1 // SPDX-License-Identifier: GPL-2.0-only
3 * xsave/xrstor support.
5 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
7 #include <linux/compat.h>
9 #include <linux/mman.h>
10 #include <linux/pkeys.h>
11 #include <linux/seq_file.h>
12 #include <linux/proc_fs.h>
14 #include <asm/fpu/api.h>
15 #include <asm/fpu/internal.h>
16 #include <asm/fpu/signal.h>
17 #include <asm/fpu/regset.h>
18 #include <asm/fpu/xstate.h>
20 #include <asm/tlbflush.h>
21 #include <asm/cpufeature.h>
24 * Although we spell it out in here, the Processor Trace
25 * xfeature is completely unused. We use other mechanisms
26 * to save/restore PT state in Linux.
28 static const char *xfeature_names[] =
30 "x87 floating point registers" ,
33 "MPX bounds registers" ,
38 "Processor Trace (unused)" ,
39 "Protection Keys User registers",
40 "unknown xstate feature" ,
43 static short xsave_cpuid_features[] __initdata = {
57 * Mask of xstate features supported by the CPU and the kernel:
59 u64 xfeatures_mask __read_mostly;
61 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
62 static unsigned int xstate_sizes[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
63 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
66 * The XSAVE area of kernel can be in standard or compacted format;
67 * it is always in standard format for user mode. This is the user
68 * mode standard format size used for signal and ptrace frames.
70 unsigned int fpu_user_xstate_size;
73 * Return whether the system supports a given xfeature.
75 * Also return the name of the (most advanced) feature that the caller requested:
77 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
79 u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
81 if (unlikely(feature_name)) {
82 long xfeature_idx, max_idx;
85 * So we use FLS here to be able to print the most advanced
86 * feature that was requested but is missing. So if a driver
87 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
88 * missing AVX feature - this is the most informative message
91 if (xfeatures_missing)
92 xfeatures_print = xfeatures_missing;
94 xfeatures_print = xfeatures_needed;
96 xfeature_idx = fls64(xfeatures_print)-1;
97 max_idx = ARRAY_SIZE(xfeature_names)-1;
98 xfeature_idx = min(xfeature_idx, max_idx);
100 *feature_name = xfeature_names[xfeature_idx];
103 if (xfeatures_missing)
108 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
110 static int xfeature_is_supervisor(int xfeature_nr)
113 * We currently do not support supervisor states, but if
114 * we did, we could find out like this.
116 * SDM says: If state component 'i' is a user state component,
117 * ECX[0] return 0; if state component i is a supervisor
118 * state component, ECX[0] returns 1.
120 u32 eax, ebx, ecx, edx;
122 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
126 static int xfeature_is_user(int xfeature_nr)
128 return !xfeature_is_supervisor(xfeature_nr);
132 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
133 * a processor implementation detects that an FPU state component is still
134 * (or is again) in its initialized state, it may clear the corresponding
135 * bit in the header.xfeatures field, and can skip the writeout of registers
136 * to the corresponding memory layout.
138 * This means that when the bit is zero, the state component might still contain
139 * some previous - non-initialized register state.
141 * Before writing xstate information to user-space we sanitize those components,
142 * to always ensure that the memory layout of a feature will be in the init state
143 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
144 * see some stale state in the memory layout during signal handling, debugging etc.
146 void fpstate_sanitize_xstate(struct fpu *fpu)
148 struct fxregs_state *fx = &fpu->state.fxsave;
155 xfeatures = fpu->state.xsave.header.xfeatures;
158 * None of the feature bits are in init state. So nothing else
159 * to do for us, as the memory layout is up to date.
