Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * xsave/xrstor support.
4 *
5 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
6 */
7#include <linux/compat.h>
8#include <linux/cpu.h>
9#include <linux/mman.h>
10#include <linux/pkeys.h>
11#include <linux/seq_file.h>
12#include <linux/proc_fs.h>
13
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>
19
20#include <asm/tlbflush.h>
21#include <asm/cpufeature.h>
22
23/*
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.
27 */
28static const char *xfeature_names[] =
29{
30 "x87 floating point registers" ,
31 "SSE registers" ,
32 "AVX registers" ,
33 "MPX bounds registers" ,
34 "MPX CSR" ,
35 "AVX-512 opmask" ,
36 "AVX-512 Hi256" ,
37 "AVX-512 ZMM_Hi256" ,
38 "Processor Trace (unused)" ,
39 "Protection Keys User registers",
40 "unknown xstate feature" ,
41};
42
43static short xsave_cpuid_features[] __initdata = {
44 X86_FEATURE_FPU,
45 X86_FEATURE_XMM,
46 X86_FEATURE_AVX,
47 X86_FEATURE_MPX,
48 X86_FEATURE_MPX,
49 X86_FEATURE_AVX512F,
50 X86_FEATURE_AVX512F,
51 X86_FEATURE_AVX512F,
52 X86_FEATURE_INTEL_PT,
53 X86_FEATURE_PKU,
54};
55
56/*
57 * Mask of xstate features supported by the CPU and the kernel:
58 */
59u64 xfeatures_mask __read_mostly;
60
61static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
62static unsigned int xstate_sizes[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
63static unsigned int xstate_comp_offsets[sizeof(xfeatures_mask)*8];
64
65/*
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.
69 */
70unsigned int fpu_user_xstate_size;
71
72/*
73 * Return whether the system supports a given xfeature.
74 *
75 * Also return the name of the (most advanced) feature that the caller requested:
76 */
77int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
78{
79 u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask;
80
81 if (unlikely(feature_name)) {
82 long xfeature_idx, max_idx;
83 u64 xfeatures_print;
84 /*
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
89 * to users:
90 */
91 if (xfeatures_missing)
92 xfeatures_print = xfeatures_missing;
93 else
94 xfeatures_print = xfeatures_needed;
95
96 xfeature_idx = fls64(xfeatures_print)-1;
97 max_idx = ARRAY_SIZE(xfeature_names)-1;
98 xfeature_idx = min(xfeature_idx, max_idx);
99
100 *feature_name = xfeature_names[xfeature_idx];
101 }
102
103 if (xfeatures_missing)
104 return 0;
105
106 return 1;
107}
108EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
109
110static int xfeature_is_supervisor(int xfeature_nr)
111{
112 /*
113 * We currently do not support supervisor states, but if
114 * we did, we could find out like this.
115 *
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.
119 */
120 u32 eax, ebx, ecx, edx;
121
122 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
123 return !!(ecx & 1);
124}
125
126static int xfeature_is_user(int xfeature_nr)
127{
128 return !xfeature_is_supervisor(xfeature_nr);
129}
130
131/*
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.
137 *
138 * This means that when the bit is zero, the state component might still contain
139 * some previous - non-initialized register state.
140 *
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.
145 */
146void fpstate_sanitize_xstate(struct fpu *fpu)
147{
148 struct fxregs_state *fx = &fpu->state.fxsave;
149 int feature_bit;
150 u64 xfeatures;
151
152 if (!use_xsaveopt())
153 return;
154
155 xfeatures = fpu->state.xsave.header.xfeatures;
156
157 /*
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.
160 */
161 if ((xfeatures & xfeatures_mask) == xfeatures_mask)
162 return;
163
164 /*
165 * FP is in init state
166 */
167 if (!(xfeatures & XFEATURE_MASK_FP)) {
168 fx->cwd = 0x37f;
169 fx->swd = 0;
170 fx->twd = 0;
171 fx->fop = 0;
172 fx->rip = 0;
173 fx->rdp = 0;
174 memset(&fx->st_space[0], 0, 128);
175 }
176
177 /*
178 * SSE is in init state
179 */
180 if (!(xfeatures & XFEATURE_MASK_SSE))
181 memset(&fx->xmm_space[0], 0, 256);
182
183 /*
184 * First two features are FPU and SSE, which above we handled
185 * in a special way already:
186 */
187 feature_bit = 0x2;
188 xfeatures = (xfeatures_mask & ~xfeatures) >> 2;
189
190 /*
191 * Update all the remaining memory layouts according to their
192 * standard xstate layout, if their header bit is in the init
193 * state:
194 */
195 while (xfeatures) {
196 if (xfeatures & 0x1) {
197 int offset = xstate_comp_offsets[feature_bit];
198 int size = xstate_sizes[feature_bit];
199
200 memcpy((void *)fx + offset,
201 (void *)&init_fpstate.xsave + offset,
202 size);
203 }
204
205 xfeatures >>= 1;
206 feature_bit++;
207 }
208}
209
210/*
211 * Enable the extended processor state save/restore feature.
212 * Called once per CPU onlining.
213 */
214void fpu__init_cpu_xstate(void)
215{
216 if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask)
217 return;
218 /*
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.
222 */
223 WARN_ONCE((xfeatures_mask & XFEATURE_MASK_SUPERVISOR),
224 "x86/fpu: XSAVES supervisor states are not yet implemented.\n");
225
226 xfeatures_mask &= ~XFEATURE_MASK_SUPERVISOR;
227
228 cr4_set_bits(X86_CR4_OSXSAVE);
229 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
230}
231
232/*
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.
236 */
237static int xfeature_enabled(enum xfeature xfeature)
238{
239 return !!(xfeatures_mask & (1UL << xfeature));
240}
241
242/*
243 * Record the offsets and sizes of various xstates contained
244 * in the XSAVE state memory layout.
245 */
246static void __init setup_xstate_features(void)
247{
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);
252 /*
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
255 * or standard form.
256 */
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);
261
262 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
263 if (!xfeature_enabled(i))
264 continue;
265
266 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
267
268 /*
269 * If an xfeature is supervisor state, the offset
270 * in EBX is invalid. We leave it to -1.
271 */
272 if (xfeature_is_user(i))
273 xstate_offsets[i] = ebx;
274
275 xstate_sizes[i] = eax;
276 /*
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.
280 */
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];
284 }
285}
286
287static void __init print_xstate_feature(u64 xstate_mask)
288{
289 const char *feature_name;
290
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);
293}
294
295/*
296 * Print out all the supported xstate features:
297 */
298static void __init print_xstate_features(void)
299{
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);
309}
310
311/*
312 * This check is important because it is easy to get XSTATE_*
313 * confused with XSTATE_BIT_*.
314 */
315#define CHECK_XFEATURE(nr) do { \
316 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
317 WARN_ON(nr >= XFEATURE_MAX); \
318} while (0)
319
320/*
321 * We could cache this like xstate_size[], but we only use
322 * it here, so it would be a waste of space.
323 */
324static int xfeature_is_aligned(int xfeature_nr)
325{
326 u32 eax, ebx, ecx, edx;
327
328 CHECK_XFEATURE(xfeature_nr);
329 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
330 /*
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:
334 */
335 return !!(ecx & 2);
336}
337
338/*
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.
342 */
343static void __init setup_xstate_comp(void)
344{
345 unsigned int xstate_comp_sizes[sizeof(xfeatures_mask)*8];
346 int i;
347
348 /*
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
351 * or standard form.
352 */
353 xstate_comp_offsets[0] = 0;
354 xstate_comp_offsets[1] = offsetof(struct fxregs_state, xmm_space);
355
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];
361 }
362 }
363 return;
364 }
365
366 xstate_comp_offsets[FIRST_EXTENDED_XFEATURE] =
367 FXSAVE_SIZE + XSAVE_HDR_SIZE;
368
369 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
370 if (xfeature_enabled(i))
371 xstate_comp_sizes[i] = xstate_sizes[i];
372 else
373 xstate_comp_sizes[i] = 0;
374
375 if (i > FIRST_EXTENDED_XFEATURE) {
376 xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
377 + xstate_comp_sizes[i-1];
378
379 if (xfeature_is_aligned(i))
380 xstate_comp_offsets[i] =
381 ALIGN(xstate_comp_offsets[i], 64);
382 }
383 }
384}
385
386/*
387 * Print out xstate component offsets and sizes
388 */
389static void __init print_xstate_offset_size(void)
390{
391 int i;
392
393 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
394 if (!xfeature_enabled(i))
395 continue;
396 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
397 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
398 }
399}
400
401/*
402 * setup the xstate image representing the init state
403 */
404static void __init setup_init_fpu_buf(void)
405{
406 static int on_boot_cpu __initdata = 1;
407
408 WARN_ON_FPU(!on_boot_cpu);
409 on_boot_cpu = 0;
410
411 if (!boot_cpu_has(X86_FEATURE_XSAVE))
412 return;
413
414 setup_xstate_features();
415 print_xstate_features();
416
417 if (boot_cpu_has(X86_FEATURE_XSAVES))
418 init_fpstate.xsave.header.xcomp_bv = (u64)1 << 63 | xfeatures_mask;
419
420 /*
421 * Init all the features state with header.xfeatures being 0x0
422 */
423 copy_kernel_to_xregs_booting(&init_fpstate.xsave);
424
425 /*
426 * Dump the init state again. This is to identify the init state
427 * of any feature which is not represented by all zero's.
428 */
429 copy_xregs_to_kernel_booting(&init_fpstate.xsave);
430}
431
432static int xfeature_uncompacted_offset(int xfeature_nr)
433{
434 u32 eax, ebx, ecx, edx;
435
436 /*
437 * Only XSAVES supports supervisor states and it uses compacted
438 * format. Checking a supervisor state's uncompacted offset is
439 * an error.
440 */
441 if (XFEATURE_MASK_SUPERVISOR & BIT_ULL(xfeature_nr)) {
442 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
443 return -1;
444 }
445
446 CHECK_XFEATURE(xfeature_nr);
447 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
448 return ebx;
449}
450
451static int xfeature_size(int xfeature_nr)
452{
453 u32 eax, ebx, ecx, edx;
454
455 CHECK_XFEATURE(xfeature_nr);
456 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
457 return eax;
458}
459
460/*
461 * 'XSAVES' implies two different things:
462 * 1. saving of supervisor/system state
463 * 2. using the compacted format
464 *
465 * Use this function when dealing with the compacted format so
466 * that it is obvious which aspect of 'XSAVES' is being handled
467 * by the calling code.
