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)" ,
35 * Mask of xstate features supported by the CPU and the kernel:
37 u64 xfeatures_mask __read_mostly
;
39 static unsigned int xstate_offsets
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
40 static unsigned int xstate_sizes
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
41 static unsigned int xstate_comp_offsets
[sizeof(xfeatures_mask
)*8];
44 * Clear all of the X86_FEATURE_* bits that are unavailable
45 * when the CPU has no XSAVE support.
47 void fpu__xstate_clear_all_cpu_caps(void)
49 setup_clear_cpu_cap(X86_FEATURE_XSAVE
);
50 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT
);
51 setup_clear_cpu_cap(X86_FEATURE_XSAVEC
);
52 setup_clear_cpu_cap(X86_FEATURE_XSAVES
);
53 setup_clear_cpu_cap(X86_FEATURE_AVX
);
54 setup_clear_cpu_cap(X86_FEATURE_AVX2
);
55 setup_clear_cpu_cap(X86_FEATURE_AVX512F
);
56 setup_clear_cpu_cap(X86_FEATURE_AVX512PF
);
57 setup_clear_cpu_cap(X86_FEATURE_AVX512ER
);
58 setup_clear_cpu_cap(X86_FEATURE_AVX512CD
);
59 setup_clear_cpu_cap(X86_FEATURE_MPX
);
60 setup_clear_cpu_cap(X86_FEATURE_XGETBV1
);
64 * Return whether the system supports a given xfeature.
66 * Also return the name of the (most advanced) feature that the caller requested:
68 int cpu_has_xfeatures(u64 xfeatures_needed
, const char **feature_name
)
70 u64 xfeatures_missing
= xfeatures_needed
& ~xfeatures_mask
;
72 if (unlikely(feature_name
)) {
73 long xfeature_idx
, max_idx
;
76 * So we use FLS here to be able to print the most advanced
77 * feature that was requested but is missing. So if a driver
78 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
79 * missing AVX feature - this is the most informative message
82 if (xfeatures_missing
)
83 xfeatures_print
= xfeatures_missing
;
85 xfeatures_print
= xfeatures_needed
;
87 xfeature_idx
= fls64(xfeatures_print
)-1;
88 max_idx
= ARRAY_SIZE(xfeature_names
)-1;
89 xfeature_idx
= min(xfeature_idx
, max_idx
);
91 *feature_name
= xfeature_names
[xfeature_idx
];
94 if (xfeatures_missing
)
99 EXPORT_SYMBOL_GPL(cpu_has_xfeatures
);
102 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
103 * a processor implementation detects that an FPU state component is still
104 * (or is again) in its initialized state, it may clear the corresponding
105 * bit in the header.xfeatures field, and can skip the writeout of registers
106 * to the corresponding memory layout.
108 * This means that when the bit is zero, the state component might still contain
109 * some previous - non-initialized register state.
111 * Before writing xstate information to user-space we sanitize those components,
112 * to always ensure that the memory layout of a feature will be in the init state
113 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
114 * see some stale state in the memory layout during signal handling, debugging etc.
116 void fpstate_sanitize_xstate(struct fpu
*fpu
)
118 struct fxregs_state
*fx
= &fpu
->state
.fxsave
;
125 xfeatures
= fpu
->state
.xsave
.header
.xfeatures
;
128 * None of the feature bits are in init state. So nothing else
129 * to do for us, as the memory layout is up to date.
131 if ((xfeatures
& xfeatures_mask
) == xfeatures_mask
)
135 * FP is in init state
137 if (!(xfeatures
& XFEATURE_MASK_FP
)) {
144 memset(&fx
->st_space
[0], 0, 128);
148 * SSE is in init state
150 if (!(xfeatures
& XFEATURE_MASK_SSE
))
151 memset(&fx
->xmm_space
[0], 0, 256);
154 * First two features are FPU and SSE, which above we handled
155 * in a special way already:
158 xfeatures
= (xfeatures_mask
& ~xfeatures
) >> 2;
161 * Update all the remaining memory layouts according to their
162 * standard xstate layout, if their header bit is in the init
166 if (xfeatures
& 0x1) {
167 int offset
= xstate_offsets
[feature_bit
];
168 int size
= xstate_sizes
[feature_bit
];
170 memcpy((void *)fx
+ offset
,
171 (void *)&init_fpstate
.xsave
+ offset
,
181 * Enable the extended processor state save/restore feature.
