2 * xsave/xrstor support.
4 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
6 #include <linux/compat.h>
8 #include <linux/pkeys.h>
10 #include <asm/fpu/api.h>
11 #include <asm/fpu/internal.h>
12 #include <asm/fpu/signal.h>
13 #include <asm/fpu/regset.h>
14 #include <asm/fpu/xstate.h>
16 #include <asm/tlbflush.h>
19 * Although we spell it out in here, the Processor Trace
20 * xfeature is completely unused. We use other mechanisms
21 * to save/restore PT state in Linux.
23 static const char *xfeature_names
[] =
25 "x87 floating point registers" ,
28 "MPX bounds registers" ,
33 "Processor Trace (unused)" ,
34 "Protection Keys User registers",
35 "unknown xstate feature" ,
39 * Mask of xstate features supported by the CPU and the kernel:
41 u64 xfeatures_mask __read_mostly
;
43 static unsigned int xstate_offsets
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
44 static unsigned int xstate_sizes
[XFEATURE_MAX
] = { [ 0 ... XFEATURE_MAX
- 1] = -1};
45 static unsigned int xstate_comp_offsets
[sizeof(xfeatures_mask
)*8];
48 * The XSAVE area of kernel can be in standard or compacted format;
49 * it is always in standard format for user mode. This is the user
50 * mode standard format size used for signal and ptrace frames.
52 unsigned int fpu_user_xstate_size
;
55 * Clear all of the X86_FEATURE_* bits that are unavailable
56 * when the CPU has no XSAVE support.
58 void fpu__xstate_clear_all_cpu_caps(void)
60 setup_clear_cpu_cap(X86_FEATURE_XSAVE
);
61 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT
);
62 setup_clear_cpu_cap(X86_FEATURE_XSAVEC
);
63 setup_clear_cpu_cap(X86_FEATURE_XSAVES
);
64 setup_clear_cpu_cap(X86_FEATURE_AVX
);
65 setup_clear_cpu_cap(X86_FEATURE_AVX2
);
66 setup_clear_cpu_cap(X86_FEATURE_AVX512F
);
67 setup_clear_cpu_cap(X86_FEATURE_AVX512PF
);
68 setup_clear_cpu_cap(X86_FEATURE_AVX512ER
);
69 setup_clear_cpu_cap(X86_FEATURE_AVX512CD
);
70 setup_clear_cpu_cap(X86_FEATURE_AVX512DQ
);
71 setup_clear_cpu_cap(X86_FEATURE_AVX512BW
);
72 setup_clear_cpu_cap(X86_FEATURE_AVX512VL
);
73 setup_clear_cpu_cap(X86_FEATURE_MPX
);
74 setup_clear_cpu_cap(X86_FEATURE_XGETBV1
);
75 setup_clear_cpu_cap(X86_FEATURE_PKU
);
79 * Return whether the system supports a given xfeature.
81 * Also return the name of the (most advanced) feature that the caller requested:
83 int cpu_has_xfeatures(u64 xfeatures_needed
, const char **feature_name
)
85 u64 xfeatures_missing
= xfeatures_needed
& ~xfeatures_mask
;
87 if (unlikely(feature_name
)) {
88 long xfeature_idx
, max_idx
;
91 * So we use FLS here to be able to print the most advanced
92 * feature that was requested but is missing. So if a driver
93 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
94 * missing AVX feature - this is the most informative message
97 if (xfeatures_missing
)
98 xfeatures_print
= xfeatures_missing
;
100 xfeatures_print
= xfeatures_needed
;
102 xfeature_idx
= fls64(xfeatures_print
)-1;
103 max_idx
= ARRAY_SIZE(xfeature_names
)-1;
104 xfeature_idx
= min(xfeature_idx
, max_idx
);
106 *feature_name
= xfeature_names
[xfeature_idx
];
109 if (xfeatures_missing
)
114 EXPORT_SYMBOL_GPL(cpu_has_xfeatures
);
116 static int xfeature_is_supervisor(int xfeature_nr
)
119 * We currently do not support supervisor states, but if
120 * we did, we could find out like this.
122 * SDM says: If state component 'i' is a user state component,
123 * ECX[0] return 0; if state component i is a supervisor
124 * state component, ECX[0] returns 1.
