GNU Linux-libre 4.4.288-gnu1
[releases.git] / arch / x86 / kernel / fpu / xstate.c
1 /*
2  * xsave/xrstor support.
3  *
4  * Author: Suresh Siddha <suresh.b.siddha@intel.com>
5  */
6 #include <linux/compat.h>
7 #include <linux/cpu.h>
8
9 #include <asm/fpu/api.h>
10 #include <asm/fpu/internal.h>
11 #include <asm/fpu/signal.h>
12 #include <asm/fpu/regset.h>
13
14 #include <asm/tlbflush.h>
15
16 static const char *xfeature_names[] =
17 {
18         "x87 floating point registers"  ,
19         "SSE registers"                 ,
20         "AVX registers"                 ,
21         "MPX bounds registers"          ,
22         "MPX CSR"                       ,
23         "AVX-512 opmask"                ,
24         "AVX-512 Hi256"                 ,
25         "AVX-512 ZMM_Hi256"             ,
26         "unknown xstate feature"        ,
27 };
28
29 /*
30  * Mask of xstate features supported by the CPU and the kernel:
31  */
32 u64 xfeatures_mask __read_mostly;
33
34 static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
35 static unsigned int xstate_sizes[XFEATURE_MAX]   = { [ 0 ... XFEATURE_MAX - 1] = -1};
36 static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
37
38 /*
39  * Clear all of the X86_FEATURE_* bits that are unavailable
40  * when the CPU has no XSAVE support.
41  */
42 void fpu__xstate_clear_all_cpu_caps(void)
43 {
44         setup_clear_cpu_cap(X86_FEATURE_XSAVE);
45         setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
46         setup_clear_cpu_cap(X86_FEATURE_XSAVEC);
47         setup_clear_cpu_cap(X86_FEATURE_XSAVES);
48         setup_clear_cpu_cap(X86_FEATURE_AVX);
49         setup_clear_cpu_cap(X86_FEATURE_AVX2);
50         setup_clear_cpu_cap(X86_FEATURE_AVX512F);
51         setup_clear_cpu_cap(X86_FEATURE_AVX512PF);
52         setup_clear_cpu_cap(X86_FEATURE_AVX512ER);
53         setup_clear_cpu_cap(X86_FEATURE_AVX512CD);
54         setup_clear_cpu_cap(X86_FEATURE_MPX);
55 }
56
57 /*
58  * Return whether the system supports a given xfeature.
59  *
60  * Also return the name of the (most advanced) feature that the caller requested:
61  */
62 int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
63 {
64         u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
65
66         if (unlikely(feature_name)) {
67                 long xfeature_idx, max_idx;
68                 u64 xfeatures_print;
69                 /*
70                  * So we use FLS here to be able to print the most advanced
71                  * feature that was requested but is missing. So if a driver
72                  * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
73                  * missing AVX feature - this is the most informative message
74                  * to users:
75                  */
76                 if (xfeatures_missing)
77                         xfeatures_print = xfeatures_missing;
78                 else
79                         xfeatures_print = xfeatures_needed;
80
81                 xfeature_idx = fls64(xfeatures_print)-1;
82                 max_idx = ARRAY_SIZE(xfeature_names)-1;
83                 xfeature_idx = min(xfeature_idx, max_idx);
84
85                 *feature_name = xfeature_names[xfeature_idx];
86         }
87
88         if (xfeatures_missing)
89                 return 0;
90
91         return 1;
92 }
93 EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
94
95 /*
96  * When executing XSAVEOPT (or other optimized XSAVE instructions), if
97  * a processor implementation detects that an FPU state component is still
98  * (or is again) in its initialized state, it may clear the corresponding
99  * bit in the header.xfeatures field, and can skip the writeout of registers
100  * to the corresponding memory layout.
101  *
102  * This means that when the bit is zero, the state component might still contain
103  * some previous - non-initialized register state.
104  *
105  * Before writing xstate information to user-space we sanitize those components,
106  * to always ensure that the memory layout of a feature will be in the init state
107  * if the corresponding header bit is zero. This is to ensure that user-space doesn't
108  * see some stale state in the memory layout during signal handling, debugging etc.
109  */
110 void fpstate_sanitize_xstate(struct fpu *fpu)
111 {
112         struct fxregs_state *fx = &fpu->state.fxsave;
113         int feature_bit;
114         u64 xfeatures;
115
116         if (!use_xsaveopt())
117                 return;
118
119         xfeatures = fpu->state.xsave.header.xfeatures;
120
121         /*
122          * None of the feature bits are in init state. So nothing else
123          * to do for us, as the memory layout is up to date.
124          */
125         if ((xfeatures & xfeatures_mask) == xfeatures_mask)
126                 return;
127
128         /*
129          * FP is in init state
130          */
131         if (!(xfeatures & XFEATURE_MASK_FP)) {
132                 fx->cwd = 0x37f;
133                 fx->swd = 0;
134                 fx->twd = 0;
135                 fx->fop = 0;
136                 fx->rip = 0;
137                 fx->rdp = 0;
138                 memset(&fx->st_space[0], 0, 128);
139         }
140
141         /*
142          * SSE is in init state
143          */
144         if (!(xfeatures & XFEATURE_MASK_SSE))
145                 memset(&fx->xmm_space[0], 0, 256);
146
147         /*
148          * First two features are FPU and SSE, which above we handled
149          * in a special way already:
150          */
151         feature_bit = 0x2;
152         xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
153
154         /*
155          * Update all the remaining memory layouts according to their
156          * standard xstate layout, if their header bit is in the init
157          * state:
158          */
159         while (xfeatures) {
160                 if (xfeatures & 0x1) {
161                         int offset = xstate_offsets[feature_bit];
162                         int size = xstate_sizes[feature_bit];
163
164                         memcpy((void *)fx + offset,
165                                (void *)&init_fpstate.xsave + offset,
166                                size);
167                 }
168
169                 xfeatures >>= 1;
170                 feature_bit++;
171         }
172 }
173
174 /*
175  * Enable the extended processor state save/restore feature.
176  * Called once per CPU onlining.
177  */
178 void fpu__init_cpu_xstate(void)
179 {
180         if (!cpu_has_xsave || !xfeatures_mask)
181                 return;
182
183         cr4_set_bits(X86_CR4_OSXSAVE);
184         xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
185 }
186
187 /*
188  * Note that in the future we will likely need a pair of
189  * functions here: one for user xstates and the other for
190  * system xstates.  For now, they are the same.
191  */
192 static int xfeature_enabled(enum xfeature xfeature)
193 {
194         return !!(xfeatures_mask & (1UL << xfeature));
195 }
196
197 /*
198  * Record the offsets and sizes of various xstates contained
199  * in the XSAVE state memory layout.
200  */
201 static void __init setup_xstate_features(void)
202 {
203         u32 eax, ebx, ecx, edx, i;
204         /* start at the beginnning of the "extended state" */
205         unsigned int last_good_offset = offsetof(struct xregs_state,
206                                                  extended_state_area);
207
208         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
209                 if (!xfeature_enabled(i))
210                         continue;
211
212                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
213                 xstate_offsets[i] = ebx;
214                 xstate_sizes[i] = eax;
215                 /*
216                  * In our xstate size checks, we assume that the
217                  * highest-numbered xstate feature has the
218                  * highest offset in the buffer.  Ensure it does.
219                  */
220                 WARN_ONCE(last_good_offset > xstate_offsets[i],
221                         "x86/fpu: misordered xstate at %d\n", last_good_offset);
222                 last_good_offset = xstate_offsets[i];
223
224                 printk(KERN_INFO "x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n", i, ebx, i, eax);
225         }
226 }
227
228 static void __init print_xstate_feature(u64 xstate_mask)
229 {
230         const char *feature_name;
231
232         if (cpu_has_xfeatures(xstate_mask, &feature_name))
233                 pr_info("x86/fpu: Supporting XSAVE feature 0x%02Lx: '%s'\n", xstate_mask, feature_name);
234 }
235
236 /*
237  * Print out all the supported xstate features:
238  */
239 static void __init print_xstate_features(void)
240 {
241         print_xstate_feature(XFEATURE_MASK_FP);
242         print_xstate_feature(XFEATURE_MASK_SSE);
243         print_xstate_feature(XFEATURE_MASK_YMM);
244         print_xstate_feature(XFEATURE_MASK_BNDREGS);
245         print_xstate_feature(XFEATURE_MASK_BNDCSR);
246         print_xstate_feature(XFEATURE_MASK_OPMASK);
247         print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
248         print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
249 }
250
251 /*
252  * This function sets up offsets and sizes of all extended states in
253  * xsave area. This supports both standard format and compacted format
254  * of the xsave aread.
255  */
256 static void __init setup_xstate_comp(void)
257 {
258         unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
259         int i;
260
261         /*
262          * The FP xstates and SSE xstates are legacy states. They are always
263          * in the fixed offsets in the xsave area in either compacted form
264          * or standard form.
265          */
266         xstate_comp_offsets[0] = 0;
267         xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
268
269         if (!cpu_has_xsaves) {
270                 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
271                         if (xfeature_enabled(i)) {
272                                 xstate_comp_offsets[i] = xstate_offsets[i];
273                                 xstate_comp_sizes[i] = xstate_sizes[i];
274                         }
275                 }
276                 return;
277         }
278
279         xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
280                 FXSAVE_SIZE + XSAVE_HDR_SIZE;
281
282         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
283                 if (xfeature_enabled(i))
284                         xstate_comp_sizes[i] = xstate_sizes[i];
285                 else
286                         xstate_comp_sizes[i] = 0;
287
288                 if (i > FIRST_EXTENDED_XFEATURE)
289                         xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
290                                         + xstate_comp_sizes[i-1];
291
292         }
293 }
294
295 /*
296  * All supported features have either init state all zeros or are
297  * handled in setup_init_fpu() individually. This is an explicit
298  * feature list and does not use XFEATURE_MASK*SUPPORTED to catch
299  * newly added supported features at build time and make people
300  * actually look at the init state for the new feature.
301  */
302 #define XFEATURES_INIT_FPSTATE_HANDLED          \
303         (XFEATURE_MASK_FP |                     \
304          XFEATURE_MASK_SSE |                    \
305          XFEATURE_MASK_YMM |                    \
306          XFEATURE_MASK_OPMASK |                 \
307          XFEATURE_MASK_ZMM_Hi256 |              \
308          XFEATURE_MASK_Hi16_ZMM  |              \
309          XFEATURE_MASK_BNDREGS |                \
310          XFEATURE_MASK_BNDCSR)
311
312 /*
313  * setup the xstate image representing the init state
314  */
315 static void __init setup_init_fpu_buf(void)
316 {
317         static int on_boot_cpu = 1;
318
319         BUILD_BUG_ON(XCNTXT_MASK != XFEATURES_INIT_FPSTATE_HANDLED);
320
321         WARN_ON_FPU(!on_boot_cpu);
322         on_boot_cpu = 0;
323
324         if (!cpu_has_xsave)
325                 return;
326
327         setup_xstate_features();
328         print_xstate_features();
329
330         if (cpu_has_xsaves) {
331                 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
332                 init_fpstate.xsave.header.xfeatures = xfeatures_mask;
333         }
334
335         /*
336          * Init all the features state with header_bv being 0x0
337          */
338         copy_kernel_to_xregs_booting(&init_fpstate.xsave);
339
340         /*
341          * All components are now in init state. Read the state back so
342          * that init_fpstate contains all non-zero init state. This only
343          * works with XSAVE, but not with XSAVEOPT and XSAVES because
344          * those use the init optimization which skips writing data for
345          * components in init state.
346          *
347          * XSAVE could be used, but that would require to reshuffle the
348          * data when XSAVES is available because XSAVES uses xstate
349          * compaction. But doing so is a pointless exercise because most
350          * components have an all zeros init state except for the legacy
351          * ones (FP and SSE). Those can be saved with FXSAVE into the
352          * legacy area. Adding new features requires to ensure that init
353          * state is all zeroes or if not to add the necessary handling
354          * here.
355          */
356         fxsave(&init_fpstate.fxsave);
357 }
358
359 static int xfeature_is_supervisor(int xfeature_nr)
360 {
361         /*
362          * We currently do not support supervisor states, but if
363          * we did, we could find out like this.
364          *
365          * SDM says: If state component i is a user state component,
366          * ECX[0] return 0; if state component i is a supervisor
367          * state component, ECX[0] returns 1.
368         u32 eax, ebx, ecx, edx;
369         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx;
370         return !!(ecx & 1);
371         */
372         return 0;
373 }
374 /*
375 static int xfeature_is_user(int xfeature_nr)
376 {
377         return !xfeature_is_supervisor(xfeature_nr);
378 }
379 */
380
381 /*
382  * This check is important because it is easy to get XSTATE_*
383  * confused with XSTATE_BIT_*.
384  */
385 #define CHECK_XFEATURE(nr) do {         \
386         WARN_ON(nr < FIRST_EXTENDED_XFEATURE);  \
387         WARN_ON(nr >= XFEATURE_MAX);    \
388 } while (0)
389
390 /*
391  * We could cache this like xstate_size[], but we only use
392  * it here, so it would be a waste of space.
393  */
394 static int xfeature_is_aligned(int xfeature_nr)
395 {
396         u32 eax, ebx, ecx, edx;
397
398         CHECK_XFEATURE(xfeature_nr);
399         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
400         /*
401          * The value returned by ECX[1] indicates the alignment
402          * of state component i when the compacted format
403          * of the extended region of an XSAVE area is used
404          */
405         return !!(ecx & 2);
406 }
407
408 static int xfeature_uncompacted_offset(int xfeature_nr)
409 {
410         u32 eax, ebx, ecx, edx;
411
412         CHECK_XFEATURE(xfeature_nr);
413         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
414         return ebx;
415 }
416
417 static int xfeature_size(int xfeature_nr)
418 {
419         u32 eax, ebx, ecx, edx;
420
421         CHECK_XFEATURE(xfeature_nr);
422         cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
423         return eax;
424 }
425
426 /*
427  * 'XSAVES' implies two different things:
428  * 1. saving of supervisor/system state
429  * 2. using the compacted format
430  *
431  * Use this function when dealing with the compacted format so
432  * that it is obvious which aspect of 'XSAVES' is being handled
433  * by the calling code.
434  */
435 static int using_compacted_format(void)
436 {
437         return cpu_has_xsaves;
438 }
439
440 static void __xstate_dump_leaves(void)
441 {
442         int i;
443         u32 eax, ebx, ecx, edx;
444         static int should_dump = 1;
445
446         if (!should_dump)
447                 return;
448         should_dump = 0;
449         /*
450          * Dump out a few leaves past the ones that we support
451          * just in case there are some goodies up there
452          */
453         for (i = 0; i < XFEATURE_MAX + 10; i++) {
454                 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
455                 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
456                         XSTATE_CPUID, i, eax, ebx, ecx, edx);
457         }
458 }
459
460 #define XSTATE_WARN_ON(x) do {                                                  \
461         if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) {        \
462                 __xstate_dump_leaves();                                         \
463         }                                                                       \
464 } while (0)
465
466 #define XCHECK_SZ(sz, nr, nr_macro, __struct) do {                      \
467         if ((nr == nr_macro) &&                                         \
468             WARN_ONCE(sz != sizeof(__struct),                           \
469                 "%s: struct is %zu bytes, cpu state %d bytes\n",        \
470                 __stringify(nr_macro), sizeof(__struct), sz)) {         \
471                 __xstate_dump_leaves();                                 \
472         }                                                               \
473 } while (0)
474
475 /*
476  * We have a C struct for each 'xstate'.  We need to ensure
477  * that our software representation matches what the CPU
478  * tells us about the state's size.
479  */
480 static void check_xstate_against_struct(int nr)
481 {
482         /*
483          * Ask the CPU for the size of the state.
484          */
485         int sz = xfeature_size(nr);
486         /*
487          * Match each CPU state with the corresponding software
488          * structure.
489          */
490         XCHECK_SZ(sz, nr, XFEATURE_YMM,       struct ymmh_struct);
491         XCHECK_SZ(sz, nr, XFEATURE_BNDREGS,   struct mpx_bndreg_state);
492         XCHECK_SZ(sz, nr, XFEATURE_BNDCSR,    struct mpx_bndcsr_state);
493         XCHECK_SZ(sz, nr, XFEATURE_OPMASK,    struct avx_512_opmask_state);
494         XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
495         XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM,  struct avx_512_hi16_state);
496
497         /*
498          * Make *SURE* to add any feature numbers in below if
499          * there are "holes" in the xsave state component
500          * numbers.
501          */
502         if ((nr < XFEATURE_YMM) ||
503             (nr >= XFEATURE_MAX)) {
504                 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
505                 XSTATE_WARN_ON(1);
506         }
507 }
508
509 /*
510  * This essentially double-checks what the cpu told us about
511  * how large the XSAVE buffer needs to be.  We are recalculating
512  * it to be safe.
513  */
514 static void do_extra_xstate_size_checks(void)
515 {
516         int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
517         int i;
518
519         for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
520                 if (!xfeature_enabled(i))
521                         continue;
522
523                 check_xstate_against_struct(i);
524                 /*
525                  * Supervisor state components can be managed only by
526                  * XSAVES, which is compacted-format only.
527                  */
528                 if (!using_compacted_format())
529                         XSTATE_WARN_ON(xfeature_is_supervisor(i));
530
531                 /* Align from the end of the previous feature */
532                 if (xfeature_is_aligned(i))
533                         paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
534                 /*
535                  * The offset of a given state in the non-compacted
536                  * format is given to us in a CPUID leaf.  We check
537                  * them for being ordered (increasing offsets) in
538                  * setup_xstate_features().
539                  */
540                 if (!using_compacted_format())
541                         paranoid_xstate_size = xfeature_uncompacted_offset(i);
542                 /*
543                  * The compacted-format offset always depends on where
544                  * the previous state ended.
545                  */
546                 paranoid_xstate_size += xfeature_size(i);
547         }
548         XSTATE_WARN_ON(paranoid_xstate_size != xstate_size);
549 }
550
551 /*
552  * Calculate total size of enabled xstates in XCR0/xfeatures_mask.
553  *
554  * Note the SDM's wording here.  "sub-function 0" only enumerates
555  * the size of the *user* states.  If we use it to size a buffer
556  * that we use 'XSAVES' on, we could potentially overflow the
557  * buffer because 'XSAVES' saves system states too.
558  *
559  * Note that we do not currently set any bits on IA32_XSS so
560  * 'XCR0 | IA32_XSS == XCR0' for now.
561  */
562 static unsigned int __init calculate_xstate_size(void)
563 {
564         unsigned int eax, ebx, ecx, edx;
565         unsigned int calculated_xstate_size;
566
567         if (!cpu_has_xsaves) {
568                 /*
569                  * - CPUID function 0DH, sub-function 0:
570                  *    EBX enumerates the size (in bytes) required by
571                  *    the XSAVE instruction for an XSAVE area
572                  *    containing all the *user* state components
573                  *    corresponding to bits currently set in XCR0.
574                  */
575                 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
576                 calculated_xstate_size = ebx;
577         } else {
578                 /*
579                  * - CPUID function 0DH, sub-function 1:
580                  *    EBX enumerates the size (in bytes) required by
581                  *    the XSAVES instruction for an XSAVE area
582                  *    containing all the state components
583                  *    corresponding to bits currently set in
584                  *    XCR0 | IA32_XSS.
585                  */
586                 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
587                 calculated_xstate_size = ebx;
588         }
589         return calculated_xstate_size;
590 }
591
592 /*
593  * Will the runtime-enumerated 'xstate_size' fit in the init
594  * task's statically-allocated buffer?
595  */
596 static bool is_supported_xstate_size(unsigned int test_xstate_size)
597 {
598         if (test_xstate_size <= sizeof(union fpregs_state))
599                 return true;
600
601         pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
602                         sizeof(union fpregs_state), test_xstate_size);
603         return false;
604 }
605
606 static int init_xstate_size(void)
607 {
608         /* Recompute the context size for enabled features: */
609         unsigned int possible_xstate_size = calculate_xstate_size();
610
611         /* Ensure we have the space to store all enabled: */
612         if (!is_supported_xstate_size(possible_xstate_size))
613                 return -EINVAL;
614
615         /*
616          * The size is OK, we are definitely going to use xsave,
617          * make it known to the world that we need more space.
618          */
619         xstate_size = possible_xstate_size;
620         do_extra_xstate_size_checks();
621         return 0;
622 }
623
624 /*
625  * We enabled the XSAVE hardware, but something went wrong and
626  * we can not use it.  Disable it.
627  */
628 static void fpu__init_disable_system_xstate(void)
629 {
630         xfeatures_mask = 0;
631         cr4_clear_bits(X86_CR4_OSXSAVE);
632         fpu__xstate_clear_all_cpu_caps();
633 }
634
635 /*
636  * Enable and initialize the xsave feature.
637  * Called once per system bootup.
