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);
80 * Return whether the system supports a given xfeature.
82 * Also return the name of the (most advanced) feature that the caller requested:
84 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
86 u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
88 if (unlikely(feature_name)) {
89 long xfeature_idx, max_idx;
92 * So we use FLS here to be able to print the most advanced
93 * feature that was requested but is missing. So if a driver
94 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
95 * missing AVX feature - this is the most informative message
98 if (xfeatures_missing)
99 xfeatures_print = xfeatures_missing;
101 xfeatures_print = xfeatures_needed;
103 xfeature_idx = fls64(xfeatures_print)-1;
104 max_idx = ARRAY_SIZE(xfeature_names)-1;
105 xfeature_idx = min(xfeature_idx, max_idx);
107 *feature_name = xfeature_names[xfeature_idx];
110 if (xfeatures_missing)
115 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
117 static int xfeature_is_supervisor(int xfeature_nr)
120 * We currently do not support supervisor states, but if
121 * we did, we could find out like this.
123 * SDM says: If state component 'i' is a user state component,
124 * ECX[0] return 0; if state component i is a supervisor
125 * state component, ECX[0] returns 1.
127 u32 eax, ebx, ecx, edx;
129 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
133 static int xfeature_is_user(int xfeature_nr)
135 return !xfeature_is_supervisor(xfeature_nr);
139 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
140 * a processor implementation detects that an FPU state component is still
141 * (or is again) in its initialized state, it may clear the corresponding
142 * bit in the header.xfeatures field, and can skip the writeout of registers
143 * to the corresponding memory layout.
145 * This means that when the bit is zero, the state component might still contain
146 * some previous - non-initialized register state.
148 * Before writing xstate information to user-space we sanitize those components,
149 * to always ensure that the memory layout of a feature will be in the init state
150 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
151 * see some stale state in the memory layout during signal handling, debugging etc.
153 void fpstate_sanitize_xstate(struct fpu *fpu)
155 struct fxregs_state *fx = &fpu->state.fxsave;
162 xfeatures = fpu->state.xsave.header.xfeatures;
165 * None of the feature bits are in init state. So nothing else
166 * to do for us, as the memory layout is up to date.
168 if ((xfeatures & xfeatures_mask) == xfeatures_mask)
172 * FP is in init state
174 if (!(xfeatures & XFEATURE_MASK_FP)) {
181 memset(&fx->st_space[0], 0, 128);
185 * SSE is in init state
187 if (!(xfeatures & XFEATURE_MASK_SSE))
188 memset(&fx->xmm_space[0], 0, 256);
191 * First two features are FPU and SSE, which above we handled
192 * in a special way already:
195 xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
198 * Update all the remaining memory layouts according to their
199 * standard xstate layout, if their header bit is in the init
203 if (xfeatures & 0x1) {
204 int offset = xstate_comp_offsets[feature_bit];
205 int size = xstate_sizes[feature_bit];
207 memcpy((void *)fx + offset,
208 (void *)&init_fpstate.xsave + offset,
218 * Enable the extended processor state save/restore feature.
219 * Called once per CPU onlining.
221 void fpu__init_cpu_xstate(void)
223 if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
226 * Make it clear that XSAVES supervisor states are not yet
227 * implemented should anyone expect it to work by changing
228 * bits in XFEATURE_MASK_* macros and XCR0.
230 WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
231 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
233 xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
235 cr4_set_bits(X86_CR4_OSXSAVE);
236 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
240 * Note that in the future we will likely need a pair of
241 * functions here: one for user xstates and the other for
242 * system xstates. For now, they are the same.
244 static int xfeature_enabled(enum xfeature xfeature)
246 return !!(xfeatures_mask & (1UL << xfeature));
250 * Record the offsets and sizes of various xstates contained
251 * in the XSAVE state memory layout.
