1 // SPDX-License-Identifier: GPL-2.0-only
3 * xsave/xrstor support.
5 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
7 #include <linux/bitops.h>
8 #include <linux/compat.h>
10 #include <linux/mman.h>
11 #include <linux/nospec.h>
12 #include <linux/pkeys.h>
13 #include <linux/seq_file.h>
14 #include <linux/proc_fs.h>
15 #include <linux/vmalloc.h>
17 #include <asm/fpu/api.h>
18 #include <asm/fpu/regset.h>
19 #include <asm/fpu/signal.h>
20 #include <asm/fpu/xcr.h>
22 #include <asm/tlbflush.h>
23 #include <asm/prctl.h>
31 #define for_each_extended_xfeature(bit, mask) \
32 (bit) = FIRST_EXTENDED_XFEATURE; \
33 for_each_set_bit_from(bit, (unsigned long *)&(mask), 8 * sizeof(mask))
36 * Although we spell it out in here, the Processor Trace
37 * xfeature is completely unused. We use other mechanisms
38 * to save/restore PT state in Linux.
40 static const char *xfeature_names
[] =
42 "x87 floating point registers",
45 "MPX bounds registers",
50 "Processor Trace (unused)",
51 "Protection Keys User registers",
53 "Control-flow User registers",
54 "Control-flow Kernel registers (unused)",
55 "unknown xstate feature",
56 "unknown xstate feature",
57 "unknown xstate feature",
58 "unknown xstate feature",
61 "unknown xstate feature",
64 static unsigned short xsave_cpuid_features
[] __initdata
= {
65 [XFEATURE_FP
] = X86_FEATURE_FPU
,
66 [XFEATURE_SSE
] = X86_FEATURE_XMM
,
67 [XFEATURE_YMM
] = X86_FEATURE_AVX
,
68 [XFEATURE_BNDREGS
] = X86_FEATURE_MPX
,
69 [XFEATURE_BNDCSR
] = X86_FEATURE_MPX
,
70 [XFEATURE_OPMASK
] = X86_FEATURE_AVX512F
,
71 [XFEATURE_ZMM_Hi256
] = X86_FEATURE_AVX512F
,
72 [XFEATURE_Hi16_ZMM
] = X86_FEATURE_AVX512F
,
73 [XFEATURE_PT_UNIMPLEMENTED_SO_FAR
] = X86_FEATURE_INTEL_PT
,
74 [XFEATURE_PKRU
] = X86_FEATURE_OSPKE
,
75 [XFEATURE_PASID
] = X86_FEATURE_ENQCMD
,
76 [XFEATURE_CET_USER
] = X86_FEATURE_SHSTK
,
77 [XFEATURE_XTILE_CFG
] = X86_FEATURE_AMX_TILE
,
78 [XFEATURE_XTILE_DATA
] = X86_FEATURE_AMX_TILE
,
81 static unsigned int xstate_offsets
[XFEATURE_MAX
] __ro_after_init
=
82 { [ 0 ... XFEATURE_MAX
- 1] = -1};
83 static unsigned int xstate_sizes
[XFEATURE_MAX
] __ro_after_init
=
84 { [ 0 ... XFEATURE_MAX
- 1] = -1};
85 static unsigned int xstate_flags
[XFEATURE_MAX
] __ro_after_init
;
87 #define XSTATE_FLAG_SUPERVISOR BIT(0)
88 #define XSTATE_FLAG_ALIGNED64 BIT(1)
91 * Return whether the system supports a given xfeature.
93 * Also return the name of the (most advanced) feature that the caller requested:
95 int cpu_has_xfeatures(u64 xfeatures_needed
, const char **feature_name
)
97 u64 xfeatures_missing
= xfeatures_needed
& ~fpu_kernel_cfg
.max_features
;
99 if (unlikely(feature_name
)) {
100 long xfeature_idx
, max_idx
;
103 * So we use FLS here to be able to print the most advanced
104 * feature that was requested but is missing. So if a driver
105 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
106 * missing AVX feature - this is the most informative message
109 if (xfeatures_missing
)
110 xfeatures_print
= xfeatures_missing
;
112 xfeatures_print
= xfeatures_needed
;
114 xfeature_idx
= fls64(xfeatures_print
)-1;
115 max_idx
= ARRAY_SIZE(xfeature_names
)-1;
116 xfeature_idx
= min(xfeature_idx
, max_idx
);
118 *feature_name
= xfeature_names
[xfeature_idx
];
121 if (xfeatures_missing
)
126 EXPORT_SYMBOL_GPL(cpu_has_xfeatures
);
128 static bool xfeature_is_aligned64(int xfeature_nr
)
130 return xstate_flags
[xfeature_nr
] & XSTATE_FLAG_ALIGNED64
;
133 static bool xfeature_is_supervisor(int xfeature_nr
)
135 return xstate_flags
[xfeature_nr
] & XSTATE_FLAG_SUPERVISOR
;
138 static unsigned int xfeature_get_offset(u64 xcomp_bv
, int xfeature
)
140 unsigned int offs
, i
;
143 * Non-compacted format and legacy features use the cached fixed
146 if (!cpu_feature_enabled(X86_FEATURE_XCOMPACTED
) ||
147 xfeature
<= XFEATURE_SSE
)
148 return xstate_offsets
[xfeature
];
151 * Compacted format offsets depend on the actual content of the
152 * compacted xsave area which is determined by the xcomp_bv header
155 offs
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
156 for_each_extended_xfeature(i
, xcomp_bv
) {
157 if (xfeature_is_aligned64(i
))
158 offs
= ALIGN(offs
, 64);
161 offs
+= xstate_sizes
[i
];
167 * Enable the extended processor state save/restore feature.
168 * Called once per CPU onlining.
170 void fpu__init_cpu_xstate(void)
172 if (!boot_cpu_has(X86_FEATURE_XSAVE
) || !fpu_kernel_cfg
.max_features
)
175 cr4_set_bits(X86_CR4_OSXSAVE
);
178 * Must happen after CR4 setup and before xsetbv() to allow KVM
179 * lazy passthrough. Write independent of the dynamic state static
180 * key as that does not work on the boot CPU. This also ensures
181 * that any stale state is wiped out from XFD.
183 if (cpu_feature_enabled(X86_FEATURE_XFD
))
184 wrmsrl(MSR_IA32_XFD
, init_fpstate
.xfd
);
187 * XCR_XFEATURE_ENABLED_MASK (aka. XCR0) sets user features
188 * managed by XSAVE{C, OPT, S} and XRSTOR{S}. Only XSAVE user
189 * states can be set here.
191 xsetbv(XCR_XFEATURE_ENABLED_MASK
, fpu_user_cfg
.max_features
);
194 * MSR_IA32_XSS sets supervisor states managed by XSAVES.
196 if (boot_cpu_has(X86_FEATURE_XSAVES
)) {
197 wrmsrl(MSR_IA32_XSS
, xfeatures_mask_supervisor() |
198 xfeatures_mask_independent());
202 static bool xfeature_enabled(enum xfeature xfeature
)
204 return fpu_kernel_cfg
.max_features
& BIT_ULL(xfeature
);
208 * Record the offsets and sizes of various xstates contained
209 * in the XSAVE state memory layout.
211 static void __init
setup_xstate_cache(void)
213 u32 eax
, ebx
, ecx
, edx
, i
;
214 /* start at the beginning of the "extended state" */
215 unsigned int last_good_offset
= offsetof(struct xregs_state
,
216 extended_state_area
);
218 * The FP xstates and SSE xstates are legacy states. They are always
219 * in the fixed offsets in the xsave area in either compacted form
222 xstate_offsets
[XFEATURE_FP
] = 0;
223 xstate_sizes
[XFEATURE_FP
] = offsetof(struct fxregs_state
,
226 xstate_offsets
[XFEATURE_SSE
] = xstate_sizes
[XFEATURE_FP
];
227 xstate_sizes
[XFEATURE_SSE
] = sizeof_field(struct fxregs_state
,
230 for_each_extended_xfeature(i
, fpu_kernel_cfg
.max_features
) {
231 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
233 xstate_sizes
[i
] = eax
;
234 xstate_flags
[i
] = ecx
;
237 * If an xfeature is supervisor state, the offset in EBX is
238 * invalid, leave it to -1.
