xref: /linux/arch/x86/kernel/fpu/core.c (revision 6f0310a126f1a46cac366327751bb7eb8941bdde)
1 /*
2  *  Copyright (C) 1994 Linus Torvalds
3  *
4  *  Pentium III FXSR, SSE support
5  *  General FPU state handling cleanups
6  *	Gareth Hughes <gareth@valinux.com>, May 2000
7  */
8 #include <asm/fpu/internal.h>
9 #include <asm/fpu/regset.h>
10 #include <asm/fpu/signal.h>
11 #include <asm/fpu/types.h>
12 #include <asm/fpu/xstate.h>
13 #include <asm/traps.h>
14 
15 #include <linux/hardirq.h>
16 #include <linux/pkeys.h>
17 
18 #define CREATE_TRACE_POINTS
19 #include <asm/trace/fpu.h>
20 
21 /*
22  * Represents the initial FPU state. It's mostly (but not completely) zeroes,
23  * depending on the FPU hardware format:
24  */
25 union fpregs_state init_fpstate __read_mostly;
26 
27 /*
28  * Track whether the kernel is using the FPU state
29  * currently.
30  *
31  * This flag is used:
32  *
33  *   - by IRQ context code to potentially use the FPU
34  *     if it's unused.
35  *
36  *   - to debug kernel_fpu_begin()/end() correctness
37  */
38 static DEFINE_PER_CPU(bool, in_kernel_fpu);
39 
40 /*
41  * Track which context is using the FPU on the CPU:
42  */
43 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
44 
45 static void kernel_fpu_disable(void)
46 {
47 	WARN_ON_FPU(this_cpu_read(in_kernel_fpu));
48 	this_cpu_write(in_kernel_fpu, true);
49 }
50 
51 static void kernel_fpu_enable(void)
52 {
53 	WARN_ON_FPU(!this_cpu_read(in_kernel_fpu));
54 	this_cpu_write(in_kernel_fpu, false);
55 }
56 
57 static bool kernel_fpu_disabled(void)
58 {
59 	return this_cpu_read(in_kernel_fpu);
60 }
61 
62 static bool interrupted_kernel_fpu_idle(void)
63 {
64 	return !kernel_fpu_disabled();
65 }
66 
67 /*
68  * Were we in user mode (or vm86 mode) when we were
69  * interrupted?
70  *
71  * Doing kernel_fpu_begin/end() is ok if we are running
72  * in an interrupt context from user mode - we'll just
73  * save the FPU state as required.
74  */
75 static bool interrupted_user_mode(void)
76 {
77 	struct pt_regs *regs = get_irq_regs();
78 	return regs && user_mode(regs);
79 }
80 
81 /*
82  * Can we use the FPU in kernel mode with the
83  * whole "kernel_fpu_begin/end()" sequence?
84  *
85  * It's always ok in process context (ie "not interrupt")
86  * but it is sometimes ok even from an irq.
87  */
88 bool irq_fpu_usable(void)
89 {
90 	return !in_interrupt() ||
91 		interrupted_user_mode() ||
92 		interrupted_kernel_fpu_idle();
93 }
94 EXPORT_SYMBOL(irq_fpu_usable);
95 
96 void __kernel_fpu_begin(void)
97 {
98 	struct fpu *fpu = &current->thread.fpu;
99 
100 	WARN_ON_FPU(!irq_fpu_usable());
101 
102 	kernel_fpu_disable();
103 
104 	if (fpu->fpregs_active) {
105 		/*
106 		 * Ignore return value -- we don't care if reg state
107 		 * is clobbered.
108 		 */
109 		copy_fpregs_to_fpstate(fpu);
110 	} else {
111 		__cpu_invalidate_fpregs_state();
112 	}
113 }
114 EXPORT_SYMBOL(__kernel_fpu_begin);
115 
116 void __kernel_fpu_end(void)
117 {
118 	struct fpu *fpu = &current->thread.fpu;
119 
120 	if (fpu->fpregs_active)
121 		copy_kernel_to_fpregs(&fpu->state);
122 
123 	kernel_fpu_enable();
124 }
125 EXPORT_SYMBOL(__kernel_fpu_end);
126 
127 void kernel_fpu_begin(void)
128 {
129 	preempt_disable();
130 	__kernel_fpu_begin();
131 }
132 EXPORT_SYMBOL_GPL(kernel_fpu_begin);
133 
134 void kernel_fpu_end(void)
135 {
136 	__kernel_fpu_end();
137 	preempt_enable();
138 }
139 EXPORT_SYMBOL_GPL(kernel_fpu_end);
140 
141 /*
142  * Save the FPU state (mark it for reload if necessary):
143  *
144  * This only ever gets called for the current task.
