+ vst1.8 {e0-e1}, [r4]!
+ vst1.8 {e2-e3}, [r4]!
+ .endm
+
+ /*
+ * void pmull_gcm_encrypt(int blocks, u64 dg[], const char *src,
+ * struct gcm_key const *k, char *dst,
+ * char *iv, int rounds, u32 counter)
+ */
+ENTRY(pmull_gcm_encrypt)
+ mov ip, sp
+ push {r4-r8, lr}
+ ldm ip, {r4-r7}
+
+ vld1.64 {SHASH}, [r3]
+
+ ghash_update p64, enc, head=0
+ vst1.64 {XL}, [r1]
+
+ pop {r4-r8, pc}
+ENDPROC(pmull_gcm_encrypt)
+
+ /*
+ * void pmull_gcm_decrypt(int blocks, u64 dg[], const char *src,
+ * struct gcm_key const *k, char *dst,
+ * char *iv, int rounds, u32 counter)
+ */
+ENTRY(pmull_gcm_decrypt)
+ mov ip, sp
+ push {r4-r8, lr}
+ ldm ip, {r4-r7}
+
+ vld1.64 {SHASH}, [r3]
+
+ ghash_update p64, dec, head=0
+ vst1.64 {XL}, [r1]
+
+ pop {r4-r8, pc}
+ENDPROC(pmull_gcm_decrypt)
+
+ /*
+ * void pmull_gcm_enc_final(int bytes, u64 dg[], char *tag,
+ * struct gcm_key const *k, char *head,
+ * char *iv, int rounds, u32 counter)
+ */
+ENTRY(pmull_gcm_enc_final)
+ mov ip, sp
+ push {r4-r8, lr}
+ ldm ip, {r4-r7}
+
+ bl pmull_aes_encrypt_final
+
+ cmp r0, #0
+ beq .Lenc_final
+
+ adr_l ip, .Lpermute
+ sub r4, r4, #16
+ add r8, ip, r0
+ add ip, ip, #32
+ add r4, r4, r0
+ sub ip, ip, r0
+
+ vld1.8 {e3}, [r8] // permute vector for key stream
+ vld1.8 {e2}, [ip] // permute vector for ghash input
+
+ vtbl.8 e3l, {e0}, e3l
+ vtbl.8 e3h, {e0}, e3h
+
+ vld1.8 {e0}, [r4] // encrypt tail block
+ veor e0, e0, e3
+ vst1.8 {e0}, [r4]
+
+ vtbl.8 T1_L, {e0}, e2l
+ vtbl.8 T1_H, {e0}, e2h
+
+ vld1.64 {XL}, [r1]
+.Lenc_final:
+ vld1.64 {SHASH}, [r3, :128]
+ vmov.i8 MASK, #0xe1
+ veor SHASH2_p64, SHASH_L, SHASH_H
+ vshl.u64 MASK, MASK, #57
+ mov r0, #1
+ bne 3f // process head block first
+ ghash_update p64, aggregate=0, head=0
+
+ vrev64.8 XL, XL
+ vext.8 XL, XL, XL, #8
+ veor XL, XL, e1
+
+ sub r2, r2, #16 // rewind src pointer
+ vst1.8 {XL}, [r2] // store tag
+
+ pop {r4-r8, pc}
+ENDPROC(pmull_gcm_enc_final)
+
+ /*
+ * int pmull_gcm_dec_final(int bytes, u64 dg[], char *tag,
+ * struct gcm_key const *k, char *head,
+ * char *iv, int rounds, u32 counter,
+ * const char *otag, int authsize)
+ */
+ENTRY(pmull_gcm_dec_final)
+ mov ip, sp
+ push {r4-r8, lr}
+ ldm ip, {r4-r7}
+
+ bl pmull_aes_encrypt_final
+
+ cmp r0, #0
+ beq .Ldec_final
+
+ adr_l ip, .Lpermute
+ sub r4, r4, #16
+ add r8, ip, r0
+ add ip, ip, #32
+ add r4, r4, r0
+ sub ip, ip, r0
+
+ vld1.8 {e3}, [r8] // permute vector for key stream
