Message-Id: <20230911-optimize_checksum-v4-2-77cc2ad9e9d7@rivosinc.com>
Date:   Mon, 11 Sep 2023 15:57:12 -0700
From:   Charlie Jenkins <charlie@...osinc.com>
To:     Charlie Jenkins <charlie@...osinc.com>,
        Palmer Dabbelt <palmer@...belt.com>,
        Conor Dooley <conor@...nel.org>,
        Samuel Holland <samuel.holland@...ive.com>,
        David Laight <David.Laight@...lab.com>,
        linux-riscv@...ts.infradead.org, linux-kernel@...r.kernel.org
Cc:     Paul Walmsley <paul.walmsley@...ive.com>,
        Albert Ou <aou@...s.berkeley.edu>
Subject: [PATCH v4 2/5] riscv: Add checksum library

Provide a 32 and 64 bit version of do_csum. When compiled for 32-bit
will load from the buffer in groups of 32 bits, and when compiled for
64-bit will load in groups of 64 bits. Benchmarking by proxy compiling
csum_ipv6_magic (64-bit version) for an x86 chip as well as running
the riscv generated code in QEMU, discovered that summing in a
tree-like structure is about 4% faster than doing 64-bit reads.

Signed-off-by: Charlie Jenkins <charlie@...osinc.com>
---
 arch/riscv/include/asm/checksum.h |  11 ++
 arch/riscv/lib/Makefile           |   1 +
 arch/riscv/lib/csum.c             | 210 ++++++++++++++++++++++++++++++++++++++
 3 files changed, 222 insertions(+)

diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h
index 0d7fc8275a5e..a09a4053fb87 100644
--- a/arch/riscv/include/asm/checksum.h
+++ b/arch/riscv/include/asm/checksum.h
@@ -16,6 +16,14 @@ typedef unsigned int csum_t;
 typedef unsigned long csum_t;
 #endif
 
+/* Default version is sufficient for 32 bit */
+#ifdef CONFIG_64BIT
+#define _HAVE_ARCH_IPV6_CSUM
+__sum16 csum_ipv6_magic(const struct in6_addr *saddr,
+			const struct in6_addr *daddr,
+			__u32 len, __u8 proto, __wsum sum);
+#endif
+
 /*
  *	Fold a partial checksum without adding pseudo headers
  */
@@ -90,6 +98,9 @@ static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
 
 #define ip_fast_csum ip_fast_csum
 
+extern unsigned int do_csum(const unsigned char *buff, int len);
+#define do_csum do_csum
+
 #include <asm-generic/checksum.h>
 