468 */
469int using_compacted_format(void)
470{
471 return boot_cpu_has(X86_FEATURE_XSAVES);
472}
473
474/* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
475int validate_xstate_header(const struct xstate_header *hdr)
476{
477 /* No unknown or supervisor features may be set */
478 if (hdr->xfeatures & (~xfeatures_mask | XFEATURE_MASK_SUPERVISOR))
479 return -EINVAL;
480
481 /* Userspace must use the uncompacted format */
482 if (hdr->xcomp_bv)
483 return -EINVAL;
484
485 /*
486 * If 'reserved' is shrunken to add a new field, make sure to validate
487 * that new field here!
488 */
489 BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
490
491 /* No reserved bits may be set */
492 if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
493 return -EINVAL;
494
495 return 0;
496}
497
498static void __xstate_dump_leaves(void)
499{
500 int i;
501 u32 eax, ebx, ecx, edx;
502 static int should_dump = 1;
503
504 if (!should_dump)
505 return;
506 should_dump = 0;
507 /*
508 * Dump out a few leaves past the ones that we support
509 * just in case there are some goodies up there
510 */
511 for (i = 0; i < XFEATURE_MAX + 10; i++) {
512 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
513 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
514 XSTATE_CPUID, i, eax, ebx, ecx, edx);
515 }
516}
517
518#define XSTATE_WARN_ON(x) do { \
519 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
520 __xstate_dump_leaves(); \
521 } \
522} while (0)
523
524#define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
525 if ((nr == nr_macro) && \
526 WARN_ONCE(sz != sizeof(__struct), \
527 "%s: struct is %zu bytes, cpu state %d bytes\n", \
528 __stringify(nr_macro), sizeof(__struct), sz)) { \
529 __xstate_dump_leaves(); \
530 } \
531} while (0)
532
533/*
534 * We have a C struct for each 'xstate'. We need to ensure
535 * that our software representation matches what the CPU
536 * tells us about the state's size.
537 */
538static void check_xstate_against_struct(int nr)
539{
540 /*
541 * Ask the CPU for the size of the state.
542 */
543 int sz = xfeature_size(nr);
544 /*
545 * Match each CPU state with the corresponding software
546 * structure.
547 */
548 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct);
549 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state);
550 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state);
551 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state);
552 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
553 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state);
554 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state);
555
556 /*
557 * Make *SURE* to add any feature numbers in below if
558 * there are "holes" in the xsave state component
559 * numbers.
560 */
561 if ((nr < XFEATURE_YMM) ||
562 (nr >= XFEATURE_MAX) ||
563 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR)) {
564 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
565 XSTATE_WARN_ON(1);
566 }
567}
568
569/*
570 * This essentially double-checks what the cpu told us about
571 * how large the XSAVE buffer needs to be. We are recalculating
572 * it to be safe.
573 */
574static void do_extra_xstate_size_checks(void)
575{
576 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
577 int i;
578
579 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
580 if (!xfeature_enabled(i))
581 continue;
582
583 check_xstate_against_struct(i);
584 /*
585 * Supervisor state components can be managed only by
586 * XSAVES, which is compacted-format only.
587 */
588 if (!using_compacted_format())
589 XSTATE_WARN_ON(xfeature_is_supervisor(i));
590
591 /* Align from the end of the previous feature */
592 if (xfeature_is_aligned(i))
593 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
594 /*
595 * The offset of a given state in the non-compacted
596 * format is given to us in a CPUID leaf. We check
597 * them for being ordered (increasing offsets) in
598 * setup_xstate_features().
599 */
600 if (!using_compacted_format())
601 paranoid_xstate_size = xfeature_uncompacted_offset(i);
602 /*
603 * The compacted-format offset always depends on where
604 * the previous state ended.
605 */
606 paranoid_xstate_size += xfeature_size(i);
607 }
608 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
609}
610
611
612/*
613 * Get total size of enabled xstates in XCR0/xfeatures_mask.
614 *
615 * Note the SDM's wording here. "sub-function 0" only enumerates
616 * the size of the *user* states. If we use it to size a buffer
617 * that we use 'XSAVES' on, we could potentially overflow the
618 * buffer because 'XSAVES' saves system states too.
619 *
620 * Note that we do not currently set any bits on IA32_XSS so
621 * 'XCR0 | IA32_XSS == XCR0' for now.
622 */
623static unsigned int __init get_xsaves_size(void)
624{
625 unsigned int eax, ebx, ecx, edx;
626 /*
627 * - CPUID function 0DH, sub-function 1:
628 * EBX enumerates the size (in bytes) required by
629 * the XSAVES instruction for an XSAVE area
630 * containing all the state components
631 * corresponding to bits currently set in
632 * XCR0 | IA32_XSS.
633 */
634 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
635 return ebx;
636}
637
638static unsigned int __init get_xsave_size(void)
639{
640 unsigned int eax, ebx, ecx, edx;
641 /*
642 * - CPUID function 0DH, sub-function 0:
643 * EBX enumerates the size (in bytes) required by
644 * the XSAVE instruction for an XSAVE area
645 * containing all the *user* state components
646 * corresponding to bits currently set in XCR0.
647 */
648 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
649 return ebx;
650}
651
652/*
653 * Will the runtime-enumerated 'xstate_size' fit in the init
654 * task's statically-allocated buffer?
655 */
656static bool is_supported_xstate_size(unsigned int test_xstate_size)
657{
658 if (test_xstate_size <= sizeof(union fpregs_state))
659 return true;
660
661 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
662 sizeof(union fpregs_state), test_xstate_size);
663 return false;
664}
665
666static int __init init_xstate_size(void)
667{
668 /* Recompute the context size for enabled features: */
669 unsigned int possible_xstate_size;
670 unsigned int xsave_size;
671
672 xsave_size = get_xsave_size();
673
674 if (boot_cpu_has(X86_FEATURE_XSAVES))
675 possible_xstate_size = get_xsaves_size();
676 else
677 possible_xstate_size = xsave_size;
678
679 /* Ensure we have the space to store all enabled: */
680 if (!is_supported_xstate_size(possible_xstate_size))
681 return -EINVAL;
682
683 /*
684 * The size is OK, we are definitely going to use xsave,
685 * make it known to the world that we need more space.
686 */
687 fpu_kernel_xstate_size = possible_xstate_size;
688 do_extra_xstate_size_checks();
689
690 /*
691 * User space is always in standard format.
692 */
693 fpu_user_xstate_size = xsave_size;
694 return 0;
695}
696
697/*
698 * We enabled the XSAVE hardware, but something went wrong and
699 * we can not use it. Disable it.
700 */
701static void fpu__init_disable_system_xstate(void)
702{
703 xfeatures_mask = 0;
704 cr4_clear_bits(X86_CR4_OSXSAVE);
705 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
706}
707
708/*
709 * Enable and initialize the xsave feature.
710 * Called once per system bootup.
711 */
712void __init fpu__init_system_xstate(void)
713{
714 unsigned int eax, ebx, ecx, edx;
715 static int on_boot_cpu __initdata = 1;
716 int err;
717 int i;
718
719 WARN_ON_FPU(!on_boot_cpu);
720 on_boot_cpu = 0;
721
722 if (!boot_cpu_has(X86_FEATURE_FPU)) {
723 pr_info("x86/fpu: No FPU detected\n");
724 return;
725 }
726
727 if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
728 pr_info("x86/fpu: x87 FPU will use %s\n",
729 boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
730 return;
731 }
732
733 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
734 WARN_ON_FPU(1);
735 return;
736 }
737
738 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
739 xfeatures_mask = eax + ((u64)edx << 32);
740
741 if ((xfeatures_mask & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
742 /*
743 * This indicates that something really unexpected happened
744 * with the enumeration. Disable XSAVE and try to continue
745 * booting without it. This is too early to BUG().
746 */
747 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n", xfeatures_mask);
748 goto out_disable;
749 }
750
751 /*
752 * Clear XSAVE features that are disabled in the normal CPUID.
753 */
754 for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
755 if (!boot_cpu_has(xsave_cpuid_features[i]))
756 xfeatures_mask &= ~BIT(i);
757 }
758
759 xfeatures_mask &= fpu__get_supported_xfeatures_mask();
760
761 /* Enable xstate instructions to be able to continue with initialization: */
762 fpu__init_cpu_xstate();
763 err = init_xstate_size();
764 if (err)
765 goto out_disable;
766
767 /*
768 * Update info used for ptrace frames; use standard-format size and no
769 * supervisor xstates:
770 */
771 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask & ~XFEATURE_MASK_SUPERVISOR);
772
773 fpu__init_prepare_fx_sw_frame();
774 setup_init_fpu_buf();
775 setup_xstate_comp();
776 print_xstate_offset_size();
777
778 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
779 xfeatures_mask,
780 fpu_kernel_xstate_size,
781 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
782 return;
783
784out_disable:
785 /* something went wrong, try to boot without any XSAVE support */
786 fpu__init_disable_system_xstate();
787}
788
789/*
790 * Restore minimal FPU state after suspend:
791 */
792void fpu__resume_cpu(void)
793{
794 /*
795 * Restore XCR0 on xsave capable CPUs:
796 */
797 if (boot_cpu_has(X86_FEATURE_XSAVE))
798 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask);
799}
800
801/*
802 * Given an xstate feature nr, calculate where in the xsave
803 * buffer the state is. Callers should ensure that the buffer
804 * is valid.
805 */
806static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
807{
808 if (!xfeature_enabled(xfeature_nr)) {
809 WARN_ON_FPU(1);
810 return NULL;
811 }
812
813 return (void *)xsave + xstate_comp_offsets[xfeature_nr];
814}
815/*
816 * Given the xsave area and a state inside, this function returns the
817 * address of the state.
818 *
819 * This is the API that is called to get xstate address in either
820 * standard format or compacted format of xsave area.
821 *
822 * Note that if there is no data for the field in the xsave buffer
823 * this will return NULL.
824 *
825 * Inputs:
826 * xstate: the thread's storage area for all FPU data
827 * xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
828 * XFEATURE_SSE, etc...)
829 * Output:
830 * address of the state in the xsave area, or NULL if the
831 * field is not present in the xsave buffer.
832 */
833void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
834{
835 /*
836 * Do we even *have* xsave state?