126 u32 eax
, ebx
, ecx
, edx
;
128 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
132 static int xfeature_is_user(int xfeature_nr
)
134 return !xfeature_is_supervisor(xfeature_nr
);
138 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
139 * a processor implementation detects that an FPU state component is still
140 * (or is again) in its initialized state, it may clear the corresponding
141 * bit in the header.xfeatures field, and can skip the writeout of registers
142 * to the corresponding memory layout.
144 * This means that when the bit is zero, the state component might still contain
145 * some previous - non-initialized register state.
147 * Before writing xstate information to user-space we sanitize those components,
148 * to always ensure that the memory layout of a feature will be in the init state
149 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
150 * see some stale state in the memory layout during signal handling, debugging etc.
152 void fpstate_sanitize_xstate(struct fpu
*fpu
)
154 struct fxregs_state
*fx
= &fpu
->state
.fxsave
;
161 xfeatures
= fpu
->state
.xsave
.header
.xfeatures
;
164 * None of the feature bits are in init state. So nothing else
165 * to do for us, as the memory layout is up to date.
167 if ((xfeatures
& xfeatures_mask
) == xfeatures_mask
)
171 * FP is in init state
173 if (!(xfeatures
& XFEATURE_MASK_FP
)) {
180 memset(&fx
->st_space
[0], 0, 128);
184 * SSE is in init state
186 if (!(xfeatures
& XFEATURE_MASK_SSE
))
187 memset(&fx
->xmm_space
[0], 0, 256);
190 * First two features are FPU and SSE, which above we handled
191 * in a special way already:
194 xfeatures
= (xfeatures_mask
& ~xfeatures
) >> 2;
197 * Update all the remaining memory layouts according to their
198 * standard xstate layout, if their header bit is in the init
202 if (xfeatures
& 0x1) {
203 int offset
= xstate_comp_offsets
[feature_bit
];
204 int size
= xstate_sizes
[feature_bit
];
206 memcpy((void *)fx
+ offset
,
207 (void *)&init_fpstate
.xsave
+ offset
,
217 * Enable the extended processor state save/restore feature.
218 * Called once per CPU onlining.
220 void fpu__init_cpu_xstate(void)
222 if (!boot_cpu_has(X86_FEATURE_XSAVE
) || !xfeatures_mask
)
225 * Make it clear that XSAVES supervisor states are not yet
226 * implemented should anyone expect it to work by changing
227 * bits in XFEATURE_MASK_* macros and XCR0.
229 WARN_ONCE((xfeatures_mask
& XFEATURE_MASK_SUPERVISOR
),
230 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
232 xfeatures_mask
&= ~XFEATURE_MASK_SUPERVISOR
;
234 cr4_set_bits(X86_CR4_OSXSAVE
);
235 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
239 * Note that in the future we will likely need a pair of
240 * functions here: one for user xstates and the other for
241 * system xstates. For now, they are the same.
243 static int xfeature_enabled(enum xfeature xfeature
)
245 return !!(xfeatures_mask
& (1UL << xfeature
));
249 * Record the offsets and sizes of various xstates contained
250 * in the XSAVE state memory layout.
252 static void __init
setup_xstate_features(void)
254 u32 eax
, ebx
, ecx
, edx
, i
;
255 /* start at the beginnning of the "extended state" */
256 unsigned int last_good_offset
= offsetof(struct xregs_state
,
257 extended_state_area
);
259 * The FP xstates and SSE xstates are legacy states. They are always
260 * in the fixed offsets in the xsave area in either compacted form
263 xstate_offsets
[0] = 0;
264 xstate_sizes
[0] = offsetof(struct fxregs_state
, xmm_space
);
265 xstate_offsets
[1] = xstate_sizes
[0];
266 xstate_sizes
[1] = FIELD_SIZEOF(struct fxregs_state
, xmm_space
);
268 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
269 if (!xfeature_enabled(i
))
272 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
275 * If an xfeature is supervisor state, the offset
276 * in EBX is invalid. We leave it to -1.
278 if (xfeature_is_user(i
))
279 xstate_offsets
[i
] = ebx
;
281 xstate_sizes
[i
] = eax
;
283 * In our xstate size checks, we assume that the
284 * highest-numbered xstate feature has the
285 * highest offset in the buffer. Ensure it does.
