2 * xsave/xrstor support.
4 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
6 #include <linux/compat.h>
8 #include <linux/mman.h>
9 #include <linux/pkeys.h>
11 #include <asm/fpu/api.h>
12 #include <asm/fpu/internal.h>
13 #include <asm/fpu/signal.h>
14 #include <asm/fpu/regset.h>
15 #include <asm/fpu/xstate.h>
17 #include <asm/tlbflush.h>
20 * Although we spell it out in here, the Processor Trace
21 * xfeature is completely unused. We use other mechanisms
22 * to save/restore PT state in Linux.
24 static const char *xfeature_names[] =
26 "x87 floating point registers" ,
29 "MPX bounds registers" ,
34 "Processor Trace (unused)" ,
35 "Protection Keys User registers",
36 "unknown xstate feature" ,
40 * Mask of xstate features supported by the CPU and the kernel:
42 u64 xfeatures_mask __read_mostly;
44 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
45 static unsigned int xstate_sizes[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
46 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
49 * The XSAVE area of kernel can be in standard or compacted format;
50 * it is always in standard format for user mode. This is the user
51 * mode standard format size used for signal and ptrace frames.
53 unsigned int fpu_user_xstate_size;
56 * Clear all of the X86_FEATURE_* bits that are unavailable
57 * when the CPU has no XSAVE support.
59 void fpu__xstate_clear_all_cpu_caps(void)
61 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
62 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
63 setup_clear_cpu_cap(X86_FEATURE_XSAVEC);
64 setup_clear_cpu_cap(X86_FEATURE_XSAVES);
65 setup_clear_cpu_cap(X86_FEATURE_AVX);
66 setup_clear_cpu_cap(X86_FEATURE_AVX2);
67 setup_clear_cpu_cap(X86_FEATURE_AVX512F);
68 setup_clear_cpu_cap(X86_FEATURE_AVX512PF);
69 setup_clear_cpu_cap(X86_FEATURE_AVX512ER);
70 setup_clear_cpu_cap(X86_FEATURE_AVX512CD);
71 setup_clear_cpu_cap(X86_FEATURE_AVX512DQ);
72 setup_clear_cpu_cap(X86_FEATURE_AVX512BW);
73 setup_clear_cpu_cap(X86_FEATURE_AVX512VL);
74 setup_clear_cpu_cap(X86_FEATURE_MPX);
75 setup_clear_cpu_cap(X86_FEATURE_XGETBV1);
76 setup_clear_cpu_cap(X86_FEATURE_PKU);
77 setup_clear_cpu_cap(X86_FEATURE_AVX512_4VNNIW);
78 setup_clear_cpu_cap(X86_FEATURE_AVX512_4FMAPS);
82 * Return whether the system supports a given xfeature.
84 * Also return the name of the (most advanced) feature that the caller requested:
86 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
88 u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
90 if (unlikely(feature_name)) {
91 long xfeature_idx, max_idx;
94 * So we use FLS here to be able to print the most advanced
95 * feature that was requested but is missing. So if a driver
96 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
97 * missing AVX feature - this is the most informative message
100 if (xfeatures_missing)
101 xfeatures_print = xfeatures_missing;
103 xfeatures_print = xfeatures_needed;
105 xfeature_idx = fls64(xfeatures_print)-1;
106 max_idx = ARRAY_SIZE(xfeature_names)-1;
107 xfeature_idx = min(xfeature_idx, max_idx);
109 *feature_name = xfeature_names[xfeature_idx];
112 if (xfeatures_missing)
117 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
119 static int xfeature_is_supervisor(int xfeature_nr)
122 * We currently do not support supervisor states, but if
123 * we did, we could find out like this.
125 * SDM says: If state component 'i' is a user state component,
126 * ECX[0] return 0; if state component i is a supervisor
127 * state component, ECX[0] returns 1.
