[tip:x86/fpu] x86/fpu: Defer FPU state load until return to userspace

[tip-bot for Rik van Riel <tipbot@zytor.com>]

Commit-ID:  5f409e20b794565e2d60ad333e79334630a6c798
Gitweb:     https://git.kernel.org/tip/5f409e20b794565e2d60ad333e79334630a6c798
Author:     Rik van Riel <riel@surriel.com>
AuthorDate: Wed, 3 Apr 2019 18:41:52 +0200
Committer:  Borislav Petkov <bp@suse.de>
CommitDate: Fri, 12 Apr 2019 19:34:47 +0200

x86/fpu: Defer FPU state load until return to userspace

Defer loading of FPU state until return to userspace. This gives
the kernel the potential to skip loading FPU state for tasks that
stay in kernel mode, or for tasks that end up with repeated
invocations of kernel_fpu_begin() & kernel_fpu_end().

The fpregs_lock/unlock() section ensures that the registers remain
unchanged. Otherwise a context switch or a bottom half could save the
registers to its FPU context and the processor's FPU registers would
became random if modified at the same time.

KVM swaps the host/guest registers on entry/exit path. This flow has
been kept as is. First it ensures that the registers are loaded and then
saves the current (host) state before it loads the guest's registers. The
swap is done at the very end with disabled interrupts so it should not
change anymore before theg guest is entered. The read/save version seems
to be cheaper compared to memcpy() in a micro benchmark.

Each thread gets TIF_NEED_FPU_LOAD set as part of fork() / fpu__copy().
For kernel threads, this flag gets never cleared which avoids saving /
restoring the FPU state for kernel threads and during in-kernel usage of
the FPU registers.

 [
   bp: Correct and update commit message and fix checkpatch warnings.
   s/register/registers/ where it is used in plural.
   minor comment corrections.
   remove unused trace_x86_fpu_activate_state() TP.
 ]

Signed-off-by: Rik van Riel <riel@surriel.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Dave Hansen <dave.hansen@intel.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Aubrey Li <aubrey.li@intel.com>
Cc: Babu Moger <Babu.Moger@amd.com>
Cc: "Chang S. Bae" <chang.seok.bae@intel.com>
Cc: Dmitry Safonov <dima@arista.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: "Jason A. Donenfeld" <Jason@zx2c4.com>
Cc: Joerg Roedel <jroedel@suse.de>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: kvm ML <kvm@vger.kernel.org>
Cc: Nicolai Stange <nstange@suse.de>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: "Radim Krčmář" <rkrcmar@redhat.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Waiman Long <longman@redhat.com>
Cc: x86-ml <x86@kernel.org>
Cc: Yi Wang <wang.yi59@zte.com.cn>
Link: https://lkml.kernel.org/r/20190403164156.19645-24-bigeasy@linutronix.de
---
 arch/x86/entry/common.c             |  10 +++-
 arch/x86/include/asm/fpu/api.h      |  22 +++++++-
 arch/x86/include/asm/fpu/internal.h |  27 +++++----
 arch/x86/include/asm/trace/fpu.h    |  10 ++--
 arch/x86/kernel/fpu/core.c          | 106 +++++++++++++++++++++++++++---------
 arch/x86/kernel/fpu/signal.c        |  49 ++++++++++-------
 arch/x86/kernel/process.c           |   2 +-
 arch/x86/kernel/process_32.c        |   5 +-
 arch/x86/kernel/process_64.c        |   5 +-
 arch/x86/kvm/x86.c                  |  20 +++++--
 10 files changed, 184 insertions(+), 72 deletions(-)

diff --git a/arch/x86/entry/common.c b/arch/x86/entry/common.c
index 7bc105f47d21..51beb8d29123 100644
--- a/arch/x86/entry/common.c
+++ b/arch/x86/entry/common.c
@@ -25,12 +25,13 @@
 #include <linux/uprobes.h>
 #include <linux/livepatch.h>
 #include <linux/syscalls.h>
+#include <linux/uaccess.h>
 
