Merge patch series "riscv: Add runtime constant support"

Alexandre Ghiti · Alexandre Ghiti · commit 74f4bf9d15ad · 2025-03-20T09:15:04.000Z
Charlie Jenkins <charlie@rivosinc.com> says: Ard brought this to my attention in this patch [1]. I benchmarked this patch on the Nezha D1 (which does not contain Zba or Zbkb so it uses the default algorithm) by navigating through a large directory structure. I created a 1000-deep directory structure and then cd and ls through it. With this patch there was a 0.57% performance improvement. [1] https://lore.kernel.org/lkml/CAMj1kXE4DJnwFejNWQu784GvyJO=aGNrzuLjSxiowX_e7nW8QA@mail.gmail.com/ * patches from https://lore.kernel.org/r/20250319-runtime_const_riscv-v10-0-745b31a11d65@rivosinc.com: riscv: Add runtime constant support riscv: Move nop definition to insn-def.h Link: https://lore.kernel.org/linux-riscv/20250319-runtime_const_riscv-v10-0-745b31a11d65@rivosinc.com/ Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
@@ -785,6 +785,28 @@ config RISCV_ISA_ZBC
 
 	   If you don't know what to do here, say Y.
 
+config TOOLCHAIN_HAS_ZBKB
+	bool
+	default y
+	depends on !64BIT || $(cc-option,-mabi=lp64 -march=rv64ima_zbkb)
+	depends on !32BIT || $(cc-option,-mabi=ilp32 -march=rv32ima_zbkb)
+	depends on LLD_VERSION >= 150000 || LD_VERSION >= 23900
+	depends on AS_HAS_OPTION_ARCH
+
+config RISCV_ISA_ZBKB
+	bool "Zbkb extension support for bit manipulation instructions"
+	depends on TOOLCHAIN_HAS_ZBKB
+	depends on RISCV_ALTERNATIVE
+	default y
+	help
+	   Adds support to dynamically detect the presence of the ZBKB
+	   extension (bit manipulation for cryptography) and enable its usage.
+
+	   The Zbkb extension provides instructions to accelerate a number
+	   of common cryptography operations (pack, zip, etc).
+
+	   If you don't know what to do here, say Y.
+
 config RISCV_ISA_ZICBOM
 	bool "Zicbom extension support for non-coherent DMA operation"
 	depends on MMU
diff --git a/arch/riscv/include/asm/asm.h b/arch/riscv/include/asm/asm.h
@@ -27,6 +27,7 @@
 #define REG_ASM		__REG_SEL(.dword, .word)
 #define SZREG		__REG_SEL(8, 4)
 #define LGREG		__REG_SEL(3, 2)
+#define SRLI		__REG_SEL(srliw, srli)
 
 #if __SIZEOF_POINTER__ == 8
 #ifdef __ASSEMBLY__
diff --git a/arch/riscv/include/asm/ftrace.h b/arch/riscv/include/asm/ftrace.h
@@ -79,7 +79,6 @@ struct dyn_arch_ftrace {
 #define AUIPC_RA		(0x00000097)
 #define JALR_T0			(0x000282e7)
 #define AUIPC_T0		(0x00000297)
-#define NOP4			(0x00000013)
 
 #define to_jalr_t0(offset)						\
 	(((offset & JALR_OFFSET_MASK) << JALR_SHIFT) | JALR_T0)
diff --git a/arch/riscv/include/asm/insn-def.h b/arch/riscv/include/asm/insn-def.h
@@ -199,5 +199,8 @@
 #define RISCV_PAUSE	".4byte 0x100000f"
 #define ZAWRS_WRS_NTO	".4byte 0x00d00073"
 #define ZAWRS_WRS_STO	".4byte 0x01d00073"
+#define RISCV_NOP4	".4byte 0x00000013"
+
+#define RISCV_INSN_NOP4	_AC(0x00000013, U)
 
