diff mbox series

[v3] riscv: Add runtime constant support

Message ID 20250128-runtime_const_riscv-v3-1-11922989e2d3@rivosinc.com (mailing list archive)
State New
Headers show
Series [v3] riscv: Add runtime constant support | expand

Checks

Context Check Description
bjorn/pre-ci_am success Success
bjorn/build-rv32-defconfig fail build-rv32-defconfig
bjorn/build-rv64-clang-allmodconfig success build-rv64-clang-allmodconfig
bjorn/build-rv64-gcc-allmodconfig success build-rv64-gcc-allmodconfig
bjorn/build-rv64-nommu-k210-defconfig success build-rv64-nommu-k210-defconfig
bjorn/build-rv64-nommu-k210-virt success build-rv64-nommu-k210-virt
bjorn/checkpatch warning checkpatch
bjorn/dtb-warn-rv64 success dtb-warn-rv64
bjorn/header-inline success header-inline
bjorn/kdoc success kdoc
bjorn/module-param success module-param
bjorn/verify-fixes success verify-fixes
bjorn/verify-signedoff success verify-signedoff

Commit Message

Charlie Jenkins Jan. 28, 2025, 8:06 p.m. UTC 
Implement the runtime constant infrastructure for riscv. Use this
infrastructure to generate constants to be used by the d_hash()
function.

This is the riscv variant of commit 94a2bc0f611c ("arm64: add 'runtime
constant' support") and commit e3c92e81711d ("runtime constants: add
x86 architecture support").

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
Ard brought this to my attention in this patch [1].

[1] https://lore.kernel.org/lkml/CAMj1kXE4DJnwFejNWQu784GvyJO=aGNrzuLjSxiowX_e7nW8QA@mail.gmail.com/
---
Changes in v3:
- Leverage "pack" instruction for runtime_const_ptr() to reduce hot path
  by 3 instructions if Zbkb is supported. Suggested by Pasha Bouzarjomehri (pasha@rivosinc.com)
- Link to v2: https://lore.kernel.org/r/20250127-runtime_const_riscv-v2-1-95ae7cf97a39@rivosinc.com

Changes in v2:
- Treat instructions as __le32 and do proper conversions (Ben)
- Link to v1: https://lore.kernel.org/r/20250127-runtime_const_riscv-v1-1-795b023ea20b@rivosinc.com
---
 arch/riscv/include/asm/runtime-const.h | 194 +++++++++++++++++++++++++++++++++
 arch/riscv/kernel/vmlinux.lds.S        |   3 +
 2 files changed, 197 insertions(+)


---
base-commit: ffd294d346d185b70e28b1a28abe367bbfe53c04
change-id: 20250123-runtime_const_riscv-6cd854ee2817