 #include <asm/desc.h>
 #include <asm/traps.h>
 #include <asm/vdso.h>
-#include <linux/uaccess.h>
 #include <asm/cpufeature.h>
+#include <asm/fpu/api.h>
 
 #define CREATE_TRACE_POINTS
 #include <trace/events/syscalls.h>
@@ -196,6 +197,13 @@ __visible inline void prepare_exit_to_usermode(struct pt_regs *regs)
 	if (unlikely(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS))
 		exit_to_usermode_loop(regs, cached_flags);
 
+	/* Reload ti->flags; we may have rescheduled above. */
+	cached_flags = READ_ONCE(ti->flags);
+
+	fpregs_assert_state_consistent();
+	if (unlikely(cached_flags & _TIF_NEED_FPU_LOAD))
+		switch_fpu_return();
+
 #ifdef CONFIG_COMPAT
 	/*
 	 * Compat syscalls set TS_COMPAT.  Make sure we clear it before
diff --git a/arch/x86/include/asm/fpu/api.h b/arch/x86/include/asm/fpu/api.h
index 73e684160f35..b774c52e5411 100644
--- a/arch/x86/include/asm/fpu/api.h
+++ b/arch/x86/include/asm/fpu/api.h
@@ -10,7 +10,7 @@
 
 #ifndef _ASM_X86_FPU_API_H
 #define _ASM_X86_FPU_API_H
-#include <linux/preempt.h>
+#include <linux/bottom_half.h>
 
 /*
  * Use kernel_fpu_begin/end() if you intend to use FPU in kernel context. It
@@ -22,17 +22,37 @@
 extern void kernel_fpu_begin(void);
 extern void kernel_fpu_end(void);
 extern bool irq_fpu_usable(void);
+extern void fpregs_mark_activate(void);
 
+/*
+ * Use fpregs_lock() while editing CPU's FPU registers or fpu->state.
+ * A context switch will (and softirq might) save CPU's FPU registers to
+ * fpu->state and set TIF_NEED_FPU_LOAD leaving CPU's FPU registers in
+ * a random state.
+ */
 static inline void fpregs_lock(void)
 {
 	preempt_disable();
+	local_bh_disable();
 }
 
 static inline void fpregs_unlock(void)
 {
+	local_bh_enable();
 	preempt_enable();
 }
 
+#ifdef CONFIG_X86_DEBUG_FPU
+extern void fpregs_assert_state_consistent(void);
+#else
+static inline void fpregs_assert_state_consistent(void) { }
+#endif
+
+/*
+ * Load the task FPU state before returning to userspace.
+ */
+extern void switch_fpu_return(void);
+
 /*
  * Query the presence of one or more xfeatures. Works on any legacy CPU as well.
  *
diff --git a/arch/x86/include/asm/fpu/internal.h b/arch/x86/include/asm/fpu/internal.h
index 2cf04fbcba5d..0c8a5093647a 100644
--- a/arch/x86/include/asm/fpu/internal.h
+++ b/arch/x86/include/asm/fpu/internal.h
@@ -30,7 +30,7 @@ extern void fpu__prepare_write(struct fpu *fpu);
 extern void fpu__save(struct fpu *fpu);
 extern int  fpu__restore_sig(void __user *buf, int ia32_frame);
 extern void fpu__drop(struct fpu *fpu);
-extern int  fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu);
+extern int  fpu__copy(struct task_struct *dst, struct task_struct *src);
 extern void fpu__clear(struct fpu *fpu);
 extern int  fpu__exception_code(struct fpu *fpu, int trap_nr);
 extern int  dump_fpu(struct pt_regs *ptregs, struct user_i387_struct *fpstate);
@@ -531,13 +531,20 @@ static inline void fpregs_activate(struct fpu *fpu)
 /*
  * Internal helper, do not use directly. Use switch_fpu_return() instead.
  */
-static inline void __fpregs_load_activate(struct fpu *fpu, int cpu)
+static inline void __fpregs_load_activate(void)
 {
+	struct fpu *fpu = &current->thread.fpu;
+	int cpu = smp_processor_id();
+
+	if (WARN_ON_ONCE(current->mm == NULL))
+		return;
+
 	if (!fpregs_state_valid(fpu, cpu)) {
-		if (current->mm)
-			copy_kernel_to_fpregs(&fpu->state);
+		copy_kernel_to_fpregs(&fpu->state);
 		fpregs_activate(fpu);
+		fpu->last_cpu = cpu;
 	}
+	clear_thread_flag(TIF_NEED_FPU_LOAD);
 }
 
 /*
@@ -548,8 +555,8 @@ static inline void __fpregs_load_activate(struct fpu *fpu, int cpu)
  *  - switch_fpu_prepare() saves the old state.
  *    This is done within the context of the old process.
  *
- *  - switch_fpu_finish() restores the new state as
- *    necessary.
+ *  - switch_fpu_finish() sets TIF_NEED_FPU_LOAD; the floating point state
+ *    will get loaded on return to userspace, or when the kernel needs it.
  *
  * If TIF_NEED_FPU_LOAD is cleared then the CPU's FPU registers
  * are saved in the current thread's FPU register state.
@@ -581,10 +588,10 @@ switch_fpu_prepare(struct fpu *old_fpu, int cpu)
  */
 
 /*
- * Set up the userspace FPU context for the new task, if the task
- * has used the FPU.
+ * Load PKRU from the FPU context if available. Delay loading of the
+ * complete FPU state until the return to userland.
  */
-static inline void switch_fpu_finish(struct fpu *new_fpu, int cpu)
+static inline void switch_fpu_finish(struct fpu *new_fpu)
 {
 	u32 pkru_val = init_pkru_value;
 	struct pkru_state *pk;
@@ -592,7 +599,7 @@ static inline void switch_fpu_finish(struct fpu *new_fpu, int cpu)
 	if (!static_cpu_has(X86_FEATURE_FPU))
 		return;
 
-	__fpregs_load_activate(new_fpu, cpu);
+	set_thread_flag(TIF_NEED_FPU_LOAD);
 
 	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
 		return;
diff --git a/arch/x86/include/asm/trace/fpu.h b/arch/x86/include/asm/trace/fpu.h
index bd65f6ba950f..879b77792f94 100644
--- a/arch/x86/include/asm/trace/fpu.h
+++ b/arch/x86/include/asm/trace/fpu.h
@@ -13,19 +13,22 @@ DECLARE_EVENT_CLASS(x86_fpu,
 
 	TP_STRUCT__entry(
 		__field(struct fpu *, fpu)
+		__field(bool, load_fpu)
 		__field(u64, xfeatures)
 		__field(u64, xcomp_bv)
 		),
 
 	TP_fast_assign(
 		__entry->fpu		= fpu;
+		__entry->load_fpu	= test_thread_flag(TIF_NEED_FPU_LOAD);
 		if (boot_cpu_has(X86_FEATURE_OSXSAVE)) {
 			__entry->xfeatures = fpu->state.xsave.header.xfeatures;
 			__entry->xcomp_bv  = fpu->state.xsave.header.xcomp_bv;
 		}
 	),
-	TP_printk("x86/fpu: %p xfeatures: %llx xcomp_bv: %llx",
+	TP_printk("x86/fpu: %p load: %d xfeatures: %llx xcomp_bv: %llx",
 			__entry->fpu,
+			__entry->load_fpu,
 			__entry->xfeatures,
 			__entry->xcomp_bv
 	)
@@ -61,11 +64,6 @@ DEFINE_EVENT(x86_fpu, x86_fpu_regs_deactivated,
 	TP_ARGS(fpu)
 );
 
-DEFINE_EVENT(x86_fpu, x86_fpu_activate_state,
-	TP_PROTO(struct fpu *fpu),
-	TP_ARGS(fpu)
-);
-
 DEFINE_EVENT(x86_fpu, x86_fpu_init_state,
 	TP_PROTO(struct fpu *fpu),
 	TP_ARGS(fpu)
diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c
index 739ca3ae2bdc..ce243f76bdb7 100644
--- a/arch/x86/kernel/fpu/core.c
+++ b/arch/x86/kernel/fpu/core.c
@@ -102,23 +102,20 @@ static void __kernel_fpu_begin(void)
 	kernel_fpu_disable();
 
 	if (current->mm) {
-		/*
-		 * Ignore return value -- we don't care if reg state
-		 * is clobbered.
-		 */
-		copy_fpregs_to_fpstate(fpu);
-	} else {
-		__cpu_invalidate_fpregs_state();
+		if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
+			set_thread_flag(TIF_NEED_FPU_LOAD);
+			/*
+			 * Ignore return value -- we don't care if reg state
+			 * is clobbered.
+			 */
+			copy_fpregs_to_fpstate(fpu);
+		}
 	}
+	__cpu_invalidate_fpregs_state();
 }
 
 static void __kernel_fpu_end(void)
 {
-	struct fpu *fpu = &current->thread.fpu;
-
-	if (current->mm)
-		copy_kernel_to_fpregs(&fpu->state);
-
 	kernel_fpu_enable();
 }
 
@@ -145,14 +142,17 @@ void fpu__save(struct fpu *fpu)
 {
 	WARN_ON_FPU(fpu != &current->thread.fpu);
 
-	preempt_disable();
+	fpregs_lock();
 	trace_x86_fpu_before_save(fpu);
 
-	if (!copy_fpregs_to_fpstate(fpu))
-		copy_kernel_to_fpregs(&fpu->state);
+	if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
+		if (!copy_fpregs_to_fpstate(fpu)) {
+			copy_kernel_to_fpregs(&fpu->state);
+		}
+	}
 
 	trace_x86_fpu_after_save(fpu);
-	preempt_enable();
+	fpregs_unlock();
 }
 EXPORT_SYMBOL_GPL(fpu__save);
 
@@ -185,8 +185,11 @@ void fpstate_init(union fpregs_state *state)
 }
 EXPORT_SYMBOL_GPL(fpstate_init);
 
-int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
+int fpu__copy(struct task_struct *dst, struct task_struct *src)
 {
+	struct fpu *dst_fpu = &dst->thread.fpu;
+	struct fpu *src_fpu = &src->thread.fpu;
+
 	dst_fpu->last_cpu = -1;
 
 	if (!static_cpu_has(X86_FEATURE_FPU))
@@ -201,16 +204,23 @@ int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
 	memset(&dst_fpu->state.xsave, 0, fpu_kernel_xstate_size);
 
 	/*
-	 * Save current FPU registers directly into the child
-	 * FPU context, without any memory-to-memory copying.
+	 * If the FPU registers are not current just memcpy() the state.
+	 * Otherwise save current FPU registers directly into the child's FPU
+	 * context, without any memory-to-memory copying.
 	 *
 	 * ( The function 'fails' in the FNSAVE case, which destroys
-	 *   register contents so we have to copy them back. )
+	 *   register contents so we have to load them back. )
 	 */
-	if (!copy_fpregs_to_fpstate(dst_fpu)) {
-		memcpy(&src_fpu->state, &dst_fpu->state, fpu_kernel_xstate_size);
-		copy_kernel_to_fpregs(&src_fpu->state);
-	}
+	fpregs_lock();
+	if (test_thread_flag(TIF_NEED_FPU_LOAD))
+		memcpy(&dst_fpu->state, &src_fpu->state, fpu_kernel_xstate_size);
+
+	else if (!copy_fpregs_to_fpstate(dst_fpu))
+		copy_kernel_to_fpregs(&dst_fpu->state);
+
+	fpregs_unlock();
+
+	set_tsk_thread_flag(dst, TIF_NEED_FPU_LOAD);
 
 	trace_x86_fpu_copy_src(src_fpu);
 	trace_x86_fpu_copy_dst(dst_fpu);
@@ -226,10 +236,9 @@ static void fpu__initialize(struct fpu *fpu)
 {
 	WARN_ON_FPU(fpu != &current->thread.fpu);
 
+	set_thread_flag(TIF_NEED_FPU_LOAD);
 	fpstate_init(&fpu->state);
 	trace_x86_fpu_init_state(fpu);
-
-	trace_x86_fpu_activate_state(fpu);
 }
 
 /*
@@ -308,6 +317,8 @@ void fpu__drop(struct fpu *fpu)
  */
 static inline void copy_init_fpstate_to_fpregs(void)
 {
+	fpregs_lock();
+
 	if (use_xsave())
 		copy_kernel_to_xregs(&init_fpstate.xsave, -1);
 	else if (static_cpu_has(X86_FEATURE_FXSR))
@@ -317,6 +328,9 @@ static inline void copy_init_fpstate_to_fpregs(void)
 
 	if (boot_cpu_has(X86_FEATURE_OSPKE))
 		copy_init_pkru_to_fpregs();
+
+	fpregs_mark_activate();
+	fpregs_unlock();
 }
 
 /*
@@ -339,6 +353,46 @@ void fpu__clear(struct fpu *fpu)
 		copy_init_fpstate_to_fpregs();
 }
 
+/*
+ * Load FPU context before returning to userspace.
+ */
+void switch_fpu_return(void)
+{
+	if (!static_cpu_has(X86_FEATURE_FPU))
+		return;
+
+	__fpregs_load_activate();
+}
+EXPORT_SYMBOL_GPL(switch_fpu_return);
+
+#ifdef CONFIG_X86_DEBUG_FPU
+/*
+ * If current FPU state according to its tracking (loaded FPU context on this
+ * CPU) is not valid then we must have TIF_NEED_FPU_LOAD set so the context is
+ * loaded on return to userland.
+ */
+void fpregs_assert_state_consistent(void)
+{
+	struct fpu *fpu = &current->thread.fpu;
+
+	if (test_thread_flag(TIF_NEED_FPU_LOAD))
+		return;
+
+	WARN_ON_FPU(!fpregs_state_valid(fpu, smp_processor_id()));
+}
+EXPORT_SYMBOL_GPL(fpregs_assert_state_consistent);
+#endif
+
+void fpregs_mark_activate(void)
+{
+	struct fpu *fpu = &current->thread.fpu;
+
+	fpregs_activate(fpu);
+	fpu->last_cpu = smp_processor_id();
+	clear_thread_flag(TIF_NEED_FPU_LOAD);
+}
+EXPORT_SYMBOL_GPL(fpregs_mark_activate);
+
 /*
  * x87 math exception handling:
  */
diff --git a/arch/x86/kernel/fpu/signal.c b/arch/x86/kernel/fpu/signal.c
index b13e86b29426..6df1f15e0cd5 100644
--- a/arch/x86/kernel/fpu/signal.c
+++ b/arch/x86/kernel/fpu/signal.c
@@ -269,11 +269,9 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
 	struct task_struct *tsk = current;
 	struct fpu *fpu = &tsk->thread.fpu;
 	struct user_i387_ia32_struct env;
-	union fpregs_state *state;
 	u64 xfeatures = 0;
 	int fx_only = 0;
 	int ret = 0;
-	void *tmp;
 
 	ia32_fxstate &= (IS_ENABLED(CONFIG_X86_32) ||
 			 IS_ENABLED(CONFIG_IA32_EMULATION));
@@ -308,14 +306,18 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
 		}
 	}
 
-	tmp = kzalloc(sizeof(*state) + fpu_kernel_xstate_size + 64, GFP_KERNEL);
-	if (!tmp)
-		return -ENOMEM;
-	state = PTR_ALIGN(tmp, 64);
+	/*
+	 * The current state of the FPU registers does not matter. By setting
+	 * TIF_NEED_FPU_LOAD unconditionally it is ensured that the our xstate
+	 * is not modified on context switch and that the xstate is considered
+	 * to be loaded again on return to userland (overriding last_cpu avoids
+	 * the optimisation).
+	 */
+	set_thread_flag(TIF_NEED_FPU_LOAD);
+	__fpu_invalidate_fpregs_state(fpu);
 
 	if ((unsigned long)buf_fx % 64)
 		fx_only = 1;
-
 	/*
 	 * For 32-bit frames with fxstate, copy the fxstate so it can be
 	 * reconstructed later.
@@ -331,43 +333,52 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
 		u64 init_bv = xfeatures_mask & ~xfeatures;
 
 		if (using_compacted_format()) {
-			ret = copy_user_to_xstate(&state->xsave, buf_fx);
+			ret = copy_user_to_xstate(&fpu->state.xsave, buf_fx);
 		} else {
-			ret = __copy_from_user(&state->xsave, buf_fx, state_size);
+			ret = __copy_from_user(&fpu->state.xsave, buf_fx, state_size);
 
 			if (!ret && state_size > offsetof(struct xregs_state, header))
-				ret = validate_xstate_header(&state->xsave.header);
+				ret = validate_xstate_header(&fpu->state.xsave.header);
 		}
 		if (ret)
 			goto err_out;
 
-		sanitize_restored_xstate(state, envp, xfeatures, fx_only);
+		sanitize_restored_xstate(&fpu->state, envp, xfeatures, fx_only);
 
+		fpregs_lock();
 		if (unlikely(init_bv))
 			copy_kernel_to_xregs(&init_fpstate.xsave, init_bv);
-		ret = copy_kernel_to_xregs_err(&state->xsave, xfeatures);
+		ret = copy_kernel_to_xregs_err(&fpu->state.xsave, xfeatures);
 
 	} else if (use_fxsr()) {
-		ret = __copy_from_user(&state->fxsave, buf_fx, state_size);
-		if (ret)
+		ret = __copy_from_user(&fpu->state.fxsave, buf_fx, state_size);
+		if (ret) {
+			ret = -EFAULT;
 			goto err_out;
+		}
+
+		sanitize_restored_xstate(&fpu->state, envp, xfeatures, fx_only);
 
-		sanitize_restored_xstate(state, envp, xfeatures, fx_only);
+		fpregs_lock();
 		if (use_xsave()) {
 			u64 init_bv = xfeatures_mask & ~XFEATURE_MASK_FPSSE;
 			copy_kernel_to_xregs(&init_fpstate.xsave, init_bv);
 		}
 
-		ret = copy_kernel_to_fxregs_err(&state->fxsave);
+		ret = copy_kernel_to_fxregs_err(&fpu->state.fxsave);
 	} else {
-		ret = __copy_from_user(&state->fsave, buf_fx, state_size);
+		ret = __copy_from_user(&fpu->state.fsave, buf_fx, state_size);
 		if (ret)
 			goto err_out;
-		ret = copy_kernel_to_fregs_err(&state->fsave);
+
+		fpregs_lock();
+		ret = copy_kernel_to_fregs_err(&fpu->state.fsave);
 	}
+	if (!ret)
+		fpregs_mark_activate();
+	fpregs_unlock();
 
 err_out:
-	kfree(tmp);
 	if (ret)
 		fpu__clear(fpu);
 	return ret;
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index 58ac7be52c7a..b9b8e6eb74f2 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -101,7 +101,7 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 	dst->thread.vm86 = NULL;
 #endif
 
-	return fpu__copy(&dst->thread.fpu, &src->thread.fpu);
+	return fpu__copy(dst, src);
 }
 
 /*
diff --git a/arch/x86/kernel/process_32.c b/arch/x86/kernel/process_32.c
index 77d9eb43ccac..1bc47f3a4885 100644
--- a/arch/x86/kernel/process_32.c
+++ b/arch/x86/kernel/process_32.c
@@ -234,7 +234,8 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 
 	/* never put a printk in __switch_to... printk() calls wake_up*() indirectly */
 
-	switch_fpu_prepare(prev_fpu, cpu);
+	if (!test_thread_flag(TIF_NEED_FPU_LOAD))
+		switch_fpu_prepare(prev_fpu, cpu);
 
 	/*
 	 * Save away %gs. No need to save %fs, as it was saved on the
@@ -290,7 +291,7 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 
 	this_cpu_write(current_task, next_p);
 
-	switch_fpu_finish(next_fpu, cpu);
+	switch_fpu_finish(next_fpu);
 
 	/* Load the Intel cache allocation PQR MSR. */
 	resctrl_sched_in();
diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c
index ffea7c557963..37b2ecef041e 100644
--- a/arch/x86/kernel/process_64.c
+++ b/arch/x86/kernel/process_64.c
@@ -520,7 +520,8 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 	WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
 		     this_cpu_read(irq_count) != -1);
 
-	switch_fpu_prepare(prev_fpu, cpu);
+	if (!test_thread_flag(TIF_NEED_FPU_LOAD))
+		switch_fpu_prepare(prev_fpu, cpu);
 
 	/* We must save %fs and %gs before load_TLS() because
 	 * %fs and %gs may be cleared by load_TLS().
@@ -572,7 +573,7 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 	this_cpu_write(current_task, next_p);
 	this_cpu_write(cpu_current_top_of_stack, task_top_of_stack(next_p));
 
-	switch_fpu_finish(next_fpu, cpu);
+	switch_fpu_finish(next_fpu);
 
 	/* Reload sp0. */
 	update_task_stack(next_p);
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 8022e7769b3a..e340c3c0cba3 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -7877,6 +7877,10 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 		wait_lapic_expire(vcpu);
 	guest_enter_irqoff();
 
+	fpregs_assert_state_consistent();
+	if (test_thread_flag(TIF_NEED_FPU_LOAD))
+		switch_fpu_return();
+
 	if (unlikely(vcpu->arch.switch_db_regs)) {
 		set_debugreg(0, 7);
 		set_debugreg(vcpu->arch.eff_db[0], 0);
@@ -8137,22 +8141,30 @@ static int complete_emulated_mmio(struct kvm_vcpu *vcpu)
 /* Swap (qemu) user FPU context for the guest FPU context. */
 static void kvm_load_guest_fpu(struct kvm_vcpu *vcpu)
 {
-	preempt_disable();
+	fpregs_lock();
+
 	copy_fpregs_to_fpstate(&current->thread.fpu);
 	/* PKRU is separately restored in kvm_x86_ops->run.  */
 	__copy_kernel_to_fpregs(&vcpu->arch.guest_fpu->state,
 				~XFEATURE_MASK_PKRU);
-	preempt_enable();
+
+	fpregs_mark_activate();
+	fpregs_unlock();
+
 	trace_kvm_fpu(1);
 }
 
 /* When vcpu_run ends, restore user space FPU context. */
 static void kvm_put_guest_fpu(struct kvm_vcpu *vcpu)
 {
-	preempt_disable();
+	fpregs_lock();
+
 	copy_fpregs_to_fpstate(vcpu->arch.guest_fpu);
 	copy_kernel_to_fpregs(&current->thread.fpu.state);
-	preempt_enable();
+
+	fpregs_mark_activate();
+	fpregs_unlock();
+
 	++vcpu->stat.fpu_reload;
 	trace_kvm_fpu(0);
 }