161 if ((xfeatures & xfeatures_mask) == xfeatures_mask)
165 * FP is in init state
167 if (!(xfeatures & XFEATURE_MASK_FP)) {
174 memset(&fx->st_space[0], 0, 128);
178 * SSE is in init state
180 if (!(xfeatures & XFEATURE_MASK_SSE))
181 memset(&fx->xmm_space[0], 0, 256);
184 * First two features are FPU and SSE, which above we handled
185 * in a special way already:
188 xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
191 * Update all the remaining memory layouts according to their
192 * standard xstate layout, if their header bit is in the init
196 if (xfeatures & 0x1) {
197 int offset = xstate_comp_offsets[feature_bit];
198 int size = xstate_sizes[feature_bit];
200 memcpy((void *)fx + offset,
201 (void *)&init_fpstate.xsave + offset,
211 * Enable the extended processor state save/restore feature.
212 * Called once per CPU onlining.
214 void fpu__init_cpu_xstate(void)
216 if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
219 * Make it clear that XSAVES supervisor states are not yet
220 * implemented should anyone expect it to work by changing
221 * bits in XFEATURE_MASK_* macros and XCR0.
223 WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
224 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
226 xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
228 cr4_set_bits(X86_CR4_OSXSAVE);
229 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
233 * Note that in the future we will likely need a pair of
234 * functions here: one for user xstates and the other for
235 * system xstates. For now, they are the same.
237 static int xfeature_enabled(enum xfeature xfeature)
239 return !!(xfeatures_mask & (1UL << xfeature));
243 * Record the offsets and sizes of various xstates contained
244 * in the XSAVE state memory layout.
246 static void __init setup_xstate_features(void)
248 u32 eax, ebx, ecx, edx, i;
249 /* start at the beginnning of the "extended state" */
250 unsigned int last_good_offset = offsetof(struct xregs_state,
251 extended_state_area);
253 * The FP xstates and SSE xstates are legacy states. They are always
254 * in the fixed offsets in the xsave area in either compacted form
257 xstate_offsets[0] = 0;
258 xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space);
259 xstate_offsets[1] = xstate_sizes[0];
260 xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space);
262 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
263 if (!xfeature_enabled(i))
266 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
269 * If an xfeature is supervisor state, the offset
270 * in EBX is invalid. We leave it to -1.
272 if (xfeature_is_user(i))
273 xstate_offsets[i] = ebx;
275 xstate_sizes[i] = eax;
277 * In our xstate size checks, we assume that the
278 * highest-numbered xstate feature has the
279 * highest offset in the buffer. Ensure it does.
281 WARN_ONCE(last_good_offset > xstate_offsets[i],
282 "x86/fpu: misordered xstate at %d\n", last_good_offset);
283 last_good_offset = xstate_offsets[i];
287 static void __init print_xstate_feature(u64 xstate_mask)
289 const char *feature_name;
291 if (cpu_has_xfeatures(xstate_mask, &feature_name))
292 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
296 * Print out all the supported xstate features:
298 static void __init print_xstate_features(void)
300 print_xstate_feature(XFEATURE_MASK_FP);
301 print_xstate_feature(XFEATURE_MASK_SSE);
302 print_xstate_feature(XFEATURE_MASK_YMM);
303 print_xstate_feature(XFEATURE_MASK_BNDREGS);
304 print_xstate_feature(XFEATURE_MASK_BNDCSR);
305 print_xstate_feature(XFEATURE_MASK_OPMASK);
306 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
307 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
308 print_xstate_feature(XFEATURE_MASK_PKRU);
312 * This check is important because it is easy to get XSTATE_*
313 * confused with XSTATE_BIT_*.
315 #define CHECK_XFEATURE(nr) do { \
316 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
317 WARN_ON(nr >= XFEATURE_MAX); \
321 * We could cache this like xstate_size[], but we only use
322 * it here, so it would be a waste of space.
324 static int xfeature_is_aligned(int xfeature_nr)
326 u32 eax, ebx, ecx, edx;
328 CHECK_XFEATURE(xfeature_nr);
329 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
331 * The value returned by ECX[1] indicates the alignment
332 * of state component 'i' when the compacted format
333 * of the extended region of an XSAVE area is used:
339 * This function sets up offsets and sizes of all extended states in
340 * xsave area. This supports both standard format and compacted format
341 * of the xsave aread.
