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>
18 #include <asm/cpufeature.h>
21 * Although we spell it out in here, the Processor Trace
22 * xfeature is completely unused. We use other mechanisms
23 * to save/restore PT state in Linux.
25 static const char *xfeature_names[] =
27 "x87 floating point registers" ,
30 "MPX bounds registers" ,
35 "Processor Trace (unused)" ,
36 "Protection Keys User registers",
37 "unknown xstate feature" ,
40 static short xsave_cpuid_features[] __initdata = {
54 * Mask of xstate features supported by the CPU and the kernel:
56 u64 xfeatures_mask __read_mostly;
58 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
59 static unsigned int xstate_sizes[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
60 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
63 * The XSAVE area of kernel can be in standard or compacted format;
64 * it is always in standard format for user mode. This is the user
65 * mode standard format size used for signal and ptrace frames.
67 unsigned int fpu_user_xstate_size;
70 * Clear all of the X86_FEATURE_* bits that are unavailable
71 * when the CPU has no XSAVE support.
73 void fpu__xstate_clear_all_cpu_caps(void)
75 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
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 * All supported features have either init state all zeros or are
409 * handled in setup_init_fpu() individually. This is an explicit
410 * feature list and does not use XFEATURE_MASK*SUPPORTED to catch
411 * newly added supported features at build time and make people
412 * actually look at the init state for the new feature.
414 #define XFEATURES_INIT_FPSTATE_HANDLED \
415 (XFEATURE_MASK_FP | \
416 XFEATURE_MASK_SSE | \
417 XFEATURE_MASK_YMM | \
418 XFEATURE_MASK_OPMASK | \
419 XFEATURE_MASK_ZMM_Hi256 | \
420 XFEATURE_MASK_Hi16_ZMM | \
421 XFEATURE_MASK_PKRU | \
422 XFEATURE_MASK_BNDREGS | \
423 XFEATURE_MASK_BNDCSR)
426 * setup the xstate image representing the init state
428 static void __init setup_init_fpu_buf(void)
430 static int on_boot_cpu __initdata = 1;
432 BUILD_BUG_ON(XCNTXT_MASK != XFEATURES_INIT_FPSTATE_HANDLED);
434 WARN_ON_FPU(!on_boot_cpu);
437 if (!boot_cpu_has(X86_FEATURE_XSAVE))
440 setup_xstate_features();
441 print_xstate_features();
443 if (boot_cpu_has(X86_FEATURE_XSAVES))
444 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
447 * Init all the features state with header.xfeatures being 0x0
449 copy_kernel_to_xregs_booting(&init_fpstate.xsave);
452 * All components are now in init state. Read the state back so
453 * that init_fpstate contains all non-zero init state. This only
454 * works with XSAVE, but not with XSAVEOPT and XSAVES because
455 * those use the init optimization which skips writing data for
456 * components in init state.
458 * XSAVE could be used, but that would require to reshuffle the
459 * data when XSAVES is available because XSAVES uses xstate
460 * compaction. But doing so is a pointless exercise because most
461 * components have an all zeros init state except for the legacy
462 * ones (FP and SSE). Those can be saved with FXSAVE into the
463 * legacy area. Adding new features requires to ensure that init
464 * state is all zeroes or if not to add the necessary handling
467 fxsave(&init_fpstate.fxsave);
470 static int xfeature_uncompacted_offset(int xfeature_nr)
472 u32 eax, ebx, ecx, edx;
475 * Only XSAVES supports supervisor states and it uses compacted
476 * format. Checking a supervisor state's uncompacted offset is
479 if (XFEATURE_MASK_SUPERVISOR & (1 << xfeature_nr)) {
480 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
484 CHECK_XFEATURE(xfeature_nr);
485 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
489 static int xfeature_size(int xfeature_nr)
491 u32 eax, ebx, ecx, edx;
493 CHECK_XFEATURE(xfeature_nr);
494 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
499 * 'XSAVES' implies two different things:
500 * 1. saving of supervisor/system state
501 * 2. using the compacted format
503 * Use this function when dealing with the compacted format so
504 * that it is obvious which aspect of 'XSAVES' is being handled
505 * by the calling code.
507 int using_compacted_format(void)
509 return boot_cpu_has(X86_FEATURE_XSAVES);
512 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
513 int validate_xstate_header(const struct xstate_header *hdr)
515 /* No unknown or supervisor features may be set */
516 if (hdr->xfeatures & (~xfeatures_mask | XFEATURE_MASK_SUPERVISOR))
519 /* Userspace must use the uncompacted format */
524 * If 'reserved' is shrunken to add a new field, make sure to validate
525 * that new field here!