129 u32 eax, ebx, ecx, edx;
131 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
135 static int xfeature_is_user(int xfeature_nr)
137 return !xfeature_is_supervisor(xfeature_nr);
141 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
142 * a processor implementation detects that an FPU state component is still
143 * (or is again) in its initialized state, it may clear the corresponding
144 * bit in the header.xfeatures field, and can skip the writeout of registers
145 * to the corresponding memory layout.
147 * This means that when the bit is zero, the state component might still contain
148 * some previous - non-initialized register state.
150 * Before writing xstate information to user-space we sanitize those components,
151 * to always ensure that the memory layout of a feature will be in the init state
152 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
153 * see some stale state in the memory layout during signal handling, debugging etc.
155 void fpstate_sanitize_xstate(struct fpu *fpu)
157 struct fxregs_state *fx = &fpu->state.fxsave;
164 xfeatures = fpu->state.xsave.header.xfeatures;
167 * None of the feature bits are in init state. So nothing else
168 * to do for us, as the memory layout is up to date.
170 if ((xfeatures & xfeatures_mask) == xfeatures_mask)
174 * FP is in init state
176 if (!(xfeatures & XFEATURE_MASK_FP)) {
183 memset(&fx->st_space[0], 0, 128);
187 * SSE is in init state
189 if (!(xfeatures & XFEATURE_MASK_SSE))
190 memset(&fx->xmm_space[0], 0, 256);
193 * First two features are FPU and SSE, which above we handled
194 * in a special way already:
197 xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
200 * Update all the remaining memory layouts according to their
201 * standard xstate layout, if their header bit is in the init
205 if (xfeatures & 0x1) {
206 int offset = xstate_comp_offsets[feature_bit];
207 int size = xstate_sizes[feature_bit];
209 memcpy((void *)fx + offset,
210 (void *)&init_fpstate.xsave + offset,
220 * Enable the extended processor state save/restore feature.
221 * Called once per CPU onlining.
223 void fpu__init_cpu_xstate(void)
225 if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
228 * Make it clear that XSAVES supervisor states are not yet
229 * implemented should anyone expect it to work by changing
230 * bits in XFEATURE_MASK_* macros and XCR0.
232 WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
233 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
235 xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
237 cr4_set_bits(X86_CR4_OSXSAVE);
238 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
242 * Note that in the future we will likely need a pair of
243 * functions here: one for user xstates and the other for
244 * system xstates. For now, they are the same.
246 static int xfeature_enabled(enum xfeature xfeature)
248 return !!(xfeatures_mask & (1UL << xfeature));
252 * Record the offsets and sizes of various xstates contained
253 * in the XSAVE state memory layout.
255 static void __init setup_xstate_features(void)
257 u32 eax, ebx, ecx, edx, i;
258 /* start at the beginnning of the "extended state" */
259 unsigned int last_good_offset = offsetof(struct xregs_state,
260 extended_state_area);
262 * The FP xstates and SSE xstates are legacy states. They are always
263 * in the fixed offsets in the xsave area in either compacted form
266 xstate_offsets[0] = 0;
267 xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space);
268 xstate_offsets[1] = xstate_sizes[0];
269 xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space);
271 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
272 if (!xfeature_enabled(i))
275 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
278 * If an xfeature is supervisor state, the offset
279 * in EBX is invalid. We leave it to -1.
281 if (xfeature_is_user(i))
282 xstate_offsets[i] = ebx;
284 xstate_sizes[i] = eax;
286 * In our xstate size checks, we assume that the
287 * highest-numbered xstate feature has the
288 * highest offset in the buffer. Ensure it does.