182 * Called once per CPU onlining.
184 void fpu__init_cpu_xstate(void)
186 if (!cpu_has_xsave
|| !xfeatures_mask
)
189 cr4_set_bits(X86_CR4_OSXSAVE
);
190 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
194 * Note that in the future we will likely need a pair of
195 * functions here: one for user xstates and the other for
196 * system xstates. For now, they are the same.
198 static int xfeature_enabled(enum xfeature xfeature
)
200 return !!(xfeatures_mask
& (1UL << xfeature
));
204 * Record the offsets and sizes of various xstates contained
205 * in the XSAVE state memory layout.
207 static void __init
setup_xstate_features(void)
209 u32 eax
, ebx
, ecx
, edx
, i
;
210 /* start at the beginnning of the "extended state" */
211 unsigned int last_good_offset
= offsetof(struct xregs_state
,
212 extended_state_area
);
214 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
215 if (!xfeature_enabled(i
))
218 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
219 xstate_offsets
[i
] = ebx
;
220 xstate_sizes
[i
] = eax
;
222 * In our xstate size checks, we assume that the
223 * highest-numbered xstate feature has the
224 * highest offset in the buffer. Ensure it does.
226 WARN_ONCE(last_good_offset
> xstate_offsets
[i
],
227 "x86/fpu: misordered xstate at %d\n", last_good_offset
);
228 last_good_offset
= xstate_offsets
[i
];
230 printk(KERN_INFO
"x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n", i
, ebx
, i
, eax
);
234 static void __init
print_xstate_feature(u64 xstate_mask
)
236 const char *feature_name
;
238 if (cpu_has_xfeatures(xstate_mask
, &feature_name
))
239 pr_info("x86/fpu: Supporting XSAVE feature 0x%02Lx: '%s'\n", xstate_mask
, feature_name
);
243 * Print out all the supported xstate features:
245 static void __init
print_xstate_features(void)
247 print_xstate_feature(XFEATURE_MASK_FP
);
248 print_xstate_feature(XFEATURE_MASK_SSE
);
249 print_xstate_feature(XFEATURE_MASK_YMM
);
250 print_xstate_feature(XFEATURE_MASK_BNDREGS
);
251 print_xstate_feature(XFEATURE_MASK_BNDCSR
);
252 print_xstate_feature(XFEATURE_MASK_OPMASK
);
253 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256
);
254 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM
);
258 * This function sets up offsets and sizes of all extended states in
259 * xsave area. This supports both standard format and compacted format
260 * of the xsave aread.
262 static void __init
setup_xstate_comp(void)
264 unsigned int xstate_comp_sizes
[sizeof(xfeatures_mask
)*8];
268 * The FP xstates and SSE xstates are legacy states. They are always
269 * in the fixed offsets in the xsave area in either compacted form
272 xstate_comp_offsets
[0] = 0;
273 xstate_comp_offsets
[1] = offsetof(struct fxregs_state
, xmm_space
);
275 if (!cpu_has_xsaves
) {
276 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
277 if (xfeature_enabled(i
)) {
278 xstate_comp_offsets
[i
] = xstate_offsets
[i
];
279 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
285 xstate_comp_offsets
[FIRST_EXTENDED_XFEATURE
] =
286 FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
288 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
289 if (xfeature_enabled(i
))
290 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
292 xstate_comp_sizes
[i
] = 0;
294 if (i
> FIRST_EXTENDED_XFEATURE
)
295 xstate_comp_offsets
[i
] = xstate_comp_offsets
[i
-1]
296 + xstate_comp_sizes
[i
-1];
302 * setup the xstate image representing the init state
304 static void __init
setup_init_fpu_buf(void)
306 static int on_boot_cpu __initdata
= 1;
308 WARN_ON_FPU(!on_boot_cpu
);
314 setup_xstate_features();
315 print_xstate_features();
317 if (cpu_has_xsaves
) {
318 init_fpstate
.xsave
.header
.xcomp_bv
= (u64
)1 << 63 | xfeatures_mask
;
319 init_fpstate
.xsave
.header
.xfeatures
= xfeatures_mask
;
323 * Init all the features state with header_bv being 0x0
325 copy_kernel_to_xregs_booting(&init_fpstate
.xsave
);
328 * Dump the init state again. This is to identify the init state
329 * of any feature which is not represented by all zero's.