+ vld1.8 {e2}, [ip] // permute vector for ghash input
+
+ vtbl.8 e3l, {e0}, e3l
+ vtbl.8 e3h, {e0}, e3h
+
+ vld1.8 {e0}, [r4]
+
+ vtbl.8 T1_L, {e0}, e2l
+ vtbl.8 T1_H, {e0}, e2h
+
+ veor e0, e0, e3
+ vst1.8 {e0}, [r4]
+
+ vld1.64 {XL}, [r1]
+.Ldec_final:
+ vld1.64 {SHASH}, [r3]
+ vmov.i8 MASK, #0xe1
+ veor SHASH2_p64, SHASH_L, SHASH_H
+ vshl.u64 MASK, MASK, #57
+ mov r0, #1
+ bne 3f // process head block first
+ ghash_update p64, aggregate=0, head=0
+
+ vrev64.8 XL, XL
+ vext.8 XL, XL, XL, #8
+ veor XL, XL, e1
+
+ adr_l ip, .Lpermute
+ ldrd r2, r3, [sp, #40] // otag and authsize
+ vld1.8 {T1}, [r2]
+ add ip, ip, r3
+ vceq.i8 T1, T1, XL // compare tags
+ vmvn T1, T1 // 0 for eq, -1 for ne
+
+ vld1.8 {e0}, [ip]
+ vtbl.8 XL_L, {T1}, e0l // keep authsize bytes only
+ vtbl.8 XL_H, {T1}, e0h
+
+ vpmin.s8 XL_L, XL_L, XL_H // take the minimum s8 across the vector
+ vpmin.s8 XL_L, XL_L, XL_L
+ vmov r0, XL_L[0] // fail if != 0x0
+
+ pop {r4-r8, pc}
+ENDPROC(pmull_gcm_dec_final)
+
+ .section ".rodata", "a", %progbits
+ .align 5
+.Lpermute:
+ .byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+ .byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+ .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07
+ .byte 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
+ .byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+ .byte 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
diff --git a/arch/arm/crypto/ghash-ce-glue.c b/arch/arm/crypto/ghash-ce-glue.c
index 3e598284498865cf..f234c6fdc0472791 100644
--- a/arch/arm/crypto/ghash-ce-glue.c
+++ b/arch/arm/crypto/ghash-ce-glue.c
@@ -9,10 +9,16 @@
#include <asm/neon.h>
#include <asm/simd.h>
#include <asm/unaligned.h>
+#include <crypto/aes.h>
+#include <crypto/gcm.h>
+#include <crypto/b128ops.h>
#include <crypto/cryptd.h>
+#include <crypto/internal/aead.h>
#include <crypto/internal/hash.h>
#include <crypto/internal/simd.h>
+#include <crypto/internal/skcipher.h>
#include <crypto/gf128mul.h>
+#include <crypto/scatterwalk.h>
#include <linux/cpufeature.h>
#include <linux/crypto.h>
#include <linux/jump_label.h>
@@ -25,12 +31,19 @@ MODULE_ALIAS_CRYPTO("ghash");
#define GHASH_BLOCK_SIZE 16
#define GHASH_DIGEST_SIZE 16
+#define GCM_IV_SIZE 12
struct ghash_key {
u64 h0[2];
u64 h[][2];
};
+struct gcm_key {
+ u64 h[4][2];
+ u32 rk[AES_MAX_KEYLENGTH_U32];
+ int rounds;
+};
+
struct ghash_desc_ctx {
u64 digest[GHASH_DIGEST_SIZE/sizeof(u64)];
u8 buf[GHASH_BLOCK_SIZE];