343 static void __init setup_xstate_comp(void)
345 unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
349 * The FP xstates and SSE xstates are legacy states. They are always
350 * in the fixed offsets in the xsave area in either compacted form
353 xstate_comp_offsets[0] = 0;
354 xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
356 if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
357 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
358 if (xfeature_enabled(i)) {
359 xstate_comp_offsets[i] = xstate_offsets[i];
360 xstate_comp_sizes[i] = xstate_sizes[i];
366 xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
367 FXSAVE_SIZE + XSAVE_HDR_SIZE;
369 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
370 if (xfeature_enabled(i))
371 xstate_comp_sizes[i] = xstate_sizes[i];
373 xstate_comp_sizes[i] = 0;
375 if (i > FIRST_EXTENDED_XFEATURE) {
376 xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
377 + xstate_comp_sizes[i-1];
379 if (xfeature_is_aligned(i))
380 xstate_comp_offsets[i] =
381 ALIGN(xstate_comp_offsets[i], 64);
387 * Print out xstate component offsets and sizes
389 static void __init print_xstate_offset_size(void)
393 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
394 if (!xfeature_enabled(i))
396 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
397 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
402 * All supported features have either init state all zeros or are
403 * handled in setup_init_fpu() individually. This is an explicit
404 * feature list and does not use XFEATURE_MASK*SUPPORTED to catch
405 * newly added supported features at build time and make people
406 * actually look at the init state for the new feature.
408 #define XFEATURES_INIT_FPSTATE_HANDLED \
409 (XFEATURE_MASK_FP | \
410 XFEATURE_MASK_SSE | \
411 XFEATURE_MASK_YMM | \
412 XFEATURE_MASK_OPMASK | \
413 XFEATURE_MASK_ZMM_Hi256 | \
414 XFEATURE_MASK_Hi16_ZMM | \
415 XFEATURE_MASK_PKRU | \
416 XFEATURE_MASK_BNDREGS | \
417 XFEATURE_MASK_BNDCSR)
420 * setup the xstate image representing the init state
422 static void __init setup_init_fpu_buf(void)
424 static int on_boot_cpu __initdata = 1;
426 BUILD_BUG_ON(XCNTXT_MASK != XFEATURES_INIT_FPSTATE_HANDLED);
428 WARN_ON_FPU(!on_boot_cpu);
431 if (!boot_cpu_has(X86_FEATURE_XSAVE))
434 setup_xstate_features();
435 print_xstate_features();
437 if (boot_cpu_has(X86_FEATURE_XSAVES))
438 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
441 * Init all the features state with header.xfeatures being 0x0
443 copy_kernel_to_xregs_booting(&init_fpstate.xsave);
446 * All components are now in init state. Read the state back so
447 * that init_fpstate contains all non-zero init state. This only
448 * works with XSAVE, but not with XSAVEOPT and XSAVES because
449 * those use the init optimization which skips writing data for
450 * components in init state.
452 * XSAVE could be used, but that would require to reshuffle the
453 * data when XSAVES is available because XSAVES uses xstate
454 * compaction. But doing so is a pointless exercise because most
455 * components have an all zeros init state except for the legacy
456 * ones (FP and SSE). Those can be saved with FXSAVE into the
457 * legacy area. Adding new features requires to ensure that init
458 * state is all zeroes or if not to add the necessary handling
461 fxsave(&init_fpstate.fxsave);
464 static int xfeature_uncompacted_offset(int xfeature_nr)
466 u32 eax, ebx, ecx, edx;
469 * Only XSAVES supports supervisor states and it uses compacted
470 * format. Checking a supervisor state's uncompacted offset is
473 if (XFEATURE_MASK_SUPERVISOR & BIT_ULL(xfeature_nr)) {
474 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
478 CHECK_XFEATURE(xfeature_nr);
479 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
483 static int xfeature_size(int xfeature_nr)
485 u32 eax, ebx, ecx, edx;
487 CHECK_XFEATURE(xfeature_nr);
488 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
493 * 'XSAVES' implies two different things:
494 * 1. saving of supervisor/system state
495 * 2. using the compacted format
497 * Use this function when dealing with the compacted format so
498 * that it is obvious which aspect of 'XSAVES' is being handled
499 * by the calling code.