290 WARN_ONCE(last_good_offset > xstate_offsets[i],
291 "x86/fpu: misordered xstate at %d\n", last_good_offset);
292 last_good_offset = xstate_offsets[i];
296 static void __init print_xstate_feature(u64 xstate_mask)
298 const char *feature_name;
300 if (cpu_has_xfeatures(xstate_mask, &feature_name))
301 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
305 * Print out all the supported xstate features:
307 static void __init print_xstate_features(void)
309 print_xstate_feature(XFEATURE_MASK_FP);
310 print_xstate_feature(XFEATURE_MASK_SSE);
311 print_xstate_feature(XFEATURE_MASK_YMM);
312 print_xstate_feature(XFEATURE_MASK_BNDREGS);
313 print_xstate_feature(XFEATURE_MASK_BNDCSR);
314 print_xstate_feature(XFEATURE_MASK_OPMASK);
315 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
316 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
317 print_xstate_feature(XFEATURE_MASK_PKRU);
321 * This check is important because it is easy to get XSTATE_*
322 * confused with XSTATE_BIT_*.
324 #define CHECK_XFEATURE(nr) do { \
325 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
326 WARN_ON(nr >= XFEATURE_MAX); \
330 * We could cache this like xstate_size[], but we only use
331 * it here, so it would be a waste of space.
333 static int xfeature_is_aligned(int xfeature_nr)
335 u32 eax, ebx, ecx, edx;
337 CHECK_XFEATURE(xfeature_nr);
338 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
340 * The value returned by ECX[1] indicates the alignment
341 * of state component 'i' when the compacted format
342 * of the extended region of an XSAVE area is used:
348 * This function sets up offsets and sizes of all extended states in
349 * xsave area. This supports both standard format and compacted format
350 * of the xsave aread.
352 static void __init setup_xstate_comp(void)
354 unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
358 * The FP xstates and SSE xstates are legacy states. They are always
359 * in the fixed offsets in the xsave area in either compacted form
362 xstate_comp_offsets[0] = 0;
363 xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
365 if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
366 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
367 if (xfeature_enabled(i)) {
368 xstate_comp_offsets[i] = xstate_offsets[i];
369 xstate_comp_sizes[i] = xstate_sizes[i];
375 xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
376 FXSAVE_SIZE + XSAVE_HDR_SIZE;
378 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
379 if (xfeature_enabled(i))
380 xstate_comp_sizes[i] = xstate_sizes[i];
382 xstate_comp_sizes[i] = 0;
384 if (i > FIRST_EXTENDED_XFEATURE) {
385 xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
386 + xstate_comp_sizes[i-1];
388 if (xfeature_is_aligned(i))
389 xstate_comp_offsets[i] =
390 ALIGN(xstate_comp_offsets[i], 64);
396 * Print out xstate component offsets and sizes
398 static void __init print_xstate_offset_size(void)
402 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
403 if (!xfeature_enabled(i))
405 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
406 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
411 * All supported features have either init state all zeros or are
412 * handled in setup_init_fpu() individually. This is an explicit
413 * feature list and does not use XFEATURE_MASK*SUPPORTED to catch
414 * newly added supported features at build time and make people
415 * actually look at the init state for the new feature.
417 #define XFEATURES_INIT_FPSTATE_HANDLED \
418 (XFEATURE_MASK_FP | \
419 XFEATURE_MASK_SSE | \
420 XFEATURE_MASK_YMM | \
421 XFEATURE_MASK_OPMASK | \
422 XFEATURE_MASK_ZMM_Hi256 | \
423 XFEATURE_MASK_Hi16_ZMM | \
424 XFEATURE_MASK_PKRU | \
425 XFEATURE_MASK_BNDREGS | \
426 XFEATURE_MASK_BNDCSR)
429 * setup the xstate image representing the init state
431 static void __init setup_init_fpu_buf(void)
433 static int on_boot_cpu __initdata = 1;
435 BUILD_BUG_ON(XCNTXT_MASK != XFEATURES_INIT_FPSTATE_HANDLED);
437 WARN_ON_FPU(!on_boot_cpu);
440 if (!boot_cpu_has(X86_FEATURE_XSAVE))
443 setup_xstate_features();
444 print_xstate_features();
446 if (boot_cpu_has(X86_FEATURE_XSAVES))
447 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
450 * Init all the features state with header.xfeatures being 0x0
452 copy_kernel_to_xregs_booting(&init_fpstate.xsave);
455 * All components are now in init state. Read the state back so
456 * that init_fpstate contains all non-zero init state. This only
457 * works with XSAVE, but not with XSAVEOPT and XSAVES because
458 * those use the init optimization which skips writing data for
459 * components in init state.