145  */
146 void fpu__save(struct fpu *fpu)
147 {
148 	WARN_ON_FPU(fpu != &current->thread.fpu);
149 
150 	preempt_disable();
151 	trace_x86_fpu_before_save(fpu);
152 	if (fpu->fpregs_active) {
153 		if (!copy_fpregs_to_fpstate(fpu)) {
154 			copy_kernel_to_fpregs(&fpu->state);
155 		}
156 	}
157 	trace_x86_fpu_after_save(fpu);
158 	preempt_enable();
159 }
160 EXPORT_SYMBOL_GPL(fpu__save);
161 
162 /*
163  * Legacy x87 fpstate state init:
164  */
165 static inline void fpstate_init_fstate(struct fregs_state *fp)
166 {
167 	fp->cwd = 0xffff037fu;
168 	fp->swd = 0xffff0000u;
169 	fp->twd = 0xffffffffu;
170 	fp->fos = 0xffff0000u;
171 }
172 
173 void fpstate_init(union fpregs_state *state)
174 {
175 	if (!static_cpu_has(X86_FEATURE_FPU)) {
176 		fpstate_init_soft(&state->soft);
177 		return;
178 	}
179 
180 	memset(state, 0, fpu_kernel_xstate_size);
181 
182 	/*
183 	 * XRSTORS requires that this bit is set in xcomp_bv, or
184 	 * it will #GP. Make sure it is replaced after the memset().
185 	 */
186 	if (static_cpu_has(X86_FEATURE_XSAVES))
187 		state->xsave.header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT |
188 					       xfeatures_mask;
189 
190 	if (static_cpu_has(X86_FEATURE_FXSR))
191 		fpstate_init_fxstate(&state->fxsave);
192 	else
193 		fpstate_init_fstate(&state->fsave);
194 }
195 EXPORT_SYMBOL_GPL(fpstate_init);
196 
197 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
198 {
199 	dst_fpu->fpregs_active = 0;
200 	dst_fpu->last_cpu = -1;
201 
202 	if (!src_fpu->fpstate_active || !static_cpu_has(X86_FEATURE_FPU))
203 		return 0;
204 
205 	WARN_ON_FPU(src_fpu != &current->thread.fpu);
206 
207 	/*
208 	 * Don't let 'init optimized' areas of the XSAVE area
209 	 * leak into the child task:
210 	 */
211 	memset(&dst_fpu->state.xsave, 0, fpu_kernel_xstate_size);
212 
213 	/*
214 	 * Save current FPU registers directly into the child
215 	 * FPU context, without any memory-to-memory copying.
216 	 * In lazy mode, if the FPU context isn't loaded into
217 	 * fpregs, CR0.TS will be set and do_device_not_available
218 	 * will load the FPU context.
219 	 *
220 	 * We have to do all this with preemption disabled,
221 	 * mostly because of the FNSAVE case, because in that
222 	 * case we must not allow preemption in the window
223 	 * between the FNSAVE and us marking the context lazy.
224 	 *
225 	 * It shouldn't be an issue as even FNSAVE is plenty
226 	 * fast in terms of critical section length.
227 	 */
228 	preempt_disable();
229 	if (!copy_fpregs_to_fpstate(dst_fpu)) {
230 		memcpy(&src_fpu->state, &dst_fpu->state,
231 		       fpu_kernel_xstate_size);
232 
233 		copy_kernel_to_fpregs(&src_fpu->state);
234 	}
235 	preempt_enable();
236 
237 	trace_x86_fpu_copy_src(src_fpu);
238 	trace_x86_fpu_copy_dst(dst_fpu);
239 
240 	return 0;
241 }
242 
243 /*
244  * Activate the current task's in-memory FPU context,
245  * if it has not been used before:
246  */
247 void fpu__activate_curr(struct fpu *fpu)
248 {
249 	WARN_ON_FPU(fpu != &current->thread.fpu);
250 
251 	if (!fpu->fpstate_active) {
252 		fpstate_init(&fpu->state);
253 		trace_x86_fpu_init_state(fpu);
254 
255 		trace_x86_fpu_activate_state(fpu);
256 		/* Safe to do for the current task: */
257 		fpu->fpstate_active = 1;
258 	}
259 }
260 EXPORT_SYMBOL_GPL(fpu__activate_curr);
261 
262 /*
263  * This function must be called before we read a task's fpstate.
264  *
265  * If the task has not used the FPU before then initialize its
266  * fpstate.
267  *
268  * If the task has used the FPU before then save it.
269  */
270 void fpu__activate_fpstate_read(struct fpu *fpu)
271 {
272 	/*
273 	 * If fpregs are active (in the current CPU), then
274 	 * copy them to the fpstate:
275 	 */
276 	if (fpu->fpregs_active) {
277 		fpu__save(fpu);
278 	} else {
279 		if (!fpu->fpstate_active) {
280 			fpstate_init(&fpu->state);
281 			trace_x86_fpu_init_state(fpu);
282 
283 			trace_x86_fpu_activate_state(fpu);
284 			/* Safe to do for current and for stopped child tasks: */
285 			fpu->fpstate_active = 1;
286 		}
287 	}
288 }
289 
290 /*
291  * This function must be called before we write a task's fpstate.
