]> git.ipfire.org Git - people/arne_f/kernel.git/blob - arch/x86/kernel/fpu/core.c
c8d6032f04d09a725cbec07dbe88a8841e1de76c
[people/arne_f/kernel.git] / arch / x86 / kernel / fpu / core.c
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/traps.h>
13
14 #include <linux/hardirq.h>
15 #include <linux/pkeys.h>
16
17 #define CREATE_TRACE_POINTS
18 #include <asm/trace/fpu.h>
19
20 /*
21 * Represents the initial FPU state. It's mostly (but not completely) zeroes,
22 * depending on the FPU hardware format:
23 */
24 union fpregs_state init_fpstate __read_mostly;
25
26 /*
27 * Track whether the kernel is using the FPU state
28 * currently.
29 *
30 * This flag is used:
31 *
32 * - by IRQ context code to potentially use the FPU
33 * if it's unused.
34 *
35 * - to debug kernel_fpu_begin()/end() correctness
36 */
37 static DEFINE_PER_CPU(bool, in_kernel_fpu);
38
39 /*
40 * Track which context is using the FPU on the CPU:
41 */
42 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
43
44 static void kernel_fpu_disable(void)
45 {
46 WARN_ON_FPU(this_cpu_read(in_kernel_fpu));
47 this_cpu_write(in_kernel_fpu, true);
48 }
49
50 static void kernel_fpu_enable(void)
51 {
52 WARN_ON_FPU(!this_cpu_read(in_kernel_fpu));
53 this_cpu_write(in_kernel_fpu, false);
54 }
55
56 static bool kernel_fpu_disabled(void)
57 {
58 return this_cpu_read(in_kernel_fpu);
59 }
60
61 static bool interrupted_kernel_fpu_idle(void)
62 {
63 return !kernel_fpu_disabled();
64 }
65
66 /*
67 * Were we in user mode (or vm86 mode) when we were
68 * interrupted?
69 *
70 * Doing kernel_fpu_begin/end() is ok if we are running
71 * in an interrupt context from user mode - we'll just
72 * save the FPU state as required.
73 */
74 static bool interrupted_user_mode(void)
75 {
76 struct pt_regs *regs = get_irq_regs();
77 return regs && user_mode(regs);
78 }
79
80 /*
81 * Can we use the FPU in kernel mode with the
82 * whole "kernel_fpu_begin/end()" sequence?
83 *
84 * It's always ok in process context (ie "not interrupt")
85 * but it is sometimes ok even from an irq.
86 */
87 bool irq_fpu_usable(void)
88 {
89 return !in_interrupt() ||
90 interrupted_user_mode() ||
91 interrupted_kernel_fpu_idle();
92 }
93 EXPORT_SYMBOL(irq_fpu_usable);
94
95 void __kernel_fpu_begin(void)
96 {
97 struct fpu *fpu = &current->thread.fpu;
98
99 WARN_ON_FPU(!irq_fpu_usable());
100
101 kernel_fpu_disable();
102
103 if (fpu->initialized) {
104 /*
105 * Ignore return value -- we don't care if reg state
106 * is clobbered.
107 */
108 copy_fpregs_to_fpstate(fpu);
109 } else {
110 __cpu_invalidate_fpregs_state();
111 }
112 }
113 EXPORT_SYMBOL(__kernel_fpu_begin);
114
115 void __kernel_fpu_end(void)
116 {
117 struct fpu *fpu = &current->thread.fpu;
118
119 if (fpu->initialized)
120 copy_kernel_to_fpregs(&fpu->state);
121
122 kernel_fpu_enable();
123 }
124 EXPORT_SYMBOL(__kernel_fpu_end);
125
126 void kernel_fpu_begin(void)
127 {
128 preempt_disable();
129 __kernel_fpu_begin();
130 }
131 EXPORT_SYMBOL_GPL(kernel_fpu_begin);
132
133 void kernel_fpu_end(void)
134 {
135 __kernel_fpu_end();
136 preempt_enable();
137 }
138 EXPORT_SYMBOL_GPL(kernel_fpu_end);
139
140 /*
141 * Save the FPU state (mark it for reload if necessary):
142 *
143 * This only ever gets called for the current task.