240 if (xfeature_is_supervisor(i
))
243 xstate_offsets
[i
] = ebx
;
246 * In our xstate size checks, we assume that the highest-numbered
247 * xstate feature has the highest offset in the buffer. Ensure
250 WARN_ONCE(last_good_offset
> xstate_offsets
[i
],
251 "x86/fpu: misordered xstate at %d\n", last_good_offset
);
253 last_good_offset
= xstate_offsets
[i
];
257 static void __init
print_xstate_feature(u64 xstate_mask
)
259 const char *feature_name
;
261 if (cpu_has_xfeatures(xstate_mask
, &feature_name
))
262 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask
, feature_name
);
266 * Print out all the supported xstate features:
268 static void __init
print_xstate_features(void)
270 print_xstate_feature(XFEATURE_MASK_FP
);
271 print_xstate_feature(XFEATURE_MASK_SSE
);
272 print_xstate_feature(XFEATURE_MASK_YMM
);
273 print_xstate_feature(XFEATURE_MASK_BNDREGS
);
274 print_xstate_feature(XFEATURE_MASK_BNDCSR
);
275 print_xstate_feature(XFEATURE_MASK_OPMASK
);
276 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256
);
277 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM
);
278 print_xstate_feature(XFEATURE_MASK_PKRU
);
279 print_xstate_feature(XFEATURE_MASK_PASID
);
280 print_xstate_feature(XFEATURE_MASK_CET_USER
);
281 print_xstate_feature(XFEATURE_MASK_XTILE_CFG
);
282 print_xstate_feature(XFEATURE_MASK_XTILE_DATA
);
286 * This check is important because it is easy to get XSTATE_*
287 * confused with XSTATE_BIT_*.
289 #define CHECK_XFEATURE(nr) do { \
290 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
291 WARN_ON(nr >= XFEATURE_MAX); \
295 * Print out xstate component offsets and sizes
297 static void __init
print_xstate_offset_size(void)
301 for_each_extended_xfeature(i
, fpu_kernel_cfg
.max_features
) {
302 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
303 i
, xfeature_get_offset(fpu_kernel_cfg
.max_features
, i
),
309 * This function is called only during boot time when x86 caps are not set
310 * up and alternative can not be used yet.
312 static __init
void os_xrstor_booting(struct xregs_state
*xstate
)
314 u64 mask
= fpu_kernel_cfg
.max_features
& XFEATURE_MASK_FPSTATE
;
316 u32 hmask
= mask
>> 32;
319 if (cpu_feature_enabled(X86_FEATURE_XSAVES
))
320 XSTATE_OP(XRSTORS
, xstate
, lmask
, hmask
, err
);
322 XSTATE_OP(XRSTOR
, xstate
, lmask
, hmask
, err
);
325 * We should never fault when copying from a kernel buffer, and the FPU
326 * state we set at boot time should be valid.
332 * All supported features have either init state all zeros or are
333 * handled in setup_init_fpu() individually. This is an explicit
334 * feature list and does not use XFEATURE_MASK*SUPPORTED to catch
335 * newly added supported features at build time and make people
336 * actually look at the init state for the new feature.
338 #define XFEATURES_INIT_FPSTATE_HANDLED \
339 (XFEATURE_MASK_FP | \
340 XFEATURE_MASK_SSE | \
341 XFEATURE_MASK_YMM | \
342 XFEATURE_MASK_OPMASK | \
343 XFEATURE_MASK_ZMM_Hi256 | \
344 XFEATURE_MASK_Hi16_ZMM | \
345 XFEATURE_MASK_PKRU | \
346 XFEATURE_MASK_BNDREGS | \
347 XFEATURE_MASK_BNDCSR | \
348 XFEATURE_MASK_PASID | \
349 XFEATURE_MASK_CET_USER | \
353 * setup the xstate image representing the init state
355 static void __init
setup_init_fpu_buf(void)
357 BUILD_BUG_ON((XFEATURE_MASK_USER_SUPPORTED
|
358 XFEATURE_MASK_SUPERVISOR_SUPPORTED
) !=
359 XFEATURES_INIT_FPSTATE_HANDLED
);
361 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
364 print_xstate_features();
366 xstate_init_xcomp_bv(&init_fpstate
.regs
.xsave
, init_fpstate
.xfeatures
);
369 * Init all the features state with header.xfeatures being 0x0
371 os_xrstor_booting(&init_fpstate
.regs
.xsave
);
374 * All components are now in init state. Read the state back so
375 * that init_fpstate contains all non-zero init state. This only
376 * works with XSAVE, but not with XSAVEOPT and XSAVEC/S because
377 * those use the init optimization which skips writing data for
378 * components in init state.
380 * XSAVE could be used, but that would require to reshuffle the
381 * data when XSAVEC/S is available because XSAVEC/S uses xstate
382 * compaction. But doing so is a pointless exercise because most
383 * components have an all zeros init state except for the legacy
384 * ones (FP and SSE). Those can be saved with FXSAVE into the
385 * legacy area. Adding new features requires to ensure that init
386 * state is all zeroes or if not to add the necessary handling
389 fxsave(&init_fpstate
.regs
.fxsave
);
392 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
);
401 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
402 static int validate_user_xstate_header(const struct xstate_header
*hdr
,
403 struct fpstate
*fpstate
)
405 /* No unknown or supervisor features may be set */
406 if (hdr
->xfeatures
& ~fpstate
->user_xfeatures
)
409 /* Userspace must use the uncompacted format */
414 * If 'reserved' is shrunken to add a new field, make sure to validate
415 * that new field here!
417 BUILD_BUG_ON(sizeof(hdr
->reserved
) != 48);
419 /* No reserved bits may be set */
420 if (memchr_inv(hdr
->reserved
, 0, sizeof(hdr
->reserved
)))
426 static void __init
__xstate_dump_leaves(void)
429 u32 eax
, ebx
, ecx
, edx
;
430 static int should_dump
= 1;
436 * Dump out a few leaves past the ones that we support
437 * just in case there are some goodies up there
439 for (i
= 0; i
< XFEATURE_MAX
+ 10; i
++) {
440 cpuid_count(XSTATE_CPUID
, i
, &eax
, &ebx
, &ecx
, &edx
);
441 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
442 XSTATE_CPUID
, i
, eax
, ebx
, ecx
, edx
);
446 #define XSTATE_WARN_ON(x, fmt, ...) do { \
447 if (WARN_ONCE(x, "XSAVE consistency problem: " fmt, ##__VA_ARGS__)) { \
448 __xstate_dump_leaves(); \
452 #define XCHECK_SZ(sz, nr, __struct) ({ \
453 if (WARN_ONCE(sz != sizeof(__struct), \
454 "[%s]: struct is %zu bytes, cpu state %d bytes\n", \
455 xfeature_names[nr], sizeof(__struct), sz)) { \
456 __xstate_dump_leaves(); \
463 * check_xtile_data_against_struct - Check tile data state size.
465 * Calculate the state size by multiplying the single tile size which is
466 * recorded in a C struct, and the number of tiles that the CPU informs.
467 * Compare the provided size with the calculation.
469 * @size: The tile data state size
471 * Returns: 0 on success, -EINVAL on mismatch.
