2 * xsave/xrstor support.
4 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/bootmem.h>
10 #include <linux/compat.h>
11 #include <linux/cpu.h>
13 #include <asm/fpu-internal.h>
14 #include <asm/sigframe.h>
18 * Supported feature mask by the CPU and the kernel.
23 * Represents init state for the supported extended state.
25 struct xsave_struct
*init_xstate_buf
;
27 static struct _fpx_sw_bytes fx_sw_reserved
, fx_sw_reserved_ia32
;
28 static unsigned int *xstate_offsets
, *xstate_sizes
;
29 static unsigned int xstate_comp_offsets
[sizeof(pcntxt_mask
)*8];
30 static unsigned int xstate_features
;
33 * If a processor implementation discern that a processor state component is
34 * in its initialized state it may modify the corresponding bit in the
35 * xsave_hdr.xstate_bv as '0', with out modifying the corresponding memory
36 * layout in the case of xsaveopt. While presenting the xstate information to
37 * the user, we always ensure that the memory layout of a feature will be in
38 * the init state if the corresponding header bit is zero. This is to ensure
39 * that the user doesn't see some stale state in the memory layout during
40 * signal handling, debugging etc.
42 void __sanitize_i387_state(struct task_struct
*tsk
)
44 struct i387_fxsave_struct
*fx
= &tsk
->thread
.fpu
.state
->fxsave
;
45 int feature_bit
= 0x2;
51 xstate_bv
= tsk
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
;
54 * None of the feature bits are in init state. So nothing else
55 * to do for us, as the memory layout is up to date.
57 if ((xstate_bv
& pcntxt_mask
) == pcntxt_mask
)
63 if (!(xstate_bv
& XSTATE_FP
)) {
70 memset(&fx
->st_space
[0], 0, 128);
74 * SSE is in init state
76 if (!(xstate_bv
& XSTATE_SSE
))
77 memset(&fx
->xmm_space
[0], 0, 256);
79 xstate_bv
= (pcntxt_mask
& ~xstate_bv
) >> 2;
82 * Update all the other memory layouts for which the corresponding
83 * header bit is in the init state.
86 if (xstate_bv
& 0x1) {
87 int offset
= xstate_offsets
[feature_bit
];
88 int size
= xstate_sizes
[feature_bit
];
90 memcpy(((void *) fx
) + offset
,
91 ((void *) init_xstate_buf
) + offset
,
101 * Check for the presence of extended state information in the
102 * user fpstate pointer in the sigcontext.
104 static inline int check_for_xstate(struct i387_fxsave_struct __user
*buf
,
105 void __user
*fpstate
,
106 struct _fpx_sw_bytes
*fx_sw
)
108 int min_xstate_size
= sizeof(struct i387_fxsave_struct
) +
109 sizeof(struct xsave_hdr_struct
);
112 if (__copy_from_user(fx_sw
, &buf
->sw_reserved
[0], sizeof(*fx_sw
)))
115 /* Check for the first magic field and other error scenarios. */
116 if (fx_sw
->magic1
!= FP_XSTATE_MAGIC1
||
117 fx_sw
->xstate_size
< min_xstate_size
||
118 fx_sw
->xstate_size
> xstate_size
||
119 fx_sw
->xstate_size
> fx_sw
->extended_size
)
123 * Check for the presence of second magic word at the end of memory
124 * layout. This detects the case where the user just copied the legacy
125 * fpstate layout with out copying the extended state information
126 * in the memory layout.
128 if (__get_user(magic2
, (__u32 __user
*)(fpstate
+ fx_sw
->xstate_size
))
129 || magic2
!= FP_XSTATE_MAGIC2
)
136 * Signal frame handlers.