501 int using_compacted_format(void)
503 return boot_cpu_has(X86_FEATURE_XSAVES);
506 /* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
507 int validate_xstate_header(const struct xstate_header *hdr)
509 /* No unknown or supervisor features may be set */
510 if (hdr->xfeatures & (~xfeatures_mask | XFEATURE_MASK_SUPERVISOR))
513 /* Userspace must use the uncompacted format */
518 * If 'reserved' is shrunken to add a new field, make sure to validate
519 * that new field here!
521 BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
523 /* No reserved bits may be set */
524 if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
530 static void __xstate_dump_leaves(void)
533 u32 eax, ebx, ecx, edx;
534 static int should_dump = 1;
540 * Dump out a few leaves past the ones that we support
541 * just in case there are some goodies up there
543 for (i = 0; i < XFEATURE_MAX + 10; i++) {
544 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
545 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
546 XSTATE_CPUID, i, eax, ebx, ecx, edx);
550 #define XSTATE_WARN_ON(x) do { \
551 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
552 __xstate_dump_leaves(); \
556 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
557 if ((nr == nr_macro) && \
558 WARN_ONCE(sz != sizeof(__struct), \
559 "%s: struct is %zu bytes, cpu state %d bytes\n", \
560 __stringify(nr_macro), sizeof(__struct), sz)) { \
561 __xstate_dump_leaves(); \
566 * We have a C struct for each 'xstate'. We need to ensure
567 * that our software representation matches what the CPU
568 * tells us about the state's size.
570 static void check_xstate_against_struct(int nr)
573 * Ask the CPU for the size of the state.
575 int sz = xfeature_size(nr);
577 * Match each CPU state with the corresponding software
580 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct);
581 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state);
582 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state);
583 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state);
584 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
585 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state);
586 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state);
589 * Make *SURE* to add any feature numbers in below if
590 * there are "holes" in the xsave state component
593 if ((nr < XFEATURE_YMM) ||
594 (nr >= XFEATURE_MAX) ||
595 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
596 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
602 * This essentially double-checks what the cpu told us about
603 * how large the XSAVE buffer needs to be. We are recalculating
606 static void do_extra_xstate_size_checks(void)
608 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
611 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
612 if (!xfeature_enabled(i))
615 check_xstate_against_struct(i);
617 * Supervisor state components can be managed only by
618 * XSAVES, which is compacted-format only.
620 if (!using_compacted_format())
621 XSTATE_WARN_ON(xfeature_is_supervisor(i));
623 /* Align from the end of the previous feature */
624 if (xfeature_is_aligned(i))
625 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
627 * The offset of a given state in the non-compacted
628 * format is given to us in a CPUID leaf. We check
629 * them for being ordered (increasing offsets) in
630 * setup_xstate_features().
632 if (!using_compacted_format())
633 paranoid_xstate_size = xfeature_uncompacted_offset(i);
635 * The compacted-format offset always depends on where
636 * the previous state ended.
638 paranoid_xstate_size += xfeature_size(i);
640 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
645 * Get total size of enabled xstates in XCR0/xfeatures_mask.
647 * Note the SDM's wording here. "sub-function 0" only enumerates
648 * the size of the *user* states. If we use it to size a buffer
649 * that we use 'XSAVES' on, we could potentially overflow the
650 * buffer because 'XSAVES' saves system states too.
652 * Note that we do not currently set any bits on IA32_XSS so
653 * 'XCR0 | IA32_XSS == XCR0' for now.
655 static unsigned int __init get_xsaves_size(void)
657 unsigned int eax, ebx, ecx, edx;
659 * - CPUID function 0DH, sub-function 1:
660 * EBX enumerates the size (in bytes) required by
661 * the XSAVES instruction for an XSAVE area
662 * containing all the state components
663 * corresponding to bits currently set in
666 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
670 static unsigned int __init get_xsave_size(void)
672 unsigned int eax, ebx, ecx, edx;
674 * - CPUID function 0DH, sub-function 0:
675 * EBX enumerates the size (in bytes) required by
676 * the XSAVE instruction for an XSAVE area
677 * containing all the *user* state components
678 * corresponding to bits currently set in XCR0.
680 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
685 * Will the runtime-enumerated 'xstate_size' fit in the init
686 * task's statically-allocated buffer?
688 static bool is_supported_xstate_size(unsigned int test_xstate_size)
690 if (test_xstate_size <= sizeof(union fpregs_state))
693 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
694 sizeof(union fpregs_state), test_xstate_size);
698 static int __init init_xstate_size(void)
700 /* Recompute the context size for enabled features: */
701 unsigned int possible_xstate_size;
702 unsigned int xsave_size;
704 xsave_size = get_xsave_size();
706 if (boot_cpu_has(X86_FEATURE_XSAVES))
707 possible_xstate_size = get_xsaves_size();
709 possible_xstate_size = xsave_size;
711 /* Ensure we have the space to store all enabled: */
712 if (!is_supported_xstate_size(possible_xstate_size))
716 * The size is OK, we are definitely going to use xsave,
717 * make it known to the world that we need more space.
719 fpu_kernel_xstate_size = possible_xstate_size;
720 do_extra_xstate_size_checks();
723 * User space is always in standard format.
725 fpu_user_xstate_size = xsave_size;
730 * We enabled the XSAVE hardware, but something went wrong and
731 * we can not use it. Disable it.
733 static void fpu__init_disable_system_xstate(void)
736 cr4_clear_bits(X86_CR4_OSXSAVE);
737 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
741 * Enable and initialize the xsave feature.
742 * Called once per system bootup.
744 void __init fpu__init_system_xstate(void)
746 unsigned int eax, ebx, ecx, edx;
747 static int on_boot_cpu __initdata = 1;
751 WARN_ON_FPU(!on_boot_cpu);
754 if (!boot_cpu_has(X86_FEATURE_FPU)) {
755 pr_info("x86/fpu: No FPU detected\n");
759 if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
760 pr_info("x86/fpu: x87 FPU will use %s\n",
761 boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
765 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
770 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
771 xfeatures_mask = eax + ((u64)edx << 32);
773 if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
775 * This indicates that something really unexpected happened
776 * with the enumeration. Disable XSAVE and try to continue
777 * booting without it. This is too early to BUG().
779 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
784 * Clear XSAVE features that are disabled in the normal CPUID.
786 for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
787 if (!boot_cpu_has(xsave_cpuid_features[i]))
788 xfeatures_mask &= ~BIT(i);
791 xfeatures_mask &= fpu__get_supported_xfeatures_mask();
793 /* Enable xstate instructions to be able to continue with initialization: */
794 fpu__init_cpu_xstate();
795 err = init_xstate_size();
800 * Update info used for ptrace frames; use standard-format size and no
801 * supervisor xstates:
803 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
805 fpu__init_prepare_fx_sw_frame();
806 setup_init_fpu_buf();
810 * CPU capabilities initialization runs before FPU init. So
811 * X86_FEATURE_OSXSAVE is not set. Now that XSAVE is completely
812 * functional, set the feature bit so depending code works.