292  *
293  * If the task has used the FPU before then unlazy it.
294  * If the task has not used the FPU before then initialize its fpstate.
295  *
296  * After this function call, after registers in the fpstate are
297  * modified and the child task has woken up, the child task will
298  * restore the modified FPU state from the modified context. If we
299  * didn't clear its lazy status here then the lazy in-registers
300  * state pending on its former CPU could be restored, corrupting
301  * the modifications.
302  */
303 void fpu__activate_fpstate_write(struct fpu *fpu)
304 {
305 	/*
306 	 * Only stopped child tasks can be used to modify the FPU
307 	 * state in the fpstate buffer:
308 	 */
309 	WARN_ON_FPU(fpu == &current->thread.fpu);
310 
311 	if (fpu->fpstate_active) {
312 		/* Invalidate any lazy state: */
313 		__fpu_invalidate_fpregs_state(fpu);
314 	} else {
315 		fpstate_init(&fpu->state);
316 		trace_x86_fpu_init_state(fpu);
317 
318 		trace_x86_fpu_activate_state(fpu);
319 		/* Safe to do for stopped child tasks: */
320 		fpu->fpstate_active = 1;
321 	}
322 }
323 
324 /*
325  * This function must be called before we write the current
326  * task's fpstate.
327  *
328  * This call gets the current FPU register state and moves
329  * it in to the 'fpstate'.  Preemption is disabled so that
330  * no writes to the 'fpstate' can occur from context
331  * swiches.
332  *
333  * Must be followed by a fpu__current_fpstate_write_end().
334  */
335 void fpu__current_fpstate_write_begin(void)
336 {
337 	struct fpu *fpu = &current->thread.fpu;
338 
339 	/*
340 	 * Ensure that the context-switching code does not write
341 	 * over the fpstate while we are doing our update.
342 	 */
343 	preempt_disable();
344 
345 	/*
346 	 * Move the fpregs in to the fpu's 'fpstate'.
347 	 */
348 	fpu__activate_fpstate_read(fpu);
349 
350 	/*
351 	 * The caller is about to write to 'fpu'.  Ensure that no
352 	 * CPU thinks that its fpregs match the fpstate.  This
353 	 * ensures we will not be lazy and skip a XRSTOR in the
354 	 * future.
355 	 */
356 	__fpu_invalidate_fpregs_state(fpu);
357 }
358 
359 /*
360  * This function must be paired with fpu__current_fpstate_write_begin()
361  *
362  * This will ensure that the modified fpstate gets placed back in
363  * the fpregs if necessary.
364  *
365  * Note: This function may be called whether or not an _actual_
366  * write to the fpstate occurred.
367  */
368 void fpu__current_fpstate_write_end(void)
369 {
370 	struct fpu *fpu = &current->thread.fpu;
371 
372 	/*
373 	 * 'fpu' now has an updated copy of the state, but the
374 	 * registers may still be out of date.  Update them with
375 	 * an XRSTOR if they are active.
376 	 */
377 	if (fpregs_active())
378 		copy_kernel_to_fpregs(&fpu->state);
379 
380 	/*
381 	 * Our update is done and the fpregs/fpstate are in sync
382 	 * if necessary.  Context switches can happen again.
383 	 */
384 	preempt_enable();
385 }
386 
387 /*
388  * 'fpu__restore()' is called to copy FPU registers from
389  * the FPU fpstate to the live hw registers and to activate
390  * access to the hardware registers, so that FPU instructions
391  * can be used afterwards.
392  *
393  * Must be called with kernel preemption disabled (for example
394  * with local interrupts disabled, as it is in the case of
395  * do_device_not_available()).
396  */
397 void fpu__restore(struct fpu *fpu)
398 {
399 	fpu__activate_curr(fpu);
400 
401 	/* Avoid __kernel_fpu_begin() right after fpregs_activate() */
402 	kernel_fpu_disable();
403 	trace_x86_fpu_before_restore(fpu);
404 	fpregs_activate(fpu);
405 	copy_kernel_to_fpregs(&fpu->state);
406 	trace_x86_fpu_after_restore(fpu);
407 	kernel_fpu_enable();
408 }
409 EXPORT_SYMBOL_GPL(fpu__restore);
410 
411 /*
412  * Drops current FPU state: deactivates the fpregs and
413  * the fpstate. NOTE: it still leaves previous contents
414  * in the fpregs in the eager-FPU case.
415  *
416  * This function can be used in cases where we know that
417  * a state-restore is coming: either an explicit one,
418  * or a reschedule.
