diff options
author | Charlie Jenkins <charlie@rivosinc.com> | 2024-01-09 02:57:05 +0300 |
---|---|---|
committer | Palmer Dabbelt <palmer@rivosinc.com> | 2024-01-18 04:52:32 +0300 |
commit | a04c192eabfb76824d00f1b4cd0f25844a59d0f0 (patch) | |
tree | 4958cc44bee3652ebefe01c374dd058151d58b36 /arch/riscv | |
parent | e11e367e9fe57164ea609807ed27184c85263355 (diff) | |
download | linux-a04c192eabfb76824d00f1b4cd0f25844a59d0f0.tar.xz |
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.
Additionally provide riscv optimized implementation of csum_ipv6_magic.
Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
Acked-by: Conor Dooley <conor.dooley@microchip.com>
Reviewed-by: Xiao Wang <xiao.w.wang@intel.com>
Link: https://lore.kernel.org/r/20240108-optimize_checksum-v15-4-1c50de5f2167@rivosinc.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
Diffstat (limited to 'arch/riscv')
-rw-r--r-- | arch/riscv/include/asm/checksum.h | 11 | ||||
-rw-r--r-- | arch/riscv/lib/Makefile | 1 | ||||
-rw-r--r-- | arch/riscv/lib/csum.c | 326 |
3 files changed, 338 insertions, 0 deletions
diff --git a/arch/riscv/include/asm/checksum.h b/arch/riscv/include/asm/checksum.h index 5a810126aac7..a5b60b54b101 100644 --- a/arch/riscv/include/asm/checksum.h +++ b/arch/riscv/include/asm/checksum.h @@ -12,6 +12,17 @@ #define ip_fast_csum ip_fast_csum +extern unsigned int do_csum(const unsigned char *buff, int len); +#define do_csum do_csum + +/* Default version is sufficient for 32 bit */ +#ifndef CONFIG_32BIT +#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 + /* Define riscv versions of functions before importing asm-generic/checksum.h */ #include <asm-generic/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..06ce8e7250d9 --- /dev/null +++ b/arch/riscv/lib/csum.c @@ -0,0 +1,326 @@ +// SPDX-License-Identifier: GPL-2.0 +/* + * Checksum library + * + * Influenced by arch/arm64/lib/csum.c + * Copyright (C) 2023 Rivos Inc. + */ +#include <linux/bitops.h> +#include <linux/compiler.h> +#include <linux/jump_label.h> +#include <linux/kasan-checks.h> +#include <linux/kernel.h> + +#include <asm/cpufeature.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 = (__force unsigned long)csum; + + sum += (__force unsigned long)saddr->s6_addr32[0]; + sum += (__force unsigned long)saddr->s6_addr32[1]; + sum += (__force unsigned long)saddr->s6_addr32[2]; + sum += (__force unsigned long)saddr->s6_addr32[3]; + + sum += (__force unsigned long)daddr->s6_addr32[0]; + sum += (__force unsigned long)daddr->s6_addr32[1]; + sum += (__force unsigned long)daddr->s6_addr32[2]; + sum += (__force unsigned long)daddr->s6_addr32[3]; + + ulen = (__force unsigned int)htonl((unsigned int)len); + sum += ulen; + + uproto = (__force unsigned int)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 + +static inline __no_sanitize_address unsigned long +do_csum_common(const unsigned long *ptr, const unsigned long *end, + unsigned long data) +{ + unsigned int shift; + unsigned long csum = 0, carry = 0; + + /* + * 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; + 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; + + return csum; +} + +/* + * 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, data; + const unsigned long *ptr, *end; + + /* + * 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 + end = (const unsigned long *)(buff + len); + csum = do_csum_common(ptr, end, data); + + /* + * 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, or when buff is known to be aligned. + */ +static inline __no_sanitize_address unsigned int +do_csum_no_alignment(const unsigned char *buff, int len) +{ + unsigned long csum, data; + const unsigned long *ptr, *end; + + ptr = (const unsigned long *)(buff); + data = *(ptr++); + + kasan_check_read(buff, len); + + end = (const unsigned long *)(buff + len); + csum = do_csum_common(ptr, end, data); + + /* + * 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\ + add %[csum], %[fold_temp], %[csum] \n\ + .option pop" + : [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp) + : + : ); + +#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) + : + : ); +#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; + + /* + * 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. The largest chunk of overlap was delegated into the + * do_csum_common function. + */ + if (static_branch_likely(&fast_misaligned_access_speed_key)) + return do_csum_no_alignment(buff, len); + + if (((unsigned long)buff & OFFSET_MASK) == 0) + return do_csum_no_alignment(buff, len); + + return do_csum_with_alignment(buff, len); +} |