2 * xsave/xrstor support.
4 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
6 #include <linux/compat.h>
9 #include <asm/fpu/api.h>
10 #include <asm/fpu/internal.h>
11 #include <asm/fpu/signal.h>
12 #include <asm/fpu/regset.h>
14 #include <asm/tlbflush.h>
16 static const char *xfeature_names
[] =
18 "x87 floating point registers" ,
21 "MPX bounds registers" ,
26 "unknown xstate feature" ,
30 * Mask of xstate features supported by the CPU and the kernel:
32 u64 xfeatures_mask __read_mostly
;
34 static unsigned int xstate_offsets
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
35 static unsigned int xstate_sizes
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
36 static unsigned int xstate_comp_offsets
[sizeof(xfeatures_mask
)*8];
39 * Clear all of the X86_FEATURE_* bits that are unavailable
40 * when the CPU has no XSAVE support.
42 void fpu__xstate_clear_all_cpu_caps(void)
44 setup_clear_cpu_cap(X86_FEATURE_XSAVE
);
45 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT
);
46 setup_clear_cpu_cap(X86_FEATURE_XSAVEC
);
47 setup_clear_cpu_cap(X86_FEATURE_XSAVES
);
48 setup_clear_cpu_cap(X86_FEATURE_AVX
);
49 setup_clear_cpu_cap(X86_FEATURE_AVX2
);
50 setup_clear_cpu_cap(X86_FEATURE_AVX512F
);
51 setup_clear_cpu_cap(X86_FEATURE_AVX512PF
);
52 setup_clear_cpu_cap(X86_FEATURE_AVX512ER
);
53 setup_clear_cpu_cap(X86_FEATURE_AVX512CD
);
54 setup_clear_cpu_cap(X86_FEATURE_MPX
);
58 * Return whether the system supports a given xfeature.
60 * Also return the name of the (most advanced) feature that the caller requested:
62 int cpu_has_xfeatures(u64 xfeatures_needed
, const char **feature_name
)
64 u64 xfeatures_missing
= xfeatures_needed
& ~xfeatures_mask
;
66 if (unlikely(feature_name
)) {
67 long xfeature_idx
, max_idx
;
70 * So we use FLS here to be able to print the most advanced
71 * feature that was requested but is missing. So if a driver
72 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
73 * missing AVX feature - this is the most informative message
76 if (xfeatures_missing
)
77 xfeatures_print
= xfeatures_missing
;
79 xfeatures_print
= xfeatures_needed
;
81 xfeature_idx
= fls64(xfeatures_print
)-1;
82 max_idx
= ARRAY_SIZE(xfeature_names
)-1;
83 xfeature_idx
= min(xfeature_idx
, max_idx
);
85 *feature_name
= xfeature_names
[xfeature_idx
];
88 if (xfeatures_missing
)
93 EXPORT_SYMBOL_GPL(cpu_has_xfeatures
);
96 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
97 * a processor implementation detects that an FPU state component is still
98 * (or is again) in its initialized state, it may clear the corresponding
99 * bit in the header.xfeatures field, and can skip the writeout of registers
100 * to the corresponding memory layout.
102 * This means that when the bit is zero, the state component might still contain
103 * some previous - non-initialized register state.
105 * Before writing xstate information to user-space we sanitize those components,
106 * to always ensure that the memory layout of a feature will be in the init state
107 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
108 * see some stale state in the memory layout during signal handling, debugging etc.
110 void fpstate_sanitize_xstate(struct fpu
*fpu
)
112 struct fxregs_state
*fx
= &fpu
->state
.fxsave
;
119 xfeatures
= fpu
->state
.xsave
.header
.xfeatures
;
122 * None of the feature bits are in init state. So nothing else
123 * to do for us, as the memory layout is up to date.