 #endif /* __ASM_INSN_DEF_H */
diff --git a/arch/riscv/include/asm/runtime-const.h b/arch/riscv/include/asm/runtime-const.h
@@ -0,0 +1,265 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_RISCV_RUNTIME_CONST_H
+#define _ASM_RISCV_RUNTIME_CONST_H
+
+#include <asm/asm.h>
+#include <asm/alternative.h>
+#include <asm/cacheflush.h>
+#include <asm/insn-def.h>
+#include <linux/memory.h>
+#include <asm/text-patching.h>
+
+#include <linux/uaccess.h>
+
+#ifdef CONFIG_32BIT
+#define runtime_const_ptr(sym)					\
+({								\
+	typeof(sym) __ret;					\
+	asm_inline(".option push\n\t"				\
+		".option norvc\n\t"				\
+		"1:\t"						\
+		"lui	%[__ret],0x89abd\n\t"			\
+		"addi	%[__ret],%[__ret],-0x211\n\t"		\
+		".option pop\n\t"				\
+		".pushsection runtime_ptr_" #sym ",\"a\"\n\t"	\
+		".long 1b - .\n\t"				\
+		".popsection"					\
+		: [__ret] "=r" (__ret));			\
+	__ret;							\
+})
+#else
+/*
+ * Loading 64-bit constants into a register from immediates is a non-trivial
+ * task on riscv64. To get it somewhat performant, load 32 bits into two
+ * different registers and then combine the results.
+ *
+ * If the processor supports the Zbkb extension, we can combine the final
+ * "slli,slli,srli,add" into the single "pack" instruction. If the processor
+ * doesn't support Zbkb but does support the Zbb extension, we can
+ * combine the final "slli,srli,add" into one instruction "add.uw".
+ */
+#define RISCV_RUNTIME_CONST_64_PREAMBLE				\
+	".option push\n\t"					\
+	".option norvc\n\t"					\
+	"1:\t"							\
+	"lui	%[__ret],0x89abd\n\t"				\
+	"lui	%[__tmp],0x1234\n\t"				\
+	"addiw	%[__ret],%[__ret],-0x211\n\t"			\
+	"addiw	%[__tmp],%[__tmp],0x567\n\t"			\
+
+#define RISCV_RUNTIME_CONST_64_BASE				\
+	"slli	%[__tmp],%[__tmp],32\n\t"			\
+	"slli	%[__ret],%[__ret],32\n\t"			\
+	"srli	%[__ret],%[__ret],32\n\t"			\
+	"add	%[__ret],%[__ret],%[__tmp]\n\t"			\
+
+#define RISCV_RUNTIME_CONST_64_ZBA				\
+	".option push\n\t"					\
+	".option arch,+zba\n\t"					\
+	"slli	%[__tmp],%[__tmp],32\n\t"			\
+	"add.uw %[__ret],%[__ret],%[__tmp]\n\t"			\
+	"nop\n\t"						\
+	"nop\n\t"						\
+	".option pop\n\t"					\
+
+#define RISCV_RUNTIME_CONST_64_ZBKB				\
+	".option push\n\t"					\
+	".option arch,+zbkb\n\t"				\
+	"pack	%[__ret],%[__ret],%[__tmp]\n\t"			\
+	"nop\n\t"						\
+	"nop\n\t"						\
+	"nop\n\t"						\
+	".option pop\n\t"					\
+
+#define RISCV_RUNTIME_CONST_64_POSTAMBLE(sym)			\
+	".option pop\n\t"					\
+	".pushsection runtime_ptr_" #sym ",\"a\"\n\t"		\
+	".long 1b - .\n\t"					\
+	".popsection"						\
+
+#if defined(CONFIG_RISCV_ISA_ZBA) && defined(CONFIG_RISCV_ISA_ZBKB)
+#define runtime_const_ptr(sym)						\
+({									\
+	typeof(sym) __ret, __tmp;					\
+	asm_inline(RISCV_RUNTIME_CONST_64_PREAMBLE			\
+		ALTERNATIVE_2(						\
+			RISCV_RUNTIME_CONST_64_BASE,			\
+			RISCV_RUNTIME_CONST_64_ZBA,			\
+			0, RISCV_ISA_EXT_ZBA, 1,			\
+			RISCV_RUNTIME_CONST_64_ZBKB,			\
+			0, RISCV_ISA_EXT_ZBKB, 1			\
+		)							\
+		RISCV_RUNTIME_CONST_64_POSTAMBLE(sym)			\
+		: [__ret] "=r" (__ret), [__tmp] "=r" (__tmp));		\
+	__ret;								\
+})
+#elif defined(CONFIG_RISCV_ISA_ZBA)
+#define runtime_const_ptr(sym)						\
+({									\
+	typeof(sym) __ret, __tmp;					\
+	asm_inline(RISCV_RUNTIME_CONST_64_PREAMBLE			\
+		ALTERNATIVE(						\
+			RISCV_RUNTIME_CONST_64_BASE,			\
+			RISCV_RUNTIME_CONST_64_ZBA,			\
+			0, RISCV_ISA_EXT_ZBA, 1				\
+		)							\
+		RISCV_RUNTIME_CONST_64_POSTAMBLE(sym)			\
+		: [__ret] "=r" (__ret), [__tmp] "=r" (__tmp));		\
+	__ret;								\
+})
+#elif defined(CONFIG_RISCV_ISA_ZBKB)
+#define runtime_const_ptr(sym)						\
+({									\
+	typeof(sym) __ret, __tmp;					\
+	asm_inline(RISCV_RUNTIME_CONST_64_PREAMBLE			\
+		ALTERNATIVE(						\
+			RISCV_RUNTIME_CONST_64_BASE,			\
+			RISCV_RUNTIME_CONST_64_ZBKB,			\
+			0, RISCV_ISA_EXT_ZBKB, 1			\
+		)							\
+		RISCV_RUNTIME_CONST_64_POSTAMBLE(sym)			\
+		: [__ret] "=r" (__ret), [__tmp] "=r" (__tmp));		\
+	__ret;								\
+})
+#else
+#define runtime_const_ptr(sym)						\
+({									\
+	typeof(sym) __ret, __tmp;					\
+	asm_inline(RISCV_RUNTIME_CONST_64_PREAMBLE			\
+		RISCV_RUNTIME_CONST_64_BASE				\
+		RISCV_RUNTIME_CONST_64_POSTAMBLE(sym)			\
+		: [__ret] "=r" (__ret), [__tmp] "=r" (__tmp));		\
+	__ret;								\
+})
+#endif
+#endif
+
+#define runtime_const_shift_right_32(val, sym)			\
