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 * For now, with eagerfpu we will return interrupted kernel FPU
26 * state as not-idle. TBD: Ideally we can change the return value
27 * to something like __thread_has_fpu(current). But we need to
28 * be careful of doing __thread_clear_has_fpu() before saving
29 * the FPU etc for supporting nested uses etc. For now, take
32 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
33 * pair does nothing at all: the thread must not have fpu (so
34 * that we don't try to save the FPU state), and TS must
35 * be set (so that the clts/stts pair does nothing that is
36 * visible in the interrupted kernel thread).
38 static inline bool interrupted_kernel_fpu_idle(void)
43 return !__thread_has_fpu(current) &&
44 (read_cr0() & X86_CR0_TS);
48 * Were we in user mode (or vm86 mode) when we were
51 * Doing kernel_fpu_begin/end() is ok if we are running
52 * in an interrupt context from user mode - we'll just
53 * save the FPU state as required.
55 static inline bool interrupted_user_mode(void)
57 struct pt_regs *regs = get_irq_regs();
58 return regs && user_mode_vm(regs);
62 * Can we use the FPU in kernel mode with the
63 * whole "kernel_fpu_begin/end()" sequence?
65 * It's always ok in process context (ie "not interrupt")
66 * but it is sometimes ok even from an irq.
68 bool irq_fpu_usable(void)
70 return !in_interrupt() ||
71 interrupted_user_mode() ||
72 interrupted_kernel_fpu_idle();
74 EXPORT_SYMBOL(irq_fpu_usable);
76 void __kernel_fpu_begin(void)
78 struct task_struct *me = current;
80 if (__thread_has_fpu(me)) {
82 __thread_clear_has_fpu(me);
83 /* We do 'stts()' in __kernel_fpu_end() */
84 } else if (!use_eager_fpu()) {
85 this_cpu_write(fpu_owner_task, NULL);
89 EXPORT_SYMBOL(__kernel_fpu_begin);
91 void __kernel_fpu_end(void)
98 EXPORT_SYMBOL(__kernel_fpu_end);
100 void unlazy_fpu(struct task_struct *tsk)
103 if (__thread_has_fpu(tsk)) {
104 __save_init_fpu(tsk);
105 __thread_fpu_end(tsk);
107 tsk->fpu_counter = 0;
110 EXPORT_SYMBOL(unlazy_fpu);
112 unsigned int mxcsr_feature_mask __read_mostly = 0xffffffffu;
113 unsigned int xstate_size;
114 EXPORT_SYMBOL_GPL(xstate_size);
115 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
117 static void __cpuinit mxcsr_feature_mask_init(void)
119 unsigned long mask = 0;
122 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
123 asm volatile("fxsave %0" : : "m" (fx_scratch));
124 mask = fx_scratch.mxcsr_mask;
128 mxcsr_feature_mask &= mask;
131 static void __cpuinit init_thread_xstate(void)
134 * Note that xstate_size might be overwriten later during
140 * Disable xsave as we do not support it if i387
141 * emulation is enabled.
143 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
144 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
145 xstate_size = sizeof(struct i387_soft_struct);
150 xstate_size = sizeof(struct i387_fxsave_struct);
152 xstate_size = sizeof(struct i387_fsave_struct);
156 * Called at bootup to set up the initial FPU state that is later cloned
157 * into all processes.
160 void __cpuinit fpu_init(void)
163 unsigned long cr4_mask = 0;
166 cr4_mask |= X86_CR4_OSFXSR;
168 cr4_mask |= X86_CR4_OSXMMEXCPT;
170 set_in_cr4(cr4_mask);
173 cr0 &= ~(X86_CR0_TS|X86_CR0_EM); /* clear TS and EM */
179 * init_thread_xstate is only called once to avoid overriding
180 * xstate_size during boot time or during CPU hotplug.