138 static inline int save_fsave_header(struct task_struct
*tsk
, void __user
*buf
)
141 struct xsave_struct
*xsave
= &tsk
->thread
.fpu
.state
->xsave
;
142 struct user_i387_ia32_struct env
;
143 struct _fpstate_ia32 __user
*fp
= buf
;
145 convert_from_fxsr(&env
, tsk
);
147 if (__copy_to_user(buf
, &env
, sizeof(env
)) ||
148 __put_user(xsave
->i387
.swd
, &fp
->status
) ||
149 __put_user(X86_FXSR_MAGIC
, &fp
->magic
))
152 struct i387_fsave_struct __user
*fp
= buf
;
154 if (__get_user(swd
, &fp
->swd
) || __put_user(swd
, &fp
->status
))
161 static inline int save_xstate_epilog(void __user
*buf
, int ia32_frame
)
163 struct xsave_struct __user
*x
= buf
;
164 struct _fpx_sw_bytes
*sw_bytes
;
168 /* Setup the bytes not touched by the [f]xsave and reserved for SW. */
169 sw_bytes
= ia32_frame
? &fx_sw_reserved_ia32
: &fx_sw_reserved
;
170 err
= __copy_to_user(&x
->i387
.sw_reserved
, sw_bytes
, sizeof(*sw_bytes
));
175 err
|= __put_user(FP_XSTATE_MAGIC2
, (__u32
*)(buf
+ xstate_size
));
178 * Read the xstate_bv which we copied (directly from the cpu or
179 * from the state in task struct) to the user buffers.
181 err
|= __get_user(xstate_bv
, (__u32
*)&x
->xsave_hdr
.xstate_bv
);
184 * For legacy compatible, we always set FP/SSE bits in the bit
185 * vector while saving the state to the user context. This will
186 * enable us capturing any changes(during sigreturn) to
187 * the FP/SSE bits by the legacy applications which don't touch
188 * xstate_bv in the xsave header.
190 * xsave aware apps can change the xstate_bv in the xsave
191 * header as well as change any contents in the memory layout.
192 * xrestore as part of sigreturn will capture all the changes.
194 xstate_bv
|= XSTATE_FPSSE
;
196 err
|= __put_user(xstate_bv
, (__u32
*)&x
->xsave_hdr
.xstate_bv
);
201 static inline int save_user_xstate(struct xsave_struct __user
*buf
)
206 err
= xsave_user(buf
);
208 err
= fxsave_user((struct i387_fxsave_struct __user
*) buf
);
210 err
= fsave_user((struct i387_fsave_struct __user
*) buf
);
212 if (unlikely(err
) && __clear_user(buf
, xstate_size
))
218 * Save the fpu, extended register state to the user signal frame.
220 * 'buf_fx' is the 64-byte aligned pointer at which the [f|fx|x]save
222 * 'buf' points to the 'buf_fx' or to the fsave header followed by 'buf_fx'.
224 * buf == buf_fx for 64-bit frames and 32-bit fsave frame.
225 * buf != buf_fx for 32-bit frames with fxstate.
227 * If the fpu, extended register state is live, save the state directly
228 * to the user frame pointed by the aligned pointer 'buf_fx'. Otherwise,
229 * copy the thread's fpu state to the user frame starting at 'buf_fx'.
231 * If this is a 32-bit frame with fxstate, put a fsave header before
232 * the aligned state at 'buf_fx'.
234 * For [f]xsave state, update the SW reserved fields in the [f]xsave frame
235 * indicating the absence/presence of the extended state to the user.
237 int save_xstate_sig(void __user
*buf
, void __user
*buf_fx
, int size
)
239 struct xsave_struct
*xsave
= ¤t
->thread
.fpu
.state
->xsave
;
240 struct task_struct
*tsk
= current
;
241 int ia32_fxstate
= (buf
!= buf_fx
);
243 ia32_fxstate
&= (config_enabled(CONFIG_X86_32
) ||
244 config_enabled(CONFIG_IA32_EMULATION
));
246 if (!access_ok(VERIFY_WRITE
, buf
, size
))
249 if (!static_cpu_has(X86_FEATURE_FPU
))
250 return fpregs_soft_get(current
, NULL
, 0,
251 sizeof(struct user_i387_ia32_struct
), NULL
,
252 (struct _fpstate_ia32 __user
*) buf
) ? -1 : 1;
254 if (user_has_fpu()) {
255 /* Save the live register state to the user directly. */
256 if (save_user_xstate(buf_fx
))
258 /* Update the thread's fxstate to save the fsave header. */
260 fpu_fxsave(&tsk
->thread
.fpu
);
262 sanitize_i387_state(tsk
);
263 if (__copy_to_user(buf_fx
, xsave
, xstate_size
))
267 /* Save the fsave header for the 32-bit frames. */
268 if ((ia32_fxstate
|| !use_fxsr()) && save_fsave_header(tsk
, buf
))
271 if (use_fxsr() && save_xstate_epilog(buf_fx
, ia32_fxstate
))
278 sanitize_restored_xstate(struct task_struct
*tsk
,
279 struct user_i387_ia32_struct
*ia32_env
,
280 u64 xstate_bv
, int fx_only
)
282 struct xsave_struct
*xsave
= &tsk
->thread
.fpu
.state
->xsave
;
283 struct xsave_hdr_struct
*xsave_hdr
= &xsave
->xsave_hdr
;
286 /* These bits must be zero. */
287 memset(xsave_hdr
->reserved
, 0, 48);
290 * Init the state that is not present in the memory
291 * layout and not enabled by the OS.