@@ -324,6 +337,332 @@ static struct ahash_alg ghash_async_alg = {
},
};
+
+void pmull_gcm_encrypt(int blocks, u64 dg[], const char *src,
+ struct gcm_key const *k, char *dst,
+ char *iv, int rounds, u32 counter);
+
+void pmull_gcm_enc_final(int blocks, u64 dg[], char *tag,
+ struct gcm_key const *k, char *head,
+ char *iv, int rounds, u32 counter);
+
+void pmull_gcm_decrypt(int bytes, u64 dg[], const char *src,
+ struct gcm_key const *k, char *dst,
+ char *iv, int rounds, u32 counter);
+
+int pmull_gcm_dec_final(int bytes, u64 dg[], char *tag,
+ struct gcm_key const *k, char *head,
+ char *iv, int rounds, u32 counter,
+ const char *otag, int authsize);
+
+static int gcm_setkey(struct crypto_aead *tfm, const u8 *inkey,
+ unsigned int keylen)
+{
+ struct gcm_key *ctx = crypto_aead_ctx(tfm);
+ struct crypto_aes_ctx aes_ctx;
+ be128 h, k;
+ int ret;
+
+ ret = aes_expandkey(&aes_ctx, inkey, keylen);
+ if (ret)
+ return -EINVAL;
+
+ aes_encrypt(&aes_ctx, (u8 *)&k, (u8[AES_BLOCK_SIZE]){});
+
+ memcpy(ctx->rk, aes_ctx.key_enc, sizeof(ctx->rk));
+ ctx->rounds = 6 + keylen / 4;
+
+ memzero_explicit(&aes_ctx, sizeof(aes_ctx));
+
+ ghash_reflect(ctx->h[0], &k);
+
+ h = k;
+ gf128mul_lle(&h, &k);
+ ghash_reflect(ctx->h[1], &h);
+
+ gf128mul_lle(&h, &k);
+ ghash_reflect(ctx->h[2], &h);
+
+ gf128mul_lle(&h, &k);
+ ghash_reflect(ctx->h[3], &h);
+
+ return 0;
+}
+
+static int gcm_setauthsize(struct crypto_aead *tfm, unsigned int authsize)
+{
+ switch (authsize) {
+ case 4:
+ case 8:
+ case 12 ... 16:
+ break;
+ default:
+ return -EINVAL;
+ }
+ return 0;
+}
+
+static void gcm_update_mac(u64 dg[], const u8 *src, int count, u8 buf[],
+ int *buf_count, struct gcm_key *ctx)
+{
+ if (*buf_count > 0) {
+ int buf_added = min(count, GHASH_BLOCK_SIZE - *buf_count);
+
+ memcpy(&buf[*buf_count], src, buf_added);
+
+ *buf_count += buf_added;
+ src += buf_added;
+ count -= buf_added;
+ }
+
+ if (count >= GHASH_BLOCK_SIZE || *buf_count == GHASH_BLOCK_SIZE) {
+ int blocks = count / GHASH_BLOCK_SIZE;
+
+ pmull_ghash_update_p64(blocks, dg, src, ctx->h,
+ *buf_count ? buf : NULL);
+
+ src += blocks * GHASH_BLOCK_SIZE;
+ count %= GHASH_BLOCK_SIZE;
+ *buf_count = 0;
+ }
+
+ if (count > 0) {
+ memcpy(buf, src, count);
+ *buf_count = count;
+ }
+}
+
+static void gcm_calculate_auth_mac(struct aead_request *req, u64 dg[])
+{
+ struct crypto_aead *aead = crypto_aead_reqtfm(req);
+ struct gcm_key *ctx = crypto_aead_ctx(aead);