125 if ((xfeatures
& xfeatures_mask
) == xfeatures_mask
)
129 * FP is in init state
131 if (!(xfeatures
& XFEATURE_MASK_FP
)) {
138 memset(&fx
->st_space
[0], 0, 128);
142 * SSE is in init state
144 if (!(xfeatures
& XFEATURE_MASK_SSE
))
145 memset(&fx
->xmm_space
[0], 0, 256);
148 * First two features are FPU and SSE, which above we handled
149 * in a special way already:
152 xfeatures
= (xfeatures_mask
& ~xfeatures
) >> 2;
155 * Update all the remaining memory layouts according to their
156 * standard xstate layout, if their header bit is in the init
160 if (xfeatures
& 0x1) {
161 int offset
= xstate_offsets
[feature_bit
];
162 int size
= xstate_sizes
[feature_bit
];
164 memcpy((void *)fx
+ offset
,
165 (void *)&init_fpstate
.xsave
+ offset
,
175 * Enable the extended processor state save/restore feature.
176 * Called once per CPU onlining.
178 void fpu__init_cpu_xstate(void)
180 if (!cpu_has_xsave
|| !xfeatures_mask
)
183 cr4_set_bits(X86_CR4_OSXSAVE
);
184 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
188 * Note that in the future we will likely need a pair of
189 * functions here: one for user xstates and the other for
190 * system xstates. For now, they are the same.
192 static int xfeature_enabled(enum xfeature xfeature
)
194 return !!(xfeatures_mask
& (1UL << xfeature
));
198 * Record the offsets and sizes of various xstates contained
199 * in the XSAVE state memory layout.
201 static void __init
setup_xstate_features(void)
203 u32 eax
, ebx
, ecx
, edx
, i
;
204 /* start at the beginnning of the "extended state" */
205 unsigned int last_good_offset
= offsetof(struct xregs_state
,
206 extended_state_area
);
208 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
209 if (!xfeature_enabled(i
))
212 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
213 xstate_offsets
[i
] = ebx
;
214 xstate_sizes
[i
] = eax
;
216 * In our xstate size checks, we assume that the
217 * highest-numbered xstate feature has the
218 * highest offset in the buffer. Ensure it does.
220 WARN_ONCE(last_good_offset
> xstate_offsets
[i
],
221 "x86/fpu: misordered xstate at %d\n", last_good_offset
);
222 last_good_offset
= xstate_offsets
[i
];
224 printk(KERN_INFO
"x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n", i
, ebx
, i
, eax
);
228 static void __init
print_xstate_feature(u64 xstate_mask
)
230 const char *feature_name
;
232 if (cpu_has_xfeatures(xstate_mask
, &feature_name
))
233 pr_info("x86/fpu: Supporting XSAVE feature 0x%02Lx: '%s'\n", xstate_mask
, feature_name
);
237 * Print out all the supported xstate features:
239 static void __init
print_xstate_features(void)
241 print_xstate_feature(XFEATURE_MASK_FP
);
242 print_xstate_feature(XFEATURE_MASK_SSE
);
243 print_xstate_feature(XFEATURE_MASK_YMM
);
244 print_xstate_feature(XFEATURE_MASK_BNDREGS
);
245 print_xstate_feature(XFEATURE_MASK_BNDCSR
);
246 print_xstate_feature(XFEATURE_MASK_OPMASK
);
247 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256
);
248 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM
);
252 * This function sets up offsets and sizes of all extended states in
253 * xsave area. This supports both standard format and compacted format
254 * of the xsave aread.
256 static void __init
setup_xstate_comp(void)
258 unsigned int xstate_comp_sizes
[sizeof(xfeatures_mask
)*8];
262 * The FP xstates and SSE xstates are legacy states. They are always
263 * in the fixed offsets in the xsave area in either compacted form
266 xstate_comp_offsets
[0] = 0;
267 xstate_comp_offsets
[1] = offsetof(struct fxregs_state
, xmm_space
);
269 if (!cpu_has_xsaves
) {
270 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
271 if (xfeature_enabled(i
)) {
272 xstate_comp_offsets
[i
] = xstate_offsets
[i
];
273 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
279 xstate_comp_offsets
[FIRST_EXTENDED_XFEATURE
] =
280 FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
282 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
283 if (xfeature_enabled(i
))
284 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
286 xstate_comp_sizes
[i
] = 0;
288 if (i
> FIRST_EXTENDED_XFEATURE
)
289 xstate_comp_offsets
[i
] = xstate_comp_offsets
[i
-1]
290 + xstate_comp_sizes
[i
-1];
296 * setup the xstate image representing the init state
298 static void __init
setup_init_fpu_buf(void)
300 static int on_boot_cpu __initdata
= 1;
302 WARN_ON_FPU(!on_boot_cpu
);
308 setup_xstate_features();
309 print_xstate_features();
311 if (cpu_has_xsaves
) {
312 init_fpstate
.xsave
.header
.xcomp_bv
= (u64
)1 << 63 | xfeatures_mask
;
313 init_fpstate
.xsave
.header
.xfeatures
= xfeatures_mask
;
317 * Init all the features state with header_bv being 0x0
319 copy_kernel_to_xregs_booting(&init_fpstate
.xsave
);
322 * Dump the init state again. This is to identify the init state
323 * of any feature which is not represented by all zero's.