287 WARN_ONCE(last_good_offset
> xstate_offsets
[i
],
288 "x86/fpu: misordered xstate at %d\n", last_good_offset
);
289 last_good_offset
= xstate_offsets
[i
];
293 static void __init
print_xstate_feature(u64 xstate_mask
)
295 const char *feature_name
;
297 if (cpu_has_xfeatures(xstate_mask
, &feature_name
))
298 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask
, feature_name
);
302 * Print out all the supported xstate features:
304 static void __init
print_xstate_features(void)
306 print_xstate_feature(XFEATURE_MASK_FP
);
307 print_xstate_feature(XFEATURE_MASK_SSE
);
308 print_xstate_feature(XFEATURE_MASK_YMM
);
309 print_xstate_feature(XFEATURE_MASK_BNDREGS
);
310 print_xstate_feature(XFEATURE_MASK_BNDCSR
);
311 print_xstate_feature(XFEATURE_MASK_OPMASK
);
312 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256
);
313 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM
);
314 print_xstate_feature(XFEATURE_MASK_PKRU
);
318 * This check is important because it is easy to get XSTATE_*
319 * confused with XSTATE_BIT_*.
321 #define CHECK_XFEATURE(nr) do { \
322 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
323 WARN_ON(nr >= XFEATURE_MAX); \
327 * We could cache this like xstate_size[], but we only use
328 * it here, so it would be a waste of space.
330 static int xfeature_is_aligned(int xfeature_nr
)
332 u32 eax
, ebx
, ecx
, edx
;
334 CHECK_XFEATURE(xfeature_nr
);
335 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
337 * The value returned by ECX[1] indicates the alignment
338 * of state component 'i' when the compacted format
339 * of the extended region of an XSAVE area is used:
345 * This function sets up offsets and sizes of all extended states in
346 * xsave area. This supports both standard format and compacted format
347 * of the xsave aread.
349 static void __init
setup_xstate_comp(void)
351 unsigned int xstate_comp_sizes
[sizeof(xfeatures_mask
)*8];
355 * The FP xstates and SSE xstates are legacy states. They are always
356 * in the fixed offsets in the xsave area in either compacted form
359 xstate_comp_offsets
[0] = 0;
360 xstate_comp_offsets
[1] = offsetof(struct fxregs_state
, xmm_space
);
362 if (!boot_cpu_has(X86_FEATURE_XSAVES
)) {
363 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
364 if (xfeature_enabled(i
)) {
365 xstate_comp_offsets
[i
] = xstate_offsets
[i
];
366 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
372 xstate_comp_offsets
[FIRST_EXTENDED_XFEATURE
] =
373 FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
375 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
376 if (xfeature_enabled(i
))
377 xstate_comp_sizes
[i
] = xstate_sizes
[i
];
379 xstate_comp_sizes
[i
] = 0;
381 if (i
> FIRST_EXTENDED_XFEATURE
) {
382 xstate_comp_offsets
[i
] = xstate_comp_offsets
[i
-1]
383 + xstate_comp_sizes
[i
-1];
385 if (xfeature_is_aligned(i
))
386 xstate_comp_offsets
[i
] =
387 ALIGN(xstate_comp_offsets
[i
], 64);
393 * Print out xstate component offsets and sizes
395 static void __init
print_xstate_offset_size(void)
399 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
400 if (!xfeature_enabled(i
))
402 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
403 i
, xstate_comp_offsets
[i
], i
, xstate_sizes
[i
]);
408 * setup the xstate image representing the init state
410 static void __init
setup_init_fpu_buf(void)
412 static int on_boot_cpu __initdata
= 1;
414 WARN_ON_FPU(!on_boot_cpu
);
417 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
420 setup_xstate_features();
421 print_xstate_features();
423 if (boot_cpu_has(X86_FEATURE_XSAVES
))
424 init_fpstate
.xsave
.header
.xcomp_bv
= (u64
)1 << 63 | xfeatures_mask
;
427 * Init all the features state with header.xfeatures being 0x0
429 copy_kernel_to_xregs_booting(&init_fpstate
.xsave
);
432 * Dump the init state again. This is to identify the init state
433 * of any feature which is not represented by all zero's.