+ u8 buf[GHASH_BLOCK_SIZE];
+ struct scatter_walk walk;
+ u32 len = req->assoclen;
+ int buf_count = 0;
+
+ scatterwalk_start(&walk, req->src);
+
+ do {
+ u32 n = scatterwalk_clamp(&walk, len);
+ u8 *p;
+
+ if (!n) {
+ scatterwalk_start(&walk, sg_next(walk.sg));
+ n = scatterwalk_clamp(&walk, len);
+ }
+ p = scatterwalk_map(&walk);
+
+ gcm_update_mac(dg, p, n, buf, &buf_count, ctx);
+
+ scatterwalk_unmap(p);
+
+ if (unlikely(len / SZ_4K > (len - n) / SZ_4K)) {
+ kernel_neon_end();
+ kernel_neon_begin();
+ }
+
+ len -= n;
+ scatterwalk_advance(&walk, n);
+ scatterwalk_done(&walk, 0, len);
+ } while (len);
+
+ if (buf_count) {
+ memset(&buf[buf_count], 0, GHASH_BLOCK_SIZE - buf_count);
+ pmull_ghash_update_p64(1, dg, buf, ctx->h, NULL);
+ }
+}
+
+static int gcm_encrypt(struct aead_request *req)
+{
+ struct crypto_aead *aead = crypto_aead_reqtfm(req);
+ struct gcm_key *ctx = crypto_aead_ctx(aead);
+ struct skcipher_walk walk;
+ u8 buf[AES_BLOCK_SIZE];
+ u32 counter = 2;
+ u64 dg[2] = {};
+ be128 lengths;
+ const u8 *src;
+ u8 *tag, *dst;
+ int tail, err;
+
+ if (WARN_ON_ONCE(!may_use_simd()))
+ return -EBUSY;
+
+ err = skcipher_walk_aead_encrypt(&walk, req, false);
+
+ kernel_neon_begin();
+
+ if (req->assoclen)
+ gcm_calculate_auth_mac(req, dg);
+
+ src = walk.src.virt.addr;
+ dst = walk.dst.virt.addr;
+
+ while (walk.nbytes >= AES_BLOCK_SIZE) {
+ int nblocks = walk.nbytes / AES_BLOCK_SIZE;
+
+ pmull_gcm_encrypt(nblocks, dg, src, ctx, dst, req->iv,
+ ctx->rounds, counter);
+ counter += nblocks;
+
+ if (walk.nbytes == walk.total && walk.nbytes % AES_BLOCK_SIZE) {
+ src += nblocks * AES_BLOCK_SIZE;
+ dst += nblocks * AES_BLOCK_SIZE;
+ break;
+ }
+
+ kernel_neon_end();
+
+ err = skcipher_walk_done(&walk,
+ walk.nbytes % AES_BLOCK_SIZE);
+ if (err)
+ return err;
+
+ src = walk.src.virt.addr;
+ dst = walk.dst.virt.addr;
+
+ kernel_neon_begin();
+ }
+
+
+ lengths.a = cpu_to_be64(req->assoclen * 8);
+ lengths.b = cpu_to_be64(req->cryptlen * 8);
+
+ tag = (u8 *)&lengths;
+ tail = walk.nbytes % AES_BLOCK_SIZE;
+
+ /*
+ * Bounce via a buffer unless we are encrypting in place and src/dst
+ * are not pointing to the start of the walk buffer. In that case, we
+ * can do a NEON load/xor/store sequence in place as long as we move
+ * the plain/ciphertext and keystream to the start of the register. If
+ * not, do a memcpy() to the end of the buffer so we can reuse the same
+ * logic.
