Merge patch series "riscv: Add fine-tuned checksum functions"

Charlie Jenkins <charlie@rivosinc.com> says:

Each architecture generally implements fine-tuned checksum functions to
leverage the instruction set. This patch adds the main checksum
functions that are used in networking. Tested on QEMU, this series
allows the CHECKSUM_KUNIT tests to complete an average of 50.9% faster.

This patch takes heavy use of the Zbb extension using alternatives
patching.

To test this patch, enable the configs for KUNIT, then CHECKSUM_KUNIT.

I have attempted to make these functions as optimal as possible, but I
have not ran anything on actual riscv hardware. My performance testing
has been limited to inspecting the assembly, running the algorithms on
x86 hardware, and running in QEMU.

ip_fast_csum is a relatively small function so even though it is
possible to read 64 bits at a time on compatible hardware, the
bottleneck becomes the clean up and setup code so loading 32 bits at a
time is actually faster.

* b4-shazam-merge:
  kunit: Add tests for csum_ipv6_magic and ip_fast_csum
  riscv: Add checksum library
  riscv: Add checksum header
  riscv: Add static key for misaligned accesses
  asm-generic: Improve csum_fold

Link: https://lore.kernel.org/r/20240108-optimize_checksum-v15-0-1c50de5f2167@rivosinc.comSigned-off-by: Palmer Dabbelt <palmer@rivosinc.com>

arch/riscv/include/asm/checksum.h

+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Checksum routines
+ *
+ * Copyright (C) 2023 Rivos Inc.
+ */
+#ifndef __ASM_RISCV_CHECKSUM_H
+#define __ASM_RISCV_CHECKSUM_H
+
+#include <linux/in6.h>
+#include <linux/uaccess.h>
+
+#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>
+
+/**
+ * Quickly compute an IP checksum with the assumption that IPv4 headers will
+ * always be in multiples of 32-bits, and have an ihl of at least 5.
+ *
+ * @ihl: the number of 32 bit segments and must be greater than or equal to 5.
+ * @iph: assumed to be word aligned given that NET_IP_ALIGN is set to 2 on
+ *  riscv, defining IP headers to be aligned.
+ */
+static inline __sum16 ip_fast_csum(const void *iph, unsigned int ihl)
+{
+	unsigned long csum = 0;
+	int pos = 0;
+
+	do {
+		csum += ((const unsigned int *)iph)[pos];
+		if (IS_ENABLED(CONFIG_32BIT))
+			csum += csum < ((const unsigned int *)iph)[pos];
+	} while (++pos < ihl);
+
+	/*
+	 * ZBB only saves three instructions on 32-bit and five on 64-bit so not
+	 * worth checking if supported without Alternatives.
+	 */
+	if (IS_ENABLED(CONFIG_RISCV_ISA_ZBB) &&
+	    IS_ENABLED(CONFIG_RISCV_ALTERNATIVE)) {
+		unsigned long fold_temp;
+
+		asm_volatile_goto(ALTERNATIVE("j %l[no_zbb]", "nop", 0,
+					      RISCV_ISA_EXT_ZBB, 1)
+		    :
+		    :
+		    :
+		    : no_zbb);
+
+		if (IS_ENABLED(CONFIG_32BIT)) {
+			asm(".option push				\n\
+			.option arch,+zbb				\n\
+				not	%[fold_temp], %[csum]		\n\
+				rori	%[csum], %[csum], 16		\n\
+				sub	%[csum], %[fold_temp], %[csum]	\n\
+			.option pop"
+			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp));
+		} else {
+			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\
+				not	%[fold_temp], %[csum]		\n\
+				roriw	%[csum], %[csum], 16		\n\
+				subw	%[csum], %[fold_temp], %[csum]	\n\
+			.option pop"
+			: [csum] "+r" (csum), [fold_temp] "=&r" (fold_temp));
+		}
+		return (__force __sum16)(csum >> 16);
+	}
+no_zbb:
+#ifndef CONFIG_32BIT
+	csum += ror64(csum, 32);
+	csum >>= 32;
+#endif
+	return csum_fold((__force __wsum)csum);
+}
+
+#endif /* __ASM_RISCV_CHECKSUM_H */

arch/riscv/include/asm/cpufeature.h

@@ -135,4 +135,6 @@ static __always_inline bool riscv_cpu_has_extension_unlikely(int cpu, const unsi
 	return __riscv_isa_extension_available(hart_isa[cpu].isa, ext);
 }
 
+DECLARE_STATIC_KEY_FALSE(fast_misaligned_access_speed_key);
+
 #endif

arch/riscv/kernel/cpufeature.c

@@ -8,8 +8,10 @@
 
 #include <linux/acpi.h>
 #include <linux/bitmap.h>
+#include <linux/cpu.h>
 #include <linux/cpuhotplug.h>
 #include <linux/ctype.h>
+#include <linux/jump_label.h>
 #include <linux/log2.h>
 #include <linux/memory.h>
 #include <linux/module.h>
@@ -44,6 +46,8 @@ struct riscv_isainfo hart_isa[NR_CPUS];
 /* Performance information */
 DEFINE_PER_CPU(long, misaligned_access_speed);
 
+static cpumask_t fast_misaligned_access;
+
 /**
  * riscv_isa_extension_base() - Get base extension word
  *
@@ -784,6 +788,16 @@ static int check_unaligned_access(void *param)
 		(speed == RISCV_HWPROBE_MISALIGNED_FAST) ? "fast" : "slow");
 
 	per_cpu(misaligned_access_speed, cpu) = speed;
+
+	/*
+	 * Set the value of fast_misaligned_access of a CPU. These operations
+	 * are atomic to avoid race conditions.
+	 */
+	if (speed == RISCV_HWPROBE_MISALIGNED_FAST)
+		cpumask_set_cpu(cpu, &fast_misaligned_access);
+	else
+		cpumask_clear_cpu(cpu, &fast_misaligned_access);
+
 	return 0;
 }
 