294 xsave_hdr
->xstate_bv
= XSTATE_FPSSE
;
296 xsave_hdr
->xstate_bv
&= (pcntxt_mask
& xstate_bv
);
301 * mscsr reserved bits must be masked to zero for security
304 xsave
->i387
.mxcsr
&= mxcsr_feature_mask
;
306 convert_to_fxsr(tsk
, ia32_env
);
311 * Restore the extended state if present. Otherwise, restore the FP/SSE state.
313 static inline int restore_user_xstate(void __user
*buf
, u64 xbv
, int fx_only
)
316 if ((unsigned long)buf
% 64 || fx_only
) {
317 u64 init_bv
= pcntxt_mask
& ~XSTATE_FPSSE
;
318 xrstor_state(init_xstate_buf
, init_bv
);
319 return fxrstor_user(buf
);
321 u64 init_bv
= pcntxt_mask
& ~xbv
;
322 if (unlikely(init_bv
))
323 xrstor_state(init_xstate_buf
, init_bv
);
324 return xrestore_user(buf
, xbv
);
326 } else if (use_fxsr()) {
327 return fxrstor_user(buf
);
329 return frstor_user(buf
);
332 int __restore_xstate_sig(void __user
*buf
, void __user
*buf_fx
, int size
)
334 int ia32_fxstate
= (buf
!= buf_fx
);
335 struct task_struct
*tsk
= current
;
336 int state_size
= xstate_size
;
340 ia32_fxstate
&= (config_enabled(CONFIG_X86_32
) ||
341 config_enabled(CONFIG_IA32_EMULATION
));
348 if (!access_ok(VERIFY_READ
, buf
, size
))
351 if (!used_math() && init_fpu(tsk
))
354 if (!static_cpu_has(X86_FEATURE_FPU
))
355 return fpregs_soft_set(current
, NULL
,
356 0, sizeof(struct user_i387_ia32_struct
),
360 struct _fpx_sw_bytes fx_sw_user
;
361 if (unlikely(check_for_xstate(buf_fx
, buf_fx
, &fx_sw_user
))) {
363 * Couldn't find the extended state information in the
364 * memory layout. Restore just the FP/SSE and init all
365 * the other extended state.
367 state_size
= sizeof(struct i387_fxsave_struct
);
370 state_size
= fx_sw_user
.xstate_size
;
371 xstate_bv
= fx_sw_user
.xstate_bv
;
377 * For 32-bit frames with fxstate, copy the user state to the
378 * thread's fpu state, reconstruct fxstate from the fsave
379 * header. Sanitize the copied state etc.
381 struct xsave_struct
*xsave
= &tsk
->thread
.fpu
.state
->xsave
;
382 struct user_i387_ia32_struct env
;
386 * Drop the current fpu which clears used_math(). This ensures
387 * that any context-switch during the copy of the new state,
388 * avoids the intermediate state from getting restored/saved.
389 * Thus avoiding the new restored state from getting corrupted.
390 * We will be ready to restore/save the state only after
391 * set_used_math() is again set.
395 if (__copy_from_user(xsave
, buf_fx
, state_size
) ||
396 __copy_from_user(&env
, buf
, sizeof(env
))) {
399 sanitize_restored_xstate(tsk
, &env
, xstate_bv
, fx_only
);
403 if (use_eager_fpu()) {
405 math_state_restore();
412 * For 64-bit frames and 32-bit fsave frames, restore the user
413 * state to the registers directly (with exceptions handled).
416 if (restore_user_xstate(buf_fx
, xstate_bv
, fx_only
)) {
426 * Prepare the SW reserved portion of the fxsave memory layout, indicating
427 * the presence of the extended state information in the memory layout
428 * pointed by the fpstate pointer in the sigcontext.