182 if (xstate_size == 0)
183 init_thread_xstate();
185 mxcsr_feature_mask_init();
190 void fpu_finit(struct fpu *fpu)
193 finit_soft_fpu(&fpu->state->soft);
198 fx_finit(&fpu->state->fxsave);
200 struct i387_fsave_struct *fp = &fpu->state->fsave;
201 memset(fp, 0, xstate_size);
202 fp->cwd = 0xffff037fu;
203 fp->swd = 0xffff0000u;
204 fp->twd = 0xffffffffu;
205 fp->fos = 0xffff0000u;
208 EXPORT_SYMBOL_GPL(fpu_finit);
211 * The _current_ task is using the FPU for the first time
212 * so initialize it and set the mxcsr to its default
213 * value at reset if we support XMM instructions and then
214 * remember the current task has used the FPU.
216 int init_fpu(struct task_struct *tsk)
220 if (tsk_used_math(tsk)) {
221 if (HAVE_HWFP && tsk == current)
223 tsk->thread.fpu.last_cpu = ~0;
228 * Memory allocation at the first usage of the FPU and other state.
230 ret = fpu_alloc(&tsk->thread.fpu);
234 fpu_finit(&tsk->thread.fpu);
236 set_stopped_child_used_math(tsk);
239 EXPORT_SYMBOL_GPL(init_fpu);
242 * The xstateregs_active() routine is the same as the fpregs_active() routine,
243 * as the "regset->n" for the xstate regset will be updated based on the feature
244 * capabilites supported by the xsave.
246 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
248 return tsk_used_math(target) ? regset->n : 0;
251 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
253 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
256 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
257 unsigned int pos, unsigned int count,
258 void *kbuf, void __user *ubuf)
265 ret = init_fpu(target);
269 sanitize_i387_state(target);
271 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
272 &target->thread.fpu.state->fxsave, 0, -1);
275 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
276 unsigned int pos, unsigned int count,
277 const void *kbuf, const void __user *ubuf)
284 ret = init_fpu(target);
288 sanitize_i387_state(target);
290 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
291 &target->thread.fpu.state->fxsave, 0, -1);
294 * mxcsr reserved bits must be masked to zero for security reasons.
296 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
299 * update the header bits in the xsave header, indicating the
300 * presence of FP and SSE state.
303 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
308 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
309 unsigned int pos, unsigned int count,
310 void *kbuf, void __user *ubuf)
317 ret = init_fpu(target);
322 * Copy the 48bytes defined by the software first into the xstate
323 * memory layout in the thread struct, so that we can copy the entire
324 * xstateregs to the user using one user_regset_copyout().
326 memcpy(&target->thread.fpu.state->fxsave.sw_reserved,
327 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
330 * Copy the xstate memory layout.
332 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf,
333 &target->thread.fpu.state->xsave, 0, -1);
337 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
338 unsigned int pos, unsigned int count,
339 const void *kbuf, const void __user *ubuf)
342 struct xsave_hdr_struct *xsave_hdr;
347 ret = init_fpu(target);
351 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
352 &target->thread.fpu.state->xsave, 0, -1);
355 * mxcsr reserved bits must be masked to zero for security reasons.
357 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
359 xsave_hdr = &target->thread.fpu.state->xsave.xsave_hdr;
361 xsave_hdr->xstate_bv &= pcntxt_mask;
363 * These bits must be zero.
365 xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
370 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
373 * FPU tag word conversions.