461 * XSAVE could be used, but that would require to reshuffle the
462 * data when XSAVES is available because XSAVES uses xstate
463 * compaction. But doing so is a pointless exercise because most
464 * components have an all zeros init state except for the legacy
465 * ones (FP and SSE). Those can be saved with FXSAVE into the
466 * legacy area. Adding new features requires to ensure that init
467 * state is all zeroes or if not to add the necessary handling
470 fxsave(&init_fpstate.fxsave);
473 static int xfeature_uncompacted_offset(int xfeature_nr)
475 u32 eax, ebx, ecx, edx;
478 * Only XSAVES supports supervisor states and it uses compacted
479 * format. Checking a supervisor state's uncompacted offset is
482 if (XFEATURE_MASK_SUPERVISOR & (1 << xfeature_nr)) {
483 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
487 CHECK_XFEATURE(xfeature_nr);
488 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
492 static int xfeature_size(int xfeature_nr)
494 u32 eax, ebx, ecx, edx;
496 CHECK_XFEATURE(xfeature_nr);
497 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
502 * 'XSAVES' implies two different things:
503 * 1. saving of supervisor/system state
504 * 2. using the compacted format
506 * Use this function when dealing with the compacted format so
507 * that it is obvious which aspect of 'XSAVES' is being handled
508 * by the calling code.
510 int using_compacted_format(void)
512 return boot_cpu_has(X86_FEATURE_XSAVES);
515 static void __xstate_dump_leaves(void)
518 u32 eax, ebx, ecx, edx;
519 static int should_dump = 1;
525 * Dump out a few leaves past the ones that we support
526 * just in case there are some goodies up there
528 for (i = 0; i < XFEATURE_MAX + 10; i++) {
529 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
530 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
531 XSTATE_CPUID, i, eax, ebx, ecx, edx);
535 #define XSTATE_WARN_ON(x) do { \
536 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
537 __xstate_dump_leaves(); \
541 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
542 if ((nr == nr_macro) && \
543 WARN_ONCE(sz != sizeof(__struct), \
544 "%s: struct is %zu bytes, cpu state %d bytes\n", \
545 __stringify(nr_macro), sizeof(__struct), sz)) { \
546 __xstate_dump_leaves(); \
551 * We have a C struct for each 'xstate'. We need to ensure
552 * that our software representation matches what the CPU
553 * tells us about the state's size.
555 static void check_xstate_against_struct(int nr)
558 * Ask the CPU for the size of the state.
560 int sz = xfeature_size(nr);
562 * Match each CPU state with the corresponding software
565 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct);
566 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state);
567 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state);
568 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state);
569 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
570 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state);
571 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state);
574 * Make *SURE* to add any feature numbers in below if
575 * there are "holes" in the xsave state component
578 if ((nr < XFEATURE_YMM) ||
579 (nr >= XFEATURE_MAX) ||
580 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
581 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
587 * This essentially double-checks what the cpu told us about
588 * how large the XSAVE buffer needs to be. We are recalculating
591 static void do_extra_xstate_size_checks(void)
593 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
596 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
597 if (!xfeature_enabled(i))
600 check_xstate_against_struct(i);
602 * Supervisor state components can be managed only by
603 * XSAVES, which is compacted-format only.
605 if (!using_compacted_format())
606 XSTATE_WARN_ON(xfeature_is_supervisor(i));
608 /* Align from the end of the previous feature */
609 if (xfeature_is_aligned(i))
610 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
612 * The offset of a given state in the non-compacted
613 * format is given to us in a CPUID leaf. We check
614 * them for being ordered (increasing offsets) in
615 * setup_xstate_features().
617 if (!using_compacted_format())
618 paranoid_xstate_size = xfeature_uncompacted_offset(i);
620 * The compacted-format offset always depends on where
621 * the previous state ended.