253 static void __init setup_xstate_features(void)
255 u32 eax, ebx, ecx, edx, i;
256 /* start at the beginnning of the "extended state" */
257 unsigned int last_good_offset = offsetof(struct xregs_state,
258 extended_state_area);
260 * The FP xstates and SSE xstates are legacy states. They are always
261 * in the fixed offsets in the xsave area in either compacted form
264 xstate_offsets[0] = 0;
265 xstate_sizes[0] = offsetof(struct fxregs_state, xmm_space);
266 xstate_offsets[1] = xstate_sizes[0];
267 xstate_sizes[1] = FIELD_SIZEOF(struct fxregs_state, xmm_space);
269 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
270 if (!xfeature_enabled(i))
273 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
276 * If an xfeature is supervisor state, the offset
277 * in EBX is invalid. We leave it to -1.
279 if (xfeature_is_user(i))
280 xstate_offsets[i] = ebx;
282 xstate_sizes[i] = eax;
284 * In our xstate size checks, we assume that the
285 * highest-numbered xstate feature has the
286 * highest offset in the buffer. Ensure it does.
288 WARN_ONCE(last_good_offset > xstate_offsets[i],
289 "x86/fpu: misordered xstate at %d\n", last_good_offset);
290 last_good_offset = xstate_offsets[i];
294 static void __init print_xstate_feature(u64 xstate_mask)
296 const char *feature_name;
298 if (cpu_has_xfeatures(xstate_mask, &feature_name))
299 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
303 * Print out all the supported xstate features:
305 static void __init print_xstate_features(void)
307 print_xstate_feature(XFEATURE_MASK_FP);
308 print_xstate_feature(XFEATURE_MASK_SSE);
309 print_xstate_feature(XFEATURE_MASK_YMM);
310 print_xstate_feature(XFEATURE_MASK_BNDREGS);
311 print_xstate_feature(XFEATURE_MASK_BNDCSR);
312 print_xstate_feature(XFEATURE_MASK_OPMASK);
313 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
314 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
315 print_xstate_feature(XFEATURE_MASK_PKRU);
319 * This check is important because it is easy to get XSTATE_*
320 * confused with XSTATE_BIT_*.
322 #define CHECK_XFEATURE(nr) do { \
323 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
324 WARN_ON(nr >= XFEATURE_MAX); \
328 * We could cache this like xstate_size[], but we only use
329 * it here, so it would be a waste of space.
331 static int xfeature_is_aligned(int xfeature_nr)
333 u32 eax, ebx, ecx, edx;
335 CHECK_XFEATURE(xfeature_nr);
336 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
338 * The value returned by ECX[1] indicates the alignment
339 * of state component 'i' when the compacted format
340 * of the extended region of an XSAVE area is used:
346 * This function sets up offsets and sizes of all extended states in
347 * xsave area. This supports both standard format and compacted format
348 * of the xsave aread.
350 static void __init setup_xstate_comp(void)
352 unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
356 * The FP xstates and SSE xstates are legacy states. They are always
357 * in the fixed offsets in the xsave area in either compacted form
360 xstate_comp_offsets[0] = 0;
361 xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
363 if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
364 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
365 if (xfeature_enabled(i)) {
366 xstate_comp_offsets[i] = xstate_offsets[i];
367 xstate_comp_sizes[i] = xstate_sizes[i];
373 xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
374 FXSAVE_SIZE + XSAVE_HDR_SIZE;
376 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
377 if (xfeature_enabled(i))
378 xstate_comp_sizes[i] = xstate_sizes[i];
380 xstate_comp_sizes[i] = 0;
382 if (i > FIRST_EXTENDED_XFEATURE) {
383 xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
384 + xstate_comp_sizes[i-1];
386 if (xfeature_is_aligned(i))
387 xstate_comp_offsets[i] =
388 ALIGN(xstate_comp_offsets[i], 64);
394 * Print out xstate component offsets and sizes
396 static void __init print_xstate_offset_size(void)
400 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
401 if (!xfeature_enabled(i))
403 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
404 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
409 * setup the xstate image representing the init state
411 static void __init setup_init_fpu_buf(void)
413 static int on_boot_cpu __initdata = 1;
415 WARN_ON_FPU(!on_boot_cpu);
418 if (!boot_cpu_has(X86_FEATURE_XSAVE))
421 setup_xstate_features();
422 print_xstate_features();
424 if (boot_cpu_has(X86_FEATURE_XSAVES))
425 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
428 * Init all the features state with header.xfeatures being 0x0
430 copy_kernel_to_xregs_booting(&init_fpstate.xsave);
433 * Dump the init state again. This is to identify the init state
434 * of any feature which is not represented by all zero's.