473 static int __init
check_xtile_data_against_struct(int size
)
475 u32 max_palid
, palid
, state_size
;
476 u32 eax
, ebx
, ecx
, edx
;
480 * Check the maximum palette id:
481 * eax: the highest numbered palette subleaf.
483 cpuid_count(TILE_CPUID
, 0, &max_palid
, &ebx
, &ecx
, &edx
);
486 * Cross-check each tile size and find the maximum number of
489 for (palid
= 1, max_tile
= 0; palid
<= max_palid
; palid
++) {
493 * Check the tile size info:
494 * eax[31:16]: bytes per title
495 * ebx[31:16]: the max names (or max number of tiles)
497 cpuid_count(TILE_CPUID
, palid
, &eax
, &ebx
, &edx
, &edx
);
498 tile_size
= eax
>> 16;
501 if (tile_size
!= sizeof(struct xtile_data
)) {
502 pr_err("%s: struct is %zu bytes, cpu xtile %d bytes\n",
503 __stringify(XFEATURE_XTILE_DATA
),
504 sizeof(struct xtile_data
), tile_size
);
505 __xstate_dump_leaves();
513 state_size
= sizeof(struct xtile_data
) * max_tile
;
514 if (size
!= state_size
) {
515 pr_err("%s: calculated size is %u bytes, cpu state %d bytes\n",
516 __stringify(XFEATURE_XTILE_DATA
), state_size
, size
);
517 __xstate_dump_leaves();
524 * We have a C struct for each 'xstate'. We need to ensure
525 * that our software representation matches what the CPU
526 * tells us about the state's size.
528 static bool __init
check_xstate_against_struct(int nr
)
531 * Ask the CPU for the size of the state.
533 int sz
= xfeature_size(nr
);
536 * Match each CPU state with the corresponding software
540 case XFEATURE_YMM
: return XCHECK_SZ(sz
, nr
, struct ymmh_struct
);
541 case XFEATURE_BNDREGS
: return XCHECK_SZ(sz
, nr
, struct mpx_bndreg_state
);
542 case XFEATURE_BNDCSR
: return XCHECK_SZ(sz
, nr
, struct mpx_bndcsr_state
);
543 case XFEATURE_OPMASK
: return XCHECK_SZ(sz
, nr
, struct avx_512_opmask_state
);
544 case XFEATURE_ZMM_Hi256
: return XCHECK_SZ(sz
, nr
, struct avx_512_zmm_uppers_state
);
545 case XFEATURE_Hi16_ZMM
: return XCHECK_SZ(sz
, nr
, struct avx_512_hi16_state
);
546 case XFEATURE_PKRU
: return XCHECK_SZ(sz
, nr
, struct pkru_state
);
547 case XFEATURE_PASID
: return XCHECK_SZ(sz
, nr
, struct ia32_pasid_state
);
548 case XFEATURE_XTILE_CFG
: return XCHECK_SZ(sz
, nr
, struct xtile_cfg
);
549 case XFEATURE_CET_USER
: return XCHECK_SZ(sz
, nr
, struct cet_user_state
);
550 case XFEATURE_XTILE_DATA
: check_xtile_data_against_struct(sz
); return true;
552 XSTATE_WARN_ON(1, "No structure for xstate: %d\n", nr
);
559 static unsigned int xstate_calculate_size(u64 xfeatures
, bool compacted
)
561 unsigned int topmost
= fls64(xfeatures
) - 1;
562 unsigned int offset
= xstate_offsets
[topmost
];
564 if (topmost
<= XFEATURE_SSE
)
565 return sizeof(struct xregs_state
);
568 offset
= xfeature_get_offset(xfeatures
, topmost
);
569 return offset
+ xstate_sizes
[topmost
];
573 * This essentially double-checks what the cpu told us about
574 * how large the XSAVE buffer needs to be. We are recalculating
577 * Independent XSAVE features allocate their own buffers and are not
578 * covered by these checks. Only the size of the buffer for task->fpu
581 static bool __init
paranoid_xstate_size_valid(unsigned int kernel_size
)
583 bool compacted
= cpu_feature_enabled(X86_FEATURE_XCOMPACTED
);
584 bool xsaves
= cpu_feature_enabled(X86_FEATURE_XSAVES
);
585 unsigned int size
= FXSAVE_SIZE
+ XSAVE_HDR_SIZE
;
588 for_each_extended_xfeature(i
, fpu_kernel_cfg
.max_features
) {
589 if (!check_xstate_against_struct(i
))
592 * Supervisor state components can be managed only by
595 if (!xsaves
&& xfeature_is_supervisor(i
)) {
596 XSTATE_WARN_ON(1, "Got supervisor feature %d, but XSAVES not advertised\n", i
);
600 size
= xstate_calculate_size(fpu_kernel_cfg
.max_features
, compacted
);
601 XSTATE_WARN_ON(size
!= kernel_size
,
602 "size %u != kernel_size %u\n", size
, kernel_size
);
603 return size
== kernel_size
;
607 * Get total size of enabled xstates in XCR0 | IA32_XSS.
609 * Note the SDM's wording here. "sub-function 0" only enumerates
610 * the size of the *user* states. If we use it to size a buffer
611 * that we use 'XSAVES' on, we could potentially overflow the
612 * buffer because 'XSAVES' saves system states too.
614 * This also takes compaction into account. So this works for
617 static unsigned int __init
get_compacted_size(void)
619 unsigned int eax
, ebx
, ecx
, edx
;
621 * - CPUID function 0DH, sub-function 1:
622 * EBX enumerates the size (in bytes) required by
623 * the XSAVES instruction for an XSAVE area
624 * containing all the state components
625 * corresponding to bits currently set in
628 * When XSAVES is not available but XSAVEC is (virt), then there
629 * are no supervisor states, but XSAVEC still uses compacted
632 cpuid_count(XSTATE_CPUID
, 1, &eax
, &ebx
, &ecx
, &edx
);
637 * Get the total size of the enabled xstates without the independent supervisor
640 static unsigned int __init
get_xsave_compacted_size(void)
642 u64 mask
= xfeatures_mask_independent();
646 return get_compacted_size();
648 /* Disable independent features. */
649 wrmsrl(MSR_IA32_XSS
, xfeatures_mask_supervisor());
652 * Ask the hardware what size is required of the buffer.
653 * This is the size required for the task->fpu buffer.
655 size
= get_compacted_size();
657 /* Re-enable independent features so XSAVES will work on them again. */
658 wrmsrl(MSR_IA32_XSS
, xfeatures_mask_supervisor() | mask
);
663 static unsigned int __init
get_xsave_size_user(void)
665 unsigned int eax
, ebx
, ecx
, edx
;
667 * - CPUID function 0DH, sub-function 0:
668 * EBX enumerates the size (in bytes) required by
669 * the XSAVE instruction for an XSAVE area
670 * containing all the *user* state components
671 * corresponding to bits currently set in XCR0.
673 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
677 static int __init
init_xstate_size(void)
679 /* Recompute the context size for enabled features: */
680 unsigned int user_size
, kernel_size
, kernel_default_size
;
681 bool compacted
= cpu_feature_enabled(X86_FEATURE_XCOMPACTED
);
683 /* Uncompacted user space size */
684 user_size
= get_xsave_size_user();
687 * XSAVES kernel size includes supervisor states and uses compacted
688 * format. XSAVEC uses compacted format, but does not save
691 * XSAVE[OPT] do not support supervisor states so kernel and user
695 kernel_size
= get_xsave_compacted_size();
697 kernel_size
= user_size
;
699 kernel_default_size
=
700 xstate_calculate_size(fpu_kernel_cfg
.default_features
, compacted
);
702 if (!paranoid_xstate_size_valid(kernel_size
))
705 fpu_kernel_cfg
.max_size
= kernel_size
;
706 fpu_user_cfg
.max_size
= user_size
;
708 fpu_kernel_cfg
.default_size
= kernel_default_size
;
709 fpu_user_cfg
.default_size
=
710 xstate_calculate_size(fpu_user_cfg
.default_features
, false);
716 * We enabled the XSAVE hardware, but something went wrong and
717 * we can not use it. Disable it.