623 paranoid_xstate_size += xfeature_size(i);
625 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
630 * Get total size of enabled xstates in XCR0/xfeatures_mask.
632 * Note the SDM's wording here. "sub-function 0" only enumerates
633 * the size of the *user* states. If we use it to size a buffer
634 * that we use 'XSAVES' on, we could potentially overflow the
635 * buffer because 'XSAVES' saves system states too.
637 * Note that we do not currently set any bits on IA32_XSS so
638 * 'XCR0 | IA32_XSS == XCR0' for now.
640 static unsigned int __init get_xsaves_size(void)
642 unsigned int eax, ebx, ecx, edx;
644 * - CPUID function 0DH, sub-function 1:
645 * EBX enumerates the size (in bytes) required by
646 * the XSAVES instruction for an XSAVE area
647 * containing all the state components
648 * corresponding to bits currently set in
651 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
655 static unsigned int __init get_xsave_size(void)
657 unsigned int eax, ebx, ecx, edx;
659 * - CPUID function 0DH, sub-function 0:
660 * EBX enumerates the size (in bytes) required by
661 * the XSAVE instruction for an XSAVE area
662 * containing all the *user* state components
663 * corresponding to bits currently set in XCR0.
665 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
670 * Will the runtime-enumerated 'xstate_size' fit in the init
671 * task's statically-allocated buffer?
673 static bool is_supported_xstate_size(unsigned int test_xstate_size)
675 if (test_xstate_size <= sizeof(union fpregs_state))
678 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
679 sizeof(union fpregs_state), test_xstate_size);
683 static int init_xstate_size(void)
685 /* Recompute the context size for enabled features: */
686 unsigned int possible_xstate_size;
687 unsigned int xsave_size;
689 xsave_size = get_xsave_size();
691 if (boot_cpu_has(X86_FEATURE_XSAVES))
692 possible_xstate_size = get_xsaves_size();
694 possible_xstate_size = xsave_size;
696 /* Ensure we have the space to store all enabled: */
697 if (!is_supported_xstate_size(possible_xstate_size))
701 * The size is OK, we are definitely going to use xsave,
702 * make it known to the world that we need more space.
704 fpu_kernel_xstate_size = possible_xstate_size;
705 do_extra_xstate_size_checks();
708 * User space is always in standard format.
710 fpu_user_xstate_size = xsave_size;
715 * We enabled the XSAVE hardware, but something went wrong and
716 * we can not use it. Disable it.
718 static void fpu__init_disable_system_xstate(void)
721 cr4_clear_bits(X86_CR4_OSXSAVE);
722 fpu__xstate_clear_all_cpu_caps();
726 * Enable and initialize the xsave feature.
727 * Called once per system bootup.
729 void __init fpu__init_system_xstate(void)
731 unsigned int eax, ebx, ecx, edx;
732 static int on_boot_cpu __initdata = 1;
735 WARN_ON_FPU(!on_boot_cpu);
738 if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
739 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
743 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
748 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
749 xfeatures_mask = eax + ((u64)edx << 32);
751 if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
753 * This indicates that something really unexpected happened
754 * with the enumeration. Disable XSAVE and try to continue
755 * booting without it. This is too early to BUG().
757 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
761 xfeatures_mask &= fpu__get_supported_xfeatures_mask();
763 /* Enable xstate instructions to be able to continue with initialization: */
764 fpu__init_cpu_xstate();
765 err = init_xstate_size();
770 * Update info used for ptrace frames; use standard-format size and no
771 * supervisor xstates:
773 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
775 fpu__init_prepare_fx_sw_frame();
776 setup_init_fpu_buf();
778 print_xstate_offset_size();
780 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
782 fpu_kernel_xstate_size,
783 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
787 /* something went wrong, try to boot without any XSAVE support */
788 fpu__init_disable_system_xstate();
792 * Restore minimal FPU state after suspend:
794 void fpu__resume_cpu(void)
797 * Restore XCR0 on xsave capable CPUs:
799 if (boot_cpu_has(X86_FEATURE_XSAVE))
800 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
804 * Given an xstate feature mask, calculate where in the xsave
805 * buffer the state is. Callers should ensure that the buffer
808 * Note: does not work for compacted buffers.