527 BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
529 /* No reserved bits may be set */
530 if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
536 static void __xstate_dump_leaves(void)
539 u32 eax, ebx, ecx, edx;
540 static int should_dump = 1;
546 * Dump out a few leaves past the ones that we support
547 * just in case there are some goodies up there
549 for (i = 0; i < XFEATURE_MAX + 10; i++) {
550 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
551 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
552 XSTATE_CPUID, i, eax, ebx, ecx, edx);
556 #define XSTATE_WARN_ON(x) do { \
557 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
558 __xstate_dump_leaves(); \
562 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
563 if ((nr == nr_macro) && \
564 WARN_ONCE(sz != sizeof(__struct), \
565 "%s: struct is %zu bytes, cpu state %d bytes\n", \
566 __stringify(nr_macro), sizeof(__struct), sz)) { \
567 __xstate_dump_leaves(); \
572 * We have a C struct for each 'xstate'. We need to ensure
573 * that our software representation matches what the CPU
574 * tells us about the state's size.
576 static void check_xstate_against_struct(int nr)
579 * Ask the CPU for the size of the state.
581 int sz = xfeature_size(nr);
583 * Match each CPU state with the corresponding software
586 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct);
587 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state);
588 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state);
589 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state);
590 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
591 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state);
592 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state);
595 * Make *SURE* to add any feature numbers in below if
596 * there are "holes" in the xsave state component
599 if ((nr < XFEATURE_YMM) ||
600 (nr >= XFEATURE_MAX) ||
601 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
602 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
608 * This essentially double-checks what the cpu told us about
609 * how large the XSAVE buffer needs to be. We are recalculating
612 static void do_extra_xstate_size_checks(void)
614 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
617 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
618 if (!xfeature_enabled(i))
621 check_xstate_against_struct(i);
623 * Supervisor state components can be managed only by
624 * XSAVES, which is compacted-format only.
626 if (!using_compacted_format())
627 XSTATE_WARN_ON(xfeature_is_supervisor(i));
629 /* Align from the end of the previous feature */
630 if (xfeature_is_aligned(i))
631 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
633 * The offset of a given state in the non-compacted
634 * format is given to us in a CPUID leaf. We check
635 * them for being ordered (increasing offsets) in
636 * setup_xstate_features().
638 if (!using_compacted_format())
639 paranoid_xstate_size = xfeature_uncompacted_offset(i);
641 * The compacted-format offset always depends on where
642 * the previous state ended.
644 paranoid_xstate_size += xfeature_size(i);
646 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
651 * Get total size of enabled xstates in XCR0/xfeatures_mask.
653 * Note the SDM's wording here. "sub-function 0" only enumerates
654 * the size of the *user* states. If we use it to size a buffer
655 * that we use 'XSAVES' on, we could potentially overflow the
656 * buffer because 'XSAVES' saves system states too.
658 * Note that we do not currently set any bits on IA32_XSS so
659 * 'XCR0 | IA32_XSS == XCR0' for now.
661 static unsigned int __init get_xsaves_size(void)
663 unsigned int eax, ebx, ecx, edx;
665 * - CPUID function 0DH, sub-function 1:
666 * EBX enumerates the size (in bytes) required by
667 * the XSAVES instruction for an XSAVE area
668 * containing all the state components
669 * corresponding to bits currently set in
672 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
676 static unsigned int __init get_xsave_size(void)
678 unsigned int eax, ebx, ecx, edx;
680 * - CPUID function 0DH, sub-function 0:
681 * EBX enumerates the size (in bytes) required by
682 * the XSAVE instruction for an XSAVE area
683 * containing all the *user* state components
684 * corresponding to bits currently set in XCR0.
686 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
691 * Will the runtime-enumerated 'xstate_size' fit in the init
692 * task's statically-allocated buffer?
694 static bool is_supported_xstate_size(unsigned int test_xstate_size)
696 if (test_xstate_size <= sizeof(union fpregs_state))
699 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
700 sizeof(union fpregs_state), test_xstate_size);
704 static int init_xstate_size(void)
706 /* Recompute the context size for enabled features: */
707 unsigned int possible_xstate_size;
708 unsigned int xsave_size;
710 xsave_size = get_xsave_size();
712 if (boot_cpu_has(X86_FEATURE_XSAVES))
713 possible_xstate_size = get_xsaves_size();
715 possible_xstate_size = xsave_size;
717 /* Ensure we have the space to store all enabled: */
718 if (!is_supported_xstate_size(possible_xstate_size))
722 * The size is OK, we are definitely going to use xsave,
723 * make it known to the world that we need more space.