719 static void __init
fpu__init_disable_system_xstate(unsigned int legacy_size
)
721 fpu_kernel_cfg
.max_features
= 0;
722 cr4_clear_bits(X86_CR4_OSXSAVE
);
723 setup_clear_cpu_cap(X86_FEATURE_XSAVE
);
725 /* Restore the legacy size.*/
726 fpu_kernel_cfg
.max_size
= legacy_size
;
727 fpu_kernel_cfg
.default_size
= legacy_size
;
728 fpu_user_cfg
.max_size
= legacy_size
;
729 fpu_user_cfg
.default_size
= legacy_size
;
732 * Prevent enabling the static branch which enables writes to the
735 init_fpstate
.xfd
= 0;
737 fpstate_reset(¤t
->thread
.fpu
);
741 * Enable and initialize the xsave feature.
742 * Called once per system bootup.
744 void __init
fpu__init_system_xstate(unsigned int legacy_size
)
746 unsigned int eax
, ebx
, ecx
, edx
;
751 if (!boot_cpu_has(X86_FEATURE_FPU
)) {
752 pr_info("x86/fpu: No FPU detected\n");
756 if (!boot_cpu_has(X86_FEATURE_XSAVE
)) {
757 pr_info("x86/fpu: x87 FPU will use %s\n",
758 boot_cpu_has(X86_FEATURE_FXSR
) ? "FXSAVE" : "FSAVE");
762 if (boot_cpu_data
.cpuid_level
< XSTATE_CPUID
) {
768 * Find user xstates supported by the processor.
770 cpuid_count(XSTATE_CPUID
, 0, &eax
, &ebx
, &ecx
, &edx
);
771 fpu_kernel_cfg
.max_features
= eax
+ ((u64
)edx
<< 32);
774 * Find supervisor xstates supported by the processor.
776 cpuid_count(XSTATE_CPUID
, 1, &eax
, &ebx
, &ecx
, &edx
);
777 fpu_kernel_cfg
.max_features
|= ecx
+ ((u64
)edx
<< 32);
779 if ((fpu_kernel_cfg
.max_features
& XFEATURE_MASK_FPSSE
) != XFEATURE_MASK_FPSSE
) {
781 * This indicates that something really unexpected happened
782 * with the enumeration. Disable XSAVE and try to continue
783 * booting without it. This is too early to BUG().
785 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n",
786 fpu_kernel_cfg
.max_features
);
791 * Clear XSAVE features that are disabled in the normal CPUID.
793 for (i
= 0; i
< ARRAY_SIZE(xsave_cpuid_features
); i
++) {
794 unsigned short cid
= xsave_cpuid_features
[i
];
796 /* Careful: X86_FEATURE_FPU is 0! */
797 if ((i
!= XFEATURE_FP
&& !cid
) || !boot_cpu_has(cid
))
798 fpu_kernel_cfg
.max_features
&= ~BIT_ULL(i
);
801 if (!cpu_feature_enabled(X86_FEATURE_XFD
))
802 fpu_kernel_cfg
.max_features
&= ~XFEATURE_MASK_USER_DYNAMIC
;
804 if (!cpu_feature_enabled(X86_FEATURE_XSAVES
))
805 fpu_kernel_cfg
.max_features
&= XFEATURE_MASK_USER_SUPPORTED
;
807 fpu_kernel_cfg
.max_features
&= XFEATURE_MASK_USER_SUPPORTED
|
808 XFEATURE_MASK_SUPERVISOR_SUPPORTED
;
810 fpu_user_cfg
.max_features
= fpu_kernel_cfg
.max_features
;
811 fpu_user_cfg
.max_features
&= XFEATURE_MASK_USER_SUPPORTED
;
813 /* Clean out dynamic features from default */
814 fpu_kernel_cfg
.default_features
= fpu_kernel_cfg
.max_features
;
815 fpu_kernel_cfg
.default_features
&= ~XFEATURE_MASK_USER_DYNAMIC
;
817 fpu_user_cfg
.default_features
= fpu_user_cfg
.max_features
;
818 fpu_user_cfg
.default_features
&= ~XFEATURE_MASK_USER_DYNAMIC
;
820 /* Store it for paranoia check at the end */
821 xfeatures
= fpu_kernel_cfg
.max_features
;
824 * Initialize the default XFD state in initfp_state and enable the
825 * dynamic sizing mechanism if dynamic states are available. The
826 * static key cannot be enabled here because this runs before
827 * jump_label_init(). This is delayed to an initcall.
829 init_fpstate
.xfd
= fpu_user_cfg
.max_features
& XFEATURE_MASK_USER_DYNAMIC
;
831 /* Set up compaction feature bit */
832 if (cpu_feature_enabled(X86_FEATURE_XSAVEC
) ||
833 cpu_feature_enabled(X86_FEATURE_XSAVES
))
834 setup_force_cpu_cap(X86_FEATURE_XCOMPACTED
);
836 /* Enable xstate instructions to be able to continue with initialization: */
837 fpu__init_cpu_xstate();
839 /* Cache size, offset and flags for initialization */
840 setup_xstate_cache();
842 err
= init_xstate_size();
846 /* Reset the state for the current task */
847 fpstate_reset(¤t
->thread
.fpu
);
850 * Update info used for ptrace frames; use standard-format size and no
851 * supervisor xstates:
853 update_regset_xstate_info(fpu_user_cfg
.max_size
,
854 fpu_user_cfg
.max_features
);
857 * init_fpstate excludes dynamic states as they are large but init
860 init_fpstate
.size
= fpu_kernel_cfg
.default_size
;
861 init_fpstate
.xfeatures
= fpu_kernel_cfg
.default_features
;
863 if (init_fpstate
.size
> sizeof(init_fpstate
.regs
)) {
864 pr_warn("x86/fpu: init_fpstate buffer too small (%zu < %d), disabling XSAVE\n",
865 sizeof(init_fpstate
.regs
), init_fpstate
.size
);
869 setup_init_fpu_buf();
872 * Paranoia check whether something in the setup modified the
875 if (xfeatures
!= fpu_kernel_cfg
.max_features
) {
876 pr_err("x86/fpu: xfeatures modified from 0x%016llx to 0x%016llx during init, disabling XSAVE\n",
877 xfeatures
, fpu_kernel_cfg
.max_features
);
882 * CPU capabilities initialization runs before FPU init. So
883 * X86_FEATURE_OSXSAVE is not set. Now that XSAVE is completely
884 * functional, set the feature bit so depending code works.
886 setup_force_cpu_cap(X86_FEATURE_OSXSAVE
);
888 print_xstate_offset_size();
889 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
890 fpu_kernel_cfg
.max_features
,
891 fpu_kernel_cfg
.max_size
,
892 boot_cpu_has(X86_FEATURE_XCOMPACTED
) ? "compacted" : "standard");
896 /* something went wrong, try to boot without any XSAVE support */
897 fpu__init_disable_system_xstate(legacy_size
);
901 * Restore minimal FPU state after suspend:
903 void fpu__resume_cpu(void)
906 * Restore XCR0 on xsave capable CPUs:
908 if (cpu_feature_enabled(X86_FEATURE_XSAVE
))
909 xsetbv(XCR_XFEATURE_ENABLED_MASK
, fpu_user_cfg
.max_features
);
912 * Restore IA32_XSS. The same CPUID bit enumerates support
913 * of XSAVES and MSR_IA32_XSS.