419  */
420 void fpu__drop(struct fpu *fpu)
421 {
422 	preempt_disable();
423 
424 	if (fpu->fpregs_active) {
425 		/* Ignore delayed exceptions from user space */
426 		asm volatile("1: fwait\n"
427 			     "2:\n"
428 			     _ASM_EXTABLE(1b, 2b));
429 		fpregs_deactivate(fpu);
430 	}
431 
432 	fpu->fpstate_active = 0;
433 
434 	trace_x86_fpu_dropped(fpu);
435 
436 	preempt_enable();
437 }
438 
439 /*
440  * Clear FPU registers by setting them up from
441  * the init fpstate:
442  */
443 static inline void copy_init_fpstate_to_fpregs(void)
444 {
445 	if (use_xsave())
446 		copy_kernel_to_xregs(&init_fpstate.xsave, -1);
447 	else if (static_cpu_has(X86_FEATURE_FXSR))
448 		copy_kernel_to_fxregs(&init_fpstate.fxsave);
449 	else
450 		copy_kernel_to_fregs(&init_fpstate.fsave);
451 
452 	if (boot_cpu_has(X86_FEATURE_OSPKE))
453 		copy_init_pkru_to_fpregs();
454 }
455 
456 /*
457  * Clear the FPU state back to init state.
458  *
459  * Called by sys_execve(), by the signal handler code and by various
460  * error paths.
461  */
462 void fpu__clear(struct fpu *fpu)
463 {
464 	WARN_ON_FPU(fpu != &current->thread.fpu); /* Almost certainly an anomaly */
465 
466 	fpu__drop(fpu);
467 
468 	/*
469 	 * Make sure fpstate is cleared and initialized.
470 	 */
471 	if (static_cpu_has(X86_FEATURE_FPU)) {
472 		fpu__activate_curr(fpu);
473 		user_fpu_begin();
474 		copy_init_fpstate_to_fpregs();
475 	}
476 }
477 
478 /*
479  * x87 math exception handling:
480  */
481 
482 int fpu__exception_code(struct fpu *fpu, int trap_nr)
483 {
484 	int err;
485 
486 	if (trap_nr == X86_TRAP_MF) {
487 		unsigned short cwd, swd;
488 		/*
489 		 * (~cwd & swd) will mask out exceptions that are not set to unmasked
490 		 * status.  0x3f is the exception bits in these regs, 0x200 is the
491 		 * C1 reg you need in case of a stack fault, 0x040 is the stack
492 		 * fault bit.  We should only be taking one exception at a time,
493 		 * so if this combination doesn't produce any single exception,
494 		 * then we have a bad program that isn't synchronizing its FPU usage
495 		 * and it will suffer the consequences since we won't be able to
496 		 * fully reproduce the context of the exception.
497 		 */
498 		if (boot_cpu_has(X86_FEATURE_FXSR)) {
499 			cwd = fpu->state.fxsave.cwd;
500 			swd = fpu->state.fxsave.swd;
501 		} else {
502 			cwd = (unsigned short)fpu->state.fsave.cwd;
503 			swd = (unsigned short)fpu->state.fsave.swd;
504 		}
505 
506 		err = swd & ~cwd;
507 	} else {
508 		/*
509 		 * The SIMD FPU exceptions are handled a little differently, as there
510 		 * is only a single status/control register.  Thus, to determine which
511 		 * unmasked exception was caught we must mask the exception mask bits
512 		 * at 0x1f80, and then use these to mask the exception bits at 0x3f.
513 		 */
514 		unsigned short mxcsr = MXCSR_DEFAULT;
515 
516 		if (boot_cpu_has(X86_FEATURE_XMM))
517 			mxcsr = fpu->state.fxsave.mxcsr;
518 
519 		err = ~(mxcsr >> 7) & mxcsr;
520 	}
521 
522 	if (err & 0x001) {	/* Invalid op */
523 		/*
524 		 * swd & 0x240 == 0x040: Stack Underflow
525 		 * swd & 0x240 == 0x240: Stack Overflow
526 		 * User must clear the SF bit (0x40) if set
527 		 */
528 		return FPE_FLTINV;
529 	} else if (err & 0x004) { /* Divide by Zero */
530 		return FPE_FLTDIV;
531 	} else if (err & 0x008) { /* Overflow */
532 		return FPE_FLTOVF;
533 	} else if (err & 0x012) { /* Denormal, Underflow */
534 		return FPE_FLTUND;
535 	} else if (err & 0x020) { /* Precision */
536 		return FPE_FLTRES;
537 	}
538 
539 	/*
540 	 * If we're using IRQ 13, or supposedly even some trap
541 	 * X86_TRAP_MF implementations, it's possible
542 	 * we get a spurious trap, which is not an error.
543 	 */
544 	return 0;
545 }
546