435 copy_xregs_to_kernel_booting(&init_fpstate
.xsave
);
438 static int xfeature_uncompacted_offset(int xfeature_nr
)
440 u32 eax
, ebx
, ecx
, edx
;
443 * Only XSAVES supports supervisor states and it uses compacted
444 * format. Checking a supervisor state's uncompacted offset is
447 if (XFEATURE_MASK_SUPERVISOR
& (1 << xfeature_nr
)) {
448 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr
);
452 CHECK_XFEATURE(xfeature_nr
);
453 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
457 static int xfeature_size(int xfeature_nr
)
459 u32 eax
, ebx
, ecx
, edx
;
461 CHECK_XFEATURE(xfeature_nr
);
462 cpuid_count(XSTATE_CPUID
, xfeature_nr
, &eax
, &ebx
, &ecx
, &edx
);
467 * 'XSAVES' implies two different things:
468 * 1. saving of supervisor/system state
469 * 2. using the compacted format
471 * Use this function when dealing with the compacted format so
472 * that it is obvious which aspect of 'XSAVES' is being handled
473 * by the calling code.
475 int using_compacted_format(void)
477 return boot_cpu_has(X86_FEATURE_XSAVES
);
480 static void __xstate_dump_leaves(void)
483 u32 eax
, ebx
, ecx
, edx
;
484 static int should_dump
= 1;
490 * Dump out a few leaves past the ones that we support
491 * just in case there are some goodies up there
493 for (i
= 0; i
< XFEATURE_MAX
+ 10; i
++) {
494 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
495 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
496 XSTATE_CPUID
, i
, eax
, ebx
, ecx
, edx
);
500 #define XSTATE_WARN_ON(x) do { \
501 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
502 __xstate_dump_leaves(); \
506 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
507 if ((nr == nr_macro) && \
508 WARN_ONCE(sz != sizeof(__struct), \
509 "%s: struct is %zu bytes, cpu state %d bytes\n", \
510 __stringify(nr_macro), sizeof(__struct), sz)) { \
511 __xstate_dump_leaves(); \
516 * We have a C struct for each 'xstate'. We need to ensure
517 * that our software representation matches what the CPU
518 * tells us about the state's size.
520 static void check_xstate_against_struct(int nr
)
523 * Ask the CPU for the size of the state.
525 int sz
= xfeature_size(nr
);
527 * Match each CPU state with the corresponding software
530 XCHECK_SZ(sz
, nr
, XFEATURE_YMM
, struct ymmh_struct
);
531 XCHECK_SZ(sz
, nr
, XFEATURE_BNDREGS
, struct mpx_bndreg_state
);
532 XCHECK_SZ(sz
, nr
, XFEATURE_BNDCSR
, struct mpx_bndcsr_state
);
533 XCHECK_SZ(sz
, nr
, XFEATURE_OPMASK
, struct avx_512_opmask_state
);
534 XCHECK_SZ(sz
, nr
, XFEATURE_ZMM_Hi256
, struct avx_512_zmm_uppers_state
);
535 XCHECK_SZ(sz
, nr
, XFEATURE_Hi16_ZMM
, struct avx_512_hi16_state
);
536 XCHECK_SZ(sz
, nr
, XFEATURE_PKRU
, struct pkru_state
);
539 * Make *SURE* to add any feature numbers in below if
540 * there are "holes" in the xsave state component
543 if ((nr
< XFEATURE_YMM
) ||
544 (nr
>= XFEATURE_MAX
) ||
545 (nr
== XFEATURE_PT_UNIMPLEMENTED_SO_FAR
)) {
546 WARN_ONCE(1, "no structure for xstate: %d\n", nr
);
552 * This essentially double-checks what the cpu told us about
553 * how large the XSAVE buffer needs to be. We are recalculating
556 static void do_extra_xstate_size_checks(void)
558 int paranoid_xstate_size
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
561 for (i
= FIRST_EXTENDED_XFEATURE
; i
< XFEATURE_MAX
; i
++) {
562 if (!xfeature_enabled(i
))
565 check_xstate_against_struct(i
);
567 * Supervisor state components can be managed only by
568 * XSAVES, which is compacted-format only.
570 if (!using_compacted_format())
571 XSTATE_WARN_ON(xfeature_is_supervisor(i
));
573 /* Align from the end of the previous feature */
574 if (xfeature_is_aligned(i
))
575 paranoid_xstate_size
= ALIGN(paranoid_xstate_size
, 64);
577 * The offset of a given state in the non-compacted
578 * format is given to us in a CPUID leaf. We check
579 * them for being ordered (increasing offsets) in
580 * setup_xstate_features().
582 if (!using_compacted_format())
583 paranoid_xstate_size
= xfeature_uncompacted_offset(i
);
585 * The compacted-format offset always depends on where
586 * the previous state ended.