376 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
378 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
380 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
382 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
383 /* and move the valid bits to the lower byte. */
384 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
385 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
386 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
391 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
392 #define FP_EXP_TAG_VALID 0
393 #define FP_EXP_TAG_ZERO 1
394 #define FP_EXP_TAG_SPECIAL 2
395 #define FP_EXP_TAG_EMPTY 3
397 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
400 u32 tos = (fxsave->swd >> 11) & 7;
401 u32 twd = (unsigned long) fxsave->twd;
403 u32 ret = 0xffff0000u;
406 for (i = 0; i < 8; i++, twd >>= 1) {
408 st = FPREG_ADDR(fxsave, (i - tos) & 7);
410 switch (st->exponent & 0x7fff) {
412 tag = FP_EXP_TAG_SPECIAL;
415 if (!st->significand[0] &&
416 !st->significand[1] &&
417 !st->significand[2] &&
419 tag = FP_EXP_TAG_ZERO;
421 tag = FP_EXP_TAG_SPECIAL;
424 if (st->significand[3] & 0x8000)
425 tag = FP_EXP_TAG_VALID;
427 tag = FP_EXP_TAG_SPECIAL;
431 tag = FP_EXP_TAG_EMPTY;
433 ret |= tag << (2 * i);
439 * FXSR floating point environment conversions.
443 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
445 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
446 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
447 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
450 env->cwd = fxsave->cwd | 0xffff0000u;
451 env->swd = fxsave->swd | 0xffff0000u;
452 env->twd = twd_fxsr_to_i387(fxsave);
455 env->fip = fxsave->rip;
456 env->foo = fxsave->rdp;
458 * should be actually ds/cs at fpu exception time, but
459 * that information is not available in 64bit mode.
461 env->fcs = task_pt_regs(tsk)->cs;
462 if (tsk == current) {
463 savesegment(ds, env->fos);
465 env->fos = tsk->thread.ds;
467 env->fos |= 0xffff0000;
469 env->fip = fxsave->fip;
470 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
471 env->foo = fxsave->foo;
472 env->fos = fxsave->fos;
475 for (i = 0; i < 8; ++i)
476 memcpy(&to[i], &from[i], sizeof(to[0]));
479 void convert_to_fxsr(struct task_struct *tsk,
480 const struct user_i387_ia32_struct *env)
483 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
484 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
485 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
488 fxsave->cwd = env->cwd;
489 fxsave->swd = env->swd;
490 fxsave->twd = twd_i387_to_fxsr(env->twd);
491 fxsave->fop = (u16) ((u32) env->fcs >> 16);
493 fxsave->rip = env->fip;
494 fxsave->rdp = env->foo;
495 /* cs and ds ignored */
497 fxsave->fip = env->fip;
498 fxsave->fcs = (env->fcs & 0xffff);
499 fxsave->foo = env->foo;
500 fxsave->fos = env->fos;
503 for (i = 0; i < 8; ++i)
504 memcpy(&to[i], &from[i], sizeof(from[0]));
507 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
508 unsigned int pos, unsigned int count,
509 void *kbuf, void __user *ubuf)
511 struct user_i387_ia32_struct env;
514 ret = init_fpu(target);
519 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
522 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
523 &target->thread.fpu.state->fsave, 0,
527 sanitize_i387_state(target);
529 if (kbuf && pos == 0 && count == sizeof(env)) {
530 convert_from_fxsr(kbuf, target);
534 convert_from_fxsr(&env, target);
536 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
539 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
540 unsigned int pos, unsigned int count,
541 const void *kbuf, const void __user *ubuf)
543 struct user_i387_ia32_struct env;
546 ret = init_fpu(target);
550 sanitize_i387_state(target);
553 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
556 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
557 &target->thread.fpu.state->fsave, 0, -1);
560 if (pos > 0 || count < sizeof(env))
561 convert_from_fxsr(&env, target);
563 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
565 convert_to_fxsr(target, &env);
568 * update the header bit in the xsave header, indicating the
572 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
577 * FPU state for core dumps.
578 * This is only used for a.out dumps now.
579 * It is declared generically using elf_fpregset_t (which is
580 * struct user_i387_struct) but is in fact only used for 32-bit
581 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
583 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
585 struct task_struct *tsk = current;
588 fpvalid = !!used_math();
590 fpvalid = !fpregs_get(tsk, NULL,
591 0, sizeof(struct user_i387_ia32_struct),
596 EXPORT_SYMBOL(dump_fpu);
598 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */