2 * Copyright (C) 1994 Linus Torvalds
4 * Pentium III FXSR, SSE support
5 * General FPU state handling cleanups
6 * Gareth Hughes <gareth@valinux.com>, May 2000
8 #include <asm/fpu/internal.h>
9 #include <asm/fpu/regset.h>
10 #include <asm/fpu/signal.h>
11 #include <asm/traps.h>
13 #include <linux/hardirq.h>
16 * Represents the initial FPU state. It's mostly (but not completely) zeroes,
17 * depending on the FPU hardware format:
19 union fpregs_state init_fpstate __read_mostly
;
22 * Track whether the kernel is using the FPU state
27 * - by IRQ context code to potentially use the FPU
30 * - to debug kernel_fpu_begin()/end() correctness
32 static DEFINE_PER_CPU(bool, in_kernel_fpu
);
35 * Track which context is using the FPU on the CPU:
37 DEFINE_PER_CPU(struct fpu
*, fpu_fpregs_owner_ctx
);
39 static void kernel_fpu_disable(void)
41 WARN_ON(this_cpu_read(in_kernel_fpu
));
42 this_cpu_write(in_kernel_fpu
, true);
45 static void kernel_fpu_enable(void)
47 WARN_ON_ONCE(!this_cpu_read(in_kernel_fpu
));
48 this_cpu_write(in_kernel_fpu
, false);
51 static bool kernel_fpu_disabled(void)
53 return this_cpu_read(in_kernel_fpu
);
57 * Were we in an interrupt that interrupted kernel mode?
59 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
60 * pair does nothing at all: the thread must not have fpu (so
61 * that we don't try to save the FPU state), and TS must
62 * be set (so that the clts/stts pair does nothing that is
63 * visible in the interrupted kernel thread).
65 * Except for the eagerfpu case when we return true; in the likely case
66 * the thread has FPU but we are not going to set/clear TS.
68 static bool interrupted_kernel_fpu_idle(void)
70 if (kernel_fpu_disabled())
76 return !current
->thread
.fpu
.fpregs_active
&& (read_cr0() & X86_CR0_TS
);
80 * Were we in user mode (or vm86 mode) when we were
83 * Doing kernel_fpu_begin/end() is ok if we are running
84 * in an interrupt context from user mode - we'll just
85 * save the FPU state as required.
87 static bool interrupted_user_mode(void)
89 struct pt_regs
*regs
= get_irq_regs();
90 return regs
&& user_mode(regs
);
94 * Can we use the FPU in kernel mode with the
95 * whole "kernel_fpu_begin/end()" sequence?
97 * It's always ok in process context (ie "not interrupt")
98 * but it is sometimes ok even from an irq.
100 bool irq_fpu_usable(void)
102 return !in_interrupt() ||
103 interrupted_user_mode() ||
104 interrupted_kernel_fpu_idle();
106 EXPORT_SYMBOL(irq_fpu_usable
);
108 void __kernel_fpu_begin(void)
110 struct fpu
*fpu
= ¤t
->thread
.fpu
;
112 WARN_ON_ONCE(!irq_fpu_usable());
114 kernel_fpu_disable();
116 if (fpu
->fpregs_active
) {
117 copy_fpregs_to_fpstate(fpu
);
119 this_cpu_write(fpu_fpregs_owner_ctx
, NULL
);
120 __fpregs_activate_hw();
123 EXPORT_SYMBOL(__kernel_fpu_begin
);
125 void __kernel_fpu_end(void)
127 struct fpu
*fpu
= ¤t
->thread
.fpu
;
129 if (fpu
->fpregs_active
) {
130 if (WARN_ON(copy_fpstate_to_fpregs(fpu
)))
133 __fpregs_deactivate_hw();
138 EXPORT_SYMBOL(__kernel_fpu_end
);
140 void kernel_fpu_begin(void)
143 __kernel_fpu_begin();
145 EXPORT_SYMBOL_GPL(kernel_fpu_begin
);
147 void kernel_fpu_end(void)
152 EXPORT_SYMBOL_GPL(kernel_fpu_end
);
155 * CR0::TS save/restore functions:
157 int irq_ts_save(void)
160 * If in process context and not atomic, we can take a spurious DNA fault.
161 * Otherwise, doing clts() in process context requires disabling preemption
162 * or some heavy lifting like kernel_fpu_begin()
167 if (read_cr0() & X86_CR0_TS
) {
174 EXPORT_SYMBOL_GPL(irq_ts_save
);
176 void irq_ts_restore(int TS_state
)
181 EXPORT_SYMBOL_GPL(irq_ts_restore
);
184 * Save the FPU state (mark it for reload if necessary):
186 * This only ever gets called for the current task.
188 void fpu__save(struct fpu
*fpu
)
190 WARN_ON(fpu
!= ¤t
->thread
.fpu
);
193 if (fpu
->fpregs_active
) {
194 if (!copy_fpregs_to_fpstate(fpu
))
195 fpregs_deactivate(fpu
);
199 EXPORT_SYMBOL_GPL(fpu__save
);
202 * Legacy x87 fpstate state init:
204 static inline void fpstate_init_fstate(struct fregs_state
*fp
)
206 fp
->cwd
= 0xffff037fu
;
207 fp
->swd
= 0xffff0000u
;
208 fp
->twd
= 0xffffffffu
;
209 fp
->fos
= 0xffff0000u
;
212 void fpstate_init(union fpregs_state
*state
)
215 fpstate_init_soft(&state
->soft
);
219 memset(state
, 0, xstate_size
);
222 fpstate_init_fxstate(&state
->fxsave
);
224 fpstate_init_fstate(&state
->fsave
);
226 EXPORT_SYMBOL_GPL(fpstate_init
);
229 * Copy the current task's FPU state to a new task's FPU context.
231 * In both the 'eager' and the 'lazy' case we save hardware registers
232 * directly to the destination buffer.
234 static void fpu_copy(struct fpu
*dst_fpu
, struct fpu
*src_fpu
)
236 WARN_ON(src_fpu
!= ¤t
->thread
.fpu
);
239 * Don't let 'init optimized' areas of the XSAVE area
240 * leak into the child task:
243 memset(&dst_fpu
->state
.xsave
, 0, xstate_size
);
246 * Save current FPU registers directly into the child
247 * FPU context, without any memory-to-memory copying.
249 * If the FPU context got destroyed in the process (FNSAVE
250 * done on old CPUs) then copy it back into the source
251 * context and mark the current task for lazy restore.
253 * We have to do all this with preemption disabled,
254 * mostly because of the FNSAVE case, because in that
255 * case we must not allow preemption in the window
256 * between the FNSAVE and us marking the context lazy.
258 * It shouldn't be an issue as even FNSAVE is plenty
259 * fast in terms of critical section length.
262 if (!copy_fpregs_to_fpstate(dst_fpu
)) {
263 memcpy(&src_fpu
->state
, &dst_fpu
->state
, xstate_size
);
264 fpregs_deactivate(src_fpu
);
269 int fpu__copy(struct fpu
*dst_fpu
, struct fpu
*src_fpu
)
271 dst_fpu
->counter
= 0;
272 dst_fpu
->fpregs_active
= 0;
273 dst_fpu
->last_cpu
= -1;
275 if (src_fpu
->fpstate_active
)
276 fpu_copy(dst_fpu
, src_fpu
);
282 * Activate the current task's in-memory FPU context,
283 * if it has not been used before:
285 void fpu__activate_curr(struct fpu
*fpu
)
287 WARN_ON_ONCE(fpu
!= ¤t
->thread
.fpu
);
289 if (!fpu
->fpstate_active
) {
290 fpstate_init(&fpu
->state
);
292 /* Safe to do for the current task: */
293 fpu
->fpstate_active
= 1;
296 EXPORT_SYMBOL_GPL(fpu__activate_curr
);
299 * This function must be called before we modify a stopped child's
302 * If the child has not used the FPU before then initialize its
305 * If the child has used the FPU before then unlazy it.
307 * [ After this function call, after registers in the fpstate are
308 * modified and the child task has woken up, the child task will
309 * restore the modified FPU state from the modified context. If we
310 * didn't clear its lazy status here then the lazy in-registers
311 * state pending on its former CPU could be restored, corrupting
312 * the modifications. ]
314 * This function is also called before we read a stopped child's
315 * FPU state - to make sure it's initialized if the child has
316 * no active FPU state.
318 * TODO: A future optimization would be to skip the unlazying in
319 * the read-only case, it's not strictly necessary for
320 * read-only access to the context.
322 void fpu__activate_stopped(struct fpu
*child_fpu
)
324 WARN_ON_ONCE(child_fpu
== ¤t
->thread
.fpu
);
326 if (child_fpu
->fpstate_active
) {
327 child_fpu
->last_cpu
= -1;
329 fpstate_init(&child_fpu
->state
);
331 /* Safe to do for stopped child tasks: */
332 child_fpu
->fpstate_active
= 1;
337 * 'fpu__restore()' is called to copy FPU registers from
338 * the FPU fpstate to the live hw registers and to activate
339 * access to the hardware registers, so that FPU instructions
340 * can be used afterwards.
342 * Must be called with kernel preemption disabled (for example
343 * with local interrupts disabled, as it is in the case of
344 * do_device_not_available()).
346 void fpu__restore(void)
348 struct task_struct
*tsk
= current
;
349 struct fpu
*fpu
= &tsk
->thread
.fpu
;
351 fpu__activate_curr(fpu
);
353 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
354 kernel_fpu_disable();
355 fpregs_activate(fpu
);
356 if (unlikely(copy_fpstate_to_fpregs(fpu
))) {
358 force_sig_info(SIGSEGV
, SEND_SIG_PRIV
, tsk
);
360 tsk
->thread
.fpu
.counter
++;
364 EXPORT_SYMBOL_GPL(fpu__restore
);
367 * Drops current FPU state: deactivates the fpregs and
368 * the fpstate. NOTE: it still leaves previous contents
369 * in the fpregs in the eager-FPU case.
371 * This function can be used in cases where we know that
372 * a state-restore is coming: either an explicit one,
375 void fpu__drop(struct fpu
*fpu
)
380 if (fpu
->fpregs_active
) {
381 /* Ignore delayed exceptions from user space */
382 asm volatile("1: fwait\n"
384 _ASM_EXTABLE(1b
, 2b
));
385 fpregs_deactivate(fpu
);
388 fpu
->fpstate_active
= 0;
394 * Clear FPU registers by setting them up from
397 static inline void copy_init_fpstate_to_fpregs(void)
400 copy_kernel_to_xregs(&init_fpstate
.xsave
, -1);
402 copy_kernel_to_fxregs(&init_fpstate
.fxsave
);
406 * Clear the FPU state back to init state.
408 * Called by sys_execve(), by the signal handler code and by various
411 void fpu__clear(struct fpu
*fpu
)
413 WARN_ON_ONCE(fpu
!= ¤t
->thread
.fpu
); /* Almost certainly an anomaly */
415 if (!use_eager_fpu()) {
416 /* FPU state will be reallocated lazily at the first use. */
419 if (!fpu
->fpstate_active
) {
420 fpu__activate_curr(fpu
);
423 copy_init_fpstate_to_fpregs();
428 * x87 math exception handling:
431 static inline unsigned short get_fpu_cwd(struct fpu
*fpu
)
434 return fpu
->state
.fxsave
.cwd
;
436 return (unsigned short)fpu
->state
.fsave
.cwd
;
440 static inline unsigned short get_fpu_swd(struct fpu
*fpu
)
443 return fpu
->state
.fxsave
.swd
;
445 return (unsigned short)fpu
->state
.fsave
.swd
;
449 static inline unsigned short get_fpu_mxcsr(struct fpu
*fpu
)
452 return fpu
->state
.fxsave
.mxcsr
;
454 return MXCSR_DEFAULT
;
458 int fpu__exception_code(struct fpu
*fpu
, int trap_nr
)
462 if (trap_nr
== X86_TRAP_MF
) {
463 unsigned short cwd
, swd
;
465 * (~cwd & swd) will mask out exceptions that are not set to unmasked
466 * status. 0x3f is the exception bits in these regs, 0x200 is the
467 * C1 reg you need in case of a stack fault, 0x040 is the stack
468 * fault bit. We should only be taking one exception at a time,
469 * so if this combination doesn't produce any single exception,
470 * then we have a bad program that isn't synchronizing its FPU usage
471 * and it will suffer the consequences since we won't be able to
472 * fully reproduce the context of the exception
474 cwd
= get_fpu_cwd(fpu
);
475 swd
= get_fpu_swd(fpu
);
480 * The SIMD FPU exceptions are handled a little differently, as there
481 * is only a single status/control register. Thus, to determine which
482 * unmasked exception was caught we must mask the exception mask bits
483 * at 0x1f80, and then use these to mask the exception bits at 0x3f.
485 unsigned short mxcsr
= get_fpu_mxcsr(fpu
);
486 err
= ~(mxcsr
>> 7) & mxcsr
;
489 if (err
& 0x001) { /* Invalid op */
491 * swd & 0x240 == 0x040: Stack Underflow
492 * swd & 0x240 == 0x240: Stack Overflow
493 * User must clear the SF bit (0x40) if set
496 } else if (err
& 0x004) { /* Divide by Zero */
498 } else if (err
& 0x008) { /* Overflow */
500 } else if (err
& 0x012) { /* Denormal, Underflow */
502 } else if (err
& 0x020) { /* Precision */
507 * If we're using IRQ 13, or supposedly even some trap
508 * X86_TRAP_MF implementations, it's possible
509 * we get a spurious trap, which is not an error.