915 if (cpu_feature_enabled(X86_FEATURE_XSAVES
)) {
916 wrmsrl(MSR_IA32_XSS
, xfeatures_mask_supervisor() |
917 xfeatures_mask_independent());
920 if (fpu_state_size_dynamic())
921 wrmsrl(MSR_IA32_XFD
, current
->thread
.fpu
.fpstate
->xfd
);
925 * Given an xstate feature nr, calculate where in the xsave
926 * buffer the state is. Callers should ensure that the buffer
929 static void *__raw_xsave_addr(struct xregs_state
*xsave
, int xfeature_nr
)
931 u64 xcomp_bv
= xsave
->header
.xcomp_bv
;
933 if (WARN_ON_ONCE(!xfeature_enabled(xfeature_nr
)))
936 if (cpu_feature_enabled(X86_FEATURE_XCOMPACTED
)) {
937 if (WARN_ON_ONCE(!(xcomp_bv
& BIT_ULL(xfeature_nr
))))
941 return (void *)xsave
+ xfeature_get_offset(xcomp_bv
, xfeature_nr
);
945 * Given the xsave area and a state inside, this function returns the
946 * address of the state.
948 * This is the API that is called to get xstate address in either
949 * standard format or compacted format of xsave area.
951 * Note that if there is no data for the field in the xsave buffer
952 * this will return NULL.
955 * xstate: the thread's storage area for all FPU data
956 * xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
957 * XFEATURE_SSE, etc...)
959 * address of the state in the xsave area, or NULL if the
960 * field is not present in the xsave buffer.
962 void *get_xsave_addr(struct xregs_state
*xsave
, int xfeature_nr
)
965 * Do we even *have* xsave state?
967 if (!boot_cpu_has(X86_FEATURE_XSAVE
))
971 * We should not ever be requesting features that we
974 if (WARN_ON_ONCE(!xfeature_enabled(xfeature_nr
)))
978 * This assumes the last 'xsave*' instruction to
979 * have requested that 'xfeature_nr' be saved.
980 * If it did not, we might be seeing and old value
981 * of the field in the buffer.
983 * This can happen because the last 'xsave' did not
984 * request that this feature be saved (unlikely)
985 * or because the "init optimization" caused it
988 if (!(xsave
->header
.xfeatures
& BIT_ULL(xfeature_nr
)))
991 return __raw_xsave_addr(xsave
, xfeature_nr
);
994 #ifdef CONFIG_ARCH_HAS_PKEYS
997 * This will go out and modify PKRU register to set the access
998 * rights for @pkey to @init_val.
1000 int arch_set_user_pkey_access(struct task_struct
*tsk
, int pkey
,
1001 unsigned long init_val
)
1003 u32 old_pkru
, new_pkru_bits
= 0;
1007 * This check implies XSAVE support. OSPKE only gets
1008 * set if we enable XSAVE and we enable PKU in XCR0.
1010 if (!cpu_feature_enabled(X86_FEATURE_OSPKE
))
1014 * This code should only be called with valid 'pkey'
1015 * values originating from in-kernel users. Complain
1016 * if a bad value is observed.
1018 if (WARN_ON_ONCE(pkey
>= arch_max_pkey()))
1021 /* Set the bits we need in PKRU: */
1022 if (init_val
& PKEY_DISABLE_ACCESS
)
1023 new_pkru_bits
|= PKRU_AD_BIT
;
1024 if (init_val
& PKEY_DISABLE_WRITE
)
1025 new_pkru_bits
|= PKRU_WD_BIT
;
1027 /* Shift the bits in to the correct place in PKRU for pkey: */
1028 pkey_shift
= pkey
* PKRU_BITS_PER_PKEY
;
1029 new_pkru_bits
<<= pkey_shift
;
1031 /* Get old PKRU and mask off any old bits in place: */
1032 old_pkru
= read_pkru();
1033 old_pkru
&= ~((PKRU_AD_BIT
|PKRU_WD_BIT
) << pkey_shift
);
1035 /* Write old part along with new part: */
1036 write_pkru(old_pkru
| new_pkru_bits
);
1040 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
1042 static void copy_feature(bool from_xstate
, struct membuf
*to
, void *xstate
,
1043 void *init_xstate
, unsigned int size
)
1045 membuf_write(to
, from_xstate
? xstate
: init_xstate
, size
);
1049 * __copy_xstate_to_uabi_buf - Copy kernel saved xstate to a UABI buffer
1050 * @to: membuf descriptor
1051 * @fpstate: The fpstate buffer from which to copy
1052 * @pkru_val: The PKRU value to store in the PKRU component
1053 * @copy_mode: The requested copy mode
1055 * Converts from kernel XSAVE or XSAVES compacted format to UABI conforming
1056 * format, i.e. from the kernel internal hardware dependent storage format
1057 * to the requested @mode. UABI XSTATE is always uncompacted!
1059 * It supports partial copy but @to.pos always starts from zero.
1061 void __copy_xstate_to_uabi_buf(struct membuf to
, struct fpstate
*fpstate
,
1062 u32 pkru_val
, enum xstate_copy_mode copy_mode
)
1064 const unsigned int off_mxcsr
= offsetof(struct fxregs_state
, mxcsr
);
1065 struct xregs_state
*xinit
= &init_fpstate
.regs
.xsave
;
1066 struct xregs_state
*xsave
= &fpstate
->regs
.xsave
;
1067 struct xstate_header header
;
1068 unsigned int zerofrom
;
1072 memset(&header
, 0, sizeof(header
));
1073 header
.xfeatures
= xsave
->header
.xfeatures
;
1075 /* Mask out the feature bits depending on copy mode */
1076 switch (copy_mode
) {
1077 case XSTATE_COPY_FP
:
1078 header
.xfeatures
&= XFEATURE_MASK_FP
;
1081 case XSTATE_COPY_FX
:
1082 header
.xfeatures
&= XFEATURE_MASK_FP
| XFEATURE_MASK_SSE
;
1085 case XSTATE_COPY_XSAVE
:
1086 header
.xfeatures
&= fpstate
->user_xfeatures
;
1090 /* Copy FP state up to MXCSR */
1091 copy_feature(header
.xfeatures
& XFEATURE_MASK_FP
, &to
, &xsave
->i387
,
1092 &xinit
->i387
, off_mxcsr
);
1094 /* Copy MXCSR when SSE or YMM are set in the feature mask */
1095 copy_feature(header
.xfeatures
& (XFEATURE_MASK_SSE
| XFEATURE_MASK_YMM
),
1096 &to
, &xsave
->i387
.mxcsr
, &xinit
->i387
.mxcsr
,
1097 MXCSR_AND_FLAGS_SIZE
);
1099 /* Copy the remaining FP state */
1100 copy_feature(header
.xfeatures
& XFEATURE_MASK_FP
,
1101 &to
, &xsave
->i387
.st_space
, &xinit
->i387
.st_space
,
1102 sizeof(xsave
->i387
.st_space
));
1104 /* Copy the SSE state - shared with YMM, but independently managed */
1105 copy_feature(header
.xfeatures
& XFEATURE_MASK_SSE
,
1106 &to
, &xsave
->i387
.xmm_space
, &xinit
->i387
.xmm_space
,
1107 sizeof(xsave
->i387
.xmm_space
));
1109 if (copy_mode
!= XSTATE_COPY_XSAVE
)
1112 /* Zero the padding area */
1113 membuf_zero(&to
, sizeof(xsave
->i387
.padding
));
1115 /* Copy xsave->i387.sw_reserved */
1116 membuf_write(&to
, xstate_fx_sw_bytes
, sizeof(xsave
->i387
.sw_reserved
));
1118 /* Copy the user space relevant state of @xsave->header */
1119 membuf_write(&to
, &header
, sizeof(header
));
1121 zerofrom
= offsetof(struct xregs_state
, extended_state_area
);
1124 * This 'mask' indicates which states to copy from fpstate.