+({								\
+	u32 __ret;						\
+	asm_inline(".option push\n\t"				\
+		".option norvc\n\t"				\
+		"1:\t"						\
+		SRLI " %[__ret],%[__val],12\n\t"		\
+		".option pop\n\t"				\
+		".pushsection runtime_shift_" #sym ",\"a\"\n\t"	\
+		".long 1b - .\n\t"				\
+		".popsection"					\
+		: [__ret] "=r" (__ret)				\
+		: [__val] "r" (val));				\
+	__ret;							\
+})
+
+#define runtime_const_init(type, sym) do {			\
+	extern s32 __start_runtime_##type##_##sym[];		\
+	extern s32 __stop_runtime_##type##_##sym[];		\
+								\
+	runtime_const_fixup(__runtime_fixup_##type,		\
+			    (unsigned long)(sym),		\
+			    __start_runtime_##type##_##sym,	\
+			    __stop_runtime_##type##_##sym);	\
+} while (0)
+
+static inline void __runtime_fixup_caches(void *where, unsigned int insns)
+{
+	/* On riscv there are currently only cache-wide flushes so va is ignored. */
+	__always_unused uintptr_t va = (uintptr_t)where;
+
+	flush_icache_range(va, va + 4 * insns);
+}
+
+/*
+ * The 32-bit immediate is stored in a lui+addi pairing.
+ * lui holds the upper 20 bits of the immediate in the first 20 bits of the instruction.
+ * addi holds the lower 12 bits of the immediate in the first 12 bits of the instruction.
+ */
+static inline void __runtime_fixup_32(__le16 *lui_parcel, __le16 *addi_parcel, unsigned int val)
+{
+	unsigned int lower_immediate, upper_immediate;
+	u32 lui_insn, addi_insn, addi_insn_mask;
+	__le32 lui_res, addi_res;
+
+	/* Mask out upper 12 bit of addi */
+	addi_insn_mask = 0x000fffff;
+
+	lui_insn = (u32)le16_to_cpu(lui_parcel[0]) | (u32)le16_to_cpu(lui_parcel[1]) << 16;
+	addi_insn = (u32)le16_to_cpu(addi_parcel[0]) | (u32)le16_to_cpu(addi_parcel[1]) << 16;
+
+	lower_immediate = sign_extend32(val, 11);
+	upper_immediate = (val - lower_immediate);
+
+	if (upper_immediate & 0xfffff000) {
+		/* replace upper 20 bits of lui with upper immediate */
+		lui_insn &= 0x00000fff;
+		lui_insn |= upper_immediate & 0xfffff000;
+	} else {
+		/* replace lui with nop if immediate is small enough to fit in addi */
+		lui_insn = RISCV_INSN_NOP4;
+		/*
+		 * lui is being skipped, so do a load instead of an add. A load
+		 * is performed by adding with the x0 register. Setting rs to
+		 * zero with the following mask will accomplish this goal.
+		 */
+		addi_insn_mask &= 0x07fff;
+	}
+
+	if (lower_immediate & 0x00000fff) {
+		/* replace upper 12 bits of addi with lower 12 bits of val */
+		addi_insn &= addi_insn_mask;
+		addi_insn |= (lower_immediate & 0x00000fff) << 20;
+	} else {
+		/* replace addi with nop if lower_immediate is empty */
+		addi_insn = RISCV_INSN_NOP4;
+	}
+
+	addi_res = cpu_to_le32(addi_insn);
+	lui_res = cpu_to_le32(lui_insn);
+	mutex_lock(&text_mutex);
+	patch_insn_write(addi_parcel, &addi_res, sizeof(addi_res));
+	patch_insn_write(lui_parcel, &lui_res, sizeof(lui_res));
+	mutex_unlock(&text_mutex);
+}
+
+static inline void __runtime_fixup_ptr(void *where, unsigned long val)
+{
+#ifdef CONFIG_32BIT
+		__runtime_fixup_32(where, where + 4, val);
+		__runtime_fixup_caches(where, 2);
+#else
+		__runtime_fixup_32(where, where + 8, val);
+		__runtime_fixup_32(where + 4, where + 12, val >> 32);
+		__runtime_fixup_caches(where, 4);
+#endif
+}
+
+/*
+ * Replace the least significant 5 bits of the srli/srliw immediate that is
+ * located at bits 20-24
+ */
+static inline void __runtime_fixup_shift(void *where, unsigned long val)
+{
+	__le16 *parcel = where;
+	__le32 res;
+	u32 insn;
+
+	insn = (u32)le16_to_cpu(parcel[0]) | (u32)le16_to_cpu(parcel[1]) << 16;
+
+	insn &= 0xfe0fffff;
+	insn |= (val & 0b11111) << 20;
+
+	res = cpu_to_le32(insn);
+	mutex_lock(&text_mutex);
+	patch_text_nosync(where, &res, sizeof(insn));
+	mutex_unlock(&text_mutex);
+}
+
+static inline void runtime_const_fixup(void (*fn)(void *, unsigned long),
+				       unsigned long val, s32 *start, s32 *end)
+{
+	while (start < end) {
+		fn(*start + (void *)start, val);
+		start++;
+	}
+}
+
+#endif /* _ASM_RISCV_RUNTIME_CONST_H */
diff --git a/arch/riscv/kernel/ftrace.c b/arch/riscv/kernel/ftrace.c
@@ -36,7 +36,7 @@ static int ftrace_check_current_call(unsigned long hook_pos,
 				     unsigned int *expected)
 {
 	unsigned int replaced[2];
-	unsigned int nops[2] = {NOP4, NOP4};
+	unsigned int nops[2] = {RISCV_INSN_NOP4, RISCV_INSN_NOP4};
 