837 */
838 if (!boot_cpu_has(X86_FEATURE_XSAVE))
839 return NULL;
840
841 /*
842 * We should not ever be requesting features that we
843 * have not enabled. Remember that pcntxt_mask is
844 * what we write to the XCR0 register.
845 */
846 WARN_ONCE(!(xfeatures_mask & BIT_ULL(xfeature_nr)),
847 "get of unsupported state");
848 /*
849 * This assumes the last 'xsave*' instruction to
850 * have requested that 'xfeature_nr' be saved.
851 * If it did not, we might be seeing and old value
852 * of the field in the buffer.
853 *
854 * This can happen because the last 'xsave' did not
855 * request that this feature be saved (unlikely)
856 * or because the "init optimization" caused it
857 * to not be saved.
858 */
859 if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
860 return NULL;
861
862 return __raw_xsave_addr(xsave, xfeature_nr);
863}
864EXPORT_SYMBOL_GPL(get_xsave_addr);
865
866/*
867 * This wraps up the common operations that need to occur when retrieving
868 * data from xsave state. It first ensures that the current task was
869 * using the FPU and retrieves the data in to a buffer. It then calculates
870 * the offset of the requested field in the buffer.
871 *
872 * This function is safe to call whether the FPU is in use or not.
873 *
874 * Note that this only works on the current task.
875 *
876 * Inputs:
877 * @xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
878 * XFEATURE_SSE, etc...)
879 * Output:
880 * address of the state in the xsave area or NULL if the state
881 * is not present or is in its 'init state'.
882 */
883const void *get_xsave_field_ptr(int xfeature_nr)
884{
885 struct fpu *fpu = ¤t->thread.fpu;
886
887 /*
888 * fpu__save() takes the CPU's xstate registers
889 * and saves them off to the 'fpu memory buffer.
890 */
891 fpu__save(fpu);
892
893 return get_xsave_addr(&fpu->state.xsave, xfeature_nr);
894}
895
896#ifdef CONFIG_ARCH_HAS_PKEYS
897
898#define NR_VALID_PKRU_BITS (CONFIG_NR_PROTECTION_KEYS * 2)
899#define PKRU_VALID_MASK (NR_VALID_PKRU_BITS - 1)
900/*
901 * This will go out and modify PKRU register to set the access
902 * rights for @pkey to @init_val.
903 */
904int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
905 unsigned long init_val)
906{
907 u32 old_pkru;
908 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
909 u32 new_pkru_bits = 0;
910
911 /*
912 * This check implies XSAVE support. OSPKE only gets
913 * set if we enable XSAVE and we enable PKU in XCR0.
914 */
915 if (!boot_cpu_has(X86_FEATURE_OSPKE))
916 return -EINVAL;
917
918 /* Set the bits we need in PKRU: */
919 if (init_val & PKEY_DISABLE_ACCESS)
920 new_pkru_bits |= PKRU_AD_BIT;
921 if (init_val & PKEY_DISABLE_WRITE)
922 new_pkru_bits |= PKRU_WD_BIT;
923
924 /* Shift the bits in to the correct place in PKRU for pkey: */
925 new_pkru_bits <<= pkey_shift;
926
927 /* Get old PKRU and mask off any old bits in place: */
928 old_pkru = read_pkru();
929 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
930
931 /* Write old part along with new part: */
932 write_pkru(old_pkru | new_pkru_bits);
933
934 return 0;
935}
936#endif /* ! CONFIG_ARCH_HAS_PKEYS */
937
938/*
939 * Weird legacy quirk: SSE and YMM states store information in the
940 * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
941 * area is marked as unused in the xfeatures header, we need to copy
942 * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
943 */
944static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
945{
946 if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
947 return false;
948
949 if (xfeatures & XFEATURE_MASK_FP)
950 return false;
951
952 return true;
953}
954
955/*
956 * This is similar to user_regset_copyout(), but will not add offset to
957 * the source data pointer or increment pos, count, kbuf, and ubuf.
958 */
959static inline void
960__copy_xstate_to_kernel(void *kbuf, const void *data,
961 unsigned int offset, unsigned int size, unsigned int size_total)
962{
963 if (offset < size_total) {
964 unsigned int copy = min(size, size_total - offset);
965
966 memcpy(kbuf + offset, data, copy);
967 }
968}
969
970/*
971 * Convert from kernel XSAVES compacted format to standard format and copy
972 * to a kernel-space ptrace buffer.
973 *
974 * It supports partial copy but pos always starts from zero. This is called
975 * from xstateregs_get() and there we check the CPU has XSAVES.
976 */
977int copy_xstate_to_kernel(void *kbuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
978{
979 unsigned int offset, size;
980 struct xstate_header header;
981 int i;
982
983 /*
984 * Currently copy_regset_to_user() starts from pos 0:
985 */
986 if (unlikely(offset_start != 0))
987 return -EFAULT;
988
989 /*
990 * The destination is a ptrace buffer; we put in only user xstates:
991 */
992 memset(&header, 0, sizeof(header));
993 header.xfeatures = xsave->header.xfeatures;
994 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
995
996 /*
997 * Copy xregs_state->header:
998 */
999 offset = offsetof(struct xregs_state, header);
1000 size = sizeof(header);
1001
1002 __copy_xstate_to_kernel(kbuf, &header, offset, size, size_total);
1003
1004 for (i = 0; i < XFEATURE_MAX; i++) {
1005 /*
1006 * Copy only in-use xstates:
1007 */
1008 if ((header.xfeatures >> i) & 1) {
1009 void *src = __raw_xsave_addr(xsave, i);
1010
1011 offset = xstate_offsets[i];
1012 size = xstate_sizes[i];
1013
1014 /* The next component has to fit fully into the output buffer: */
1015 if (offset + size > size_total)
1016 break;
1017
1018 __copy_xstate_to_kernel(kbuf, src, offset, size, size_total);
1019 }
1020
1021 }
1022
1023 if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1024 offset = offsetof(struct fxregs_state, mxcsr);
1025 size = MXCSR_AND_FLAGS_SIZE;
1026 __copy_xstate_to_kernel(kbuf, &xsave->i387.mxcsr, offset, size, size_total);
1027 }
1028
1029 /*
1030 * Fill xsave->i387.sw_reserved value for ptrace frame:
1031 */
1032 offset = offsetof(struct fxregs_state, sw_reserved);
1033 size = sizeof(xstate_fx_sw_bytes);
1034
1035 __copy_xstate_to_kernel(kbuf, xstate_fx_sw_bytes, offset, size, size_total);
1036
1037 return 0;
1038}
1039
1040static inline int
1041__copy_xstate_to_user(void __user *ubuf, const void *data, unsigned int offset, unsigned int size, unsigned int size_total)
1042{
1043 if (!size)
1044 return 0;
1045
1046 if (offset < size_total) {
1047 unsigned int copy = min(size, size_total - offset);
1048
1049 if (__copy_to_user(ubuf + offset, data, copy))
1050 return -EFAULT;
1051 }
1052 return 0;
1053}
1054
1055/*
1056 * Convert from kernel XSAVES compacted format to standard format and copy
1057 * to a user-space buffer. It supports partial copy but pos always starts from
1058 * zero. This is called from xstateregs_get() and there we check the CPU
1059 * has XSAVES.
1060 */
1061int copy_xstate_to_user(void __user *ubuf, struct xregs_state *xsave, unsigned int offset_start, unsigned int size_total)
1062{
1063 unsigned int offset, size;
1064 int ret, i;
1065 struct xstate_header header;
1066
1067 /*
1068 * Currently copy_regset_to_user() starts from pos 0:
1069 */
1070 if (unlikely(offset_start != 0))
1071 return -EFAULT;
1072
1073 /*
1074 * The destination is a ptrace buffer; we put in only user xstates:
1075 */
1076 memset(&header, 0, sizeof(header));
1077 header.xfeatures = xsave->header.xfeatures;
1078 header.xfeatures &= ~XFEATURE_MASK_SUPERVISOR;
1079
1080 /*
1081 * Copy xregs_state->header:
1082 */
1083 offset = offsetof(struct xregs_state, header);
1084 size = sizeof(header);
1085
1086 ret = __copy_xstate_to_user(ubuf, &header, offset, size, size_total);
1087 if (ret)
1088 return ret;
1089
1090 for (i = 0; i < XFEATURE_MAX; i++) {
1091 /*
1092 * Copy only in-use xstates:
1093 */
1094 if ((header.xfeatures >> i) & 1) {
1095 void *src = __raw_xsave_addr(xsave, i);
1096
1097 offset = xstate_offsets[i];
1098 size = xstate_sizes[i];
1099
1100 /* The next component has to fit fully into the output buffer: */
1101 if (offset + size > size_total)
1102 break;
1103
1104 ret = __copy_xstate_to_user(ubuf, src, offset, size, size_total);
1105 if (ret)
1106 return ret;
1107 }
1108
1109 }
1110
1111 if (xfeatures_mxcsr_quirk(header.xfeatures)) {
1112 offset = offsetof(struct fxregs_state, mxcsr);
1113 size = MXCSR_AND_FLAGS_SIZE;
1114 __copy_xstate_to_user(ubuf, &xsave->i387.mxcsr, offset, size, size_total);
1115 }
1116
1117 /*
1118 * Fill xsave->i387.sw_reserved value for ptrace frame:
1119 */
1120 offset = offsetof(struct fxregs_state, sw_reserved);
1121 size = sizeof(xstate_fx_sw_bytes);
1122
1123 ret = __copy_xstate_to_user(ubuf, xstate_fx_sw_bytes, offset, size, size_total);
1124 if (ret)
1125 return ret;
1126
1127 return 0;
1128}
1129
1130/*
1131 * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1132 * and copy to the target thread. This is called from xstateregs_set().
1133 */
1134int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1135{
1136 unsigned int offset, size;
1137 int i;
1138 struct xstate_header hdr;
1139
1140 offset = offsetof(struct xregs_state, header);
1141 size = sizeof(hdr);
1142
1143 memcpy(&hdr, kbuf + offset, size);
1144
1145 if (validate_xstate_header(&hdr))
1146 return -EINVAL;
1147
1148 for (i = 0; i < XFEATURE_MAX; i++) {
1149 u64 mask = ((u64)1 << i);
1150
1151 if (hdr.xfeatures & mask) {
1152 void *dst = __raw_xsave_addr(xsave, i);
1153
1154 offset = xstate_offsets[i];
1155 size = xstate_sizes[i];
1156
1157 memcpy(dst, kbuf + offset, size);
1158 }
1159 }
1160
1161 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1162 offset = offsetof(struct fxregs_state, mxcsr);
1163 size = MXCSR_AND_FLAGS_SIZE;
1164 memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1165 }
1166
1167 /*
1168 * The state that came in from userspace was user-state only.