725 fpu_kernel_xstate_size = possible_xstate_size;
726 do_extra_xstate_size_checks();
729 * User space is always in standard format.
731 fpu_user_xstate_size = xsave_size;
736 * We enabled the XSAVE hardware, but something went wrong and
737 * we can not use it. Disable it.
739 static void fpu__init_disable_system_xstate(void)
742 cr4_clear_bits(X86_CR4_OSXSAVE);
743 fpu__xstate_clear_all_cpu_caps();
747 * Enable and initialize the xsave feature.
748 * Called once per system bootup.
750 void __init fpu__init_system_xstate(void)
752 unsigned int eax, ebx, ecx, edx;
753 static int on_boot_cpu __initdata = 1;
757 WARN_ON_FPU(!on_boot_cpu);
760 if (!boot_cpu_has(X86_FEATURE_FPU)) {
761 pr_info("x86/fpu: No FPU detected\n");
765 if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
766 pr_info("x86/fpu: x87 FPU will use %s\n",
767 boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
771 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
776 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
777 xfeatures_mask = eax + ((u64)edx << 32);
779 if ((xfeatures_mask & 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", xfeatures_mask);
790 * Clear XSAVE features that are disabled in the normal CPUID.
792 for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
793 if (!boot_cpu_has(xsave_cpuid_features[i]))
794 xfeatures_mask &= ~BIT(i);
797 xfeatures_mask &= fpu__get_supported_xfeatures_mask();
799 /* Enable xstate instructions to be able to continue with initialization: */
800 fpu__init_cpu_xstate();
801 err = init_xstate_size();
806 * Update info used for ptrace frames; use standard-format size and no
807 * supervisor xstates:
809 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
811 fpu__init_prepare_fx_sw_frame();
812 setup_init_fpu_buf();
816 * CPU capabilities initialization runs before FPU init. So
817 * X86_FEATURE_OSXSAVE is not set. Now that XSAVE is completely
818 * functional, set the feature bit so depending code works.
820 setup_force_cpu_cap(X86_FEATURE_OSXSAVE);
822 print_xstate_offset_size();
824 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
826 fpu_kernel_xstate_size,
827 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
831 /* something went wrong, try to boot without any XSAVE support */
832 fpu__init_disable_system_xstate();
836 * Restore minimal FPU state after suspend:
838 void fpu__resume_cpu(void)
841 * Restore XCR0 on xsave capable CPUs:
843 if (boot_cpu_has(X86_FEATURE_XSAVE))
844 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
848 * Given an xstate feature mask, calculate where in the xsave
849 * buffer the state is. Callers should ensure that the buffer
852 * Note: does not work for compacted buffers.
854 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
856 int feature_nr = fls64(xstate_feature_mask) - 1;
858 if (!xfeature_enabled(feature_nr)) {
863 return (void *)xsave + xstate_comp_offsets[feature_nr];
866 * Given the xsave area and a state inside, this function returns the
867 * address of the state.
869 * This is the API that is called to get xstate address in either
870 * standard format or compacted format of xsave area.
872 * Note that if there is no data for the field in the xsave buffer
873 * this will return NULL.
876 * xstate: the thread's storage area for all FPU data
877 * xstate_feature: state which is defined in xsave.h (e.g.
878 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
880 * address of the state in the xsave area, or NULL if the
881 * field is not present in the xsave buffer.
883 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
886 * Do we even *have* xsave state?
888 if (!boot_cpu_has(X86_FEATURE_XSAVE))
892 * We should not ever be requesting features that we
893 * have not enabled. Remember that pcntxt_mask is
894 * what we write to the XCR0 register.
896 WARN_ONCE(!(xfeatures_mask & xstate_feature),
897 "get of unsupported state");
899 * This assumes the last 'xsave*' instruction to
900 * have requested that 'xstate_feature' be saved.
901 * If it did not, we might be seeing and old value
902 * of the field in the buffer.
904 * This can happen because the last 'xsave' did not
905 * request that this feature be saved (unlikely)
906 * or because the "init optimization" caused it
909 if (!(xsave->header.xfeatures & xstate_feature))
912 return __raw_xsave_addr(xsave, xstate_feature);
914 EXPORT_SYMBOL_GPL(get_xsave_addr);
917 * This wraps up the common operations that need to occur when retrieving
918 * data from xsave state. It first ensures that the current task was
919 * using the FPU and retrieves the data in to a buffer. It then calculates
920 * the offset of the requested field in the buffer.
922 * This function is safe to call whether the FPU is in use or not.
924 * Note that this only works on the current task.
927 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
928 * XFEATURE_MASK_SSE, etc...)
930 * address of the state in the xsave area or NULL if the state
931 * is not present or is in its 'init state'.
933 const void *get_xsave_field_ptr(int xsave_state)
935 struct fpu *fpu = ¤t->thread.fpu;
937 if (!fpu->initialized)
940 * fpu__save() takes the CPU's xstate registers
941 * and saves them off to the 'fpu memory buffer.
945 return get_xsave_addr(&fpu->state.xsave, xsave_state);
948 #ifdef CONFIG_ARCH_HAS_PKEYS
951 * This will go out and modify PKRU register to set the access
952 * rights for @pkey to @init_val.
954 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
955 unsigned long init_val)
958 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
959 u32 new_pkru_bits = 0;
962 * This check implies XSAVE support. OSPKE only gets
963 * set if we enable XSAVE and we enable PKU in XCR0.
965 if (!boot_cpu_has(X86_FEATURE_OSPKE))
969 * This code should only be called with valid 'pkey'
970 * values originating from in-kernel users. Complain
971 * if a bad value is observed.
973 WARN_ON_ONCE(pkey >= arch_max_pkey());
975 /* Set the bits we need in PKRU: */
976 if (init_val & PKEY_DISABLE_ACCESS)
977 new_pkru_bits |= PKRU_AD_BIT;
978 if (init_val & PKEY_DISABLE_WRITE)
979 new_pkru_bits |= PKRU_WD_BIT;
981 /* Shift the bits in to the correct place in PKRU for pkey: */
982 new_pkru_bits <<= pkey_shift;
984 /* Get old PKRU and mask off any old bits in place: */
985 old_pkru = read_pkru();
986 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
988 /* Write old part along with new part: */
989 write_pkru(old_pkru | new_pkru_bits);
993 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
996 * Weird legacy quirk: SSE and YMM states store information in the
997 * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
998 * area is marked as unused in the xfeatures header, we need to copy
999 * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
1001 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
1003 if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
1006 if (xfeatures & XFEATURE_MASK_FP)
1012 static void fill_gap(unsigned to, void **kbuf, unsigned *pos, unsigned *count)
1015 unsigned size = to - *pos;
1019 memcpy(*kbuf, (void *)&init_fpstate.xsave + *pos, size);
1026 static void copy_part(unsigned offset, unsigned size, void *from,
1027 void **kbuf, unsigned *pos, unsigned *count)
1029 fill_gap(offset, kbuf, pos, count);
1033 memcpy(*kbuf, from, size);
1041 * Convert from kernel XSAVES compacted format to standard format and copy
1042 * to a kernel-space ptrace buffer.
1044 * It supports partial copy but pos always starts from zero. This is called
1045 * from xstateregs_get() and there we check the CPU has XSAVES.
1047 int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1049 struct xstate_header header;
1050 const unsigned off_mxcsr = offsetof(struct fxregs_state, mxcsr);
1051 unsigned count = size_total;
1055 * Currently copy_regset_to_user() starts from pos 0:
1057 if (unlikely(offset_start != 0))
1061 * The destination is a ptrace buffer; we put in only user xstates:
1063 memset(&header, 0, sizeof(header));
1064 header.xfeatures = xsave->header.xfeatures;
1065 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1067 if (header.xfeatures & XFEATURE_MASK_FP)
1068 copy_part(0, off_mxcsr,
1069 &xsave->i387, &kbuf, &offset_start, &count);
1070 if (header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM))
1071 copy_part(off_mxcsr, MXCSR_AND_FLAGS_SIZE,
1072 &xsave->i387.mxcsr, &kbuf, &offset_start, &count);
1073 if (header.xfeatures & XFEATURE_MASK_FP)
1074 copy_part(offsetof(struct fxregs_state, st_space), 128,
1075 &xsave->i387.st_space, &kbuf, &offset_start, &count);
1076 if (header.xfeatures & XFEATURE_MASK_SSE)
1077 copy_part(xstate_offsets[XFEATURE_SSE], 256,
1078 &xsave->i387.xmm_space, &kbuf, &offset_start, &count);
1080 * Fill xsave->i387.sw_reserved value for ptrace frame:
1082 copy_part(offsetof(struct fxregs_state, sw_reserved), 48,
1083 xstate_fx_sw_bytes, &kbuf, &offset_start, &count);
1085 * Copy xregs_state->header:
1087 copy_part(offsetof(struct xregs_state, header), sizeof(header),
1088 &header, &kbuf, &offset_start, &count);
1090 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
1092 * Copy only in-use xstates:
1094 if ((header.xfeatures >> i) & 1) {
