Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:7 as permitted sender) client-ip=2620:137:e000::3:7;
From:   Charlie Jenkins <charlie@rivosinc.com>
Date:   Fri, 27 Oct 2023 15:43:54 -0700
Subject: [PATCH v8 4/5] riscv: Add checksum library
MIME-Version: 1.0
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
Message-Id: <20231027-optimize_checksum-v8-4-feb7101d128d@rivosinc.com>
References: <20231027-optimize_checksum-v8-0-feb7101d128d@rivosinc.com>
In-Reply-To: <20231027-optimize_checksum-v8-0-feb7101d128d@rivosinc.com>
To:     Charlie Jenkins <charlie@rivosinc.com>,
        Palmer Dabbelt <palmer@dabbelt.com>,
        Conor Dooley <conor@kernel.org>,
        Samuel Holland <samuel.holland@sifive.com>,
        David Laight <David.Laight@aculab.com>,
        Xiao Wang <xiao.w.wang@intel.com>,
        Evan Green <evan@rivosinc.com>,
        linux-riscv@lists.infradead.org, linux-kernel@vger.kernel.org,
        linux-arch@vger.kernel.org
Cc:     Paul Walmsley <paul.walmsley@sifive.com>,
        Albert Ou <aou@eecs.berkeley.edu>,
        Arnd Bergmann <arnd@arndb.de>,
        Conor Dooley <conor.dooley@microchip.com>
Precedence: bulk

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.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
Acked-by: Conor Dooley <conor.dooley@microchip.com>
---
 arch/riscv/lib/Makefile |   1 +
 arch/riscv/lib/csum.c   | 339 ++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 340 insertions(+)