Comments

Emil Renner Berthing Jan. 29, 2025, 9:30 a.m. UTC | #1 
Charlie Jenkins wrote:
> Implement the runtime constant infrastructure for riscv. Use this
> infrastructure to generate constants to be used by the d_hash()
> function.
>
> This is the riscv variant of commit 94a2bc0f611c ("arm64: add 'runtime
> constant' support") and commit e3c92e81711d ("runtime constants: add
> x86 architecture support").
>
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
> ---
> Ard brought this to my attention in this patch [1].
>
> [1] https://lore.kernel.org/lkml/CAMj1kXE4DJnwFejNWQu784GvyJO=aGNrzuLjSxiowX_e7nW8QA@mail.gmail.com/
> ---
> Changes in v3:
> - Leverage "pack" instruction for runtime_const_ptr() to reduce hot path
>   by 3 instructions if Zbkb is supported. Suggested by Pasha Bouzarjomehri (pasha@rivosinc.com)
> - Link to v2: https://lore.kernel.org/r/20250127-runtime_const_riscv-v2-1-95ae7cf97a39@rivosinc.com
>
> Changes in v2:
> - Treat instructions as __le32 and do proper conversions (Ben)
> - Link to v1: https://lore.kernel.org/r/20250127-runtime_const_riscv-v1-1-795b023ea20b@rivosinc.com
> ---
>  arch/riscv/include/asm/runtime-const.h | 194 +++++++++++++++++++++++++++++++++
>  arch/riscv/kernel/vmlinux.lds.S        |   3 +
>  2 files changed, 197 insertions(+)
>
> diff --git a/arch/riscv/include/asm/runtime-const.h b/arch/riscv/include/asm/runtime-const.h
> new file mode 100644
> index 0000000000000000000000000000000000000000..0ecbe6967013900781b0b1048d4622f676b64076
> --- /dev/null
> +++ b/arch/riscv/include/asm/runtime-const.h
> @@ -0,0 +1,194 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ASM_RISCV_RUNTIME_CONST_H
> +#define _ASM_RISCV_RUNTIME_CONST_H
> +
> +#include <asm/alternative.h>
> +#include <asm/cacheflush.h>
> +#include <asm/text-patching.h>
> +#include <linux/uaccess.h>
> +
> +#ifdef CONFIG_32BIT
> +#define runtime_const_ptr(sym)					\
> +({								\
> +	typeof(sym) __ret, __tmp;				\
> +	asm_inline("1:\t"					\
> +		".option push"					\
> +		".option norvc"					\
> +		"lui	%[__ret],0x89abd\n\t"			\
> +		"addi	%[__ret],-0x211\n\t"			\
> +		".option pop"					\
> +		".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 runtime_const_ptr(sym)						\
> +({									\
> +	typeof(sym) __ret, __tmp;					\
> +	asm_inline("1:\t"						\
> +		".option push\n\t"					\
> +		".option norvc\n\t"					\
> +		"lui	%[__ret],0x89abd\n\t"				\
> +		"lui	%[__tmp],0x1234\n\t"				\
> +		"addiw	%[__ret],%[__ret],-0x211\n\t"			\
> +		"addiw	%[__tmp],%[__tmp],0x567\n\t"			\
> +		ALTERNATIVE_2(						\
> +			"slli	%[__tmp],%[__tmp],32\n\t"		\
> +			"slli	%[__ret],%[__ret],32\n\t"		\
> +			"srli	%[__ret],%[__ret],32\n\t"		\
> +			"add	%[__ret],%[__ret],%[__tmp]\n\t",	\
> +			".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",				\
> +			0, RISCV_ISA_EXT_ZBA, 1,			\
> +			".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",				\
> +			0, RISCV_ISA_EXT_ZBKB, 1			\
> +		)							\
> +		".option pop\n\t"					\
> +		".pushsection runtime_ptr_" #sym ",\"a\"\n\t"		\
> +		".long 1b - .\n\t"					\
> +		".popsection"						\
> +		: [__ret] "=r" (__ret), [__tmp] "=r" (__tmp));		\
> +	__ret;								\
> +})
> +#endif
> +
> +#ifdef CONFIG_32BIT
> +#define SRLI "srli "
> +#else
> +#define SRLI "srliw "
> +#endif
> +
> +#define runtime_const_shift_right_32(val, sym)			\
> +({								\
> +	u32 __ret;						\
> +	asm_inline("1:\t"					\
> +		".option push\n\t"				\
> +		".option norvc\n\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(u32 *lui, u32 *addi, unsigned int val)
> +{
> +	unsigned int lower_immediate, upper_immediate;
> +	u32 lui_insn = le32_to_cpu(*lui);
> +	u32 addi_insn = le32_to_cpu(*addi);

Because of the compressed extensions RISC-V instructions are only aligned on
16bit boundaries, so is there another reason you know that these two
instructions are 32bit aligned? Otherwise you're adding unaligned accesses
here.

/Emil

> +	__le32 addi_res, lui_res;
> +
> +	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 = 0x00000013;
> +	}
> +
> +	if (lower_immediate & 0x00000fff) {
> +		/* replace upper 12 bits of addi with lower 12 bits of val */
> +		addi_insn &= 0x000fffff;
> +		addi_insn |= (lower_immediate & 0x00000fff) << 20;
> +	} else {
> +		/* replace addi with nop if lower_immediate is empty */
> +		addi_insn = 0x00000013;
> +	}
> +
> +	addi_res = cpu_to_le32(addi_insn);
> +	lui_res = cpu_to_le32(lui_insn);
> +	patch_insn_write(addi, &addi_res, sizeof(addi_res));
> +	patch_insn_write(lui, &lui_res, sizeof(lui_res));
> +}
> +
> +static inline void __runtime_fixup_ptr(void *where, unsigned long val)
> +{
> +	if (IS_ENABLED(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);
> +	}
> +}
> +
> +/*
> + * 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)
> +{
> +	u32 insn = le32_to_cpu(*(__le32 *)where);
> +	__le32 res;
> +
> +	insn &= 0xfe0fffff;
> +	insn |= (val & 0b11111) << 20;
> +
> +	res = cpu_to_le32(insn);
> +	patch_text_nosync(where, &res, sizeof(insn));
> +}
> +
> +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/vmlinux.lds.S b/arch/riscv/kernel/vmlinux.lds.S
> index 002ca58dd998cb78b662837b5ebac988fb6c77bb..61bd5ba6680a786bf1db7dc37bf1acda0639b5c7 100644
> --- 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 : {
>
> ---
> base-commit: ffd294d346d185b70e28b1a28abe367bbfe53c04
> change-id: 20250123-runtime_const_riscv-6cd854ee2817
> --
> - Charlie
>
>
> _______________________________________________
> linux-riscv mailing list
> linux-riscv@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-riscv
diff mbox series