1169 * Mask all the user states out of 'xfeatures':
1170 */
1171 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1172
1173 /*
1174 * Add back in the features that came in from userspace:
1175 */
1176 xsave->header.xfeatures |= hdr.xfeatures;
1177
1178 return 0;
1179}
1180
1181/*
1182 * Convert from a ptrace or sigreturn standard-format user-space buffer to
1183 * kernel XSAVES format and copy to the target thread. This is called from
1184 * xstateregs_set(), as well as potentially from the sigreturn() and
1185 * rt_sigreturn() system calls.
1186 */
1187int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1188{
1189 unsigned int offset, size;
1190 int i;
1191 struct xstate_header hdr;
1192
1193 offset = offsetof(struct xregs_state, header);
1194 size = sizeof(hdr);
1195
1196 if (__copy_from_user(&hdr, ubuf + offset, size))
1197 return -EFAULT;
1198
1199 if (validate_xstate_header(&hdr))
1200 return -EINVAL;
1201
1202 for (i = 0; i < XFEATURE_MAX; i++) {
1203 u64 mask = ((u64)1 << i);
1204
1205 if (hdr.xfeatures & mask) {
1206 void *dst = __raw_xsave_addr(xsave, i);
1207
1208 offset = xstate_offsets[i];
1209 size = xstate_sizes[i];
1210
1211 if (__copy_from_user(dst, ubuf + offset, size))
1212 return -EFAULT;
1213 }
1214 }
1215
1216 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1217 offset = offsetof(struct fxregs_state, mxcsr);
1218 size = MXCSR_AND_FLAGS_SIZE;
1219 if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1220 return -EFAULT;
1221 }
1222
1223 /*
1224 * The state that came in from userspace was user-state only.
1225 * Mask all the user states out of 'xfeatures':
1226 */
1227 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR;
1228
1229 /*
1230 * Add back in the features that came in from userspace:
1231 */
1232 xsave->header.xfeatures |= hdr.xfeatures;
1233
1234 return 0;
1235}
1236
1237#ifdef CONFIG_PROC_PID_ARCH_STATUS
1238/*
1239 * Report the amount of time elapsed in millisecond since last AVX512
1240 * use in the task.
1241 */
1242static void avx512_status(struct seq_file *m, struct task_struct *task)
1243{
1244 unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1245 long delta;
1246
1247 if (!timestamp) {
1248 /*
1249 * Report -1 if no AVX512 usage
1250 */
1251 delta = -1;
1252 } else {
1253 delta = (long)(jiffies - timestamp);
1254 /*
1255 * Cap to LONG_MAX if time difference > LONG_MAX
1256 */
1257 if (delta < 0)
1258 delta = LONG_MAX;
1259 delta = jiffies_to_msecs(delta);
1260 }
1261
1262 seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1263 seq_putc(m, '\n');
1264}
1265
1266/*
1267 * Report architecture specific information
1268 */
1269int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1270 struct pid *pid, struct task_struct *task)
1271{
1272 /*
1273 * Report AVX512 state if the processor and build option supported.
1274 */
1275 if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1276 avx512_status(m, task);
1277
1278 return 0;
1279}
1280#endif /* CONFIG_PROC_PID_ARCH_STATUS */
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * xsave/xrstor support.
4 *
5 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
6 */
7#include <linux/compat.h>
8#include <linux/cpu.h>
9#include <linux/mman.h>
10#include <linux/pkeys.h>
11#include <linux/seq_file.h>
12#include <linux/proc_fs.h>
13
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>
19
20#include <asm/tlbflush.h>
21#include <asm/cpufeature.h>
22
23/*
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.
27 */
28static const char *xfeature_names[] =
29{
30 "x87 floating point registers" ,
31 "SSE registers" ,
32 "AVX registers" ,
33 "MPX bounds registers" ,
34 "MPX CSR" ,
35 "AVX-512 opmask" ,
36 "AVX-512 Hi256" ,
37 "AVX-512 ZMM_Hi256" ,
38 "Processor Trace (unused)" ,
39 "Protection Keys User registers",
40 "unknown xstate feature" ,
41};
42
43static short xsave_cpuid_features[] __initdata = {
44 X86_FEATURE_FPU,
45 X86_FEATURE_XMM,
46 X86_FEATURE_AVX,
47 X86_FEATURE_MPX,
48 X86_FEATURE_MPX,
49 X86_FEATURE_AVX512F,
50 X86_FEATURE_AVX512F,
51 X86_FEATURE_AVX512F,
52 X86_FEATURE_INTEL_PT,
53 X86_FEATURE_PKU,
54};
55
56/*
57 * This represents the full set of bits that should ever be set in a kernel
58 * XSAVE buffer, both supervisor and user xstates.
59 */
60u64 xfeatures_mask_all __read_mostly;
61
62static unsigned int xstate_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
63static unsigned int xstate_sizes[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
64static unsigned int xstate_comp_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
65static unsigned int xstate_supervisor_only_offsets[XFEATURE_MAX] = { [ 0 ... XFEATURE_MAX - 1] = -1};
66
67/*
68 * The XSAVE area of kernel can be in standard or compacted format;
69 * it is always in standard format for user mode. This is the user
70 * mode standard format size used for signal and ptrace frames.
71 */
72unsigned int fpu_user_xstate_size;
73
74/*
75 * Return whether the system supports a given xfeature.
76 *
77 * Also return the name of the (most advanced) feature that the caller requested:
78 */
79int cpu_has_xfeatures(u64 xfeatures_needed, const char **feature_name)
80{
81 u64 xfeatures_missing = xfeatures_needed & ~xfeatures_mask_all;
82
83 if (unlikely(feature_name)) {
84 long xfeature_idx, max_idx;
85 u64 xfeatures_print;
86 /*
87 * So we use FLS here to be able to print the most advanced
88 * feature that was requested but is missing. So if a driver
89 * asks about "XFEATURE_MASK_SSE | XFEATURE_MASK_YMM" we'll print the
90 * missing AVX feature - this is the most informative message
91 * to users:
92 */
93 if (xfeatures_missing)
94 xfeatures_print = xfeatures_missing;
95 else
96 xfeatures_print = xfeatures_needed;
97
98 xfeature_idx = fls64(xfeatures_print)-1;
99 max_idx = ARRAY_SIZE(xfeature_names)-1;
100 xfeature_idx = min(xfeature_idx, max_idx);
101
102 *feature_name = xfeature_names[xfeature_idx];
103 }
104
105 if (xfeatures_missing)
106 return 0;
107
108 return 1;
109}
110EXPORT_SYMBOL_GPL(cpu_has_xfeatures);
111
112static bool xfeature_is_supervisor(int xfeature_nr)
113{
114 /*
115 * Extended State Enumeration Sub-leaves (EAX = 0DH, ECX = n, n > 1)
116 * returns ECX[0] set to (1) for a supervisor state, and cleared (0)
117 * for a user state.
118 */
119 u32 eax, ebx, ecx, edx;
120
121 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
122 return ecx & 1;
123}
124
125/*
126 * When executing XSAVEOPT (or other optimized XSAVE instructions), if
127 * a processor implementation detects that an FPU state component is still
128 * (or is again) in its initialized state, it may clear the corresponding
129 * bit in the header.xfeatures field, and can skip the writeout of registers
130 * to the corresponding memory layout.
131 *
132 * This means that when the bit is zero, the state component might still contain
133 * some previous - non-initialized register state.
134 *
135 * Before writing xstate information to user-space we sanitize those components,
136 * to always ensure that the memory layout of a feature will be in the init state
137 * if the corresponding header bit is zero. This is to ensure that user-space doesn't
138 * see some stale state in the memory layout during signal handling, debugging etc.
139 */
140void fpstate_sanitize_xstate(struct fpu *fpu)
141{
142 struct fxregs_state *fx = &fpu->state.fxsave;
143 int feature_bit;
144 u64 xfeatures;
145
146 if (!use_xsaveopt())
147 return;
148
149 xfeatures = fpu->state.xsave.header.xfeatures;
150
151 /*
152 * None of the feature bits are in init state. So nothing else
153 * to do for us, as the memory layout is up to date.
154 */
155 if ((xfeatures & xfeatures_mask_all) == xfeatures_mask_all)
156 return;
157
158 /*
159 * FP is in init state
160 */
161 if (!(xfeatures & XFEATURE_MASK_FP)) {
162 fx->cwd = 0x37f;
163 fx->swd = 0;
164 fx->twd = 0;
165 fx->fop = 0;
166 fx->rip = 0;
167 fx->rdp = 0;
168 memset(&fx->st_space[0], 0, 128);
169 }
170
171 /*
172 * SSE is in init state
173 */
174 if (!(xfeatures & XFEATURE_MASK_SSE))
175 memset(&fx->xmm_space[0], 0, 256);
176
177 /*
178 * First two features are FPU and SSE, which above we handled
179 * in a special way already:
180 */
181 feature_bit = 0x2;
182 xfeatures = (xfeatures_mask_user() & ~xfeatures) >> 2;
183
184 /*
185 * Update all the remaining memory layouts according to their
186 * standard xstate layout, if their header bit is in the init
187 * state:
188 */
189 while (xfeatures) {
190 if (xfeatures & 0x1) {
191 int offset = xstate_comp_offsets[feature_bit];
192 int size = xstate_sizes[feature_bit];
193
194 memcpy((void *)fx + offset,
195 (void *)&init_fpstate.xsave + offset,
196 size);
197 }
198
199 xfeatures >>= 1;
200 feature_bit++;
201 }
202}
203
204/*
205 * Enable the extended processor state save/restore feature.
206 * Called once per CPU onlining.
207 */
208void fpu__init_cpu_xstate(void)
209{
210 u64 unsup_bits;
211
212 if (!boot_cpu_has(X86_FEATURE_XSAVE) || !xfeatures_mask_all)
213 return;
214 /*
215 * Unsupported supervisor xstates should not be found in
216 * the xfeatures mask.
