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>
17 * Although we spell it out in here, the Processor Trace
18 * xfeature is completely unused. We use other mechanisms
19 * to save/restore PT state in Linux.
21 static const char *xfeature_names
[] =
23 "x87 floating point registers" ,
26 "MPX bounds registers" ,
31 "Processor Trace (unused)" ,
32 "Protection Keys User registers",
33 "unknown xstate feature" ,
37 * Mask of xstate features supported by the CPU and the kernel:
39 u64 xfeatures_mask __read_mostly
;
41 static unsigned int xstate_offsets
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
42 static unsigned int xstate_sizes
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
43 static unsigned int xstate_comp_offsets
[sizeof(xfeatures_mask
)*8];
46 * Clear all of the X86_FEATURE_* bits that are unavailable
47 * when the CPU has no XSAVE support.
49 void fpu__xstate_clear_all_cpu_caps(void)
51 setup_clear_cpu_cap(X86_FEATURE_XSAVE
);
52 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT
);
53 setup_clear_cpu_cap(X86_FEATURE_XSAVEC
);
54 setup_clear_cpu_cap(X86_FEATURE_XSAVES
);
55 setup_clear_cpu_cap(X86_FEATURE_AVX
);
56 setup_clear_cpu_cap(X86_FEATURE_AVX2
);
57 setup_clear_cpu_cap(X86_FEATURE_AVX512F
);
58 setup_clear_cpu_cap(X86_FEATURE_AVX512PF
);
59 setup_clear_cpu_cap(X86_FEATURE_AVX512ER
);
60 setup_clear_cpu_cap(X86_FEATURE_AVX512CD
);
61 setup_clear_cpu_cap(X86_FEATURE_MPX
);
62 setup_clear_cpu_cap(X86_FEATURE_XGETBV1
);
63 setup_clear_cpu_cap(X86_FEATURE_PKU
);
67 * Return whether the system supports a given xfeature.
69 * Also return the name of the (most advanced) feature that the caller requested:
71 int cpu_has_xfeatures(u64 xfeatures_needed
, const char **feature_name
)
73 u64 xfeatures_missing
= xfeatures_needed
& ~xfeatures_mask
;
75 if (unlikely(feature_name
)) {
76 long xfeature_idx
, max_idx
;
79 * So we use FLS here to be able to print the most advanced
80 * feature that was requested but is missing. So if a driver
81 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
82 * missing AVX feature - this is the most informative message
85 if (xfeatures_missing
)
86 xfeatures_print
= xfeatures_missing
;
88 xfeatures_print
= xfeatures_needed
;
90 xfeature_idx
= fls64(xfeatures_print
)-1;
91 max_idx
= ARRAY_SIZE(xfeature_names
)-1;
92 xfeature_idx
= min(xfeature_idx
, max_idx
);
94 *feature_name
= xfeature_names
[xfeature_idx
];
97 if (xfeatures_missing
)
102 EXPORT_SYMBOL_GPL(cpu_has_xfeatures
);
105 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
106 * a processor implementation detects that an FPU state component is still
107 * (or is again) in its initialized state, it may clear the corresponding
108 * bit in the header.xfeatures field, and can skip the writeout of registers
109 * to the corresponding memory layout.
111 * This means that when the bit is zero, the state component might still contain
112 * some previous - non-initialized register state.
114 * Before writing xstate information to user-space we sanitize those components,
115 * to always ensure that the memory layout of a feature will be in the init state
116 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
117 * see some stale state in the memory layout during signal handling, debugging etc.
119 void fpstate_sanitize_xstate(struct fpu
*fpu
)
121 struct fxregs_state
*fx
= &fpu
->state
.fxsave
;
128 xfeatures
= fpu
->state
.xsave
.header
.xfeatures
;
131 * None of the feature bits are in init state. So nothing else
132 * to do for us, as the memory layout is up to date.
