diff --git a/sys/compat/linuxkpi/common/include/linux/compat.h b/sys/compat/linuxkpi/common/include/linux/compat.h
--- a/sys/compat/linuxkpi/common/include/linux/compat.h
+++ b/sys/compat/linuxkpi/common/include/linux/compat.h
@@ -33,6 +33,10 @@
 #include <sys/proc.h>
 #include <sys/malloc.h>
 
+#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
+#define	LKPI_HAVE_FPU_CTX
+#endif
+
 struct domainset;
 struct thread;
 struct task_struct;
@@ -40,6 +44,7 @@
 extern int linux_alloc_current(struct thread *, int flags);
 extern void linux_free_current(struct task_struct *);
 extern struct domainset *linux_get_vm_domain_set(int node);
+extern int linux_set_fpu_ctx(struct task_struct *);
 
 static inline void
 linux_set_current(struct thread *td)
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
@@ -64,6 +64,7 @@
 #define	TASK_COMM_LEN		(MAXCOMLEN + 1)
 
 struct seq_file;
+struct fpu_kern_ctx;
 
 struct work_struct;
 struct task_struct {
@@ -89,6 +90,7 @@
 	struct task_struct *group_leader;
 	unsigned rcu_section[TS_RCU_TYPE_MAX];
 	unsigned int fpu_ctx_level;
+	struct fpu_kern_ctx *fpu_ctx;
 };
 
 #define	current	({ \
diff --git a/sys/compat/linuxkpi/common/src/linux_current.c b/sys/compat/linuxkpi/common/src/linux_current.c
--- a/sys/compat/linuxkpi/common/src/linux_current.c
+++ b/sys/compat/linuxkpi/common/src/linux_current.c
@@ -43,6 +43,10 @@
 #include <sys/sysctl.h>
 #include <vm/uma.h>
 
+#ifdef LKPI_HAVE_FPU_CTX
+#include <machine/fpu.h>
+#endif
+
 #ifdef DEV_APIC
 extern u_int first_msi_irq, num_msi_irqs;
 #endif
@@ -153,6 +157,21 @@
 	return (0);
 }
 
+int linux_set_fpu_ctx(struct task_struct *task)
+{
+#ifdef LKPI_HAVE_FPU_CTX
+	uint flags = 0;
+
+	if (task->fpu_ctx == NULL) {
+		if ((curthread->td_pflags & TDP_ITHREAD) != 0 ||
+		    !THREAD_CAN_SLEEP())
+			flags |= FPU_KERN_NOWAIT;
+		task->fpu_ctx = fpu_kern_alloc_ctx(flags);
+	}
+#endif
+	return (task->fpu_ctx != NULL ? 0 : ENOMEM);
+}
+
 struct mm_struct *
 linux_get_task_mm(struct task_struct *task)
 {
@@ -176,6 +195,10 @@
 linux_free_current(struct task_struct *ts)
 {
 	mmput(ts->mm);
+#ifdef LKPI_HAVE_FPU_CTX
+	if (ts->fpu_ctx != NULL)
+		fpu_kern_free_ctx(ts->fpu_ctx);
+#endif
 	uma_zfree(linux_current_zone, ts);
 }
 
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
@@ -30,11 +30,12 @@
 #include <sys/proc.h>
 #include <sys/kernel.h>
 
+#include <linux/compat.h>
 #include <linux/sched.h>
 
 #include <asm/fpu/api.h>
 
-#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
+#ifdef LKPI_HAVE_FPU_CTX
 
 #include <machine/fpu.h>
 
@@ -47,15 +48,20 @@
 void
 lkpi_kernel_fpu_begin(void)
 {
-	if ((current->fpu_ctx_level)++ == 0)
-		fpu_kern_enter(curthread, NULL, FPU_KERN_NOCTX);
+	int err;
+
+	if ((current->fpu_ctx_level)++ == 0) {
+		err = linux_set_fpu_ctx(current);
+		fpu_kern_enter(curthread, current->fpu_ctx,
+		    err == 0 ? FPU_KERN_KTHR : FPU_KERN_NOCTX);
+	}
 }
 
 void
 lkpi_kernel_fpu_end(void)
 {
 	if (--(current->fpu_ctx_level) == 0)
-		fpu_kern_leave(curthread, NULL);
+		fpu_kern_leave(curthread, current->fpu_ctx);
 }
 
 #else