325 copy_xregs_to_kernel_booting(&init_fpstate
.xsave
);
328 static int xfeature_is_supervisor(int xfeature_nr
)
331 * We currently do not support supervisor states, but if
332 * we did, we could find out like this.
334 * SDM says: If state component i is a user state component,
335 * ECX[0] return 0; if state component i is a supervisor
336 * state component, ECX[0] returns 1.
337 u32 eax, ebx, ecx, edx;
338 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx;
344 static int xfeature_is_user(int xfeature_nr)
346 return !xfeature_is_supervisor(xfeature_nr);
351 * This check is important because it is easy to get XSTATE_*
352 * confused with XSTATE_BIT_*.
354 #define CHECK_XFEATURE(nr) do { \
355 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
356 WARN_ON(nr >= XFEATURE_MAX); \
360 * We could cache this like xstate_size[], but we only use
361 * it here, so it would be a waste of space.
363 static int xfeature_is_aligned(int xfeature_nr
)
365 u32 eax
, ebx
, ecx
, edx
;
367 CHECK_XFEATURE(xfeature_nr
);
368 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
370 * The value returned by ECX[1] indicates the alignment
371 * of state component i when the compacted format
372 * of the extended region of an XSAVE area is used
377 static int xfeature_uncompacted_offset(int xfeature_nr
)
379 u32 eax
, ebx
, ecx
, edx
;
381 CHECK_XFEATURE(xfeature_nr
);
382 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
386 static int xfeature_size(int xfeature_nr
)
388 u32 eax
, ebx
, ecx
, edx
;
390 CHECK_XFEATURE(xfeature_nr
);
391 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
396 * 'XSAVES' implies two different things:
397 * 1. saving of supervisor/system state
398 * 2. using the compacted format
400 * Use this function when dealing with the compacted format so
401 * that it is obvious which aspect of 'XSAVES' is being handled
402 * by the calling code.
404 static int using_compacted_format(void)
406 return cpu_has_xsaves
;
409 static void __xstate_dump_leaves(void)
412 u32 eax
, ebx
, ecx
, edx
;
413 static int should_dump
= 1;
419 * Dump out a few leaves past the ones that we support
420 * just in case there are some goodies up there
422 for (i
= 0; i
< XFEATURE_MAX
+ 10; i
++) {
423 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
424 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
425 XSTATE_CPUID
, i
, eax
, ebx
, ecx
, edx
);
429 #define XSTATE_WARN_ON(x) do { \
430 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
431 __xstate_dump_leaves(); \
435 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
436 if ((nr == nr_macro) && \
437 WARN_ONCE(sz != sizeof(__struct), \
438 "%s: struct is %zu bytes, cpu state %d bytes\n", \
439 __stringify(nr_macro), sizeof(__struct), sz)) { \
440 __xstate_dump_leaves(); \
445 * We have a C struct for each 'xstate'. We need to ensure
446 * that our software representation matches what the CPU
447 * tells us about the state's size.
449 static void check_xstate_against_struct(int nr
)
452 * Ask the CPU for the size of the state.