+ */
+ if (unlikely(tail && (tail == walk.nbytes || src != dst)))
+ src = memcpy(buf + sizeof(buf) - tail, src, tail);
+
+ pmull_gcm_enc_final(tail, dg, tag, ctx, (u8 *)src, req->iv,
+ ctx->rounds, counter);
+ kernel_neon_end();
+
+ if (unlikely(tail && src != dst))
+ memcpy(dst, src, tail);
+
+ if (walk.nbytes) {
+ err = skcipher_walk_done(&walk, 0);
+ if (err)
+ return err;
+ }
+
+ /* copy authtag to end of dst */
+ scatterwalk_map_and_copy(tag, req->dst, req->assoclen + req->cryptlen,
+ crypto_aead_authsize(aead), 1);
+
+ return 0;
+}
+
+static int gcm_decrypt(struct aead_request *req)
+{
+ struct crypto_aead *aead = crypto_aead_reqtfm(req);
+ struct gcm_key *ctx = crypto_aead_ctx(aead);
+ int authsize = crypto_aead_authsize(aead);
+ struct skcipher_walk walk;
+ u8 otag[AES_BLOCK_SIZE];
+ u8 buf[AES_BLOCK_SIZE];
+ u32 counter = 2;
+ u64 dg[2] = {};
+ be128 lengths;
+ const u8 *src;
+ u8 *tag, *dst;
+ int tail, err, ret;
+
+ if (WARN_ON_ONCE(!may_use_simd()))
+ return -EBUSY;
+
+ scatterwalk_map_and_copy(otag, req->src,
+ req->assoclen + req->cryptlen - authsize,
+ authsize, 0);
+
+ err = skcipher_walk_aead_decrypt(&walk, req, false);
+
+ kernel_neon_begin();
+
+ if (req->assoclen)
+ gcm_calculate_auth_mac(req, dg);
+
+ src = walk.src.virt.addr;
+ dst = walk.dst.virt.addr;
+
+ while (walk.nbytes >= AES_BLOCK_SIZE) {
+ int nblocks = walk.nbytes / AES_BLOCK_SIZE;
+
+ pmull_gcm_decrypt(nblocks, dg, src, ctx, dst, req->iv,
+ ctx->rounds, counter);
+ counter += nblocks;
+
+ if (walk.nbytes == walk.total && walk.nbytes % AES_BLOCK_SIZE) {
+ src += nblocks * AES_BLOCK_SIZE;
+ dst += nblocks * AES_BLOCK_SIZE;
+ break;
+ }
+
+ kernel_neon_end();
+
+ err = skcipher_walk_done(&walk,
+ walk.nbytes % AES_BLOCK_SIZE);
+ if (err)
+ return err;
+
+ src = walk.src.virt.addr;
+ dst = walk.dst.virt.addr;
+
+ kernel_neon_begin();
+ }
+
+ lengths.a = cpu_to_be64(req->assoclen * 8);
+ lengths.b = cpu_to_be64((req->cryptlen - authsize) * 8);
+
+ tag = (u8 *)&lengths;
+ tail = walk.nbytes % AES_BLOCK_SIZE;
+
+ if (unlikely(tail && (tail == walk.nbytes || src != dst)))
+ src = memcpy(buf + sizeof(buf) - tail, src, tail);
+
+ ret = pmull_gcm_dec_final(tail, dg, tag, ctx, (u8 *)src, req->iv,
+ ctx->rounds, counter, otag, authsize);
+ kernel_neon_end();
+
+ if (unlikely(tail && src != dst))
+ memcpy(dst, src, tail);
+
+ if (walk.nbytes) {
+ err = skcipher_walk_done(&walk, 0);
+ if (err)
+ return err;
+ }
+
+ return ret ? -EBADMSG : 0;
+}
+
+static struct aead_alg gcm_aes_alg = {
+ .ivsize = GCM_IV_SIZE,
+ .chunksize = AES_BLOCK_SIZE,
+ .maxauthsize = AES_BLOCK_SIZE,
+ .setkey = gcm_setkey,
+ .setauthsize = gcm_setauthsize,
+ .encrypt = gcm_encrypt,
+ .decrypt = gcm_decrypt,
+
+ .base.cra_name = "gcm(aes)",
+ .base.cra_driver_name = "gcm-aes-ce",
+ .base.cra_priority = 400,
+ .base.cra_blocksize = 1,
+ .base.cra_ctxsize = sizeof(struct gcm_key),
+ .base.cra_module = THIS_MODULE,
+};
+
static int __init ghash_ce_mod_init(void)
{
int err;
@@ -332,21 +671,26 @@ static int __init ghash_ce_mod_init(void)
return -ENODEV;
if (elf_hwcap2 & HWCAP2_PMULL) {
+ err = crypto_register_aead(&gcm_aes_alg);
+ if (err)
+ return err;
ghash_alg.base.cra_ctxsize += 3 * sizeof(u64[2]);
static_branch_enable(&use_p64);
}
err = crypto_register_shash(&ghash_alg);
if (err)
- return err;
+ goto err_aead;
err = crypto_register_ahash(&ghash_async_alg);
if (err)
goto err_shash;
return 0;
-
err_shash:
crypto_unregister_shash(&ghash_alg);
+err_aead:
+ if (elf_hwcap2 & HWCAP2_PMULL)
+ crypto_unregister_aead(&gcm_aes_alg);
return err;
}
@@ -354,6 +698,8 @@ static void __exit ghash_ce_mod_exit(void)
{
crypto_unregister_ahash(&ghash_async_alg);
crypto_unregister_shash(&ghash_alg);
+ if (elf_hwcap2 & HWCAP2_PMULL)
+ crypto_unregister_aead(&gcm_aes_alg);
}
module_init(ghash_ce_mod_init);
--
2.35.1
On Fri, Dec 30, 2022 at 11:03:43PM +0800, Herbert Xu wrote:
> On Mon, Dec 12, 2022 at 07:37:58PM +0100, Ard Biesheuvel wrote:
> > On 32-bit ARM, AES in GCM mode takes full advantage of the ARMv8 Crypto
> > Extensions when available, resulting in a performance of 6-7 cycles per
> > byte for typical IPsec frames on cores such as Cortex-A53, using the
> > generic GCM template encapsulating the accelerated AES-CTR and GHASH
> > implementations.