217 */
218 unsup_bits = xfeatures_mask_all & XFEATURE_MASK_SUPERVISOR_UNSUPPORTED;
219 WARN_ONCE(unsup_bits, "x86/fpu: Found unsupported supervisor xstates: 0x%llx\n",
220 unsup_bits);
221
222 xfeatures_mask_all &= ~XFEATURE_MASK_SUPERVISOR_UNSUPPORTED;
223
224 cr4_set_bits(X86_CR4_OSXSAVE);
225
226 /*
227 * XCR_XFEATURE_ENABLED_MASK (aka. XCR0) sets user features
228 * managed by XSAVE{C, OPT, S} and XRSTOR{S}. Only XSAVE user
229 * states can be set here.
230 */
231 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask_user());
232
233 /*
234 * MSR_IA32_XSS sets supervisor states managed by XSAVES.
235 */
236 if (boot_cpu_has(X86_FEATURE_XSAVES)) {
237 wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() |
238 xfeatures_mask_dynamic());
239 }
240}
241
242static bool xfeature_enabled(enum xfeature xfeature)
243{
244 return xfeatures_mask_all & BIT_ULL(xfeature);
245}
246
247/*
248 * Record the offsets and sizes of various xstates contained
249 * in the XSAVE state memory layout.
250 */
251static void __init setup_xstate_features(void)
252{
253 u32 eax, ebx, ecx, edx, i;
254 /* start at the beginnning of the "extended state" */
255 unsigned int last_good_offset = offsetof(struct xregs_state,
256 extended_state_area);
257 /*
258 * The FP xstates and SSE xstates are legacy states. They are always
259 * in the fixed offsets in the xsave area in either compacted form
260 * or standard form.
261 */
262 xstate_offsets[XFEATURE_FP] = 0;
263 xstate_sizes[XFEATURE_FP] = offsetof(struct fxregs_state,
264 xmm_space);
265
266 xstate_offsets[XFEATURE_SSE] = xstate_sizes[XFEATURE_FP];
267 xstate_sizes[XFEATURE_SSE] = sizeof_field(struct fxregs_state,
268 xmm_space);
269
270 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
271 if (!xfeature_enabled(i))
272 continue;
273
274 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
275
276 xstate_sizes[i] = eax;
277
278 /*
279 * If an xfeature is supervisor state, the offset in EBX is
280 * invalid, leave it to -1.
281 */
282 if (xfeature_is_supervisor(i))
283 continue;
284
285 xstate_offsets[i] = ebx;
286
287 /*
288 * In our xstate size checks, we assume that the highest-numbered
289 * xstate feature has the highest offset in the buffer. Ensure
290 * it does.
291 */
292 WARN_ONCE(last_good_offset > xstate_offsets[i],
293 "x86/fpu: misordered xstate at %d\n", last_good_offset);
294
295 last_good_offset = xstate_offsets[i];
296 }
297}
298
299static void __init print_xstate_feature(u64 xstate_mask)
300{
301 const char *feature_name;
302
303 if (cpu_has_xfeatures(xstate_mask, &feature_name))
304 pr_info("x86/fpu: Supporting XSAVE feature 0x%03Lx: '%s'\n", xstate_mask, feature_name);
305}
306
307/*
308 * Print out all the supported xstate features:
309 */
310static void __init print_xstate_features(void)
311{
312 print_xstate_feature(XFEATURE_MASK_FP);
313 print_xstate_feature(XFEATURE_MASK_SSE);
314 print_xstate_feature(XFEATURE_MASK_YMM);
315 print_xstate_feature(XFEATURE_MASK_BNDREGS);
316 print_xstate_feature(XFEATURE_MASK_BNDCSR);
317 print_xstate_feature(XFEATURE_MASK_OPMASK);
318 print_xstate_feature(XFEATURE_MASK_ZMM_Hi256);
319 print_xstate_feature(XFEATURE_MASK_Hi16_ZMM);
320 print_xstate_feature(XFEATURE_MASK_PKRU);
321}
322
323/*
324 * This check is important because it is easy to get XSTATE_*
325 * confused with XSTATE_BIT_*.
326 */
327#define CHECK_XFEATURE(nr) do { \
328 WARN_ON(nr < FIRST_EXTENDED_XFEATURE); \
329 WARN_ON(nr >= XFEATURE_MAX); \
330} while (0)
331
332/*
333 * We could cache this like xstate_size[], but we only use
334 * it here, so it would be a waste of space.
335 */
336static int xfeature_is_aligned(int xfeature_nr)
337{
338 u32 eax, ebx, ecx, edx;
339
340 CHECK_XFEATURE(xfeature_nr);
341
342 if (!xfeature_enabled(xfeature_nr)) {
343 WARN_ONCE(1, "Checking alignment of disabled xfeature %d\n",
344 xfeature_nr);
345 return 0;
346 }
347
348 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
349 /*
350 * The value returned by ECX[1] indicates the alignment
351 * of state component 'i' when the compacted format
352 * of the extended region of an XSAVE area is used:
353 */
354 return !!(ecx & 2);
355}
356
357/*
358 * This function sets up offsets and sizes of all extended states in
359 * xsave area. This supports both standard format and compacted format
360 * of the xsave area.
361 */
362static void __init setup_xstate_comp_offsets(void)
363{
364 unsigned int next_offset;
365 int i;
366
367 /*
368 * The FP xstates and SSE xstates are legacy states. They are always
369 * in the fixed offsets in the xsave area in either compacted form
370 * or standard form.
371 */
372 xstate_comp_offsets[XFEATURE_FP] = 0;
373 xstate_comp_offsets[XFEATURE_SSE] = offsetof(struct fxregs_state,
374 xmm_space);
375
376 if (!boot_cpu_has(X86_FEATURE_XSAVES)) {
377 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
378 if (xfeature_enabled(i))
379 xstate_comp_offsets[i] = xstate_offsets[i];
380 }
381 return;
382 }
383
384 next_offset = FXSAVE_SIZE + XSAVE_HDR_SIZE;
385
386 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
387 if (!xfeature_enabled(i))
388 continue;
389
390 if (xfeature_is_aligned(i))
391 next_offset = ALIGN(next_offset, 64);
392
393 xstate_comp_offsets[i] = next_offset;
394 next_offset += xstate_sizes[i];
395 }
396}
397
398/*
399 * Setup offsets of a supervisor-state-only XSAVES buffer:
400 *
401 * The offsets stored in xstate_comp_offsets[] only work for one specific
402 * value of the Requested Feature BitMap (RFBM). In cases where a different
403 * RFBM value is used, a different set of offsets is required. This set of
404 * offsets is for when RFBM=xfeatures_mask_supervisor().
405 */
406static void __init setup_supervisor_only_offsets(void)
407{
408 unsigned int next_offset;
409 int i;
410
411 next_offset = FXSAVE_SIZE + XSAVE_HDR_SIZE;
412
413 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
414 if (!xfeature_enabled(i) || !xfeature_is_supervisor(i))
415 continue;
416
417 if (xfeature_is_aligned(i))
418 next_offset = ALIGN(next_offset, 64);
419
420 xstate_supervisor_only_offsets[i] = next_offset;
421 next_offset += xstate_sizes[i];
422 }
423}
424
425/*
426 * Print out xstate component offsets and sizes
427 */
428static void __init print_xstate_offset_size(void)
429{
430 int i;
431
432 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
433 if (!xfeature_enabled(i))
434 continue;
435 pr_info("x86/fpu: xstate_offset[%d]: %4d, xstate_sizes[%d]: %4d\n",
436 i, xstate_comp_offsets[i], i, xstate_sizes[i]);
437 }
438}
439
440/*
441 * setup the xstate image representing the init state
442 */
443static void __init setup_init_fpu_buf(void)
444{
445 static int on_boot_cpu __initdata = 1;
446
447 WARN_ON_FPU(!on_boot_cpu);
448 on_boot_cpu = 0;
449
450 if (!boot_cpu_has(X86_FEATURE_XSAVE))
451 return;
452
453 setup_xstate_features();
454 print_xstate_features();
455
456 if (boot_cpu_has(X86_FEATURE_XSAVES))
457 init_fpstate.xsave.header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT |
458 xfeatures_mask_all;
459
460 /*
461 * Init all the features state with header.xfeatures being 0x0
462 */
463 copy_kernel_to_xregs_booting(&init_fpstate.xsave);
464
465 /*
466 * Dump the init state again. This is to identify the init state
467 * of any feature which is not represented by all zero's.
468 */
469 copy_xregs_to_kernel_booting(&init_fpstate.xsave);
470}
471
472static int xfeature_uncompacted_offset(int xfeature_nr)
473{
474 u32 eax, ebx, ecx, edx;
475
476 /*
477 * Only XSAVES supports supervisor states and it uses compacted
478 * format. Checking a supervisor state's uncompacted offset is
479 * an error.
480 */
481 if (XFEATURE_MASK_SUPERVISOR_ALL & BIT_ULL(xfeature_nr)) {
482 WARN_ONCE(1, "No fixed offset for xstate %d\n", xfeature_nr);
483 return -1;
484 }
485
486 CHECK_XFEATURE(xfeature_nr);
487 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
488 return ebx;
489}
490
491int xfeature_size(int xfeature_nr)
492{
493 u32 eax, ebx, ecx, edx;
494
495 CHECK_XFEATURE(xfeature_nr);
496 cpuid_count(XSTATE_CPUID, xfeature_nr, &eax, &ebx, &ecx, &edx);
497 return eax;
498}
499
500/*
501 * 'XSAVES' implies two different things:
502 * 1. saving of supervisor/system state
503 * 2. using the compacted format
504 *
505 * Use this function when dealing with the compacted format so
506 * that it is obvious which aspect of 'XSAVES' is being handled
507 * by the calling code.
508 */
509int using_compacted_format(void)
510{
511 return boot_cpu_has(X86_FEATURE_XSAVES);
512}
513
514/* Validate an xstate header supplied by userspace (ptrace or sigreturn) */
515int validate_user_xstate_header(const struct xstate_header *hdr)
516{
517 /* No unknown or supervisor features may be set */
518 if (hdr->xfeatures & ~xfeatures_mask_user())
519 return -EINVAL;
520
521 /* Userspace must use the uncompacted format */
522 if (hdr->xcomp_bv)
523 return -EINVAL;
524
525 /*
526 * If 'reserved' is shrunken to add a new field, make sure to validate
527 * that new field here!