454 int sz
= xfeature_size(nr
);
456 * Match each CPU state with the corresponding software
459 XCHECK_SZ(sz
, nr
, XFEATURE_YMM
, struct ymmh_struct
);
460 XCHECK_SZ(sz
, nr
, XFEATURE_BNDREGS
, struct mpx_bndreg_state
);
461 XCHECK_SZ(sz
, nr
, XFEATURE_BNDCSR
, struct mpx_bndcsr_state
);
462 XCHECK_SZ(sz
, nr
, XFEATURE_OPMASK
, struct avx_512_opmask_state
);
463 XCHECK_SZ(sz
, nr
, XFEATURE_ZMM_Hi256
, struct avx_512_zmm_uppers_state
);
464 XCHECK_SZ(sz
, nr
, XFEATURE_Hi16_ZMM
, struct avx_512_hi16_state
);
467 * Make *SURE* to add any feature numbers in below if
468 * there are "holes" in the xsave state component
471 if ((nr
< XFEATURE_YMM
) ||
472 (nr
>= XFEATURE_MAX
)) {
473 WARN_ONCE(1, "no structure for xstate: %d\n", nr
);
479 * This essentially double-checks what the cpu told us about
480 * how large the XSAVE buffer needs to be. We are recalculating
483 static void do_extra_xstate_size_checks(void)
485 int paranoid_xstate_size
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
488 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
489 if (!xfeature_enabled(i
))
492 check_xstate_against_struct(i
);
494 * Supervisor state components can be managed only by
495 * XSAVES, which is compacted-format only.
497 if (!using_compacted_format())
498 XSTATE_WARN_ON(xfeature_is_supervisor(i
));
500 /* Align from the end of the previous feature */
501 if (xfeature_is_aligned(i
))
502 paranoid_xstate_size
= ALIGN(paranoid_xstate_size
, 64);
504 * The offset of a given state in the non-compacted
505 * format is given to us in a CPUID leaf. We check
506 * them for being ordered (increasing offsets) in
507 * setup_xstate_features().
509 if (!using_compacted_format())
510 paranoid_xstate_size
= xfeature_uncompacted_offset(i
);
512 * The compacted-format offset always depends on where
513 * the previous state ended.
515 paranoid_xstate_size
+= xfeature_size(i
);
517 XSTATE_WARN_ON(paranoid_xstate_size
!= xstate_size
);
521 * Calculate total size of enabled xstates in XCR0/xfeatures_mask.
523 * Note the SDM's wording here. "sub-function 0" only enumerates
524 * the size of the *user* states. If we use it to size a buffer
525 * that we use 'XSAVES' on, we could potentially overflow the
526 * buffer because 'XSAVES' saves system states too.
528 * Note that we do not currently set any bits on IA32_XSS so
529 * 'XCR0 | IA32_XSS == XCR0' for now.
531 static unsigned int __init
calculate_xstate_size(void)
533 unsigned int eax
, ebx
, ecx
, edx
;
534 unsigned int calculated_xstate_size
;
536 if (!cpu_has_xsaves
) {
538 * - CPUID function 0DH, sub-function 0:
539 * EBX enumerates the size (in bytes) required by
540 * the XSAVE instruction for an XSAVE area
541 * containing all the *user* state components
542 * corresponding to bits currently set in XCR0.
544 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
545 calculated_xstate_size
= ebx
;
548 * - CPUID function 0DH, sub-function 1:
549 * EBX enumerates the size (in bytes) required by
550 * the XSAVES instruction for an XSAVE area
551 * containing all the state components
552 * corresponding to bits currently set in
555 cpuid_count(XSTATE_CPUID
, 1, &eax
, &ebx
, &ecx
, &edx
);
556 calculated_xstate_size
= ebx
;
558 return calculated_xstate_size
;
562 * Will the runtime-enumerated 'xstate_size' fit in the init
563 * task's statically-allocated buffer?
565 static bool is_supported_xstate_size(unsigned int test_xstate_size
)
567 if (test_xstate_size
<= sizeof(union fpregs_state
))
570 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
571 sizeof(union fpregs_state
), test_xstate_size
);
575 static int init_xstate_size(void)
577 /* Recompute the context size for enabled features: */
578 unsigned int possible_xstate_size
= calculate_xstate_size();
580 /* Ensure we have the space to store all enabled: */
581 if (!is_supported_xstate_size(possible_xstate_size
))
585 * The size is OK, we are definitely going to use xsave,
586 * make it known to the world that we need more space.