810 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
812 int feature_nr = fls64(xstate_feature_mask) - 1;
814 if (!xfeature_enabled(feature_nr)) {
819 return (void *)xsave + xstate_comp_offsets[feature_nr];
822 * Given the xsave area and a state inside, this function returns the
823 * address of the state.
825 * This is the API that is called to get xstate address in either
826 * standard format or compacted format of xsave area.
828 * Note that if there is no data for the field in the xsave buffer
829 * this will return NULL.
832 * xstate: the thread's storage area for all FPU data
833 * xstate_feature: state which is defined in xsave.h (e.g.
834 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
836 * address of the state in the xsave area, or NULL if the
837 * field is not present in the xsave buffer.
839 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
842 * Do we even *have* xsave state?
844 if (!boot_cpu_has(X86_FEATURE_XSAVE))
848 * We should not ever be requesting features that we
849 * have not enabled. Remember that pcntxt_mask is
850 * what we write to the XCR0 register.
852 WARN_ONCE(!(xfeatures_mask & xstate_feature),
853 "get of unsupported state");
855 * This assumes the last 'xsave*' instruction to
856 * have requested that 'xstate_feature' be saved.
857 * If it did not, we might be seeing and old value
858 * of the field in the buffer.
860 * This can happen because the last 'xsave' did not
861 * request that this feature be saved (unlikely)
862 * or because the "init optimization" caused it
865 if (!(xsave->header.xfeatures & xstate_feature))
868 return __raw_xsave_addr(xsave, xstate_feature);
870 EXPORT_SYMBOL_GPL(get_xsave_addr);
873 * This wraps up the common operations that need to occur when retrieving
874 * data from xsave state. It first ensures that the current task was
875 * using the FPU and retrieves the data in to a buffer. It then calculates
876 * the offset of the requested field in the buffer.
878 * This function is safe to call whether the FPU is in use or not.
880 * Note that this only works on the current task.
883 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
884 * XFEATURE_MASK_SSE, etc...)
886 * address of the state in the xsave area or NULL if the state
887 * is not present or is in its 'init state'.
889 const void *get_xsave_field_ptr(int xsave_state)
891 struct fpu *fpu = ¤t->thread.fpu;
893 if (!fpu->fpstate_active)
896 * fpu__save() takes the CPU's xstate registers
897 * and saves them off to the 'fpu memory buffer.
901 return get_xsave_addr(&fpu->state.xsave, xsave_state);
904 #ifdef CONFIG_ARCH_HAS_PKEYS
907 * This will go out and modify PKRU register to set the access
908 * rights for @pkey to @init_val.
910 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
911 unsigned long init_val)
914 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
915 u32 new_pkru_bits = 0;
918 * This check implies XSAVE support. OSPKE only gets
919 * set if we enable XSAVE and we enable PKU in XCR0.
921 if (!boot_cpu_has(X86_FEATURE_OSPKE))
925 * This code should only be called with valid 'pkey'
926 * values originating from in-kernel users. Complain
927 * if a bad value is observed.
929 WARN_ON_ONCE(pkey >= arch_max_pkey());
931 /* Set the bits we need in PKRU: */
932 if (init_val & PKEY_DISABLE_ACCESS)
933 new_pkru_bits |= PKRU_AD_BIT;
934 if (init_val & PKEY_DISABLE_WRITE)
935 new_pkru_bits |= PKRU_WD_BIT;
937 /* Shift the bits in to the correct place in PKRU for pkey: */
938 new_pkru_bits <<= pkey_shift;
940 /* Get old PKRU and mask off any old bits in place: */
941 old_pkru = read_pkru();
942 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
944 /* Write old part along with new part: */
945 write_pkru(old_pkru | new_pkru_bits);
949 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
952 * This is similar to user_regset_copyout(), but will not add offset to
953 * the source data pointer or increment pos, count, kbuf, and ubuf.