 #endif // __ASM_RISCV_CHECKSUM_H
diff --git a/arch/riscv/lib/Makefile b/arch/riscv/lib/Makefile
index 26cb2502ecf8..2aa1a4ad361f 100644
--- a/arch/riscv/lib/Makefile
+++ b/arch/riscv/lib/Makefile
@@ -6,6 +6,7 @@ lib-y			+= memmove.o
 lib-y			+= strcmp.o
 lib-y			+= strlen.o
 lib-y			+= strncmp.o
+lib-y			+= csum.o
 lib-$(CONFIG_MMU)	+= uaccess.o
 lib-$(CONFIG_64BIT)	+= tishift.o
 lib-$(CONFIG_RISCV_ISA_ZICBOZ)	+= clear_page.o
diff --git a/arch/riscv/lib/csum.c b/arch/riscv/lib/csum.c
new file mode 100644
index 000000000000..47d98c51bab2
--- /dev/null
+++ b/arch/riscv/lib/csum.c
@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * IP checksum library
+ *
+ * Influenced by arch/arm64/lib/csum.c
+ * Copyright (C) 2023 Rivos Inc.
+ */
+#include <linux/bitops.h>
+#include <linux/compiler.h>
+#include <linux/kasan-checks.h>
+#include <linux/kernel.h>
+
+#include <net/checksum.h>
+
+/* Default version is sufficient for 32 bit */
+#ifndef CONFIG_32BIT
+__sum16 csum_ipv6_magic(const struct in6_addr *saddr,
+			const struct in6_addr *daddr,
+			__u32 len, __u8 proto, __wsum csum)
+{
+	/*
+	 * Inform the compiler/processor that the operation we are performing is
+	 * "Commutative and Associative" by summing parts of the checksum in a
+	 * tree-like structure (Section 2(A) of "Computing the Internet
+	 * Checksum"). Furthermore, defer the overflow until the end of the
+	 * computation which is shown to be valid in Section 2(C)(1) of the
+	 * same handbook.
+	 */
+	unsigned long sum, sum1, sum2, sum3, sum4, ulen, uproto;
+
+	uproto = htonl(proto);
+	ulen = htonl(len);
+
+	sum   = saddr->s6_addr32[0];
+	sum  += saddr->s6_addr32[1];
+	sum1  = saddr->s6_addr32[2];
+	sum1 += saddr->s6_addr32[3];
+
+	sum2  = daddr->s6_addr32[0];
+	sum2 += daddr->s6_addr32[1];
+	sum3  = daddr->s6_addr32[2];
+	sum3 += daddr->s6_addr32[3];
+
+	sum4  = csum;
+	sum4 += ulen;
+	sum4 += uproto;
+
+	sum  += sum1;
+	sum2 += sum3;
+
+	sum += sum2;
+	sum += sum4;
+
+	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
+	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
+		csum_t fold_temp;
+
+		/*
+		 * Zbb is likely available when the kernel is compiled with Zbb
+		 * support, so nop when Zbb is available and jump when Zbb is
+		 * not available.
+		 */
+		asm_volatile_goto(ALTERNATIVE("j %l[no_zbb]", "nop", 0,
+					      RISCV_ISA_EXT_ZBB, 1)
+				  :
+				  :
+				  :
+				  : no_zbb);
+		asm(".option push					\n\
+		.option arch,+zbb					\n\
+			rori	%[fold_temp], %[sum], 32		\n\
+			add	%[sum], %[fold_temp], %[sum]		\n\
+			srli	%[sum], %[sum], 32			\n\
+			not	%[fold_temp], %[sum]			\n\
+			roriw	%[sum], %[sum], 16			\n\
+			subw	%[sum], %[fold_temp], %[sum]		\n\
+		.option pop"
+		: [sum] "+r" (sum), [fold_temp] "=&r" (fold_temp));
+		return (__force __sum16)(sum >> 16);
+	}
+no_zbb:
+	sum += (sum >> 32) | (sum << 32);
+	sum >>= 32;
+	return csum_fold((__force __wsum)sum);
+}
+EXPORT_SYMBOL(csum_ipv6_magic);
+#endif // !CONFIG_32BIT
+
+#ifdef CONFIG_32BIT
+#define OFFSET_MASK 3
+#elif CONFIG_64BIT
+#define OFFSET_MASK 7
+#endif
+
+/*
+ * Perform a checksum on an arbitrary memory address.
+ * Algorithm accounts for buff being misaligned.
+ * If buff is not aligned, will over-read bytes but not use the bytes that it
+ * shouldn't. The same thing will occur on the tail-end of the read.
+ */
+unsigned int __no_sanitize_address do_csum(const unsigned char *buff, int len)
+{
+	unsigned int offset, shift;
+	csum_t csum, data;
+	const csum_t *ptr;
+
+	if (unlikely(len <= 0))
+		return 0;
+	/*
+	 * To align the address, grab the whole first byte in buff.
+	 * Since it is inside of a same byte, it will never cross pages or cache
+	 * lines.
+	 * Directly call KASAN with the alignment we will be using.
+	 */
+	offset = (csum_t)buff & OFFSET_MASK;
+	kasan_check_read(buff, len);
+	ptr = (const csum_t *)(buff - offset);
+	len = len + offset - sizeof(csum_t);
+
+	/*
+	 * Clear the most signifant bits that were over-read if buff was not
+	 * aligned.
+	 */
+	shift = offset * 8;
+	data = *ptr;
+#ifdef __LITTLE_ENDIAN
+	data = (data >> shift) << shift;
+#else
+	data = (data << shift) >> shift;
+#endif
+	/*
+	 * Do 32-bit reads on RV32 and 64-bit reads otherwise. This should be
+	 * faster than doing 32-bit reads on architectures that support larger
+	 * reads.
+	 */
+	while (len > 0) {
+		csum += data;
+		csum += csum < data;
+		len -= sizeof(csum_t);
+		ptr += 1;
+		data = *ptr;
+	}
+
+	/*
+	 * Perform alignment (and over-read) bytes on the tail if any bytes
+	 * leftover.
+	 */
+	shift = len * -8;
+#ifdef __LITTLE_ENDIAN
+	data = (data << shift) >> shift;
+#else
+	data = (data >> shift) << shift;
+#endif
+	csum += data;
+	csum += csum < data;
+
+	if (!riscv_has_extension_likely(RISCV_ISA_EXT_ZBB))
+		goto no_zbb;
+
+	unsigned int fold_temp;
+
+	if (IS_ENABLED(CONFIG_32BIT)) {
+		asm_volatile_goto(".option push			\n\
+		.option arch,+zbb				\n\
+			rori	%[fold_temp], %[csum], 16	\n\
+			andi	%[offset], %[offset], 1		\n\
+			add	%[csum], %[fold_temp], %[csum]	\n\
+			beq	%[offset], zero, %l[end]	\n\
+			rev8	%[csum], %[csum]		\n\
+			zext.h	%[csum], %[csum]		\n\
+		.option pop"
+			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
+			: [offset] "r" (offset)
+			:
+			: end);
+
+		return csum;
+	} else {
+		asm_volatile_goto(".option push			\n\
+		.option arch,+zbb				\n\
+			rori	%[fold_temp], %[csum], 32	\n\
+			add	%[csum], %[fold_temp], %[csum]	\n\
+			srli	%[csum], %[csum], 32		\n\
+			roriw	%[fold_temp], %[csum], 16	\n\
+			addw	%[csum], %[fold_temp], %[csum]	\n\
+			andi	%[offset], %[offset], 1		\n\
+			beq	%[offset], zero, %l[end]	\n\
+			rev8	%[csum], %[csum]		\n\
+			srli	%[csum], %[csum], 32		\n\
+			zext.h	%[csum], %[csum]		\n\
+		.option pop"
+			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
+			: [offset] "r" (offset)
+			:
+			: end);
+
+		return csum;
+	}
+end:
+		return csum >> 16;
+no_zbb:
+#ifndef CONFIG_32BIT
+		csum += (csum >> 32) | (csum << 32);
+		csum >>= 32;
+#endif
+	csum = (unsigned int)csum + (((unsigned int)csum >> 16) | ((unsigned int)csum << 16));
+	if (offset & 1)
+		return (unsigned short)swab32(csum);
+	return csum >> 16;
+}

-- 
2.42.0

Powered by blists - more mailing lists