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 <linux/module.h>
9 #include <linux/regset.h>
10 #include <linux/sched.h>
11 #include <linux/slab.h>
13 #include <asm/sigcontext.h>
14 #include <asm/processor.h>
15 #include <asm/math_emu.h>
16 #include <asm/uaccess.h>
17 #include <asm/ptrace.h>
19 #include <asm/fpu-internal.h>
23 * Were we in an interrupt that interrupted kernel mode?
25 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
26 * pair does nothing at all: the thread must not have fpu (so
27 * that we don't try to save the FPU state), and TS must
28 * be set (so that the clts/stts pair does nothing that is
29 * visible in the interrupted kernel thread).
31 * Except for the eagerfpu case when we return 1 unless we've already
32 * been eager and saved the state in kernel_fpu_begin().
34 static inline bool interrupted_kernel_fpu_idle(void)
37 return __thread_has_fpu(current);
39 return !__thread_has_fpu(current) &&
40 (read_cr0() & X86_CR0_TS);
44 * Were we in user mode (or vm86 mode) when we were
47 * Doing kernel_fpu_begin/end() is ok if we are running
48 * in an interrupt context from user mode - we'll just
49 * save the FPU state as required.
51 static inline bool interrupted_user_mode(void)
53 struct pt_regs *regs = get_irq_regs();
54 return regs && user_mode_vm(regs);
58 * Can we use the FPU in kernel mode with the
59 * whole "kernel_fpu_begin/end()" sequence?
61 * It's always ok in process context (ie "not interrupt")
62 * but it is sometimes ok even from an irq.
64 bool irq_fpu_usable(void)
66 return !in_interrupt() ||
67 interrupted_user_mode() ||
68 interrupted_kernel_fpu_idle();
70 EXPORT_SYMBOL(irq_fpu_usable);
72 void __kernel_fpu_begin(void)
74 struct task_struct *me = current;
76 if (__thread_has_fpu(me)) {
77 __thread_clear_has_fpu(me);
79 /* We do 'stts()' in __kernel_fpu_end() */
80 } else if (!use_eager_fpu()) {
81 this_cpu_write(fpu_owner_task, NULL);
85 EXPORT_SYMBOL(__kernel_fpu_begin);
87 void __kernel_fpu_end(void)
89 if (use_eager_fpu()) {
91 * For eager fpu, most the time, tsk_used_math() is true.
92 * Restore the user math as we are done with the kernel usage.
93 * At few instances during thread exit, signal handling etc,
94 * tsk_used_math() is false. Those few places will take proper
95 * actions, so we don't need to restore the math here.
97 if (likely(tsk_used_math(current)))
103 EXPORT_SYMBOL(__kernel_fpu_end);
105 void unlazy_fpu(struct task_struct *tsk)
108 if (__thread_has_fpu(tsk)) {
109 if (use_eager_fpu()) {
112 __save_init_fpu(tsk);
113 __thread_fpu_end(tsk);
118 EXPORT_SYMBOL(unlazy_fpu);
120 unsigned int mxcsr_feature_mask __read_mostly = 0xffffffffu;
121 unsigned int xstate_size;
122 EXPORT_SYMBOL_GPL(xstate_size);
123 static struct i387_fxsave_struct fx_scratch;
125 static void mxcsr_feature_mask_init(void)
127 unsigned long mask = 0;
130 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
131 asm volatile("fxsave %0" : "+m" (fx_scratch));
132 mask = fx_scratch.mxcsr_mask;
136 mxcsr_feature_mask &= mask;
139 static void init_thread_xstate(void)
142 * Note that xstate_size might be overwriten later during
148 * Disable xsave as we do not support it if i387
149 * emulation is enabled.
151 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
152 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
153 xstate_size = sizeof(struct i387_soft_struct);
158 xstate_size = sizeof(struct i387_fxsave_struct);
160 xstate_size = sizeof(struct i387_fsave_struct);
164 * Called at bootup to set up the initial FPU state that is later cloned
165 * into all processes.
171 unsigned long cr4_mask = 0;
173 #ifndef CONFIG_MATH_EMULATION
175 pr_emerg("No FPU found and no math emulation present\n");
176 pr_emerg("Giving up\n");
182 cr4_mask |= X86_CR4_OSFXSR;
184 cr4_mask |= X86_CR4_OSXMMEXCPT;
186 set_in_cr4(cr4_mask);
189 cr0 &= ~(X86_CR0_TS|X86_CR0_EM); /* clear TS and EM */
195 * init_thread_xstate is only called once to avoid overriding
196 * xstate_size during boot time or during CPU hotplug.