@@ -796,13 +810,69 @@ static void check_unaligned_access_nonboot_cpu(void *param)
 		check_unaligned_access(pages[cpu]);
 }
 
+DEFINE_STATIC_KEY_FALSE(fast_misaligned_access_speed_key);
+
+static void modify_unaligned_access_branches(cpumask_t *mask, int weight)
+{
+	if (cpumask_weight(mask) == weight)
+		static_branch_enable_cpuslocked(&fast_misaligned_access_speed_key);
+	else
+		static_branch_disable_cpuslocked(&fast_misaligned_access_speed_key);
+}
+
+static void set_unaligned_access_static_branches_except_cpu(int cpu)
+{
+	/*
+	 * Same as set_unaligned_access_static_branches, except excludes the
+	 * given CPU from the result. When a CPU is hotplugged into an offline
+	 * state, this function is called before the CPU is set to offline in
+	 * the cpumask, and thus the CPU needs to be explicitly excluded.
+	 */
+
+	cpumask_t fast_except_me;
+
+	cpumask_and(&fast_except_me, &fast_misaligned_access, cpu_online_mask);
+	cpumask_clear_cpu(cpu, &fast_except_me);
+
+	modify_unaligned_access_branches(&fast_except_me, num_online_cpus() - 1);
+}
+
+static void set_unaligned_access_static_branches(void)
+{
+	/*
+	 * This will be called after check_unaligned_access_all_cpus so the
+	 * result of unaligned access speed for all CPUs will be available.
+	 *
+	 * To avoid the number of online cpus changing between reading
+	 * cpu_online_mask and calling num_online_cpus, cpus_read_lock must be
+	 * held before calling this function.
+	 */
+
+	cpumask_t fast_and_online;
+
+	cpumask_and(&fast_and_online, &fast_misaligned_access, cpu_online_mask);
+
+	modify_unaligned_access_branches(&fast_and_online, num_online_cpus());
+}
+
+static int lock_and_set_unaligned_access_static_branch(void)
+{
+	cpus_read_lock();
+	set_unaligned_access_static_branches();
+	cpus_read_unlock();
+
+	return 0;
+}
+
+arch_initcall_sync(lock_and_set_unaligned_access_static_branch);
+
 static int riscv_online_cpu(unsigned int cpu)
 {
 	static struct page *buf;
 
 	/* We are already set since the last check */
 	if (per_cpu(misaligned_access_speed, cpu) != RISCV_HWPROBE_MISALIGNED_UNKNOWN)
-		return 0;
+		goto exit;
 
 	buf = alloc_pages(GFP_KERNEL, MISALIGNED_BUFFER_ORDER);
 	if (!buf) {
@@ -812,6 +882,17 @@ static int riscv_online_cpu(unsigned int cpu)
 
 	check_unaligned_access(buf);
 	__free_pages(buf, MISALIGNED_BUFFER_ORDER);
+
+exit:
+	set_unaligned_access_static_branches();
+
+	return 0;
+}
+
+static int riscv_offline_cpu(unsigned int cpu)
+{
+	set_unaligned_access_static_branches_except_cpu(cpu);
+
 	return 0;
 }
 
@@ -846,9 +927,12 @@ static int check_unaligned_access_all_cpus(void)
 	/* Check core 0. */
 	smp_call_on_cpu(0, check_unaligned_access, bufs[0], true);
 
-	/* Setup hotplug callback for any new CPUs that come online. */
+	/*
+	 * Setup hotplug callbacks for any new CPUs that come online or go
+	 * offline.
+	 */
 	cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "riscv:online",
-				  riscv_online_cpu, NULL);
+				  riscv_online_cpu, riscv_offline_cpu);
 
 out:
 	unaligned_emulation_finish();

arch/riscv/lib/Makefile

@@ -6,10 +6,11 @@ lib-y			+= memmove.o
 lib-y			+= strcmp.o
 lib-y			+= strlen.o
 lib-y			+= strncmp.o
+lib-y			+= csum.o
 ifeq ($(CONFIG_MMU), y)
-lib-y				+= uaccess.o
 lib-$(CONFIG_RISCV_ISA_V)	+= uaccess_vector.o
 endif
+lib-$(CONFIG_MMU)	+= uaccess.o
 lib-$(CONFIG_64BIT)	+= tishift.o
 lib-$(CONFIG_RISCV_ISA_ZICBOZ)	+= clear_page.o
 

arch/riscv/lib/csum.c

+// 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);
+}

include/asm-generic/checksum.h

@@ -2,6 +2,8 @@
 #ifndef __ASM_GENERIC_CHECKSUM_H
 #define __ASM_GENERIC_CHECKSUM_H
 
+#include <linux/bitops.h>
+
 /*
  * computes the checksum of a memory block at buff, length len,
  * and adds in "sum" (32-bit)
@@ -31,9 +33,7 @@ extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
 static inline __sum16 csum_fold(__wsum csum)
 {
 	u32 sum = (__force u32)csum;
-	sum = (sum & 0xffff) + (sum >> 16);
-	sum = (sum & 0xffff) + (sum >> 16);
-	return (__force __sum16)~sum;
+	return (__force __sum16)((~sum - ror32(sum, 16)) >> 16);
 }
 #endif