1095 void *src = __raw_xsave_addr(xsave, 1 << i);
1097 copy_part(xstate_offsets[i], xstate_sizes[i],
1098 src, &kbuf, &offset_start, &count);
1102 fill_gap(size_total, &kbuf, &offset_start, &count);
1108 __copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total)
1113 if (offset < size_total) {
1114 unsigned int copy = min(size, size_total - offset);
1116 if (__copy_to_user(ubuf + offset, data, copy))
1123 * Convert from kernel XSAVES compacted format to standard format and copy
1124 * to a user-space buffer. It supports partial copy but pos always starts from
1125 * zero. This is called from xstateregs_get() and there we check the CPU
1128 int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1130 unsigned int offset, size;
1132 struct xstate_header header;
1135 * Currently copy_regset_to_user() starts from pos 0:
1137 if (unlikely(offset_start != 0))
1141 * The destination is a ptrace buffer; we put in only user xstates:
1143 memset(&header, 0, sizeof(header));
1144 header.xfeatures = xsave->header.xfeatures;
1145 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1148 * Copy xregs_state->header:
1150 offset = offsetof(struct xregs_state, header);
1151 size = sizeof(header);
1153 ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total);
1157 for (i = 0; i < XFEATURE_MAX; i++) {
1159 * Copy only in-use xstates:
1161 if ((header.xfeatures >> i) & 1) {
1162 void *src = __raw_xsave_addr(xsave, 1 << i);
1164 offset = xstate_offsets[i];
1165 size = xstate_sizes[i];
1167 /* The next component has to fit fully into the output buffer: */
1168 if (offset + size > size_total)
1171 ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total);
1178 if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1179 offset = offsetof(struct fxregs_state, mxcsr);
1180 size = MXCSR_AND_FLAGS_SIZE;
1181 __copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total);
1185 * Fill xsave->i387.sw_reserved value for ptrace frame:
1187 offset = offsetof(struct fxregs_state, sw_reserved);
1188 size = sizeof(xstate_fx_sw_bytes);
1190 ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total);
1198 * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1199 * and copy to the target thread. This is called from xstateregs_set().
1201 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1203 unsigned int offset, size;
1205 struct xstate_header hdr;
1207 offset = offsetof(struct xregs_state, header);
1210 memcpy(&hdr, kbuf + offset, size);
1212 if (validate_xstate_header(&hdr))
1215 for (i = 0; i < XFEATURE_MAX; i++) {
1216 u64 mask = ((u64)1 << i);
1218 if (hdr.xfeatures & mask) {
1219 void *dst = __raw_xsave_addr(xsave, 1 << i);
1221 offset = xstate_offsets[i];
1222 size = xstate_sizes[i];
1224 memcpy(dst, kbuf + offset, size);
1228 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1229 offset = offsetof(struct fxregs_state, mxcsr);
1230 size = MXCSR_AND_FLAGS_SIZE;
1231 memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1235 * The state that came in from userspace was user-state only.
1236 * Mask all the user states out of 'xfeatures':
1238 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1241 * Add back in the features that came in from userspace:
1243 xsave->header.xfeatures |= hdr.xfeatures;
1249 * Convert from a ptrace or sigreturn standard-format user-space buffer to
1250 * kernel XSAVES format and copy to the target thread. This is called from
1251 * xstateregs_set(), as well as potentially from the sigreturn() and
1252 * rt_sigreturn() system calls.
1254 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1256 unsigned int offset, size;
1258 struct xstate_header hdr;
1260 offset = offsetof(struct xregs_state, header);
1263 if (__copy_from_user(&hdr, ubuf + offset, size))
1266 if (validate_xstate_header(&hdr))
1269 for (i = 0; i < XFEATURE_MAX; i++) {
1270 u64 mask = ((u64)1 << i);
1272 if (hdr.xfeatures & mask) {
1273 void *dst = __raw_xsave_addr(xsave, 1 << i);
1275 offset = xstate_offsets[i];
1276 size = xstate_sizes[i];
1278 if (__copy_from_user(dst, ubuf + offset, size))
1283 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1284 offset = offsetof(struct fxregs_state, mxcsr);
1285 size = MXCSR_AND_FLAGS_SIZE;
1286 if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1291 * The state that came in from userspace was user-state only.
1292 * Mask all the user states out of 'xfeatures':
1294 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1297 * Add back in the features that came in from userspace:
1299 xsave->header.xfeatures |= hdr.xfeatures;