144 */
145 void fpu__save(struct fpu *fpu)
146 {
147 WARN_ON_FPU(fpu != &current->thread.fpu);
148
149 preempt_disable();
150 trace_x86_fpu_before_save(fpu);
151 if (fpu->initialized) {
152 if (!copy_fpregs_to_fpstate(fpu)) {
153 copy_kernel_to_fpregs(&fpu->state);
154 }
155 }
156 trace_x86_fpu_after_save(fpu);
157 preempt_enable();
158 }
159 EXPORT_SYMBOL_GPL(fpu__save);
160
161 /*
162 * Legacy x87 fpstate state init:
163 */
164 static inline void fpstate_init_fstate(struct fregs_state *fp)
165 {
166 fp->cwd = 0xffff037fu;
167 fp->swd = 0xffff0000u;
168 fp->twd = 0xffffffffu;
169 fp->fos = 0xffff0000u;
170 }
171
172 void fpstate_init(union fpregs_state *state)
173 {
174 if (!static_cpu_has(X86_FEATURE_FPU)) {
175 fpstate_init_soft(&state->soft);
176 return;
177 }
178
179 memset(state, 0, fpu_kernel_xstate_size);
180
181 if (static_cpu_has(X86_FEATURE_XSAVES))
182 fpstate_init_xstate(&state->xsave);
183 if (static_cpu_has(X86_FEATURE_FXSR))
184 fpstate_init_fxstate(&state->fxsave);
185 else
186 fpstate_init_fstate(&state->fsave);
187 }
188 EXPORT_SYMBOL_GPL(fpstate_init);
189
190 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
191 {
192 dst_fpu->last_cpu = -1;
193
194 if (!src_fpu->initialized || !static_cpu_has(X86_FEATURE_FPU))
195 return 0;
196
197 WARN_ON_FPU(src_fpu != &current->thread.fpu);
198
199 /*
200 * Don't let 'init optimized' areas of the XSAVE area
201 * leak into the child task:
202 */
203 memset(&dst_fpu->state.xsave, 0, fpu_kernel_xstate_size);
204
205 /*
206 * Save current FPU registers directly into the child
207 * FPU context, without any memory-to-memory copying.
208 * In lazy mode, if the FPU context isn't loaded into
209 * fpregs, CR0.TS will be set and do_device_not_available
210 * will load the FPU context.
211 *
212 * We have to do all this with preemption disabled,
213 * mostly because of the FNSAVE case, because in that
214 * case we must not allow preemption in the window
215 * between the FNSAVE and us marking the context lazy.
216 *
217 * It shouldn't be an issue as even FNSAVE is plenty
218 * fast in terms of critical section length.
219 */
220 preempt_disable();
221 if (!copy_fpregs_to_fpstate(dst_fpu)) {
222 memcpy(&src_fpu->state, &dst_fpu->state,
223 fpu_kernel_xstate_size);
224
225 copy_kernel_to_fpregs(&src_fpu->state);
226 }
227 preempt_enable();
228
229 trace_x86_fpu_copy_src(src_fpu);
230 trace_x86_fpu_copy_dst(dst_fpu);
231
232 return 0;
233 }
234
235 /*
236 * Activate the current task's in-memory FPU context,
237 * if it has not been used before:
238 */
239 void fpu__activate_curr(struct fpu *fpu)
240 {
241 WARN_ON_FPU(fpu != &current->thread.fpu);
242
243 if (!fpu->initialized) {
244 fpstate_init(&fpu->state);
245 trace_x86_fpu_init_state(fpu);
246
247 trace_x86_fpu_activate_state(fpu);
248 /* Safe to do for the current task: */
249 fpu->initialized = 1;
250 }
251 }
252 EXPORT_SYMBOL_GPL(fpu__activate_curr);
253
254 /*
255 * This function must be called before we read a task's fpstate.