814 setup_force_cpu_cap(X86_FEATURE_OSXSAVE);
816 print_xstate_offset_size();
818 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
820 fpu_kernel_xstate_size,
821 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
825 /* something went wrong, try to boot without any XSAVE support */
826 fpu__init_disable_system_xstate();
830 * Restore minimal FPU state after suspend:
832 void fpu__resume_cpu(void)
835 * Restore XCR0 on xsave capable CPUs:
837 if (boot_cpu_has(X86_FEATURE_XSAVE))
838 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
842 * Given an xstate feature nr, calculate where in the xsave
843 * buffer the state is. Callers should ensure that the buffer
846 static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
848 if (!xfeature_enabled(xfeature_nr)) {
853 return (void *)xsave + xstate_comp_offsets[xfeature_nr];
856 * Given the xsave area and a state inside, this function returns the
857 * address of the state.
859 * This is the API that is called to get xstate address in either
860 * standard format or compacted format of xsave area.
862 * Note that if there is no data for the field in the xsave buffer
863 * this will return NULL.
866 * xstate: the thread's storage area for all FPU data
867 * xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
868 * XFEATURE_SSE, etc...)
870 * address of the state in the xsave area, or NULL if the
871 * field is not present in the xsave buffer.
873 void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
876 * Do we even *have* xsave state?
878 if (!boot_cpu_has(X86_FEATURE_XSAVE))
882 * We should not ever be requesting features that we
883 * have not enabled. Remember that pcntxt_mask is
884 * what we write to the XCR0 register.
886 WARN_ONCE(!(xfeatures_mask & BIT_ULL(xfeature_nr)),
887 "get of unsupported state");
889 * This assumes the last 'xsave*' instruction to
890 * have requested that 'xfeature_nr' be saved.
891 * If it did not, we might be seeing and old value
892 * of the field in the buffer.
894 * This can happen because the last 'xsave' did not
895 * request that this feature be saved (unlikely)
896 * or because the "init optimization" caused it
899 if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
902 return __raw_xsave_addr(xsave, xfeature_nr);
904 EXPORT_SYMBOL_GPL(get_xsave_addr);
907 * This wraps up the common operations that need to occur when retrieving
908 * data from xsave state. It first ensures that the current task was
909 * using the FPU and retrieves the data in to a buffer. It then calculates
910 * the offset of the requested field in the buffer.
912 * This function is safe to call whether the FPU is in use or not.
914 * Note that this only works on the current task.
917 * @xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
918 * XFEATURE_SSE, etc...)
920 * address of the state in the xsave area or NULL if the state
921 * is not present or is in its 'init state'.
923 const void *get_xsave_field_ptr(int xfeature_nr)
925 struct fpu *fpu = ¤t->thread.fpu;
928 * fpu__save() takes the CPU's xstate registers
929 * and saves them off to the 'fpu memory buffer.
933 return get_xsave_addr(&fpu->state.xsave, xfeature_nr);
936 #ifdef CONFIG_ARCH_HAS_PKEYS
939 * This will go out and modify PKRU register to set the access
940 * rights for @pkey to @init_val.
942 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
943 unsigned long init_val)
946 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
947 u32 new_pkru_bits = 0;
950 * This check implies XSAVE support. OSPKE only gets
951 * set if we enable XSAVE and we enable PKU in XCR0.
953 if (!boot_cpu_has(X86_FEATURE_OSPKE))
957 * This code should only be called with valid 'pkey'
958 * values originating from in-kernel users. Complain
959 * if a bad value is observed.
961 WARN_ON_ONCE(pkey >= arch_max_pkey());
963 /* Set the bits we need in PKRU: */
964 if (init_val & PKEY_DISABLE_ACCESS)
965 new_pkru_bits |= PKRU_AD_BIT;
966 if (init_val & PKEY_DISABLE_WRITE)
967 new_pkru_bits |= PKRU_WD_BIT;
969 /* Shift the bits in to the correct place in PKRU for pkey: */
970 new_pkru_bits <<= pkey_shift;
972 /* Get old PKRU and mask off any old bits in place: */
973 old_pkru = read_pkru();
974 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
976 /* Write old part along with new part: */
977 write_pkru(old_pkru | new_pkru_bits);
981 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
984 * Weird legacy quirk: SSE and YMM states store information in the
985 * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
986 * area is marked as unused in the xfeatures header, we need to copy
987 * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
989 static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
991 if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
994 if (xfeatures & XFEATURE_MASK_FP)
1000 static void fill_gap(unsigned to, void **kbuf, unsigned *pos, unsigned *count)
1003 unsigned size = to - *pos;
1007 memcpy(*kbuf, (void *)&init_fpstate.xsave + *pos, size);
1014 static void copy_part(unsigned offset, unsigned size, void *from,
1015 void **kbuf, unsigned *pos, unsigned *count)
1017 fill_gap(offset, kbuf, pos, count);
1021 memcpy(*kbuf, from, size);
1029 * Convert from kernel XSAVES compacted format to standard format and copy
1030 * to a kernel-space ptrace buffer.