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..f90e73606597
--- /dev/null
+++ b/arch/riscv/lib/csum.c
@@ -0,0 +1,339 @@
+// 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 <asm/cpufeature.h>
+#include <linux/jump_label.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)
+{
+	unsigned int ulen, uproto;
+	unsigned long sum = csum;
+
+	sum += saddr->s6_addr32[0];
+	sum += saddr->s6_addr32[1];
+	sum += saddr->s6_addr32[2];
+	sum += saddr->s6_addr32[3];
+
+	sum += daddr->s6_addr32[0];
+	sum += daddr->s6_addr32[1];
+	sum += daddr->s6_addr32[2];
+	sum += daddr->s6_addr32[3];
+
+	ulen = htonl((unsigned int)len);
+	sum += ulen;
+
+	uproto = htonl(proto);
+	sum += uproto;
+
+	/*
+	 * Zbb support saves 4 instructions, so not worth checking without
+	 * alternatives if supported
+	 */
+	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
+	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
+		unsigned long 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 += ror64(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
+
+/*
+ * 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.
+ */
+static inline __no_sanitize_address unsigned int do_csum_with_alignment(const unsigned char *buff, int len)
+{
+	unsigned int offset, shift;
+	unsigned long csum = 0, carry = 0, data;
+	const unsigned long *ptr, *end;
+
+	end = (const unsigned long *)(buff + len);
+
+	/*
+	 * Align address to closest word (double word on rv64) that comes before
+	 * buff. This should always be in the same page and cache line.
+	 * Directly call KASAN with the alignment we will be using.
+	 */
+	offset = (unsigned long)buff & OFFSET_MASK;
+	kasan_check_read(buff, len);
+	ptr = (const unsigned long *)(buff - offset);
+
+	/*
+	 * Clear the most significant bytes 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 (ptr < end) {
+		csum += data;
+		carry += csum < data;
+		len -= sizeof(long);
+		data = *(ptr++);
+	}
+
+	/*
+	 * Perform alignment (and over-read) bytes on the tail if any bytes
+	 * leftover.
+	 */
+	shift = ((long)ptr - (long)end) * 8;
+#ifdef __LITTLE_ENDIAN
+	data = (data << shift) >> shift;
+#else
+	data = (data >> shift) << shift;
+#endif
+	csum += data;
+	carry += csum < data;
+	csum += carry;
+	csum += csum < carry;
+
+	/*
+	 * Zbb support saves 6 instructions, so not worth checking without
+	 * alternatives if supported
+	 */
+	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
+	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
+		unsigned long 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);
+
+#ifdef 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\
+		.option pop"
+			: [csum] "+r" (csum),
+				[fold_temp] "=&r" (fold_temp)
+			: [offset] "r" (offset)
+			:
+			: end);
+
+		return (unsigned short)csum;
+#else /* !CONFIG_32BIT */
+		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\
+		.option pop"
+			: [csum] "+r" (csum),
+				[fold_temp] "=&r" (fold_temp)
+			: [offset] "r" (offset)
+			:
+			: end);
+
+		return (csum << 16) >> 48;
+#endif /* !CONFIG_32BIT */
+end:
+		return csum >> 16;
+	}
+no_zbb:
+#ifndef CONFIG_32BIT
+	csum += ror64(csum, 32);
+	csum >>= 32;
+#endif
+	csum = (u32)csum + ror32((u32)csum, 16);
+	if (offset & 1)
+		return (u16)swab32(csum);
+	return csum >> 16;
+}
+
+/*
+ * Does not perform alignment, should only be used if machine has fast
+ * misaligned accesses, because buff may be misaligned.
+ */
+static inline unsigned int do_csum_no_alignment(const unsigned char *buff, int len)
+{
+	unsigned int offset, shift;
+	unsigned long csum = 0, carry = 0, data;
+	const unsigned long *ptr, *end;
+
+	end = (const unsigned long *)(buff + len);
+
+	ptr = (const unsigned long *)(buff);
+
+	data = *(ptr++);
+
+	/*
+	 * 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 (ptr < end) {
+		csum += data;
+		carry += csum < data;
+		len -= sizeof(long);
+		data = *(ptr++);
+	}
+
+	/*
+	 * Perform alignment (and over-read) bytes on the tail if any bytes
+	 * leftover.
+	 */
+	shift = ((long)ptr - (long)end) * 8;
+#ifdef __LITTLE_ENDIAN
+	data = (data << shift) >> shift;
+#else
+	data = (data >> shift) << shift;
+#endif
+	csum += data;
+	carry += csum < data;
+	csum += carry;
+	csum += csum < carry;
+
+	/*
+	 * Zbb support saves 6 instructions, so not worth checking without
+	 * alternatives if supported
+	 */
+	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
+	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
+		unsigned long 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);
+
+#ifdef CONFIG_32BIT
+		asm (".option push				\n\
+		.option arch,+zbb				\n\
+			rori	%[fold_temp], %[csum], 16	\n\
+			andi	%[offset], %[offset], 1		\n\
+			add	%[csum], %[fold_temp], %[csum]	\n\
+		.option pop"
+			: [csum] "+r" (csum),
+				[fold_temp] "=&r" (fold_temp)
+			: [offset] "r" (offset)
+			: );
+
+#else /* !CONFIG_32BIT */
+		asm (".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\
+		.option pop"
+			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp)
+			: [offset] "r" (offset)
+			: );
+#endif /* !CONFIG_32BIT */
+		return csum >> 16;
+	}
+no_zbb:
+#ifndef CONFIG_32BIT
+	csum += ror64(csum, 32);
+	csum >>= 32;
+#endif
+	csum = (u32)csum + ror32((u32)csum, 16);
+	return csum >> 16;
+}
+
+/*
+ * Perform a checksum on an arbitrary memory address.
+ * Will do a light-weight address alignment if buff is misaligned, unless
+ * cpu supports fast misaligned accesses.
+ */
+unsigned int do_csum(const unsigned char *buff, int len)
+{
+	if (unlikely(len <= 0))
+		return 0;
+
+	/*
+	 * Very significant performance gains can be seen by not doing alignment
+	 * on machines with fast misaligned accesses.
+	 *
+	 * There is some duplicate code between the "with_alignment" and
+	 * "no_alignment" implmentations, but the overlap is too awkward to be
+	 * able to fit in one function without introducing multiple static
+	 * branches.
+	 */
+	if (static_branch_likely(&fast_misaligned_access_speed_key))
+		return do_csum_no_alignment(buff, len);
+
+	return do_csum_with_alignment(buff, len);
+}

-- 
2.42.0