588 paranoid_xstate_size
+= xfeature_size(i
);
590 XSTATE_WARN_ON(paranoid_xstate_size
!= fpu_kernel_xstate_size
);
595 * Get total size of enabled xstates in XCR0/xfeatures_mask.
597 * Note the SDM's wording here. "sub-function 0" only enumerates
598 * the size of the *user* states. If we use it to size a buffer
599 * that we use 'XSAVES' on, we could potentially overflow the
600 * buffer because 'XSAVES' saves system states too.
602 * Note that we do not currently set any bits on IA32_XSS so
603 * 'XCR0 | IA32_XSS == XCR0' for now.
605 static unsigned int __init
get_xsaves_size(void)
607 unsigned int eax
, ebx
, ecx
, edx
;
609 * - CPUID function 0DH, sub-function 1:
610 * EBX enumerates the size (in bytes) required by
611 * the XSAVES instruction for an XSAVE area
612 * containing all the state components
613 * corresponding to bits currently set in
616 cpuid_count(XSTATE_CPUID
, 1, &eax
, &ebx
, &ecx
, &edx
);
620 static unsigned int __init
get_xsave_size(void)
622 unsigned int eax
, ebx
, ecx
, edx
;
624 * - CPUID function 0DH, sub-function 0:
625 * EBX enumerates the size (in bytes) required by
626 * the XSAVE instruction for an XSAVE area
627 * containing all the *user* state components
628 * corresponding to bits currently set in XCR0.
630 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
635 * Will the runtime-enumerated 'xstate_size' fit in the init
636 * task's statically-allocated buffer?
638 static bool is_supported_xstate_size(unsigned int test_xstate_size
)
640 if (test_xstate_size
<= sizeof(union fpregs_state
))
643 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
644 sizeof(union fpregs_state
), test_xstate_size
);
648 static int init_xstate_size(void)
650 /* Recompute the context size for enabled features: */
651 unsigned int possible_xstate_size
;
652 unsigned int xsave_size
;
654 xsave_size
= get_xsave_size();
656 if (boot_cpu_has(X86_FEATURE_XSAVES
))
657 possible_xstate_size
= get_xsaves_size();
659 possible_xstate_size
= xsave_size
;
661 /* Ensure we have the space to store all enabled: */
662 if (!is_supported_xstate_size(possible_xstate_size
))
666 * The size is OK, we are definitely going to use xsave,
667 * make it known to the world that we need more space.
669 fpu_kernel_xstate_size
= possible_xstate_size
;
670 do_extra_xstate_size_checks();
673 * User space is always in standard format.
675 fpu_user_xstate_size
= xsave_size
;
680 * We enabled the XSAVE hardware, but something went wrong and
681 * we can not use it. Disable it.
683 static void fpu__init_disable_system_xstate(void)
686 cr4_clear_bits(X86_CR4_OSXSAVE
);
687 fpu__xstate_clear_all_cpu_caps();
691 * Enable and initialize the xsave feature.
692 * Called once per system bootup.
694 void __init
fpu__init_system_xstate(void)
696 unsigned int eax
, ebx
, ecx
, edx
;
697 static int on_boot_cpu __initdata
= 1;
700 WARN_ON_FPU(!on_boot_cpu
);
703 if (!boot_cpu_has(X86_FEATURE_XSAVE
)) {
704 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
708 if (boot_cpu_data
.cpuid_level
< XSTATE_CPUID
) {
713 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
714 xfeatures_mask
= eax
+ ((u64
)edx
<< 32);
716 if ((xfeatures_mask
& XFEATURE_MASK_FPSSE
) != XFEATURE_MASK_FPSSE
) {
718 * This indicates that something really unexpected happened
719 * with the enumeration. Disable XSAVE and try to continue
720 * booting without it. This is too early to BUG().