436 copy_xregs_to_kernel_booting(&init_fpstate.xsave);
439 static int xfeature_uncompacted_offset(int xfeature_nr)
441 u32 eax, ebx, ecx, edx;
444 * Only XSAVES supports supervisor states and it uses compacted
445 * format. Checking a supervisor state's uncompacted offset is
448 if (XFEATURE_MASK_SUPERVISOR & (1 << xfeature_nr)) {
449 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
453 CHECK_XFEATURE(xfeature_nr);
454 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
458 static int xfeature_size(int xfeature_nr)
460 u32 eax, ebx, ecx, edx;
462 CHECK_XFEATURE(xfeature_nr);
463 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
468 * 'XSAVES' implies two different things:
469 * 1. saving of supervisor/system state
470 * 2. using the compacted format
472 * Use this function when dealing with the compacted format so
473 * that it is obvious which aspect of 'XSAVES' is being handled
474 * by the calling code.
476 int using_compacted_format(void)
478 return boot_cpu_has(X86_FEATURE_XSAVES);
481 static void __xstate_dump_leaves(void)
484 u32 eax, ebx, ecx, edx;
485 static int should_dump = 1;
491 * Dump out a few leaves past the ones that we support
492 * just in case there are some goodies up there
494 for (i = 0; i < XFEATURE_MAX + 10; i++) {
495 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
496 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
497 XSTATE_CPUID, i, eax, ebx, ecx, edx);
501 #define XSTATE_WARN_ON(x) do { \
502 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
503 __xstate_dump_leaves(); \
507 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
508 if ((nr == nr_macro) && \
509 WARN_ONCE(sz != sizeof(__struct), \
510 "%s: struct is %zu bytes, cpu state %d bytes\n", \
511 __stringify(nr_macro), sizeof(__struct), sz)) { \
512 __xstate_dump_leaves(); \
517 * We have a C struct for each 'xstate'. We need to ensure
518 * that our software representation matches what the CPU
519 * tells us about the state's size.
521 static void check_xstate_against_struct(int nr)
524 * Ask the CPU for the size of the state.
526 int sz = xfeature_size(nr);
528 * Match each CPU state with the corresponding software
531 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct);
532 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state);
533 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state);
534 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state);
535 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
536 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state);
537 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state);
540 * Make *SURE* to add any feature numbers in below if
541 * there are "holes" in the xsave state component
544 if ((nr < XFEATURE_YMM) ||
545 (nr >= XFEATURE_MAX) ||
546 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
547 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
553 * This essentially double-checks what the cpu told us about
554 * how large the XSAVE buffer needs to be. We are recalculating
557 static void do_extra_xstate_size_checks(void)
559 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
562 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
563 if (!xfeature_enabled(i))
566 check_xstate_against_struct(i);
568 * Supervisor state components can be managed only by
569 * XSAVES, which is compacted-format only.
571 if (!using_compacted_format())
572 XSTATE_WARN_ON(xfeature_is_supervisor(i));
574 /* Align from the end of the previous feature */
575 if (xfeature_is_aligned(i))
576 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
578 * The offset of a given state in the non-compacted
579 * format is given to us in a CPUID leaf. We check
580 * them for being ordered (increasing offsets) in
581 * setup_xstate_features().
583 if (!using_compacted_format())
584 paranoid_xstate_size = xfeature_uncompacted_offset(i);
586 * The compacted-format offset always depends on where
587 * the previous state ended.
589 paranoid_xstate_size += xfeature_size(i);
591 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
596 * Get total size of enabled xstates in XCR0/xfeatures_mask.