198 if (xstate_size == 0)
199 init_thread_xstate();
201 mxcsr_feature_mask_init();
206 void fpu_finit(struct fpu *fpu)
209 finit_soft_fpu(&fpu->state->soft);
214 fx_finit(&fpu->state->fxsave);
216 struct i387_fsave_struct *fp = &fpu->state->fsave;
217 memset(fp, 0, xstate_size);
218 fp->cwd = 0xffff037fu;
219 fp->swd = 0xffff0000u;
220 fp->twd = 0xffffffffu;
221 fp->fos = 0xffff0000u;
224 EXPORT_SYMBOL_GPL(fpu_finit);
227 * The _current_ task is using the FPU for the first time
228 * so initialize it and set the mxcsr to its default
229 * value at reset if we support XMM instructions and then
230 * remember the current task has used the FPU.
232 int init_fpu(struct task_struct *tsk)
236 if (tsk_used_math(tsk)) {
237 if (cpu_has_fpu && tsk == current)
239 task_disable_lazy_fpu_restore(tsk);
244 * Memory allocation at the first usage of the FPU and other state.
246 ret = fpu_alloc(&tsk->thread.fpu);
250 fpu_finit(&tsk->thread.fpu);
252 set_stopped_child_used_math(tsk);
255 EXPORT_SYMBOL_GPL(init_fpu);
258 * The xstateregs_active() routine is the same as the fpregs_active() routine,
259 * as the "regset->n" for the xstate regset will be updated based on the feature
260 * capabilites supported by the xsave.
262 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
264 return tsk_used_math(target) ? regset->n : 0;
267 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
269 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
272 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
273 unsigned int pos, unsigned int count,
274 void *kbuf, void __user *ubuf)
281 ret = init_fpu(target);
285 sanitize_i387_state(target);
287 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
288 &target->thread.fpu.state->fxsave, 0, -1);
291 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
292 unsigned int pos, unsigned int count,
293 const void *kbuf, const void __user *ubuf)
300 ret = init_fpu(target);
304 sanitize_i387_state(target);
306 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
307 &target->thread.fpu.state->fxsave, 0, -1);
310 * mxcsr reserved bits must be masked to zero for security reasons.
312 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
315 * update the header bits in the xsave header, indicating the
316 * presence of FP and SSE state.
319 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
324 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
325 unsigned int pos, unsigned int count,
326 void *kbuf, void __user *ubuf)
333 ret = init_fpu(target);
338 * Copy the 48bytes defined by the software first into the xstate
339 * memory layout in the thread struct, so that we can copy the entire
340 * xstateregs to the user using one user_regset_copyout().
342 memcpy(&target->thread.fpu.state->fxsave.sw_reserved,
343 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
346 * Copy the xstate memory layout.
348 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf,
349 &target->thread.fpu.state->xsave, 0, -1);
353 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
354 unsigned int pos, unsigned int count,
355 const void *kbuf, const void __user *ubuf)
358 struct xsave_hdr_struct *xsave_hdr;
363 ret = init_fpu(target);
367 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
368 &target->thread.fpu.state->xsave, 0, -1);
371 * mxcsr reserved bits must be masked to zero for security reasons.
373 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
375 xsave_hdr = &target->thread.fpu.state->xsave.xsave_hdr;
377 xsave_hdr->xstate_bv &= pcntxt_mask;
379 * These bits must be zero.
381 memset(xsave_hdr->reserved, 0, 48);
386 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
389 * FPU tag word conversions.
392 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
394 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
396 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
398 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
399 /* and move the valid bits to the lower byte. */
400 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
401 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
402 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
407 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
408 #define FP_EXP_TAG_VALID 0
409 #define FP_EXP_TAG_ZERO 1
410 #define FP_EXP_TAG_SPECIAL 2
411 #define FP_EXP_TAG_EMPTY 3
413 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
416 u32 tos = (fxsave->swd >> 11) & 7;
417 u32 twd = (unsigned long) fxsave->twd;
419 u32 ret = 0xffff0000u;
422 for (i = 0; i < 8; i++, twd >>= 1) {
424 st = FPREG_ADDR(fxsave, (i - tos) & 7);
426 switch (st->exponent & 0x7fff) {
428 tag = FP_EXP_TAG_SPECIAL;
431 if (!st->significand[0] &&
432 !st->significand[1] &&
433 !st->significand[2] &&
435 tag = FP_EXP_TAG_ZERO;
437 tag = FP_EXP_TAG_SPECIAL;
440 if (st->significand[3] & 0x8000)
441 tag = FP_EXP_TAG_VALID;
443 tag = FP_EXP_TAG_SPECIAL;
447 tag = FP_EXP_TAG_EMPTY;
449 ret |= tag << (2 * i);
455 * FXSR floating point environment conversions.