256 *
257 * There's two cases where this gets called:
258 *
259 * - for the current task (when coredumping), in which case we have
260 * to save the latest FPU registers into the fpstate,
261 *
262 * - or it's called for stopped tasks (ptrace), in which case the
263 * registers were already saved by the context-switch code when
264 * the task scheduled out - we only have to initialize the registers
265 * if they've never been initialized.
266 *
267 * If the task has used the FPU before then save it.
268 */
269 void fpu__activate_fpstate_read(struct fpu *fpu)
270 {
271 if (fpu == &current->thread.fpu) {
272 fpu__save(fpu);
273 } else {
274 if (!fpu->initialized) {
275 fpstate_init(&fpu->state);
276 trace_x86_fpu_init_state(fpu);
277
278 trace_x86_fpu_activate_state(fpu);
279 /* Safe to do for current and for stopped child tasks: */
280 fpu->initialized = 1;
281 }
282 }
283 }
284
285 /*
286 * This function must be called before we write a task's fpstate.
287 *
288 * If the task has used the FPU before then unlazy it.
289 * If the task has not used the FPU before then initialize its fpstate.
290 *
291 * After this function call, after registers in the fpstate are
292 * modified and the child task has woken up, the child task will
293 * restore the modified FPU state from the modified context. If we
294 * didn't clear its lazy status here then the lazy in-registers
295 * state pending on its former CPU could be restored, corrupting
296 * the modifications.
297 */
298 void fpu__activate_fpstate_write(struct fpu *fpu)
299 {
300 /*
301 * Only stopped child tasks can be used to modify the FPU
302 * state in the fpstate buffer:
303 */
304 WARN_ON_FPU(fpu == &current->thread.fpu);
305
306 if (fpu->initialized) {
307 /* Invalidate any lazy state: */
308 __fpu_invalidate_fpregs_state(fpu);
309 } else {
310 fpstate_init(&fpu->state);
311 trace_x86_fpu_init_state(fpu);
312
313 trace_x86_fpu_activate_state(fpu);
314 /* Safe to do for stopped child tasks: */
315 fpu->initialized = 1;
316 }
317 }
318
319 /*
320 * 'fpu__restore()' is called to copy FPU registers from
321 * the FPU fpstate to the live hw registers and to activate
322 * access to the hardware registers, so that FPU instructions
323 * can be used afterwards.
324 *
325 * Must be called with kernel preemption disabled (for example
326 * with local interrupts disabled, as it is in the case of
327 * do_device_not_available()).
328 */
329 void fpu__restore(struct fpu *fpu)
330 {
331 fpu__activate_curr(fpu);
332
333 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
334 kernel_fpu_disable();
335 trace_x86_fpu_before_restore(fpu);
336 fpregs_activate(fpu);
337 copy_kernel_to_fpregs(&fpu->state);
338 trace_x86_fpu_after_restore(fpu);
339 kernel_fpu_enable();
340 }
341 EXPORT_SYMBOL_GPL(fpu__restore);
342
343 /*
344 * Drops current FPU state: deactivates the fpregs and
345 * the fpstate. NOTE: it still leaves previous contents
346 * in the fpregs in the eager-FPU case.
347 *
348 * This function can be used in cases where we know that
349 * a state-restore is coming: either an explicit one,
350 * or a reschedule.