331 copy_xregs_to_kernel_booting(&init_fpstate
.xsave
);
334 static int xfeature_is_supervisor(int xfeature_nr
)
337 * We currently do not support supervisor states, but if
338 * we did, we could find out like this.
340 * SDM says: If state component i is a user state component,
341 * ECX[0] return 0; if state component i is a supervisor
342 * state component, ECX[0] returns 1.
343 u32 eax, ebx, ecx, edx;
344 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx;
350 static int xfeature_is_user(int xfeature_nr)
352 return !xfeature_is_supervisor(xfeature_nr);
357 * This check is important because it is easy to get XSTATE_*
358 * confused with XSTATE_BIT_*.
360 #define CHECK_XFEATURE(nr) do { \
361 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
362 WARN_ON(nr >= XFEATURE_MAX); \
366 * We could cache this like xstate_size[], but we only use
367 * it here, so it would be a waste of space.
369 static int xfeature_is_aligned(int xfeature_nr
)
371 u32 eax
, ebx
, ecx
, edx
;
373 CHECK_XFEATURE(xfeature_nr
);
374 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
376 * The value returned by ECX[1] indicates the alignment
377 * of state component i when the compacted format
378 * of the extended region of an XSAVE area is used
383 static int xfeature_uncompacted_offset(int xfeature_nr
)
385 u32 eax
, ebx
, ecx
, edx
;
387 CHECK_XFEATURE(xfeature_nr
);
388 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
392 static int xfeature_size(int xfeature_nr
)
394 u32 eax
, ebx
, ecx
, edx
;
396 CHECK_XFEATURE(xfeature_nr
);
397 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
402 * 'XSAVES' implies two different things:
403 * 1. saving of supervisor/system state
404 * 2. using the compacted format
406 * Use this function when dealing with the compacted format so
407 * that it is obvious which aspect of 'XSAVES' is being handled
408 * by the calling code.
410 static int using_compacted_format(void)
412 return cpu_has_xsaves
;
415 static void __xstate_dump_leaves(void)
418 u32 eax
, ebx
, ecx
, edx
;
419 static int should_dump
= 1;
425 * Dump out a few leaves past the ones that we support
426 * just in case there are some goodies up there
428 for (i
= 0; i
< XFEATURE_MAX
+ 10; i
++) {
429 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
430 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
431 XSTATE_CPUID
, i
, eax
, ebx
, ecx
, edx
);
435 #define XSTATE_WARN_ON(x) do { \
436 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
437 __xstate_dump_leaves(); \
441 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
442 if ((nr == nr_macro) && \
443 WARN_ONCE(sz != sizeof(__struct), \
444 "%s: struct is %zu bytes, cpu state %d bytes\n", \
445 __stringify(nr_macro), sizeof(__struct), sz)) { \
446 __xstate_dump_leaves(); \
451 * We have a C struct for each 'xstate'. We need to ensure
452 * that our software representation matches what the CPU
453 * tells us about the state's size.
455 static void check_xstate_against_struct(int nr
)
458 * Ask the CPU for the size of the state.