722 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask
);
726 xfeatures_mask
&= fpu__get_supported_xfeatures_mask();
728 /* Enable xstate instructions to be able to continue with initialization: */
729 fpu__init_cpu_xstate();
730 err
= init_xstate_size();
735 * Update info used for ptrace frames; use standard-format size and no
736 * supervisor xstates:
738 update_regset_xstate_info(fpu_user_xstate_size
, xfeatures_mask
& ~XFEATURE_MASK_SUPERVISOR
);
740 fpu__init_prepare_fx_sw_frame();
741 setup_init_fpu_buf();
743 print_xstate_offset_size();
745 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
747 fpu_kernel_xstate_size
,
748 boot_cpu_has(X86_FEATURE_XSAVES
) ? "compacted" : "standard");
752 /* something went wrong, try to boot without any XSAVE support */
753 fpu__init_disable_system_xstate();
757 * Restore minimal FPU state after suspend:
759 void fpu__resume_cpu(void)
762 * Restore XCR0 on xsave capable CPUs:
764 if (boot_cpu_has(X86_FEATURE_XSAVE
))
765 xsetbv(XCR_XFEATURE_ENABLED_MASK
, xfeatures_mask
);
769 * Given an xstate feature mask, calculate where in the xsave
770 * buffer the state is. Callers should ensure that the buffer
773 * Note: does not work for compacted buffers.
775 void *__raw_xsave_addr(struct xregs_state
*xsave
, int xstate_feature_mask
)
777 int feature_nr
= fls64(xstate_feature_mask
) - 1;
779 if (!xfeature_enabled(feature_nr
)) {
784 return (void *)xsave
+ xstate_comp_offsets
[feature_nr
];
787 * Given the xsave area and a state inside, this function returns the
788 * address of the state.
790 * This is the API that is called to get xstate address in either
791 * standard format or compacted format of xsave area.
793 * Note that if there is no data for the field in the xsave buffer
794 * this will return NULL.
797 * xstate: the thread's storage area for all FPU data
798 * xstate_feature: state which is defined in xsave.h (e.g.
799 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
801 * address of the state in the xsave area, or NULL if the
802 * field is not present in the xsave buffer.
804 void *get_xsave_addr(struct xregs_state
*xsave
, int xstate_feature
)
807 * Do we even *have* xsave state?
809 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
813 * We should not ever be requesting features that we
814 * have not enabled. Remember that pcntxt_mask is
815 * what we write to the XCR0 register.
817 WARN_ONCE(!(xfeatures_mask
& xstate_feature
),
818 "get of unsupported state");
820 * This assumes the last 'xsave*' instruction to
821 * have requested that 'xstate_feature' be saved.
822 * If it did not, we might be seeing and old value
823 * of the field in the buffer.
825 * This can happen because the last 'xsave' did not
826 * request that this feature be saved (unlikely)
827 * or because the "init optimization" caused it
830 if (!(xsave
->header
.xfeatures
& xstate_feature
))
833 return __raw_xsave_addr(xsave
, xstate_feature
);
835 EXPORT_SYMBOL_GPL(get_xsave_addr
);
838 * This wraps up the common operations that need to occur when retrieving
839 * data from xsave state. It first ensures that the current task was
840 * using the FPU and retrieves the data in to a buffer. It then calculates
841 * the offset of the requested field in the buffer.
843 * This function is safe to call whether the FPU is in use or not.
845 * Note that this only works on the current task.
848 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
849 * XFEATURE_MASK_SSE, etc...)
851 * address of the state in the xsave area or NULL if the state
852 * is not present or is in its 'init state'.
854 const void *get_xsave_field_ptr(int xsave_state
)
856 struct fpu
*fpu
= ¤t
->thread
.fpu
;
858 if (!fpu
->fpstate_active
)
861 * fpu__save() takes the CPU's xstate registers
862 * and saves them off to the 'fpu memory buffer.
866 return get_xsave_addr(&fpu
->state
.xsave
, xsave_state
);
871 * Set xfeatures (aka XSTATE_BV) bit for a feature that we want
872 * to take out of its "init state". This will ensure that an
873 * XRSTOR actually restores the state.
875 static void fpu__xfeature_set_non_init(struct xregs_state
*xsave
,
876 int xstate_feature_mask
)
878 xsave
->header
.xfeatures
|= xstate_feature_mask
;
882 * This function is safe to call whether the FPU is in use or not.
884 * Note that this only works on the current task.
887 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
888 * XFEATURE_MASK_SSE, etc...)
889 * @xsave_state_ptr: a pointer to a copy of the state that you would
890 * like written in to the current task's FPU xsave state. This pointer
891 * must not be located in the current tasks's xsave area.
893 * address of the state in the xsave area or NULL if the state
894 * is not present or is in its 'init state'.