598 * Note the SDM's wording here. "sub-function 0" only enumerates
599 * the size of the *user* states. If we use it to size a buffer
600 * that we use 'XSAVES' on, we could potentially overflow the
601 * buffer because 'XSAVES' saves system states too.
603 * Note that we do not currently set any bits on IA32_XSS so
604 * 'XCR0 | IA32_XSS == XCR0' for now.
606 static unsigned int __init get_xsaves_size(void)
608 unsigned int eax, ebx, ecx, edx;
610 * - CPUID function 0DH, sub-function 1:
611 * EBX enumerates the size (in bytes) required by
612 * the XSAVES instruction for an XSAVE area
613 * containing all the state components
614 * corresponding to bits currently set in
617 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
621 static unsigned int __init get_xsave_size(void)
623 unsigned int eax, ebx, ecx, edx;
625 * - CPUID function 0DH, sub-function 0:
626 * EBX enumerates the size (in bytes) required by
627 * the XSAVE instruction for an XSAVE area
628 * containing all the *user* state components
629 * corresponding to bits currently set in XCR0.
631 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
636 * Will the runtime-enumerated 'xstate_size' fit in the init
637 * task's statically-allocated buffer?
639 static bool is_supported_xstate_size(unsigned int test_xstate_size)
641 if (test_xstate_size <= sizeof(union fpregs_state))
644 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
645 sizeof(union fpregs_state), test_xstate_size);
649 static int init_xstate_size(void)
651 /* Recompute the context size for enabled features: */
652 unsigned int possible_xstate_size;
653 unsigned int xsave_size;
655 xsave_size = get_xsave_size();
657 if (boot_cpu_has(X86_FEATURE_XSAVES))
658 possible_xstate_size = get_xsaves_size();
660 possible_xstate_size = xsave_size;
662 /* Ensure we have the space to store all enabled: */
663 if (!is_supported_xstate_size(possible_xstate_size))
667 * The size is OK, we are definitely going to use xsave,
668 * make it known to the world that we need more space.
670 fpu_kernel_xstate_size = possible_xstate_size;
671 do_extra_xstate_size_checks();
674 * User space is always in standard format.
676 fpu_user_xstate_size = xsave_size;
681 * We enabled the XSAVE hardware, but something went wrong and
682 * we can not use it. Disable it.
684 static void fpu__init_disable_system_xstate(void)
687 cr4_clear_bits(X86_CR4_OSXSAVE);
688 fpu__xstate_clear_all_cpu_caps();
692 * Enable and initialize the xsave feature.
693 * Called once per system bootup.
695 void __init fpu__init_system_xstate(void)
697 unsigned int eax, ebx, ecx, edx;
698 static int on_boot_cpu __initdata = 1;
701 WARN_ON_FPU(!on_boot_cpu);
704 if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
705 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
709 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
714 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
715 xfeatures_mask = eax + ((u64)edx << 32);
717 if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
719 * This indicates that something really unexpected happened
720 * with the enumeration. Disable XSAVE and try to continue
721 * booting without it. This is too early to BUG().
723 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
727 xfeatures_mask &= fpu__get_supported_xfeatures_mask();
729 /* Enable xstate instructions to be able to continue with initialization: */
730 fpu__init_cpu_xstate();
731 err = init_xstate_size();
736 * Update info used for ptrace frames; use standard-format size and no
737 * supervisor xstates:
739 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
741 fpu__init_prepare_fx_sw_frame();
742 setup_init_fpu_buf();
744 print_xstate_offset_size();
746 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
748 fpu_kernel_xstate_size,
749 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
753 /* something went wrong, try to boot without any XSAVE support */
754 fpu__init_disable_system_xstate();
758 * Restore minimal FPU state after suspend:
760 void fpu__resume_cpu(void)
763 * Restore XCR0 on xsave capable CPUs:
765 if (boot_cpu_has(X86_FEATURE_XSAVE))
766 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
770 * Given an xstate feature mask, calculate where in the xsave
771 * buffer the state is. Callers should ensure that the buffer
774 * Note: does not work for compacted buffers.
776 void *__raw_xsave_addr(struct xregs_state *xsave, int xstate_feature_mask)
778 int feature_nr = fls64(xstate_feature_mask) - 1;
780 if (!xfeature_enabled(feature_nr)) {
785 return (void *)xsave + xstate_comp_offsets[feature_nr];
788 * Given the xsave area and a state inside, this function returns the
789 * address of the state.
791 * This is the API that is called to get xstate address in either
792 * standard format or compacted format of xsave area.
794 * Note that if there is no data for the field in the xsave buffer
795 * this will return NULL.
798 * xstate: the thread's storage area for all FPU data
799 * xstate_feature: state which is defined in xsave.h (e.g.
800 * XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
802 * address of the state in the xsave area, or NULL if the
803 * field is not present in the xsave buffer.
805 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
808 * Do we even *have* xsave state?
810 if (!boot_cpu_has(X86_FEATURE_XSAVE))
814 * We should not ever be requesting features that we
815 * have not enabled. Remember that pcntxt_mask is
816 * what we write to the XCR0 register.
818 WARN_ONCE(!(xfeatures_mask & xstate_feature),
819 "get of unsupported state");
821 * This assumes the last 'xsave*' instruction to
822 * have requested that 'xstate_feature' be saved.
823 * If it did not, we might be seeing and old value
824 * of the field in the buffer.
826 * This can happen because the last 'xsave' did not
827 * request that this feature be saved (unlikely)
828 * or because the "init optimization" caused it
831 if (!(xsave->header.xfeatures & xstate_feature))
834 return __raw_xsave_addr(xsave, xstate_feature);
836 EXPORT_SYMBOL_GPL(get_xsave_addr);
839 * This wraps up the common operations that need to occur when retrieving
840 * data from xsave state. It first ensures that the current task was
841 * using the FPU and retrieves the data in to a buffer. It then calculates
842 * the offset of the requested field in the buffer.
844 * This function is safe to call whether the FPU is in use or not.
846 * Note that this only works on the current task.
849 * @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
850 * XFEATURE_MASK_SSE, etc...)
852 * address of the state in the xsave area or NULL if the state
853 * is not present or is in its 'init state'.
855 const void *get_xsave_field_ptr(int xsave_state)
857 struct fpu *fpu = ¤t->thread.fpu;
859 if (!fpu->fpstate_active)
862 * fpu__save() takes the CPU's xstate registers
863 * and saves them off to the 'fpu memory buffer.
867 return get_xsave_addr(&fpu->state.xsave, xsave_state);
870 #ifdef CONFIG_ARCH_HAS_PKEYS
872 #define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
873 #define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
875 * This will go out and modify PKRU register to set the access
876 * rights for @pkey to @init_val.
878 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
879 unsigned long init_val)
882 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
883 u32 new_pkru_bits = 0;
886 * This check implies XSAVE support. OSPKE only gets
887 * set if we enable XSAVE and we enable PKU in XCR0.
889 if (!boot_cpu_has(X86_FEATURE_OSPKE))
892 * For most XSAVE components, this would be an arduous task:
893 * brining fpstate up to date with fpregs, updating fpstate,
894 * then re-populating fpregs. But, for components that are
895 * never lazily managed, we can just access the fpregs
896 * directly. PKRU is never managed lazily, so we can just
897 * manipulate it directly. Make sure it stays that way.
899 WARN_ON_ONCE(!use_eager_fpu());
901 /* Set the bits we need in PKRU: */
902 if (init_val & PKEY_DISABLE_ACCESS)
903 new_pkru_bits |= PKRU_AD_BIT;
904 if (init_val & PKEY_DISABLE_WRITE)
905 new_pkru_bits |= PKRU_WD_BIT;
907 /* Shift the bits in to the correct place in PKRU for pkey: */
908 new_pkru_bits <<= pkey_shift;
910 /* Get old PKRU and mask off any old bits in place: */
911 old_pkru = read_pkru();
912 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
914 /* Write old part along with new part: */
915 write_pkru(old_pkru | new_pkru_bits);
919 #endif /* ! CONFIG_ARCH_HAS_PKEYS */
922 * This is similar to user_regset_copyout(), but will not add offset to
923 * the source data pointer or increment pos, count, kbuf, and ubuf.