528 */
529 BUILD_BUG_ON(sizeof(hdr->reserved) != 48);
530
531 /* No reserved bits may be set */
532 if (memchr_inv(hdr->reserved, 0, sizeof(hdr->reserved)))
533 return -EINVAL;
534
535 return 0;
536}
537
538static void __xstate_dump_leaves(void)
539{
540 int i;
541 u32 eax, ebx, ecx, edx;
542 static int should_dump = 1;
543
544 if (!should_dump)
545 return;
546 should_dump = 0;
547 /*
548 * Dump out a few leaves past the ones that we support
549 * just in case there are some goodies up there
550 */
551 for (i = 0; i < XFEATURE_MAX + 10; i++) {
552 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
553 pr_warn("CPUID[%02x, %02x]: eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
554 XSTATE_CPUID, i, eax, ebx, ecx, edx);
555 }
556}
557
558#define XSTATE_WARN_ON(x) do { \
559 if (WARN_ONCE(x, "XSAVE consistency problem, dumping leaves")) { \
560 __xstate_dump_leaves(); \
561 } \
562} while (0)
563
564#define XCHECK_SZ(sz, nr, nr_macro, __struct) do { \
565 if ((nr == nr_macro) && \
566 WARN_ONCE(sz != sizeof(__struct), \
567 "%s: struct is %zu bytes, cpu state %d bytes\n", \
568 __stringify(nr_macro), sizeof(__struct), sz)) { \
569 __xstate_dump_leaves(); \
570 } \
571} while (0)
572
573/*
574 * We have a C struct for each 'xstate'. We need to ensure
575 * that our software representation matches what the CPU
576 * tells us about the state's size.
577 */
578static void check_xstate_against_struct(int nr)
579{
580 /*
581 * Ask the CPU for the size of the state.
582 */
583 int sz = xfeature_size(nr);
584 /*
585 * Match each CPU state with the corresponding software
586 * structure.
587 */
588 XCHECK_SZ(sz, nr, XFEATURE_YMM, struct ymmh_struct);
589 XCHECK_SZ(sz, nr, XFEATURE_BNDREGS, struct mpx_bndreg_state);
590 XCHECK_SZ(sz, nr, XFEATURE_BNDCSR, struct mpx_bndcsr_state);
591 XCHECK_SZ(sz, nr, XFEATURE_OPMASK, struct avx_512_opmask_state);
592 XCHECK_SZ(sz, nr, XFEATURE_ZMM_Hi256, struct avx_512_zmm_uppers_state);
593 XCHECK_SZ(sz, nr, XFEATURE_Hi16_ZMM, struct avx_512_hi16_state);
594 XCHECK_SZ(sz, nr, XFEATURE_PKRU, struct pkru_state);
595
596 /*
597 * Make *SURE* to add any feature numbers in below if
598 * there are "holes" in the xsave state component
599 * numbers.
600 */
601 if ((nr < XFEATURE_YMM) ||
602 (nr >= XFEATURE_MAX) ||
603 (nr == XFEATURE_PT_UNIMPLEMENTED_SO_FAR) ||
604 ((nr >= XFEATURE_RSRVD_COMP_10) && (nr <= XFEATURE_LBR))) {
605 WARN_ONCE(1, "no structure for xstate: %d\n", nr);
606 XSTATE_WARN_ON(1);
607 }
608}
609
610/*
611 * This essentially double-checks what the cpu told us about
612 * how large the XSAVE buffer needs to be. We are recalculating
613 * it to be safe.
614 *
615 * Dynamic XSAVE features allocate their own buffers and are not
616 * covered by these checks. Only the size of the buffer for task->fpu
617 * is checked here.
618 */
619static void do_extra_xstate_size_checks(void)
620{
621 int paranoid_xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
622 int i;
623
624 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
625 if (!xfeature_enabled(i))
626 continue;
627
628 check_xstate_against_struct(i);
629 /*
630 * Supervisor state components can be managed only by
631 * XSAVES, which is compacted-format only.
632 */
633 if (!using_compacted_format())
634 XSTATE_WARN_ON(xfeature_is_supervisor(i));
635
636 /* Align from the end of the previous feature */
637 if (xfeature_is_aligned(i))
638 paranoid_xstate_size = ALIGN(paranoid_xstate_size, 64);
639 /*
640 * The offset of a given state in the non-compacted
641 * format is given to us in a CPUID leaf. We check
642 * them for being ordered (increasing offsets) in
643 * setup_xstate_features().
644 */
645 if (!using_compacted_format())
646 paranoid_xstate_size = xfeature_uncompacted_offset(i);
647 /*
648 * The compacted-format offset always depends on where
649 * the previous state ended.
650 */
651 paranoid_xstate_size += xfeature_size(i);
652 }
653 XSTATE_WARN_ON(paranoid_xstate_size != fpu_kernel_xstate_size);
654}
655
656
657/*
658 * Get total size of enabled xstates in XCR0 | IA32_XSS.
659 *
660 * Note the SDM's wording here. "sub-function 0" only enumerates
661 * the size of the *user* states. If we use it to size a buffer
662 * that we use 'XSAVES' on, we could potentially overflow the
663 * buffer because 'XSAVES' saves system states too.
664 */
665static unsigned int __init get_xsaves_size(void)
666{
667 unsigned int eax, ebx, ecx, edx;
668 /*
669 * - CPUID function 0DH, sub-function 1:
670 * EBX enumerates the size (in bytes) required by
671 * the XSAVES instruction for an XSAVE area
672 * containing all the state components
673 * corresponding to bits currently set in
674 * XCR0 | IA32_XSS.
675 */
676 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
677 return ebx;
678}
679
680/*
681 * Get the total size of the enabled xstates without the dynamic supervisor
682 * features.
683 */
684static unsigned int __init get_xsaves_size_no_dynamic(void)
685{
686 u64 mask = xfeatures_mask_dynamic();
687 unsigned int size;
688
689 if (!mask)
690 return get_xsaves_size();
691
692 /* Disable dynamic features. */
693 wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor());
694
695 /*
696 * Ask the hardware what size is required of the buffer.
697 * This is the size required for the task->fpu buffer.
698 */
699 size = get_xsaves_size();
700
701 /* Re-enable dynamic features so XSAVES will work on them again. */
702 wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() | mask);
703
704 return size;
705}
706
707static unsigned int __init get_xsave_size(void)
708{
709 unsigned int eax, ebx, ecx, edx;
710 /*
711 * - CPUID function 0DH, sub-function 0:
712 * EBX enumerates the size (in bytes) required by
713 * the XSAVE instruction for an XSAVE area
714 * containing all the *user* state components
715 * corresponding to bits currently set in XCR0.
716 */
717 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
718 return ebx;
719}
720
721/*
722 * Will the runtime-enumerated 'xstate_size' fit in the init
723 * task's statically-allocated buffer?
724 */
725static bool is_supported_xstate_size(unsigned int test_xstate_size)
726{
727 if (test_xstate_size <= sizeof(union fpregs_state))
728 return true;
729
730 pr_warn("x86/fpu: xstate buffer too small (%zu < %d), disabling xsave\n",
731 sizeof(union fpregs_state), test_xstate_size);
732 return false;
733}
734
735static int __init init_xstate_size(void)
736{
737 /* Recompute the context size for enabled features: */
738 unsigned int possible_xstate_size;
739 unsigned int xsave_size;
740
741 xsave_size = get_xsave_size();
742
743 if (boot_cpu_has(X86_FEATURE_XSAVES))
744 possible_xstate_size = get_xsaves_size_no_dynamic();
745 else
746 possible_xstate_size = xsave_size;
747
748 /* Ensure we have the space to store all enabled: */
749 if (!is_supported_xstate_size(possible_xstate_size))
750 return -EINVAL;
751
752 /*
753 * The size is OK, we are definitely going to use xsave,
754 * make it known to the world that we need more space.
755 */
756 fpu_kernel_xstate_size = possible_xstate_size;
757 do_extra_xstate_size_checks();
758
759 /*
760 * User space is always in standard format.
761 */
762 fpu_user_xstate_size = xsave_size;
763 return 0;
764}
765
766/*
767 * We enabled the XSAVE hardware, but something went wrong and
768 * we can not use it. Disable it.
769 */
770static void fpu__init_disable_system_xstate(void)
771{
772 xfeatures_mask_all = 0;
773 cr4_clear_bits(X86_CR4_OSXSAVE);
774 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
775}
776
777/*
778 * Enable and initialize the xsave feature.
779 * Called once per system bootup.
780 */
781void __init fpu__init_system_xstate(void)
782{
783 unsigned int eax, ebx, ecx, edx;
784 static int on_boot_cpu __initdata = 1;
785 int err;
786 int i;
787
788 WARN_ON_FPU(!on_boot_cpu);
789 on_boot_cpu = 0;
790
791 if (!boot_cpu_has(X86_FEATURE_FPU)) {
792 pr_info("x86/fpu: No FPU detected\n");
793 return;
794 }
795
796 if (!boot_cpu_has(X86_FEATURE_XSAVE)) {
797 pr_info("x86/fpu: x87 FPU will use %s\n",
798 boot_cpu_has(X86_FEATURE_FXSR) ? "FXSAVE" : "FSAVE");
799 return;
800 }
801
802 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
803 WARN_ON_FPU(1);
804 return;
805 }
806
807 /*
808 * Find user xstates supported by the processor.
809 */
810 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
811 xfeatures_mask_all = eax + ((u64)edx << 32);
812
813 /*
814 * Find supervisor xstates supported by the processor.
815 */
816 cpuid_count(XSTATE_CPUID, 1, &eax, &ebx, &ecx, &edx);
817 xfeatures_mask_all |= ecx + ((u64)edx << 32);
818
819 if ((xfeatures_mask_user() & XFEATURE_MASK_FPSSE) != XFEATURE_MASK_FPSSE) {
820 /*
821 * This indicates that something really unexpected happened
822 * with the enumeration. Disable XSAVE and try to continue
823 * booting without it. This is too early to BUG().
824 */
825 pr_err("x86/fpu: FP/SSE not present amongst the CPU's xstate features: 0x%llx.\n",
826 xfeatures_mask_all);
827 goto out_disable;
828 }
829
830 /*
831 * Clear XSAVE features that are disabled in the normal CPUID.