1125 * Those extended states that are not present in fpstate are
1126 * either disabled or initialized:
1128 * In non-compacted format, disabled features still occupy
1129 * state space but there is no state to copy from in the
1130 * compacted init_fpstate. The gap tracking will zero these
1133 * The extended features have an all zeroes init state. Thus,
1134 * remove them from 'mask' to zero those features in the user
1135 * buffer instead of retrieving them from init_fpstate.
1137 mask
= header
.xfeatures
;
1139 for_each_extended_xfeature(i
, mask
) {
1141 * If there was a feature or alignment gap, zero the space
1142 * in the destination buffer.
1144 if (zerofrom
< xstate_offsets
[i
])
1145 membuf_zero(&to
, xstate_offsets
[i
] - zerofrom
);
1147 if (i
== XFEATURE_PKRU
) {
1148 struct pkru_state pkru
= {0};
1150 * PKRU is not necessarily up to date in the
1151 * XSAVE buffer. Use the provided value.
1153 pkru
.pkru
= pkru_val
;
1154 membuf_write(&to
, &pkru
, sizeof(pkru
));
1157 __raw_xsave_addr(xsave
, i
),
1161 * Keep track of the last copied state in the non-compacted
1162 * target buffer for gap zeroing.
1164 zerofrom
= xstate_offsets
[i
] + xstate_sizes
[i
];
1169 membuf_zero(&to
, to
.left
);
1173 * copy_xstate_to_uabi_buf - Copy kernel saved xstate to a UABI buffer
1174 * @to: membuf descriptor
1175 * @tsk: The task from which to copy the saved xstate
1176 * @copy_mode: The requested copy mode
1178 * Converts from kernel XSAVE or XSAVES compacted format to UABI conforming
1179 * format, i.e. from the kernel internal hardware dependent storage format
1180 * to the requested @mode. UABI XSTATE is always uncompacted!
1182 * It supports partial copy but @to.pos always starts from zero.
1184 void copy_xstate_to_uabi_buf(struct membuf to
, struct task_struct
*tsk
,
1185 enum xstate_copy_mode copy_mode
)
1187 __copy_xstate_to_uabi_buf(to
, tsk
->thread
.fpu
.fpstate
,
1188 tsk
->thread
.pkru
, copy_mode
);
1191 static int copy_from_buffer(void *dst
, unsigned int offset
, unsigned int size
,
1192 const void *kbuf
, const void __user
*ubuf
)
1195 memcpy(dst
, kbuf
+ offset
, size
);
1197 if (copy_from_user(dst
, ubuf
+ offset
, size
))
1205 * copy_uabi_to_xstate - Copy a UABI format buffer to the kernel xstate
1206 * @fpstate: The fpstate buffer to copy to
1207 * @kbuf: The UABI format buffer, if it comes from the kernel
1208 * @ubuf: The UABI format buffer, if it comes from userspace
1209 * @pkru: The location to write the PKRU value to
1211 * Converts from the UABI format into the kernel internal hardware
1214 * This function ultimately has three different callers with distinct PKRU
1216 * 1. When called from sigreturn the PKRU register will be restored from
1217 * @fpstate via an XRSTOR. Correctly copying the UABI format buffer to
1218 * @fpstate is sufficient to cover this case, but the caller will also
1219 * pass a pointer to the thread_struct's pkru field in @pkru and updating
1221 * 2. When called from ptrace the PKRU register will be restored from the
1222 * thread_struct's pkru field. A pointer to that is passed in @pkru.
1223 * The kernel will restore it manually, so the XRSTOR behavior that resets
1224 * the PKRU register to the hardware init value (0) if the corresponding
1225 * xfeatures bit is not set is emulated here.
1226 * 3. When called from KVM the PKRU register will be restored from the vcpu's
1227 * pkru field. A pointer to that is passed in @pkru. KVM hasn't used
1228 * XRSTOR and hasn't had the PKRU resetting behavior described above. To
1229 * preserve that KVM behavior, it passes NULL for @pkru if the xfeatures
1232 static int copy_uabi_to_xstate(struct fpstate
*fpstate
, const void *kbuf
,
1233 const void __user
*ubuf
, u32
*pkru
)
1235 struct xregs_state
*xsave
= &fpstate
->regs
.xsave
;
1236 unsigned int offset
, size
;
1237 struct xstate_header hdr
;
1241 offset
= offsetof(struct xregs_state
, header
);
1242 if (copy_from_buffer(&hdr
, offset
, sizeof(hdr
), kbuf
, ubuf
))
1245 if (validate_user_xstate_header(&hdr
, fpstate
))
1248 /* Validate MXCSR when any of the related features is in use */
1249 mask
= XFEATURE_MASK_FP
| XFEATURE_MASK_SSE
| XFEATURE_MASK_YMM
;
1250 if (hdr
.xfeatures
& mask
) {
1253 offset
= offsetof(struct fxregs_state
, mxcsr
);
1254 if (copy_from_buffer(mxcsr
, offset
, sizeof(mxcsr
), kbuf
, ubuf
))
1257 /* Reserved bits in MXCSR must be zero. */
1258 if (mxcsr
[0] & ~mxcsr_feature_mask
)
1261 /* SSE and YMM require MXCSR even when FP is not in use. */
1262 if (!(hdr
.xfeatures
& XFEATURE_MASK_FP
)) {
1263 xsave
->i387
.mxcsr
= mxcsr
[0];
1264 xsave
->i387
.mxcsr_mask
= mxcsr
[1];
1268 for (i
= 0; i
< XFEATURE_MAX
; i
++) {
1271 if (hdr
.xfeatures
& mask
) {
1272 void *dst
= __raw_xsave_addr(xsave
, i
);
1274 offset
= xstate_offsets
[i
];
1275 size
= xstate_sizes
[i
];
1277 if (copy_from_buffer(dst
, offset
, size
, kbuf
, ubuf
))
1282 if (hdr
.xfeatures
& XFEATURE_MASK_PKRU
) {
1283 struct pkru_state
*xpkru
;
1285 xpkru
= __raw_xsave_addr(xsave
, XFEATURE_PKRU
);
1286 *pkru
= xpkru
->pkru
;
1289 * KVM may pass NULL here to indicate that it does not need
1297 * The state that came in from userspace was user-state only.
1298 * Mask all the user states out of 'xfeatures':
1300 xsave
->header
.xfeatures
&= XFEATURE_MASK_SUPERVISOR_ALL
;
1303 * Add back in the features that came in from userspace:
1305 xsave
->header
.xfeatures
|= hdr
.xfeatures
;
1311 * Convert from a ptrace standard-format kernel buffer to kernel XSAVE[S]
1312 * format and copy to the target thread. Used by ptrace and KVM.
1314 int copy_uabi_from_kernel_to_xstate(struct fpstate
*fpstate
, const void *kbuf
, u32
*pkru
)
1316 return copy_uabi_to_xstate(fpstate
, kbuf
, NULL
, pkru
);
1320 * Convert from a sigreturn standard-format user-space buffer to kernel
1321 * XSAVE[S] format and copy to the target thread. This is called from the
1322 * sigreturn() and rt_sigreturn() system calls.
1324 int copy_sigframe_from_user_to_xstate(struct task_struct
*tsk
,
1325 const void __user
*ubuf
)
1327 return copy_uabi_to_xstate(tsk
->thread
.fpu
.fpstate
, NULL
, ubuf
, &tsk
->thread
.pkru
);
1330 static bool validate_independent_components(u64 mask
)
1334 if (WARN_ON_FPU(!cpu_feature_enabled(X86_FEATURE_XSAVES
)))
1337 xchk
= ~xfeatures_mask_independent();
1339 if (WARN_ON_ONCE(!mask
|| mask
& xchk
))
1346 * xsaves - Save selected components to a kernel xstate buffer
1347 * @xstate: Pointer to the buffer
1348 * @mask: Feature mask to select the components to save
1350 * The @xstate buffer must be 64 byte aligned and correctly initialized as
1351 * XSAVES does not write the full xstate header. Before first use the
1352 * buffer should be zeroed otherwise a consecutive XRSTORS from that buffer
1355 * The feature mask must be a subset of the independent features.