925 static inline int xstate_copyout(unsigned int pos, unsigned int count,
926 void *kbuf, void __user *ubuf,
927 const void *data, const int start_pos,
930 if ((count == 0) || (pos < start_pos))
933 if (end_pos < 0 || pos < end_pos) {
934 unsigned int copy = (end_pos < 0 ? count : min(count, end_pos - pos));
937 memcpy(kbuf + pos, data, copy);
939 if (__copy_to_user(ubuf + pos, data, copy))
947 * Convert from kernel XSAVES compacted format to standard format and copy
948 * to a ptrace buffer. It supports partial copy but pos always starts from
949 * zero. This is called from xstateregs_get() and there we check the CPU
952 int copyout_from_xsaves(unsigned int pos, unsigned int count, void *kbuf,
953 void __user *ubuf, struct xregs_state *xsave)
955 unsigned int offset, size;
957 struct xstate_header header;
960 * Currently copy_regset_to_user() starts from pos 0:
962 if (unlikely(pos != 0))
966 * The destination is a ptrace buffer; we put in only user xstates:
968 memset(&header, 0, sizeof(header));
969 header.xfeatures = xsave->header.xfeatures;
970 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
973 * Copy xregs_state->header:
975 offset = offsetof(struct xregs_state, header);
976 size = sizeof(header);
978 ret = xstate_copyout(offset, size, kbuf, ubuf, &header, 0, count);
983 for (i = 0; i < XFEATURE_MAX; i++) {
985 * Copy only in-use xstates:
987 if ((header.xfeatures >> i) & 1) {
988 void *src = __raw_xsave_addr(xsave, 1 << i);
990 offset = xstate_offsets[i];
991 size = xstate_sizes[i];
993 ret = xstate_copyout(offset, size, kbuf, ubuf, src, 0, count);
998 if (offset + size >= count)
1005 * Fill xsave->i387.sw_reserved value for ptrace frame:
1007 offset = offsetof(struct fxregs_state, sw_reserved);
1008 size = sizeof(xstate_fx_sw_bytes);
1010 ret = xstate_copyout(offset, size, kbuf, ubuf, xstate_fx_sw_bytes, 0, count);
1019 * Convert from a ptrace standard-format buffer to kernel XSAVES format
1020 * and copy to the target thread. This is called from xstateregs_set() and
1021 * there we check the CPU has XSAVES and a whole standard-sized buffer
1024 int copyin_to_xsaves(const void *kbuf, const void __user *ubuf,
1025 struct xregs_state *xsave)
1027 unsigned int offset, size;
1030 u64 allowed_features;
1032 offset = offsetof(struct xregs_state, header);
1033 size = sizeof(xfeatures);
1036 memcpy(&xfeatures, kbuf + offset, size);
1038 if (__copy_from_user(&xfeatures, ubuf + offset, size))
1043 * Reject if the user sets any disabled or supervisor features:
1045 allowed_features = xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR;
1047 if (xfeatures & ~allowed_features)
1050 for (i = 0; i < XFEATURE_MAX; i++) {
1051 u64 mask = ((u64)1 << i);
1053 if (xfeatures & mask) {
1054 void *dst = __raw_xsave_addr(xsave, 1 << i);
1056 offset = xstate_offsets[i];
1057 size = xstate_sizes[i];
1060 memcpy(dst, kbuf + offset, size);
1062 if (__copy_from_user(dst, ubuf + offset, size))
1069 * The state that came in from userspace was user-state only.
1070 * Mask all the user states out of 'xfeatures':
1072 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1075 * Add back in the features that came in from userspace:
1077 xsave->header.xfeatures |= xfeatures;