 	/* we expect nops at the hook position */
 	if (!expected)
@@ -68,7 +68,7 @@ static int __ftrace_modify_call(unsigned long hook_pos, unsigned long target,
 				bool enable, bool ra)
 {
 	unsigned int call[2];
-	unsigned int nops[2] = {NOP4, NOP4};
+	unsigned int nops[2] = {RISCV_INSN_NOP4, RISCV_INSN_NOP4};
 
 	if (ra)
 		make_call_ra(hook_pos, target, call);
@@ -97,7 +97,7 @@ int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
 		    unsigned long addr)
 {
-	unsigned int nops[2] = {NOP4, NOP4};
+	unsigned int nops[2] = {RISCV_INSN_NOP4, RISCV_INSN_NOP4};
 
 	if (patch_insn_write((void *)rec->ip, nops, MCOUNT_INSN_SIZE))
 		return -EPERM;
diff --git a/arch/riscv/kernel/jump_label.c b/arch/riscv/kernel/jump_label.c
@@ -11,8 +11,8 @@
 #include <asm/bug.h>
 #include <asm/cacheflush.h>
 #include <asm/text-patching.h>
+#include <asm/insn-def.h>
 
-#define RISCV_INSN_NOP 0x00000013U
 #define RISCV_INSN_JAL 0x0000006fU
 
 bool arch_jump_label_transform_queue(struct jump_entry *entry,
@@ -33,7 +33,7 @@ bool arch_jump_label_transform_queue(struct jump_entry *entry,
 			(((u32)offset & GENMASK(10,  1)) << (21 -  1)) |
 			(((u32)offset & GENMASK(20, 20)) << (31 - 20));
 	} else {
-		insn = RISCV_INSN_NOP;
+		insn = RISCV_INSN_NOP4;
 	}
 
 	if (early_boot_irqs_disabled) {
diff --git a/arch/riscv/kernel/vmlinux.lds.S b/arch/riscv/kernel/vmlinux.lds.S
@@ -97,6 +97,9 @@ SECTIONS
 	{
 		EXIT_DATA
 	}
+
+	RUNTIME_CONST_VARIABLES
+
 	PERCPU_SECTION(L1_CACHE_BYTES)
 
 	.rel.dyn : {

Original file line number	Diff line number	Diff line change
`@@ -97,6 +97,9 @@ SECTIONS`
`97`	`97`	`{`
`98`	`98`	`EXIT_DATA`
`99`	`99`	`}`
	`100`	`+`
	`101`	`+ RUNTIME_CONST_VARIABLES`
	`102`	`+`
`100`	`103`	`PERCPU_SECTION(L1_CACHE_BYTES)`
`101`	`104`
`102`	`105`	`.rel.dyn : {`