Several architectures provide an API to enable the FPU and run
floating-point SIMD code in kernel space. However, the function names,
header locations, and semantics are inconsistent across architectures,
and FPU support may be gated behind other Kconfig options.
Provide a standard way for architectures to declare that kernel space
FPU support is available. Architectures selecting this option must
implement what is currently the most common API (kernel_fpu_begin() and
kernel_fpu_end(), plus a new function kernel_fpu_available()) and
provide the appropriate CFLAGS for compiling floating-point C code.
Suggested-by: Christoph Hellwig <hch@xxxxxx>
Reviewed-by: Christoph Hellwig <hch@xxxxxx>
Signed-off-by: Samuel Holland <samuel.holland@xxxxxxxxxx>
---
(no changes since v2)
Changes in v2:
- Add documentation explaining the built-time and runtime APIs
- Add a linux/fpu.h header for generic isolation enforcement
Documentation/core-api/floating-point.rst | 78 +++++++++++++++++++++++
Documentation/core-api/index.rst | 1 +
Makefile | 5 ++
arch/Kconfig | 6 ++
include/linux/fpu.h | 12 ++++
5 files changed, 102 insertions(+)
create mode 100644 Documentation/core-api/floating-point.rst
create mode 100644 include/linux/fpu.h
diff --git a/Documentation/core-api/floating-point.rst b/Documentation/core-api/floating-point.rst
new file mode 100644
index 000000000000..a8d0d4b05052
--- /dev/null
+++ b/Documentation/core-api/floating-point.rst
@@ -0,0 +1,78 @@
+.. SPDX-License-Identifier: GPL-2.0+
+
+Floating-point API
+==================
+
+Kernel code is normally prohibited from using floating-point (FP) registers or
+instructions, including the C float and double data types. This rule reduces
+system call overhead, because the kernel does not need to save and restore the
+userspace floating-point register state.
+
+However, occasionally drivers or library functions may need to include FP code.
+This is supported by isolating the functions containing FP code to a separate
+translation unit (a separate source file), and saving/restoring the FP register
+state around calls to those functions. This creates "critical sections" of
+floating-point usage.
+
+The reason for this isolation is to prevent the compiler from generating code
+touching the FP registers outside these critical sections. Compilers sometimes
+use FP registers to optimize inlined ``memcpy`` or variable assignment, as
+floating-point registers may be wider than general-purpose registers.
+
+Usability of floating-point code within the kernel is architecture-specific.
+Additionally, because a single kernel may be configured to support platforms
+both with and without a floating-point unit, FPU availability must be checked
+both at build time and at run time.
+
+Several architectures implement the generic kernel floating-point API from
+``linux/fpu.h``, as described below. Some other architectures implement their
+own unique APIs, which are documented separately.
+
+Build-time API
+--------------
+
+Floating-point code may be built if the option ``ARCH_HAS_KERNEL_FPU_SUPPORT``
+is enabled. For C code, such code must be placed in a separate file, and that
+file must have its compilation flags adjusted using the following pattern::
+
+ CFLAGS_foo.o += $(CC_FLAGS_FPU)
+ CFLAGS_REMOVE_foo.o += $(CC_FLAGS_NO_FPU)
+
+Architectures are expected to define one or both of these variables in their
+top-level Makefile as needed. For example::
+
+ CC_FLAGS_FPU := -mhard-float
+
+or::
+
+ CC_FLAGS_NO_FPU := -msoft-float
+
+Normal kernel code is assumed to use the equivalent of ``CC_FLAGS_NO_FPU``.
+
+Runtime API
+-----------
+
+The runtime API is provided in ``linux/fpu.h``. This header cannot be included
+from files implementing FP code (those with their compilation flags adjusted as
+above). Instead, it must be included when defining the FP critical sections.
+
+.. c:function:: bool kernel_fpu_available( void )
+
+ This function reports if floating-point code can be used on this CPU or
+ platform. The value returned by this function is not expected to change
+ at runtime, so it only needs to be called once, not before every
+ critical section.
+
+.. c:function:: void kernel_fpu_begin( void )
+ void kernel_fpu_end( void )
+
+ These functions create a floating-point critical section. It is only
+ valid to call ``kernel_fpu_begin()`` after a previous call to
+ ``kernel_fpu_available()`` returned ``true``. These functions are only
+ guaranteed to be callable from (preemptible or non-preemptible) process
+ context.
+
+ Preemption may be disabled inside critical sections, so their size
+ should be minimized. They are *not* required to be reentrant. If the
+ caller expects to nest critical sections, it must implement its own
+ reference counting.
diff --git a/Documentation/core-api/index.rst b/Documentation/core-api/index.rst
index 7a3a08d81f11..974beccd671f 100644
--- a/Documentation/core-api/index.rst
+++ b/Documentation/core-api/index.rst
@@ -48,6 +48,7 @@ Library functionality that is used throughout the kernel.
errseq
wrappers/atomic_t
wrappers/atomic_bitops
+ floating-point
Low level entry and exit
========================
diff --git a/Makefile b/Makefile
index 763b6792d3d5..710f65e4249d 100644
--- a/Makefile
+++ b/Makefile
@@ -964,6 +964,11 @@ KBUILD_CFLAGS += $(CC_FLAGS_CFI)
export CC_FLAGS_CFI
endif
+# Architectures can define flags to add/remove for floating-point support
+CC_FLAGS_FPU += -D_LINUX_FPU_COMPILATION_UNIT
+export CC_FLAGS_FPU
+export CC_FLAGS_NO_FPU
+
ifneq ($(CONFIG_FUNCTION_ALIGNMENT),0)
# Set the minimal function alignment. Use the newer GCC option
# -fmin-function-alignment if it is available, or fall back to -falign-funtions.
diff --git a/arch/Kconfig b/arch/Kconfig
index 9f066785bb71..8e34b3acf73d 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -1569,6 +1569,12 @@ config ARCH_HAS_NONLEAF_PMD_YOUNG
address translations. Page table walkers that clear the accessed bit
may use this capability to reduce their search space.
+config ARCH_HAS_KERNEL_FPU_SUPPORT
+ bool
+ help
+ Architectures that select this option can run floating-point code in
+ the kernel, as described in Documentation/core-api/floating-point.rst.
+
source "kernel/gcov/Kconfig"
source "scripts/gcc-plugins/Kconfig"
diff --git a/include/linux/fpu.h b/include/linux/fpu.h
new file mode 100644
index 000000000000..2fb63e22913b
--- /dev/null
+++ b/include/linux/fpu.h
@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef _LINUX_FPU_H
+#define _LINUX_FPU_H
+
+#ifdef _LINUX_FPU_COMPILATION_UNIT
+#error FP code must be compiled separately. See Documentation/core-api/floating-point.rst.
+#endif
+
+#include <asm/fpu.h>
+
+#endif
--
2.44.0