1032 * It supports partial copy but pos always starts from zero. This is called
1033 * from xstateregs_get() and there we check the CPU has XSAVES.
1035 int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1037 struct xstate_header header;
1038 const unsigned off_mxcsr = offsetof(struct fxregs_state, mxcsr);
1039 unsigned count = size_total;
1043 * Currently copy_regset_to_user() starts from pos 0:
1045 if (unlikely(offset_start != 0))
1049 * The destination is a ptrace buffer; we put in only user xstates:
1051 memset(&header, 0, sizeof(header));
1052 header.xfeatures = xsave->header.xfeatures;
1053 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1055 if (header.xfeatures & XFEATURE_MASK_FP)
1056 copy_part(0, off_mxcsr,
1057 &xsave->i387, &kbuf, &offset_start, &count);
1058 if (header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM))
1059 copy_part(off_mxcsr, MXCSR_AND_FLAGS_SIZE,
1060 &xsave->i387.mxcsr, &kbuf, &offset_start, &count);
1061 if (header.xfeatures & XFEATURE_MASK_FP)
1062 copy_part(offsetof(struct fxregs_state, st_space), 128,
1063 &xsave->i387.st_space, &kbuf, &offset_start, &count);
1064 if (header.xfeatures & XFEATURE_MASK_SSE)
1065 copy_part(xstate_offsets[XFEATURE_SSE], 256,
1066 &xsave->i387.xmm_space, &kbuf, &offset_start, &count);
1068 * Fill xsave->i387.sw_reserved value for ptrace frame:
1070 copy_part(offsetof(struct fxregs_state, sw_reserved), 48,
1071 xstate_fx_sw_bytes, &kbuf, &offset_start, &count);
1073 * Copy xregs_state->header:
1075 copy_part(offsetof(struct xregs_state, header), sizeof(header),
1076 &header, &kbuf, &offset_start, &count);
1078 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
1080 * Copy only in-use xstates:
1082 if ((header.xfeatures >> i) & 1) {
1083 void *src = __raw_xsave_addr(xsave, i);
1085 copy_part(xstate_offsets[i], xstate_sizes[i],
1086 src, &kbuf, &offset_start, &count);
1090 fill_gap(size_total, &kbuf, &offset_start, &count);
1096 __copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total)
1101 if (offset < size_total) {
1102 unsigned int copy = min(size, size_total - offset);
1104 if (__copy_to_user(ubuf + offset, data, copy))
1111 * Convert from kernel XSAVES compacted format to standard format and copy
1112 * to a user-space buffer. It supports partial copy but pos always starts from
1113 * zero. This is called from xstateregs_get() and there we check the CPU
1116 int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1118 unsigned int offset, size;
1120 struct xstate_header header;
1123 * Currently copy_regset_to_user() starts from pos 0:
1125 if (unlikely(offset_start != 0))
1129 * The destination is a ptrace buffer; we put in only user xstates:
1131 memset(&header, 0, sizeof(header));
1132 header.xfeatures = xsave->header.xfeatures;
1133 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1136 * Copy xregs_state->header:
1138 offset = offsetof(struct xregs_state, header);
1139 size = sizeof(header);
1141 ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total);
1145 for (i = 0; i < XFEATURE_MAX; i++) {
1147 * Copy only in-use xstates:
1149 if ((header.xfeatures >> i) & 1) {
1150 void *src = __raw_xsave_addr(xsave, i);
1152 offset = xstate_offsets[i];
1153 size = xstate_sizes[i];
1155 /* The next component has to fit fully into the output buffer: */
1156 if (offset + size > size_total)
1159 ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total);
1166 if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1167 offset = offsetof(struct fxregs_state, mxcsr);
1168 size = MXCSR_AND_FLAGS_SIZE;
1169 __copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total);
1173 * Fill xsave->i387.sw_reserved value for ptrace frame:
1175 offset = offsetof(struct fxregs_state, sw_reserved);
1176 size = sizeof(xstate_fx_sw_bytes);
1178 ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total);
1186 * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1187 * and copy to the target thread. This is called from xstateregs_set().