134 if ((xfeatures
& xfeatures_mask
) == xfeatures_mask
)
138 * FP is in init state
140 if (!(xfeatures
& XFEATURE_MASK_FP
)) {
147 memset(&fx
->st_space
[0], 0, 128);
151 * SSE is in init state
153 if (!(xfeatures
& XFEATURE_MASK_SSE
))
154 memset(&fx
->xmm_space
[0], 0, 256);
157 * First two features are FPU and SSE, which above we handled
158 * in a special way already:
161 xfeatures
= (xfeatures_mask
& ~xfeatures
) >> 2;
164 * Update all the remaining memory layouts according to their
165 * standard xstate layout, if their header bit is in the init
169 if (xfeatures
& 0x1) {
170 int offset
= xstate_offsets
[feature_bit
];
171 int size
= xstate_sizes
[feature_bit
];
173 memcpy((void *)fx
+ offset
,
174 (void *)&init_fpstate
.xsave
+ offset
,
184 * Enable the extended processor state save/restore feature.
185 * Called once per CPU onlining.
187 void fpu__init_cpu_xstate(void)
189 if (!cpu_has_xsave
|| !xfeatures_mask
)
192 cr4_set_bits(X86_CR4_OSXSAVE
);
193 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
197 * Note that in the future we will likely need a pair of
198 * functions here: one for user xstates and the other for
199 * system xstates. For now, they are the same.
201 static int xfeature_enabled(enum xfeature xfeature
)
203 return !!(xfeatures_mask
& (1UL << xfeature
));
207 * Record the offsets and sizes of various xstates contained
208 * in the XSAVE state memory layout.
210 static void __init
setup_xstate_features(void)
212 u32 eax
, ebx
, ecx
, edx
, i
;
213 /* start at the beginnning of the "extended state" */
214 unsigned int last_good_offset
= offsetof(struct xregs_state
,
215 extended_state_area
);
217 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
218 if (!xfeature_enabled(i
))
221 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
222 xstate_offsets
[i
] = ebx
;
223 xstate_sizes
[i
] = eax
;
225 * In our xstate size checks, we assume that the
226 * highest-numbered xstate feature has the
227 * highest offset in the buffer. Ensure it does.
229 WARN_ONCE(last_good_offset
> xstate_offsets
[i
],
230 "x86/fpu: misordered xstate at %d\n", last_good_offset
);
231 last_good_offset
= xstate_offsets
[i
];
233 printk(KERN_INFO
"x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n", i
, ebx
, i
, eax
);
237 static void __init
print_xstate_feature(u64 xstate_mask
)
239 const char *feature_name
;
241 if (cpu_has_xfeatures(xstate_mask
, &feature_name
))
242 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask
, feature_name
);
246 * Print out all the supported xstate features:
248 static void __init
print_xstate_features(void)
250 print_xstate_feature(XFEATURE_MASK_FP
);
251 print_xstate_feature(XFEATURE_MASK_SSE
);
252 print_xstate_feature(XFEATURE_MASK_YMM
);
253 print_xstate_feature(XFEATURE_MASK_BNDREGS
);
254 print_xstate_feature(XFEATURE_MASK_BNDCSR
);
255 print_xstate_feature(XFEATURE_MASK_OPMASK
);
256 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256
);
257 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM
);
258 print_xstate_feature(XFEATURE_MASK_PKRU
);
262 * This function sets up offsets and sizes of all extended states in
263 * xsave area. This supports both standard format and compacted format
264 * of the xsave aread.
266 static void __init
setup_xstate_comp(void)
268 unsigned int xstate_comp_sizes
[sizeof(xfeatures_mask
)*8];
272 * The FP xstates and SSE xstates are legacy states. They are always
273 * in the fixed offsets in the xsave area in either compacted form
276 xstate_comp_offsets
[0] = 0;
277 xstate_comp_offsets
[1] = offsetof(struct fxregs_state
, xmm_space
);
279 if (!cpu_has_xsaves
) {
280 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
281 if (xfeature_enabled(i
)) {
282 xstate_comp_offsets
[i
] = xstate_offsets
[i
];
283 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
289 xstate_comp_offsets
[FIRST_EXTENDED_XFEATURE
] =
290 FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
292 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
293 if (xfeature_enabled(i
))
294 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
296 xstate_comp_sizes
[i
] = 0;
298 if (i
> FIRST_EXTENDED_XFEATURE
)
299 xstate_comp_offsets
[i
] = xstate_comp_offsets
[i
-1]
300 + xstate_comp_sizes
[i
-1];
306 * setup the xstate image representing the init state
308 static void __init
setup_init_fpu_buf(void)
310 static int on_boot_cpu __initdata
= 1;
312 WARN_ON_FPU(!on_boot_cpu
);
318 setup_xstate_features();
319 print_xstate_features();
321 if (cpu_has_xsaves
) {
322 init_fpstate
.xsave
.header
.xcomp_bv
= (u64
)1 << 63 | xfeatures_mask
;
323 init_fpstate
.xsave
.header
.xfeatures
= xfeatures_mask
;
327 * Init all the features state with header_bv being 0x0
329 copy_kernel_to_xregs_booting(&init_fpstate
.xsave
);
332 * Dump the init state again. This is to identify the init state
333 * of any feature which is not represented by all zero's.
335 copy_xregs_to_kernel_booting(&init_fpstate
.xsave
);
338 static int xfeature_is_supervisor(int xfeature_nr
)
341 * We currently do not support supervisor states, but if
342 * we did, we could find out like this.
344 * SDM says: If state component i is a user state component,
345 * ECX[0] return 0; if state component i is a supervisor
346 * state component, ECX[0] returns 1.
347 u32 eax, ebx, ecx, edx;
348 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx;
354 static int xfeature_is_user(int xfeature_nr)
356 return !xfeature_is_supervisor(xfeature_nr);
361 * This check is important because it is easy to get XSTATE_*
362 * confused with XSTATE_BIT_*.
364 #define CHECK_XFEATURE(nr) do { \
365 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
366 WARN_ON(nr >= XFEATURE_MAX); \
370 * We could cache this like xstate_size[], but we only use
371 * it here, so it would be a waste of space.
373 static int xfeature_is_aligned(int xfeature_nr
)
375 u32 eax
, ebx
, ecx
, edx
;
377 CHECK_XFEATURE(xfeature_nr
);
378 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
380 * The value returned by ECX[1] indicates the alignment
381 * of state component i when the compacted format
382 * of the extended region of an XSAVE area is used
387 static int xfeature_uncompacted_offset(int xfeature_nr
)
389 u32 eax
, ebx
, ecx
, edx
;
391 CHECK_XFEATURE(xfeature_nr
);
392 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
396 static int xfeature_size(int xfeature_nr
)
398 u32 eax
, ebx
, ecx
, edx
;
400 CHECK_XFEATURE(xfeature_nr
);
401 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
406 * 'XSAVES' implies two different things:
407 * 1. saving of supervisor/system state
408 * 2. using the compacted format
410 * Use this function when dealing with the compacted format so
411 * that it is obvious which aspect of 'XSAVES' is being handled
412 * by the calling code.
414 static int using_compacted_format(void)
416 return cpu_has_xsaves
;
419 static void __xstate_dump_leaves(void)
422 u32 eax
, ebx
, ecx
, edx
;
423 static int should_dump
= 1;
429 * Dump out a few leaves past the ones that we support
430 * just in case there are some goodies up there
432 for (i
= 0; i
< XFEATURE_MAX
+ 10; i
++) {
433 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
434 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
435 XSTATE_CPUID
, i
, eax
, ebx
, ecx
, edx
);
439 #define XSTATE_WARN_ON(x) do { \
440 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
441 __xstate_dump_leaves(); \
445 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
446 if ((nr == nr_macro) && \
447 WARN_ONCE(sz != sizeof(__struct), \
448 "%s: struct is %zu bytes, cpu state %d bytes\n", \
449 __stringify(nr_macro), sizeof(__struct), sz)) { \
450 __xstate_dump_leaves(); \
455 * We have a C struct for each 'xstate'. We need to ensure
456 * that our software representation matches what the CPU
457 * tells us about the state's size.
459 static void check_xstate_against_struct(int nr
)
462 * Ask the CPU for the size of the state.
464 int sz
= xfeature_size(nr
);
466 * Match each CPU state with the corresponding software
469 XCHECK_SZ(sz
, nr
, XFEATURE_YMM
, struct ymmh_struct
);
470 XCHECK_SZ(sz
, nr
, XFEATURE_BNDREGS
, struct mpx_bndreg_state
);
471 XCHECK_SZ(sz
, nr
, XFEATURE_BNDCSR
, struct mpx_bndcsr_state
);
472 XCHECK_SZ(sz
, nr
, XFEATURE_OPMASK
, struct avx_512_opmask_state
);
473 XCHECK_SZ(sz
, nr
, XFEATURE_ZMM_Hi256
, struct avx_512_zmm_uppers_state
);
474 XCHECK_SZ(sz
, nr
, XFEATURE_Hi16_ZMM
, struct avx_512_hi16_state
);
475 XCHECK_SZ(sz
, nr
, XFEATURE_PKRU
, struct pkru_state
);
478 * Make *SURE* to add any feature numbers in below if
479 * there are "holes" in the xsave state component
482 if ((nr
< XFEATURE_YMM
) ||
483 (nr
>= XFEATURE_MAX
) ||
484 (nr
== XFEATURE_PT_UNIMPLEMENTED_SO_FAR
)) {
485 WARN_ONCE(1, "no structure for xstate: %d\n", nr
);
491 * This essentially double-checks what the cpu told us about
492 * how large the XSAVE buffer needs to be. We are recalculating
495 static void do_extra_xstate_size_checks(void)
497 int paranoid_xstate_size
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
500 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
501 if (!xfeature_enabled(i
))
504 check_xstate_against_struct(i
);
506 * Supervisor state components can be managed only by
507 * XSAVES, which is compacted-format only.
509 if (!using_compacted_format())
510 XSTATE_WARN_ON(xfeature_is_supervisor(i
));
512 /* Align from the end of the previous feature */
513 if (xfeature_is_aligned(i
))
514 paranoid_xstate_size
= ALIGN(paranoid_xstate_size
, 64);
516 * The offset of a given state in the non-compacted
517 * format is given to us in a CPUID leaf. We check
518 * them for being ordered (increasing offsets) in
519 * setup_xstate_features().
521 if (!using_compacted_format())
522 paranoid_xstate_size
= xfeature_uncompacted_offset(i
);
524 * The compacted-format offset always depends on where
525 * the previous state ended.
527 paranoid_xstate_size
+= xfeature_size(i
);
529 XSTATE_WARN_ON(paranoid_xstate_size
!= xstate_size
);
533 * Calculate total size of enabled xstates in XCR0/xfeatures_mask.
535 * Note the SDM's wording here. "sub-function 0" only enumerates
536 * the size of the *user* states. If we use it to size a buffer
537 * that we use 'XSAVES' on, we could potentially overflow the
538 * buffer because 'XSAVES' saves system states too.
540 * Note that we do not currently set any bits on IA32_XSS so
541 * 'XCR0 | IA32_XSS == XCR0' for now.
543 static unsigned int __init
calculate_xstate_size(void)
545 unsigned int eax
, ebx
, ecx
, edx
;
546 unsigned int calculated_xstate_size
;
548 if (!cpu_has_xsaves
) {
550 * - CPUID function 0DH, sub-function 0:
551 * EBX enumerates the size (in bytes) required by
552 * the XSAVE instruction for an XSAVE area
553 * containing all the *user* state components
554 * corresponding to bits currently set in XCR0.
556 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
557 calculated_xstate_size
= ebx
;
560 * - CPUID function 0DH, sub-function 1:
561 * EBX enumerates the size (in bytes) required by
562 * the XSAVES instruction for an XSAVE area
563 * containing all the state components
564 * corresponding to bits currently set in
567 cpuid_count(XSTATE_CPUID
, 1, &eax
, &ebx
, &ecx
, &edx
);
568 calculated_xstate_size
= ebx
;
570 return calculated_xstate_size
;
574 * Will the runtime-enumerated 'xstate_size' fit in the init
575 * task's statically-allocated buffer?
577 static bool is_supported_xstate_size(unsigned int test_xstate_size
)
579 if (test_xstate_size
<= sizeof(union fpregs_state
))
582 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
583 sizeof(union fpregs_state
), test_xstate_size
);
587 static int init_xstate_size(void)
589 /* Recompute the context size for enabled features: */
590 unsigned int possible_xstate_size
= calculate_xstate_size();
592 /* Ensure we have the space to store all enabled: */
593 if (!is_supported_xstate_size(possible_xstate_size
))
597 * The size is OK, we are definitely going to use xsave,
598 * make it known to the world that we need more space.
600 xstate_size
= possible_xstate_size
;
601 do_extra_xstate_size_checks();
606 * We enabled the XSAVE hardware, but something went wrong and
607 * we can not use it. Disable it.
609 static void fpu__init_disable_system_xstate(void)
612 cr4_clear_bits(X86_CR4_OSXSAVE
);
613 fpu__xstate_clear_all_cpu_caps();
617 * Enable and initialize the xsave feature.
618 * Called once per system bootup.
620 void __init
fpu__init_system_xstate(void)
622 unsigned int eax
, ebx
, ecx
, edx
;
623 static int on_boot_cpu __initdata
= 1;
626 WARN_ON_FPU(!on_boot_cpu
);
629 if (!cpu_has_xsave
) {
630 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
634 if (boot_cpu_data
.cpuid_level
< XSTATE_CPUID
) {
639 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
640 xfeatures_mask
= eax
+ ((u64
)edx
<< 32);
642 if ((xfeatures_mask
& XFEATURE_MASK_FPSSE
) != XFEATURE_MASK_FPSSE
) {
643 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask
);
647 xfeatures_mask
&= fpu__get_supported_xfeatures_mask();
649 /* Enable xstate instructions to be able to continue with initialization: */
650 fpu__init_cpu_xstate();
651 err
= init_xstate_size();
653 /* something went wrong, boot without any XSAVE support */
654 fpu__init_disable_system_xstate();
658 update_regset_xstate_info(xstate_size
, xfeatures_mask
);
659 fpu__init_prepare_fx_sw_frame();
660 setup_init_fpu_buf();
663 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
666 cpu_has_xsaves
? "compacted" : "standard");
670 * Restore minimal FPU state after suspend:
672 void fpu__resume_cpu(void)
675 * Restore XCR0 on xsave capable CPUs:
678 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
682 * Given the xsave area and a state inside, this function returns the
683 * address of the state.
685 * This is the API that is called to get xstate address in either
686 * standard format or compacted format of xsave area.
688 * Note that if there is no data for the field in the xsave buffer
689 * this will return NULL.
692 * xstate: the thread's storage area for all FPU data
693 * xstate_feature: state which is defined in xsave.h (e.g.
694 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
696 * address of the state in the xsave area, or NULL if the
697 * field is not present in the xsave buffer.
699 void *get_xsave_addr(struct xregs_state
*xsave
, int xstate_feature
)
701 int feature_nr
= fls64(xstate_feature
) - 1;
703 * Do we even *have* xsave state?
705 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
709 * We should not ever be requesting features that we
710 * have not enabled. Remember that pcntxt_mask is
711 * what we write to the XCR0 register.
713 WARN_ONCE(!(xfeatures_mask
& xstate_feature
),
714 "get of unsupported state");
716 * This assumes the last 'xsave*' instruction to
717 * have requested that 'xstate_feature' be saved.
718 * If it did not, we might be seeing and old value
719 * of the field in the buffer.
721 * This can happen because the last 'xsave' did not
722 * request that this feature be saved (unlikely)
723 * or because the "init optimization" caused it
726 if (!(xsave
->header
.xfeatures
& xstate_feature
))
729 return (void *)xsave
+ xstate_comp_offsets
[feature_nr
];
731 EXPORT_SYMBOL_GPL(get_xsave_addr
);
734 * This wraps up the common operations that need to occur when retrieving
735 * data from xsave state. It first ensures that the current task was
736 * using the FPU and retrieves the data in to a buffer. It then calculates
737 * the offset of the requested field in the buffer.
739 * This function is safe to call whether the FPU is in use or not.
741 * Note that this only works on the current task.
744 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
745 * XFEATURE_MASK_SSE, etc...)
747 * address of the state in the xsave area or NULL if the state
748 * is not present or is in its 'init state'.
750 const void *get_xsave_field_ptr(int xsave_state
)
752 struct fpu
*fpu
= ¤t
->thread
.fpu
;
754 if (!fpu
->fpstate_active
)
757 * fpu__save() takes the CPU's xstate registers
758 * and saves them off to the 'fpu memory buffer.
762 return get_xsave_addr(&fpu
->state
.xsave
, xsave_state
);