429 * This will be saved when ever the FP and extended state context is
430 * saved on the user stack during the signal handler delivery to the user.
432 static void prepare_fx_sw_frame(void)
434 int fsave_header_size
= sizeof(struct i387_fsave_struct
);
435 int size
= xstate_size
+ FP_XSTATE_MAGIC2_SIZE
;
437 if (config_enabled(CONFIG_X86_32
))
438 size
+= fsave_header_size
;
440 fx_sw_reserved
.magic1
= FP_XSTATE_MAGIC1
;
441 fx_sw_reserved
.extended_size
= size
;
442 fx_sw_reserved
.xstate_bv
= pcntxt_mask
;
443 fx_sw_reserved
.xstate_size
= xstate_size
;
445 if (config_enabled(CONFIG_IA32_EMULATION
)) {
446 fx_sw_reserved_ia32
= fx_sw_reserved
;
447 fx_sw_reserved_ia32
.extended_size
+= fsave_header_size
;
452 * Enable the extended processor state save/restore feature
454 static inline void xstate_enable(void)
456 set_in_cr4(X86_CR4_OSXSAVE
);
457 xsetbv(XCR_XFEATURE_ENABLED_MASK
, pcntxt_mask
);
461 * Record the offsets and sizes of different state managed by the xsave
464 static void __init
setup_xstate_features(void)
466 int eax
, ebx
, ecx
, edx
, leaf
= 0x2;
468 xstate_features
= fls64(pcntxt_mask
);
469 xstate_offsets
= alloc_bootmem(xstate_features
* sizeof(int));
470 xstate_sizes
= alloc_bootmem(xstate_features
* sizeof(int));
473 cpuid_count(XSTATE_CPUID
, leaf
, &eax
, &ebx
, &ecx
, &edx
);
478 xstate_offsets
[leaf
] = ebx
;
479 xstate_sizes
[leaf
] = eax
;
486 * This function sets up offsets and sizes of all extended states in
487 * xsave area. This supports both standard format and compacted format
488 * of the xsave aread.
493 void setup_xstate_comp(void)
495 unsigned int xstate_comp_sizes
[sizeof(pcntxt_mask
)*8];
499 * The FP xstates and SSE xstates are legacy states. They are always
500 * in the fixed offsets in the xsave area in either compacted form
503 xstate_comp_offsets
[0] = 0;
504 xstate_comp_offsets
[1] = offsetof(struct i387_fxsave_struct
, xmm_space
);
506 if (!cpu_has_xsaves
) {
507 for (i
= 2; i
< xstate_features
; i
++) {
508 if (test_bit(i
, (unsigned long *)&pcntxt_mask
)) {
509 xstate_comp_offsets
[i
] = xstate_offsets
[i
];
510 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
516 xstate_comp_offsets
[2] = FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
518 for (i
= 2; i
< xstate_features
; i
++) {
519 if (test_bit(i
, (unsigned long *)&pcntxt_mask
))
520 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
522 xstate_comp_sizes
[i
] = 0;
525 xstate_comp_offsets
[i
] = xstate_comp_offsets
[i
-1]
526 + xstate_comp_sizes
[i
-1];
532 * setup the xstate image representing the init state
534 static void __init
setup_init_fpu_buf(void)
537 * Setup init_xstate_buf to represent the init state of
538 * all the features managed by the xsave
540 init_xstate_buf
= alloc_bootmem_align(xstate_size
,
541 __alignof__(struct xsave_struct
));
542 fx_finit(&init_xstate_buf
->i387
);
547 setup_xstate_features();
549 if (cpu_has_xsaves
) {
550 init_xstate_buf
->xsave_hdr
.xcomp_bv
=
551 (u64
)1 << 63 | pcntxt_mask
;
552 init_xstate_buf
->xsave_hdr
.xstate_bv
= pcntxt_mask
;
556 * Init all the features state with header_bv being 0x0
558 xrstor_state_booting(init_xstate_buf
, -1);
560 * Dump the init state again. This is to identify the init state
561 * of any feature which is not represented by all zero's.