638  */
639 void __init fpu__init_system_xstate(void)
640 {
641         unsigned int eax, ebx, ecx, edx;
642         static int on_boot_cpu = 1;
643         int err;
644
645         WARN_ON_FPU(!on_boot_cpu);
646         on_boot_cpu = 0;
647
648         if (!cpu_has_xsave) {
649                 pr_info("x86/fpu: Legacy x87 FPU detected.\n");
650                 return;
651         }
652
653         if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
654                 WARN_ON_FPU(1);
655                 return;
656         }
657
658         cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
659         xfeatures_mask = eax + ((u64)edx << 32);
660
661         if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
662                 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
663                 BUG();
664         }
665
666         xfeatures_mask &= fpu__get_supported_xfeatures_mask();
667
668         /* Enable xstate instructions to be able to continue with initialization: */
669         fpu__init_cpu_xstate();
670         err = init_xstate_size();
671         if (err) {
672                 /* something went wrong, boot without any XSAVE support */
673                 fpu__init_disable_system_xstate();
674                 return;
675         }
676
677         update_regset_xstate_info(xstate_size, xfeatures_mask);
678         fpu__init_prepare_fx_sw_frame();
679         setup_init_fpu_buf();
680         setup_xstate_comp();
681
682         pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
683                 xfeatures_mask,
684                 xstate_size,
685                 cpu_has_xsaves ? "compacted" : "standard");
686 }
687
688 /*
689  * Restore minimal FPU state after suspend:
690  */
691 void fpu__resume_cpu(void)
692 {
693         /*
694          * Restore XCR0 on xsave capable CPUs:
695          */
696         if (cpu_has_xsave)
697                 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
698 }
699
700 /*
701  * Given the xsave area and a state inside, this function returns the
702  * address of the state.
703  *
704  * This is the API that is called to get xstate address in either
705  * standard format or compacted format of xsave area.
706  *
707  * Note that if there is no data for the field in the xsave buffer
708  * this will return NULL.
709  *
710  * Inputs:
711  *      xstate: the thread's storage area for all FPU data
712  *      xstate_feature: state which is defined in xsave.h (e.g.
713  *      XFEATURE_MASK_FP, XFEATURE_MASK_SSE, etc...)
714  * Output:
715  *      address of the state in the xsave area, or NULL if the
716  *      field is not present in the xsave buffer.
717  */
718 void *get_xsave_addr(struct xregs_state *xsave, int xstate_feature)
719 {
720         int feature_nr = fls64(xstate_feature) - 1;
721         /*
722          * Do we even *have* xsave state?
723          */
724         if (!boot_cpu_has(X86_FEATURE_XSAVE))
725                 return NULL;
726
727         /*
728          * We should not ever be requesting features that we
729          * have not enabled.  Remember that pcntxt_mask is
730          * what we write to the XCR0 register.
731          */
732         WARN_ONCE(!(xfeatures_mask & xstate_feature),
733                   "get of unsupported state");
734         /*
735          * This assumes the last 'xsave*' instruction to
736          * have requested that 'xstate_feature' be saved.
737          * If it did not, we might be seeing and old value
738          * of the field in the buffer.
739          *
740          * This can happen because the last 'xsave' did not
741          * request that this feature be saved (unlikely)
742          * or because the "init optimization" caused it
743          * to not be saved.
744          */
745         if (!(xsave->header.xfeatures & xstate_feature))
746                 return NULL;
747
748         return (void *)xsave + xstate_comp_offsets[feature_nr];
749 }
750 EXPORT_SYMBOL_GPL(get_xsave_addr);
751
752 /*
753  * This wraps up the common operations that need to occur when retrieving
754  * data from xsave state.  It first ensures that the current task was
755  * using the FPU and retrieves the data in to a buffer.  It then calculates
756  * the offset of the requested field in the buffer.
757  *
758  * This function is safe to call whether the FPU is in use or not.
759  *
760  * Note that this only works on the current task.
761  *
762  * Inputs:
763  *      @xsave_state: state which is defined in xsave.h (e.g. XFEATURE_MASK_FP,
764  *      XFEATURE_MASK_SSE, etc...)
765  * Output:
766  *      address of the state in the xsave area or NULL if the state
767  *      is not present or is in its 'init state'.
768  */
769 const void *get_xsave_field_ptr(int xsave_state)
770 {
771         struct fpu *fpu = &current->thread.fpu;
772
773         if (!fpu->fpstate_active)
774                 return NULL;
775         /*
776          * fpu__save() takes the CPU's xstate registers
777          * and saves them off to the 'fpu memory buffer.
778          */
779         fpu__save(fpu);
780
781         return get_xsave_addr(&fpu->state.xsave, xsave_state);
782 }