1189 int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1191 unsigned int offset, size;
1193 struct xstate_header hdr;
1195 offset = offsetof(struct xregs_state, header);
1198 memcpy(&hdr, kbuf + offset, size);
1200 if (validate_xstate_header(&hdr))
1203 for (i = 0; i < XFEATURE_MAX; i++) {
1204 u64 mask = ((u64)1 << i);
1206 if (hdr.xfeatures & mask) {
1207 void *dst = __raw_xsave_addr(xsave, i);
1209 offset = xstate_offsets[i];
1210 size = xstate_sizes[i];
1212 memcpy(dst, kbuf + offset, size);
1216 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1217 offset = offsetof(struct fxregs_state, mxcsr);
1218 size = MXCSR_AND_FLAGS_SIZE;
1219 memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1223 * The state that came in from userspace was user-state only.
1224 * Mask all the user states out of 'xfeatures':
1226 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1229 * Add back in the features that came in from userspace:
1231 xsave->header.xfeatures |= hdr.xfeatures;
1237 * Convert from a ptrace or sigreturn standard-format user-space buffer to
1238 * kernel XSAVES format and copy to the target thread. This is called from
1239 * xstateregs_set(), as well as potentially from the sigreturn() and
1240 * rt_sigreturn() system calls.
1242 int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1244 unsigned int offset, size;
1246 struct xstate_header hdr;
1248 offset = offsetof(struct xregs_state, header);
1251 if (__copy_from_user(&hdr, ubuf + offset, size))
1254 if (validate_xstate_header(&hdr))
1257 for (i = 0; i < XFEATURE_MAX; i++) {
1258 u64 mask = ((u64)1 << i);
1260 if (hdr.xfeatures & mask) {
1261 void *dst = __raw_xsave_addr(xsave, i);
1263 offset = xstate_offsets[i];
1264 size = xstate_sizes[i];
1266 if (__copy_from_user(dst, ubuf + offset, size))
1271 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1272 offset = offsetof(struct fxregs_state, mxcsr);
1273 size = MXCSR_AND_FLAGS_SIZE;
1274 if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1279 * The state that came in from userspace was user-state only.
1280 * Mask all the user states out of 'xfeatures':
1282 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1285 * Add back in the features that came in from userspace:
1287 xsave->header.xfeatures |= hdr.xfeatures;
1292 #ifdef CONFIG_PROC_PID_ARCH_STATUS
1294 * Report the amount of time elapsed in millisecond since last AVX512
1297 static void avx512_status(struct seq_file *m, struct task_struct *task)
1299 unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1304 * Report -1 if no AVX512 usage
1308 delta = (long)(jiffies - timestamp);
1310 * Cap to LONG_MAX if time difference > LONG_MAX
1314 delta = jiffies_to_msecs(delta);
1317 seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1322 * Report architecture specific information
1324 int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1325 struct pid *pid, struct task_struct *task)
1328 * Report AVX512 state if the processor and build option supported.
1330 if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1331 avx512_status(m, task);
1335 #endif /* CONFIG_PROC_PID_ARCH_STATUS */