832 */
833 for (i = 0; i < ARRAY_SIZE(xsave_cpuid_features); i++) {
834 if (!boot_cpu_has(xsave_cpuid_features[i]))
835 xfeatures_mask_all &= ~BIT_ULL(i);
836 }
837
838 xfeatures_mask_all &= fpu__get_supported_xfeatures_mask();
839
840 /* Enable xstate instructions to be able to continue with initialization: */
841 fpu__init_cpu_xstate();
842 err = init_xstate_size();
843 if (err)
844 goto out_disable;
845
846 /*
847 * Update info used for ptrace frames; use standard-format size and no
848 * supervisor xstates:
849 */
850 update_regset_xstate_info(fpu_user_xstate_size, xfeatures_mask_user());
851
852 fpu__init_prepare_fx_sw_frame();
853 setup_init_fpu_buf();
854 setup_xstate_comp_offsets();
855 setup_supervisor_only_offsets();
856 print_xstate_offset_size();
857
858 pr_info("x86/fpu: Enabled xstate features 0x%llx, context size is %d bytes, using '%s' format.\n",
859 xfeatures_mask_all,
860 fpu_kernel_xstate_size,
861 boot_cpu_has(X86_FEATURE_XSAVES) ? "compacted" : "standard");
862 return;
863
864out_disable:
865 /* something went wrong, try to boot without any XSAVE support */
866 fpu__init_disable_system_xstate();
867}
868
869/*
870 * Restore minimal FPU state after suspend:
871 */
872void fpu__resume_cpu(void)
873{
874 /*
875 * Restore XCR0 on xsave capable CPUs:
876 */
877 if (boot_cpu_has(X86_FEATURE_XSAVE))
878 xsetbv(XCR_XFEATURE_ENABLED_MASK, xfeatures_mask_user());
879
880 /*
881 * Restore IA32_XSS. The same CPUID bit enumerates support
882 * of XSAVES and MSR_IA32_XSS.
883 */
884 if (boot_cpu_has(X86_FEATURE_XSAVES)) {
885 wrmsrl(MSR_IA32_XSS, xfeatures_mask_supervisor() |
886 xfeatures_mask_dynamic());
887 }
888}
889
890/*
891 * Given an xstate feature nr, calculate where in the xsave
892 * buffer the state is. Callers should ensure that the buffer
893 * is valid.
894 */
895static void *__raw_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
896{
897 if (!xfeature_enabled(xfeature_nr)) {
898 WARN_ON_FPU(1);
899 return NULL;
900 }
901
902 return (void *)xsave + xstate_comp_offsets[xfeature_nr];
903}
904/*
905 * Given the xsave area and a state inside, this function returns the
906 * address of the state.
907 *
908 * This is the API that is called to get xstate address in either
909 * standard format or compacted format of xsave area.
910 *
911 * Note that if there is no data for the field in the xsave buffer
912 * this will return NULL.
913 *
914 * Inputs:
915 * xstate: the thread's storage area for all FPU data
916 * xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
917 * XFEATURE_SSE, etc...)
918 * Output:
919 * address of the state in the xsave area, or NULL if the
920 * field is not present in the xsave buffer.
921 */
922void *get_xsave_addr(struct xregs_state *xsave, int xfeature_nr)
923{
924 /*
925 * Do we even *have* xsave state?
926 */
927 if (!boot_cpu_has(X86_FEATURE_XSAVE))
928 return NULL;
929
930 /*
931 * We should not ever be requesting features that we
932 * have not enabled.
933 */
934 WARN_ONCE(!(xfeatures_mask_all & BIT_ULL(xfeature_nr)),
935 "get of unsupported state");
936 /*
937 * This assumes the last 'xsave*' instruction to
938 * have requested that 'xfeature_nr' be saved.
939 * If it did not, we might be seeing and old value
940 * of the field in the buffer.
941 *
942 * This can happen because the last 'xsave' did not
943 * request that this feature be saved (unlikely)
944 * or because the "init optimization" caused it
945 * to not be saved.
946 */
947 if (!(xsave->header.xfeatures & BIT_ULL(xfeature_nr)))
948 return NULL;
949
950 return __raw_xsave_addr(xsave, xfeature_nr);
951}
952EXPORT_SYMBOL_GPL(get_xsave_addr);
953
954/*
955 * This wraps up the common operations that need to occur when retrieving
956 * data from xsave state. It first ensures that the current task was
957 * using the FPU and retrieves the data in to a buffer. It then calculates
958 * the offset of the requested field in the buffer.
959 *
960 * This function is safe to call whether the FPU is in use or not.
961 *
962 * Note that this only works on the current task.
963 *
964 * Inputs:
965 * @xfeature_nr: state which is defined in xsave.h (e.g. XFEATURE_FP,
966 * XFEATURE_SSE, etc...)
967 * Output:
968 * address of the state in the xsave area or NULL if the state
969 * is not present or is in its 'init state'.
970 */
971const void *get_xsave_field_ptr(int xfeature_nr)
972{
973 struct fpu *fpu = ¤t->thread.fpu;
974
975 /*
976 * fpu__save() takes the CPU's xstate registers
977 * and saves them off to the 'fpu memory buffer.
978 */
979 fpu__save(fpu);
980
981 return get_xsave_addr(&fpu->state.xsave, xfeature_nr);
982}
983
984#ifdef CONFIG_ARCH_HAS_PKEYS
985
986/*
987 * This will go out and modify PKRU register to set the access
988 * rights for @pkey to @init_val.
989 */
990int arch_set_user_pkey_access(struct task_struct *tsk, int pkey,
991 unsigned long init_val)
992{
993 u32 old_pkru;
994 int pkey_shift = (pkey * PKRU_BITS_PER_PKEY);
995 u32 new_pkru_bits = 0;
996
997 /*
998 * This check implies XSAVE support. OSPKE only gets
999 * set if we enable XSAVE and we enable PKU in XCR0.
1000 */
1001 if (!boot_cpu_has(X86_FEATURE_OSPKE))
1002 return -EINVAL;
1003
1004 /*
1005 * This code should only be called with valid 'pkey'
1006 * values originating from in-kernel users. Complain
1007 * if a bad value is observed.
1008 */
1009 WARN_ON_ONCE(pkey >= arch_max_pkey());
1010
1011 /* Set the bits we need in PKRU: */
1012 if (init_val & PKEY_DISABLE_ACCESS)
1013 new_pkru_bits |= PKRU_AD_BIT;
1014 if (init_val & PKEY_DISABLE_WRITE)
1015 new_pkru_bits |= PKRU_WD_BIT;
1016
1017 /* Shift the bits in to the correct place in PKRU for pkey: */
1018 new_pkru_bits <<= pkey_shift;
1019
1020 /* Get old PKRU and mask off any old bits in place: */
1021 old_pkru = read_pkru();
1022 old_pkru &= ~((PKRU_AD_BIT|PKRU_WD_BIT) << pkey_shift);
1023
1024 /* Write old part along with new part: */
1025 write_pkru(old_pkru | new_pkru_bits);
1026
1027 return 0;
1028}
1029#endif /* ! CONFIG_ARCH_HAS_PKEYS */
1030
1031/*
1032 * Weird legacy quirk: SSE and YMM states store information in the
1033 * MXCSR and MXCSR_FLAGS fields of the FP area. That means if the FP
1034 * area is marked as unused in the xfeatures header, we need to copy
1035 * MXCSR and MXCSR_FLAGS if either SSE or YMM are in use.
1036 */
1037static inline bool xfeatures_mxcsr_quirk(u64 xfeatures)
1038{
1039 if (!(xfeatures & (XFEATURE_MASK_SSE|XFEATURE_MASK_YMM)))
1040 return false;
1041
1042 if (xfeatures & XFEATURE_MASK_FP)
1043 return false;
1044
1045 return true;
1046}
1047
1048static void fill_gap(struct membuf *to, unsigned *last, unsigned offset)
1049{
1050 if (*last >= offset)
1051 return;
1052 membuf_write(to, (void *)&init_fpstate.xsave + *last, offset - *last);
1053 *last = offset;
1054}
1055
1056static void copy_part(struct membuf *to, unsigned *last, unsigned offset,
1057 unsigned size, void *from)
1058{
1059 fill_gap(to, last, offset);
1060 membuf_write(to, from, size);
1061 *last = offset + size;
1062}
1063
1064/*
1065 * Convert from kernel XSAVES compacted format to standard format and copy
1066 * to a kernel-space ptrace buffer.
1067 *
1068 * It supports partial copy but pos always starts from zero. This is called
1069 * from xstateregs_get() and there we check the CPU has XSAVES.
1070 */
1071void copy_xstate_to_kernel(struct membuf to, struct xregs_state *xsave)
1072{
1073 struct xstate_header header;
1074 const unsigned off_mxcsr = offsetof(struct fxregs_state, mxcsr);
1075 unsigned size = to.left;
1076 unsigned last = 0;
1077 int i;
1078
1079 /*
1080 * The destination is a ptrace buffer; we put in only user xstates:
1081 */
1082 memset(&header, 0, sizeof(header));
1083 header.xfeatures = xsave->header.xfeatures;
1084 header.xfeatures &= xfeatures_mask_user();
1085
1086 if (header.xfeatures & XFEATURE_MASK_FP)
1087 copy_part(&to, &last, 0, off_mxcsr, &xsave->i387);
1088 if (header.xfeatures & (XFEATURE_MASK_SSE | XFEATURE_MASK_YMM))
1089 copy_part(&to, &last, off_mxcsr,
1090 MXCSR_AND_FLAGS_SIZE, &xsave->i387.mxcsr);
1091 if (header.xfeatures & XFEATURE_MASK_FP)
1092 copy_part(&to, &last, offsetof(struct fxregs_state, st_space),
1093 128, &xsave->i387.st_space);
1094 if (header.xfeatures & XFEATURE_MASK_SSE)
1095 copy_part(&to, &last, xstate_offsets[XFEATURE_SSE],
1096 256, &xsave->i387.xmm_space);
1097 /*
1098 * Fill xsave->i387.sw_reserved value for ptrace frame:
1099 */
1100 copy_part(&to, &last, offsetof(struct fxregs_state, sw_reserved),
1101 48, xstate_fx_sw_bytes);
1102 /*
1103 * Copy xregs_state->header:
1104 */
1105 copy_part(&to, &last, offsetof(struct xregs_state, header),
1106 sizeof(header), &header);
1107
1108 for (i = FIRST_EXTENDED_XFEATURE; i < XFEATURE_MAX; i++) {
1109 /*
1110 * Copy only in-use xstates:
1111 */
1112 if ((header.xfeatures >> i) & 1) {
1113 void *src = __raw_xsave_addr(xsave, i);
1114
1115 copy_part(&to, &last, xstate_offsets[i],
1116 xstate_sizes[i], src);
1117 }
1118
1119 }
1120 fill_gap(&to, &last, size);
1121}
1122
1123/*
1124 * Convert from a ptrace standard-format kernel buffer to kernel XSAVES format
1125 * and copy to the target thread. This is called from xstateregs_set().