1357 void xsaves(struct xregs_state
*xstate
, u64 mask
)
1361 if (!validate_independent_components(mask
))
1364 XSTATE_OP(XSAVES
, xstate
, (u32
)mask
, (u32
)(mask
>> 32), err
);
1369 * xrstors - Restore selected components from a kernel xstate buffer
1370 * @xstate: Pointer to the buffer
1371 * @mask: Feature mask to select the components to restore
1373 * The @xstate buffer must be 64 byte aligned and correctly initialized
1374 * otherwise XRSTORS from that buffer can #GP.
1376 * Proper usage is to restore the state which was saved with
1377 * xsaves() into @xstate.
1379 * The feature mask must be a subset of the independent features.
1381 void xrstors(struct xregs_state
*xstate
, u64 mask
)
1385 if (!validate_independent_components(mask
))
1388 XSTATE_OP(XRSTORS
, xstate
, (u32
)mask
, (u32
)(mask
>> 32), err
);
1392 #if IS_ENABLED(CONFIG_KVM)
1393 void fpstate_clear_xstate_component(struct fpstate
*fps
, unsigned int xfeature
)
1395 void *addr
= get_xsave_addr(&fps
->regs
.xsave
, xfeature
);
1398 memset(addr
, 0, xstate_sizes
[xfeature
]);
1400 EXPORT_SYMBOL_GPL(fpstate_clear_xstate_component
);
1403 #ifdef CONFIG_X86_64
1405 #ifdef CONFIG_X86_DEBUG_FPU
1407 * Ensure that a subsequent XSAVE* or XRSTOR* instruction with RFBM=@mask
1408 * can safely operate on the @fpstate buffer.
1410 static bool xstate_op_valid(struct fpstate
*fpstate
, u64 mask
, bool rstor
)
1412 u64 xfd
= __this_cpu_read(xfd_state
);
1414 if (fpstate
->xfd
== xfd
)
1418 * The XFD MSR does not match fpstate->xfd. That's invalid when
1419 * the passed in fpstate is current's fpstate.
1421 if (fpstate
->xfd
== current
->thread
.fpu
.fpstate
->xfd
)
1425 * XRSTOR(S) from init_fpstate are always correct as it will just
1426 * bring all components into init state and not read from the
1427 * buffer. XSAVE(S) raises #PF after init.
1429 if (fpstate
== &init_fpstate
)
1433 * XSAVE(S): clone(), fpu_swap_kvm_fpu()
1434 * XRSTORS(S): fpu_swap_kvm_fpu()
1438 * No XSAVE/XRSTOR instructions (except XSAVE itself) touch
1439 * the buffer area for XFD-disabled state components.
1444 * Remove features which are valid in fpstate. They
1445 * have space allocated in fpstate.
1447 mask
&= ~fpstate
->xfeatures
;
1450 * Any remaining state components in 'mask' might be written
1451 * by XSAVE/XRSTOR. Fail validation it found.
1456 void xfd_validate_state(struct fpstate
*fpstate
, u64 mask
, bool rstor
)
1458 WARN_ON_ONCE(!xstate_op_valid(fpstate
, mask
, rstor
));
1460 #endif /* CONFIG_X86_DEBUG_FPU */
1462 static int __init
xfd_update_static_branch(void)
1465 * If init_fpstate.xfd has bits set then dynamic features are
1466 * available and the dynamic sizing must be enabled.
1468 if (init_fpstate
.xfd
)
1469 static_branch_enable(&__fpu_state_size_dynamic
);
1472 arch_initcall(xfd_update_static_branch
)
1474 void fpstate_free(struct fpu
*fpu
)
1476 if (fpu
->fpstate
&& fpu
->fpstate
!= &fpu
->__fpstate
)
1477 vfree(fpu
->fpstate
);
1481 * fpstate_realloc - Reallocate struct fpstate for the requested new features
1483 * @xfeatures: A bitmap of xstate features which extend the enabled features
1485 * @ksize: The required size for the kernel buffer
1486 * @usize: The required size for user space buffers
1487 * @guest_fpu: Pointer to a guest FPU container. NULL for host allocations
1489 * Note vs. vmalloc(): If the task with a vzalloc()-allocated buffer
1490 * terminates quickly, vfree()-induced IPIs may be a concern, but tasks
1491 * with large states are likely to live longer.
1493 * Returns: 0 on success, -ENOMEM on allocation error.
1495 static int fpstate_realloc(u64 xfeatures
, unsigned int ksize
,
1496 unsigned int usize
, struct fpu_guest
*guest_fpu
)
1498 struct fpu
*fpu
= ¤t
->thread
.fpu
;
1499 struct fpstate
*curfps
, *newfps
= NULL
;
1500 unsigned int fpsize
;
1503 fpsize
= ksize
+ ALIGN(offsetof(struct fpstate
, regs
), 64);
1505 newfps
= vzalloc(fpsize
);
1508 newfps
->size
= ksize
;
1509 newfps
->user_size
= usize
;
1510 newfps
->is_valloc
= true;
1513 * When a guest FPU is supplied, use @guest_fpu->fpstate
1514 * as reference independent whether it is in use or not.
1516 curfps
= guest_fpu
? guest_fpu
->fpstate
: fpu
->fpstate
;
1518 /* Determine whether @curfps is the active fpstate */
1519 in_use
= fpu
->fpstate
== curfps
;
1522 newfps
->is_guest
= true;
1523 newfps
->is_confidential
= curfps
->is_confidential
;
1524 newfps
->in_use
= curfps
->in_use
;
1525 guest_fpu
->xfeatures
|= xfeatures
;
1526 guest_fpu
->uabi_size
= usize
;
1531 * If @curfps is in use, ensure that the current state is in the
1532 * registers before swapping fpstate as that might invalidate it
1533 * due to layout changes.