896 static void fpu__xfeature_set_state(int xstate_feature_mask
,
897 void *xstate_feature_src
, size_t len
)
899 struct xregs_state
*xsave
= ¤t
->thread
.fpu
.state
.xsave
;
900 struct fpu
*fpu
= ¤t
->thread
.fpu
;
903 if (!boot_cpu_has(X86_FEATURE_XSAVE
)) {
904 WARN_ONCE(1, "%s() attempted with no xsave support", __func__
);
909 * Tell the FPU code that we need the FPU state to be in
910 * 'fpu' (not in the registers), and that we need it to
911 * be stable while we write to it.
913 fpu__current_fpstate_write_begin();
916 * This method *WILL* *NOT* work for compact-format
917 * buffers. If the 'xstate_feature_mask' is unset in
918 * xcomp_bv then we may need to move other feature state
919 * "up" in the buffer.
921 if (xsave
->header
.xcomp_bv
& xstate_feature_mask
) {
926 /* find the location in the xsave buffer of the desired state */
927 dst
= __raw_xsave_addr(&fpu
->state
.xsave
, xstate_feature_mask
);
930 * Make sure that the pointer being passed in did not
931 * come from the xsave buffer itself.
933 WARN_ONCE(xstate_feature_src
== dst
, "set from xsave buffer itself");
935 /* put the caller-provided data in the location */
936 memcpy(dst
, xstate_feature_src
, len
);
939 * Mark the xfeature so that the CPU knows there is state
942 fpu__xfeature_set_non_init(xsave
, xstate_feature_mask
);
945 * We are done writing to the 'fpu'. Reenable preeption
946 * and (possibly) move the fpstate back in to the fpregs.
948 fpu__current_fpstate_write_end();
951 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
952 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
955 * This will go out and modify the XSAVE buffer so that PKRU is
956 * set to a particular state for access to 'pkey'.
958 * PKRU state does affect kernel access to user memory. We do
959 * not modfiy PKRU *itself* here, only the XSAVE state that will
960 * be restored in to PKRU when we return back to userspace.
962 int arch_set_user_pkey_access(struct task_struct
*tsk
, int pkey
,
963 unsigned long init_val
)
965 struct xregs_state
*xsave
= &tsk
->thread
.fpu
.state
.xsave
;
966 struct pkru_state
*old_pkru_state
;
967 struct pkru_state new_pkru_state
;
968 int pkey_shift
= (pkey
* PKRU_BITS_PER_PKEY
);
969 u32 new_pkru_bits
= 0;
972 * This check implies XSAVE support. OSPKE only gets
973 * set if we enable XSAVE and we enable PKU in XCR0.
975 if (!boot_cpu_has(X86_FEATURE_OSPKE
))
978 /* Set the bits we need in PKRU: */
979 if (init_val
& PKEY_DISABLE_ACCESS
)
980 new_pkru_bits
|= PKRU_AD_BIT
;
981 if (init_val
& PKEY_DISABLE_WRITE
)
982 new_pkru_bits
|= PKRU_WD_BIT
;
984 /* Shift the bits in to the correct place in PKRU for pkey: */
985 new_pkru_bits
<<= pkey_shift
;
987 /* Locate old copy of the state in the xsave buffer: */
988 old_pkru_state
= get_xsave_addr(xsave
, XFEATURE_MASK_PKRU
);
991 * When state is not in the buffer, it is in the init
992 * state, set it manually. Otherwise, copy out the old
996 new_pkru_state
.pkru
= 0;
998 new_pkru_state
.pkru
= old_pkru_state
->pkru
;
1000 /* Mask off any old bits in place: */
1001 new_pkru_state
.pkru
&= ~((PKRU_AD_BIT
|PKRU_WD_BIT
) << pkey_shift
);
1003 /* Set the newly-requested bits: */
1004 new_pkru_state
.pkru
|= new_pkru_bits
;
1007 * We could theoretically live without zeroing pkru.pad.
1008 * The current XSAVE feature state definition says that
1009 * only bytes 0->3 are used. But we do not want to
1010 * chance leaking kernel stack out to userspace in case a
1011 * memcpy() of the whole xsave buffer was done.
1013 * They're in the same cacheline anyway.
1015 new_pkru_state
.pad
= 0;
1017 fpu__xfeature_set_state(XFEATURE_MASK_PKRU
, &new_pkru_state
, sizeof(new_pkru_state
));
1023 * This is similar to user_regset_copyout(), but will not add offset to
1024 * the source data pointer or increment pos, count, kbuf, and ubuf.