> >
> > At such high rates, any time spent copying data or doing other poorly
> > optimized work in the generic layer hurts disproportionately, and we can
> > get a significant performance improvement by combining the optimized
> > AES-CTR and GHASH implementations into a single one.
> >
> > On Cortex-A53, this results in a performance improvement of around 70%,
> > or 4.2 cycles per byte for AES-256-GCM-128 with RFC4106 encapsulation.
> > The fastest mode on this core is bare AES-128-GCM using 8k blocks, which
> > manages 2.66 cycles per byte.
> >
> > Signed-off-by: Ard Biesheuvel <[email protected]>
> > ---
> > Note: this patch depends on the softirq context patches for kernel mode
> > NEON I sent last week. More specifically, this implements a sync AEAD
> > that does not implement a !simd fallback, as AEADs are not callable in
> > IRQ context anyway.
> >
> > arch/arm/crypto/Kconfig | 2 +
> > arch/arm/crypto/ghash-ce-core.S | 381 +++++++++++++++++++-
> > arch/arm/crypto/ghash-ce-glue.c | 350 +++++++++++++++++-
> > 3 files changed, 718 insertions(+), 15 deletions(-)
>
> Patch applied. Thanks.
Oops, this email was an error. This patch hasn't been applied
as it was already superceded.
--
Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
On Mon, Dec 12, 2022 at 07:37:58PM +0100, Ard Biesheuvel wrote:
> On 32-bit ARM, AES in GCM mode takes full advantage of the ARMv8 Crypto
> Extensions when available, resulting in a performance of 6-7 cycles per
> byte for typical IPsec frames on cores such as Cortex-A53, using the
> generic GCM template encapsulating the accelerated AES-CTR and GHASH
> implementations.
>
> At such high rates, any time spent copying data or doing other poorly
> optimized work in the generic layer hurts disproportionately, and we can
> get a significant performance improvement by combining the optimized
> AES-CTR and GHASH implementations into a single one.
>
> On Cortex-A53, this results in a performance improvement of around 70%,
> or 4.2 cycles per byte for AES-256-GCM-128 with RFC4106 encapsulation.
> The fastest mode on this core is bare AES-128-GCM using 8k blocks, which
> manages 2.66 cycles per byte.
>
> Signed-off-by: Ard Biesheuvel <[email protected]>
> ---
> Note: this patch depends on the softirq context patches for kernel mode
> NEON I sent last week. More specifically, this implements a sync AEAD
> that does not implement a !simd fallback, as AEADs are not callable in
> IRQ context anyway.
>
> arch/arm/crypto/Kconfig | 2 +
> arch/arm/crypto/ghash-ce-core.S | 381 +++++++++++++++++++-
> arch/arm/crypto/ghash-ce-glue.c | 350 +++++++++++++++++-
> 3 files changed, 718 insertions(+), 15 deletions(-)
Patch applied. Thanks.
--
Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt