Linux 3.2.8

commit 34ddc81a23 upstream.

After all the FPU state cleanups and finally finding the problem that
caused all our FPU save/restore problems, this re-introduces the
preloading of FPU state that was removed in commit b3b0870ef3 ("i387:
do not preload FPU state at task switch time").

However, instead of simply reverting the removal, this reimplements
preloading with several fixes, most notably

 - properly abstracted as a true FPU state switch, rather than as
   open-coded save and restore with various hacks.

   In particular, implementing it as a proper FPU state switch allows us
   to optimize the CR0.TS flag accesses: there is no reason to set the
   TS bit only to then almost immediately clear it again.  CR0 accesses
   are quite slow and expensive, don't flip the bit back and forth for
   no good reason.

 - Make sure that the same model works for both x86-32 and x86-64, so
   that there are no gratuitous differences between the two due to the
   way they save and restore segment state differently due to
   architectural differences that really don't matter to the FPU state.

 - Avoid exposing the "preload" state to the context switch routines,
   and in particular allow the concept of lazy state restore: if nothing
   else has used the FPU in the meantime, and the process is still on
   the same CPU, we can avoid restoring state from memory entirely, just
   re-expose the state that is still in the FPU unit.

   That optimized lazy restore isn't actually implemented here, but the
   infrastructure is set up for it.  Of course, older CPU's that use
   'fnsave' to save the state cannot take advantage of this, since the
   state saving also trashes the state.

In other words, there is now an actual _design_ to the FPU state saving,
rather than just random historical baggage.  Hopefully it's easier to
follow as a result.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

Makefile

@@ -1,6 +1,6 @@
 VERSION = 3
 PATCHLEVEL = 2
-SUBLEVEL = 7
+SUBLEVEL = 8
 EXTRAVERSION =
 NAME = Saber-toothed Squirrel
 

arch/x86/include/asm/i387.h

@@ -29,8 +29,8 @@ extern unsigned int sig_xstate_size;
 extern void fpu_init(void);
 extern void mxcsr_feature_mask_init(void);
 extern int init_fpu(struct task_struct *child);
-extern asmlinkage void math_state_restore(void);
-extern void __math_state_restore(void);
+extern void __math_state_restore(struct task_struct *);
+extern void math_state_restore(void);
 extern int dump_fpu(struct pt_regs *, struct user_i387_struct *);
 
 extern user_regset_active_fn fpregs_active, xfpregs_active;
@@ -212,19 +212,11 @@ static inline void fpu_fxsave(struct fpu *fpu)
 
 #endif	/* CONFIG_X86_64 */
 
-/* We need a safe address that is cheap to find and that is already
-   in L1 during context switch. The best choices are unfortunately
-   different for UP and SMP */
-#ifdef CONFIG_SMP
-#define safe_address (__per_cpu_offset[0])
-#else
-#define safe_address (kstat_cpu(0).cpustat.user)
-#endif
-
 /*
- * These must be called with preempt disabled
+ * These must be called with preempt disabled. Returns
+ * 'true' if the FPU state is still intact.
  */
-static inline void fpu_save_init(struct fpu *fpu)
+static inline int fpu_save_init(struct fpu *fpu)
 {
 	if (use_xsave()) {
 		fpu_xsave(fpu);
@@ -233,33 +225,33 @@ static inline void fpu_save_init(struct fpu *fpu)
 		 * xsave header may indicate the init state of the FP.
 		 */
 		if (!(fpu->state->xsave.xsave_hdr.xstate_bv & XSTATE_FP))
-			return;
+			return 1;
 	} else if (use_fxsr()) {
 		fpu_fxsave(fpu);
 	} else {
 		asm volatile("fnsave %[fx]; fwait"
 			     : [fx] "=m" (fpu->state->fsave));
-		return;
+		return 0;
 	}
 
-	if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES))
+	/*
+	 * If exceptions are pending, we need to clear them so
+	 * that we don't randomly get exceptions later.
+	 *
+	 * FIXME! Is this perhaps only true for the old-style
+	 * irq13 case? Maybe we could leave the x87 state
+	 * intact otherwise?
+	 */
+	if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) {
 		asm volatile("fnclex");
-
-	/* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception
-	   is pending.  Clear the x87 state here by setting it to fixed
-	   values. safe_address is a random variable that should be in L1 */
-	alternative_input(
-		ASM_NOP8 ASM_NOP2,
-		"emms\n\t"	  	/* clear stack tags */
-		"fildl %P[addr]",	/* set F?P to defined value */
-		X86_FEATURE_FXSAVE_LEAK,
-		[addr] "m" (safe_address));
+		return 0;
+	}
+	return 1;
 }
 
-static inline void __save_init_fpu(struct task_struct *tsk)
+static inline int __save_init_fpu(struct task_struct *tsk)
 {
-	fpu_save_init(&tsk->thread.fpu);
-	task_thread_info(tsk)->status &= ~TS_USEDFPU;
+	return fpu_save_init(&tsk->thread.fpu);
 }
 
 static inline int fpu_fxrstor_checking(struct fpu *fpu)
@@ -280,40 +272,186 @@ static inline int restore_fpu_checking(struct task_struct *tsk)
 	return fpu_restore_checking(&tsk->thread.fpu);
 }
 
+/*
+ * Software FPU state helpers. Careful: these need to
+ * be preemption protection *and* they need to be
+ * properly paired with the CR0.TS changes!
+ */
+static inline int __thread_has_fpu(struct task_struct *tsk)
+{
+	return tsk->thread.has_fpu;
+}
+
+/* Must be paired with an 'stts' after! */
+static inline void __thread_clear_has_fpu(struct task_struct *tsk)
+{
+	tsk->thread.has_fpu = 0;
+}
+
+/* Must be paired with a 'clts' before! */
+static inline void __thread_set_has_fpu(struct task_struct *tsk)
+{
+	tsk->thread.has_fpu = 1;
+}
+
+/*
+ * Encapsulate the CR0.TS handling together with the
+ * software flag.
+ *
+ * These generally need preemption protection to work,
+ * do try to avoid using these on their own.
+ */
+static inline void __thread_fpu_end(struct task_struct *tsk)
+{
+	__thread_clear_has_fpu(tsk);
+	stts();
+}
+
+static inline void __thread_fpu_begin(struct task_struct *tsk)
+{
+	clts();
+	__thread_set_has_fpu(tsk);
+}
+
+/*
+ * FPU state switching for scheduling.
+ *
+ * This is a two-stage process:
+ *
+ *  - switch_fpu_prepare() saves the old state and
+ *    sets the new state of the CR0.TS bit. This is
+ *    done within the context of the old process.
+ *
+ *  - switch_fpu_finish() restores the new state as
+ *    necessary.
+ */
+typedef struct { int preload; } fpu_switch_t;
+
+/*
+ * FIXME! We could do a totally lazy restore, but we need to
+ * add a per-cpu "this was the task that last touched the FPU
+ * on this CPU" variable, and the task needs to have a "I last
+ * touched the FPU on this CPU" and check them.
+ *
+ * We don't do that yet, so "fpu_lazy_restore()" always returns
+ * false, but some day..
+ */
+#define fpu_lazy_restore(tsk) (0)
+#define fpu_lazy_state_intact(tsk) do { } while (0)
+
+static inline fpu_switch_t switch_fpu_prepare(struct task_struct *old, struct task_struct *new)
+{
+	fpu_switch_t fpu;
+
+	fpu.preload = tsk_used_math(new) && new->fpu_counter > 5;
+	if (__thread_has_fpu(old)) {
+		if (__save_init_fpu(old))
+			fpu_lazy_state_intact(old);
+		__thread_clear_has_fpu(old);
+		old->fpu_counter++;
+
+		/* Don't change CR0.TS if we just switch! */
+		if (fpu.preload) {
+			__thread_set_has_fpu(new);
+			prefetch(new->thread.fpu.state);
+		} else
+			stts();
+	} else {
+		old->fpu_counter = 0;
+		if (fpu.preload) {
+			if (fpu_lazy_restore(new))
+				fpu.preload = 0;
+			else
+				prefetch(new->thread.fpu.state);
+			__thread_fpu_begin(new);
+		}
+	}
+	return fpu;
+}
+
+/*
+ * By the time this gets called, we've already cleared CR0.TS and
+ * given the process the FPU if we are going to preload the FPU
+ * state - all we need to do is to conditionally restore the register
+ * state itself.
+ */
+static inline void switch_fpu_finish(struct task_struct *new, fpu_switch_t fpu)
+{
+	if (fpu.preload)
+		__math_state_restore(new);
+}
+
 /*
  * Signal frame handlers...
  */
 extern int save_i387_xstate(void __user *buf);
 extern int restore_i387_xstate(void __user *buf);
 
-static inline void __unlazy_fpu(struct task_struct *tsk)
-{
-	if (task_thread_info(tsk)->status & TS_USEDFPU) {
-		__save_init_fpu(tsk);
-		stts();
-	} else
-		tsk->fpu_counter = 0;
-}
-
 static inline void __clear_fpu(struct task_struct *tsk)
 {
-	if (task_thread_info(tsk)->status & TS_USEDFPU) {
+	if (__thread_has_fpu(tsk)) {
 		/* Ignore delayed exceptions from user space */
 		asm volatile("1: fwait\n"
 			     "2:\n"
 			     _ASM_EXTABLE(1b, 2b));
-		task_thread_info(tsk)->status &= ~TS_USEDFPU;
-		stts();
+		__thread_fpu_end(tsk);
 	}
 }
 
+/*
+ * Were we in an interrupt that interrupted kernel mode?
+ *
+ * We can do a kernel_fpu_begin/end() pair *ONLY* if that
+ * pair does nothing at all: the thread must not have fpu (so
+ * that we don't try to save the FPU state), and TS must
+ * be set (so that the clts/stts pair does nothing that is
+ * visible in the interrupted kernel thread).
+ */
+static inline bool interrupted_kernel_fpu_idle(void)
+{
+	return !__thread_has_fpu(current) &&
+		(read_cr0() & X86_CR0_TS);
+}
+
+/*
+ * Were we in user mode (or vm86 mode) when we were
+ * interrupted?
+ *
+ * Doing kernel_fpu_begin/end() is ok if we are running
+ * in an interrupt context from user mode - we'll just
+ * save the FPU state as required.
+ */
+static inline bool interrupted_user_mode(void)
+{
+	struct pt_regs *regs = get_irq_regs();
+	return regs && user_mode_vm(regs);
+}
+
+/*
+ * Can we use the FPU in kernel mode with the
+ * whole "kernel_fpu_begin/end()" sequence?
+ *
+ * It's always ok in process context (ie "not interrupt")
+ * but it is sometimes ok even from an irq.
+ */
+static inline bool irq_fpu_usable(void)
+{
+	return !in_interrupt() ||
+		interrupted_user_mode() ||
+		interrupted_kernel_fpu_idle();
+}
+
 static inline void kernel_fpu_begin(void)
 {
-	struct thread_info *me = current_thread_info();
+	struct task_struct *me = current;
+
+	WARN_ON_ONCE(!irq_fpu_usable());
 	preempt_disable();
-	if (me->status & TS_USEDFPU)
-		__save_init_fpu(me->task);
-	else
+	if (__thread_has_fpu(me)) {
+		__save_init_fpu(me);
+		__thread_clear_has_fpu(me);
+		/* We do 'stts()' in kernel_fpu_end() */
+	} else
 		clts();
 }
 
@@ -323,14 +461,6 @@ static inline void kernel_fpu_end(void)
 	preempt_enable();
 }
 
-static inline bool irq_fpu_usable(void)
-{
-	struct pt_regs *regs;
-
-	return !in_interrupt() || !(regs = get_irq_regs()) || \
-		user_mode(regs) || (read_cr0() & X86_CR0_TS);
-}
-
 /*
  * Some instructions like VIA's padlock instructions generate a spurious
  * DNA fault but don't modify SSE registers. And these instructions
@@ -362,21 +492,65 @@ static inline void irq_ts_restore(int TS_state)
 		stts();
 }
 
+/*
+ * The question "does this thread have fpu access?"
+ * is slightly racy, since preemption could come in
+ * and revoke it immediately after the test.
+ *
+ * However, even in that very unlikely scenario,
+ * we can just assume we have FPU access - typically
+ * to save the FP state - we'll just take a #NM
+ * fault and get the FPU access back.
+ *
+ * The actual user_fpu_begin/end() functions
+ * need to be preemption-safe, though.
+ *
+ * NOTE! user_fpu_end() must be used only after you
+ * have saved the FP state, and user_fpu_begin() must
+ * be used only immediately before restoring it.
+ * These functions do not do any save/restore on
+ * their own.
+ */
+static inline int user_has_fpu(void)
+{
+	return __thread_has_fpu(current);
+}
+
+static inline void user_fpu_end(void)
+{
+	preempt_disable();
+	__thread_fpu_end(current);
+	preempt_enable();
+}
+
+static inline void user_fpu_begin(void)
+{
+	preempt_disable();
+	if (!user_has_fpu())
+		__thread_fpu_begin(current);
+	preempt_enable();
+}
+
 /*
  * These disable preemption on their own and are safe
  */
 static inline void save_init_fpu(struct task_struct *tsk)
 {
+	WARN_ON_ONCE(!__thread_has_fpu(tsk));
 	preempt_disable();
 	__save_init_fpu(tsk);
-	stts();
+	__thread_fpu_end(tsk);
 	preempt_enable();
 }
 
 static inline void unlazy_fpu(struct task_struct *tsk)
 {
 	preempt_disable();
-	__unlazy_fpu(tsk);
+	if (__thread_has_fpu(tsk)) {
+		__save_init_fpu(tsk);
+		__thread_fpu_end(tsk);
+	} else
+		tsk->fpu_counter = 0;
 	preempt_enable();
 }
 

arch/x86/include/asm/processor.h

@@ -456,6 +456,7 @@ struct thread_struct {
 	unsigned long		trap_no;
 	unsigned long		error_code;
 	/* floating point and extended processor state */
+	unsigned long		has_fpu;
 	struct fpu		fpu;
 #ifdef CONFIG_X86_32
 	/* Virtual 86 mode info */

arch/x86/include/asm/thread_info.h

@@ -242,8 +242,6 @@ static inline struct thread_info *current_thread_info(void)
  * ever touches our thread-synchronous status, so we don't
  * have to worry about atomic accesses.
  */
-#define TS_USEDFPU		0x0001	/* FPU was used by this task
-					   this quantum (SMP) */
 #define TS_COMPAT		0x0002	/* 32bit syscall active (64BIT)*/
 #define TS_POLLING		0x0004	/* idle task polling need_resched,
 					   skip sending interrupt */

arch/x86/kernel/process_32.c

@@ -297,22 +297,11 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 				 *next = &next_p->thread;
 	int cpu = smp_processor_id();
 	struct tss_struct *tss = &per_cpu(init_tss, cpu);
-	bool preload_fpu;
+	fpu_switch_t fpu;
 
 	/* never put a printk in __switch_to... printk() calls wake_up*() indirectly */
 
-	/*
-	 * If the task has used fpu the last 5 timeslices, just do a full
-	 * restore of the math state immediately to avoid the trap; the
-	 * chances of needing FPU soon are obviously high now
-	 */
-	preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5;
-
-	__unlazy_fpu(prev_p);
-
-	/* we're going to use this soon, after a few expensive things */
-	if (preload_fpu)
-		prefetch(next->fpu.state);
+	fpu = switch_fpu_prepare(prev_p, next_p);
 
 	/*
 	 * Reload esp0.
@@ -352,11 +341,6 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 		     task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT))
 		__switch_to_xtra(prev_p, next_p, tss);
 
-	/* If we're going to preload the fpu context, make sure clts
-	   is run while we're batching the cpu state updates. */
-	if (preload_fpu)
-		clts();
-
 	/*
 	 * Leave lazy mode, flushing any hypercalls made here.
 	 * This must be done before restoring TLS segments so
@@ -366,15 +350,14 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 	 */
 	arch_end_context_switch(next_p);
 
-	if (preload_fpu)
-		__math_state_restore();
-
 	/*
 	 * Restore %gs if needed (which is common)
 	 */
 	if (prev->gs | next->gs)
 		lazy_load_gs(next->gs);
 
+	switch_fpu_finish(next_p, fpu);
+
 	percpu_write(current_task, next_p);
 
 	return prev_p;

arch/x86/kernel/process_64.c

@@ -381,18 +381,9 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 	int cpu = smp_processor_id();
 	struct tss_struct *tss = &per_cpu(init_tss, cpu);
 	unsigned fsindex, gsindex;
-	bool preload_fpu;
+	fpu_switch_t fpu;
 
-	/*
-	 * If the task has used fpu the last 5 timeslices, just do a full
-	 * restore of the math state immediately to avoid the trap; the
-	 * chances of needing FPU soon are obviously high now
-	 */
-	preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5;
-
-	/* we're going to use this soon, after a few expensive things */
-	if (preload_fpu)
-		prefetch(next->fpu.state);
+	fpu = switch_fpu_prepare(prev_p, next_p);
 
 	/*
 	 * Reload esp0, LDT and the page table pointer:
@@ -422,13 +413,6 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 
 	load_TLS(next, cpu);
 
-	/* Must be after DS reload */
-	__unlazy_fpu(prev_p);
-
-	/* Make sure cpu is ready for new context */
-	if (preload_fpu)
-		clts();
-
 	/*
 	 * Leave lazy mode, flushing any hypercalls made here.
 	 * This must be done before restoring TLS segments so
@@ -469,6 +453,8 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 		wrmsrl(MSR_KERNEL_GS_BASE, next->gs);
 	prev->gsindex = gsindex;
 
+	switch_fpu_finish(next_p, fpu);
+
 	/*
 	 * Switch the PDA and FPU contexts.
 	 */
@@ -487,13 +473,6 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 		     task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
 		__switch_to_xtra(prev_p, next_p, tss);
 
-	/*
-	 * Preload the FPU context, now that we've determined that the
-	 * task is likely to be using it. 
-	 */
-	if (preload_fpu)
-		__math_state_restore();
-
 	return prev_p;
 }
 

arch/x86/kernel/traps.c

@@ -562,25 +562,34 @@ asmlinkage void __attribute__((weak)) smp_threshold_interrupt(void)
 }
 
 /*
- * __math_state_restore assumes that cr0.TS is already clear and the
- * fpu state is all ready for use.  Used during context switch.
+ * This gets called with the process already owning the
+ * FPU state, and with CR0.TS cleared. It just needs to
+ * restore the FPU register state.
  */
-void __math_state_restore(void)
+void __math_state_restore(struct task_struct *tsk)
 {
-	struct thread_info *thread = current_thread_info();
-	struct task_struct *tsk = thread->task;
+	/* We need a safe address that is cheap to find and that is already
+	   in L1. We've just brought in "tsk->thread.has_fpu", so use that */
+#define safe_address (tsk->thread.has_fpu)
+
+	/* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception
+	   is pending.  Clear the x87 state here by setting it to fixed
+	   values. safe_address is a random variable that should be in L1 */
+	alternative_input(
+		ASM_NOP8 ASM_NOP2,
+		"emms\n\t"	  	/* clear stack tags */
+		"fildl %P[addr]",	/* set F?P to defined value */
+		X86_FEATURE_FXSAVE_LEAK,
+		[addr] "m" (safe_address));
 
 	/*
 	 * Paranoid restore. send a SIGSEGV if we fail to restore the state.
 	 */
 	if (unlikely(restore_fpu_checking(tsk))) {
-		stts();
+		__thread_fpu_end(tsk);
 		force_sig(SIGSEGV, tsk);
 		return;
 	}
-
-	thread->status |= TS_USEDFPU;	/* So we fnsave on switch_to() */
-	tsk->fpu_counter++;
 }
 
 /*
@@ -590,13 +599,12 @@ void __math_state_restore(void)
  * Careful.. There are problems with IBM-designed IRQ13 behaviour.
  * Don't touch unless you *really* know how it works.
  *
- * Must be called with kernel preemption disabled (in this case,
- * local interrupts are disabled at the call-site in entry.S).
+ * Must be called with kernel preemption disabled (eg with local
+ * local interrupts as in the case of do_device_not_available).
  */
-asmlinkage void math_state_restore(void)
+void math_state_restore(void)
 {
-	struct thread_info *thread = current_thread_info();
-	struct task_struct *tsk = thread->task;
+	struct task_struct *tsk = current;
 
 	if (!tsk_used_math(tsk)) {
 		local_irq_enable();
@@ -613,9 +621,10 @@ asmlinkage void math_state_restore(void)
 		local_irq_disable();
 	}
 
-	clts();				/* Allow maths ops (or we recurse) */
+	__thread_fpu_begin(tsk);
+	__math_state_restore(tsk);
 
-	__math_state_restore();
+	tsk->fpu_counter++;
 }
 EXPORT_SYMBOL_GPL(math_state_restore);
 

arch/x86/kernel/xsave.c

@@ -47,7 +47,7 @@ void __sanitize_i387_state(struct task_struct *tsk)
 	if (!fx)
 		return;
 
-	BUG_ON(task_thread_info(tsk)->status & TS_USEDFPU);
+	BUG_ON(__thread_has_fpu(tsk));
 
 	xstate_bv = tsk->thread.fpu.state->xsave.xsave_hdr.xstate_bv;
 
@@ -168,7 +168,7 @@ int save_i387_xstate(void __user *buf)
 	if (!used_math())
 		return 0;
 
-	if (task_thread_info(tsk)->status & TS_USEDFPU) {
+	if (user_has_fpu()) {
 		if (use_xsave())
 			err = xsave_user(buf);
 		else
@@ -176,8 +176,7 @@ int save_i387_xstate(void __user *buf)
 
 		if (err)
 			return err;
-		task_thread_info(tsk)->status &= ~TS_USEDFPU;
-		stts();
+		user_fpu_end();
 	} else {
 		sanitize_i387_state(tsk);
 		if (__copy_to_user(buf, &tsk->thread.fpu.state->fxsave,
@@ -292,10 +291,7 @@ int restore_i387_xstate(void __user *buf)
 			return err;
 	}
 
-	if (!(task_thread_info(current)->status & TS_USEDFPU)) {
-		clts();
-		task_thread_info(current)->status |= TS_USEDFPU;
-	}
+	user_fpu_begin();
 	if (use_xsave())
 		err = restore_user_xstate(buf);
 	else

arch/x86/kvm/vmx.c

@@ -1456,7 +1456,7 @@ static void __vmx_load_host_state(struct vcpu_vmx *vmx)
 #ifdef CONFIG_X86_64
 	wrmsrl(MSR_KERNEL_GS_BASE, vmx->msr_host_kernel_gs_base);
 #endif
-	if (current_thread_info()->status & TS_USEDFPU)
+	if (__thread_has_fpu(current))
 		clts();
 	load_gdt(&__get_cpu_var(host_gdt));
 }