1126 */
1127int copy_kernel_to_xstate(struct xregs_state *xsave, const void *kbuf)
1128{
1129 unsigned int offset, size;
1130 int i;
1131 struct xstate_header hdr;
1132
1133 offset = offsetof(struct xregs_state, header);
1134 size = sizeof(hdr);
1135
1136 memcpy(&hdr, kbuf + offset, size);
1137
1138 if (validate_user_xstate_header(&hdr))
1139 return -EINVAL;
1140
1141 for (i = 0; i < XFEATURE_MAX; i++) {
1142 u64 mask = ((u64)1 << i);
1143
1144 if (hdr.xfeatures & mask) {
1145 void *dst = __raw_xsave_addr(xsave, i);
1146
1147 offset = xstate_offsets[i];
1148 size = xstate_sizes[i];
1149
1150 memcpy(dst, kbuf + offset, size);
1151 }
1152 }
1153
1154 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1155 offset = offsetof(struct fxregs_state, mxcsr);
1156 size = MXCSR_AND_FLAGS_SIZE;
1157 memcpy(&xsave->i387.mxcsr, kbuf + offset, size);
1158 }
1159
1160 /*
1161 * The state that came in from userspace was user-state only.
1162 * Mask all the user states out of 'xfeatures':
1163 */
1164 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1165
1166 /*
1167 * Add back in the features that came in from userspace:
1168 */
1169 xsave->header.xfeatures |= hdr.xfeatures;
1170
1171 return 0;
1172}
1173
1174/*
1175 * Convert from a ptrace or sigreturn standard-format user-space buffer to
1176 * kernel XSAVES format and copy to the target thread. This is called from
1177 * xstateregs_set(), as well as potentially from the sigreturn() and
1178 * rt_sigreturn() system calls.
1179 */
1180int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)
1181{
1182 unsigned int offset, size;
1183 int i;
1184 struct xstate_header hdr;
1185
1186 offset = offsetof(struct xregs_state, header);
1187 size = sizeof(hdr);
1188
1189 if (__copy_from_user(&hdr, ubuf + offset, size))
1190 return -EFAULT;
1191
1192 if (validate_user_xstate_header(&hdr))
1193 return -EINVAL;
1194
1195 for (i = 0; i < XFEATURE_MAX; i++) {
1196 u64 mask = ((u64)1 << i);
1197
1198 if (hdr.xfeatures & mask) {
1199 void *dst = __raw_xsave_addr(xsave, i);
1200
1201 offset = xstate_offsets[i];
1202 size = xstate_sizes[i];
1203
1204 if (__copy_from_user(dst, ubuf + offset, size))
1205 return -EFAULT;
1206 }
1207 }
1208
1209 if (xfeatures_mxcsr_quirk(hdr.xfeatures)) {
1210 offset = offsetof(struct fxregs_state, mxcsr);
1211 size = MXCSR_AND_FLAGS_SIZE;
1212 if (__copy_from_user(&xsave->i387.mxcsr, ubuf + offset, size))
1213 return -EFAULT;
1214 }
1215
1216 /*
1217 * The state that came in from userspace was user-state only.
1218 * Mask all the user states out of 'xfeatures':
1219 */
1220 xsave->header.xfeatures &= XFEATURE_MASK_SUPERVISOR_ALL;
1221
1222 /*
1223 * Add back in the features that came in from userspace:
1224 */
1225 xsave->header.xfeatures |= hdr.xfeatures;
1226
1227 return 0;
1228}
1229
1230/*
1231 * Save only supervisor states to the kernel buffer. This blows away all
1232 * old states, and is intended to be used only in __fpu__restore_sig(), where
1233 * user states are restored from the user buffer.
1234 */
1235void copy_supervisor_to_kernel(struct xregs_state *xstate)
1236{
1237 struct xstate_header *header;
1238 u64 max_bit, min_bit;
1239 u32 lmask, hmask;
1240 int err, i;
1241
1242 if (WARN_ON(!boot_cpu_has(X86_FEATURE_XSAVES)))
1243 return;
1244
1245 if (!xfeatures_mask_supervisor())
1246 return;
1247
1248 max_bit = __fls(xfeatures_mask_supervisor());
1249 min_bit = __ffs(xfeatures_mask_supervisor());
1250
1251 lmask = xfeatures_mask_supervisor();
1252 hmask = xfeatures_mask_supervisor() >> 32;
1253 XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
1254
1255 /* We should never fault when copying to a kernel buffer: */
1256 if (WARN_ON_FPU(err))
1257 return;
1258
1259 /*
1260 * At this point, the buffer has only supervisor states and must be
1261 * converted back to normal kernel format.
1262 */
1263 header = &xstate->header;
1264 header->xcomp_bv |= xfeatures_mask_all;
1265
1266 /*
1267 * This only moves states up in the buffer. Start with
1268 * the last state and move backwards so that states are
1269 * not overwritten until after they are moved. Note:
1270 * memmove() allows overlapping src/dst buffers.
1271 */
1272 for (i = max_bit; i >= min_bit; i--) {
1273 u8 *xbuf = (u8 *)xstate;
1274
1275 if (!((header->xfeatures >> i) & 1))
1276 continue;
1277
1278 /* Move xfeature 'i' into its normal location */
1279 memmove(xbuf + xstate_comp_offsets[i],
1280 xbuf + xstate_supervisor_only_offsets[i],
1281 xstate_sizes[i]);
1282 }
1283}
1284
1285/**
1286 * copy_dynamic_supervisor_to_kernel() - Save dynamic supervisor states to
1287 * an xsave area
1288 * @xstate: A pointer to an xsave area
1289 * @mask: Represent the dynamic supervisor features saved into the xsave area
1290 *
1291 * Only the dynamic supervisor states sets in the mask are saved into the xsave
1292 * area (See the comment in XFEATURE_MASK_DYNAMIC for the details of dynamic
1293 * supervisor feature). Besides the dynamic supervisor states, the legacy
1294 * region and XSAVE header are also saved into the xsave area. The supervisor
1295 * features in the XFEATURE_MASK_SUPERVISOR_SUPPORTED and
1296 * XFEATURE_MASK_SUPERVISOR_UNSUPPORTED are not saved.
1297 *
1298 * The xsave area must be 64-bytes aligned.
1299 */
1300void copy_dynamic_supervisor_to_kernel(struct xregs_state *xstate, u64 mask)
1301{
1302 u64 dynamic_mask = xfeatures_mask_dynamic() & mask;
1303 u32 lmask, hmask;
1304 int err;
1305
1306 if (WARN_ON_FPU(!boot_cpu_has(X86_FEATURE_XSAVES)))
1307 return;
1308
1309 if (WARN_ON_FPU(!dynamic_mask))
1310 return;
1311
1312 lmask = dynamic_mask;
1313 hmask = dynamic_mask >> 32;
1314
1315 XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
1316
1317 /* Should never fault when copying to a kernel buffer */
1318 WARN_ON_FPU(err);
1319}
1320
1321/**
1322 * copy_kernel_to_dynamic_supervisor() - Restore dynamic supervisor states from
1323 * an xsave area
1324 * @xstate: A pointer to an xsave area
1325 * @mask: Represent the dynamic supervisor features restored from the xsave area
1326 *
1327 * Only the dynamic supervisor states sets in the mask are restored from the
1328 * xsave area (See the comment in XFEATURE_MASK_DYNAMIC for the details of
1329 * dynamic supervisor feature). Besides the dynamic supervisor states, the
1330 * legacy region and XSAVE header are also restored from the xsave area. The
1331 * supervisor features in the XFEATURE_MASK_SUPERVISOR_SUPPORTED and
1332 * XFEATURE_MASK_SUPERVISOR_UNSUPPORTED are not restored.
1333 *
1334 * The xsave area must be 64-bytes aligned.
1335 */
1336void copy_kernel_to_dynamic_supervisor(struct xregs_state *xstate, u64 mask)
1337{
1338 u64 dynamic_mask = xfeatures_mask_dynamic() & mask;
1339 u32 lmask, hmask;
1340 int err;
1341
1342 if (WARN_ON_FPU(!boot_cpu_has(X86_FEATURE_XSAVES)))
1343 return;
1344
1345 if (WARN_ON_FPU(!dynamic_mask))
1346 return;
1347
1348 lmask = dynamic_mask;
1349 hmask = dynamic_mask >> 32;
1350
1351 XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
1352
1353 /* Should never fault when copying from a kernel buffer */
1354 WARN_ON_FPU(err);
1355}
1356
1357#ifdef CONFIG_PROC_PID_ARCH_STATUS
1358/*
1359 * Report the amount of time elapsed in millisecond since last AVX512
1360 * use in the task.
1361 */
1362static void avx512_status(struct seq_file *m, struct task_struct *task)
1363{
1364 unsigned long timestamp = READ_ONCE(task->thread.fpu.avx512_timestamp);
1365 long delta;
1366
1367 if (!timestamp) {
1368 /*
1369 * Report -1 if no AVX512 usage
1370 */
1371 delta = -1;
1372 } else {
1373 delta = (long)(jiffies - timestamp);
1374 /*
1375 * Cap to LONG_MAX if time difference > LONG_MAX
1376 */
1377 if (delta < 0)
1378 delta = LONG_MAX;
1379 delta = jiffies_to_msecs(delta);
1380 }
1381
1382 seq_put_decimal_ll(m, "AVX512_elapsed_ms:\t", delta);
1383 seq_putc(m, '\n');
1384}
1385
1386/*
1387 * Report architecture specific information
1388 */
1389int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
1390 struct pid *pid, struct task_struct *task)
1391{
1392 /*
1393 * Report AVX512 state if the processor and build option supported.
1394 */
1395 if (cpu_feature_enabled(X86_FEATURE_AVX512F))
1396 avx512_status(m, task);
1397
1398 return 0;
1399}
1400#endif /* CONFIG_PROC_PID_ARCH_STATUS */