diff --git a/sys/compat/linuxkpi/common/include/asm/fpu/api.h b/sys/compat/linuxkpi/common/include/asm/fpu/api.h
--- a/sys/compat/linuxkpi/common/include/asm/fpu/api.h
+++ b/sys/compat/linuxkpi/common/include/asm/fpu/api.h
@@ -28,41 +28,13 @@
 #ifndef	_FPU_API_H_
 #define	_FPU_API_H_
 
-#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
+#define	kernel_fpu_begin() \
+   lkpi_kernel_fpu_begin()
 
-#include <machine/fpu.h>
+#define	kernel_fpu_end() \
+   lkpi_kernel_fpu_end()
 
-extern struct fpu_kern_ctx *__lkpi_fpu_ctx;
-extern unsigned int __lkpi_fpu_ctx_level;
-
-static inline void
-kernel_fpu_begin()
-{
-	if (__lkpi_fpu_ctx_level++ == 0) {
-		fpu_kern_enter(curthread, __lkpi_fpu_ctx, FPU_KERN_NORMAL);
-	}
-}
-
-static inline void
-kernel_fpu_end()
-{
-	if (--__lkpi_fpu_ctx_level == 0) {
-		fpu_kern_leave(curthread, __lkpi_fpu_ctx);
-	}
-}
-
-#else
-
-static inline void
-kernel_fpu_begin()
-{
-}
-
-static inline void
-kernel_fpu_end()
-{
-}
-
-#endif
+extern void lkpi_kernel_fpu_begin(void);
+extern void lkpi_kernel_fpu_end(void);
 
 #endif /* _FPU_API_H_ */
diff --git a/sys/compat/linuxkpi/common/include/linux/sched.h b/sys/compat/linuxkpi/common/include/linux/sched.h
--- a/sys/compat/linuxkpi/common/include/linux/sched.h
+++ b/sys/compat/linuxkpi/common/include/linux/sched.h
@@ -82,7 +82,8 @@
 	int bsd_interrupt_value;
 	struct work_struct *work;	/* current work struct, if set */
 	struct task_struct *group_leader;
-  	unsigned rcu_section[TS_RCU_TYPE_MAX];
+	unsigned rcu_section[TS_RCU_TYPE_MAX];
+	unsigned int fpu_ctx_level;
 };
 
 #define	current	({ \
diff --git a/sys/compat/linuxkpi/common/src/linux_fpu.c b/sys/compat/linuxkpi/common/src/linux_fpu.c
--- a/sys/compat/linuxkpi/common/src/linux_fpu.c
+++ b/sys/compat/linuxkpi/common/src/linux_fpu.c
@@ -30,21 +30,44 @@
 #include <sys/proc.h>
 #include <sys/kernel.h>
 
+#include <linux/sched.h>
+
+#include <asm/fpu/api.h>
+
+#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
+
 #include <machine/fpu.h>
 
-struct fpu_kern_ctx *__lkpi_fpu_ctx;
-unsigned int __lkpi_fpu_ctx_level = 0;
+/*
+ * Technically the Linux API isn't supposed to allow nesting sections
+ * either, but currently used versions of GPU drivers rely on nesting
+ * working, so we only enter the section on the outermost level.
+ */
+
+void
+lkpi_kernel_fpu_begin(void)
+{
+	if ((current->fpu_ctx_level)++ == 0)
+		fpu_kern_enter(curthread, NULL, FPU_KERN_NOCTX);
+}
+
+void
+lkpi_kernel_fpu_end(void)
+{
+	if (--(current->fpu_ctx_level) == 0)
+		fpu_kern_leave(curthread, NULL);
+}
+
+#else
 
-static void
-linux_fpu_init(void *arg __unused)
+void
+lkpi_kernel_fpu_begin(void)
 {
-	__lkpi_fpu_ctx = fpu_kern_alloc_ctx(0);
 }
-SYSINIT(linux_fpu, SI_SUB_EVENTHANDLER, SI_ORDER_SECOND, linux_fpu_init, NULL);
 
-static void
-linux_fpu_uninit(void *arg __unused)
+void
+lkpi_kernel_fpu_end(void)
 {
-	fpu_kern_free_ctx(__lkpi_fpu_ctx);
 }
-SYSUNINIT(linux_fpu, SI_SUB_EVENTHANDLER, SI_ORDER_SECOND, linux_fpu_uninit, NULL);
+
+#endif
diff --git a/sys/conf/files.amd64 b/sys/conf/files.amd64
--- a/sys/conf/files.amd64
+++ b/sys/conf/files.amd64
@@ -398,9 +398,6 @@
 x86/x86/msi.c			optional	pci
 x86/xen/pv.c			optional	xenhvm
 
-compat/linuxkpi/common/src/linux_fpu.c		optional compat_linuxkpi \
-	compile-with "${LINUXKPI_C}"
-
 contrib/openzfs/module/zcommon/zfs_fletcher_avx512.c		optional zfs compile-with "${ZFS_C}"
 contrib/openzfs/module/zcommon/zfs_fletcher_intel.c		optional zfs compile-with "${ZFS_C}"
 contrib/openzfs/module/zcommon/zfs_fletcher_sse.c		optional zfs compile-with "${ZFS_C}"
diff --git a/sys/conf/files.arm64 b/sys/conf/files.arm64
--- a/sys/conf/files.arm64
+++ b/sys/conf/files.arm64
@@ -560,10 +560,6 @@
 # Xilinx
 arm/xilinx/uart_dev_cdnc.c			optional uart soc_xilinx_zynq
 
-# Linuxkpi
-compat/linuxkpi/common/src/linux_fpu.c		optional compat_linuxkpi \
-	compile-with "${LINUXKPI_C}"
-
 # Cloudabi
 arm64/cloudabi32/cloudabi32_sysvec.c		optional compat_cloudabi32
 arm64/cloudabi64/cloudabi64_sysvec.c		optional compat_cloudabi64
diff --git a/sys/conf/files.i386 b/sys/conf/files.i386
--- a/sys/conf/files.i386
+++ b/sys/conf/files.i386
@@ -182,6 +182,3 @@
 x86/x86/mptable.c		optional apic
 x86/x86/mptable_pci.c		optional apic pci
 x86/x86/msi.c			optional apic pci
-
-compat/linuxkpi/common/src/linux_fpu.c		optional compat_linuxkpi \
-	compile-with "${LINUXKPI_C}"
diff --git a/sys/modules/linuxkpi/Makefile b/sys/modules/linuxkpi/Makefile
--- a/sys/modules/linuxkpi/Makefile
+++ b/sys/modules/linuxkpi/Makefile
@@ -8,6 +8,7 @@
 	linux_dmi.c \
 	linux_domain.c \
 	linux_firmware.c \
+	linux_fpu.c \
 	linux_hrtimer.c \
 	linux_idr.c \
 	linux_kmod.c \
@@ -29,7 +30,7 @@
 
 .if ${MACHINE_CPUARCH} == "aarch64" || ${MACHINE_CPUARCH} == "amd64" || \
     ${MACHINE_CPUARCH} == "i386"
-SRCS+=	opt_acpi.h acpi_if.h linux_acpi.c linux_fpu.c
+SRCS+=	opt_acpi.h acpi_if.h linux_acpi.c
 .endif
 
 SRCS+=	${LINUXKPI_GENSRCS}