563 xsave_state_booting(init_xstate_buf
, -1);
566 static enum { AUTO
, ENABLE
, DISABLE
} eagerfpu
= AUTO
;
567 static int __init
eager_fpu_setup(char *s
)
569 if (!strcmp(s
, "on"))
571 else if (!strcmp(s
, "off"))
573 else if (!strcmp(s
, "auto"))
577 __setup("eagerfpu=", eager_fpu_setup
);
581 * Calculate total size of enabled xstates in XCR0/pcntxt_mask.
583 static void __init
init_xstate_size(void)
585 unsigned int eax
, ebx
, ecx
, edx
;
588 if (!cpu_has_xsaves
) {
589 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
594 xstate_size
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
595 for (i
= 2; i
< 64; i
++) {
596 if (test_bit(i
, (unsigned long *)&pcntxt_mask
)) {
597 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
604 * Enable and initialize the xsave feature.
606 static void __init
xstate_enable_boot_cpu(void)
608 unsigned int eax
, ebx
, ecx
, edx
;
610 if (boot_cpu_data
.cpuid_level
< XSTATE_CPUID
) {
611 WARN(1, KERN_ERR
"XSTATE_CPUID missing\n");
615 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
616 pcntxt_mask
= eax
+ ((u64
)edx
<< 32);
618 if ((pcntxt_mask
& XSTATE_FPSSE
) != XSTATE_FPSSE
) {
619 pr_err("FP/SSE not shown under xsave features 0x%llx\n",
625 * Support only the state known to OS.
627 pcntxt_mask
= pcntxt_mask
& XCNTXT_MASK
;
632 * Recompute the context size for enabled features
636 update_regset_xstate_info(xstate_size
, pcntxt_mask
);
637 prepare_fx_sw_frame();
638 setup_init_fpu_buf();
640 /* Auto enable eagerfpu for xsaveopt */
641 if (cpu_has_xsaveopt
&& eagerfpu
!= DISABLE
)
644 if (pcntxt_mask
& XSTATE_EAGER
) {
645 if (eagerfpu
== DISABLE
) {
646 pr_err("eagerfpu not present, disabling some xstate features: 0x%llx\n",
647 pcntxt_mask
& XSTATE_EAGER
);
648 pcntxt_mask
&= ~XSTATE_EAGER
;
654 pr_info("enabled xstate_bv 0x%llx, cntxt size 0x%x using %s\n",
655 pcntxt_mask
, xstate_size
,
656 cpu_has_xsaves
? "compacted form" : "standard form");
660 * For the very first instance, this calls xstate_enable_boot_cpu();
661 * for all subsequent instances, this calls xstate_enable().
663 * This is somewhat obfuscated due to the lack of powerful enough
664 * overrides for the section checks.
666 void xsave_init(void)
668 static __refdata
void (*next_func
)(void) = xstate_enable_boot_cpu
;
669 void (*this_func
)(void);
674 this_func
= next_func
;
675 next_func
= xstate_enable
;
679 static inline void __init
eager_fpu_init_bp(void)
681 current
->thread
.fpu
.state
=
682 alloc_bootmem_align(xstate_size
, __alignof__(struct xsave_struct
));
683 if (!init_xstate_buf
)
684 setup_init_fpu_buf();
687 void eager_fpu_init(void)
689 static __refdata
void (*boot_func
)(void) = eager_fpu_init_bp
;
692 current_thread_info()->status
= 0;
694 if (eagerfpu
== ENABLE
)
695 setup_force_cpu_cap(X86_FEATURE_EAGER_FPU
);
697 if (!cpu_has_eager_fpu
) {
708 * This is same as math_state_restore(). But use_xsave() is
709 * not yet patched to use math_state_restore().
712 __thread_fpu_begin(current
);
714 xrstor_state(init_xstate_buf
, -1);
716 fxrstor_checking(&init_xstate_buf
->i387
);
720 * Given the xsave area and a state inside, this function returns the
721 * address of the state.
723 * This is the API that is called to get xstate address in either
724 * standard format or compacted format of xsave area.
727 * xsave: base address of the xsave area;
728 * xstate: state which is defined in xsave.h (e.g. XSTATE_FP, XSTATE_SSE,
731 * address of the state in the xsave area.
733 void *get_xsave_addr(struct xsave_struct
*xsave
, int xstate
)
735 int feature
= fls64(xstate
) - 1;
736 if (!test_bit(feature
, (unsigned long *)&pcntxt_mask
))
739 return (void *)xsave
+ xstate_comp_offsets
[feature
];