460 int sz
= xfeature_size(nr
);
462 * Match each CPU state with the corresponding software
465 XCHECK_SZ(sz
, nr
, XFEATURE_YMM
, struct ymmh_struct
);
466 XCHECK_SZ(sz
, nr
, XFEATURE_BNDREGS
, struct mpx_bndreg_state
);
467 XCHECK_SZ(sz
, nr
, XFEATURE_BNDCSR
, struct mpx_bndcsr_state
);
468 XCHECK_SZ(sz
, nr
, XFEATURE_OPMASK
, struct avx_512_opmask_state
);
469 XCHECK_SZ(sz
, nr
, XFEATURE_ZMM_Hi256
, struct avx_512_zmm_uppers_state
);
470 XCHECK_SZ(sz
, nr
, XFEATURE_Hi16_ZMM
, struct avx_512_hi16_state
);
473 * Make *SURE* to add any feature numbers in below if
474 * there are "holes" in the xsave state component
477 if ((nr
< XFEATURE_YMM
) ||
478 (nr
>= XFEATURE_MAX
) ||
479 (nr
== XFEATURE_PT_UNIMPLEMENTED_SO_FAR
)) {
480 WARN_ONCE(1, "no structure for xstate: %d\n", nr
);
486 * This essentially double-checks what the cpu told us about
487 * how large the XSAVE buffer needs to be. We are recalculating
490 static void do_extra_xstate_size_checks(void)
492 int paranoid_xstate_size
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
495 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
496 if (!xfeature_enabled(i
))
499 check_xstate_against_struct(i
);
501 * Supervisor state components can be managed only by
502 * XSAVES, which is compacted-format only.
504 if (!using_compacted_format())
505 XSTATE_WARN_ON(xfeature_is_supervisor(i
));
507 /* Align from the end of the previous feature */
508 if (xfeature_is_aligned(i
))
509 paranoid_xstate_size
= ALIGN(paranoid_xstate_size
, 64);
511 * The offset of a given state in the non-compacted
512 * format is given to us in a CPUID leaf. We check
513 * them for being ordered (increasing offsets) in
514 * setup_xstate_features().
516 if (!using_compacted_format())
517 paranoid_xstate_size
= xfeature_uncompacted_offset(i
);
519 * The compacted-format offset always depends on where
520 * the previous state ended.
522 paranoid_xstate_size
+= xfeature_size(i
);
524 XSTATE_WARN_ON(paranoid_xstate_size
!= xstate_size
);
528 * Calculate total size of enabled xstates in XCR0/xfeatures_mask.
530 * Note the SDM's wording here. "sub-function 0" only enumerates
531 * the size of the *user* states. If we use it to size a buffer
532 * that we use 'XSAVES' on, we could potentially overflow the
533 * buffer because 'XSAVES' saves system states too.
535 * Note that we do not currently set any bits on IA32_XSS so
536 * 'XCR0 | IA32_XSS == XCR0' for now.
538 static unsigned int __init
calculate_xstate_size(void)
540 unsigned int eax
, ebx
, ecx
, edx
;
541 unsigned int calculated_xstate_size
;
543 if (!cpu_has_xsaves
) {
545 * - CPUID function 0DH, sub-function 0:
546 * EBX enumerates the size (in bytes) required by
547 * the XSAVE instruction for an XSAVE area
548 * containing all the *user* state components
549 * corresponding to bits currently set in XCR0.
551 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
552 calculated_xstate_size
= ebx
;
555 * - CPUID function 0DH, sub-function 1:
556 * EBX enumerates the size (in bytes) required by
557 * the XSAVES instruction for an XSAVE area
558 * containing all the state components
559 * corresponding to bits currently set in
562 cpuid_count(XSTATE_CPUID
, 1, &eax
, &ebx
, &ecx
, &edx
);
563 calculated_xstate_size
= ebx
;
565 return calculated_xstate_size
;
569 * Will the runtime-enumerated 'xstate_size' fit in the init
570 * task's statically-allocated buffer?
572 static bool is_supported_xstate_size(unsigned int test_xstate_size
)
574 if (test_xstate_size
<= sizeof(union fpregs_state
))
577 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
578 sizeof(union fpregs_state
), test_xstate_size
);
582 static int init_xstate_size(void)
584 /* Recompute the context size for enabled features: */
585 unsigned int possible_xstate_size
= calculate_xstate_size();
587 /* Ensure we have the space to store all enabled: */
588 if (!is_supported_xstate_size(possible_xstate_size
))
592 * The size is OK, we are definitely going to use xsave,
593 * make it known to the world that we need more space.