Patch

diff --git a/arch/riscv/include/asm/runtime-const.h b/arch/riscv/include/asm/runtime-const.h
new file mode 100644
index 0000000000000000000000000000000000000000..0ecbe6967013900781b0b1048d4622f676b64076
--- /dev/null
+++ b/arch/riscv/include/asm/runtime-const.h
@@ -0,0 +1,194 @@ 
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_RISCV_RUNTIME_CONST_H
+#define _ASM_RISCV_RUNTIME_CONST_H
+
+#include <asm/alternative.h>
+#include <asm/cacheflush.h>
+#include <asm/text-patching.h>
+#include <linux/uaccess.h>
+
+#ifdef CONFIG_32BIT
+#define runtime_const_ptr(sym)					\
+({								\
+	typeof(sym) __ret, __tmp;				\
+	asm_inline("1:\t"					\
+		".option push"					\
+		".option norvc"					\
+		"lui	%[__ret],0x89abd\n\t"			\
+		"addi	%[__ret],-0x211\n\t"			\
+		".option pop"					\
+		".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 runtime_const_ptr(sym)						\
+({									\
+	typeof(sym) __ret, __tmp;					\
+	asm_inline("1:\t"						\
+		".option push\n\t"					\
+		".option norvc\n\t"					\
+		"lui	%[__ret],0x89abd\n\t"				\
+		"lui	%[__tmp],0x1234\n\t"				\
+		"addiw	%[__ret],%[__ret],-0x211\n\t"			\
+		"addiw	%[__tmp],%[__tmp],0x567\n\t"			\
+		ALTERNATIVE_2(						\
+			"slli	%[__tmp],%[__tmp],32\n\t"		\
+			"slli	%[__ret],%[__ret],32\n\t"		\
+			"srli	%[__ret],%[__ret],32\n\t"		\
+			"add	%[__ret],%[__ret],%[__tmp]\n\t",	\
+			".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",				\
+			0, RISCV_ISA_EXT_ZBA, 1,			\
+			".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",				\
+			0, RISCV_ISA_EXT_ZBKB, 1			\
+		)							\
+		".option pop\n\t"					\
+		".pushsection runtime_ptr_" #sym ",\"a\"\n\t"		\
+		".long 1b - .\n\t"					\
+		".popsection"						\
+		: [__ret] "=r" (__ret), [__tmp] "=r" (__tmp));		\
+	__ret;								\
+})
+#endif
+
+#ifdef CONFIG_32BIT
+#define SRLI "srli "
+#else
+#define SRLI "srliw "
+#endif
+
+#define runtime_const_shift_right_32(val, sym)			\
+({								\
+	u32 __ret;						\
+	asm_inline("1:\t"					\
+		".option push\n\t"				\
+		".option norvc\n\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(u32 *lui, u32 *addi, unsigned int val)
+{
+	unsigned int lower_immediate, upper_immediate;
+	u32 lui_insn = le32_to_cpu(*lui);
+	u32 addi_insn = le32_to_cpu(*addi);
+	__le32 addi_res, lui_res;
+
+	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 = 0x00000013;
+	}
+
+	if (lower_immediate & 0x00000fff) {
+		/* replace upper 12 bits of addi with lower 12 bits of val */
+		addi_insn &= 0x000fffff;
+		addi_insn |= (lower_immediate & 0x00000fff) << 20;
+	} else {
+		/* replace addi with nop if lower_immediate is empty */
+		addi_insn = 0x00000013;
+	}
+
+	addi_res = cpu_to_le32(addi_insn);
+	lui_res = cpu_to_le32(lui_insn);
+	patch_insn_write(addi, &addi_res, sizeof(addi_res));
+	patch_insn_write(lui, &lui_res, sizeof(lui_res));
+}
+
+static inline void __runtime_fixup_ptr(void *where, unsigned long val)
+{
+	if (IS_ENABLED(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);
+	}
+}
+
+/*
+ * 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)
+{
+	u32 insn = le32_to_cpu(*(__le32 *)where);
+	__le32 res;
+
+	insn &= 0xfe0fffff;
+	insn |= (val & 0b11111) << 20;
+
+	res = cpu_to_le32(insn);
+	patch_text_nosync(where, &res, sizeof(insn));
+}
+
+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/vmlinux.lds.S b/arch/riscv/kernel/vmlinux.lds.S
index 002ca58dd998cb78b662837b5ebac988fb6c77bb..61bd5ba6680a786bf1db7dc37bf1acda0639b5c7 100644
--- 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 : {