351 */
352 void fpu__drop(struct fpu *fpu)
353 {
354 preempt_disable();
355
356 if (fpu == &current->thread.fpu) {
357 if (fpu->initialized) {
358 /* Ignore delayed exceptions from user space */
359 asm volatile("1: fwait\n"
360 "2:\n"
361 _ASM_EXTABLE(1b, 2b));
362 fpregs_deactivate(fpu);
363 }
364 }
365
366 fpu->initialized = 0;
367
368 trace_x86_fpu_dropped(fpu);
369
370 preempt_enable();
371 }
372
373 /*
374 * Clear FPU registers by setting them up from
375 * the init fpstate:
376 */
377 static inline void copy_init_fpstate_to_fpregs(void)
378 {
379 if (use_xsave())
380 copy_kernel_to_xregs(&init_fpstate.xsave, -1);
381 else if (static_cpu_has(X86_FEATURE_FXSR))
382 copy_kernel_to_fxregs(&init_fpstate.fxsave);
383 else
384 copy_kernel_to_fregs(&init_fpstate.fsave);
385
386 if (boot_cpu_has(X86_FEATURE_OSPKE))
387 copy_init_pkru_to_fpregs();
388 }
389
390 /*
391 * Clear the FPU state back to init state.
392 *
393 * Called by sys_execve(), by the signal handler code and by various
394 * error paths.
395 */
396 void fpu__clear(struct fpu *fpu)
397 {
398 WARN_ON_FPU(fpu != &current->thread.fpu); /* Almost certainly an anomaly */
399
400 fpu__drop(fpu);
401
402 /*
403 * Make sure fpstate is cleared and initialized.
404 */
405 if (static_cpu_has(X86_FEATURE_FPU)) {
406 preempt_disable();
407 fpu__activate_curr(fpu);
408 user_fpu_begin();
409 copy_init_fpstate_to_fpregs();
410 preempt_enable();
411 }
412 }
413
414 /*
415 * x87 math exception handling:
416 */
417
418 int fpu__exception_code(struct fpu *fpu, int trap_nr)
419 {
420 int err;
421
422 if (trap_nr == X86_TRAP_MF) {
423 unsigned short cwd, swd;
424 /*
425 * (~cwd & swd) will mask out exceptions that are not set to unmasked
426 * status. 0x3f is the exception bits in these regs, 0x200 is the
427 * C1 reg you need in case of a stack fault, 0x040 is the stack
428 * fault bit. We should only be taking one exception at a time,
429 * so if this combination doesn't produce any single exception,
430 * then we have a bad program that isn't synchronizing its FPU usage
431 * and it will suffer the consequences since we won't be able to
432 * fully reproduce the context of the exception.
433 */
434 if (boot_cpu_has(X86_FEATURE_FXSR)) {
435 cwd = fpu->state.fxsave.cwd;
436 swd = fpu->state.fxsave.swd;
437 } else {
438 cwd = (unsigned short)fpu->state.fsave.cwd;
439 swd = (unsigned short)fpu->state.fsave.swd;
440 }
441
442 err = swd & ~cwd;
443 } else {
444 /*
445 * The SIMD FPU exceptions are handled a little differently, as there
446 * is only a single status/control register. Thus, to determine which
447 * unmasked exception was caught we must mask the exception mask bits
448 * at 0x1f80, and then use these to mask the exception bits at 0x3f.
449 */
450 unsigned short mxcsr = MXCSR_DEFAULT;
451
452 if (boot_cpu_has(X86_FEATURE_XMM))
453 mxcsr = fpu->state.fxsave.mxcsr;
454
455 err = ~(mxcsr >> 7) & mxcsr;
456 }
457
458 if (err & 0x001) { /* Invalid op */
459 /*
460 * swd & 0x240 == 0x040: Stack Underflow
461 * swd & 0x240 == 0x240: Stack Overflow
462 * User must clear the SF bit (0x40) if set
463 */
464 return FPE_FLTINV;
465 } else if (err & 0x004) { /* Divide by Zero */
466 return FPE_FLTDIV;
467 } else if (err & 0x008) { /* Overflow */
468 return FPE_FLTOVF;
469 } else if (err & 0x012) { /* Denormal, Underflow */
470 return FPE_FLTUND;
471 } else if (err & 0x020) { /* Precision */
472 return FPE_FLTRES;
473 }
474
475 /*
476 * If we're using IRQ 13, or supposedly even some trap
477 * X86_TRAP_MF implementations, it's possible
478 * we get a spurious trap, which is not an error.
479 */
480 return 0;
481 }