459 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
461 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
462 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
463 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
466 env->cwd = fxsave->cwd | 0xffff0000u;
467 env->swd = fxsave->swd | 0xffff0000u;
468 env->twd = twd_fxsr_to_i387(fxsave);
471 env->fip = fxsave->rip;
472 env->foo = fxsave->rdp;
474 * should be actually ds/cs at fpu exception time, but
475 * that information is not available in 64bit mode.
477 env->fcs = task_pt_regs(tsk)->cs;
478 if (tsk == current) {
479 savesegment(ds, env->fos);
481 env->fos = tsk->thread.ds;
483 env->fos |= 0xffff0000;
485 env->fip = fxsave->fip;
486 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
487 env->foo = fxsave->foo;
488 env->fos = fxsave->fos;
491 for (i = 0; i < 8; ++i)
492 memcpy(&to[i], &from[i], sizeof(to[0]));
495 void convert_to_fxsr(struct task_struct *tsk,
496 const struct user_i387_ia32_struct *env)
499 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
500 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
501 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
504 fxsave->cwd = env->cwd;
505 fxsave->swd = env->swd;
506 fxsave->twd = twd_i387_to_fxsr(env->twd);
507 fxsave->fop = (u16) ((u32) env->fcs >> 16);
509 fxsave->rip = env->fip;
510 fxsave->rdp = env->foo;
511 /* cs and ds ignored */
513 fxsave->fip = env->fip;
514 fxsave->fcs = (env->fcs & 0xffff);
515 fxsave->foo = env->foo;
516 fxsave->fos = env->fos;
519 for (i = 0; i < 8; ++i)
520 memcpy(&to[i], &from[i], sizeof(from[0]));
523 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
524 unsigned int pos, unsigned int count,
525 void *kbuf, void __user *ubuf)
527 struct user_i387_ia32_struct env;
530 ret = init_fpu(target);
534 if (!static_cpu_has(X86_FEATURE_FPU))
535 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
538 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
539 &target->thread.fpu.state->fsave, 0,
542 sanitize_i387_state(target);
544 if (kbuf && pos == 0 && count == sizeof(env)) {
545 convert_from_fxsr(kbuf, target);
549 convert_from_fxsr(&env, target);
551 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
554 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
555 unsigned int pos, unsigned int count,
556 const void *kbuf, const void __user *ubuf)
558 struct user_i387_ia32_struct env;
561 ret = init_fpu(target);
565 sanitize_i387_state(target);
567 if (!static_cpu_has(X86_FEATURE_FPU))
568 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
571 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
572 &target->thread.fpu.state->fsave, 0,
575 if (pos > 0 || count < sizeof(env))
576 convert_from_fxsr(&env, target);
578 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
580 convert_to_fxsr(target, &env);
583 * update the header bit in the xsave header, indicating the
587 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
592 * FPU state for core dumps.
593 * This is only used for a.out dumps now.
594 * It is declared generically using elf_fpregset_t (which is
595 * struct user_i387_struct) but is in fact only used for 32-bit
596 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
598 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
600 struct task_struct *tsk = current;
603 fpvalid = !!used_math();
605 fpvalid = !fpregs_get(tsk, NULL,
606 0, sizeof(struct user_i387_ia32_struct),
611 EXPORT_SYMBOL(dump_fpu);
613 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */
615 static int __init no_387(char *s)
617 setup_clear_cpu_cap(X86_FEATURE_FPU);
621 __setup("no387", no_387);
623 void fpu_detect(struct cpuinfo_x86 *c)
631 cr0 &= ~(X86_CR0_TS | X86_CR0_EM);
634 asm volatile("fninit ; fnstsw %0 ; fnstcw %1"
635 : "+m" (fsw), "+m" (fcw));
637 if (fsw == 0 && (fcw & 0x103f) == 0x003f)
638 set_cpu_cap(c, X86_FEATURE_FPU);
640 clear_cpu_cap(c, X86_FEATURE_FPU);
642 /* The final cr0 value is set in fpu_init() */