588 xstate_size
= possible_xstate_size
;
589 do_extra_xstate_size_checks();
594 * We enabled the XSAVE hardware, but something went wrong and
595 * we can not use it. Disable it.
597 static void fpu__init_disable_system_xstate(void)
600 cr4_clear_bits(X86_CR4_OSXSAVE
);
601 fpu__xstate_clear_all_cpu_caps();
605 * Enable and initialize the xsave feature.
606 * Called once per system bootup.
608 void __init
fpu__init_system_xstate(void)
610 unsigned int eax
, ebx
, ecx
, edx
;
611 static int on_boot_cpu __initdata
= 1;
614 WARN_ON_FPU(!on_boot_cpu
);
617 if (!cpu_has_xsave
) {
618 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
622 if (boot_cpu_data
.cpuid_level
< XSTATE_CPUID
) {
627 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
628 xfeatures_mask
= eax
+ ((u64
)edx
<< 32);
630 if ((xfeatures_mask
& XFEATURE_MASK_FPSSE
) != XFEATURE_MASK_FPSSE
) {
631 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask
);
635 /* Support only the state known to the OS: */
636 xfeatures_mask
= xfeatures_mask
& XCNTXT_MASK
;
638 /* Enable xstate instructions to be able to continue with initialization: */
639 fpu__init_cpu_xstate();
640 err
= init_xstate_size();
642 /* something went wrong, boot without any XSAVE support */
643 fpu__init_disable_system_xstate();
647 update_regset_xstate_info(xstate_size
, xfeatures_mask
);
648 fpu__init_prepare_fx_sw_frame();
649 setup_init_fpu_buf();
652 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
655 cpu_has_xsaves
? "compacted" : "standard");
659 * Restore minimal FPU state after suspend:
661 void fpu__resume_cpu(void)
664 * Restore XCR0 on xsave capable CPUs:
667 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
671 * Given the xsave area and a state inside, this function returns the
672 * address of the state.
674 * This is the API that is called to get xstate address in either
675 * standard format or compacted format of xsave area.
677 * Note that if there is no data for the field in the xsave buffer
678 * this will return NULL.
681 * xstate: the thread's storage area for all FPU data
682 * xstate_feature: state which is defined in xsave.h (e.g.
683 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
685 * address of the state in the xsave area, or NULL if the
686 * field is not present in the xsave buffer.
688 void *get_xsave_addr(struct xregs_state
*xsave
, int xstate_feature
)
690 int feature_nr
= fls64(xstate_feature
) - 1;
692 * Do we even *have* xsave state?
694 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
698 * We should not ever be requesting features that we
699 * have not enabled. Remember that pcntxt_mask is
700 * what we write to the XCR0 register.
702 WARN_ONCE(!(xfeatures_mask
& xstate_feature
),
703 "get of unsupported state");
705 * This assumes the last 'xsave*' instruction to
706 * have requested that 'xstate_feature' be saved.
707 * If it did not, we might be seeing and old value
708 * of the field in the buffer.
710 * This can happen because the last 'xsave' did not
711 * request that this feature be saved (unlikely)
712 * or because the "init optimization" caused it
715 if (!(xsave
->header
.xfeatures
& xstate_feature
))
718 return (void *)xsave
+ xstate_comp_offsets
[feature_nr
];
720 EXPORT_SYMBOL_GPL(get_xsave_addr
);
723 * This wraps up the common operations that need to occur when retrieving
724 * data from xsave state. It first ensures that the current task was
725 * using the FPU and retrieves the data in to a buffer. It then calculates
726 * the offset of the requested field in the buffer.
728 * This function is safe to call whether the FPU is in use or not.
730 * Note that this only works on the current task.
733 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
734 * XFEATURE_MASK_SSE, etc...)
736 * address of the state in the xsave area or NULL if the state
737 * is not present or is in its 'init state'.
739 const void *get_xsave_field_ptr(int xsave_state
)
741 struct fpu
*fpu
= ¤t
->thread
.fpu
;
743 if (!fpu
->fpstate_active
)
746 * fpu__save() takes the CPU's xstate registers
747 * and saves them off to the 'fpu memory buffer.
751 return get_xsave_addr(&fpu
->state
.xsave
, xsave_state
);