1026 static inline int xstate_copyout(unsigned int pos
, unsigned int count
,
1027 void *kbuf
, void __user
*ubuf
,
1028 const void *data
, const int start_pos
,
1031 if ((count
== 0) || (pos
< start_pos
))
1034 if (end_pos
< 0 || pos
< end_pos
) {
1035 unsigned int copy
= (end_pos
< 0 ? count
: min(count
, end_pos
- pos
));
1038 memcpy(kbuf
+ pos
, data
, copy
);
1040 if (__copy_to_user(ubuf
+ pos
, data
, copy
))
1048 * Convert from kernel XSAVES compacted format to standard format and copy
1049 * to a ptrace buffer. It supports partial copy but pos always starts from
1050 * zero. This is called from xstateregs_get() and there we check the CPU
1053 int copyout_from_xsaves(unsigned int pos
, unsigned int count
, void *kbuf
,
1054 void __user
*ubuf
, struct xregs_state
*xsave
)
1056 unsigned int offset
, size
;
1058 struct xstate_header header
;
1061 * Currently copy_regset_to_user() starts from pos 0:
1063 if (unlikely(pos
!= 0))
1067 * The destination is a ptrace buffer; we put in only user xstates:
1069 memset(&header
, 0, sizeof(header
));
1070 header
.xfeatures
= xsave
->header
.xfeatures
;
1071 header
.xfeatures
&= ~XFEATURE_MASK_SUPERVISOR
;
1074 * Copy xregs_state->header:
1076 offset
= offsetof(struct xregs_state
, header
);
1077 size
= sizeof(header
);
1079 ret
= xstate_copyout(offset
, size
, kbuf
, ubuf
, &header
, 0, count
);
1084 for (i
= 0; i
< XFEATURE_MAX
; i
++) {
1086 * Copy only in-use xstates:
1088 if ((header
.xfeatures
>> i
) & 1) {
1089 void *src
= __raw_xsave_addr(xsave
, 1 << i
);
1091 offset
= xstate_offsets
[i
];
1092 size
= xstate_sizes
[i
];
1094 ret
= xstate_copyout(offset
, size
, kbuf
, ubuf
, src
, 0, count
);
1099 if (offset
+ size
>= count
)
1106 * Fill xsave->i387.sw_reserved value for ptrace frame:
1108 offset
= offsetof(struct fxregs_state
, sw_reserved
);
1109 size
= sizeof(xstate_fx_sw_bytes
);
1111 ret
= xstate_copyout(offset
, size
, kbuf
, ubuf
, xstate_fx_sw_bytes
, 0, count
);
1120 * Convert from a ptrace standard-format buffer to kernel XSAVES format
1121 * and copy to the target thread. This is called from xstateregs_set() and
1122 * there we check the CPU has XSAVES and a whole standard-sized buffer
1125 int copyin_to_xsaves(const void *kbuf
, const void __user
*ubuf
,
1126 struct xregs_state
*xsave
)
1128 unsigned int offset
, size
;
1131 u64 allowed_features
;
1133 offset
= offsetof(struct xregs_state
, header
);
1134 size
= sizeof(xfeatures
);
1137 memcpy(&xfeatures
, kbuf
+ offset
, size
);
1139 if (__copy_from_user(&xfeatures
, ubuf
+ offset
, size
))
1144 * Reject if the user sets any disabled or supervisor features:
1146 allowed_features
= xfeatures_mask
& ~XFEATURE_MASK_SUPERVISOR
;
1148 if (xfeatures
& ~allowed_features
)
1151 for (i
= 0; i
< XFEATURE_MAX
; i
++) {
1152 u64 mask
= ((u64
)1 << i
);
1154 if (xfeatures
& mask
) {
1155 void *dst
= __raw_xsave_addr(xsave
, 1 << i
);
1157 offset
= xstate_offsets
[i
];
1158 size
= xstate_sizes
[i
];
1161 memcpy(dst
, kbuf
+ offset
, size
);
1163 if (__copy_from_user(dst
, ubuf
+ offset
, size
))
1170 * The state that came in from userspace was user-state only.
1171 * Mask all the user states out of 'xfeatures':
1173 xsave
->header
.xfeatures
&= XFEATURE_MASK_SUPERVISOR
;
1176 * Add back in the features that came in from userspace:
1178 xsave
->header
.xfeatures
|= xfeatures
;