1535 if (in_use
&& test_thread_flag(TIF_NEED_FPU_LOAD
))
1536 fpregs_restore_userregs();
1538 newfps
->xfeatures
= curfps
->xfeatures
| xfeatures
;
1541 newfps
->user_xfeatures
= curfps
->user_xfeatures
| xfeatures
;
1543 newfps
->xfd
= curfps
->xfd
& ~xfeatures
;
1545 /* Do the final updates within the locked region */
1546 xstate_init_xcomp_bv(&newfps
->regs
.xsave
, newfps
->xfeatures
);
1549 guest_fpu
->fpstate
= newfps
;
1550 /* If curfps is active, update the FPU fpstate pointer */
1552 fpu
->fpstate
= newfps
;
1554 fpu
->fpstate
= newfps
;
1558 xfd_update_state(fpu
->fpstate
);
1561 /* Only free valloc'ed state */
1562 if (curfps
&& curfps
->is_valloc
)
1568 static int validate_sigaltstack(unsigned int usize
)
1570 struct task_struct
*thread
, *leader
= current
->group_leader
;
1571 unsigned long framesize
= get_sigframe_size();
1573 lockdep_assert_held(¤t
->sighand
->siglock
);
1575 /* get_sigframe_size() is based on fpu_user_cfg.max_size */
1576 framesize
-= fpu_user_cfg
.max_size
;
1578 for_each_thread(leader
, thread
) {
1579 if (thread
->sas_ss_size
&& thread
->sas_ss_size
< framesize
)
1585 static int __xstate_request_perm(u64 permitted
, u64 requested
, bool guest
)
1588 * This deliberately does not exclude !XSAVES as we still might
1589 * decide to optionally context switch XCR0 or talk the silicon
1590 * vendors into extending XFD for the pre AMX states, especially
1593 bool compacted
= cpu_feature_enabled(X86_FEATURE_XCOMPACTED
);
1594 struct fpu
*fpu
= ¤t
->group_leader
->thread
.fpu
;
1595 struct fpu_state_perm
*perm
;
1596 unsigned int ksize
, usize
;
1600 /* Check whether fully enabled */
1601 if ((permitted
& requested
) == requested
)
1604 /* Calculate the resulting kernel state size */
1605 mask
= permitted
| requested
;
1606 /* Take supervisor states into account on the host */
1608 mask
|= xfeatures_mask_supervisor();
1609 ksize
= xstate_calculate_size(mask
, compacted
);
1611 /* Calculate the resulting user state size */
1612 mask
&= XFEATURE_MASK_USER_SUPPORTED
;
1613 usize
= xstate_calculate_size(mask
, false);
1616 ret
= validate_sigaltstack(usize
);
1621 perm
= guest
? &fpu
->guest_perm
: &fpu
->perm
;
1622 /* Pairs with the READ_ONCE() in xstate_get_group_perm() */
1623 WRITE_ONCE(perm
->__state_perm
, mask
);
1624 /* Protected by sighand lock */
1625 perm
->__state_size
= ksize
;
1626 perm
->__user_state_size
= usize
;
1631 * Permissions array to map facilities with more than one component
1633 static const u64 xstate_prctl_req
[XFEATURE_MAX
] = {
1634 [XFEATURE_XTILE_DATA
] = XFEATURE_MASK_XTILE_DATA
,
1637 static int xstate_request_perm(unsigned long idx
, bool guest
)
1639 u64 permitted
, requested
;
1642 if (idx
>= XFEATURE_MAX
)
1646 * Look up the facility mask which can require more than
1647 * one xstate component.
1649 idx
= array_index_nospec(idx
, ARRAY_SIZE(xstate_prctl_req
));
1650 requested
= xstate_prctl_req
[idx
];
1654 if ((fpu_user_cfg
.max_features
& requested
) != requested
)
1657 /* Lockless quick check */
1658 permitted
= xstate_get_group_perm(guest
);
1659 if ((permitted
& requested
) == requested
)
1662 /* Protect against concurrent modifications */
1663 spin_lock_irq(¤t
->sighand
->siglock
);
1664 permitted
= xstate_get_group_perm(guest
);
1666 /* First vCPU allocation locks the permissions. */
1667 if (guest
&& (permitted
& FPU_GUEST_PERM_LOCKED
))
1670 ret
= __xstate_request_perm(permitted
, requested
, guest
);
1671 spin_unlock_irq(¤t
->sighand
->siglock
);
1675 int __xfd_enable_feature(u64 xfd_err
, struct fpu_guest
*guest_fpu
)
1677 u64 xfd_event
= xfd_err
& XFEATURE_MASK_USER_DYNAMIC
;
1678 struct fpu_state_perm
*perm
;
1679 unsigned int ksize
, usize
;
1684 pr_err_once("XFD: Invalid xfd error: %016llx\n", xfd_err
);
1688 /* Protect against concurrent modifications */
1689 spin_lock_irq(¤t
->sighand
->siglock
);
1691 /* If not permitted let it die */
1692 if ((xstate_get_group_perm(!!guest_fpu
) & xfd_event
) != xfd_event
) {
1693 spin_unlock_irq(¤t
->sighand
->siglock
);
1697 fpu
= ¤t
->group_leader
->thread
.fpu
;
1698 perm
= guest_fpu
? &fpu
->guest_perm
: &fpu
->perm
;
1699 ksize
= perm
->__state_size
;
1700 usize
= perm
->__user_state_size
;
1703 * The feature is permitted. State size is sufficient. Dropping
1704 * the lock is safe here even if more features are added from
1705 * another task, the retrieved buffer sizes are valid for the
1706 * currently requested feature(s).
1708 spin_unlock_irq(¤t
->sighand
->siglock
);
1711 * Try to allocate a new fpstate. If that fails there is no way
1714 if (fpstate_realloc(xfd_event
, ksize
, usize
, guest_fpu
))
1719 int xfd_enable_feature(u64 xfd_err
)
1721 return __xfd_enable_feature(xfd_err
, NULL
);
1724 #else /* CONFIG_X86_64 */
1725 static inline int xstate_request_perm(unsigned long idx
, bool guest
)
1729 #endif /* !CONFIG_X86_64 */
1731 u64
xstate_get_guest_group_perm(void)
1733 return xstate_get_group_perm(true);
1735 EXPORT_SYMBOL_GPL(xstate_get_guest_group_perm
);
1738 * fpu_xstate_prctl - xstate permission operations
1739 * @tsk: Redundant pointer to current
1740 * @option: A subfunction of arch_prctl()
1741 * @arg2: option argument
1742 * Return: 0 if successful; otherwise, an error code
1746 * ARCH_GET_XCOMP_SUPP: Pointer to user space u64 to store the info
1747 * ARCH_GET_XCOMP_PERM: Pointer to user space u64 to store the info
1748 * ARCH_REQ_XCOMP_PERM: Facility number requested
1750 * For facilities which require more than one XSTATE component, the request
1751 * must be the highest state component number related to that facility,
1752 * e.g. for AMX which requires XFEATURE_XTILE_CFG(17) and
1753 * XFEATURE_XTILE_DATA(18) this would be XFEATURE_XTILE_DATA(18).
1755 long fpu_xstate_prctl(int option
, unsigned long arg2
)
1757 u64 __user
*uptr
= (u64 __user
*)arg2
;
1758 u64 permitted
, supported
;
1759 unsigned long idx
= arg2
;
1763 case ARCH_GET_XCOMP_SUPP
:
1764 supported
= fpu_user_cfg
.max_features
| fpu_user_cfg
.legacy_features
;
1765 return put_user(supported
, uptr
);
1767 case ARCH_GET_XCOMP_PERM
:
1769 * Lockless snapshot as it can also change right after the
1770 * dropping the lock.
1772 permitted
= xstate_get_host_group_perm();
1773 permitted
&= XFEATURE_MASK_USER_SUPPORTED
;
1774 return put_user(permitted
, uptr
);
1776 case ARCH_GET_XCOMP_GUEST_PERM
:
1777 permitted
= xstate_get_guest_group_perm();
1778 permitted
&= XFEATURE_MASK_USER_SUPPORTED
;
1779 return put_user(permitted
, uptr
);
1781 case ARCH_REQ_XCOMP_GUEST_PERM
:
1785 case ARCH_REQ_XCOMP_PERM
:
1786 if (!IS_ENABLED(CONFIG_X86_64
))
1789 return xstate_request_perm(idx
, guest
);
1796 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1798 * Report the amount of time elapsed in millisecond since last AVX512
1801 static void avx512_status(struct seq_file
*m
, struct task_struct
*task
)
1803 unsigned long timestamp
= READ_ONCE(task
->thread
.fpu
.avx512_timestamp
);
1808 * Report -1 if no AVX512 usage
1812 delta
= (long)(jiffies
- timestamp
);
1814 * Cap to LONG_MAX if time difference > LONG_MAX
1818 delta
= jiffies_to_msecs(delta
);
1821 seq_put_decimal_ll(m
, "AVX512_elapsed_ms:\t", delta
);
1826 * Report architecture specific information
1828 int proc_pid_arch_status(struct seq_file
*m
, struct pid_namespace
*ns
,
1829 struct pid
*pid
, struct task_struct
*task
)
1832 * Report AVX512 state if the processor and build option supported.
1834 if (cpu_feature_enabled(X86_FEATURE_AVX512F
))
1835 avx512_status(m
, task
);
1839 #endif /* CONFIG_PROC_PID_ARCH_STATUS */