955 static inline int xstate_copyout(unsigned int pos, unsigned int count,
956 void *kbuf, void __user *ubuf,
957 const void *data, const int start_pos,
960 if ((count == 0) || (pos < start_pos))
963 if (end_pos < 0 || pos < end_pos) {
964 unsigned int copy = (end_pos < 0 ? count : min(count, end_pos - pos));
967 memcpy(kbuf + pos, data, copy);
969 if (__copy_to_user(ubuf + pos, data, copy))
977 * Convert from kernel XSAVES compacted format to standard format and copy
978 * to a ptrace buffer. It supports partial copy but pos always starts from
979 * zero. This is called from xstateregs_get() and there we check the CPU
982 int copyout_from_xsaves(unsigned int pos, unsigned int count, void *kbuf,
983 void __user *ubuf, struct xregs_state *xsave)
985 unsigned int offset, size;
987 struct xstate_header header;
990 * Currently copy_regset_to_user() starts from pos 0:
992 if (unlikely(pos != 0))
996 * The destination is a ptrace buffer; we put in only user xstates:
998 memset(&header, 0, sizeof(header));
999 header.xfeatures = xsave->header.xfeatures;
1000 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1003 * Copy xregs_state->header:
1005 offset = offsetof(struct xregs_state, header);
1006 size = sizeof(header);
1008 ret = xstate_copyout(offset, size, kbuf, ubuf, &header, 0, count);
1013 for (i = 0; i < XFEATURE_MAX; i++) {
1015 * Copy only in-use xstates:
1017 if ((header.xfeatures >> i) & 1) {
1018 void *src = __raw_xsave_addr(xsave, 1 << i);
1020 offset = xstate_offsets[i];
1021 size = xstate_sizes[i];
1023 ret = xstate_copyout(offset, size, kbuf, ubuf, src, 0, count);
1028 if (offset + size >= count)
1035 * Fill xsave->i387.sw_reserved value for ptrace frame:
1037 offset = offsetof(struct fxregs_state, sw_reserved);
1038 size = sizeof(xstate_fx_sw_bytes);
1040 ret = xstate_copyout(offset, size, kbuf, ubuf, xstate_fx_sw_bytes, 0, count);
1049 * Convert from a ptrace standard-format buffer to kernel XSAVES format
1050 * and copy to the target thread. This is called from xstateregs_set() and
1051 * there we check the CPU has XSAVES and a whole standard-sized buffer
1054 int copyin_to_xsaves(const void *kbuf, const void __user *ubuf,
1055 struct xregs_state *xsave)
1057 unsigned int offset, size;
1060 u64 allowed_features;
1062 offset = offsetof(struct xregs_state, header);
1063 size = sizeof(xfeatures);
1066 memcpy(&xfeatures, kbuf + offset, size);
1068 if (__copy_from_user(&xfeatures, ubuf + offset, size))
1073 * Reject if the user sets any disabled or supervisor features:
1075 allowed_features = xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR;
1077 if (xfeatures & ~allowed_features)
1080 for (i = 0; i < XFEATURE_MAX; i++) {
1081 u64 mask = ((u64)1 << i);
1083 if (xfeatures & mask) {
1084 void *dst = __raw_xsave_addr(xsave, 1 << i);
1086 offset = xstate_offsets[i];
1087 size = xstate_sizes[i];
1090 memcpy(dst, kbuf + offset, size);
1092 if (__copy_from_user(dst, ubuf + offset, size))
1099 * The state that came in from userspace was user-state only.
1100 * Mask all the user states out of 'xfeatures':
1102 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1105 * Add back in the features that came in from userspace:
1107 xsave->header.xfeatures |= xfeatures;
1108 xsave->header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT | xsave->header.xfeatures;