[RFC,09/12] riscv: Add support for kernel-mode FPU
Commit Message
This is motivated by the amdgpu DRM driver, which needs floating-point
code to support recent hardware. That code is not performance-critical,
so only provide a minimal non-preemptible implementation for now.
Use a similar trick as ARM to force placing floating-point code in a
separate translation unit, so it is not possible for compiler-generated
floating-point code to appear outside kernel_fpu_{begin,end}().
Signed-off-by: Samuel Holland <samuel.holland@sifive.com>
---
arch/riscv/Kconfig | 1 +
arch/riscv/Makefile | 3 +++
arch/riscv/include/asm/fpu.h | 26 ++++++++++++++++++++++++++
arch/riscv/kernel/Makefile | 1 +
arch/riscv/kernel/kernel_mode_fpu.c | 28 ++++++++++++++++++++++++++++
5 files changed, 59 insertions(+)
create mode 100644 arch/riscv/include/asm/fpu.h
create mode 100644 arch/riscv/kernel/kernel_mode_fpu.c
Comments
> +#ifdef __riscv_f
> +
> +#define kernel_fpu_begin() \
> + static_assert(false, "floating-point code must use a separate translation unit")
> +#define kernel_fpu_end() kernel_fpu_begin()
> +
> +#else
> +
> +void kernel_fpu_begin(void);
> +void kernel_fpu_end(void);
> +
> +#endif
I'll assume this is related to trick that places code in a separate
translation unit, but I fail to understand it. Can you add a comment
explaining it?
On 2023-12-11 10:11 AM, Christoph Hellwig wrote:
>> +#ifdef __riscv_f
>> +
>> +#define kernel_fpu_begin() \
>> + static_assert(false, "floating-point code must use a separate translation unit")
>> +#define kernel_fpu_end() kernel_fpu_begin()
>> +
>> +#else
>> +
>> +void kernel_fpu_begin(void);
>> +void kernel_fpu_end(void);
>> +
>> +#endif
>
> I'll assume this is related to trick that places code in a separate
> translation unit, but I fail to understand it. Can you add a comment
> explaining it?
Yes, I can add a comment. Here, __riscv_f refers to RISC-V's F extension for
single-precision floating point, which is enabled by CC_FLAGS_FPU.
@@ -27,6 +27,7 @@ config RISCV
select ARCH_HAS_GCOV_PROFILE_ALL
select ARCH_HAS_GIGANTIC_PAGE
select ARCH_HAS_KCOV
+ select ARCH_HAS_KERNEL_FPU_SUPPORT if FPU
select ARCH_HAS_MMIOWB
select ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE
select ARCH_HAS_PMEM_API
@@ -81,6 +81,9 @@ KBUILD_CFLAGS += -march=$(shell echo $(riscv-march-y) | sed -E 's/(rv32ima|rv64i
KBUILD_AFLAGS += -march=$(riscv-march-y)
+# For C code built with floating-point support, exclude V but keep F and D.
+CC_FLAGS_FPU := -march=$(shell echo $(riscv-march-y) | sed -E 's/(rv32ima|rv64ima)([^v_]*)v?/\1\2/')
+
KBUILD_CFLAGS += -mno-save-restore
KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
new file mode 100644
@@ -0,0 +1,26 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2023 SiFive
+ */
+
+#ifndef _ASM_RISCV_FPU_H
+#define _ASM_RISCV_FPU_H
+
+#include <asm/switch_to.h>
+
+#define kernel_fpu_available() has_fpu()
+
+#ifdef __riscv_f
+
+#define kernel_fpu_begin() \
+ static_assert(false, "floating-point code must use a separate translation unit")
+#define kernel_fpu_end() kernel_fpu_begin()
+
+#else
+
+void kernel_fpu_begin(void);
+void kernel_fpu_end(void);
+
+#endif
+
+#endif /* ! _ASM_RISCV_FPU_H */
@@ -62,6 +62,7 @@ obj-$(CONFIG_MMU) += vdso.o vdso/
obj-$(CONFIG_RISCV_MISALIGNED) += traps_misaligned.o
obj-$(CONFIG_FPU) += fpu.o
+obj-$(CONFIG_FPU) += kernel_mode_fpu.o
obj-$(CONFIG_RISCV_ISA_V) += vector.o
obj-$(CONFIG_SMP) += smpboot.o
obj-$(CONFIG_SMP) += smp.o
new file mode 100644
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (C) 2023 SiFive
+ */
+
+#include <linux/export.h>
+#include <linux/preempt.h>
+
+#include <asm/csr.h>
+#include <asm/fpu.h>
+#include <asm/processor.h>
+#include <asm/switch_to.h>
+
+void kernel_fpu_begin(void)
+{
+ preempt_disable();
+ fstate_save(current, task_pt_regs(current));
+ csr_set(CSR_SSTATUS, SR_FS);
+}
+EXPORT_SYMBOL_GPL(kernel_fpu_begin);
+
+void kernel_fpu_end(void)
+{
+ csr_clear(CSR_SSTATUS, SR_FS);
+ fstate_restore(current, task_pt_regs(current));
+ preempt_enable();
+}
+EXPORT_SYMBOL_GPL(kernel_fpu_end);