595 xstate_size
= possible_xstate_size
;
596 do_extra_xstate_size_checks();
601 * We enabled the XSAVE hardware, but something went wrong and
602 * we can not use it. Disable it.
604 static void fpu__init_disable_system_xstate(void)
607 cr4_clear_bits(X86_CR4_OSXSAVE
);
608 fpu__xstate_clear_all_cpu_caps();
612 * Enable and initialize the xsave feature.
613 * Called once per system bootup.
615 void __init
fpu__init_system_xstate(void)
617 unsigned int eax
, ebx
, ecx
, edx
;
618 static int on_boot_cpu __initdata
= 1;
621 WARN_ON_FPU(!on_boot_cpu
);
624 if (!cpu_has_xsave
) {
625 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
629 if (boot_cpu_data
.cpuid_level
< XSTATE_CPUID
) {
634 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
635 xfeatures_mask
= eax
+ ((u64
)edx
<< 32);
637 if ((xfeatures_mask
& XFEATURE_MASK_FPSSE
) != XFEATURE_MASK_FPSSE
) {
638 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask
);
642 xfeatures_mask
&= fpu__get_supported_xfeatures_mask();
644 /* Enable xstate instructions to be able to continue with initialization: */
645 fpu__init_cpu_xstate();
646 err
= init_xstate_size();
648 /* something went wrong, boot without any XSAVE support */
649 fpu__init_disable_system_xstate();
653 update_regset_xstate_info(xstate_size
, xfeatures_mask
);
654 fpu__init_prepare_fx_sw_frame();
655 setup_init_fpu_buf();
658 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
661 cpu_has_xsaves
? "compacted" : "standard");
665 * Restore minimal FPU state after suspend:
667 void fpu__resume_cpu(void)
670 * Restore XCR0 on xsave capable CPUs:
673 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
677 * Given the xsave area and a state inside, this function returns the
678 * address of the state.
680 * This is the API that is called to get xstate address in either
681 * standard format or compacted format of xsave area.
683 * Note that if there is no data for the field in the xsave buffer
684 * this will return NULL.
687 * xstate: the thread's storage area for all FPU data
688 * xstate_feature: state which is defined in xsave.h (e.g.
689 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
691 * address of the state in the xsave area, or NULL if the
692 * field is not present in the xsave buffer.
694 void *get_xsave_addr(struct xregs_state
*xsave
, int xstate_feature
)
696 int feature_nr
= fls64(xstate_feature
) - 1;
698 * Do we even *have* xsave state?
700 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
704 * We should not ever be requesting features that we
705 * have not enabled. Remember that pcntxt_mask is
706 * what we write to the XCR0 register.
708 WARN_ONCE(!(xfeatures_mask
& xstate_feature
),
709 "get of unsupported state");
711 * This assumes the last 'xsave*' instruction to
712 * have requested that 'xstate_feature' be saved.
713 * If it did not, we might be seeing and old value
714 * of the field in the buffer.
716 * This can happen because the last 'xsave' did not
717 * request that this feature be saved (unlikely)
718 * or because the "init optimization" caused it
721 if (!(xsave
->header
.xfeatures
& xstate_feature
))
724 return (void *)xsave
+ xstate_comp_offsets
[feature_nr
];
726 EXPORT_SYMBOL_GPL(get_xsave_addr
);
729 * This wraps up the common operations that need to occur when retrieving
730 * data from xsave state. It first ensures that the current task was
731 * using the FPU and retrieves the data in to a buffer. It then calculates
732 * the offset of the requested field in the buffer.
734 * This function is safe to call whether the FPU is in use or not.
736 * Note that this only works on the current task.
739 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
740 * XFEATURE_MASK_SSE, etc...)
742 * address of the state in the xsave area or NULL if the state
743 * is not present or is in its 'init state'.
745 const void *get_xsave_field_ptr(int xsave_state
)
747 struct fpu
*fpu
= ¤t
->thread
.fpu
;
749 if (!fpu
->fpstate_active
)
752 * fpu__save() takes the CPU's xstate registers
753 * and saves them off to the 'fpu memory buffer.
757 return get_xsave_addr(&fpu
->state
.xsave
, xsave_state
);