ICP: Improve AES-GCM performance

Currently SIMD accelerated AES-GCM performance is limited by two
factors:

a. The need to disable preemption and interrupts and save the FPU
state before using it and to do the reverse when done. Due to the
way the code is organized (see (b) below) we have to pay this price
twice for each 16 byte GCM block processed.

b. Most processing is done in C, operating on single GCM blocks.
The use of SIMD instructions is limited to the AES encryption of the
counter block (AES-NI) and the Galois multiplication (PCLMULQDQ).
This leads to the FPU not being fully utilized for crypto
operations.

To solve (a) we do crypto processing in larger chunks while owning
the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced
to make this chunk size tweakable. It defaults to 32 KiB. This step
alone roughly doubles performance. (b) is tackled by porting and
using the highly optimized openssl AES-GCM assembler routines, which
do all the processing (CTR, AES, GMULT) in a single routine. Both
steps together result in up to 32x reduction of the time spend in
the en/decryption routines, leading up to approximately 12x
throughput increase for large (128 KiB) blocks.

Signed-off-by: Attila Fülöp <[email protected]>

config/toolchain-simd.m4

@@ -401,3 +401,23 @@ AC_DEFUN([ZFS_AC_CONFIG_TOOLCHAIN_CAN_BUILD_PCLMULQDQ], [
 		AC_MSG_RESULT([no])
 	])
 ])
+
+dnl #
+dnl # ZFS_AC_CONFIG_TOOLCHAIN_CAN_BUILD_MOVBE
+dnl #
+AC_DEFUN([ZFS_AC_CONFIG_TOOLCHAIN_CAN_BUILD_MOVBE], [
+	AC_MSG_CHECKING([whether host toolchain supports MOVBE])
+
+	AC_LINK_IFELSE([AC_LANG_SOURCE([
+	[
+		void main()
+		{
+			__asm__ __volatile__("movbe 0(%eax), %eax");
+		}
+	]])], [
+		AC_MSG_RESULT([yes])
+		AC_DEFINE([HAVE_MOVBE], 1, [Define if host toolchain supports MOVBE])
+	], [
+		AC_MSG_RESULT([no])
+	])
+])

include/os/linux/kernel/linux/simd_x86.h

@@ -477,6 +477,19 @@ zfs_pclmulqdq_available(void)
 #endif
 }
 
+/*
+ * Check if MOVBE instruction is available
+ */
+static inline boolean_t
+zfs_movbe_available(void)
+{
+#if defined(X86_FEATURE_MOVBE)
+	return (!!boot_cpu_has(X86_FEATURE_MOVBE));
+#else
+	return (B_FALSE);
+#endif
+}
+
 /*
  * AVX-512 family of instruction sets:
  *

lib/libicp/Makefile.am

@@ -15,6 +15,8 @@ ASM_SOURCES_AS = \
 	asm-x86_64/aes/aes_amd64.S \
 	asm-x86_64/aes/aes_aesni.S \
 	asm-x86_64/modes/gcm_pclmulqdq.S \
+	asm-x86_64/modes/aesni-gcm-x86_64.S \
+	asm-x86_64/modes/ghash-x86_64.S \
 	asm-x86_64/sha1/sha1-x86_64.S \
 	asm-x86_64/sha2/sha256_impl.S \
 	asm-x86_64/sha2/sha512_impl.S

lib/libspl/include/sys/simd.h

@@ -77,7 +77,8 @@ typedef enum cpuid_inst_sets {
 	AVX512ER,
 	AVX512VL,
 	AES,
-	PCLMULQDQ
+	PCLMULQDQ,
+	MOVBE
 } cpuid_inst_sets_t;
 
 /*
@@ -101,6 +102,7 @@ typedef struct cpuid_feature_desc {
 #define	_AVX512VL_BIT		(1U << 31) /* if used also check other levels */
 #define	_AES_BIT		(1U << 25)
 #define	_PCLMULQDQ_BIT		(1U << 1)
+#define	_MOVBE_BIT		(1U << 22)
 
 /*
  * Descriptions of supported instruction sets
@@ -128,6 +130,7 @@ static const cpuid_feature_desc_t cpuid_features[] = {
 	[AVX512VL]	= {7U, 0U, _AVX512ER_BIT,	EBX	},
 	[AES]		= {1U, 0U, _AES_BIT,		ECX	},
 	[PCLMULQDQ]	= {1U, 0U, _PCLMULQDQ_BIT,	ECX	},
+	[MOVBE]		= {1U, 0U, _MOVBE_BIT,		ECX	},
 };
 
 /*
@@ -200,6 +203,7 @@ CPUID_FEATURE_CHECK(avx512er, AVX512ER);
 CPUID_FEATURE_CHECK(avx512vl, AVX512VL);
 CPUID_FEATURE_CHECK(aes, AES);
 CPUID_FEATURE_CHECK(pclmulqdq, PCLMULQDQ);
+CPUID_FEATURE_CHECK(movbe, MOVBE);
 
 /*
  * Detect register set support
@@ -332,6 +336,15 @@ zfs_pclmulqdq_available(void)
 	return (__cpuid_has_pclmulqdq());
 }
 
+/*
+ * Check if MOVBE instruction is available
+ */
+static inline boolean_t
+zfs_movbe_available(void)
+{
+	return (__cpuid_has_movbe());
+}
+
 /*
  * AVX-512 family of instruction sets:
  *

module/icp/Makefile.in

@@ -10,6 +10,8 @@ ASM_SOURCES := asm-x86_64/aes/aeskey.o
 ASM_SOURCES += asm-x86_64/aes/aes_amd64.o
 ASM_SOURCES += asm-x86_64/aes/aes_aesni.o
 ASM_SOURCES += asm-x86_64/modes/gcm_pclmulqdq.o
+ASM_SOURCES += asm-x86_64/modes/aesni-gcm-x86_64.o
+ASM_SOURCES += asm-x86_64/modes/ghash-x86_64.o
 ASM_SOURCES += asm-x86_64/sha1/sha1-x86_64.o
 ASM_SOURCES += asm-x86_64/sha2/sha256_impl.o
 ASM_SOURCES += asm-x86_64/sha2/sha512_impl.o